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 buildTree_rec(Left, Depth + 1);
918 buildTree_rec(Right, Depth + 1);
922 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
924 // Prepare the operand vector.
925 for (unsigned j = 0; j < VL.size(); ++j)
926 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
928 buildTree_rec(Operands, Depth+1);
932 case Instruction::Store: {
933 // Check if the stores are consecutive or of we need to swizzle them.
934 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
935 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
936 newTreeEntry(VL, false);
937 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
941 newTreeEntry(VL, true);
942 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
945 for (unsigned j = 0; j < VL.size(); ++j)
946 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
948 // We can ignore these values because we are sinking them down.
949 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
950 buildTree_rec(Operands, Depth + 1);
953 case Instruction::Call: {
954 // Check if the calls are all to the same vectorizable intrinsic.
955 CallInst *CI = cast<CallInst>(VL[0]);
956 // Check if this is an Intrinsic call or something that can be
957 // represented by an intrinsic call
958 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
959 if (!isTriviallyVectorizable(ID)) {
960 newTreeEntry(VL, false);
961 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
965 Function *Int = CI->getCalledFunction();
967 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
968 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
969 if (!CI2 || CI2->getCalledFunction() != Int ||
970 getIntrinsicIDForCall(CI2, TLI) != ID) {
971 newTreeEntry(VL, false);
972 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]
978 newTreeEntry(VL, true);
979 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
981 // Prepare the operand vector.
982 for (unsigned j = 0; j < VL.size(); ++j) {
983 CallInst *CI2 = dyn_cast<CallInst>(VL[j]);
984 Operands.push_back(CI2->getArgOperand(i));
986 buildTree_rec(Operands, Depth + 1);
991 newTreeEntry(VL, false);
992 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
997 int BoUpSLP::getEntryCost(TreeEntry *E) {
998 ArrayRef<Value*> VL = E->Scalars;
1000 Type *ScalarTy = VL[0]->getType();
1001 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1002 ScalarTy = SI->getValueOperand()->getType();
1003 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1005 if (E->NeedToGather) {
1006 if (allConstant(VL))
1009 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1011 return getGatherCost(E->Scalars);
1014 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1016 Instruction *VL0 = cast<Instruction>(VL[0]);
1017 unsigned Opcode = VL0->getOpcode();
1019 case Instruction::PHI: {
1022 case Instruction::ExtractElement: {
1023 if (CanReuseExtract(VL)) {
1025 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1026 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1028 // Take credit for instruction that will become dead.
1030 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1034 return getGatherCost(VecTy);
1036 case Instruction::ZExt:
1037 case Instruction::SExt:
1038 case Instruction::FPToUI:
1039 case Instruction::FPToSI:
1040 case Instruction::FPExt:
1041 case Instruction::PtrToInt:
1042 case Instruction::IntToPtr:
1043 case Instruction::SIToFP:
1044 case Instruction::UIToFP:
1045 case Instruction::Trunc:
1046 case Instruction::FPTrunc:
1047 case Instruction::BitCast: {
1048 Type *SrcTy = VL0->getOperand(0)->getType();
1050 // Calculate the cost of this instruction.
1051 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1052 VL0->getType(), SrcTy);
1054 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1055 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1056 return VecCost - ScalarCost;
1058 case Instruction::FCmp:
1059 case Instruction::ICmp:
1060 case Instruction::Select:
1061 case Instruction::Add:
1062 case Instruction::FAdd:
1063 case Instruction::Sub:
1064 case Instruction::FSub:
1065 case Instruction::Mul:
1066 case Instruction::FMul:
1067 case Instruction::UDiv:
1068 case Instruction::SDiv:
1069 case Instruction::FDiv:
1070 case Instruction::URem:
1071 case Instruction::SRem:
1072 case Instruction::FRem:
1073 case Instruction::Shl:
1074 case Instruction::LShr:
1075 case Instruction::AShr:
1076 case Instruction::And:
1077 case Instruction::Or:
1078 case Instruction::Xor: {
1079 // Calculate the cost of this instruction.
1082 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1083 Opcode == Instruction::Select) {
1084 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1085 ScalarCost = VecTy->getNumElements() *
1086 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1087 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1089 // Certain instructions can be cheaper to vectorize if they have a
1090 // constant second vector operand.
1091 TargetTransformInfo::OperandValueKind Op1VK =
1092 TargetTransformInfo::OK_AnyValue;
1093 TargetTransformInfo::OperandValueKind Op2VK =
1094 TargetTransformInfo::OK_UniformConstantValue;
1096 // If all operands are exactly the same ConstantInt then set the
1097 // operand kind to OK_UniformConstantValue.
1098 // If instead not all operands are constants, then set the operand kind
1099 // to OK_AnyValue. If all operands are constants but not the same,
1100 // then set the operand kind to OK_NonUniformConstantValue.
1101 ConstantInt *CInt = nullptr;
1102 for (unsigned i = 0; i < VL.size(); ++i) {
1103 const Instruction *I = cast<Instruction>(VL[i]);
1104 if (!isa<ConstantInt>(I->getOperand(1))) {
1105 Op2VK = TargetTransformInfo::OK_AnyValue;
1109 CInt = cast<ConstantInt>(I->getOperand(1));
1112 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1113 CInt != cast<ConstantInt>(I->getOperand(1)))
1114 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1118 VecTy->getNumElements() *
1119 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1120 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1122 return VecCost - ScalarCost;
1124 case Instruction::Load: {
1125 // Cost of wide load - cost of scalar loads.
1126 int ScalarLdCost = VecTy->getNumElements() *
1127 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1128 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1129 return VecLdCost - ScalarLdCost;
1131 case Instruction::Store: {
1132 // We know that we can merge the stores. Calculate the cost.
1133 int ScalarStCost = VecTy->getNumElements() *
1134 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1135 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1136 return VecStCost - ScalarStCost;
1138 case Instruction::Call: {
1139 CallInst *CI = cast<CallInst>(VL0);
1140 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1142 // Calculate the cost of the scalar and vector calls.
1143 SmallVector<Type*, 4> ScalarTys, VecTys;
1144 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1145 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1146 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1147 VecTy->getNumElements()));
1150 int ScalarCallCost = VecTy->getNumElements() *
1151 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1153 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1155 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1156 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1157 << " for " << *CI << "\n");
1159 return VecCallCost - ScalarCallCost;
1162 llvm_unreachable("Unknown instruction");
1166 bool BoUpSLP::isFullyVectorizableTinyTree() {
1167 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1168 VectorizableTree.size() << " is fully vectorizable .\n");
1170 // We only handle trees of height 2.
1171 if (VectorizableTree.size() != 2)
1174 // Handle splat stores.
1175 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1178 // Gathering cost would be too much for tiny trees.
1179 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1185 int BoUpSLP::getTreeCost() {
1187 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1188 VectorizableTree.size() << ".\n");
1190 // We only vectorize tiny trees if it is fully vectorizable.
1191 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1192 if (!VectorizableTree.size()) {
1193 assert(!ExternalUses.size() && "We should not have any external users");
1198 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1200 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1201 int C = getEntryCost(&VectorizableTree[i]);
1202 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1203 << *VectorizableTree[i].Scalars[0] << " .\n");
1207 SmallSet<Value *, 16> ExtractCostCalculated;
1208 int ExtractCost = 0;
1209 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1211 // We only add extract cost once for the same scalar.
1212 if (!ExtractCostCalculated.insert(I->Scalar))
1215 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1216 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1220 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1221 return Cost + ExtractCost;
1224 int BoUpSLP::getGatherCost(Type *Ty) {
1226 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1227 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1231 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1232 // Find the type of the operands in VL.
1233 Type *ScalarTy = VL[0]->getType();
1234 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1235 ScalarTy = SI->getValueOperand()->getType();
1236 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1237 // Find the cost of inserting/extracting values from the vector.
1238 return getGatherCost(VecTy);
1241 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1242 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1243 return AA->getLocation(SI);
1244 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1245 return AA->getLocation(LI);
1246 return AliasAnalysis::Location();
1249 Value *BoUpSLP::getPointerOperand(Value *I) {
1250 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1251 return LI->getPointerOperand();
1252 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1253 return SI->getPointerOperand();
1257 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1258 if (LoadInst *L = dyn_cast<LoadInst>(I))
1259 return L->getPointerAddressSpace();
1260 if (StoreInst *S = dyn_cast<StoreInst>(I))
1261 return S->getPointerAddressSpace();
1265 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1266 Value *PtrA = getPointerOperand(A);
1267 Value *PtrB = getPointerOperand(B);
1268 unsigned ASA = getAddressSpaceOperand(A);
1269 unsigned ASB = getAddressSpaceOperand(B);
1271 // Check that the address spaces match and that the pointers are valid.
1272 if (!PtrA || !PtrB || (ASA != ASB))
1275 // Make sure that A and B are different pointers of the same type.
1276 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1279 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1280 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1281 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1283 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1284 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1285 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1287 APInt OffsetDelta = OffsetB - OffsetA;
1289 // Check if they are based on the same pointer. That makes the offsets
1292 return OffsetDelta == Size;
1294 // Compute the necessary base pointer delta to have the necessary final delta
1295 // equal to the size.
1296 APInt BaseDelta = Size - OffsetDelta;
1298 // Otherwise compute the distance with SCEV between the base pointers.
1299 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1300 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1301 const SCEV *C = SE->getConstant(BaseDelta);
1302 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1303 return X == PtrSCEVB;
1306 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1307 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1308 BasicBlock::iterator I = Src, E = Dst;
1309 /// Scan all of the instruction from SRC to DST and check if
1310 /// the source may alias.
1311 for (++I; I != E; ++I) {
1312 // Ignore store instructions that are marked as 'ignore'.
1313 if (MemBarrierIgnoreList.count(I))
1315 if (Src->mayWriteToMemory()) /* Write */ {
1316 if (!I->mayReadOrWriteMemory())
1319 if (!I->mayWriteToMemory())
1322 AliasAnalysis::Location A = getLocation(&*I);
1323 AliasAnalysis::Location B = getLocation(Src);
1325 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1331 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1332 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1333 assert(BB == getSameBlock(VL) && "Invalid block");
1334 BlockNumbering &BN = getBlockNumbering(BB);
1336 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1337 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1338 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1342 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1343 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1344 assert(BB == getSameBlock(VL) && "Invalid block");
1345 BlockNumbering &BN = getBlockNumbering(BB);
1347 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1348 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1349 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1350 Instruction *I = BN.getInstruction(MaxIdx);
1351 assert(I && "bad location");
1355 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1356 Instruction *VL0 = cast<Instruction>(VL[0]);
1357 Instruction *LastInst = getLastInstruction(VL);
1358 BasicBlock::iterator NextInst = LastInst;
1360 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1361 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1364 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1365 Value *Vec = UndefValue::get(Ty);
1366 // Generate the 'InsertElement' instruction.
1367 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1368 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1369 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1370 GatherSeq.insert(Insrt);
1371 CSEBlocks.insert(Insrt->getParent());
1373 // Add to our 'need-to-extract' list.
1374 if (ScalarToTreeEntry.count(VL[i])) {
1375 int Idx = ScalarToTreeEntry[VL[i]];
1376 TreeEntry *E = &VectorizableTree[Idx];
1377 // Find which lane we need to extract.
1379 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1380 // Is this the lane of the scalar that we are looking for ?
1381 if (E->Scalars[Lane] == VL[i]) {
1386 assert(FoundLane >= 0 && "Could not find the correct lane");
1387 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1395 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1396 SmallDenseMap<Value*, int>::const_iterator Entry
1397 = ScalarToTreeEntry.find(VL[0]);
1398 if (Entry != ScalarToTreeEntry.end()) {
1399 int Idx = Entry->second;
1400 const TreeEntry *En = &VectorizableTree[Idx];
1401 if (En->isSame(VL) && En->VectorizedValue)
1402 return En->VectorizedValue;
1407 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1408 if (ScalarToTreeEntry.count(VL[0])) {
1409 int Idx = ScalarToTreeEntry[VL[0]];
1410 TreeEntry *E = &VectorizableTree[Idx];
1412 return vectorizeTree(E);
1415 Type *ScalarTy = VL[0]->getType();
1416 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1417 ScalarTy = SI->getValueOperand()->getType();
1418 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1420 return Gather(VL, VecTy);
1423 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1424 IRBuilder<>::InsertPointGuard Guard(Builder);
1426 if (E->VectorizedValue) {
1427 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1428 return E->VectorizedValue;
1431 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1432 Type *ScalarTy = VL0->getType();
1433 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1434 ScalarTy = SI->getValueOperand()->getType();
1435 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1437 if (E->NeedToGather) {
1438 setInsertPointAfterBundle(E->Scalars);
1439 return Gather(E->Scalars, VecTy);
1442 unsigned Opcode = VL0->getOpcode();
1443 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1446 case Instruction::PHI: {
1447 PHINode *PH = dyn_cast<PHINode>(VL0);
1448 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1449 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1450 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1451 E->VectorizedValue = NewPhi;
1453 // PHINodes may have multiple entries from the same block. We want to
1454 // visit every block once.
1455 SmallSet<BasicBlock*, 4> VisitedBBs;
1457 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1459 BasicBlock *IBB = PH->getIncomingBlock(i);
1461 if (!VisitedBBs.insert(IBB)) {
1462 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1466 // Prepare the operand vector.
1467 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1468 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1469 getIncomingValueForBlock(IBB));
1471 Builder.SetInsertPoint(IBB->getTerminator());
1472 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1473 Value *Vec = vectorizeTree(Operands);
1474 NewPhi->addIncoming(Vec, IBB);
1477 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1478 "Invalid number of incoming values");
1482 case Instruction::ExtractElement: {
1483 if (CanReuseExtract(E->Scalars)) {
1484 Value *V = VL0->getOperand(0);
1485 E->VectorizedValue = V;
1488 return Gather(E->Scalars, VecTy);
1490 case Instruction::ZExt:
1491 case Instruction::SExt:
1492 case Instruction::FPToUI:
1493 case Instruction::FPToSI:
1494 case Instruction::FPExt:
1495 case Instruction::PtrToInt:
1496 case Instruction::IntToPtr:
1497 case Instruction::SIToFP:
1498 case Instruction::UIToFP:
1499 case Instruction::Trunc:
1500 case Instruction::FPTrunc:
1501 case Instruction::BitCast: {
1503 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1504 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1506 setInsertPointAfterBundle(E->Scalars);
1508 Value *InVec = vectorizeTree(INVL);
1510 if (Value *V = alreadyVectorized(E->Scalars))
1513 CastInst *CI = dyn_cast<CastInst>(VL0);
1514 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1515 E->VectorizedValue = V;
1518 case Instruction::FCmp:
1519 case Instruction::ICmp: {
1520 ValueList LHSV, RHSV;
1521 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1522 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1523 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1526 setInsertPointAfterBundle(E->Scalars);
1528 Value *L = vectorizeTree(LHSV);
1529 Value *R = vectorizeTree(RHSV);
1531 if (Value *V = alreadyVectorized(E->Scalars))
1534 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1536 if (Opcode == Instruction::FCmp)
1537 V = Builder.CreateFCmp(P0, L, R);
1539 V = Builder.CreateICmp(P0, L, R);
1541 E->VectorizedValue = V;
1544 case Instruction::Select: {
1545 ValueList TrueVec, FalseVec, CondVec;
1546 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1547 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1548 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1549 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1552 setInsertPointAfterBundle(E->Scalars);
1554 Value *Cond = vectorizeTree(CondVec);
1555 Value *True = vectorizeTree(TrueVec);
1556 Value *False = vectorizeTree(FalseVec);
1558 if (Value *V = alreadyVectorized(E->Scalars))
1561 Value *V = Builder.CreateSelect(Cond, True, False);
1562 E->VectorizedValue = V;
1565 case Instruction::Add:
1566 case Instruction::FAdd:
1567 case Instruction::Sub:
1568 case Instruction::FSub:
1569 case Instruction::Mul:
1570 case Instruction::FMul:
1571 case Instruction::UDiv:
1572 case Instruction::SDiv:
1573 case Instruction::FDiv:
1574 case Instruction::URem:
1575 case Instruction::SRem:
1576 case Instruction::FRem:
1577 case Instruction::Shl:
1578 case Instruction::LShr:
1579 case Instruction::AShr:
1580 case Instruction::And:
1581 case Instruction::Or:
1582 case Instruction::Xor: {
1583 ValueList LHSVL, RHSVL;
1584 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1585 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1587 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1588 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1589 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1592 setInsertPointAfterBundle(E->Scalars);
1594 Value *LHS = vectorizeTree(LHSVL);
1595 Value *RHS = vectorizeTree(RHSVL);
1597 if (LHS == RHS && isa<Instruction>(LHS)) {
1598 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1601 if (Value *V = alreadyVectorized(E->Scalars))
1604 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1605 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1606 E->VectorizedValue = V;
1608 if (Instruction *I = dyn_cast<Instruction>(V))
1609 return propagateMetadata(I, E->Scalars);
1613 case Instruction::Load: {
1614 // Loads are inserted at the head of the tree because we don't want to
1615 // sink them all the way down past store instructions.
1616 setInsertPointAfterBundle(E->Scalars);
1618 LoadInst *LI = cast<LoadInst>(VL0);
1619 unsigned AS = LI->getPointerAddressSpace();
1621 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1622 VecTy->getPointerTo(AS));
1623 unsigned Alignment = LI->getAlignment();
1624 LI = Builder.CreateLoad(VecPtr);
1626 Alignment = DL->getABITypeAlignment(LI->getPointerOperand()->getType());
1627 LI->setAlignment(Alignment);
1628 E->VectorizedValue = LI;
1629 return propagateMetadata(LI, E->Scalars);
1631 case Instruction::Store: {
1632 StoreInst *SI = cast<StoreInst>(VL0);
1633 unsigned Alignment = SI->getAlignment();
1634 unsigned AS = SI->getPointerAddressSpace();
1637 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1638 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1640 setInsertPointAfterBundle(E->Scalars);
1642 Value *VecValue = vectorizeTree(ValueOp);
1643 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1644 VecTy->getPointerTo(AS));
1645 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1647 Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType());
1648 S->setAlignment(Alignment);
1649 E->VectorizedValue = S;
1650 return propagateMetadata(S, E->Scalars);
1652 case Instruction::Call: {
1653 CallInst *CI = cast<CallInst>(VL0);
1654 setInsertPointAfterBundle(E->Scalars);
1655 std::vector<Value *> OpVecs;
1656 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1658 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1659 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1660 OpVL.push_back(CEI->getArgOperand(j));
1663 Value *OpVec = vectorizeTree(OpVL);
1664 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1665 OpVecs.push_back(OpVec);
1668 Module *M = F->getParent();
1669 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1670 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
1671 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
1672 Value *V = Builder.CreateCall(CF, OpVecs);
1673 E->VectorizedValue = V;
1677 llvm_unreachable("unknown inst");
1682 Value *BoUpSLP::vectorizeTree() {
1683 Builder.SetInsertPoint(F->getEntryBlock().begin());
1684 vectorizeTree(&VectorizableTree[0]);
1686 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1688 // Extract all of the elements with the external uses.
1689 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1691 Value *Scalar = it->Scalar;
1692 llvm::User *User = it->User;
1694 // Skip users that we already RAUW. This happens when one instruction
1695 // has multiple uses of the same value.
1696 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
1699 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1701 int Idx = ScalarToTreeEntry[Scalar];
1702 TreeEntry *E = &VectorizableTree[Idx];
1703 assert(!E->NeedToGather && "Extracting from a gather list");
1705 Value *Vec = E->VectorizedValue;
1706 assert(Vec && "Can't find vectorizable value");
1708 Value *Lane = Builder.getInt32(it->Lane);
1709 // Generate extracts for out-of-tree users.
1710 // Find the insertion point for the extractelement lane.
1711 if (isa<Instruction>(Vec)){
1712 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1713 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1714 if (PH->getIncomingValue(i) == Scalar) {
1715 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1716 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1717 CSEBlocks.insert(PH->getIncomingBlock(i));
1718 PH->setOperand(i, Ex);
1722 Builder.SetInsertPoint(cast<Instruction>(User));
1723 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1724 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1725 User->replaceUsesOfWith(Scalar, Ex);
1728 Builder.SetInsertPoint(F->getEntryBlock().begin());
1729 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1730 CSEBlocks.insert(&F->getEntryBlock());
1731 User->replaceUsesOfWith(Scalar, Ex);
1734 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1737 // For each vectorized value:
1738 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1739 TreeEntry *Entry = &VectorizableTree[EIdx];
1742 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1743 Value *Scalar = Entry->Scalars[Lane];
1745 // No need to handle users of gathered values.
1746 if (Entry->NeedToGather)
1749 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1751 Type *Ty = Scalar->getType();
1752 if (!Ty->isVoidTy()) {
1754 for (User *U : Scalar->users()) {
1755 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
1757 assert((ScalarToTreeEntry.count(U) ||
1758 // It is legal to replace users in the ignorelist by undef.
1759 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
1760 UserIgnoreList.end())) &&
1761 "Replacing out-of-tree value with undef");
1764 Value *Undef = UndefValue::get(Ty);
1765 Scalar->replaceAllUsesWith(Undef);
1767 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1768 cast<Instruction>(Scalar)->eraseFromParent();
1772 for (auto &BN : BlocksNumbers)
1775 Builder.ClearInsertionPoint();
1777 return VectorizableTree[0].VectorizedValue;
1780 void BoUpSLP::optimizeGatherSequence() {
1781 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1782 << " gather sequences instructions.\n");
1783 // LICM InsertElementInst sequences.
1784 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1785 e = GatherSeq.end(); it != e; ++it) {
1786 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1791 // Check if this block is inside a loop.
1792 Loop *L = LI->getLoopFor(Insert->getParent());
1796 // Check if it has a preheader.
1797 BasicBlock *PreHeader = L->getLoopPreheader();
1801 // If the vector or the element that we insert into it are
1802 // instructions that are defined in this basic block then we can't
1803 // hoist this instruction.
1804 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1805 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1806 if (CurrVec && L->contains(CurrVec))
1808 if (NewElem && L->contains(NewElem))
1811 // We can hoist this instruction. Move it to the pre-header.
1812 Insert->moveBefore(PreHeader->getTerminator());
1815 // Make a list of all reachable blocks in our CSE queue.
1816 SmallVector<const DomTreeNode *, 8> CSEWorkList;
1817 CSEWorkList.reserve(CSEBlocks.size());
1818 for (BasicBlock *BB : CSEBlocks)
1819 if (DomTreeNode *N = DT->getNode(BB)) {
1820 assert(DT->isReachableFromEntry(N));
1821 CSEWorkList.push_back(N);
1824 // Sort blocks by domination. This ensures we visit a block after all blocks
1825 // dominating it are visited.
1826 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1827 [this](const DomTreeNode *A, const DomTreeNode *B) {
1828 return DT->properlyDominates(A, B);
1831 // Perform O(N^2) search over the gather sequences and merge identical
1832 // instructions. TODO: We can further optimize this scan if we split the
1833 // instructions into different buckets based on the insert lane.
1834 SmallVector<Instruction *, 16> Visited;
1835 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
1836 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1837 "Worklist not sorted properly!");
1838 BasicBlock *BB = (*I)->getBlock();
1839 // For all instructions in blocks containing gather sequences:
1840 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1841 Instruction *In = it++;
1842 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1845 // Check if we can replace this instruction with any of the
1846 // visited instructions.
1847 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1850 if (In->isIdenticalTo(*v) &&
1851 DT->dominates((*v)->getParent(), In->getParent())) {
1852 In->replaceAllUsesWith(*v);
1853 In->eraseFromParent();
1859 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1860 Visited.push_back(In);
1868 /// The SLPVectorizer Pass.
1869 struct SLPVectorizer : public FunctionPass {
1870 typedef SmallVector<StoreInst *, 8> StoreList;
1871 typedef MapVector<Value *, StoreList> StoreListMap;
1873 /// Pass identification, replacement for typeid
1876 explicit SLPVectorizer() : FunctionPass(ID) {
1877 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1880 ScalarEvolution *SE;
1881 const DataLayout *DL;
1882 TargetTransformInfo *TTI;
1883 TargetLibraryInfo *TLI;
1888 bool runOnFunction(Function &F) override {
1889 if (skipOptnoneFunction(F))
1892 SE = &getAnalysis<ScalarEvolution>();
1893 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1894 DL = DLP ? &DLP->getDataLayout() : nullptr;
1895 TTI = &getAnalysis<TargetTransformInfo>();
1896 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1897 AA = &getAnalysis<AliasAnalysis>();
1898 LI = &getAnalysis<LoopInfo>();
1899 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1902 bool Changed = false;
1904 // If the target claims to have no vector registers don't attempt
1906 if (!TTI->getNumberOfRegisters(true))
1909 // Must have DataLayout. We can't require it because some tests run w/o
1914 // Don't vectorize when the attribute NoImplicitFloat is used.
1915 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1918 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1920 // Use the bottom up slp vectorizer to construct chains that start with
1921 // he store instructions.
1922 BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT);
1924 // Scan the blocks in the function in post order.
1925 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1926 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1927 BasicBlock *BB = *it;
1929 // Vectorize trees that end at stores.
1930 if (unsigned count = collectStores(BB, R)) {
1932 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1933 Changed |= vectorizeStoreChains(R);
1936 // Vectorize trees that end at reductions.
1937 Changed |= vectorizeChainsInBlock(BB, R);
1941 R.optimizeGatherSequence();
1942 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1943 DEBUG(verifyFunction(F));
1948 void getAnalysisUsage(AnalysisUsage &AU) const override {
1949 FunctionPass::getAnalysisUsage(AU);
1950 AU.addRequired<ScalarEvolution>();
1951 AU.addRequired<AliasAnalysis>();
1952 AU.addRequired<TargetTransformInfo>();
1953 AU.addRequired<LoopInfo>();
1954 AU.addRequired<DominatorTreeWrapperPass>();
1955 AU.addPreserved<LoopInfo>();
1956 AU.addPreserved<DominatorTreeWrapperPass>();
1957 AU.setPreservesCFG();
1962 /// \brief Collect memory references and sort them according to their base
1963 /// object. We sort the stores to their base objects to reduce the cost of the
1964 /// quadratic search on the stores. TODO: We can further reduce this cost
1965 /// if we flush the chain creation every time we run into a memory barrier.
1966 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1968 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1969 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1971 /// \brief Try to vectorize a list of operands.
1972 /// \@param BuildVector A list of users to ignore for the purpose of
1973 /// scheduling and that don't need extracting.
1974 /// \returns true if a value was vectorized.
1975 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
1976 ArrayRef<Value *> BuildVector = None);
1978 /// \brief Try to vectorize a chain that may start at the operands of \V;
1979 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1981 /// \brief Vectorize the stores that were collected in StoreRefs.
1982 bool vectorizeStoreChains(BoUpSLP &R);
1984 /// \brief Scan the basic block and look for patterns that are likely to start
1985 /// a vectorization chain.
1986 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1988 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1991 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1994 StoreListMap StoreRefs;
1997 /// \brief Check that the Values in the slice in VL array are still existent in
1998 /// the WeakVH array.
1999 /// Vectorization of part of the VL array may cause later values in the VL array
2000 /// to become invalid. We track when this has happened in the WeakVH array.
2001 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
2002 SmallVectorImpl<WeakVH> &VH,
2003 unsigned SliceBegin,
2004 unsigned SliceSize) {
2005 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
2012 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
2013 int CostThreshold, BoUpSLP &R) {
2014 unsigned ChainLen = Chain.size();
2015 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2017 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2018 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2019 unsigned VF = MinVecRegSize / Sz;
2021 if (!isPowerOf2_32(Sz) || VF < 2)
2024 // Keep track of values that were deleted by vectorizing in the loop below.
2025 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2027 bool Changed = false;
2028 // Look for profitable vectorizable trees at all offsets, starting at zero.
2029 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2033 // Check that a previous iteration of this loop did not delete the Value.
2034 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2037 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2039 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2041 R.buildTree(Operands);
2043 int Cost = R.getTreeCost();
2045 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2046 if (Cost < CostThreshold) {
2047 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2050 // Move to the next bundle.
2059 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2060 int costThreshold, BoUpSLP &R) {
2061 SetVector<Value *> Heads, Tails;
2062 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2064 // We may run into multiple chains that merge into a single chain. We mark the
2065 // stores that we vectorized so that we don't visit the same store twice.
2066 BoUpSLP::ValueSet VectorizedStores;
2067 bool Changed = false;
2069 // Do a quadratic search on all of the given stores and find
2070 // all of the pairs of stores that follow each other.
2071 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2072 for (unsigned j = 0; j < e; ++j) {
2076 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2077 Tails.insert(Stores[j]);
2078 Heads.insert(Stores[i]);
2079 ConsecutiveChain[Stores[i]] = Stores[j];
2084 // For stores that start but don't end a link in the chain:
2085 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2087 if (Tails.count(*it))
2090 // We found a store instr that starts a chain. Now follow the chain and try
2092 BoUpSLP::ValueList Operands;
2094 // Collect the chain into a list.
2095 while (Tails.count(I) || Heads.count(I)) {
2096 if (VectorizedStores.count(I))
2098 Operands.push_back(I);
2099 // Move to the next value in the chain.
2100 I = ConsecutiveChain[I];
2103 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2105 // Mark the vectorized stores so that we don't vectorize them again.
2107 VectorizedStores.insert(Operands.begin(), Operands.end());
2108 Changed |= Vectorized;
2115 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2118 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2119 StoreInst *SI = dyn_cast<StoreInst>(it);
2123 // Don't touch volatile stores.
2124 if (!SI->isSimple())
2127 // Check that the pointer points to scalars.
2128 Type *Ty = SI->getValueOperand()->getType();
2129 if (Ty->isAggregateType() || Ty->isVectorTy())
2132 // Find the base pointer.
2133 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2135 // Save the store locations.
2136 StoreRefs[Ptr].push_back(SI);
2142 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2145 Value *VL[] = { A, B };
2146 return tryToVectorizeList(VL, R);
2149 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2150 ArrayRef<Value *> BuildVector) {
2154 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2156 // Check that all of the parts are scalar instructions of the same type.
2157 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2161 unsigned Opcode0 = I0->getOpcode();
2163 Type *Ty0 = I0->getType();
2164 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2165 unsigned VF = MinVecRegSize / Sz;
2167 for (int i = 0, e = VL.size(); i < e; ++i) {
2168 Type *Ty = VL[i]->getType();
2169 if (Ty->isAggregateType() || Ty->isVectorTy())
2171 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2172 if (!Inst || Inst->getOpcode() != Opcode0)
2176 bool Changed = false;
2178 // Keep track of values that were deleted by vectorizing in the loop below.
2179 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2181 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2182 unsigned OpsWidth = 0;
2189 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2192 // Check that a previous iteration of this loop did not delete the Value.
2193 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2196 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2198 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2200 ArrayRef<Value *> BuildVectorSlice;
2201 if (!BuildVector.empty())
2202 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
2204 R.buildTree(Ops, BuildVectorSlice);
2205 int Cost = R.getTreeCost();
2207 if (Cost < -SLPCostThreshold) {
2208 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
2209 Value *VectorizedRoot = R.vectorizeTree();
2211 // Reconstruct the build vector by extracting the vectorized root. This
2212 // way we handle the case where some elements of the vector are undefined.
2213 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
2214 if (!BuildVectorSlice.empty()) {
2215 // The insert point is the last build vector instruction. The vectorized
2216 // root will precede it. This guarantees that we get an instruction. The
2217 // vectorized tree could have been constant folded.
2218 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
2219 unsigned VecIdx = 0;
2220 for (auto &V : BuildVectorSlice) {
2221 IRBuilder<true, NoFolder> Builder(
2222 ++BasicBlock::iterator(InsertAfter));
2223 InsertElementInst *IE = cast<InsertElementInst>(V);
2224 Instruction *Extract = cast<Instruction>(Builder.CreateExtractElement(
2225 VectorizedRoot, Builder.getInt32(VecIdx++)));
2226 IE->setOperand(1, Extract);
2227 IE->removeFromParent();
2228 IE->insertAfter(Extract);
2232 // Move to the next bundle.
2241 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2245 // Try to vectorize V.
2246 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2249 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2250 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2252 if (B && B->hasOneUse()) {
2253 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2254 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2255 if (tryToVectorizePair(A, B0, R)) {
2259 if (tryToVectorizePair(A, B1, R)) {
2266 if (A && A->hasOneUse()) {
2267 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2268 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2269 if (tryToVectorizePair(A0, B, R)) {
2273 if (tryToVectorizePair(A1, B, R)) {
2281 /// \brief Generate a shuffle mask to be used in a reduction tree.
2283 /// \param VecLen The length of the vector to be reduced.
2284 /// \param NumEltsToRdx The number of elements that should be reduced in the
2286 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2287 /// reduction. A pairwise reduction will generate a mask of
2288 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2289 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2290 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2291 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2292 bool IsPairwise, bool IsLeft,
2293 IRBuilder<> &Builder) {
2294 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2296 SmallVector<Constant *, 32> ShuffleMask(
2297 VecLen, UndefValue::get(Builder.getInt32Ty()));
2300 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2301 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2302 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2304 // Move the upper half of the vector to the lower half.
2305 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2306 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2308 return ConstantVector::get(ShuffleMask);
2312 /// Model horizontal reductions.
2314 /// A horizontal reduction is a tree of reduction operations (currently add and
2315 /// fadd) that has operations that can be put into a vector as its leaf.
2316 /// For example, this tree:
2323 /// This tree has "mul" as its reduced values and "+" as its reduction
2324 /// operations. A reduction might be feeding into a store or a binary operation
2339 class HorizontalReduction {
2340 SmallVector<Value *, 16> ReductionOps;
2341 SmallVector<Value *, 32> ReducedVals;
2343 BinaryOperator *ReductionRoot;
2344 PHINode *ReductionPHI;
2346 /// The opcode of the reduction.
2347 unsigned ReductionOpcode;
2348 /// The opcode of the values we perform a reduction on.
2349 unsigned ReducedValueOpcode;
2350 /// The width of one full horizontal reduction operation.
2351 unsigned ReduxWidth;
2352 /// Should we model this reduction as a pairwise reduction tree or a tree that
2353 /// splits the vector in halves and adds those halves.
2354 bool IsPairwiseReduction;
2357 HorizontalReduction()
2358 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
2359 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2361 /// \brief Try to find a reduction tree.
2362 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2363 const DataLayout *DL) {
2365 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2366 "Thi phi needs to use the binary operator");
2368 // We could have a initial reductions that is not an add.
2369 // r *= v1 + v2 + v3 + v4
2370 // In such a case start looking for a tree rooted in the first '+'.
2372 if (B->getOperand(0) == Phi) {
2374 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2375 } else if (B->getOperand(1) == Phi) {
2377 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2384 Type *Ty = B->getType();
2385 if (Ty->isVectorTy())
2388 ReductionOpcode = B->getOpcode();
2389 ReducedValueOpcode = 0;
2390 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2397 // We currently only support adds.
2398 if (ReductionOpcode != Instruction::Add &&
2399 ReductionOpcode != Instruction::FAdd)
2402 // Post order traverse the reduction tree starting at B. We only handle true
2403 // trees containing only binary operators.
2404 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2405 Stack.push_back(std::make_pair(B, 0));
2406 while (!Stack.empty()) {
2407 BinaryOperator *TreeN = Stack.back().first;
2408 unsigned EdgeToVist = Stack.back().second++;
2409 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2411 // Only handle trees in the current basic block.
2412 if (TreeN->getParent() != B->getParent())
2415 // Each tree node needs to have one user except for the ultimate
2417 if (!TreeN->hasOneUse() && TreeN != B)
2421 if (EdgeToVist == 2 || IsReducedValue) {
2422 if (IsReducedValue) {
2423 // Make sure that the opcodes of the operations that we are going to
2425 if (!ReducedValueOpcode)
2426 ReducedValueOpcode = TreeN->getOpcode();
2427 else if (ReducedValueOpcode != TreeN->getOpcode())
2429 ReducedVals.push_back(TreeN);
2431 // We need to be able to reassociate the adds.
2432 if (!TreeN->isAssociative())
2434 ReductionOps.push_back(TreeN);
2441 // Visit left or right.
2442 Value *NextV = TreeN->getOperand(EdgeToVist);
2443 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2445 Stack.push_back(std::make_pair(Next, 0));
2446 else if (NextV != Phi)
2452 /// \brief Attempt to vectorize the tree found by
2453 /// matchAssociativeReduction.
2454 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2455 if (ReducedVals.empty())
2458 unsigned NumReducedVals = ReducedVals.size();
2459 if (NumReducedVals < ReduxWidth)
2462 Value *VectorizedTree = nullptr;
2463 IRBuilder<> Builder(ReductionRoot);
2464 FastMathFlags Unsafe;
2465 Unsafe.setUnsafeAlgebra();
2466 Builder.SetFastMathFlags(Unsafe);
2469 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2470 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2471 V.buildTree(ValsToReduce, ReductionOps);
2474 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2475 if (Cost >= -SLPCostThreshold)
2478 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2481 // Vectorize a tree.
2482 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2483 Value *VectorizedRoot = V.vectorizeTree();
2485 // Emit a reduction.
2486 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2487 if (VectorizedTree) {
2488 Builder.SetCurrentDebugLocation(Loc);
2489 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2490 ReducedSubTree, "bin.rdx");
2492 VectorizedTree = ReducedSubTree;
2495 if (VectorizedTree) {
2496 // Finish the reduction.
2497 for (; i < NumReducedVals; ++i) {
2498 Builder.SetCurrentDebugLocation(
2499 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2500 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2505 assert(ReductionRoot && "Need a reduction operation");
2506 ReductionRoot->setOperand(0, VectorizedTree);
2507 ReductionRoot->setOperand(1, ReductionPHI);
2509 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2511 return VectorizedTree != nullptr;
2516 /// \brief Calcuate the cost of a reduction.
2517 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2518 Type *ScalarTy = FirstReducedVal->getType();
2519 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2521 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2522 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2524 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2525 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2527 int ScalarReduxCost =
2528 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2530 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2531 << " for reduction that starts with " << *FirstReducedVal
2533 << (IsPairwiseReduction ? "pairwise" : "splitting")
2534 << " reduction)\n");
2536 return VecReduxCost - ScalarReduxCost;
2539 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2540 Value *R, const Twine &Name = "") {
2541 if (Opcode == Instruction::FAdd)
2542 return Builder.CreateFAdd(L, R, Name);
2543 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2546 /// \brief Emit a horizontal reduction of the vectorized value.
2547 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2548 assert(VectorizedValue && "Need to have a vectorized tree node");
2549 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2550 assert(isPowerOf2_32(ReduxWidth) &&
2551 "We only handle power-of-two reductions for now");
2553 Value *TmpVec = ValToReduce;
2554 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2555 if (IsPairwiseReduction) {
2557 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2559 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2561 Value *LeftShuf = Builder.CreateShuffleVector(
2562 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2563 Value *RightShuf = Builder.CreateShuffleVector(
2564 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2566 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2570 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2571 Value *Shuf = Builder.CreateShuffleVector(
2572 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2573 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2577 // The result is in the first element of the vector.
2578 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2582 /// \brief Recognize construction of vectors like
2583 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2584 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2585 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2586 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2588 /// Returns true if it matches
2590 static bool findBuildVector(InsertElementInst *FirstInsertElem,
2591 SmallVectorImpl<Value *> &BuildVector,
2592 SmallVectorImpl<Value *> &BuildVectorOpds) {
2593 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
2596 InsertElementInst *IE = FirstInsertElem;
2598 BuildVector.push_back(IE);
2599 BuildVectorOpds.push_back(IE->getOperand(1));
2601 if (IE->use_empty())
2604 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2608 // If this isn't the final use, make sure the next insertelement is the only
2609 // use. It's OK if the final constructed vector is used multiple times
2610 if (!IE->hasOneUse())
2619 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2620 return V->getType() < V2->getType();
2623 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2624 bool Changed = false;
2625 SmallVector<Value *, 4> Incoming;
2626 SmallSet<Value *, 16> VisitedInstrs;
2628 bool HaveVectorizedPhiNodes = true;
2629 while (HaveVectorizedPhiNodes) {
2630 HaveVectorizedPhiNodes = false;
2632 // Collect the incoming values from the PHIs.
2634 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2636 PHINode *P = dyn_cast<PHINode>(instr);
2640 if (!VisitedInstrs.count(P))
2641 Incoming.push_back(P);
2645 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2647 // Try to vectorize elements base on their type.
2648 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2652 // Look for the next elements with the same type.
2653 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2654 while (SameTypeIt != E &&
2655 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2656 VisitedInstrs.insert(*SameTypeIt);
2660 // Try to vectorize them.
2661 unsigned NumElts = (SameTypeIt - IncIt);
2662 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2664 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2665 // Success start over because instructions might have been changed.
2666 HaveVectorizedPhiNodes = true;
2671 // Start over at the next instruction of a different type (or the end).
2676 VisitedInstrs.clear();
2678 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2679 // We may go through BB multiple times so skip the one we have checked.
2680 if (!VisitedInstrs.insert(it))
2683 if (isa<DbgInfoIntrinsic>(it))
2686 // Try to vectorize reductions that use PHINodes.
2687 if (PHINode *P = dyn_cast<PHINode>(it)) {
2688 // Check that the PHI is a reduction PHI.
2689 if (P->getNumIncomingValues() != 2)
2692 (P->getIncomingBlock(0) == BB
2693 ? (P->getIncomingValue(0))
2694 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
2696 // Check if this is a Binary Operator.
2697 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2701 // Try to match and vectorize a horizontal reduction.
2702 HorizontalReduction HorRdx;
2703 if (ShouldVectorizeHor &&
2704 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2705 HorRdx.tryToReduce(R, TTI)) {
2712 Value *Inst = BI->getOperand(0);
2714 Inst = BI->getOperand(1);
2716 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2717 // We would like to start over since some instructions are deleted
2718 // and the iterator may become invalid value.
2728 // Try to vectorize horizontal reductions feeding into a store.
2729 if (ShouldStartVectorizeHorAtStore)
2730 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2731 if (BinaryOperator *BinOp =
2732 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2733 HorizontalReduction HorRdx;
2734 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
2735 HorRdx.tryToReduce(R, TTI)) ||
2736 tryToVectorize(BinOp, R))) {
2744 // Try to vectorize trees that start at compare instructions.
2745 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2746 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2748 // We would like to start over since some instructions are deleted
2749 // and the iterator may become invalid value.
2755 for (int i = 0; i < 2; ++i) {
2756 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2757 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2759 // We would like to start over since some instructions are deleted
2760 // and the iterator may become invalid value.
2769 // Try to vectorize trees that start at insertelement instructions.
2770 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
2771 SmallVector<Value *, 16> BuildVector;
2772 SmallVector<Value *, 16> BuildVectorOpds;
2773 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
2776 // Vectorize starting with the build vector operands ignoring the
2777 // BuildVector instructions for the purpose of scheduling and user
2779 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
2792 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2793 bool Changed = false;
2794 // Attempt to sort and vectorize each of the store-groups.
2795 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2797 if (it->second.size() < 2)
2800 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2801 << it->second.size() << ".\n");
2803 // Process the stores in chunks of 16.
2804 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2805 unsigned Len = std::min<unsigned>(CE - CI, 16);
2806 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2807 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2813 } // end anonymous namespace
2815 char SLPVectorizer::ID = 0;
2816 static const char lv_name[] = "SLP Vectorizer";
2817 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2818 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2819 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2820 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2821 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2822 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2825 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }