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/NoFolder.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Utils/VectorUtils.h"
48 #define SV_NAME "slp-vectorizer"
49 #define DEBUG_TYPE "SLP"
52 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
53 cl::desc("Only vectorize if you gain more than this "
57 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
58 cl::desc("Attempt to vectorize horizontal reductions"));
60 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
61 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
63 "Attempt to vectorize horizontal reductions feeding into a store"));
67 static const unsigned MinVecRegSize = 128;
69 static const unsigned RecursionMaxDepth = 12;
71 /// A helper class for numbering instructions in multiple blocks.
72 /// Numbers start at zero for each basic block.
73 struct BlockNumbering {
75 BlockNumbering(BasicBlock *Bb) : BB(Bb), 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 = nullptr; // 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, TargetLibraryInfo *TLi, AliasAnalysis *Aa,
349 LoopInfo *Li, DominatorTree *Dt)
350 : F(Func), SE(Se), DL(Dl), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt),
351 Builder(Se->getContext()) {}
353 /// \brief Vectorize the tree that starts with the elements in \p VL.
354 /// Returns the vectorized root.
355 Value *vectorizeTree();
357 /// \returns the vectorization cost of the subtree that starts at \p VL.
358 /// A negative number means that this is profitable.
361 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
362 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
363 void buildTree(ArrayRef<Value *> Roots,
364 ArrayRef<Value *> UserIgnoreLst = None);
366 /// Clear the internal data structures that are created by 'buildTree'.
368 VectorizableTree.clear();
369 ScalarToTreeEntry.clear();
371 ExternalUses.clear();
372 MemBarrierIgnoreList.clear();
375 /// \returns true if the memory operations A and B are consecutive.
376 bool isConsecutiveAccess(Value *A, Value *B);
378 /// \brief Perform LICM and CSE on the newly generated gather sequences.
379 void optimizeGatherSequence();
383 /// \returns the cost of the vectorizable entry.
384 int getEntryCost(TreeEntry *E);
386 /// This is the recursive part of buildTree.
387 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
389 /// Vectorize a single entry in the tree.
390 Value *vectorizeTree(TreeEntry *E);
392 /// Vectorize a single entry in the tree, starting in \p VL.
393 Value *vectorizeTree(ArrayRef<Value *> VL);
395 /// \returns the pointer to the vectorized value if \p VL is already
396 /// vectorized, or NULL. They may happen in cycles.
397 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
399 /// \brief Take the pointer operand from the Load/Store instruction.
400 /// \returns NULL if this is not a valid Load/Store instruction.
401 static Value *getPointerOperand(Value *I);
403 /// \brief Take the address space operand from the Load/Store instruction.
404 /// \returns -1 if this is not a valid Load/Store instruction.
405 static unsigned getAddressSpaceOperand(Value *I);
407 /// \returns the scalarization cost for this type. Scalarization in this
408 /// context means the creation of vectors from a group of scalars.
409 int getGatherCost(Type *Ty);
411 /// \returns the scalarization cost for this list of values. Assuming that
412 /// this subtree gets vectorized, we may need to extract the values from the
413 /// roots. This method calculates the cost of extracting the values.
414 int getGatherCost(ArrayRef<Value *> VL);
416 /// \returns the AA location that is being access by the instruction.
417 AliasAnalysis::Location getLocation(Instruction *I);
419 /// \brief Checks if it is possible to sink an instruction from
420 /// \p Src to \p Dst.
421 /// \returns the pointer to the barrier instruction if we can't sink.
422 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
424 /// \returns the index of the last instruction in the BB from \p VL.
425 int getLastIndex(ArrayRef<Value *> VL);
427 /// \returns the Instruction in the bundle \p VL.
428 Instruction *getLastInstruction(ArrayRef<Value *> VL);
430 /// \brief Set the Builder insert point to one after the last instruction in
432 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
434 /// \returns a vector from a collection of scalars in \p VL.
435 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
437 /// \returns whether the VectorizableTree is fully vectoriable and will
438 /// be beneficial even the tree height is tiny.
439 bool isFullyVectorizableTinyTree();
442 TreeEntry() : Scalars(), VectorizedValue(nullptr), LastScalarIndex(0),
445 /// \returns true if the scalars in VL are equal to this entry.
446 bool isSame(ArrayRef<Value *> VL) const {
447 assert(VL.size() == Scalars.size() && "Invalid size");
448 return std::equal(VL.begin(), VL.end(), Scalars.begin());
451 /// A vector of scalars.
454 /// The Scalars are vectorized into this value. It is initialized to Null.
455 Value *VectorizedValue;
457 /// The index in the basic block of the last scalar.
460 /// Do we need to gather this sequence ?
464 /// Create a new VectorizableTree entry.
465 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
466 VectorizableTree.push_back(TreeEntry());
467 int idx = VectorizableTree.size() - 1;
468 TreeEntry *Last = &VectorizableTree[idx];
469 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
470 Last->NeedToGather = !Vectorized;
472 Last->LastScalarIndex = getLastIndex(VL);
473 for (int i = 0, e = VL.size(); i != e; ++i) {
474 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
475 ScalarToTreeEntry[VL[i]] = idx;
478 Last->LastScalarIndex = 0;
479 MustGather.insert(VL.begin(), VL.end());
484 /// -- Vectorization State --
485 /// Holds all of the tree entries.
486 std::vector<TreeEntry> VectorizableTree;
488 /// Maps a specific scalar to its tree entry.
489 SmallDenseMap<Value*, int> ScalarToTreeEntry;
491 /// A list of scalars that we found that we need to keep as scalars.
494 /// This POD struct describes one external user in the vectorized tree.
495 struct ExternalUser {
496 ExternalUser (Value *S, llvm::User *U, int L) :
497 Scalar(S), User(U), Lane(L){};
498 // Which scalar in our function.
500 // Which user that uses the scalar.
502 // Which lane does the scalar belong to.
505 typedef SmallVector<ExternalUser, 16> UserList;
507 /// A list of values that need to extracted out of the tree.
508 /// This list holds pairs of (Internal Scalar : External User).
509 UserList ExternalUses;
511 /// A list of instructions to ignore while sinking
512 /// memory instructions. This map must be reset between runs of getCost.
513 ValueSet MemBarrierIgnoreList;
515 /// Holds all of the instructions that we gathered.
516 SetVector<Instruction *> GatherSeq;
517 /// A list of blocks that we are going to CSE.
518 SetVector<BasicBlock *> CSEBlocks;
520 /// Numbers instructions in different blocks.
521 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
523 /// \brief Get the corresponding instruction numbering list for a given
524 /// BasicBlock. The list is allocated lazily.
525 BlockNumbering &getBlockNumbering(BasicBlock *BB) {
526 auto I = BlocksNumbers.insert(std::make_pair(BB, BlockNumbering(BB)));
527 return I.first->second;
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;
538 TargetLibraryInfo *TLI;
542 /// Instruction builder to construct the vectorized tree.
546 void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
547 ArrayRef<Value *> UserIgnoreLst) {
549 UserIgnoreList = UserIgnoreLst;
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 users in the user ignore list.
582 if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UserInst) !=
583 UserIgnoreList.end())
586 DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<
587 Lane << " from " << *Scalar << ".\n");
588 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
595 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
596 bool SameTy = getSameType(VL); (void)SameTy;
597 assert(SameTy && "Invalid types!");
599 if (Depth == RecursionMaxDepth) {
600 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
601 newTreeEntry(VL, false);
605 // Don't handle vectors.
606 if (VL[0]->getType()->isVectorTy()) {
607 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
608 newTreeEntry(VL, false);
612 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
613 if (SI->getValueOperand()->getType()->isVectorTy()) {
614 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
615 newTreeEntry(VL, false);
619 // If all of the operands are identical or constant we have a simple solution.
620 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
621 !getSameOpcode(VL)) {
622 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
623 newTreeEntry(VL, false);
627 // We now know that this is a vector of instructions of the same type from
630 // Check if this is a duplicate of another entry.
631 if (ScalarToTreeEntry.count(VL[0])) {
632 int Idx = ScalarToTreeEntry[VL[0]];
633 TreeEntry *E = &VectorizableTree[Idx];
634 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
635 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
636 if (E->Scalars[i] != VL[i]) {
637 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
638 newTreeEntry(VL, false);
642 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
646 // Check that none of the instructions in the bundle are already in the tree.
647 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
648 if (ScalarToTreeEntry.count(VL[i])) {
649 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
650 ") is already in tree.\n");
651 newTreeEntry(VL, false);
656 // If any of the scalars appears in the table OR it is marked as a value that
657 // needs to stat scalar then we need to gather the scalars.
658 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
659 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
660 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
661 newTreeEntry(VL, false);
666 // Check that all of the users of the scalars that we want to vectorize are
668 Instruction *VL0 = cast<Instruction>(VL[0]);
669 int MyLastIndex = getLastIndex(VL);
670 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
672 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
673 Instruction *Scalar = cast<Instruction>(VL[i]);
674 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
675 for (User *U : Scalar->users()) {
676 DEBUG(dbgs() << "SLP: \tUser " << *U << ". \n");
677 Instruction *UI = dyn_cast<Instruction>(U);
679 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
680 newTreeEntry(VL, false);
684 // We don't care if the user is in a different basic block.
685 BasicBlock *UserBlock = UI->getParent();
686 if (UserBlock != BB) {
687 DEBUG(dbgs() << "SLP: User from a different basic block "
692 // If this is a PHINode within this basic block then we can place the
693 // extract wherever we want.
694 if (isa<PHINode>(*UI)) {
695 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *UI << ". \n");
699 // Check if this is a safe in-tree user.
700 if (ScalarToTreeEntry.count(UI)) {
701 int Idx = ScalarToTreeEntry[UI];
702 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
703 if (VecLocation <= MyLastIndex) {
704 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
705 newTreeEntry(VL, false);
708 DEBUG(dbgs() << "SLP: In-tree user (" << *UI << ") at #" <<
709 VecLocation << " vector value (" << *Scalar << ") at #"
710 << MyLastIndex << ".\n");
714 // Ignore users in the user ignore list.
715 if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UI) !=
716 UserIgnoreList.end())
719 // Make sure that we can schedule this unknown user.
720 BlockNumbering &BN = getBlockNumbering(BB);
721 int UserIndex = BN.getIndex(UI);
722 if (UserIndex < MyLastIndex) {
724 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
726 newTreeEntry(VL, false);
732 // Check that every instructions appears once in this bundle.
733 for (unsigned i = 0, e = VL.size(); i < e; ++i)
734 for (unsigned j = i+1; j < e; ++j)
735 if (VL[i] == VL[j]) {
736 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
737 newTreeEntry(VL, false);
741 // Check that instructions in this bundle don't reference other instructions.
742 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
743 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
744 for (User *U : VL[i]->users()) {
745 for (unsigned j = 0; j < e; ++j) {
746 if (i != j && U == VL[j]) {
747 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n");
748 newTreeEntry(VL, false);
755 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
757 unsigned Opcode = getSameOpcode(VL);
759 // Check if it is safe to sink the loads or the stores.
760 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
761 Instruction *Last = getLastInstruction(VL);
763 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
766 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
768 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
769 << "\n because of " << *Barrier << ". Gathering.\n");
770 newTreeEntry(VL, false);
777 case Instruction::PHI: {
778 PHINode *PH = dyn_cast<PHINode>(VL0);
780 // Check for terminator values (e.g. invoke).
781 for (unsigned j = 0; j < VL.size(); ++j)
782 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
783 TerminatorInst *Term = dyn_cast<TerminatorInst>(
784 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
786 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
787 newTreeEntry(VL, false);
792 newTreeEntry(VL, true);
793 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
795 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
797 // Prepare the operand vector.
798 for (unsigned j = 0; j < VL.size(); ++j)
799 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
800 PH->getIncomingBlock(i)));
802 buildTree_rec(Operands, Depth + 1);
806 case Instruction::ExtractElement: {
807 bool Reuse = CanReuseExtract(VL);
809 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
811 newTreeEntry(VL, Reuse);
814 case Instruction::Load: {
815 // Check if the loads are consecutive or of we need to swizzle them.
816 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
817 LoadInst *L = cast<LoadInst>(VL[i]);
818 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
819 newTreeEntry(VL, false);
820 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
824 newTreeEntry(VL, true);
825 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
828 case Instruction::ZExt:
829 case Instruction::SExt:
830 case Instruction::FPToUI:
831 case Instruction::FPToSI:
832 case Instruction::FPExt:
833 case Instruction::PtrToInt:
834 case Instruction::IntToPtr:
835 case Instruction::SIToFP:
836 case Instruction::UIToFP:
837 case Instruction::Trunc:
838 case Instruction::FPTrunc:
839 case Instruction::BitCast: {
840 Type *SrcTy = VL0->getOperand(0)->getType();
841 for (unsigned i = 0; i < VL.size(); ++i) {
842 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
843 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
844 newTreeEntry(VL, false);
845 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
849 newTreeEntry(VL, true);
850 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
852 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
854 // Prepare the operand vector.
855 for (unsigned j = 0; j < VL.size(); ++j)
856 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
858 buildTree_rec(Operands, Depth+1);
862 case Instruction::ICmp:
863 case Instruction::FCmp: {
864 // Check that all of the compares have the same predicate.
865 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
866 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
867 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
868 CmpInst *Cmp = cast<CmpInst>(VL[i]);
869 if (Cmp->getPredicate() != P0 ||
870 Cmp->getOperand(0)->getType() != ComparedTy) {
871 newTreeEntry(VL, false);
872 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
877 newTreeEntry(VL, true);
878 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
880 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
882 // Prepare the operand vector.
883 for (unsigned j = 0; j < VL.size(); ++j)
884 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
886 buildTree_rec(Operands, Depth+1);
890 case Instruction::Select:
891 case Instruction::Add:
892 case Instruction::FAdd:
893 case Instruction::Sub:
894 case Instruction::FSub:
895 case Instruction::Mul:
896 case Instruction::FMul:
897 case Instruction::UDiv:
898 case Instruction::SDiv:
899 case Instruction::FDiv:
900 case Instruction::URem:
901 case Instruction::SRem:
902 case Instruction::FRem:
903 case Instruction::Shl:
904 case Instruction::LShr:
905 case Instruction::AShr:
906 case Instruction::And:
907 case Instruction::Or:
908 case Instruction::Xor: {
909 newTreeEntry(VL, true);
910 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
912 // Sort operands of the instructions so that each side is more likely to
913 // have the same opcode.
914 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
915 ValueList Left, Right;
916 reorderInputsAccordingToOpcode(VL, Left, Right);
917 BasicBlock *LeftBB = getSameBlock(Left);
918 BasicBlock *RightBB = getSameBlock(Right);
919 // If we have common uses on separate paths in the tree make sure we
920 // process the one with greater common depth first.
921 // We can use block numbering to determine the subtree traversal as
922 // earler user has to come in between the common use and the later user.
923 if (LeftBB && RightBB && LeftBB == RightBB &&
924 getLastIndex(Right) > getLastIndex(Left)) {
925 buildTree_rec(Right, Depth + 1);
926 buildTree_rec(Left, Depth + 1);
928 buildTree_rec(Left, Depth + 1);
929 buildTree_rec(Right, Depth + 1);
934 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
936 // Prepare the operand vector.
937 for (unsigned j = 0; j < VL.size(); ++j)
938 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
940 buildTree_rec(Operands, Depth+1);
944 case Instruction::Store: {
945 // Check if the stores are consecutive or of we need to swizzle them.
946 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
947 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
948 newTreeEntry(VL, false);
949 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
953 newTreeEntry(VL, true);
954 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
957 for (unsigned j = 0; j < VL.size(); ++j)
958 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
960 // We can ignore these values because we are sinking them down.
961 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
962 buildTree_rec(Operands, Depth + 1);
965 case Instruction::Call: {
966 // Check if the calls are all to the same vectorizable intrinsic.
967 CallInst *CI = cast<CallInst>(VL[0]);
968 // Check if this is an Intrinsic call or something that can be
969 // represented by an intrinsic call
970 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
971 if (!isTriviallyVectorizable(ID)) {
972 newTreeEntry(VL, false);
973 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
976 Function *Int = CI->getCalledFunction();
977 Value *A1I = nullptr;
978 if (hasVectorInstrinsicScalarOpd(ID, 1))
979 A1I = CI->getArgOperand(1);
980 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
981 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
982 if (!CI2 || CI2->getCalledFunction() != Int ||
983 getIntrinsicIDForCall(CI2, TLI) != ID) {
984 newTreeEntry(VL, false);
985 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]
989 // ctlz,cttz and powi are special intrinsics whose second argument
990 // should be same in order for them to be vectorized.
991 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
992 Value *A1J = CI2->getArgOperand(1);
994 newTreeEntry(VL, false);
995 DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI
996 << " argument "<< A1I<<"!=" << A1J
1003 newTreeEntry(VL, true);
1004 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
1006 // Prepare the operand vector.
1007 for (unsigned j = 0; j < VL.size(); ++j) {
1008 CallInst *CI2 = dyn_cast<CallInst>(VL[j]);
1009 Operands.push_back(CI2->getArgOperand(i));
1011 buildTree_rec(Operands, Depth + 1);
1016 newTreeEntry(VL, false);
1017 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
1022 int BoUpSLP::getEntryCost(TreeEntry *E) {
1023 ArrayRef<Value*> VL = E->Scalars;
1025 Type *ScalarTy = VL[0]->getType();
1026 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1027 ScalarTy = SI->getValueOperand()->getType();
1028 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1030 if (E->NeedToGather) {
1031 if (allConstant(VL))
1034 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1036 return getGatherCost(E->Scalars);
1039 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1041 Instruction *VL0 = cast<Instruction>(VL[0]);
1042 unsigned Opcode = VL0->getOpcode();
1044 case Instruction::PHI: {
1047 case Instruction::ExtractElement: {
1048 if (CanReuseExtract(VL)) {
1050 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1051 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1053 // Take credit for instruction that will become dead.
1055 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1059 return getGatherCost(VecTy);
1061 case Instruction::ZExt:
1062 case Instruction::SExt:
1063 case Instruction::FPToUI:
1064 case Instruction::FPToSI:
1065 case Instruction::FPExt:
1066 case Instruction::PtrToInt:
1067 case Instruction::IntToPtr:
1068 case Instruction::SIToFP:
1069 case Instruction::UIToFP:
1070 case Instruction::Trunc:
1071 case Instruction::FPTrunc:
1072 case Instruction::BitCast: {
1073 Type *SrcTy = VL0->getOperand(0)->getType();
1075 // Calculate the cost of this instruction.
1076 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1077 VL0->getType(), SrcTy);
1079 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1080 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1081 return VecCost - ScalarCost;
1083 case Instruction::FCmp:
1084 case Instruction::ICmp:
1085 case Instruction::Select:
1086 case Instruction::Add:
1087 case Instruction::FAdd:
1088 case Instruction::Sub:
1089 case Instruction::FSub:
1090 case Instruction::Mul:
1091 case Instruction::FMul:
1092 case Instruction::UDiv:
1093 case Instruction::SDiv:
1094 case Instruction::FDiv:
1095 case Instruction::URem:
1096 case Instruction::SRem:
1097 case Instruction::FRem:
1098 case Instruction::Shl:
1099 case Instruction::LShr:
1100 case Instruction::AShr:
1101 case Instruction::And:
1102 case Instruction::Or:
1103 case Instruction::Xor: {
1104 // Calculate the cost of this instruction.
1107 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1108 Opcode == Instruction::Select) {
1109 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1110 ScalarCost = VecTy->getNumElements() *
1111 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1112 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1114 // Certain instructions can be cheaper to vectorize if they have a
1115 // constant second vector operand.
1116 TargetTransformInfo::OperandValueKind Op1VK =
1117 TargetTransformInfo::OK_AnyValue;
1118 TargetTransformInfo::OperandValueKind Op2VK =
1119 TargetTransformInfo::OK_UniformConstantValue;
1121 // If all operands are exactly the same ConstantInt then set the
1122 // operand kind to OK_UniformConstantValue.
1123 // If instead not all operands are constants, then set the operand kind
1124 // to OK_AnyValue. If all operands are constants but not the same,
1125 // then set the operand kind to OK_NonUniformConstantValue.
1126 ConstantInt *CInt = nullptr;
1127 for (unsigned i = 0; i < VL.size(); ++i) {
1128 const Instruction *I = cast<Instruction>(VL[i]);
1129 if (!isa<ConstantInt>(I->getOperand(1))) {
1130 Op2VK = TargetTransformInfo::OK_AnyValue;
1134 CInt = cast<ConstantInt>(I->getOperand(1));
1137 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1138 CInt != cast<ConstantInt>(I->getOperand(1)))
1139 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1143 VecTy->getNumElements() *
1144 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1145 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1147 return VecCost - ScalarCost;
1149 case Instruction::Load: {
1150 // Cost of wide load - cost of scalar loads.
1151 int ScalarLdCost = VecTy->getNumElements() *
1152 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1153 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1154 return VecLdCost - ScalarLdCost;
1156 case Instruction::Store: {
1157 // We know that we can merge the stores. Calculate the cost.
1158 int ScalarStCost = VecTy->getNumElements() *
1159 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1160 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1161 return VecStCost - ScalarStCost;
1163 case Instruction::Call: {
1164 CallInst *CI = cast<CallInst>(VL0);
1165 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1167 // Calculate the cost of the scalar and vector calls.
1168 SmallVector<Type*, 4> ScalarTys, VecTys;
1169 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1170 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1171 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1172 VecTy->getNumElements()));
1175 int ScalarCallCost = VecTy->getNumElements() *
1176 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1178 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1180 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1181 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1182 << " for " << *CI << "\n");
1184 return VecCallCost - ScalarCallCost;
1187 llvm_unreachable("Unknown instruction");
1191 bool BoUpSLP::isFullyVectorizableTinyTree() {
1192 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1193 VectorizableTree.size() << " is fully vectorizable .\n");
1195 // We only handle trees of height 2.
1196 if (VectorizableTree.size() != 2)
1199 // Handle splat stores.
1200 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1203 // Gathering cost would be too much for tiny trees.
1204 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1210 int BoUpSLP::getTreeCost() {
1212 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1213 VectorizableTree.size() << ".\n");
1215 // We only vectorize tiny trees if it is fully vectorizable.
1216 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1217 if (!VectorizableTree.size()) {
1218 assert(!ExternalUses.size() && "We should not have any external users");
1223 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1225 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1226 int C = getEntryCost(&VectorizableTree[i]);
1227 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1228 << *VectorizableTree[i].Scalars[0] << " .\n");
1232 SmallSet<Value *, 16> ExtractCostCalculated;
1233 int ExtractCost = 0;
1234 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1236 // We only add extract cost once for the same scalar.
1237 if (!ExtractCostCalculated.insert(I->Scalar))
1240 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1241 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1245 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1246 return Cost + ExtractCost;
1249 int BoUpSLP::getGatherCost(Type *Ty) {
1251 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1252 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1256 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1257 // Find the type of the operands in VL.
1258 Type *ScalarTy = VL[0]->getType();
1259 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1260 ScalarTy = SI->getValueOperand()->getType();
1261 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1262 // Find the cost of inserting/extracting values from the vector.
1263 return getGatherCost(VecTy);
1266 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1267 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1268 return AA->getLocation(SI);
1269 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1270 return AA->getLocation(LI);
1271 return AliasAnalysis::Location();
1274 Value *BoUpSLP::getPointerOperand(Value *I) {
1275 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1276 return LI->getPointerOperand();
1277 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1278 return SI->getPointerOperand();
1282 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1283 if (LoadInst *L = dyn_cast<LoadInst>(I))
1284 return L->getPointerAddressSpace();
1285 if (StoreInst *S = dyn_cast<StoreInst>(I))
1286 return S->getPointerAddressSpace();
1290 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1291 Value *PtrA = getPointerOperand(A);
1292 Value *PtrB = getPointerOperand(B);
1293 unsigned ASA = getAddressSpaceOperand(A);
1294 unsigned ASB = getAddressSpaceOperand(B);
1296 // Check that the address spaces match and that the pointers are valid.
1297 if (!PtrA || !PtrB || (ASA != ASB))
1300 // Make sure that A and B are different pointers of the same type.
1301 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1304 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1305 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1306 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1308 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1309 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1310 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1312 APInt OffsetDelta = OffsetB - OffsetA;
1314 // Check if they are based on the same pointer. That makes the offsets
1317 return OffsetDelta == Size;
1319 // Compute the necessary base pointer delta to have the necessary final delta
1320 // equal to the size.
1321 APInt BaseDelta = Size - OffsetDelta;
1323 // Otherwise compute the distance with SCEV between the base pointers.
1324 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1325 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1326 const SCEV *C = SE->getConstant(BaseDelta);
1327 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1328 return X == PtrSCEVB;
1331 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1332 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1333 BasicBlock::iterator I = Src, E = Dst;
1334 /// Scan all of the instruction from SRC to DST and check if
1335 /// the source may alias.
1336 for (++I; I != E; ++I) {
1337 // Ignore store instructions that are marked as 'ignore'.
1338 if (MemBarrierIgnoreList.count(I))
1340 if (Src->mayWriteToMemory()) /* Write */ {
1341 if (!I->mayReadOrWriteMemory())
1344 if (!I->mayWriteToMemory())
1347 AliasAnalysis::Location A = getLocation(&*I);
1348 AliasAnalysis::Location B = getLocation(Src);
1350 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1356 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1357 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1358 assert(BB == getSameBlock(VL) && "Invalid block");
1359 BlockNumbering &BN = getBlockNumbering(BB);
1361 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1362 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1363 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1367 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1368 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1369 assert(BB == getSameBlock(VL) && "Invalid block");
1370 BlockNumbering &BN = getBlockNumbering(BB);
1372 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1373 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1374 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1375 Instruction *I = BN.getInstruction(MaxIdx);
1376 assert(I && "bad location");
1380 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1381 Instruction *VL0 = cast<Instruction>(VL[0]);
1382 Instruction *LastInst = getLastInstruction(VL);
1383 BasicBlock::iterator NextInst = LastInst;
1385 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1386 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1389 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1390 Value *Vec = UndefValue::get(Ty);
1391 // Generate the 'InsertElement' instruction.
1392 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1393 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1394 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1395 GatherSeq.insert(Insrt);
1396 CSEBlocks.insert(Insrt->getParent());
1398 // Add to our 'need-to-extract' list.
1399 if (ScalarToTreeEntry.count(VL[i])) {
1400 int Idx = ScalarToTreeEntry[VL[i]];
1401 TreeEntry *E = &VectorizableTree[Idx];
1402 // Find which lane we need to extract.
1404 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1405 // Is this the lane of the scalar that we are looking for ?
1406 if (E->Scalars[Lane] == VL[i]) {
1411 assert(FoundLane >= 0 && "Could not find the correct lane");
1412 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1420 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1421 SmallDenseMap<Value*, int>::const_iterator Entry
1422 = ScalarToTreeEntry.find(VL[0]);
1423 if (Entry != ScalarToTreeEntry.end()) {
1424 int Idx = Entry->second;
1425 const TreeEntry *En = &VectorizableTree[Idx];
1426 if (En->isSame(VL) && En->VectorizedValue)
1427 return En->VectorizedValue;
1432 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1433 if (ScalarToTreeEntry.count(VL[0])) {
1434 int Idx = ScalarToTreeEntry[VL[0]];
1435 TreeEntry *E = &VectorizableTree[Idx];
1437 return vectorizeTree(E);
1440 Type *ScalarTy = VL[0]->getType();
1441 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1442 ScalarTy = SI->getValueOperand()->getType();
1443 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1445 return Gather(VL, VecTy);
1448 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1449 IRBuilder<>::InsertPointGuard Guard(Builder);
1451 if (E->VectorizedValue) {
1452 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1453 return E->VectorizedValue;
1456 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1457 Type *ScalarTy = VL0->getType();
1458 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1459 ScalarTy = SI->getValueOperand()->getType();
1460 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1462 if (E->NeedToGather) {
1463 setInsertPointAfterBundle(E->Scalars);
1464 return Gather(E->Scalars, VecTy);
1467 unsigned Opcode = VL0->getOpcode();
1468 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1471 case Instruction::PHI: {
1472 PHINode *PH = dyn_cast<PHINode>(VL0);
1473 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1474 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1475 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1476 E->VectorizedValue = NewPhi;
1478 // PHINodes may have multiple entries from the same block. We want to
1479 // visit every block once.
1480 SmallSet<BasicBlock*, 4> VisitedBBs;
1482 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1484 BasicBlock *IBB = PH->getIncomingBlock(i);
1486 if (!VisitedBBs.insert(IBB)) {
1487 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1491 // Prepare the operand vector.
1492 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1493 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1494 getIncomingValueForBlock(IBB));
1496 Builder.SetInsertPoint(IBB->getTerminator());
1497 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1498 Value *Vec = vectorizeTree(Operands);
1499 NewPhi->addIncoming(Vec, IBB);
1502 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1503 "Invalid number of incoming values");
1507 case Instruction::ExtractElement: {
1508 if (CanReuseExtract(E->Scalars)) {
1509 Value *V = VL0->getOperand(0);
1510 E->VectorizedValue = V;
1513 return Gather(E->Scalars, VecTy);
1515 case Instruction::ZExt:
1516 case Instruction::SExt:
1517 case Instruction::FPToUI:
1518 case Instruction::FPToSI:
1519 case Instruction::FPExt:
1520 case Instruction::PtrToInt:
1521 case Instruction::IntToPtr:
1522 case Instruction::SIToFP:
1523 case Instruction::UIToFP:
1524 case Instruction::Trunc:
1525 case Instruction::FPTrunc:
1526 case Instruction::BitCast: {
1528 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1529 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1531 setInsertPointAfterBundle(E->Scalars);
1533 Value *InVec = vectorizeTree(INVL);
1535 if (Value *V = alreadyVectorized(E->Scalars))
1538 CastInst *CI = dyn_cast<CastInst>(VL0);
1539 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1540 E->VectorizedValue = V;
1543 case Instruction::FCmp:
1544 case Instruction::ICmp: {
1545 ValueList LHSV, RHSV;
1546 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1547 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1548 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1551 setInsertPointAfterBundle(E->Scalars);
1553 Value *L = vectorizeTree(LHSV);
1554 Value *R = vectorizeTree(RHSV);
1556 if (Value *V = alreadyVectorized(E->Scalars))
1559 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1561 if (Opcode == Instruction::FCmp)
1562 V = Builder.CreateFCmp(P0, L, R);
1564 V = Builder.CreateICmp(P0, L, R);
1566 E->VectorizedValue = V;
1569 case Instruction::Select: {
1570 ValueList TrueVec, FalseVec, CondVec;
1571 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1572 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1573 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1574 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1577 setInsertPointAfterBundle(E->Scalars);
1579 Value *Cond = vectorizeTree(CondVec);
1580 Value *True = vectorizeTree(TrueVec);
1581 Value *False = vectorizeTree(FalseVec);
1583 if (Value *V = alreadyVectorized(E->Scalars))
1586 Value *V = Builder.CreateSelect(Cond, True, False);
1587 E->VectorizedValue = V;
1590 case Instruction::Add:
1591 case Instruction::FAdd:
1592 case Instruction::Sub:
1593 case Instruction::FSub:
1594 case Instruction::Mul:
1595 case Instruction::FMul:
1596 case Instruction::UDiv:
1597 case Instruction::SDiv:
1598 case Instruction::FDiv:
1599 case Instruction::URem:
1600 case Instruction::SRem:
1601 case Instruction::FRem:
1602 case Instruction::Shl:
1603 case Instruction::LShr:
1604 case Instruction::AShr:
1605 case Instruction::And:
1606 case Instruction::Or:
1607 case Instruction::Xor: {
1608 ValueList LHSVL, RHSVL;
1609 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1610 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1612 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1613 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1614 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1617 setInsertPointAfterBundle(E->Scalars);
1619 Value *LHS = vectorizeTree(LHSVL);
1620 Value *RHS = vectorizeTree(RHSVL);
1622 if (LHS == RHS && isa<Instruction>(LHS)) {
1623 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1626 if (Value *V = alreadyVectorized(E->Scalars))
1629 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1630 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1631 E->VectorizedValue = V;
1633 if (Instruction *I = dyn_cast<Instruction>(V))
1634 return propagateMetadata(I, E->Scalars);
1638 case Instruction::Load: {
1639 // Loads are inserted at the head of the tree because we don't want to
1640 // sink them all the way down past store instructions.
1641 setInsertPointAfterBundle(E->Scalars);
1643 LoadInst *LI = cast<LoadInst>(VL0);
1644 unsigned AS = LI->getPointerAddressSpace();
1646 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1647 VecTy->getPointerTo(AS));
1648 unsigned Alignment = LI->getAlignment();
1649 LI = Builder.CreateLoad(VecPtr);
1651 Alignment = DL->getABITypeAlignment(LI->getPointerOperand()->getType());
1652 LI->setAlignment(Alignment);
1653 E->VectorizedValue = LI;
1654 return propagateMetadata(LI, E->Scalars);
1656 case Instruction::Store: {
1657 StoreInst *SI = cast<StoreInst>(VL0);
1658 unsigned Alignment = SI->getAlignment();
1659 unsigned AS = SI->getPointerAddressSpace();
1662 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1663 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1665 setInsertPointAfterBundle(E->Scalars);
1667 Value *VecValue = vectorizeTree(ValueOp);
1668 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1669 VecTy->getPointerTo(AS));
1670 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1672 Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType());
1673 S->setAlignment(Alignment);
1674 E->VectorizedValue = S;
1675 return propagateMetadata(S, E->Scalars);
1677 case Instruction::Call: {
1678 CallInst *CI = cast<CallInst>(VL0);
1679 setInsertPointAfterBundle(E->Scalars);
1681 Intrinsic::ID IID = Intrinsic::not_intrinsic;
1682 if (CI && (FI = CI->getCalledFunction())) {
1683 IID = (Intrinsic::ID) FI->getIntrinsicID();
1685 std::vector<Value *> OpVecs;
1686 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1688 // ctlz,cttz and powi are special intrinsics whose second argument is
1689 // a scalar. This argument should not be vectorized.
1690 if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) {
1691 CallInst *CEI = cast<CallInst>(E->Scalars[0]);
1692 OpVecs.push_back(CEI->getArgOperand(j));
1695 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1696 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1697 OpVL.push_back(CEI->getArgOperand(j));
1700 Value *OpVec = vectorizeTree(OpVL);
1701 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1702 OpVecs.push_back(OpVec);
1705 Module *M = F->getParent();
1706 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1707 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
1708 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
1709 Value *V = Builder.CreateCall(CF, OpVecs);
1710 E->VectorizedValue = V;
1714 llvm_unreachable("unknown inst");
1719 Value *BoUpSLP::vectorizeTree() {
1720 Builder.SetInsertPoint(F->getEntryBlock().begin());
1721 vectorizeTree(&VectorizableTree[0]);
1723 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1725 // Extract all of the elements with the external uses.
1726 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1728 Value *Scalar = it->Scalar;
1729 llvm::User *User = it->User;
1731 // Skip users that we already RAUW. This happens when one instruction
1732 // has multiple uses of the same value.
1733 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
1736 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1738 int Idx = ScalarToTreeEntry[Scalar];
1739 TreeEntry *E = &VectorizableTree[Idx];
1740 assert(!E->NeedToGather && "Extracting from a gather list");
1742 Value *Vec = E->VectorizedValue;
1743 assert(Vec && "Can't find vectorizable value");
1745 Value *Lane = Builder.getInt32(it->Lane);
1746 // Generate extracts for out-of-tree users.
1747 // Find the insertion point for the extractelement lane.
1748 if (isa<Instruction>(Vec)){
1749 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1750 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1751 if (PH->getIncomingValue(i) == Scalar) {
1752 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1753 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1754 CSEBlocks.insert(PH->getIncomingBlock(i));
1755 PH->setOperand(i, Ex);
1759 Builder.SetInsertPoint(cast<Instruction>(User));
1760 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1761 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1762 User->replaceUsesOfWith(Scalar, Ex);
1765 Builder.SetInsertPoint(F->getEntryBlock().begin());
1766 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1767 CSEBlocks.insert(&F->getEntryBlock());
1768 User->replaceUsesOfWith(Scalar, Ex);
1771 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1774 // For each vectorized value:
1775 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1776 TreeEntry *Entry = &VectorizableTree[EIdx];
1779 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1780 Value *Scalar = Entry->Scalars[Lane];
1782 // No need to handle users of gathered values.
1783 if (Entry->NeedToGather)
1786 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1788 Type *Ty = Scalar->getType();
1789 if (!Ty->isVoidTy()) {
1791 for (User *U : Scalar->users()) {
1792 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
1794 assert((ScalarToTreeEntry.count(U) ||
1795 // It is legal to replace users in the ignorelist by undef.
1796 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
1797 UserIgnoreList.end())) &&
1798 "Replacing out-of-tree value with undef");
1801 Value *Undef = UndefValue::get(Ty);
1802 Scalar->replaceAllUsesWith(Undef);
1804 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1805 cast<Instruction>(Scalar)->eraseFromParent();
1809 for (auto &BN : BlocksNumbers)
1812 Builder.ClearInsertionPoint();
1814 return VectorizableTree[0].VectorizedValue;
1817 void BoUpSLP::optimizeGatherSequence() {
1818 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1819 << " gather sequences instructions.\n");
1820 // LICM InsertElementInst sequences.
1821 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1822 e = GatherSeq.end(); it != e; ++it) {
1823 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1828 // Check if this block is inside a loop.
1829 Loop *L = LI->getLoopFor(Insert->getParent());
1833 // Check if it has a preheader.
1834 BasicBlock *PreHeader = L->getLoopPreheader();
1838 // If the vector or the element that we insert into it are
1839 // instructions that are defined in this basic block then we can't
1840 // hoist this instruction.
1841 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1842 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1843 if (CurrVec && L->contains(CurrVec))
1845 if (NewElem && L->contains(NewElem))
1848 // We can hoist this instruction. Move it to the pre-header.
1849 Insert->moveBefore(PreHeader->getTerminator());
1852 // Make a list of all reachable blocks in our CSE queue.
1853 SmallVector<const DomTreeNode *, 8> CSEWorkList;
1854 CSEWorkList.reserve(CSEBlocks.size());
1855 for (BasicBlock *BB : CSEBlocks)
1856 if (DomTreeNode *N = DT->getNode(BB)) {
1857 assert(DT->isReachableFromEntry(N));
1858 CSEWorkList.push_back(N);
1861 // Sort blocks by domination. This ensures we visit a block after all blocks
1862 // dominating it are visited.
1863 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1864 [this](const DomTreeNode *A, const DomTreeNode *B) {
1865 return DT->properlyDominates(A, B);
1868 // Perform O(N^2) search over the gather sequences and merge identical
1869 // instructions. TODO: We can further optimize this scan if we split the
1870 // instructions into different buckets based on the insert lane.
1871 SmallVector<Instruction *, 16> Visited;
1872 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
1873 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1874 "Worklist not sorted properly!");
1875 BasicBlock *BB = (*I)->getBlock();
1876 // For all instructions in blocks containing gather sequences:
1877 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1878 Instruction *In = it++;
1879 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1882 // Check if we can replace this instruction with any of the
1883 // visited instructions.
1884 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1887 if (In->isIdenticalTo(*v) &&
1888 DT->dominates((*v)->getParent(), In->getParent())) {
1889 In->replaceAllUsesWith(*v);
1890 In->eraseFromParent();
1896 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1897 Visited.push_back(In);
1905 /// The SLPVectorizer Pass.
1906 struct SLPVectorizer : public FunctionPass {
1907 typedef SmallVector<StoreInst *, 8> StoreList;
1908 typedef MapVector<Value *, StoreList> StoreListMap;
1910 /// Pass identification, replacement for typeid
1913 explicit SLPVectorizer() : FunctionPass(ID) {
1914 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1917 ScalarEvolution *SE;
1918 const DataLayout *DL;
1919 TargetTransformInfo *TTI;
1920 TargetLibraryInfo *TLI;
1925 bool runOnFunction(Function &F) override {
1926 if (skipOptnoneFunction(F))
1929 SE = &getAnalysis<ScalarEvolution>();
1930 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1931 DL = DLP ? &DLP->getDataLayout() : nullptr;
1932 TTI = &getAnalysis<TargetTransformInfo>();
1933 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1934 AA = &getAnalysis<AliasAnalysis>();
1935 LI = &getAnalysis<LoopInfo>();
1936 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1939 bool Changed = false;
1941 // If the target claims to have no vector registers don't attempt
1943 if (!TTI->getNumberOfRegisters(true))
1946 // Must have DataLayout. We can't require it because some tests run w/o
1951 // Don't vectorize when the attribute NoImplicitFloat is used.
1952 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1955 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1957 // Use the bottom up slp vectorizer to construct chains that start with
1958 // store instructions.
1959 BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT);
1961 // Scan the blocks in the function in post order.
1962 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1963 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1964 BasicBlock *BB = *it;
1966 // Vectorize trees that end at stores.
1967 if (unsigned count = collectStores(BB, R)) {
1969 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1970 Changed |= vectorizeStoreChains(R);
1973 // Vectorize trees that end at reductions.
1974 Changed |= vectorizeChainsInBlock(BB, R);
1978 R.optimizeGatherSequence();
1979 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1980 DEBUG(verifyFunction(F));
1985 void getAnalysisUsage(AnalysisUsage &AU) const override {
1986 FunctionPass::getAnalysisUsage(AU);
1987 AU.addRequired<ScalarEvolution>();
1988 AU.addRequired<AliasAnalysis>();
1989 AU.addRequired<TargetTransformInfo>();
1990 AU.addRequired<LoopInfo>();
1991 AU.addRequired<DominatorTreeWrapperPass>();
1992 AU.addPreserved<LoopInfo>();
1993 AU.addPreserved<DominatorTreeWrapperPass>();
1994 AU.setPreservesCFG();
1999 /// \brief Collect memory references and sort them according to their base
2000 /// object. We sort the stores to their base objects to reduce the cost of the
2001 /// quadratic search on the stores. TODO: We can further reduce this cost
2002 /// if we flush the chain creation every time we run into a memory barrier.
2003 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
2005 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
2006 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
2008 /// \brief Try to vectorize a list of operands.
2009 /// \@param BuildVector A list of users to ignore for the purpose of
2010 /// scheduling and that don't need extracting.
2011 /// \returns true if a value was vectorized.
2012 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2013 ArrayRef<Value *> BuildVector = None);
2015 /// \brief Try to vectorize a chain that may start at the operands of \V;
2016 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
2018 /// \brief Vectorize the stores that were collected in StoreRefs.
2019 bool vectorizeStoreChains(BoUpSLP &R);
2021 /// \brief Scan the basic block and look for patterns that are likely to start
2022 /// a vectorization chain.
2023 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
2025 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
2028 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
2031 StoreListMap StoreRefs;
2034 /// \brief Check that the Values in the slice in VL array are still existent in
2035 /// the WeakVH array.
2036 /// Vectorization of part of the VL array may cause later values in the VL array
2037 /// to become invalid. We track when this has happened in the WeakVH array.
2038 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
2039 SmallVectorImpl<WeakVH> &VH,
2040 unsigned SliceBegin,
2041 unsigned SliceSize) {
2042 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
2049 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
2050 int CostThreshold, BoUpSLP &R) {
2051 unsigned ChainLen = Chain.size();
2052 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2054 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2055 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2056 unsigned VF = MinVecRegSize / Sz;
2058 if (!isPowerOf2_32(Sz) || VF < 2)
2061 // Keep track of values that were deleted by vectorizing in the loop below.
2062 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2064 bool Changed = false;
2065 // Look for profitable vectorizable trees at all offsets, starting at zero.
2066 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2070 // Check that a previous iteration of this loop did not delete the Value.
2071 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2074 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2076 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2078 R.buildTree(Operands);
2080 int Cost = R.getTreeCost();
2082 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2083 if (Cost < CostThreshold) {
2084 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2087 // Move to the next bundle.
2096 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2097 int costThreshold, BoUpSLP &R) {
2098 SetVector<Value *> Heads, Tails;
2099 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2101 // We may run into multiple chains that merge into a single chain. We mark the
2102 // stores that we vectorized so that we don't visit the same store twice.
2103 BoUpSLP::ValueSet VectorizedStores;
2104 bool Changed = false;
2106 // Do a quadratic search on all of the given stores and find
2107 // all of the pairs of stores that follow each other.
2108 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2109 for (unsigned j = 0; j < e; ++j) {
2113 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2114 Tails.insert(Stores[j]);
2115 Heads.insert(Stores[i]);
2116 ConsecutiveChain[Stores[i]] = Stores[j];
2121 // For stores that start but don't end a link in the chain:
2122 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2124 if (Tails.count(*it))
2127 // We found a store instr that starts a chain. Now follow the chain and try
2129 BoUpSLP::ValueList Operands;
2131 // Collect the chain into a list.
2132 while (Tails.count(I) || Heads.count(I)) {
2133 if (VectorizedStores.count(I))
2135 Operands.push_back(I);
2136 // Move to the next value in the chain.
2137 I = ConsecutiveChain[I];
2140 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2142 // Mark the vectorized stores so that we don't vectorize them again.
2144 VectorizedStores.insert(Operands.begin(), Operands.end());
2145 Changed |= Vectorized;
2152 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2155 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2156 StoreInst *SI = dyn_cast<StoreInst>(it);
2160 // Don't touch volatile stores.
2161 if (!SI->isSimple())
2164 // Check that the pointer points to scalars.
2165 Type *Ty = SI->getValueOperand()->getType();
2166 if (Ty->isAggregateType() || Ty->isVectorTy())
2169 // Find the base pointer.
2170 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2172 // Save the store locations.
2173 StoreRefs[Ptr].push_back(SI);
2179 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2182 Value *VL[] = { A, B };
2183 return tryToVectorizeList(VL, R);
2186 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2187 ArrayRef<Value *> BuildVector) {
2191 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2193 // Check that all of the parts are scalar instructions of the same type.
2194 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2198 unsigned Opcode0 = I0->getOpcode();
2200 Type *Ty0 = I0->getType();
2201 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2202 unsigned VF = MinVecRegSize / Sz;
2204 for (int i = 0, e = VL.size(); i < e; ++i) {
2205 Type *Ty = VL[i]->getType();
2206 if (Ty->isAggregateType() || Ty->isVectorTy())
2208 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2209 if (!Inst || Inst->getOpcode() != Opcode0)
2213 bool Changed = false;
2215 // Keep track of values that were deleted by vectorizing in the loop below.
2216 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2218 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2219 unsigned OpsWidth = 0;
2226 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2229 // Check that a previous iteration of this loop did not delete the Value.
2230 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2233 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2235 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2237 ArrayRef<Value *> BuildVectorSlice;
2238 if (!BuildVector.empty())
2239 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
2241 R.buildTree(Ops, BuildVectorSlice);
2242 int Cost = R.getTreeCost();
2244 if (Cost < -SLPCostThreshold) {
2245 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
2246 Value *VectorizedRoot = R.vectorizeTree();
2248 // Reconstruct the build vector by extracting the vectorized root. This
2249 // way we handle the case where some elements of the vector are undefined.
2250 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
2251 if (!BuildVectorSlice.empty()) {
2252 // The insert point is the last build vector instruction. The vectorized
2253 // root will precede it. This guarantees that we get an instruction. The
2254 // vectorized tree could have been constant folded.
2255 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
2256 unsigned VecIdx = 0;
2257 for (auto &V : BuildVectorSlice) {
2258 IRBuilder<true, NoFolder> Builder(
2259 ++BasicBlock::iterator(InsertAfter));
2260 InsertElementInst *IE = cast<InsertElementInst>(V);
2261 Instruction *Extract = cast<Instruction>(Builder.CreateExtractElement(
2262 VectorizedRoot, Builder.getInt32(VecIdx++)));
2263 IE->setOperand(1, Extract);
2264 IE->removeFromParent();
2265 IE->insertAfter(Extract);
2269 // Move to the next bundle.
2278 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2282 // Try to vectorize V.
2283 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2286 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2287 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2289 if (B && B->hasOneUse()) {
2290 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2291 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2292 if (tryToVectorizePair(A, B0, R)) {
2296 if (tryToVectorizePair(A, B1, R)) {
2303 if (A && A->hasOneUse()) {
2304 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2305 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2306 if (tryToVectorizePair(A0, B, R)) {
2310 if (tryToVectorizePair(A1, B, R)) {
2318 /// \brief Generate a shuffle mask to be used in a reduction tree.
2320 /// \param VecLen The length of the vector to be reduced.
2321 /// \param NumEltsToRdx The number of elements that should be reduced in the
2323 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2324 /// reduction. A pairwise reduction will generate a mask of
2325 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2326 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2327 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2328 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2329 bool IsPairwise, bool IsLeft,
2330 IRBuilder<> &Builder) {
2331 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2333 SmallVector<Constant *, 32> ShuffleMask(
2334 VecLen, UndefValue::get(Builder.getInt32Ty()));
2337 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2338 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2339 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2341 // Move the upper half of the vector to the lower half.
2342 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2343 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2345 return ConstantVector::get(ShuffleMask);
2349 /// Model horizontal reductions.
2351 /// A horizontal reduction is a tree of reduction operations (currently add and
2352 /// fadd) that has operations that can be put into a vector as its leaf.
2353 /// For example, this tree:
2360 /// This tree has "mul" as its reduced values and "+" as its reduction
2361 /// operations. A reduction might be feeding into a store or a binary operation
2376 class HorizontalReduction {
2377 SmallVector<Value *, 16> ReductionOps;
2378 SmallVector<Value *, 32> ReducedVals;
2380 BinaryOperator *ReductionRoot;
2381 PHINode *ReductionPHI;
2383 /// The opcode of the reduction.
2384 unsigned ReductionOpcode;
2385 /// The opcode of the values we perform a reduction on.
2386 unsigned ReducedValueOpcode;
2387 /// The width of one full horizontal reduction operation.
2388 unsigned ReduxWidth;
2389 /// Should we model this reduction as a pairwise reduction tree or a tree that
2390 /// splits the vector in halves and adds those halves.
2391 bool IsPairwiseReduction;
2394 HorizontalReduction()
2395 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
2396 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2398 /// \brief Try to find a reduction tree.
2399 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2400 const DataLayout *DL) {
2402 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2403 "Thi phi needs to use the binary operator");
2405 // We could have a initial reductions that is not an add.
2406 // r *= v1 + v2 + v3 + v4
2407 // In such a case start looking for a tree rooted in the first '+'.
2409 if (B->getOperand(0) == Phi) {
2411 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2412 } else if (B->getOperand(1) == Phi) {
2414 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2421 Type *Ty = B->getType();
2422 if (Ty->isVectorTy())
2425 ReductionOpcode = B->getOpcode();
2426 ReducedValueOpcode = 0;
2427 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2434 // We currently only support adds.
2435 if (ReductionOpcode != Instruction::Add &&
2436 ReductionOpcode != Instruction::FAdd)
2439 // Post order traverse the reduction tree starting at B. We only handle true
2440 // trees containing only binary operators.
2441 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2442 Stack.push_back(std::make_pair(B, 0));
2443 while (!Stack.empty()) {
2444 BinaryOperator *TreeN = Stack.back().first;
2445 unsigned EdgeToVist = Stack.back().second++;
2446 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2448 // Only handle trees in the current basic block.
2449 if (TreeN->getParent() != B->getParent())
2452 // Each tree node needs to have one user except for the ultimate
2454 if (!TreeN->hasOneUse() && TreeN != B)
2458 if (EdgeToVist == 2 || IsReducedValue) {
2459 if (IsReducedValue) {
2460 // Make sure that the opcodes of the operations that we are going to
2462 if (!ReducedValueOpcode)
2463 ReducedValueOpcode = TreeN->getOpcode();
2464 else if (ReducedValueOpcode != TreeN->getOpcode())
2466 ReducedVals.push_back(TreeN);
2468 // We need to be able to reassociate the adds.
2469 if (!TreeN->isAssociative())
2471 ReductionOps.push_back(TreeN);
2478 // Visit left or right.
2479 Value *NextV = TreeN->getOperand(EdgeToVist);
2480 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2482 Stack.push_back(std::make_pair(Next, 0));
2483 else if (NextV != Phi)
2489 /// \brief Attempt to vectorize the tree found by
2490 /// matchAssociativeReduction.
2491 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2492 if (ReducedVals.empty())
2495 unsigned NumReducedVals = ReducedVals.size();
2496 if (NumReducedVals < ReduxWidth)
2499 Value *VectorizedTree = nullptr;
2500 IRBuilder<> Builder(ReductionRoot);
2501 FastMathFlags Unsafe;
2502 Unsafe.setUnsafeAlgebra();
2503 Builder.SetFastMathFlags(Unsafe);
2506 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2507 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2508 V.buildTree(ValsToReduce, ReductionOps);
2511 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2512 if (Cost >= -SLPCostThreshold)
2515 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2518 // Vectorize a tree.
2519 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2520 Value *VectorizedRoot = V.vectorizeTree();
2522 // Emit a reduction.
2523 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2524 if (VectorizedTree) {
2525 Builder.SetCurrentDebugLocation(Loc);
2526 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2527 ReducedSubTree, "bin.rdx");
2529 VectorizedTree = ReducedSubTree;
2532 if (VectorizedTree) {
2533 // Finish the reduction.
2534 for (; i < NumReducedVals; ++i) {
2535 Builder.SetCurrentDebugLocation(
2536 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2537 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2542 assert(ReductionRoot && "Need a reduction operation");
2543 ReductionRoot->setOperand(0, VectorizedTree);
2544 ReductionRoot->setOperand(1, ReductionPHI);
2546 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2548 return VectorizedTree != nullptr;
2553 /// \brief Calcuate the cost of a reduction.
2554 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2555 Type *ScalarTy = FirstReducedVal->getType();
2556 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2558 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2559 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2561 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2562 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2564 int ScalarReduxCost =
2565 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2567 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2568 << " for reduction that starts with " << *FirstReducedVal
2570 << (IsPairwiseReduction ? "pairwise" : "splitting")
2571 << " reduction)\n");
2573 return VecReduxCost - ScalarReduxCost;
2576 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2577 Value *R, const Twine &Name = "") {
2578 if (Opcode == Instruction::FAdd)
2579 return Builder.CreateFAdd(L, R, Name);
2580 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2583 /// \brief Emit a horizontal reduction of the vectorized value.
2584 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2585 assert(VectorizedValue && "Need to have a vectorized tree node");
2586 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2587 assert(isPowerOf2_32(ReduxWidth) &&
2588 "We only handle power-of-two reductions for now");
2590 Value *TmpVec = ValToReduce;
2591 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2592 if (IsPairwiseReduction) {
2594 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2596 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2598 Value *LeftShuf = Builder.CreateShuffleVector(
2599 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2600 Value *RightShuf = Builder.CreateShuffleVector(
2601 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2603 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2607 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2608 Value *Shuf = Builder.CreateShuffleVector(
2609 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2610 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2614 // The result is in the first element of the vector.
2615 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2619 /// \brief Recognize construction of vectors like
2620 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2621 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2622 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2623 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2625 /// Returns true if it matches
2627 static bool findBuildVector(InsertElementInst *FirstInsertElem,
2628 SmallVectorImpl<Value *> &BuildVector,
2629 SmallVectorImpl<Value *> &BuildVectorOpds) {
2630 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
2633 InsertElementInst *IE = FirstInsertElem;
2635 BuildVector.push_back(IE);
2636 BuildVectorOpds.push_back(IE->getOperand(1));
2638 if (IE->use_empty())
2641 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2645 // If this isn't the final use, make sure the next insertelement is the only
2646 // use. It's OK if the final constructed vector is used multiple times
2647 if (!IE->hasOneUse())
2656 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2657 return V->getType() < V2->getType();
2660 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2661 bool Changed = false;
2662 SmallVector<Value *, 4> Incoming;
2663 SmallSet<Value *, 16> VisitedInstrs;
2665 bool HaveVectorizedPhiNodes = true;
2666 while (HaveVectorizedPhiNodes) {
2667 HaveVectorizedPhiNodes = false;
2669 // Collect the incoming values from the PHIs.
2671 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2673 PHINode *P = dyn_cast<PHINode>(instr);
2677 if (!VisitedInstrs.count(P))
2678 Incoming.push_back(P);
2682 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2684 // Try to vectorize elements base on their type.
2685 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2689 // Look for the next elements with the same type.
2690 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2691 while (SameTypeIt != E &&
2692 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2693 VisitedInstrs.insert(*SameTypeIt);
2697 // Try to vectorize them.
2698 unsigned NumElts = (SameTypeIt - IncIt);
2699 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2701 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2702 // Success start over because instructions might have been changed.
2703 HaveVectorizedPhiNodes = true;
2708 // Start over at the next instruction of a different type (or the end).
2713 VisitedInstrs.clear();
2715 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2716 // We may go through BB multiple times so skip the one we have checked.
2717 if (!VisitedInstrs.insert(it))
2720 if (isa<DbgInfoIntrinsic>(it))
2723 // Try to vectorize reductions that use PHINodes.
2724 if (PHINode *P = dyn_cast<PHINode>(it)) {
2725 // Check that the PHI is a reduction PHI.
2726 if (P->getNumIncomingValues() != 2)
2729 (P->getIncomingBlock(0) == BB
2730 ? (P->getIncomingValue(0))
2731 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
2733 // Check if this is a Binary Operator.
2734 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2738 // Try to match and vectorize a horizontal reduction.
2739 HorizontalReduction HorRdx;
2740 if (ShouldVectorizeHor &&
2741 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2742 HorRdx.tryToReduce(R, TTI)) {
2749 Value *Inst = BI->getOperand(0);
2751 Inst = BI->getOperand(1);
2753 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2754 // We would like to start over since some instructions are deleted
2755 // and the iterator may become invalid value.
2765 // Try to vectorize horizontal reductions feeding into a store.
2766 if (ShouldStartVectorizeHorAtStore)
2767 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2768 if (BinaryOperator *BinOp =
2769 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2770 HorizontalReduction HorRdx;
2771 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
2772 HorRdx.tryToReduce(R, TTI)) ||
2773 tryToVectorize(BinOp, R))) {
2781 // Try to vectorize trees that start at compare instructions.
2782 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2783 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2785 // We would like to start over since some instructions are deleted
2786 // and the iterator may become invalid value.
2792 for (int i = 0; i < 2; ++i) {
2793 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2794 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2796 // We would like to start over since some instructions are deleted
2797 // and the iterator may become invalid value.
2806 // Try to vectorize trees that start at insertelement instructions.
2807 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
2808 SmallVector<Value *, 16> BuildVector;
2809 SmallVector<Value *, 16> BuildVectorOpds;
2810 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
2813 // Vectorize starting with the build vector operands ignoring the
2814 // BuildVector instructions for the purpose of scheduling and user
2816 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
2829 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2830 bool Changed = false;
2831 // Attempt to sort and vectorize each of the store-groups.
2832 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2834 if (it->second.size() < 2)
2837 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2838 << it->second.size() << ".\n");
2840 // Process the stores in chunks of 16.
2841 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2842 unsigned Len = std::min<unsigned>(CE - CI, 16);
2843 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2844 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2850 } // end anonymous namespace
2852 char SLPVectorizer::ID = 0;
2853 static const char lv_name[] = "SLP Vectorizer";
2854 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2855 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2856 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2857 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2858 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2859 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2862 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }