1 //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // This pass implements the Bottom Up SLP vectorizer. It detects consecutive
10 // stores that can be put together into vector-stores. Next, it attempts to
11 // construct vectorizable tree using the use-def chains. If a profitable tree
12 // was found, the SLP vectorizer performs vectorization on the tree.
14 // The pass is inspired by the work described in the paper:
15 // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
17 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/MapVector.h"
20 #include "llvm/ADT/PostOrderIterator.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Type.h"
35 #include "llvm/IR/Value.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Utils/VectorUtils.h"
47 #define SV_NAME "slp-vectorizer"
48 #define DEBUG_TYPE "SLP"
51 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
52 cl::desc("Only vectorize if you gain more than this "
56 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
57 cl::desc("Attempt to vectorize horizontal reductions"));
59 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
60 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
62 "Attempt to vectorize horizontal reductions feeding into a store"));
66 static const unsigned MinVecRegSize = 128;
68 static const unsigned RecursionMaxDepth = 12;
70 /// A helper class for numbering instructions in multiple blocks.
71 /// Numbers start at zero for each basic block.
72 struct BlockNumbering {
74 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
76 BlockNumbering() : BB(nullptr), Valid(false) {}
78 void numberInstructions() {
82 // Number the instructions in the block.
83 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
85 InstrVec.push_back(it);
86 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
91 int getIndex(Instruction *I) {
92 assert(I->getParent() == BB && "Invalid instruction");
95 assert(InstrIdx.count(I) && "Unknown instruction");
99 Instruction *getInstruction(unsigned loc) {
101 numberInstructions();
102 assert(InstrVec.size() > loc && "Invalid Index");
103 return InstrVec[loc];
106 void forget() { Valid = false; }
109 /// The block we are numbering.
111 /// Is the block numbered.
113 /// Maps instructions to numbers and back.
114 SmallDenseMap<Instruction *, int> InstrIdx;
115 /// Maps integers to Instructions.
116 SmallVector<Instruction *, 32> InstrVec;
119 /// \returns the parent basic block if all of the instructions in \p VL
120 /// are in the same block or null otherwise.
121 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
122 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
125 BasicBlock *BB = I0->getParent();
126 for (int i = 1, e = VL.size(); i < e; i++) {
127 Instruction *I = dyn_cast<Instruction>(VL[i]);
131 if (BB != I->getParent())
137 /// \returns True if all of the values in \p VL are constants.
138 static bool allConstant(ArrayRef<Value *> VL) {
139 for (unsigned i = 0, e = VL.size(); i < e; ++i)
140 if (!isa<Constant>(VL[i]))
145 /// \returns True if all of the values in \p VL are identical.
146 static bool isSplat(ArrayRef<Value *> VL) {
147 for (unsigned i = 1, e = VL.size(); i < e; ++i)
153 /// \returns The opcode if all of the Instructions in \p VL have the same
155 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
156 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
159 unsigned Opcode = I0->getOpcode();
160 for (int i = 1, e = VL.size(); i < e; i++) {
161 Instruction *I = dyn_cast<Instruction>(VL[i]);
162 if (!I || Opcode != I->getOpcode())
168 /// \returns \p I after propagating metadata from \p VL.
169 static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) {
170 Instruction *I0 = cast<Instruction>(VL[0]);
171 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
172 I0->getAllMetadataOtherThanDebugLoc(Metadata);
174 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) {
175 unsigned Kind = Metadata[i].first;
176 MDNode *MD = Metadata[i].second;
178 for (int i = 1, e = VL.size(); MD && i != e; i++) {
179 Instruction *I = cast<Instruction>(VL[i]);
180 MDNode *IMD = I->getMetadata(Kind);
184 MD = nullptr; // Remove unknown metadata
186 case LLVMContext::MD_tbaa:
187 MD = MDNode::getMostGenericTBAA(MD, IMD);
189 case LLVMContext::MD_fpmath:
190 MD = MDNode::getMostGenericFPMath(MD, IMD);
194 I->setMetadata(Kind, MD);
199 /// \returns The type that all of the values in \p VL have or null if there
200 /// are different types.
201 static Type* getSameType(ArrayRef<Value *> VL) {
202 Type *Ty = VL[0]->getType();
203 for (int i = 1, e = VL.size(); i < e; i++)
204 if (VL[i]->getType() != Ty)
210 /// \returns True if the ExtractElement instructions in VL can be vectorized
211 /// to use the original vector.
212 static bool CanReuseExtract(ArrayRef<Value *> VL) {
213 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
214 // Check if all of the extracts come from the same vector and from the
217 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
218 Value *Vec = E0->getOperand(0);
220 // We have to extract from the same vector type.
221 unsigned NElts = Vec->getType()->getVectorNumElements();
223 if (NElts != VL.size())
226 // Check that all of the indices extract from the correct offset.
227 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
228 if (!CI || CI->getZExtValue())
231 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
232 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
233 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
235 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
242 static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
243 SmallVectorImpl<Value *> &Left,
244 SmallVectorImpl<Value *> &Right) {
246 SmallVector<Value *, 16> OrigLeft, OrigRight;
248 bool AllSameOpcodeLeft = true;
249 bool AllSameOpcodeRight = true;
250 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
251 Instruction *I = cast<Instruction>(VL[i]);
252 Value *V0 = I->getOperand(0);
253 Value *V1 = I->getOperand(1);
255 OrigLeft.push_back(V0);
256 OrigRight.push_back(V1);
258 Instruction *I0 = dyn_cast<Instruction>(V0);
259 Instruction *I1 = dyn_cast<Instruction>(V1);
261 // Check whether all operands on one side have the same opcode. In this case
262 // we want to preserve the original order and not make things worse by
264 AllSameOpcodeLeft = I0;
265 AllSameOpcodeRight = I1;
267 if (i && AllSameOpcodeLeft) {
268 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) {
269 if(P0->getOpcode() != I0->getOpcode())
270 AllSameOpcodeLeft = false;
272 AllSameOpcodeLeft = false;
274 if (i && AllSameOpcodeRight) {
275 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) {
276 if(P1->getOpcode() != I1->getOpcode())
277 AllSameOpcodeRight = false;
279 AllSameOpcodeRight = false;
282 // Sort two opcodes. In the code below we try to preserve the ability to use
283 // broadcast of values instead of individual inserts.
290 // If we just sorted according to opcode we would leave the first line in
291 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load).
294 // Because vr2 and vr1 are from the same load we loose the opportunity of a
295 // broadcast for the packed right side in the backend: we have [vr1, vl2]
296 // instead of [vr1, vr2=vr1].
298 if(!i && I0->getOpcode() > I1->getOpcode()) {
301 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) {
302 // Try not to destroy a broad cast for no apparent benefit.
305 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) {
306 // Try preserve broadcasts.
309 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) {
310 // Try preserve broadcasts.
319 // One opcode, put the instruction on the right.
329 bool LeftBroadcast = isSplat(Left);
330 bool RightBroadcast = isSplat(Right);
332 // Don't reorder if the operands where good to begin with.
333 if (!(LeftBroadcast || RightBroadcast) &&
334 (AllSameOpcodeRight || AllSameOpcodeLeft)) {
340 /// Bottom Up SLP Vectorizer.
343 typedef SmallVector<Value *, 8> ValueList;
344 typedef SmallVector<Instruction *, 16> InstrList;
345 typedef SmallPtrSet<Value *, 16> ValueSet;
346 typedef SmallVector<StoreInst *, 8> StoreList;
348 BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl,
349 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li,
351 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt),
352 Builder(Se->getContext()) {
353 // Setup the block numbering utility for all of the blocks in the
355 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
357 BlocksNumbers[BB] = BlockNumbering(BB);
361 /// \brief Vectorize the tree that starts with the elements in \p VL.
362 /// Returns the vectorized root.
363 Value *vectorizeTree();
365 /// \returns the vectorization cost of the subtree that starts at \p VL.
366 /// A negative number means that this is profitable.
369 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
370 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
371 void buildTree(ArrayRef<Value *> Roots,
372 ArrayRef<Value *> UserIgnoreLst = None);
374 /// Clear the internal data structures that are created by 'buildTree'.
376 VectorizableTree.clear();
377 ScalarToTreeEntry.clear();
379 ExternalUses.clear();
380 MemBarrierIgnoreList.clear();
383 /// \returns true if the memory operations A and B are consecutive.
384 bool isConsecutiveAccess(Value *A, Value *B);
386 /// \brief Perform LICM and CSE on the newly generated gather sequences.
387 void optimizeGatherSequence();
391 /// \returns the cost of the vectorizable entry.
392 int getEntryCost(TreeEntry *E);
394 /// This is the recursive part of buildTree.
395 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
397 /// Vectorize a single entry in the tree.
398 Value *vectorizeTree(TreeEntry *E);
400 /// Vectorize a single entry in the tree, starting in \p VL.
401 Value *vectorizeTree(ArrayRef<Value *> VL);
403 /// \returns the pointer to the vectorized value if \p VL is already
404 /// vectorized, or NULL. They may happen in cycles.
405 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
407 /// \brief Take the pointer operand from the Load/Store instruction.
408 /// \returns NULL if this is not a valid Load/Store instruction.
409 static Value *getPointerOperand(Value *I);
411 /// \brief Take the address space operand from the Load/Store instruction.
412 /// \returns -1 if this is not a valid Load/Store instruction.
413 static unsigned getAddressSpaceOperand(Value *I);
415 /// \returns the scalarization cost for this type. Scalarization in this
416 /// context means the creation of vectors from a group of scalars.
417 int getGatherCost(Type *Ty);
419 /// \returns the scalarization cost for this list of values. Assuming that
420 /// this subtree gets vectorized, we may need to extract the values from the
421 /// roots. This method calculates the cost of extracting the values.
422 int getGatherCost(ArrayRef<Value *> VL);
424 /// \returns the AA location that is being access by the instruction.
425 AliasAnalysis::Location getLocation(Instruction *I);
427 /// \brief Checks if it is possible to sink an instruction from
428 /// \p Src to \p Dst.
429 /// \returns the pointer to the barrier instruction if we can't sink.
430 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
432 /// \returns the index of the last instruction in the BB from \p VL.
433 int getLastIndex(ArrayRef<Value *> VL);
435 /// \returns the Instruction in the bundle \p VL.
436 Instruction *getLastInstruction(ArrayRef<Value *> VL);
438 /// \brief Set the Builder insert point to one after the last instruction in
440 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
442 /// \returns a vector from a collection of scalars in \p VL.
443 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
445 /// \returns whether the VectorizableTree is fully vectoriable and will
446 /// be beneficial even the tree height is tiny.
447 bool isFullyVectorizableTinyTree();
450 TreeEntry() : Scalars(), VectorizedValue(nullptr), LastScalarIndex(0),
453 /// \returns true if the scalars in VL are equal to this entry.
454 bool isSame(ArrayRef<Value *> VL) const {
455 assert(VL.size() == Scalars.size() && "Invalid size");
456 return std::equal(VL.begin(), VL.end(), Scalars.begin());
459 /// A vector of scalars.
462 /// The Scalars are vectorized into this value. It is initialized to Null.
463 Value *VectorizedValue;
465 /// The index in the basic block of the last scalar.
468 /// Do we need to gather this sequence ?
472 /// Create a new VectorizableTree entry.
473 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
474 VectorizableTree.push_back(TreeEntry());
475 int idx = VectorizableTree.size() - 1;
476 TreeEntry *Last = &VectorizableTree[idx];
477 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
478 Last->NeedToGather = !Vectorized;
480 Last->LastScalarIndex = getLastIndex(VL);
481 for (int i = 0, e = VL.size(); i != e; ++i) {
482 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
483 ScalarToTreeEntry[VL[i]] = idx;
486 Last->LastScalarIndex = 0;
487 MustGather.insert(VL.begin(), VL.end());
492 /// -- Vectorization State --
493 /// Holds all of the tree entries.
494 std::vector<TreeEntry> VectorizableTree;
496 /// Maps a specific scalar to its tree entry.
497 SmallDenseMap<Value*, int> ScalarToTreeEntry;
499 /// A list of scalars that we found that we need to keep as scalars.
502 /// This POD struct describes one external user in the vectorized tree.
503 struct ExternalUser {
504 ExternalUser (Value *S, llvm::User *U, int L) :
505 Scalar(S), User(U), Lane(L){};
506 // Which scalar in our function.
508 // Which user that uses the scalar.
510 // Which lane does the scalar belong to.
513 typedef SmallVector<ExternalUser, 16> UserList;
515 /// A list of values that need to extracted out of the tree.
516 /// This list holds pairs of (Internal Scalar : External User).
517 UserList ExternalUses;
519 /// A list of instructions to ignore while sinking
520 /// memory instructions. This map must be reset between runs of getCost.
521 ValueSet MemBarrierIgnoreList;
523 /// Holds all of the instructions that we gathered.
524 SetVector<Instruction *> GatherSeq;
525 /// A list of blocks that we are going to CSE.
526 SetVector<BasicBlock *> CSEBlocks;
528 /// Numbers instructions in different blocks.
529 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
531 /// List of users to ignore during scheduling and that don't need extracting.
532 ArrayRef<Value *> UserIgnoreList;
534 // Analysis and block reference.
537 const DataLayout *DL;
538 TargetTransformInfo *TTI;
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 = BlocksNumbers[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 IntrinsicInst *II = dyn_cast<IntrinsicInst>(VL[0]);
956 Intrinsic::ID ID = II ? II->getIntrinsicID() : Intrinsic::not_intrinsic;
958 if (!isTriviallyVectorizable(ID)) {
959 newTreeEntry(VL, false);
960 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
964 Function *Int = II->getCalledFunction();
966 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
967 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[i]);
968 if (!II2 || II2->getCalledFunction() != Int) {
969 newTreeEntry(VL, false);
970 DEBUG(dbgs() << "SLP: mismatched calls:" << *II << "!=" << *VL[i]
976 newTreeEntry(VL, true);
977 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i) {
979 // Prepare the operand vector.
980 for (unsigned j = 0; j < VL.size(); ++j) {
981 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[j]);
982 Operands.push_back(II2->getArgOperand(i));
984 buildTree_rec(Operands, Depth + 1);
989 newTreeEntry(VL, false);
990 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
995 int BoUpSLP::getEntryCost(TreeEntry *E) {
996 ArrayRef<Value*> VL = E->Scalars;
998 Type *ScalarTy = VL[0]->getType();
999 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1000 ScalarTy = SI->getValueOperand()->getType();
1001 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1003 if (E->NeedToGather) {
1004 if (allConstant(VL))
1007 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1009 return getGatherCost(E->Scalars);
1012 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1014 Instruction *VL0 = cast<Instruction>(VL[0]);
1015 unsigned Opcode = VL0->getOpcode();
1017 case Instruction::PHI: {
1020 case Instruction::ExtractElement: {
1021 if (CanReuseExtract(VL)) {
1023 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1024 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1026 // Take credit for instruction that will become dead.
1028 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1032 return getGatherCost(VecTy);
1034 case Instruction::ZExt:
1035 case Instruction::SExt:
1036 case Instruction::FPToUI:
1037 case Instruction::FPToSI:
1038 case Instruction::FPExt:
1039 case Instruction::PtrToInt:
1040 case Instruction::IntToPtr:
1041 case Instruction::SIToFP:
1042 case Instruction::UIToFP:
1043 case Instruction::Trunc:
1044 case Instruction::FPTrunc:
1045 case Instruction::BitCast: {
1046 Type *SrcTy = VL0->getOperand(0)->getType();
1048 // Calculate the cost of this instruction.
1049 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1050 VL0->getType(), SrcTy);
1052 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1053 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1054 return VecCost - ScalarCost;
1056 case Instruction::FCmp:
1057 case Instruction::ICmp:
1058 case Instruction::Select:
1059 case Instruction::Add:
1060 case Instruction::FAdd:
1061 case Instruction::Sub:
1062 case Instruction::FSub:
1063 case Instruction::Mul:
1064 case Instruction::FMul:
1065 case Instruction::UDiv:
1066 case Instruction::SDiv:
1067 case Instruction::FDiv:
1068 case Instruction::URem:
1069 case Instruction::SRem:
1070 case Instruction::FRem:
1071 case Instruction::Shl:
1072 case Instruction::LShr:
1073 case Instruction::AShr:
1074 case Instruction::And:
1075 case Instruction::Or:
1076 case Instruction::Xor: {
1077 // Calculate the cost of this instruction.
1080 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1081 Opcode == Instruction::Select) {
1082 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1083 ScalarCost = VecTy->getNumElements() *
1084 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1085 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1087 // Certain instructions can be cheaper to vectorize if they have a
1088 // constant second vector operand.
1089 TargetTransformInfo::OperandValueKind Op1VK =
1090 TargetTransformInfo::OK_AnyValue;
1091 TargetTransformInfo::OperandValueKind Op2VK =
1092 TargetTransformInfo::OK_UniformConstantValue;
1094 // If all operands are exactly the same ConstantInt then set the
1095 // operand kind to OK_UniformConstantValue.
1096 // If instead not all operands are constants, then set the operand kind
1097 // to OK_AnyValue. If all operands are constants but not the same,
1098 // then set the operand kind to OK_NonUniformConstantValue.
1099 ConstantInt *CInt = nullptr;
1100 for (unsigned i = 0; i < VL.size(); ++i) {
1101 const Instruction *I = cast<Instruction>(VL[i]);
1102 if (!isa<ConstantInt>(I->getOperand(1))) {
1103 Op2VK = TargetTransformInfo::OK_AnyValue;
1107 CInt = cast<ConstantInt>(I->getOperand(1));
1110 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1111 CInt != cast<ConstantInt>(I->getOperand(1)))
1112 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1116 VecTy->getNumElements() *
1117 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1118 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1120 return VecCost - ScalarCost;
1122 case Instruction::Load: {
1123 // Cost of wide load - cost of scalar loads.
1124 int ScalarLdCost = VecTy->getNumElements() *
1125 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1126 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1127 return VecLdCost - ScalarLdCost;
1129 case Instruction::Store: {
1130 // We know that we can merge the stores. Calculate the cost.
1131 int ScalarStCost = VecTy->getNumElements() *
1132 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1133 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1134 return VecStCost - ScalarStCost;
1136 case Instruction::Call: {
1137 CallInst *CI = cast<CallInst>(VL0);
1138 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1139 Intrinsic::ID ID = II->getIntrinsicID();
1141 // Calculate the cost of the scalar and vector calls.
1142 SmallVector<Type*, 4> ScalarTys, VecTys;
1143 for (unsigned op = 0, opc = II->getNumArgOperands(); op!= opc; ++op) {
1144 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1145 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1146 VecTy->getNumElements()));
1149 int ScalarCallCost = VecTy->getNumElements() *
1150 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1152 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1154 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1155 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1156 << " for " << *II << "\n");
1158 return VecCallCost - ScalarCallCost;
1161 llvm_unreachable("Unknown instruction");
1165 bool BoUpSLP::isFullyVectorizableTinyTree() {
1166 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1167 VectorizableTree.size() << " is fully vectorizable .\n");
1169 // We only handle trees of height 2.
1170 if (VectorizableTree.size() != 2)
1173 // Handle splat stores.
1174 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1177 // Gathering cost would be too much for tiny trees.
1178 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1184 int BoUpSLP::getTreeCost() {
1186 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1187 VectorizableTree.size() << ".\n");
1189 // We only vectorize tiny trees if it is fully vectorizable.
1190 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1191 if (!VectorizableTree.size()) {
1192 assert(!ExternalUses.size() && "We should not have any external users");
1197 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1199 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1200 int C = getEntryCost(&VectorizableTree[i]);
1201 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1202 << *VectorizableTree[i].Scalars[0] << " .\n");
1206 SmallSet<Value *, 16> ExtractCostCalculated;
1207 int ExtractCost = 0;
1208 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1210 // We only add extract cost once for the same scalar.
1211 if (!ExtractCostCalculated.insert(I->Scalar))
1214 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1215 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1219 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1220 return Cost + ExtractCost;
1223 int BoUpSLP::getGatherCost(Type *Ty) {
1225 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1226 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1230 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1231 // Find the type of the operands in VL.
1232 Type *ScalarTy = VL[0]->getType();
1233 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1234 ScalarTy = SI->getValueOperand()->getType();
1235 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1236 // Find the cost of inserting/extracting values from the vector.
1237 return getGatherCost(VecTy);
1240 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1241 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1242 return AA->getLocation(SI);
1243 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1244 return AA->getLocation(LI);
1245 return AliasAnalysis::Location();
1248 Value *BoUpSLP::getPointerOperand(Value *I) {
1249 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1250 return LI->getPointerOperand();
1251 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1252 return SI->getPointerOperand();
1256 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1257 if (LoadInst *L = dyn_cast<LoadInst>(I))
1258 return L->getPointerAddressSpace();
1259 if (StoreInst *S = dyn_cast<StoreInst>(I))
1260 return S->getPointerAddressSpace();
1264 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1265 Value *PtrA = getPointerOperand(A);
1266 Value *PtrB = getPointerOperand(B);
1267 unsigned ASA = getAddressSpaceOperand(A);
1268 unsigned ASB = getAddressSpaceOperand(B);
1270 // Check that the address spaces match and that the pointers are valid.
1271 if (!PtrA || !PtrB || (ASA != ASB))
1274 // Make sure that A and B are different pointers of the same type.
1275 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1278 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1279 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1280 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1282 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1283 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1284 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1286 APInt OffsetDelta = OffsetB - OffsetA;
1288 // Check if they are based on the same pointer. That makes the offsets
1291 return OffsetDelta == Size;
1293 // Compute the necessary base pointer delta to have the necessary final delta
1294 // equal to the size.
1295 APInt BaseDelta = Size - OffsetDelta;
1297 // Otherwise compute the distance with SCEV between the base pointers.
1298 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1299 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1300 const SCEV *C = SE->getConstant(BaseDelta);
1301 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1302 return X == PtrSCEVB;
1305 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1306 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1307 BasicBlock::iterator I = Src, E = Dst;
1308 /// Scan all of the instruction from SRC to DST and check if
1309 /// the source may alias.
1310 for (++I; I != E; ++I) {
1311 // Ignore store instructions that are marked as 'ignore'.
1312 if (MemBarrierIgnoreList.count(I))
1314 if (Src->mayWriteToMemory()) /* Write */ {
1315 if (!I->mayReadOrWriteMemory())
1318 if (!I->mayWriteToMemory())
1321 AliasAnalysis::Location A = getLocation(&*I);
1322 AliasAnalysis::Location B = getLocation(Src);
1324 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1330 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1331 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1332 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1333 BlockNumbering &BN = BlocksNumbers[BB];
1335 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1336 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1337 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1341 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1342 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1343 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1344 BlockNumbering &BN = BlocksNumbers[BB];
1346 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1347 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1348 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1349 Instruction *I = BN.getInstruction(MaxIdx);
1350 assert(I && "bad location");
1354 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1355 Instruction *VL0 = cast<Instruction>(VL[0]);
1356 Instruction *LastInst = getLastInstruction(VL);
1357 BasicBlock::iterator NextInst = LastInst;
1359 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1360 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1363 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1364 Value *Vec = UndefValue::get(Ty);
1365 // Generate the 'InsertElement' instruction.
1366 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1367 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1368 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1369 GatherSeq.insert(Insrt);
1370 CSEBlocks.insert(Insrt->getParent());
1372 // Add to our 'need-to-extract' list.
1373 if (ScalarToTreeEntry.count(VL[i])) {
1374 int Idx = ScalarToTreeEntry[VL[i]];
1375 TreeEntry *E = &VectorizableTree[Idx];
1376 // Find which lane we need to extract.
1378 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1379 // Is this the lane of the scalar that we are looking for ?
1380 if (E->Scalars[Lane] == VL[i]) {
1385 assert(FoundLane >= 0 && "Could not find the correct lane");
1386 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1394 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1395 SmallDenseMap<Value*, int>::const_iterator Entry
1396 = ScalarToTreeEntry.find(VL[0]);
1397 if (Entry != ScalarToTreeEntry.end()) {
1398 int Idx = Entry->second;
1399 const TreeEntry *En = &VectorizableTree[Idx];
1400 if (En->isSame(VL) && En->VectorizedValue)
1401 return En->VectorizedValue;
1406 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1407 if (ScalarToTreeEntry.count(VL[0])) {
1408 int Idx = ScalarToTreeEntry[VL[0]];
1409 TreeEntry *E = &VectorizableTree[Idx];
1411 return vectorizeTree(E);
1414 Type *ScalarTy = VL[0]->getType();
1415 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1416 ScalarTy = SI->getValueOperand()->getType();
1417 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1419 return Gather(VL, VecTy);
1422 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1423 IRBuilder<>::InsertPointGuard Guard(Builder);
1425 if (E->VectorizedValue) {
1426 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1427 return E->VectorizedValue;
1430 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1431 Type *ScalarTy = VL0->getType();
1432 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1433 ScalarTy = SI->getValueOperand()->getType();
1434 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1436 if (E->NeedToGather) {
1437 setInsertPointAfterBundle(E->Scalars);
1438 return Gather(E->Scalars, VecTy);
1441 unsigned Opcode = VL0->getOpcode();
1442 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1445 case Instruction::PHI: {
1446 PHINode *PH = dyn_cast<PHINode>(VL0);
1447 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1448 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1449 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1450 E->VectorizedValue = NewPhi;
1452 // PHINodes may have multiple entries from the same block. We want to
1453 // visit every block once.
1454 SmallSet<BasicBlock*, 4> VisitedBBs;
1456 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1458 BasicBlock *IBB = PH->getIncomingBlock(i);
1460 if (!VisitedBBs.insert(IBB)) {
1461 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1465 // Prepare the operand vector.
1466 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1467 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1468 getIncomingValueForBlock(IBB));
1470 Builder.SetInsertPoint(IBB->getTerminator());
1471 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1472 Value *Vec = vectorizeTree(Operands);
1473 NewPhi->addIncoming(Vec, IBB);
1476 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1477 "Invalid number of incoming values");
1481 case Instruction::ExtractElement: {
1482 if (CanReuseExtract(E->Scalars)) {
1483 Value *V = VL0->getOperand(0);
1484 E->VectorizedValue = V;
1487 return Gather(E->Scalars, VecTy);
1489 case Instruction::ZExt:
1490 case Instruction::SExt:
1491 case Instruction::FPToUI:
1492 case Instruction::FPToSI:
1493 case Instruction::FPExt:
1494 case Instruction::PtrToInt:
1495 case Instruction::IntToPtr:
1496 case Instruction::SIToFP:
1497 case Instruction::UIToFP:
1498 case Instruction::Trunc:
1499 case Instruction::FPTrunc:
1500 case Instruction::BitCast: {
1502 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1503 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1505 setInsertPointAfterBundle(E->Scalars);
1507 Value *InVec = vectorizeTree(INVL);
1509 if (Value *V = alreadyVectorized(E->Scalars))
1512 CastInst *CI = dyn_cast<CastInst>(VL0);
1513 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1514 E->VectorizedValue = V;
1517 case Instruction::FCmp:
1518 case Instruction::ICmp: {
1519 ValueList LHSV, RHSV;
1520 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1521 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1522 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1525 setInsertPointAfterBundle(E->Scalars);
1527 Value *L = vectorizeTree(LHSV);
1528 Value *R = vectorizeTree(RHSV);
1530 if (Value *V = alreadyVectorized(E->Scalars))
1533 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1535 if (Opcode == Instruction::FCmp)
1536 V = Builder.CreateFCmp(P0, L, R);
1538 V = Builder.CreateICmp(P0, L, R);
1540 E->VectorizedValue = V;
1543 case Instruction::Select: {
1544 ValueList TrueVec, FalseVec, CondVec;
1545 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1546 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1547 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1548 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1551 setInsertPointAfterBundle(E->Scalars);
1553 Value *Cond = vectorizeTree(CondVec);
1554 Value *True = vectorizeTree(TrueVec);
1555 Value *False = vectorizeTree(FalseVec);
1557 if (Value *V = alreadyVectorized(E->Scalars))
1560 Value *V = Builder.CreateSelect(Cond, True, False);
1561 E->VectorizedValue = V;
1564 case Instruction::Add:
1565 case Instruction::FAdd:
1566 case Instruction::Sub:
1567 case Instruction::FSub:
1568 case Instruction::Mul:
1569 case Instruction::FMul:
1570 case Instruction::UDiv:
1571 case Instruction::SDiv:
1572 case Instruction::FDiv:
1573 case Instruction::URem:
1574 case Instruction::SRem:
1575 case Instruction::FRem:
1576 case Instruction::Shl:
1577 case Instruction::LShr:
1578 case Instruction::AShr:
1579 case Instruction::And:
1580 case Instruction::Or:
1581 case Instruction::Xor: {
1582 ValueList LHSVL, RHSVL;
1583 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1584 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1586 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1587 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1588 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1591 setInsertPointAfterBundle(E->Scalars);
1593 Value *LHS = vectorizeTree(LHSVL);
1594 Value *RHS = vectorizeTree(RHSVL);
1596 if (LHS == RHS && isa<Instruction>(LHS)) {
1597 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1600 if (Value *V = alreadyVectorized(E->Scalars))
1603 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1604 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1605 E->VectorizedValue = V;
1607 if (Instruction *I = dyn_cast<Instruction>(V))
1608 return propagateMetadata(I, E->Scalars);
1612 case Instruction::Load: {
1613 // Loads are inserted at the head of the tree because we don't want to
1614 // sink them all the way down past store instructions.
1615 setInsertPointAfterBundle(E->Scalars);
1617 LoadInst *LI = cast<LoadInst>(VL0);
1618 unsigned AS = LI->getPointerAddressSpace();
1620 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1621 VecTy->getPointerTo(AS));
1622 unsigned Alignment = LI->getAlignment();
1623 LI = Builder.CreateLoad(VecPtr);
1624 LI->setAlignment(Alignment);
1625 E->VectorizedValue = LI;
1626 return propagateMetadata(LI, E->Scalars);
1628 case Instruction::Store: {
1629 StoreInst *SI = cast<StoreInst>(VL0);
1630 unsigned Alignment = SI->getAlignment();
1631 unsigned AS = SI->getPointerAddressSpace();
1634 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1635 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1637 setInsertPointAfterBundle(E->Scalars);
1639 Value *VecValue = vectorizeTree(ValueOp);
1640 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1641 VecTy->getPointerTo(AS));
1642 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1643 S->setAlignment(Alignment);
1644 E->VectorizedValue = S;
1645 return propagateMetadata(S, E->Scalars);
1647 case Instruction::Call: {
1648 CallInst *CI = cast<CallInst>(VL0);
1650 setInsertPointAfterBundle(E->Scalars);
1651 std::vector<Value *> OpVecs;
1652 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1654 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1655 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1656 OpVL.push_back(CEI->getArgOperand(j));
1659 Value *OpVec = vectorizeTree(OpVL);
1660 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1661 OpVecs.push_back(OpVec);
1664 Module *M = F->getParent();
1665 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1666 Intrinsic::ID ID = II->getIntrinsicID();
1667 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
1668 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
1669 Value *V = Builder.CreateCall(CF, OpVecs);
1670 E->VectorizedValue = V;
1674 llvm_unreachable("unknown inst");
1679 Value *BoUpSLP::vectorizeTree() {
1680 Builder.SetInsertPoint(F->getEntryBlock().begin());
1681 vectorizeTree(&VectorizableTree[0]);
1683 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1685 // Extract all of the elements with the external uses.
1686 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1688 Value *Scalar = it->Scalar;
1689 llvm::User *User = it->User;
1691 // Skip users that we already RAUW. This happens when one instruction
1692 // has multiple uses of the same value.
1693 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
1696 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1698 int Idx = ScalarToTreeEntry[Scalar];
1699 TreeEntry *E = &VectorizableTree[Idx];
1700 assert(!E->NeedToGather && "Extracting from a gather list");
1702 Value *Vec = E->VectorizedValue;
1703 assert(Vec && "Can't find vectorizable value");
1705 Value *Lane = Builder.getInt32(it->Lane);
1706 // Generate extracts for out-of-tree users.
1707 // Find the insertion point for the extractelement lane.
1708 if (isa<Instruction>(Vec)){
1709 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1710 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1711 if (PH->getIncomingValue(i) == Scalar) {
1712 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1713 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1714 CSEBlocks.insert(PH->getIncomingBlock(i));
1715 PH->setOperand(i, Ex);
1719 Builder.SetInsertPoint(cast<Instruction>(User));
1720 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1721 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1722 User->replaceUsesOfWith(Scalar, Ex);
1725 Builder.SetInsertPoint(F->getEntryBlock().begin());
1726 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1727 CSEBlocks.insert(&F->getEntryBlock());
1728 User->replaceUsesOfWith(Scalar, Ex);
1731 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1734 // For each vectorized value:
1735 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1736 TreeEntry *Entry = &VectorizableTree[EIdx];
1739 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1740 Value *Scalar = Entry->Scalars[Lane];
1742 // No need to handle users of gathered values.
1743 if (Entry->NeedToGather)
1746 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1748 Type *Ty = Scalar->getType();
1749 if (!Ty->isVoidTy()) {
1751 for (User *U : Scalar->users()) {
1752 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
1754 assert((ScalarToTreeEntry.count(U) ||
1755 // It is legal to replace users in the ignorelist by undef.
1756 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
1757 UserIgnoreList.end())) &&
1758 "Replacing out-of-tree value with undef");
1761 Value *Undef = UndefValue::get(Ty);
1762 Scalar->replaceAllUsesWith(Undef);
1764 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1765 cast<Instruction>(Scalar)->eraseFromParent();
1769 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1770 BlocksNumbers[it].forget();
1772 Builder.ClearInsertionPoint();
1774 return VectorizableTree[0].VectorizedValue;
1777 void BoUpSLP::optimizeGatherSequence() {
1778 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1779 << " gather sequences instructions.\n");
1780 // LICM InsertElementInst sequences.
1781 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1782 e = GatherSeq.end(); it != e; ++it) {
1783 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1788 // Check if this block is inside a loop.
1789 Loop *L = LI->getLoopFor(Insert->getParent());
1793 // Check if it has a preheader.
1794 BasicBlock *PreHeader = L->getLoopPreheader();
1798 // If the vector or the element that we insert into it are
1799 // instructions that are defined in this basic block then we can't
1800 // hoist this instruction.
1801 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1802 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1803 if (CurrVec && L->contains(CurrVec))
1805 if (NewElem && L->contains(NewElem))
1808 // We can hoist this instruction. Move it to the pre-header.
1809 Insert->moveBefore(PreHeader->getTerminator());
1812 // Sort blocks by domination. This ensures we visit a block after all blocks
1813 // dominating it are visited.
1814 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1815 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1816 [this](const BasicBlock *A, const BasicBlock *B) {
1817 return DT->properlyDominates(A, B);
1820 // Perform O(N^2) search over the gather sequences and merge identical
1821 // instructions. TODO: We can further optimize this scan if we split the
1822 // instructions into different buckets based on the insert lane.
1823 SmallVector<Instruction *, 16> Visited;
1824 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1825 E = CSEWorkList.end();
1827 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1828 "Worklist not sorted properly!");
1829 BasicBlock *BB = *I;
1830 // For all instructions in blocks containing gather sequences:
1831 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1832 Instruction *In = it++;
1833 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1836 // Check if we can replace this instruction with any of the
1837 // visited instructions.
1838 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1841 if (In->isIdenticalTo(*v) &&
1842 DT->dominates((*v)->getParent(), In->getParent())) {
1843 In->replaceAllUsesWith(*v);
1844 In->eraseFromParent();
1850 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1851 Visited.push_back(In);
1859 /// The SLPVectorizer Pass.
1860 struct SLPVectorizer : public FunctionPass {
1861 typedef SmallVector<StoreInst *, 8> StoreList;
1862 typedef MapVector<Value *, StoreList> StoreListMap;
1864 /// Pass identification, replacement for typeid
1867 explicit SLPVectorizer() : FunctionPass(ID) {
1868 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1871 ScalarEvolution *SE;
1872 const DataLayout *DL;
1873 TargetTransformInfo *TTI;
1878 bool runOnFunction(Function &F) override {
1879 if (skipOptnoneFunction(F))
1882 SE = &getAnalysis<ScalarEvolution>();
1883 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1884 DL = DLP ? &DLP->getDataLayout() : nullptr;
1885 TTI = &getAnalysis<TargetTransformInfo>();
1886 AA = &getAnalysis<AliasAnalysis>();
1887 LI = &getAnalysis<LoopInfo>();
1888 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1891 bool Changed = false;
1893 // If the target claims to have no vector registers don't attempt
1895 if (!TTI->getNumberOfRegisters(true))
1898 // Must have DataLayout. We can't require it because some tests run w/o
1903 // Don't vectorize when the attribute NoImplicitFloat is used.
1904 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1907 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1909 // Use the bottom up slp vectorizer to construct chains that start with
1910 // he store instructions.
1911 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1913 // Scan the blocks in the function in post order.
1914 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1915 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1916 BasicBlock *BB = *it;
1918 // Vectorize trees that end at stores.
1919 if (unsigned count = collectStores(BB, R)) {
1921 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1922 Changed |= vectorizeStoreChains(R);
1925 // Vectorize trees that end at reductions.
1926 Changed |= vectorizeChainsInBlock(BB, R);
1930 R.optimizeGatherSequence();
1931 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1932 DEBUG(verifyFunction(F));
1937 void getAnalysisUsage(AnalysisUsage &AU) const override {
1938 FunctionPass::getAnalysisUsage(AU);
1939 AU.addRequired<ScalarEvolution>();
1940 AU.addRequired<AliasAnalysis>();
1941 AU.addRequired<TargetTransformInfo>();
1942 AU.addRequired<LoopInfo>();
1943 AU.addRequired<DominatorTreeWrapperPass>();
1944 AU.addPreserved<LoopInfo>();
1945 AU.addPreserved<DominatorTreeWrapperPass>();
1946 AU.setPreservesCFG();
1951 /// \brief Collect memory references and sort them according to their base
1952 /// object. We sort the stores to their base objects to reduce the cost of the
1953 /// quadratic search on the stores. TODO: We can further reduce this cost
1954 /// if we flush the chain creation every time we run into a memory barrier.
1955 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1957 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1958 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1960 /// \brief Try to vectorize a list of operands.
1961 /// \@param BuildVector A list of users to ignore for the purpose of
1962 /// scheduling and that don't need extracting.
1963 /// \returns true if a value was vectorized.
1964 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
1965 ArrayRef<Value *> BuildVector = None);
1967 /// \brief Try to vectorize a chain that may start at the operands of \V;
1968 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1970 /// \brief Vectorize the stores that were collected in StoreRefs.
1971 bool vectorizeStoreChains(BoUpSLP &R);
1973 /// \brief Scan the basic block and look for patterns that are likely to start
1974 /// a vectorization chain.
1975 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1977 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1980 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1983 StoreListMap StoreRefs;
1986 /// \brief Check that the Values in the slice in VL array are still existent in
1987 /// the WeakVH array.
1988 /// Vectorization of part of the VL array may cause later values in the VL array
1989 /// to become invalid. We track when this has happened in the WeakVH array.
1990 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1991 SmallVectorImpl<WeakVH> &VH,
1992 unsigned SliceBegin,
1993 unsigned SliceSize) {
1994 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
2001 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
2002 int CostThreshold, BoUpSLP &R) {
2003 unsigned ChainLen = Chain.size();
2004 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2006 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2007 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2008 unsigned VF = MinVecRegSize / Sz;
2010 if (!isPowerOf2_32(Sz) || VF < 2)
2013 // Keep track of values that were deleted by vectorizing in the loop below.
2014 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2016 bool Changed = false;
2017 // Look for profitable vectorizable trees at all offsets, starting at zero.
2018 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2022 // Check that a previous iteration of this loop did not delete the Value.
2023 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2026 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2028 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2030 R.buildTree(Operands);
2032 int Cost = R.getTreeCost();
2034 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2035 if (Cost < CostThreshold) {
2036 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2039 // Move to the next bundle.
2048 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2049 int costThreshold, BoUpSLP &R) {
2050 SetVector<Value *> Heads, Tails;
2051 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2053 // We may run into multiple chains that merge into a single chain. We mark the
2054 // stores that we vectorized so that we don't visit the same store twice.
2055 BoUpSLP::ValueSet VectorizedStores;
2056 bool Changed = false;
2058 // Do a quadratic search on all of the given stores and find
2059 // all of the pairs of stores that follow each other.
2060 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2061 for (unsigned j = 0; j < e; ++j) {
2065 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2066 Tails.insert(Stores[j]);
2067 Heads.insert(Stores[i]);
2068 ConsecutiveChain[Stores[i]] = Stores[j];
2073 // For stores that start but don't end a link in the chain:
2074 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2076 if (Tails.count(*it))
2079 // We found a store instr that starts a chain. Now follow the chain and try
2081 BoUpSLP::ValueList Operands;
2083 // Collect the chain into a list.
2084 while (Tails.count(I) || Heads.count(I)) {
2085 if (VectorizedStores.count(I))
2087 Operands.push_back(I);
2088 // Move to the next value in the chain.
2089 I = ConsecutiveChain[I];
2092 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2094 // Mark the vectorized stores so that we don't vectorize them again.
2096 VectorizedStores.insert(Operands.begin(), Operands.end());
2097 Changed |= Vectorized;
2104 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2107 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2108 StoreInst *SI = dyn_cast<StoreInst>(it);
2112 // Don't touch volatile stores.
2113 if (!SI->isSimple())
2116 // Check that the pointer points to scalars.
2117 Type *Ty = SI->getValueOperand()->getType();
2118 if (Ty->isAggregateType() || Ty->isVectorTy())
2121 // Find the base pointer.
2122 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2124 // Save the store locations.
2125 StoreRefs[Ptr].push_back(SI);
2131 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2134 Value *VL[] = { A, B };
2135 return tryToVectorizeList(VL, R);
2138 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2139 ArrayRef<Value *> BuildVector) {
2143 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2145 // Check that all of the parts are scalar instructions of the same type.
2146 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2150 unsigned Opcode0 = I0->getOpcode();
2152 Type *Ty0 = I0->getType();
2153 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2154 unsigned VF = MinVecRegSize / Sz;
2156 for (int i = 0, e = VL.size(); i < e; ++i) {
2157 Type *Ty = VL[i]->getType();
2158 if (Ty->isAggregateType() || Ty->isVectorTy())
2160 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2161 if (!Inst || Inst->getOpcode() != Opcode0)
2165 bool Changed = false;
2167 // Keep track of values that were deleted by vectorizing in the loop below.
2168 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2170 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2171 unsigned OpsWidth = 0;
2178 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2181 // Check that a previous iteration of this loop did not delete the Value.
2182 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2185 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2187 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2189 ArrayRef<Value *> BuildVectorSlice;
2190 if (!BuildVector.empty())
2191 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
2193 R.buildTree(Ops, BuildVectorSlice);
2194 int Cost = R.getTreeCost();
2196 if (Cost < -SLPCostThreshold) {
2197 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
2198 Value *VectorizedRoot = R.vectorizeTree();
2200 // Reconstruct the build vector by extracting the vectorized root. This
2201 // way we handle the case where some elements of the vector are undefined.
2202 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
2203 if (!BuildVectorSlice.empty()) {
2204 Instruction *InsertAfter = cast<Instruction>(VectorizedRoot);
2205 for (auto &V : BuildVectorSlice) {
2206 InsertElementInst *IE = cast<InsertElementInst>(V);
2207 IRBuilder<> Builder(++BasicBlock::iterator(InsertAfter));
2208 Instruction *Extract = cast<Instruction>(
2209 Builder.CreateExtractElement(VectorizedRoot, IE->getOperand(2)));
2210 IE->setOperand(1, Extract);
2211 IE->removeFromParent();
2212 IE->insertAfter(Extract);
2216 // Move to the next bundle.
2225 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2229 // Try to vectorize V.
2230 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2233 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2234 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2236 if (B && B->hasOneUse()) {
2237 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2238 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2239 if (tryToVectorizePair(A, B0, R)) {
2243 if (tryToVectorizePair(A, B1, R)) {
2250 if (A && A->hasOneUse()) {
2251 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2252 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2253 if (tryToVectorizePair(A0, B, R)) {
2257 if (tryToVectorizePair(A1, B, R)) {
2265 /// \brief Generate a shuffle mask to be used in a reduction tree.
2267 /// \param VecLen The length of the vector to be reduced.
2268 /// \param NumEltsToRdx The number of elements that should be reduced in the
2270 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2271 /// reduction. A pairwise reduction will generate a mask of
2272 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2273 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2274 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2275 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2276 bool IsPairwise, bool IsLeft,
2277 IRBuilder<> &Builder) {
2278 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2280 SmallVector<Constant *, 32> ShuffleMask(
2281 VecLen, UndefValue::get(Builder.getInt32Ty()));
2284 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2285 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2286 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2288 // Move the upper half of the vector to the lower half.
2289 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2290 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2292 return ConstantVector::get(ShuffleMask);
2296 /// Model horizontal reductions.
2298 /// A horizontal reduction is a tree of reduction operations (currently add and
2299 /// fadd) that has operations that can be put into a vector as its leaf.
2300 /// For example, this tree:
2307 /// This tree has "mul" as its reduced values and "+" as its reduction
2308 /// operations. A reduction might be feeding into a store or a binary operation
2323 class HorizontalReduction {
2324 SmallVector<Value *, 16> ReductionOps;
2325 SmallVector<Value *, 32> ReducedVals;
2327 BinaryOperator *ReductionRoot;
2328 PHINode *ReductionPHI;
2330 /// The opcode of the reduction.
2331 unsigned ReductionOpcode;
2332 /// The opcode of the values we perform a reduction on.
2333 unsigned ReducedValueOpcode;
2334 /// The width of one full horizontal reduction operation.
2335 unsigned ReduxWidth;
2336 /// Should we model this reduction as a pairwise reduction tree or a tree that
2337 /// splits the vector in halves and adds those halves.
2338 bool IsPairwiseReduction;
2341 HorizontalReduction()
2342 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
2343 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2345 /// \brief Try to find a reduction tree.
2346 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2347 const DataLayout *DL) {
2349 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2350 "Thi phi needs to use the binary operator");
2352 // We could have a initial reductions that is not an add.
2353 // r *= v1 + v2 + v3 + v4
2354 // In such a case start looking for a tree rooted in the first '+'.
2356 if (B->getOperand(0) == Phi) {
2358 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2359 } else if (B->getOperand(1) == Phi) {
2361 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2368 Type *Ty = B->getType();
2369 if (Ty->isVectorTy())
2372 ReductionOpcode = B->getOpcode();
2373 ReducedValueOpcode = 0;
2374 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2381 // We currently only support adds.
2382 if (ReductionOpcode != Instruction::Add &&
2383 ReductionOpcode != Instruction::FAdd)
2386 // Post order traverse the reduction tree starting at B. We only handle true
2387 // trees containing only binary operators.
2388 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2389 Stack.push_back(std::make_pair(B, 0));
2390 while (!Stack.empty()) {
2391 BinaryOperator *TreeN = Stack.back().first;
2392 unsigned EdgeToVist = Stack.back().second++;
2393 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2395 // Only handle trees in the current basic block.
2396 if (TreeN->getParent() != B->getParent())
2399 // Each tree node needs to have one user except for the ultimate
2401 if (!TreeN->hasOneUse() && TreeN != B)
2405 if (EdgeToVist == 2 || IsReducedValue) {
2406 if (IsReducedValue) {
2407 // Make sure that the opcodes of the operations that we are going to
2409 if (!ReducedValueOpcode)
2410 ReducedValueOpcode = TreeN->getOpcode();
2411 else if (ReducedValueOpcode != TreeN->getOpcode())
2413 ReducedVals.push_back(TreeN);
2415 // We need to be able to reassociate the adds.
2416 if (!TreeN->isAssociative())
2418 ReductionOps.push_back(TreeN);
2425 // Visit left or right.
2426 Value *NextV = TreeN->getOperand(EdgeToVist);
2427 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2429 Stack.push_back(std::make_pair(Next, 0));
2430 else if (NextV != Phi)
2436 /// \brief Attempt to vectorize the tree found by
2437 /// matchAssociativeReduction.
2438 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2439 if (ReducedVals.empty())
2442 unsigned NumReducedVals = ReducedVals.size();
2443 if (NumReducedVals < ReduxWidth)
2446 Value *VectorizedTree = nullptr;
2447 IRBuilder<> Builder(ReductionRoot);
2448 FastMathFlags Unsafe;
2449 Unsafe.setUnsafeAlgebra();
2450 Builder.SetFastMathFlags(Unsafe);
2453 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2454 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2455 V.buildTree(ValsToReduce, ReductionOps);
2458 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2459 if (Cost >= -SLPCostThreshold)
2462 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2465 // Vectorize a tree.
2466 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2467 Value *VectorizedRoot = V.vectorizeTree();
2469 // Emit a reduction.
2470 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2471 if (VectorizedTree) {
2472 Builder.SetCurrentDebugLocation(Loc);
2473 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2474 ReducedSubTree, "bin.rdx");
2476 VectorizedTree = ReducedSubTree;
2479 if (VectorizedTree) {
2480 // Finish the reduction.
2481 for (; i < NumReducedVals; ++i) {
2482 Builder.SetCurrentDebugLocation(
2483 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2484 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2489 assert(ReductionRoot && "Need a reduction operation");
2490 ReductionRoot->setOperand(0, VectorizedTree);
2491 ReductionRoot->setOperand(1, ReductionPHI);
2493 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2495 return VectorizedTree != nullptr;
2500 /// \brief Calcuate the cost of a reduction.
2501 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2502 Type *ScalarTy = FirstReducedVal->getType();
2503 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2505 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2506 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2508 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2509 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2511 int ScalarReduxCost =
2512 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2514 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2515 << " for reduction that starts with " << *FirstReducedVal
2517 << (IsPairwiseReduction ? "pairwise" : "splitting")
2518 << " reduction)\n");
2520 return VecReduxCost - ScalarReduxCost;
2523 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2524 Value *R, const Twine &Name = "") {
2525 if (Opcode == Instruction::FAdd)
2526 return Builder.CreateFAdd(L, R, Name);
2527 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2530 /// \brief Emit a horizontal reduction of the vectorized value.
2531 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2532 assert(VectorizedValue && "Need to have a vectorized tree node");
2533 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2534 assert(isPowerOf2_32(ReduxWidth) &&
2535 "We only handle power-of-two reductions for now");
2537 Value *TmpVec = ValToReduce;
2538 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2539 if (IsPairwiseReduction) {
2541 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2543 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2545 Value *LeftShuf = Builder.CreateShuffleVector(
2546 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2547 Value *RightShuf = Builder.CreateShuffleVector(
2548 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2550 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2554 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2555 Value *Shuf = Builder.CreateShuffleVector(
2556 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2557 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2561 // The result is in the first element of the vector.
2562 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2566 /// \brief Recognize construction of vectors like
2567 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2568 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2569 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2570 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2572 /// Returns true if it matches
2574 static bool findBuildVector(InsertElementInst *FirstInsertElem,
2575 SmallVectorImpl<Value *> &BuildVector,
2576 SmallVectorImpl<Value *> &BuildVectorOpds) {
2577 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
2580 InsertElementInst *IE = FirstInsertElem;
2582 BuildVector.push_back(IE);
2583 BuildVectorOpds.push_back(IE->getOperand(1));
2585 if (IE->use_empty())
2588 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2592 // If this isn't the final use, make sure the next insertelement is the only
2593 // use. It's OK if the final constructed vector is used multiple times
2594 if (!IE->hasOneUse())
2603 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2604 return V->getType() < V2->getType();
2607 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2608 bool Changed = false;
2609 SmallVector<Value *, 4> Incoming;
2610 SmallSet<Value *, 16> VisitedInstrs;
2612 bool HaveVectorizedPhiNodes = true;
2613 while (HaveVectorizedPhiNodes) {
2614 HaveVectorizedPhiNodes = false;
2616 // Collect the incoming values from the PHIs.
2618 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2620 PHINode *P = dyn_cast<PHINode>(instr);
2624 if (!VisitedInstrs.count(P))
2625 Incoming.push_back(P);
2629 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2631 // Try to vectorize elements base on their type.
2632 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2636 // Look for the next elements with the same type.
2637 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2638 while (SameTypeIt != E &&
2639 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2640 VisitedInstrs.insert(*SameTypeIt);
2644 // Try to vectorize them.
2645 unsigned NumElts = (SameTypeIt - IncIt);
2646 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2648 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2649 // Success start over because instructions might have been changed.
2650 HaveVectorizedPhiNodes = true;
2655 // Start over at the next instruction of a different type (or the end).
2660 VisitedInstrs.clear();
2662 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2663 // We may go through BB multiple times so skip the one we have checked.
2664 if (!VisitedInstrs.insert(it))
2667 if (isa<DbgInfoIntrinsic>(it))
2670 // Try to vectorize reductions that use PHINodes.
2671 if (PHINode *P = dyn_cast<PHINode>(it)) {
2672 // Check that the PHI is a reduction PHI.
2673 if (P->getNumIncomingValues() != 2)
2676 (P->getIncomingBlock(0) == BB
2677 ? (P->getIncomingValue(0))
2678 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
2680 // Check if this is a Binary Operator.
2681 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2685 // Try to match and vectorize a horizontal reduction.
2686 HorizontalReduction HorRdx;
2687 if (ShouldVectorizeHor &&
2688 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2689 HorRdx.tryToReduce(R, TTI)) {
2696 Value *Inst = BI->getOperand(0);
2698 Inst = BI->getOperand(1);
2700 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2701 // We would like to start over since some instructions are deleted
2702 // and the iterator may become invalid value.
2712 // Try to vectorize horizontal reductions feeding into a store.
2713 if (ShouldStartVectorizeHorAtStore)
2714 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2715 if (BinaryOperator *BinOp =
2716 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2717 HorizontalReduction HorRdx;
2718 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
2719 HorRdx.tryToReduce(R, TTI)) ||
2720 tryToVectorize(BinOp, R))) {
2728 // Try to vectorize trees that start at compare instructions.
2729 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2730 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2732 // We would like to start over since some instructions are deleted
2733 // and the iterator may become invalid value.
2739 for (int i = 0; i < 2; ++i) {
2740 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2741 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2743 // We would like to start over since some instructions are deleted
2744 // and the iterator may become invalid value.
2753 // Try to vectorize trees that start at insertelement instructions.
2754 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
2755 SmallVector<Value *, 16> BuildVector;
2756 SmallVector<Value *, 16> BuildVectorOpds;
2757 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
2760 // Vectorize starting with the build vector operands ignoring the
2761 // BuildVector instructions for the purpose of scheduling and user
2763 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
2776 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2777 bool Changed = false;
2778 // Attempt to sort and vectorize each of the store-groups.
2779 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2781 if (it->second.size() < 2)
2784 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2785 << it->second.size() << ".\n");
2787 // Process the stores in chunks of 16.
2788 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2789 unsigned Len = std::min<unsigned>(CE - CI, 16);
2790 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2791 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2797 } // end anonymous namespace
2799 char SLPVectorizer::ID = 0;
2800 static const char lv_name[] = "SLP Vectorizer";
2801 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2802 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2803 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2804 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2805 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2806 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2809 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }