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 #define SV_NAME "slp-vectorizer"
19 #define DEBUG_TYPE "SLP"
21 #include "llvm/Transforms/Vectorize.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/IR/Verifier.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
50 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
51 cl::desc("Only vectorize if you gain more than this "
55 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
56 cl::desc("Attempt to vectorize horizontal reductions"));
58 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
59 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
61 "Attempt to vectorize horizontal reductions feeding into a store"));
65 static const unsigned MinVecRegSize = 128;
67 static const unsigned RecursionMaxDepth = 12;
69 /// A helper class for numbering instructions in multiple blocks.
70 /// Numbers start at zero for each basic block.
71 struct BlockNumbering {
73 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
75 BlockNumbering() : BB(0), Valid(false) {}
77 void numberInstructions() {
81 // Number the instructions in the block.
82 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
84 InstrVec.push_back(it);
85 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
90 int getIndex(Instruction *I) {
91 assert(I->getParent() == BB && "Invalid instruction");
94 assert(InstrIdx.count(I) && "Unknown instruction");
98 Instruction *getInstruction(unsigned loc) {
100 numberInstructions();
101 assert(InstrVec.size() > loc && "Invalid Index");
102 return InstrVec[loc];
105 void forget() { Valid = false; }
108 /// The block we are numbering.
110 /// Is the block numbered.
112 /// Maps instructions to numbers and back.
113 SmallDenseMap<Instruction *, int> InstrIdx;
114 /// Maps integers to Instructions.
115 SmallVector<Instruction *, 32> InstrVec;
118 /// \returns the parent basic block if all of the instructions in \p VL
119 /// are in the same block or null otherwise.
120 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
121 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
124 BasicBlock *BB = I0->getParent();
125 for (int i = 1, e = VL.size(); i < e; i++) {
126 Instruction *I = dyn_cast<Instruction>(VL[i]);
130 if (BB != I->getParent())
136 /// \returns True if all of the values in \p VL are constants.
137 static bool allConstant(ArrayRef<Value *> VL) {
138 for (unsigned i = 0, e = VL.size(); i < e; ++i)
139 if (!isa<Constant>(VL[i]))
144 /// \returns True if all of the values in \p VL are identical.
145 static bool isSplat(ArrayRef<Value *> VL) {
146 for (unsigned i = 1, e = VL.size(); i < e; ++i)
152 /// \returns The opcode if all of the Instructions in \p VL have the same
154 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
155 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
158 unsigned Opcode = I0->getOpcode();
159 for (int i = 1, e = VL.size(); i < e; i++) {
160 Instruction *I = dyn_cast<Instruction>(VL[i]);
161 if (!I || Opcode != I->getOpcode())
167 /// \returns \p I after propagating metadata from \p VL.
168 static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) {
169 Instruction *I0 = cast<Instruction>(VL[0]);
170 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
171 I0->getAllMetadataOtherThanDebugLoc(Metadata);
173 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) {
174 unsigned Kind = Metadata[i].first;
175 MDNode *MD = Metadata[i].second;
177 for (int i = 1, e = VL.size(); MD && i != e; i++) {
178 Instruction *I = cast<Instruction>(VL[i]);
179 MDNode *IMD = I->getMetadata(Kind);
183 MD = 0; // Remove unknown metadata
185 case LLVMContext::MD_tbaa:
186 MD = MDNode::getMostGenericTBAA(MD, IMD);
188 case LLVMContext::MD_fpmath:
189 MD = MDNode::getMostGenericFPMath(MD, IMD);
193 I->setMetadata(Kind, MD);
198 /// \returns The type that all of the values in \p VL have or null if there
199 /// are different types.
200 static Type* getSameType(ArrayRef<Value *> VL) {
201 Type *Ty = VL[0]->getType();
202 for (int i = 1, e = VL.size(); i < e; i++)
203 if (VL[i]->getType() != Ty)
209 /// \returns True if the ExtractElement instructions in VL can be vectorized
210 /// to use the original vector.
211 static bool CanReuseExtract(ArrayRef<Value *> VL) {
212 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
213 // Check if all of the extracts come from the same vector and from the
216 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
217 Value *Vec = E0->getOperand(0);
219 // We have to extract from the same vector type.
220 unsigned NElts = Vec->getType()->getVectorNumElements();
222 if (NElts != VL.size())
225 // Check that all of the indices extract from the correct offset.
226 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
227 if (!CI || CI->getZExtValue())
230 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
231 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
232 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
234 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
241 static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
242 SmallVectorImpl<Value *> &Left,
243 SmallVectorImpl<Value *> &Right) {
245 SmallVector<Value *, 16> OrigLeft, OrigRight;
247 bool AllSameOpcodeLeft = true;
248 bool AllSameOpcodeRight = true;
249 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
250 Instruction *I = cast<Instruction>(VL[i]);
251 Value *V0 = I->getOperand(0);
252 Value *V1 = I->getOperand(1);
254 OrigLeft.push_back(V0);
255 OrigRight.push_back(V1);
257 Instruction *I0 = dyn_cast<Instruction>(V0);
258 Instruction *I1 = dyn_cast<Instruction>(V1);
260 // Check whether all operands on one side have the same opcode. In this case
261 // we want to preserve the original order and not make things worse by
263 AllSameOpcodeLeft = I0;
264 AllSameOpcodeRight = I1;
266 if (i && AllSameOpcodeLeft) {
267 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) {
268 if(P0->getOpcode() != I0->getOpcode())
269 AllSameOpcodeLeft = false;
271 AllSameOpcodeLeft = false;
273 if (i && AllSameOpcodeRight) {
274 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) {
275 if(P1->getOpcode() != I1->getOpcode())
276 AllSameOpcodeRight = false;
278 AllSameOpcodeRight = false;
281 // Sort two opcodes. In the code below we try to preserve the ability to use
282 // broadcast of values instead of individual inserts.
289 // If we just sorted according to opcode we would leave the first line in
290 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load).
293 // Because vr2 and vr1 are from the same load we loose the opportunity of a
294 // broadcast for the packed right side in the backend: we have [vr1, vl2]
295 // instead of [vr1, vr2=vr1].
297 if(!i && I0->getOpcode() > I1->getOpcode()) {
300 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) {
301 // Try not to destroy a broad cast for no apparent benefit.
304 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) {
305 // Try preserve broadcasts.
308 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) {
309 // Try preserve broadcasts.
318 // One opcode, put the instruction on the right.
328 bool LeftBroadcast = isSplat(Left);
329 bool RightBroadcast = isSplat(Right);
331 // Don't reorder if the operands where good to begin with.
332 if (!(LeftBroadcast || RightBroadcast) &&
333 (AllSameOpcodeRight || AllSameOpcodeLeft)) {
339 /// Bottom Up SLP Vectorizer.
342 typedef SmallVector<Value *, 8> ValueList;
343 typedef SmallVector<Instruction *, 16> InstrList;
344 typedef SmallPtrSet<Value *, 16> ValueSet;
345 typedef SmallVector<StoreInst *, 8> StoreList;
347 BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl,
348 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li,
350 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt),
351 Builder(Se->getContext()) {
352 // Setup the block numbering utility for all of the blocks in the
354 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
356 BlocksNumbers[BB] = BlockNumbering(BB);
360 /// \brief Vectorize the tree that starts with the elements in \p VL.
361 /// Returns the vectorized root.
362 Value *vectorizeTree();
364 /// \returns the vectorization cost of the subtree that starts at \p VL.
365 /// A negative number means that this is profitable.
368 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
369 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
370 void buildTree(ArrayRef<Value *> Roots,
371 ArrayRef<Value *> UserIgnoreLst = None);
373 /// Clear the internal data structures that are created by 'buildTree'.
375 VectorizableTree.clear();
376 ScalarToTreeEntry.clear();
378 ExternalUses.clear();
379 MemBarrierIgnoreList.clear();
382 /// \returns true if the memory operations A and B are consecutive.
383 bool isConsecutiveAccess(Value *A, Value *B);
385 /// \brief Perform LICM and CSE on the newly generated gather sequences.
386 void optimizeGatherSequence();
390 /// \returns the cost of the vectorizable entry.
391 int getEntryCost(TreeEntry *E);
393 /// This is the recursive part of buildTree.
394 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
396 /// Vectorize a single entry in the tree.
397 Value *vectorizeTree(TreeEntry *E);
399 /// Vectorize a single entry in the tree, starting in \p VL.
400 Value *vectorizeTree(ArrayRef<Value *> VL);
402 /// \returns the pointer to the vectorized value if \p VL is already
403 /// vectorized, or NULL. They may happen in cycles.
404 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
406 /// \brief Take the pointer operand from the Load/Store instruction.
407 /// \returns NULL if this is not a valid Load/Store instruction.
408 static Value *getPointerOperand(Value *I);
410 /// \brief Take the address space operand from the Load/Store instruction.
411 /// \returns -1 if this is not a valid Load/Store instruction.
412 static unsigned getAddressSpaceOperand(Value *I);
414 /// \returns the scalarization cost for this type. Scalarization in this
415 /// context means the creation of vectors from a group of scalars.
416 int getGatherCost(Type *Ty);
418 /// \returns the scalarization cost for this list of values. Assuming that
419 /// this subtree gets vectorized, we may need to extract the values from the
420 /// roots. This method calculates the cost of extracting the values.
421 int getGatherCost(ArrayRef<Value *> VL);
423 /// \returns the AA location that is being access by the instruction.
424 AliasAnalysis::Location getLocation(Instruction *I);
426 /// \brief Checks if it is possible to sink an instruction from
427 /// \p Src to \p Dst.
428 /// \returns the pointer to the barrier instruction if we can't sink.
429 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
431 /// \returns the index of the last instruction in the BB from \p VL.
432 int getLastIndex(ArrayRef<Value *> VL);
434 /// \returns the Instruction in the bundle \p VL.
435 Instruction *getLastInstruction(ArrayRef<Value *> VL);
437 /// \brief Set the Builder insert point to one after the last instruction in
439 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
441 /// \returns a vector from a collection of scalars in \p VL.
442 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
444 /// \returns whether the VectorizableTree is fully vectoriable and will
445 /// be beneficial even the tree height is tiny.
446 bool isFullyVectorizableTinyTree();
449 TreeEntry() : Scalars(), VectorizedValue(0), LastScalarIndex(0),
452 /// \returns true if the scalars in VL are equal to this entry.
453 bool isSame(ArrayRef<Value *> VL) const {
454 assert(VL.size() == Scalars.size() && "Invalid size");
455 return std::equal(VL.begin(), VL.end(), Scalars.begin());
458 /// A vector of scalars.
461 /// The Scalars are vectorized into this value. It is initialized to Null.
462 Value *VectorizedValue;
464 /// The index in the basic block of the last scalar.
467 /// Do we need to gather this sequence ?
471 /// Create a new VectorizableTree entry.
472 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
473 VectorizableTree.push_back(TreeEntry());
474 int idx = VectorizableTree.size() - 1;
475 TreeEntry *Last = &VectorizableTree[idx];
476 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
477 Last->NeedToGather = !Vectorized;
479 Last->LastScalarIndex = getLastIndex(VL);
480 for (int i = 0, e = VL.size(); i != e; ++i) {
481 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
482 ScalarToTreeEntry[VL[i]] = idx;
485 Last->LastScalarIndex = 0;
486 MustGather.insert(VL.begin(), VL.end());
491 /// -- Vectorization State --
492 /// Holds all of the tree entries.
493 std::vector<TreeEntry> VectorizableTree;
495 /// Maps a specific scalar to its tree entry.
496 SmallDenseMap<Value*, int> ScalarToTreeEntry;
498 /// A list of scalars that we found that we need to keep as scalars.
501 /// This POD struct describes one external user in the vectorized tree.
502 struct ExternalUser {
503 ExternalUser (Value *S, llvm::User *U, int L) :
504 Scalar(S), User(U), Lane(L){};
505 // Which scalar in our function.
507 // Which user that uses the scalar.
509 // Which lane does the scalar belong to.
512 typedef SmallVector<ExternalUser, 16> UserList;
514 /// A list of values that need to extracted out of the tree.
515 /// This list holds pairs of (Internal Scalar : External User).
516 UserList ExternalUses;
518 /// A list of instructions to ignore while sinking
519 /// memory instructions. This map must be reset between runs of getCost.
520 ValueSet MemBarrierIgnoreList;
522 /// Holds all of the instructions that we gathered.
523 SetVector<Instruction *> GatherSeq;
524 /// A list of blocks that we are going to CSE.
525 SetVector<BasicBlock *> CSEBlocks;
527 /// Numbers instructions in different blocks.
528 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
530 /// List of users to ignore during scheduling and that don't need extracting.
531 ArrayRef<Value *> UserIgnoreList;
533 // Analysis and block reference.
536 const DataLayout *DL;
537 TargetTransformInfo *TTI;
541 /// Instruction builder to construct the vectorized tree.
545 void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
546 ArrayRef<Value *> UserIgnoreLst) {
548 UserIgnoreList = UserIgnoreLst;
549 if (!getSameType(Roots))
551 buildTree_rec(Roots, 0);
553 // Collect the values that we need to extract from the tree.
554 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
555 TreeEntry *Entry = &VectorizableTree[EIdx];
558 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
559 Value *Scalar = Entry->Scalars[Lane];
561 // No need to handle users of gathered values.
562 if (Entry->NeedToGather)
565 for (User *U : Scalar->users()) {
566 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
568 // Skip in-tree scalars that become vectors.
569 if (ScalarToTreeEntry.count(U)) {
570 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
572 int Idx = ScalarToTreeEntry[U]; (void) Idx;
573 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
576 Instruction *UserInst = dyn_cast<Instruction>(U);
580 // Ignore users in the user ignore list.
581 if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UserInst) !=
582 UserIgnoreList.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 // Ignore users in the user ignore list.
714 if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UI) !=
715 UserIgnoreList.end())
718 // Make sure that we can schedule this unknown user.
719 BlockNumbering &BN = BlocksNumbers[BB];
720 int UserIndex = BN.getIndex(UI);
721 if (UserIndex < MyLastIndex) {
723 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
725 newTreeEntry(VL, false);
731 // Check that every instructions appears once in this bundle.
732 for (unsigned i = 0, e = VL.size(); i < e; ++i)
733 for (unsigned j = i+1; j < e; ++j)
734 if (VL[i] == VL[j]) {
735 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
736 newTreeEntry(VL, false);
740 // Check that instructions in this bundle don't reference other instructions.
741 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
742 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
743 for (User *U : VL[i]->users()) {
744 for (unsigned j = 0; j < e; ++j) {
745 if (i != j && U == VL[j]) {
746 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n");
747 newTreeEntry(VL, false);
754 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
756 unsigned Opcode = getSameOpcode(VL);
758 // Check if it is safe to sink the loads or the stores.
759 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
760 Instruction *Last = getLastInstruction(VL);
762 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
765 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
767 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
768 << "\n because of " << *Barrier << ". Gathering.\n");
769 newTreeEntry(VL, false);
776 case Instruction::PHI: {
777 PHINode *PH = dyn_cast<PHINode>(VL0);
779 // Check for terminator values (e.g. invoke).
780 for (unsigned j = 0; j < VL.size(); ++j)
781 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
782 TerminatorInst *Term = dyn_cast<TerminatorInst>(
783 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
785 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
786 newTreeEntry(VL, false);
791 newTreeEntry(VL, true);
792 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
794 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
796 // Prepare the operand vector.
797 for (unsigned j = 0; j < VL.size(); ++j)
798 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
799 PH->getIncomingBlock(i)));
801 buildTree_rec(Operands, Depth + 1);
805 case Instruction::ExtractElement: {
806 bool Reuse = CanReuseExtract(VL);
808 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
810 newTreeEntry(VL, Reuse);
813 case Instruction::Load: {
814 // Check if the loads are consecutive or of we need to swizzle them.
815 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
816 LoadInst *L = cast<LoadInst>(VL[i]);
817 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
818 newTreeEntry(VL, false);
819 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
823 newTreeEntry(VL, true);
824 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
827 case Instruction::ZExt:
828 case Instruction::SExt:
829 case Instruction::FPToUI:
830 case Instruction::FPToSI:
831 case Instruction::FPExt:
832 case Instruction::PtrToInt:
833 case Instruction::IntToPtr:
834 case Instruction::SIToFP:
835 case Instruction::UIToFP:
836 case Instruction::Trunc:
837 case Instruction::FPTrunc:
838 case Instruction::BitCast: {
839 Type *SrcTy = VL0->getOperand(0)->getType();
840 for (unsigned i = 0; i < VL.size(); ++i) {
841 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
842 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
843 newTreeEntry(VL, false);
844 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
848 newTreeEntry(VL, true);
849 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
851 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
853 // Prepare the operand vector.
854 for (unsigned j = 0; j < VL.size(); ++j)
855 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
857 buildTree_rec(Operands, Depth+1);
861 case Instruction::ICmp:
862 case Instruction::FCmp: {
863 // Check that all of the compares have the same predicate.
864 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
865 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
866 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
867 CmpInst *Cmp = cast<CmpInst>(VL[i]);
868 if (Cmp->getPredicate() != P0 ||
869 Cmp->getOperand(0)->getType() != ComparedTy) {
870 newTreeEntry(VL, false);
871 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
876 newTreeEntry(VL, true);
877 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
879 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
881 // Prepare the operand vector.
882 for (unsigned j = 0; j < VL.size(); ++j)
883 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
885 buildTree_rec(Operands, Depth+1);
889 case Instruction::Select:
890 case Instruction::Add:
891 case Instruction::FAdd:
892 case Instruction::Sub:
893 case Instruction::FSub:
894 case Instruction::Mul:
895 case Instruction::FMul:
896 case Instruction::UDiv:
897 case Instruction::SDiv:
898 case Instruction::FDiv:
899 case Instruction::URem:
900 case Instruction::SRem:
901 case Instruction::FRem:
902 case Instruction::Shl:
903 case Instruction::LShr:
904 case Instruction::AShr:
905 case Instruction::And:
906 case Instruction::Or:
907 case Instruction::Xor: {
908 newTreeEntry(VL, true);
909 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
911 // Sort operands of the instructions so that each side is more likely to
912 // have the same opcode.
913 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
914 ValueList Left, Right;
915 reorderInputsAccordingToOpcode(VL, Left, Right);
916 buildTree_rec(Left, Depth + 1);
917 buildTree_rec(Right, Depth + 1);
921 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
923 // Prepare the operand vector.
924 for (unsigned j = 0; j < VL.size(); ++j)
925 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
927 buildTree_rec(Operands, Depth+1);
931 case Instruction::Store: {
932 // Check if the stores are consecutive or of we need to swizzle them.
933 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
934 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
935 newTreeEntry(VL, false);
936 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
940 newTreeEntry(VL, true);
941 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
944 for (unsigned j = 0; j < VL.size(); ++j)
945 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
947 // We can ignore these values because we are sinking them down.
948 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
949 buildTree_rec(Operands, Depth + 1);
952 case Instruction::Call: {
953 // Check if the calls are all to the same vectorizable intrinsic.
954 IntrinsicInst *II = dyn_cast<IntrinsicInst>(VL[0]);
956 newTreeEntry(VL, false);
957 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
961 Intrinsic::ID ID = II->getIntrinsicID();
963 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
964 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[i]);
965 if (!II2 || II2->getIntrinsicID() != ID) {
966 newTreeEntry(VL, false);
967 DEBUG(dbgs() << "SLP: mismatched calls:" << *II << "!=" << *VL[i]
973 newTreeEntry(VL, true);
974 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i) {
976 // Prepare the operand vector.
977 for (unsigned j = 0; j < VL.size(); ++j) {
978 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[j]);
979 Operands.push_back(II2->getArgOperand(i));
981 buildTree_rec(Operands, Depth + 1);
986 newTreeEntry(VL, false);
987 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
992 int BoUpSLP::getEntryCost(TreeEntry *E) {
993 ArrayRef<Value*> VL = E->Scalars;
995 Type *ScalarTy = VL[0]->getType();
996 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
997 ScalarTy = SI->getValueOperand()->getType();
998 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1000 if (E->NeedToGather) {
1001 if (allConstant(VL))
1004 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1006 return getGatherCost(E->Scalars);
1009 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1011 Instruction *VL0 = cast<Instruction>(VL[0]);
1012 unsigned Opcode = VL0->getOpcode();
1014 case Instruction::PHI: {
1017 case Instruction::ExtractElement: {
1018 if (CanReuseExtract(VL)) {
1020 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1021 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1023 // Take credit for instruction that will become dead.
1025 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1029 return getGatherCost(VecTy);
1031 case Instruction::ZExt:
1032 case Instruction::SExt:
1033 case Instruction::FPToUI:
1034 case Instruction::FPToSI:
1035 case Instruction::FPExt:
1036 case Instruction::PtrToInt:
1037 case Instruction::IntToPtr:
1038 case Instruction::SIToFP:
1039 case Instruction::UIToFP:
1040 case Instruction::Trunc:
1041 case Instruction::FPTrunc:
1042 case Instruction::BitCast: {
1043 Type *SrcTy = VL0->getOperand(0)->getType();
1045 // Calculate the cost of this instruction.
1046 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1047 VL0->getType(), SrcTy);
1049 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1050 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1051 return VecCost - ScalarCost;
1053 case Instruction::FCmp:
1054 case Instruction::ICmp:
1055 case Instruction::Select:
1056 case Instruction::Add:
1057 case Instruction::FAdd:
1058 case Instruction::Sub:
1059 case Instruction::FSub:
1060 case Instruction::Mul:
1061 case Instruction::FMul:
1062 case Instruction::UDiv:
1063 case Instruction::SDiv:
1064 case Instruction::FDiv:
1065 case Instruction::URem:
1066 case Instruction::SRem:
1067 case Instruction::FRem:
1068 case Instruction::Shl:
1069 case Instruction::LShr:
1070 case Instruction::AShr:
1071 case Instruction::And:
1072 case Instruction::Or:
1073 case Instruction::Xor: {
1074 // Calculate the cost of this instruction.
1077 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1078 Opcode == Instruction::Select) {
1079 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1080 ScalarCost = VecTy->getNumElements() *
1081 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1082 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1084 // Certain instructions can be cheaper to vectorize if they have a
1085 // constant second vector operand.
1086 TargetTransformInfo::OperandValueKind Op1VK =
1087 TargetTransformInfo::OK_AnyValue;
1088 TargetTransformInfo::OperandValueKind Op2VK =
1089 TargetTransformInfo::OK_UniformConstantValue;
1091 // If all operands are exactly the same ConstantInt then set the
1092 // operand kind to OK_UniformConstantValue.
1093 // If instead not all operands are constants, then set the operand kind
1094 // to OK_AnyValue. If all operands are constants but not the same,
1095 // then set the operand kind to OK_NonUniformConstantValue.
1096 ConstantInt *CInt = NULL;
1097 for (unsigned i = 0; i < VL.size(); ++i) {
1098 const Instruction *I = cast<Instruction>(VL[i]);
1099 if (!isa<ConstantInt>(I->getOperand(1))) {
1100 Op2VK = TargetTransformInfo::OK_AnyValue;
1104 CInt = cast<ConstantInt>(I->getOperand(1));
1107 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1108 CInt != cast<ConstantInt>(I->getOperand(1)))
1109 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1113 VecTy->getNumElements() *
1114 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1115 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1117 return VecCost - ScalarCost;
1119 case Instruction::Load: {
1120 // Cost of wide load - cost of scalar loads.
1121 int ScalarLdCost = VecTy->getNumElements() *
1122 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1123 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1124 return VecLdCost - ScalarLdCost;
1126 case Instruction::Store: {
1127 // We know that we can merge the stores. Calculate the cost.
1128 int ScalarStCost = VecTy->getNumElements() *
1129 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1130 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1131 return VecStCost - ScalarStCost;
1133 case Instruction::Call: {
1134 CallInst *CI = cast<CallInst>(VL0);
1135 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1136 Intrinsic::ID ID = II->getIntrinsicID();
1138 // Calculate the cost of the scalar and vector calls.
1139 SmallVector<Type*, 4> ScalarTys, VecTys;
1140 for (unsigned op = 0, opc = II->getNumArgOperands(); op!= opc; ++op) {
1141 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1142 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1143 VecTy->getNumElements()));
1146 int ScalarCallCost = VecTy->getNumElements() *
1147 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1149 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1151 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1152 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1153 << " for " << *II << "\n");
1155 return VecCallCost - ScalarCallCost;
1158 llvm_unreachable("Unknown instruction");
1162 bool BoUpSLP::isFullyVectorizableTinyTree() {
1163 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1164 VectorizableTree.size() << " is fully vectorizable .\n");
1166 // We only handle trees of height 2.
1167 if (VectorizableTree.size() != 2)
1170 // Handle splat stores.
1171 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1174 // Gathering cost would be too much for tiny trees.
1175 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1181 int BoUpSLP::getTreeCost() {
1183 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1184 VectorizableTree.size() << ".\n");
1186 // We only vectorize tiny trees if it is fully vectorizable.
1187 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1188 if (!VectorizableTree.size()) {
1189 assert(!ExternalUses.size() && "We should not have any external users");
1194 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1196 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1197 int C = getEntryCost(&VectorizableTree[i]);
1198 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1199 << *VectorizableTree[i].Scalars[0] << " .\n");
1203 SmallSet<Value *, 16> ExtractCostCalculated;
1204 int ExtractCost = 0;
1205 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1207 // We only add extract cost once for the same scalar.
1208 if (!ExtractCostCalculated.insert(I->Scalar))
1211 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1212 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1216 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1217 return Cost + ExtractCost;
1220 int BoUpSLP::getGatherCost(Type *Ty) {
1222 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1223 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1227 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1228 // Find the type of the operands in VL.
1229 Type *ScalarTy = VL[0]->getType();
1230 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1231 ScalarTy = SI->getValueOperand()->getType();
1232 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1233 // Find the cost of inserting/extracting values from the vector.
1234 return getGatherCost(VecTy);
1237 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1238 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1239 return AA->getLocation(SI);
1240 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1241 return AA->getLocation(LI);
1242 return AliasAnalysis::Location();
1245 Value *BoUpSLP::getPointerOperand(Value *I) {
1246 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1247 return LI->getPointerOperand();
1248 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1249 return SI->getPointerOperand();
1253 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1254 if (LoadInst *L = dyn_cast<LoadInst>(I))
1255 return L->getPointerAddressSpace();
1256 if (StoreInst *S = dyn_cast<StoreInst>(I))
1257 return S->getPointerAddressSpace();
1261 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1262 Value *PtrA = getPointerOperand(A);
1263 Value *PtrB = getPointerOperand(B);
1264 unsigned ASA = getAddressSpaceOperand(A);
1265 unsigned ASB = getAddressSpaceOperand(B);
1267 // Check that the address spaces match and that the pointers are valid.
1268 if (!PtrA || !PtrB || (ASA != ASB))
1271 // Make sure that A and B are different pointers of the same type.
1272 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1275 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1276 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1277 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1279 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1280 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1281 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1283 APInt OffsetDelta = OffsetB - OffsetA;
1285 // Check if they are based on the same pointer. That makes the offsets
1288 return OffsetDelta == Size;
1290 // Compute the necessary base pointer delta to have the necessary final delta
1291 // equal to the size.
1292 APInt BaseDelta = Size - OffsetDelta;
1294 // Otherwise compute the distance with SCEV between the base pointers.
1295 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1296 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1297 const SCEV *C = SE->getConstant(BaseDelta);
1298 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1299 return X == PtrSCEVB;
1302 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1303 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1304 BasicBlock::iterator I = Src, E = Dst;
1305 /// Scan all of the instruction from SRC to DST and check if
1306 /// the source may alias.
1307 for (++I; I != E; ++I) {
1308 // Ignore store instructions that are marked as 'ignore'.
1309 if (MemBarrierIgnoreList.count(I))
1311 if (Src->mayWriteToMemory()) /* Write */ {
1312 if (!I->mayReadOrWriteMemory())
1315 if (!I->mayWriteToMemory())
1318 AliasAnalysis::Location A = getLocation(&*I);
1319 AliasAnalysis::Location B = getLocation(Src);
1321 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1327 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1328 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1329 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1330 BlockNumbering &BN = BlocksNumbers[BB];
1332 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1333 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1334 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1338 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1339 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1340 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1341 BlockNumbering &BN = BlocksNumbers[BB];
1343 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1344 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1345 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1346 Instruction *I = BN.getInstruction(MaxIdx);
1347 assert(I && "bad location");
1351 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1352 Instruction *VL0 = cast<Instruction>(VL[0]);
1353 Instruction *LastInst = getLastInstruction(VL);
1354 BasicBlock::iterator NextInst = LastInst;
1356 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1357 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1360 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1361 Value *Vec = UndefValue::get(Ty);
1362 // Generate the 'InsertElement' instruction.
1363 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1364 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1365 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1366 GatherSeq.insert(Insrt);
1367 CSEBlocks.insert(Insrt->getParent());
1369 // Add to our 'need-to-extract' list.
1370 if (ScalarToTreeEntry.count(VL[i])) {
1371 int Idx = ScalarToTreeEntry[VL[i]];
1372 TreeEntry *E = &VectorizableTree[Idx];
1373 // Find which lane we need to extract.
1375 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1376 // Is this the lane of the scalar that we are looking for ?
1377 if (E->Scalars[Lane] == VL[i]) {
1382 assert(FoundLane >= 0 && "Could not find the correct lane");
1383 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1391 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1392 SmallDenseMap<Value*, int>::const_iterator Entry
1393 = ScalarToTreeEntry.find(VL[0]);
1394 if (Entry != ScalarToTreeEntry.end()) {
1395 int Idx = Entry->second;
1396 const TreeEntry *En = &VectorizableTree[Idx];
1397 if (En->isSame(VL) && En->VectorizedValue)
1398 return En->VectorizedValue;
1403 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1404 if (ScalarToTreeEntry.count(VL[0])) {
1405 int Idx = ScalarToTreeEntry[VL[0]];
1406 TreeEntry *E = &VectorizableTree[Idx];
1408 return vectorizeTree(E);
1411 Type *ScalarTy = VL[0]->getType();
1412 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1413 ScalarTy = SI->getValueOperand()->getType();
1414 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1416 return Gather(VL, VecTy);
1419 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1420 IRBuilder<>::InsertPointGuard Guard(Builder);
1422 if (E->VectorizedValue) {
1423 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1424 return E->VectorizedValue;
1427 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1428 Type *ScalarTy = VL0->getType();
1429 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1430 ScalarTy = SI->getValueOperand()->getType();
1431 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1433 if (E->NeedToGather) {
1434 setInsertPointAfterBundle(E->Scalars);
1435 return Gather(E->Scalars, VecTy);
1438 unsigned Opcode = VL0->getOpcode();
1439 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1442 case Instruction::PHI: {
1443 PHINode *PH = dyn_cast<PHINode>(VL0);
1444 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1445 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1446 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1447 E->VectorizedValue = NewPhi;
1449 // PHINodes may have multiple entries from the same block. We want to
1450 // visit every block once.
1451 SmallSet<BasicBlock*, 4> VisitedBBs;
1453 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1455 BasicBlock *IBB = PH->getIncomingBlock(i);
1457 if (!VisitedBBs.insert(IBB)) {
1458 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1462 // Prepare the operand vector.
1463 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1464 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1465 getIncomingValueForBlock(IBB));
1467 Builder.SetInsertPoint(IBB->getTerminator());
1468 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1469 Value *Vec = vectorizeTree(Operands);
1470 NewPhi->addIncoming(Vec, IBB);
1473 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1474 "Invalid number of incoming values");
1478 case Instruction::ExtractElement: {
1479 if (CanReuseExtract(E->Scalars)) {
1480 Value *V = VL0->getOperand(0);
1481 E->VectorizedValue = V;
1484 return Gather(E->Scalars, VecTy);
1486 case Instruction::ZExt:
1487 case Instruction::SExt:
1488 case Instruction::FPToUI:
1489 case Instruction::FPToSI:
1490 case Instruction::FPExt:
1491 case Instruction::PtrToInt:
1492 case Instruction::IntToPtr:
1493 case Instruction::SIToFP:
1494 case Instruction::UIToFP:
1495 case Instruction::Trunc:
1496 case Instruction::FPTrunc:
1497 case Instruction::BitCast: {
1499 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1500 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1502 setInsertPointAfterBundle(E->Scalars);
1504 Value *InVec = vectorizeTree(INVL);
1506 if (Value *V = alreadyVectorized(E->Scalars))
1509 CastInst *CI = dyn_cast<CastInst>(VL0);
1510 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1511 E->VectorizedValue = V;
1514 case Instruction::FCmp:
1515 case Instruction::ICmp: {
1516 ValueList LHSV, RHSV;
1517 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1518 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1519 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1522 setInsertPointAfterBundle(E->Scalars);
1524 Value *L = vectorizeTree(LHSV);
1525 Value *R = vectorizeTree(RHSV);
1527 if (Value *V = alreadyVectorized(E->Scalars))
1530 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1532 if (Opcode == Instruction::FCmp)
1533 V = Builder.CreateFCmp(P0, L, R);
1535 V = Builder.CreateICmp(P0, L, R);
1537 E->VectorizedValue = V;
1540 case Instruction::Select: {
1541 ValueList TrueVec, FalseVec, CondVec;
1542 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1543 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1544 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1545 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1548 setInsertPointAfterBundle(E->Scalars);
1550 Value *Cond = vectorizeTree(CondVec);
1551 Value *True = vectorizeTree(TrueVec);
1552 Value *False = vectorizeTree(FalseVec);
1554 if (Value *V = alreadyVectorized(E->Scalars))
1557 Value *V = Builder.CreateSelect(Cond, True, False);
1558 E->VectorizedValue = V;
1561 case Instruction::Add:
1562 case Instruction::FAdd:
1563 case Instruction::Sub:
1564 case Instruction::FSub:
1565 case Instruction::Mul:
1566 case Instruction::FMul:
1567 case Instruction::UDiv:
1568 case Instruction::SDiv:
1569 case Instruction::FDiv:
1570 case Instruction::URem:
1571 case Instruction::SRem:
1572 case Instruction::FRem:
1573 case Instruction::Shl:
1574 case Instruction::LShr:
1575 case Instruction::AShr:
1576 case Instruction::And:
1577 case Instruction::Or:
1578 case Instruction::Xor: {
1579 ValueList LHSVL, RHSVL;
1580 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1581 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1583 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1584 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1585 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1588 setInsertPointAfterBundle(E->Scalars);
1590 Value *LHS = vectorizeTree(LHSVL);
1591 Value *RHS = vectorizeTree(RHSVL);
1593 if (LHS == RHS && isa<Instruction>(LHS)) {
1594 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1597 if (Value *V = alreadyVectorized(E->Scalars))
1600 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1601 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1602 E->VectorizedValue = V;
1604 if (Instruction *I = dyn_cast<Instruction>(V))
1605 return propagateMetadata(I, E->Scalars);
1609 case Instruction::Load: {
1610 // Loads are inserted at the head of the tree because we don't want to
1611 // sink them all the way down past store instructions.
1612 setInsertPointAfterBundle(E->Scalars);
1614 LoadInst *LI = cast<LoadInst>(VL0);
1615 unsigned AS = LI->getPointerAddressSpace();
1617 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1618 VecTy->getPointerTo(AS));
1619 unsigned Alignment = LI->getAlignment();
1620 LI = Builder.CreateLoad(VecPtr);
1621 LI->setAlignment(Alignment);
1622 E->VectorizedValue = LI;
1623 return propagateMetadata(LI, E->Scalars);
1625 case Instruction::Store: {
1626 StoreInst *SI = cast<StoreInst>(VL0);
1627 unsigned Alignment = SI->getAlignment();
1628 unsigned AS = SI->getPointerAddressSpace();
1631 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1632 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1634 setInsertPointAfterBundle(E->Scalars);
1636 Value *VecValue = vectorizeTree(ValueOp);
1637 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1638 VecTy->getPointerTo(AS));
1639 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1640 S->setAlignment(Alignment);
1641 E->VectorizedValue = S;
1642 return propagateMetadata(S, E->Scalars);
1644 case Instruction::Call: {
1645 CallInst *CI = cast<CallInst>(VL0);
1647 setInsertPointAfterBundle(E->Scalars);
1648 std::vector<Value *> OpVecs;
1649 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1651 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1652 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1653 OpVL.push_back(CEI->getArgOperand(j));
1656 Value *OpVec = vectorizeTree(OpVL);
1657 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1658 OpVecs.push_back(OpVec);
1661 Module *M = F->getParent();
1662 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1663 Intrinsic::ID ID = II->getIntrinsicID();
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 users in the ignorelist by undef.
1753 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
1754 UserIgnoreList.end())) &&
1755 "Replacing out-of-tree value with undef");
1758 Value *Undef = UndefValue::get(Ty);
1759 Scalar->replaceAllUsesWith(Undef);
1761 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1762 cast<Instruction>(Scalar)->eraseFromParent();
1766 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1767 BlocksNumbers[it].forget();
1769 Builder.ClearInsertionPoint();
1771 return VectorizableTree[0].VectorizedValue;
1774 void BoUpSLP::optimizeGatherSequence() {
1775 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1776 << " gather sequences instructions.\n");
1777 // LICM InsertElementInst sequences.
1778 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1779 e = GatherSeq.end(); it != e; ++it) {
1780 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1785 // Check if this block is inside a loop.
1786 Loop *L = LI->getLoopFor(Insert->getParent());
1790 // Check if it has a preheader.
1791 BasicBlock *PreHeader = L->getLoopPreheader();
1795 // If the vector or the element that we insert into it are
1796 // instructions that are defined in this basic block then we can't
1797 // hoist this instruction.
1798 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1799 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1800 if (CurrVec && L->contains(CurrVec))
1802 if (NewElem && L->contains(NewElem))
1805 // We can hoist this instruction. Move it to the pre-header.
1806 Insert->moveBefore(PreHeader->getTerminator());
1809 // Sort blocks by domination. This ensures we visit a block after all blocks
1810 // dominating it are visited.
1811 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1812 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1813 [this](const BasicBlock *A, const BasicBlock *B) {
1814 return DT->properlyDominates(A, B);
1817 // Perform O(N^2) search over the gather sequences and merge identical
1818 // instructions. TODO: We can further optimize this scan if we split the
1819 // instructions into different buckets based on the insert lane.
1820 SmallVector<Instruction *, 16> Visited;
1821 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1822 E = CSEWorkList.end();
1824 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1825 "Worklist not sorted properly!");
1826 BasicBlock *BB = *I;
1827 // For all instructions in blocks containing gather sequences:
1828 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1829 Instruction *In = it++;
1830 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1833 // Check if we can replace this instruction with any of the
1834 // visited instructions.
1835 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1838 if (In->isIdenticalTo(*v) &&
1839 DT->dominates((*v)->getParent(), In->getParent())) {
1840 In->replaceAllUsesWith(*v);
1841 In->eraseFromParent();
1847 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1848 Visited.push_back(In);
1856 /// The SLPVectorizer Pass.
1857 struct SLPVectorizer : public FunctionPass {
1858 typedef SmallVector<StoreInst *, 8> StoreList;
1859 typedef MapVector<Value *, StoreList> StoreListMap;
1861 /// Pass identification, replacement for typeid
1864 explicit SLPVectorizer() : FunctionPass(ID) {
1865 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1868 ScalarEvolution *SE;
1869 const DataLayout *DL;
1870 TargetTransformInfo *TTI;
1875 bool runOnFunction(Function &F) override {
1876 if (skipOptnoneFunction(F))
1879 SE = &getAnalysis<ScalarEvolution>();
1880 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1881 DL = DLP ? &DLP->getDataLayout() : 0;
1882 TTI = &getAnalysis<TargetTransformInfo>();
1883 AA = &getAnalysis<AliasAnalysis>();
1884 LI = &getAnalysis<LoopInfo>();
1885 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1888 bool Changed = false;
1890 // If the target claims to have no vector registers don't attempt
1892 if (!TTI->getNumberOfRegisters(true))
1895 // Must have DataLayout. We can't require it because some tests run w/o
1900 // Don't vectorize when the attribute NoImplicitFloat is used.
1901 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1904 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1906 // Use the bottom up slp vectorizer to construct chains that start with
1907 // he store instructions.
1908 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1910 // Scan the blocks in the function in post order.
1911 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1912 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1913 BasicBlock *BB = *it;
1915 // Vectorize trees that end at stores.
1916 if (unsigned count = collectStores(BB, R)) {
1918 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1919 Changed |= vectorizeStoreChains(R);
1922 // Vectorize trees that end at reductions.
1923 Changed |= vectorizeChainsInBlock(BB, R);
1927 R.optimizeGatherSequence();
1928 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1929 DEBUG(verifyFunction(F));
1934 void getAnalysisUsage(AnalysisUsage &AU) const override {
1935 FunctionPass::getAnalysisUsage(AU);
1936 AU.addRequired<ScalarEvolution>();
1937 AU.addRequired<AliasAnalysis>();
1938 AU.addRequired<TargetTransformInfo>();
1939 AU.addRequired<LoopInfo>();
1940 AU.addRequired<DominatorTreeWrapperPass>();
1941 AU.addPreserved<LoopInfo>();
1942 AU.addPreserved<DominatorTreeWrapperPass>();
1943 AU.setPreservesCFG();
1948 /// \brief Collect memory references and sort them according to their base
1949 /// object. We sort the stores to their base objects to reduce the cost of the
1950 /// quadratic search on the stores. TODO: We can further reduce this cost
1951 /// if we flush the chain creation every time we run into a memory barrier.
1952 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1954 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1955 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1957 /// \brief Try to vectorize a list of operands.
1958 /// \@param BuildVector A list of users to ignore for the purpose of
1959 /// scheduling and that don't need extracting.
1960 /// \returns true if a value was vectorized.
1961 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
1962 ArrayRef<Value *> BuildVector = None);
1964 /// \brief Try to vectorize a chain that may start at the operands of \V;
1965 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1967 /// \brief Vectorize the stores that were collected in StoreRefs.
1968 bool vectorizeStoreChains(BoUpSLP &R);
1970 /// \brief Scan the basic block and look for patterns that are likely to start
1971 /// a vectorization chain.
1972 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1974 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1977 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1980 StoreListMap StoreRefs;
1983 /// \brief Check that the Values in the slice in VL array are still existent in
1984 /// the WeakVH array.
1985 /// Vectorization of part of the VL array may cause later values in the VL array
1986 /// to become invalid. We track when this has happened in the WeakVH array.
1987 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1988 SmallVectorImpl<WeakVH> &VH,
1989 unsigned SliceBegin,
1990 unsigned SliceSize) {
1991 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1998 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1999 int CostThreshold, BoUpSLP &R) {
2000 unsigned ChainLen = Chain.size();
2001 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2003 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2004 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2005 unsigned VF = MinVecRegSize / Sz;
2007 if (!isPowerOf2_32(Sz) || VF < 2)
2010 // Keep track of values that were deleted by vectorizing in the loop below.
2011 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2013 bool Changed = false;
2014 // Look for profitable vectorizable trees at all offsets, starting at zero.
2015 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2019 // Check that a previous iteration of this loop did not delete the Value.
2020 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2023 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2025 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2027 R.buildTree(Operands);
2029 int Cost = R.getTreeCost();
2031 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2032 if (Cost < CostThreshold) {
2033 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2036 // Move to the next bundle.
2045 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2046 int costThreshold, BoUpSLP &R) {
2047 SetVector<Value *> Heads, Tails;
2048 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2050 // We may run into multiple chains that merge into a single chain. We mark the
2051 // stores that we vectorized so that we don't visit the same store twice.
2052 BoUpSLP::ValueSet VectorizedStores;
2053 bool Changed = false;
2055 // Do a quadratic search on all of the given stores and find
2056 // all of the pairs of stores that follow each other.
2057 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2058 for (unsigned j = 0; j < e; ++j) {
2062 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2063 Tails.insert(Stores[j]);
2064 Heads.insert(Stores[i]);
2065 ConsecutiveChain[Stores[i]] = Stores[j];
2070 // For stores that start but don't end a link in the chain:
2071 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2073 if (Tails.count(*it))
2076 // We found a store instr that starts a chain. Now follow the chain and try
2078 BoUpSLP::ValueList Operands;
2080 // Collect the chain into a list.
2081 while (Tails.count(I) || Heads.count(I)) {
2082 if (VectorizedStores.count(I))
2084 Operands.push_back(I);
2085 // Move to the next value in the chain.
2086 I = ConsecutiveChain[I];
2089 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2091 // Mark the vectorized stores so that we don't vectorize them again.
2093 VectorizedStores.insert(Operands.begin(), Operands.end());
2094 Changed |= Vectorized;
2101 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2104 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2105 StoreInst *SI = dyn_cast<StoreInst>(it);
2109 // Don't touch volatile stores.
2110 if (!SI->isSimple())
2113 // Check that the pointer points to scalars.
2114 Type *Ty = SI->getValueOperand()->getType();
2115 if (Ty->isAggregateType() || Ty->isVectorTy())
2118 // Find the base pointer.
2119 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2121 // Save the store locations.
2122 StoreRefs[Ptr].push_back(SI);
2128 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2131 Value *VL[] = { A, B };
2132 return tryToVectorizeList(VL, R);
2135 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2136 ArrayRef<Value *> BuildVector) {
2140 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2142 // Check that all of the parts are scalar instructions of the same type.
2143 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2147 unsigned Opcode0 = I0->getOpcode();
2149 Type *Ty0 = I0->getType();
2150 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2151 unsigned VF = MinVecRegSize / Sz;
2153 for (int i = 0, e = VL.size(); i < e; ++i) {
2154 Type *Ty = VL[i]->getType();
2155 if (Ty->isAggregateType() || Ty->isVectorTy())
2157 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2158 if (!Inst || Inst->getOpcode() != Opcode0)
2162 bool Changed = false;
2164 // Keep track of values that were delete by vectorizing in the loop below.
2165 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2167 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2168 unsigned OpsWidth = 0;
2175 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2178 // Check that a previous iteration of this loop did not delete the Value.
2179 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2182 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2184 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2186 ArrayRef<Value *> BuildVectorSlice;
2187 if (!BuildVector.empty())
2188 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
2190 R.buildTree(Ops, BuildVectorSlice);
2191 int Cost = R.getTreeCost();
2193 if (Cost < -SLPCostThreshold) {
2194 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
2195 Value *VectorizedRoot = R.vectorizeTree();
2197 // Reconstruct the build vector by extracting the vectorized root. This
2198 // way we handle the case where some elements of the vector are undefined.
2199 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
2200 if (!BuildVectorSlice.empty()) {
2201 Instruction *InsertAfter = cast<Instruction>(VectorizedRoot);
2202 for (auto &V : BuildVectorSlice) {
2203 InsertElementInst *IE = cast<InsertElementInst>(V);
2204 IRBuilder<> Builder(++BasicBlock::iterator(InsertAfter));
2205 Instruction *Extract = cast<Instruction>(
2206 Builder.CreateExtractElement(VectorizedRoot, IE->getOperand(2)));
2207 IE->setOperand(1, Extract);
2208 IE->removeFromParent();
2209 IE->insertAfter(Extract);
2213 // Move to the next bundle.
2222 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2226 // Try to vectorize V.
2227 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2230 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2231 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2233 if (B && B->hasOneUse()) {
2234 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2235 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2236 if (tryToVectorizePair(A, B0, R)) {
2240 if (tryToVectorizePair(A, B1, R)) {
2247 if (A && A->hasOneUse()) {
2248 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2249 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2250 if (tryToVectorizePair(A0, B, R)) {
2254 if (tryToVectorizePair(A1, B, R)) {
2262 /// \brief Generate a shuffle mask to be used in a reduction tree.
2264 /// \param VecLen The length of the vector to be reduced.
2265 /// \param NumEltsToRdx The number of elements that should be reduced in the
2267 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2268 /// reduction. A pairwise reduction will generate a mask of
2269 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2270 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2271 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2272 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2273 bool IsPairwise, bool IsLeft,
2274 IRBuilder<> &Builder) {
2275 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2277 SmallVector<Constant *, 32> ShuffleMask(
2278 VecLen, UndefValue::get(Builder.getInt32Ty()));
2281 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2282 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2283 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2285 // Move the upper half of the vector to the lower half.
2286 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2287 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2289 return ConstantVector::get(ShuffleMask);
2293 /// Model horizontal reductions.
2295 /// A horizontal reduction is a tree of reduction operations (currently add and
2296 /// fadd) that has operations that can be put into a vector as its leaf.
2297 /// For example, this tree:
2304 /// This tree has "mul" as its reduced values and "+" as its reduction
2305 /// operations. A reduction might be feeding into a store or a binary operation
2320 class HorizontalReduction {
2321 SmallVector<Value *, 16> ReductionOps;
2322 SmallVector<Value *, 32> ReducedVals;
2324 BinaryOperator *ReductionRoot;
2325 PHINode *ReductionPHI;
2327 /// The opcode of the reduction.
2328 unsigned ReductionOpcode;
2329 /// The opcode of the values we perform a reduction on.
2330 unsigned ReducedValueOpcode;
2331 /// The width of one full horizontal reduction operation.
2332 unsigned ReduxWidth;
2333 /// Should we model this reduction as a pairwise reduction tree or a tree that
2334 /// splits the vector in halves and adds those halves.
2335 bool IsPairwiseReduction;
2338 HorizontalReduction()
2339 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0),
2340 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2342 /// \brief Try to find a reduction tree.
2343 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2344 const DataLayout *DL) {
2346 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2347 "Thi phi needs to use the binary operator");
2349 // We could have a initial reductions that is not an add.
2350 // r *= v1 + v2 + v3 + v4
2351 // In such a case start looking for a tree rooted in the first '+'.
2353 if (B->getOperand(0) == Phi) {
2355 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2356 } else if (B->getOperand(1) == Phi) {
2358 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2365 Type *Ty = B->getType();
2366 if (Ty->isVectorTy())
2369 ReductionOpcode = B->getOpcode();
2370 ReducedValueOpcode = 0;
2371 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2378 // We currently only support adds.
2379 if (ReductionOpcode != Instruction::Add &&
2380 ReductionOpcode != Instruction::FAdd)
2383 // Post order traverse the reduction tree starting at B. We only handle true
2384 // trees containing only binary operators.
2385 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2386 Stack.push_back(std::make_pair(B, 0));
2387 while (!Stack.empty()) {
2388 BinaryOperator *TreeN = Stack.back().first;
2389 unsigned EdgeToVist = Stack.back().second++;
2390 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2392 // Only handle trees in the current basic block.
2393 if (TreeN->getParent() != B->getParent())
2396 // Each tree node needs to have one user except for the ultimate
2398 if (!TreeN->hasOneUse() && TreeN != B)
2402 if (EdgeToVist == 2 || IsReducedValue) {
2403 if (IsReducedValue) {
2404 // Make sure that the opcodes of the operations that we are going to
2406 if (!ReducedValueOpcode)
2407 ReducedValueOpcode = TreeN->getOpcode();
2408 else if (ReducedValueOpcode != TreeN->getOpcode())
2410 ReducedVals.push_back(TreeN);
2412 // We need to be able to reassociate the adds.
2413 if (!TreeN->isAssociative())
2415 ReductionOps.push_back(TreeN);
2422 // Visit left or right.
2423 Value *NextV = TreeN->getOperand(EdgeToVist);
2424 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2426 Stack.push_back(std::make_pair(Next, 0));
2427 else if (NextV != Phi)
2433 /// \brief Attempt to vectorize the tree found by
2434 /// matchAssociativeReduction.
2435 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2436 if (ReducedVals.empty())
2439 unsigned NumReducedVals = ReducedVals.size();
2440 if (NumReducedVals < ReduxWidth)
2443 Value *VectorizedTree = 0;
2444 IRBuilder<> Builder(ReductionRoot);
2445 FastMathFlags Unsafe;
2446 Unsafe.setUnsafeAlgebra();
2447 Builder.SetFastMathFlags(Unsafe);
2450 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2451 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2452 V.buildTree(ValsToReduce, ReductionOps);
2455 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2456 if (Cost >= -SLPCostThreshold)
2459 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2462 // Vectorize a tree.
2463 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2464 Value *VectorizedRoot = V.vectorizeTree();
2466 // Emit a reduction.
2467 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2468 if (VectorizedTree) {
2469 Builder.SetCurrentDebugLocation(Loc);
2470 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2471 ReducedSubTree, "bin.rdx");
2473 VectorizedTree = ReducedSubTree;
2476 if (VectorizedTree) {
2477 // Finish the reduction.
2478 for (; i < NumReducedVals; ++i) {
2479 Builder.SetCurrentDebugLocation(
2480 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2481 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2486 assert(ReductionRoot != NULL && "Need a reduction operation");
2487 ReductionRoot->setOperand(0, VectorizedTree);
2488 ReductionRoot->setOperand(1, ReductionPHI);
2490 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2492 return VectorizedTree != 0;
2497 /// \brief Calcuate the cost of a reduction.
2498 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2499 Type *ScalarTy = FirstReducedVal->getType();
2500 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2502 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2503 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2505 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2506 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2508 int ScalarReduxCost =
2509 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2511 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2512 << " for reduction that starts with " << *FirstReducedVal
2514 << (IsPairwiseReduction ? "pairwise" : "splitting")
2515 << " reduction)\n");
2517 return VecReduxCost - ScalarReduxCost;
2520 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2521 Value *R, const Twine &Name = "") {
2522 if (Opcode == Instruction::FAdd)
2523 return Builder.CreateFAdd(L, R, Name);
2524 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2527 /// \brief Emit a horizontal reduction of the vectorized value.
2528 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2529 assert(VectorizedValue && "Need to have a vectorized tree node");
2530 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2531 assert(isPowerOf2_32(ReduxWidth) &&
2532 "We only handle power-of-two reductions for now");
2534 Value *TmpVec = ValToReduce;
2535 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2536 if (IsPairwiseReduction) {
2538 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2540 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2542 Value *LeftShuf = Builder.CreateShuffleVector(
2543 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2544 Value *RightShuf = Builder.CreateShuffleVector(
2545 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2547 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2551 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2552 Value *Shuf = Builder.CreateShuffleVector(
2553 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2554 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2558 // The result is in the first element of the vector.
2559 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2563 /// \brief Recognize construction of vectors like
2564 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2565 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2566 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2567 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2569 /// Returns true if it matches
2571 static bool findBuildVector(InsertElementInst *FirstInsertElem,
2572 SmallVectorImpl<Value *> &BuildVector,
2573 SmallVectorImpl<Value *> &BuildVectorOpds) {
2574 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
2577 InsertElementInst *IE = FirstInsertElem;
2579 BuildVector.push_back(IE);
2580 BuildVectorOpds.push_back(IE->getOperand(1));
2582 if (IE->use_empty())
2585 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2589 // If this isn't the final use, make sure the next insertelement is the only
2590 // use. It's OK if the final constructed vector is used multiple times
2591 if (!IE->hasOneUse())
2600 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2601 return V->getType() < V2->getType();
2604 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2605 bool Changed = false;
2606 SmallVector<Value *, 4> Incoming;
2607 SmallSet<Value *, 16> VisitedInstrs;
2609 bool HaveVectorizedPhiNodes = true;
2610 while (HaveVectorizedPhiNodes) {
2611 HaveVectorizedPhiNodes = false;
2613 // Collect the incoming values from the PHIs.
2615 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2617 PHINode *P = dyn_cast<PHINode>(instr);
2621 if (!VisitedInstrs.count(P))
2622 Incoming.push_back(P);
2626 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2628 // Try to vectorize elements base on their type.
2629 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2633 // Look for the next elements with the same type.
2634 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2635 while (SameTypeIt != E &&
2636 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2637 VisitedInstrs.insert(*SameTypeIt);
2641 // Try to vectorize them.
2642 unsigned NumElts = (SameTypeIt - IncIt);
2643 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2645 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2646 // Success start over because instructions might have been changed.
2647 HaveVectorizedPhiNodes = true;
2652 // Start over at the next instruction of a different type (or the end).
2657 VisitedInstrs.clear();
2659 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2660 // We may go through BB multiple times so skip the one we have checked.
2661 if (!VisitedInstrs.insert(it))
2664 if (isa<DbgInfoIntrinsic>(it))
2667 // Try to vectorize reductions that use PHINodes.
2668 if (PHINode *P = dyn_cast<PHINode>(it)) {
2669 // Check that the PHI is a reduction PHI.
2670 if (P->getNumIncomingValues() != 2)
2673 (P->getIncomingBlock(0) == BB
2674 ? (P->getIncomingValue(0))
2675 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
2676 // Check if this is a Binary Operator.
2677 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2681 // Try to match and vectorize a horizontal reduction.
2682 HorizontalReduction HorRdx;
2683 if (ShouldVectorizeHor &&
2684 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2685 HorRdx.tryToReduce(R, TTI)) {
2692 Value *Inst = BI->getOperand(0);
2694 Inst = BI->getOperand(1);
2696 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2697 // We would like to start over since some instructions are deleted
2698 // and the iterator may become invalid value.
2708 // Try to vectorize horizontal reductions feeding into a store.
2709 if (ShouldStartVectorizeHorAtStore)
2710 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2711 if (BinaryOperator *BinOp =
2712 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2713 HorizontalReduction HorRdx;
2714 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) &&
2715 HorRdx.tryToReduce(R, TTI)) ||
2716 tryToVectorize(BinOp, R))) {
2724 // Try to vectorize trees that start at compare instructions.
2725 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2726 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2728 // We would like to start over since some instructions are deleted
2729 // and the iterator may become invalid value.
2735 for (int i = 0; i < 2; ++i) {
2736 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2737 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2739 // We would like to start over since some instructions are deleted
2740 // and the iterator may become invalid value.
2749 // Try to vectorize trees that start at insertelement instructions.
2750 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
2751 SmallVector<Value *, 16> BuildVector;
2752 SmallVector<Value *, 16> BuildVectorOpds;
2753 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
2756 // Vectorize starting with the build vector operands ignoring the
2757 // BuildVector instructions for the purpose of scheduling and user
2759 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
2772 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2773 bool Changed = false;
2774 // Attempt to sort and vectorize each of the store-groups.
2775 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2777 if (it->second.size() < 2)
2780 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2781 << it->second.size() << ".\n");
2783 // Process the stores in chunks of 16.
2784 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2785 unsigned Len = std::min<unsigned>(CE - CI, 16);
2786 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2787 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2793 } // end anonymous namespace
2795 char SLPVectorizer::ID = 0;
2796 static const char lv_name[] = "SLP Vectorizer";
2797 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2798 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2799 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2800 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2801 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2802 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2805 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }