1 //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // This pass implements the Bottom Up SLP vectorizer. It detects consecutive
10 // stores that can be put together into vector-stores. Next, it attempts to
11 // construct vectorizable tree using the use-def chains. If a profitable tree
12 // was found, the SLP vectorizer performs vectorization on the tree.
14 // The pass is inspired by the work described in the paper:
15 // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
17 //===----------------------------------------------------------------------===//
18 #define SV_NAME "slp-vectorizer"
19 #define DEBUG_TYPE "SLP"
21 #include "llvm/Transforms/Vectorize.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/IR/Verifier.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
50 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
51 cl::desc("Only vectorize if you gain more than this "
55 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
56 cl::desc("Attempt to vectorize horizontal reductions"));
58 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
59 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
61 "Attempt to vectorize horizontal reductions feeding into a store"));
65 static const unsigned MinVecRegSize = 128;
67 static const unsigned RecursionMaxDepth = 12;
69 /// A helper class for numbering instructions in multiple blocks.
70 /// Numbers start at zero for each basic block.
71 struct BlockNumbering {
73 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
75 BlockNumbering() : BB(0), Valid(false) {}
77 void numberInstructions() {
81 // Number the instructions in the block.
82 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
84 InstrVec.push_back(it);
85 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
90 int getIndex(Instruction *I) {
91 assert(I->getParent() == BB && "Invalid instruction");
94 assert(InstrIdx.count(I) && "Unknown instruction");
98 Instruction *getInstruction(unsigned loc) {
100 numberInstructions();
101 assert(InstrVec.size() > loc && "Invalid Index");
102 return InstrVec[loc];
105 void forget() { Valid = false; }
108 /// The block we are numbering.
110 /// Is the block numbered.
112 /// Maps instructions to numbers and back.
113 SmallDenseMap<Instruction *, int> InstrIdx;
114 /// Maps integers to Instructions.
115 SmallVector<Instruction *, 32> InstrVec;
118 /// \returns the parent basic block if all of the instructions in \p VL
119 /// are in the same block or null otherwise.
120 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
121 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
124 BasicBlock *BB = I0->getParent();
125 for (int i = 1, e = VL.size(); i < e; i++) {
126 Instruction *I = dyn_cast<Instruction>(VL[i]);
130 if (BB != I->getParent())
136 /// \returns True if all of the values in \p VL are constants.
137 static bool allConstant(ArrayRef<Value *> VL) {
138 for (unsigned i = 0, e = VL.size(); i < e; ++i)
139 if (!isa<Constant>(VL[i]))
144 /// \returns True if all of the values in \p VL are identical.
145 static bool isSplat(ArrayRef<Value *> VL) {
146 for (unsigned i = 1, e = VL.size(); i < e; ++i)
152 /// \returns The opcode if all of the Instructions in \p VL have the same
154 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
155 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
158 unsigned Opcode = I0->getOpcode();
159 for (int i = 1, e = VL.size(); i < e; i++) {
160 Instruction *I = dyn_cast<Instruction>(VL[i]);
161 if (!I || Opcode != I->getOpcode())
167 /// \returns \p I after propagating metadata from \p VL.
168 static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) {
169 Instruction *I0 = cast<Instruction>(VL[0]);
170 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
171 I0->getAllMetadataOtherThanDebugLoc(Metadata);
173 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) {
174 unsigned Kind = Metadata[i].first;
175 MDNode *MD = Metadata[i].second;
177 for (int i = 1, e = VL.size(); MD && i != e; i++) {
178 Instruction *I = cast<Instruction>(VL[i]);
179 MDNode *IMD = I->getMetadata(Kind);
183 MD = 0; // Remove unknown metadata
185 case LLVMContext::MD_tbaa:
186 MD = MDNode::getMostGenericTBAA(MD, IMD);
188 case LLVMContext::MD_fpmath:
189 MD = MDNode::getMostGenericFPMath(MD, IMD);
193 I->setMetadata(Kind, MD);
198 /// \returns The type that all of the values in \p VL have or null if there
199 /// are different types.
200 static Type* getSameType(ArrayRef<Value *> VL) {
201 Type *Ty = VL[0]->getType();
202 for (int i = 1, e = VL.size(); i < e; i++)
203 if (VL[i]->getType() != Ty)
209 /// \returns True if the ExtractElement instructions in VL can be vectorized
210 /// to use the original vector.
211 static bool CanReuseExtract(ArrayRef<Value *> VL) {
212 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
213 // Check if all of the extracts come from the same vector and from the
216 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
217 Value *Vec = E0->getOperand(0);
219 // We have to extract from the same vector type.
220 unsigned NElts = Vec->getType()->getVectorNumElements();
222 if (NElts != VL.size())
225 // Check that all of the indices extract from the correct offset.
226 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
227 if (!CI || CI->getZExtValue())
230 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
231 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
232 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
234 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
241 static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
242 SmallVectorImpl<Value *> &Left,
243 SmallVectorImpl<Value *> &Right) {
245 SmallVector<Value *, 16> OrigLeft, OrigRight;
247 bool AllSameOpcodeLeft = true;
248 bool AllSameOpcodeRight = true;
249 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
250 Instruction *I = cast<Instruction>(VL[i]);
251 Value *V0 = I->getOperand(0);
252 Value *V1 = I->getOperand(1);
254 OrigLeft.push_back(V0);
255 OrigRight.push_back(V1);
257 Instruction *I0 = dyn_cast<Instruction>(V0);
258 Instruction *I1 = dyn_cast<Instruction>(V1);
260 // Check whether all operands on one side have the same opcode. In this case
261 // we want to preserve the original order and not make things worse by
263 AllSameOpcodeLeft = I0;
264 AllSameOpcodeRight = I1;
266 if (i && AllSameOpcodeLeft) {
267 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) {
268 if(P0->getOpcode() != I0->getOpcode())
269 AllSameOpcodeLeft = false;
271 AllSameOpcodeLeft = false;
273 if (i && AllSameOpcodeRight) {
274 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) {
275 if(P1->getOpcode() != I1->getOpcode())
276 AllSameOpcodeRight = false;
278 AllSameOpcodeRight = false;
281 // Sort two opcodes. In the code below we try to preserve the ability to use
282 // broadcast of values instead of individual inserts.
289 // If we just sorted according to opcode we would leave the first line in
290 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load).
293 // Because vr2 and vr1 are from the same load we loose the opportunity of a
294 // broadcast for the packed right side in the backend: we have [vr1, vl2]
295 // instead of [vr1, vr2=vr1].
297 if(!i && I0->getOpcode() > I1->getOpcode()) {
300 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) {
301 // Try not to destroy a broad cast for no apparent benefit.
304 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) {
305 // Try preserve broadcasts.
308 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) {
309 // Try preserve broadcasts.
318 // One opcode, put the instruction on the right.
328 bool LeftBroadcast = isSplat(Left);
329 bool RightBroadcast = isSplat(Right);
331 // Don't reorder if the operands where good to begin with.
332 if (!(LeftBroadcast || RightBroadcast) &&
333 (AllSameOpcodeRight || AllSameOpcodeLeft)) {
339 /// Bottom Up SLP Vectorizer.
342 typedef SmallVector<Value *, 8> ValueList;
343 typedef SmallVector<Instruction *, 16> InstrList;
344 typedef SmallPtrSet<Value *, 16> ValueSet;
345 typedef SmallVector<StoreInst *, 8> StoreList;
347 BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl,
348 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li,
350 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt),
351 Builder(Se->getContext()) {
352 // Setup the block numbering utility for all of the blocks in the
354 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
356 BlocksNumbers[BB] = BlockNumbering(BB);
360 /// \brief Vectorize the tree that starts with the elements in \p VL.
361 /// Returns the vectorized root.
362 Value *vectorizeTree();
364 /// \returns the vectorization cost of the subtree that starts at \p VL.
365 /// A negative number means that this is profitable.
368 /// Construct a vectorizable tree that starts at \p Roots and is possibly
369 /// used by a reduction of \p RdxOps.
370 void buildTree(ArrayRef<Value *> Roots, ValueSet *RdxOps = 0);
372 /// Clear the internal data structures that are created by 'buildTree'.
375 VectorizableTree.clear();
376 ScalarToTreeEntry.clear();
378 ExternalUses.clear();
379 MemBarrierIgnoreList.clear();
382 /// \returns true if the memory operations A and B are consecutive.
383 bool isConsecutiveAccess(Value *A, Value *B);
385 /// \brief Perform LICM and CSE on the newly generated gather sequences.
386 void optimizeGatherSequence();
390 /// \returns the cost of the vectorizable entry.
391 int getEntryCost(TreeEntry *E);
393 /// This is the recursive part of buildTree.
394 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
396 /// Vectorize a single entry in the tree.
397 Value *vectorizeTree(TreeEntry *E);
399 /// Vectorize a single entry in the tree, starting in \p VL.
400 Value *vectorizeTree(ArrayRef<Value *> VL);
402 /// \returns the pointer to the vectorized value if \p VL is already
403 /// vectorized, or NULL. They may happen in cycles.
404 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
406 /// \brief Take the pointer operand from the Load/Store instruction.
407 /// \returns NULL if this is not a valid Load/Store instruction.
408 static Value *getPointerOperand(Value *I);
410 /// \brief Take the address space operand from the Load/Store instruction.
411 /// \returns -1 if this is not a valid Load/Store instruction.
412 static unsigned getAddressSpaceOperand(Value *I);
414 /// \returns the scalarization cost for this type. Scalarization in this
415 /// context means the creation of vectors from a group of scalars.
416 int getGatherCost(Type *Ty);
418 /// \returns the scalarization cost for this list of values. Assuming that
419 /// this subtree gets vectorized, we may need to extract the values from the
420 /// roots. This method calculates the cost of extracting the values.
421 int getGatherCost(ArrayRef<Value *> VL);
423 /// \returns the AA location that is being access by the instruction.
424 AliasAnalysis::Location getLocation(Instruction *I);
426 /// \brief Checks if it is possible to sink an instruction from
427 /// \p Src to \p Dst.
428 /// \returns the pointer to the barrier instruction if we can't sink.
429 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
431 /// \returns the index of the last instruction in the BB from \p VL.
432 int getLastIndex(ArrayRef<Value *> VL);
434 /// \returns the Instruction in the bundle \p VL.
435 Instruction *getLastInstruction(ArrayRef<Value *> VL);
437 /// \brief Set the Builder insert point to one after the last instruction in
439 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
441 /// \returns a vector from a collection of scalars in \p VL.
442 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
444 /// \returns whether the VectorizableTree is fully vectoriable and will
445 /// be beneficial even the tree height is tiny.
446 bool isFullyVectorizableTinyTree();
449 TreeEntry() : Scalars(), VectorizedValue(0), LastScalarIndex(0),
452 /// \returns true if the scalars in VL are equal to this entry.
453 bool isSame(ArrayRef<Value *> VL) const {
454 assert(VL.size() == Scalars.size() && "Invalid size");
455 return std::equal(VL.begin(), VL.end(), Scalars.begin());
458 /// A vector of scalars.
461 /// The Scalars are vectorized into this value. It is initialized to Null.
462 Value *VectorizedValue;
464 /// The index in the basic block of the last scalar.
467 /// Do we need to gather this sequence ?
471 /// Create a new VectorizableTree entry.
472 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
473 VectorizableTree.push_back(TreeEntry());
474 int idx = VectorizableTree.size() - 1;
475 TreeEntry *Last = &VectorizableTree[idx];
476 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
477 Last->NeedToGather = !Vectorized;
479 Last->LastScalarIndex = getLastIndex(VL);
480 for (int i = 0, e = VL.size(); i != e; ++i) {
481 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
482 ScalarToTreeEntry[VL[i]] = idx;
485 Last->LastScalarIndex = 0;
486 MustGather.insert(VL.begin(), VL.end());
491 /// -- Vectorization State --
492 /// Holds all of the tree entries.
493 std::vector<TreeEntry> VectorizableTree;
495 /// Maps a specific scalar to its tree entry.
496 SmallDenseMap<Value*, int> ScalarToTreeEntry;
498 /// A list of scalars that we found that we need to keep as scalars.
501 /// This POD struct describes one external user in the vectorized tree.
502 struct ExternalUser {
503 ExternalUser (Value *S, llvm::User *U, int L) :
504 Scalar(S), User(U), Lane(L){};
505 // Which scalar in our function.
507 // Which user that uses the scalar.
509 // Which lane does the scalar belong to.
512 typedef SmallVector<ExternalUser, 16> UserList;
514 /// A list of values that need to extracted out of the tree.
515 /// This list holds pairs of (Internal Scalar : External User).
516 UserList ExternalUses;
518 /// A list of instructions to ignore while sinking
519 /// memory instructions. This map must be reset between runs of getCost.
520 ValueSet MemBarrierIgnoreList;
522 /// Holds all of the instructions that we gathered.
523 SetVector<Instruction *> GatherSeq;
524 /// A list of blocks that we are going to CSE.
525 SetVector<BasicBlock *> CSEBlocks;
527 /// Numbers instructions in different blocks.
528 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
530 /// Reduction operators.
533 // Analysis and block reference.
536 const DataLayout *DL;
537 TargetTransformInfo *TTI;
541 /// Instruction builder to construct the vectorized tree.
545 void BoUpSLP::buildTree(ArrayRef<Value *> Roots, ValueSet *Rdx) {
548 if (!getSameType(Roots))
550 buildTree_rec(Roots, 0);
552 // Collect the values that we need to extract from the tree.
553 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
554 TreeEntry *Entry = &VectorizableTree[EIdx];
557 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
558 Value *Scalar = Entry->Scalars[Lane];
560 // No need to handle users of gathered values.
561 if (Entry->NeedToGather)
564 for (User *U : Scalar->users()) {
565 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
567 // Skip in-tree scalars that become vectors.
568 if (ScalarToTreeEntry.count(U)) {
569 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
571 int Idx = ScalarToTreeEntry[U]; (void) Idx;
572 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
575 Instruction *UserInst = dyn_cast<Instruction>(U);
579 // Ignore uses that are part of the reduction.
580 if (Rdx && std::find(Rdx->begin(), Rdx->end(), UserInst) != Rdx->end())
583 DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<
584 Lane << " from " << *Scalar << ".\n");
585 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
592 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
593 bool SameTy = getSameType(VL); (void)SameTy;
594 assert(SameTy && "Invalid types!");
596 if (Depth == RecursionMaxDepth) {
597 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
598 newTreeEntry(VL, false);
602 // Don't handle vectors.
603 if (VL[0]->getType()->isVectorTy()) {
604 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
605 newTreeEntry(VL, false);
609 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
610 if (SI->getValueOperand()->getType()->isVectorTy()) {
611 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
612 newTreeEntry(VL, false);
616 // If all of the operands are identical or constant we have a simple solution.
617 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
618 !getSameOpcode(VL)) {
619 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
620 newTreeEntry(VL, false);
624 // We now know that this is a vector of instructions of the same type from
627 // Check if this is a duplicate of another entry.
628 if (ScalarToTreeEntry.count(VL[0])) {
629 int Idx = ScalarToTreeEntry[VL[0]];
630 TreeEntry *E = &VectorizableTree[Idx];
631 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
632 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
633 if (E->Scalars[i] != VL[i]) {
634 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
635 newTreeEntry(VL, false);
639 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
643 // Check that none of the instructions in the bundle are already in the tree.
644 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
645 if (ScalarToTreeEntry.count(VL[i])) {
646 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
647 ") is already in tree.\n");
648 newTreeEntry(VL, false);
653 // If any of the scalars appears in the table OR it is marked as a value that
654 // needs to stat scalar then we need to gather the scalars.
655 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
656 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
657 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
658 newTreeEntry(VL, false);
663 // Check that all of the users of the scalars that we want to vectorize are
665 Instruction *VL0 = cast<Instruction>(VL[0]);
666 int MyLastIndex = getLastIndex(VL);
667 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
669 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
670 Instruction *Scalar = cast<Instruction>(VL[i]);
671 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
672 for (User *U : Scalar->users()) {
673 DEBUG(dbgs() << "SLP: \tUser " << *U << ". \n");
674 Instruction *UI = dyn_cast<Instruction>(U);
676 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
677 newTreeEntry(VL, false);
681 // We don't care if the user is in a different basic block.
682 BasicBlock *UserBlock = UI->getParent();
683 if (UserBlock != BB) {
684 DEBUG(dbgs() << "SLP: User from a different basic block "
689 // If this is a PHINode within this basic block then we can place the
690 // extract wherever we want.
691 if (isa<PHINode>(*UI)) {
692 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *UI << ". \n");
696 // Check if this is a safe in-tree user.
697 if (ScalarToTreeEntry.count(UI)) {
698 int Idx = ScalarToTreeEntry[UI];
699 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
700 if (VecLocation <= MyLastIndex) {
701 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
702 newTreeEntry(VL, false);
705 DEBUG(dbgs() << "SLP: In-tree user (" << *UI << ") at #" <<
706 VecLocation << " vector value (" << *Scalar << ") at #"
707 << MyLastIndex << ".\n");
711 // This user is part of the reduction.
712 if (RdxOps && RdxOps->count(UI))
715 // Make sure that we can schedule this unknown user.
716 BlockNumbering &BN = BlocksNumbers[BB];
717 int UserIndex = BN.getIndex(UI);
718 if (UserIndex < MyLastIndex) {
720 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
722 newTreeEntry(VL, false);
728 // Check that every instructions appears once in this bundle.
729 for (unsigned i = 0, e = VL.size(); i < e; ++i)
730 for (unsigned j = i+1; j < e; ++j)
731 if (VL[i] == VL[j]) {
732 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
733 newTreeEntry(VL, false);
737 // Check that instructions in this bundle don't reference other instructions.
738 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
739 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
740 for (User *U : VL[i]->users()) {
741 for (unsigned j = 0; j < e; ++j) {
742 if (i != j && U == VL[j]) {
743 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n");
744 newTreeEntry(VL, false);
751 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
753 unsigned Opcode = getSameOpcode(VL);
755 // Check if it is safe to sink the loads or the stores.
756 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
757 Instruction *Last = getLastInstruction(VL);
759 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
762 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
764 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
765 << "\n because of " << *Barrier << ". Gathering.\n");
766 newTreeEntry(VL, false);
773 case Instruction::PHI: {
774 PHINode *PH = dyn_cast<PHINode>(VL0);
776 // Check for terminator values (e.g. invoke).
777 for (unsigned j = 0; j < VL.size(); ++j)
778 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
779 TerminatorInst *Term = dyn_cast<TerminatorInst>(
780 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
782 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
783 newTreeEntry(VL, false);
788 newTreeEntry(VL, true);
789 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
791 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
793 // Prepare the operand vector.
794 for (unsigned j = 0; j < VL.size(); ++j)
795 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
796 PH->getIncomingBlock(i)));
798 buildTree_rec(Operands, Depth + 1);
802 case Instruction::ExtractElement: {
803 bool Reuse = CanReuseExtract(VL);
805 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
807 newTreeEntry(VL, Reuse);
810 case Instruction::Load: {
811 // Check if the loads are consecutive or of we need to swizzle them.
812 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
813 LoadInst *L = cast<LoadInst>(VL[i]);
814 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
815 newTreeEntry(VL, false);
816 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
820 newTreeEntry(VL, true);
821 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
824 case Instruction::ZExt:
825 case Instruction::SExt:
826 case Instruction::FPToUI:
827 case Instruction::FPToSI:
828 case Instruction::FPExt:
829 case Instruction::PtrToInt:
830 case Instruction::IntToPtr:
831 case Instruction::SIToFP:
832 case Instruction::UIToFP:
833 case Instruction::Trunc:
834 case Instruction::FPTrunc:
835 case Instruction::BitCast: {
836 Type *SrcTy = VL0->getOperand(0)->getType();
837 for (unsigned i = 0; i < VL.size(); ++i) {
838 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
839 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
840 newTreeEntry(VL, false);
841 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
845 newTreeEntry(VL, true);
846 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
848 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
850 // Prepare the operand vector.
851 for (unsigned j = 0; j < VL.size(); ++j)
852 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
854 buildTree_rec(Operands, Depth+1);
858 case Instruction::ICmp:
859 case Instruction::FCmp: {
860 // Check that all of the compares have the same predicate.
861 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
862 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
863 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
864 CmpInst *Cmp = cast<CmpInst>(VL[i]);
865 if (Cmp->getPredicate() != P0 ||
866 Cmp->getOperand(0)->getType() != ComparedTy) {
867 newTreeEntry(VL, false);
868 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
873 newTreeEntry(VL, true);
874 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
876 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
878 // Prepare the operand vector.
879 for (unsigned j = 0; j < VL.size(); ++j)
880 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
882 buildTree_rec(Operands, Depth+1);
886 case Instruction::Select:
887 case Instruction::Add:
888 case Instruction::FAdd:
889 case Instruction::Sub:
890 case Instruction::FSub:
891 case Instruction::Mul:
892 case Instruction::FMul:
893 case Instruction::UDiv:
894 case Instruction::SDiv:
895 case Instruction::FDiv:
896 case Instruction::URem:
897 case Instruction::SRem:
898 case Instruction::FRem:
899 case Instruction::Shl:
900 case Instruction::LShr:
901 case Instruction::AShr:
902 case Instruction::And:
903 case Instruction::Or:
904 case Instruction::Xor: {
905 newTreeEntry(VL, true);
906 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
908 // Sort operands of the instructions so that each side is more likely to
909 // have the same opcode.
910 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
911 ValueList Left, Right;
912 reorderInputsAccordingToOpcode(VL, Left, Right);
913 buildTree_rec(Left, Depth + 1);
914 buildTree_rec(Right, Depth + 1);
918 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
920 // Prepare the operand vector.
921 for (unsigned j = 0; j < VL.size(); ++j)
922 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
924 buildTree_rec(Operands, Depth+1);
928 case Instruction::Store: {
929 // Check if the stores are consecutive or of we need to swizzle them.
930 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
931 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
932 newTreeEntry(VL, false);
933 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
937 newTreeEntry(VL, true);
938 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
941 for (unsigned j = 0; j < VL.size(); ++j)
942 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
944 // We can ignore these values because we are sinking them down.
945 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
946 buildTree_rec(Operands, Depth + 1);
950 newTreeEntry(VL, false);
951 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
956 int BoUpSLP::getEntryCost(TreeEntry *E) {
957 ArrayRef<Value*> VL = E->Scalars;
959 Type *ScalarTy = VL[0]->getType();
960 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
961 ScalarTy = SI->getValueOperand()->getType();
962 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
964 if (E->NeedToGather) {
968 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
970 return getGatherCost(E->Scalars);
973 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
975 Instruction *VL0 = cast<Instruction>(VL[0]);
976 unsigned Opcode = VL0->getOpcode();
978 case Instruction::PHI: {
981 case Instruction::ExtractElement: {
982 if (CanReuseExtract(VL))
984 return getGatherCost(VecTy);
986 case Instruction::ZExt:
987 case Instruction::SExt:
988 case Instruction::FPToUI:
989 case Instruction::FPToSI:
990 case Instruction::FPExt:
991 case Instruction::PtrToInt:
992 case Instruction::IntToPtr:
993 case Instruction::SIToFP:
994 case Instruction::UIToFP:
995 case Instruction::Trunc:
996 case Instruction::FPTrunc:
997 case Instruction::BitCast: {
998 Type *SrcTy = VL0->getOperand(0)->getType();
1000 // Calculate the cost of this instruction.
1001 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1002 VL0->getType(), SrcTy);
1004 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1005 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1006 return VecCost - ScalarCost;
1008 case Instruction::FCmp:
1009 case Instruction::ICmp:
1010 case Instruction::Select:
1011 case Instruction::Add:
1012 case Instruction::FAdd:
1013 case Instruction::Sub:
1014 case Instruction::FSub:
1015 case Instruction::Mul:
1016 case Instruction::FMul:
1017 case Instruction::UDiv:
1018 case Instruction::SDiv:
1019 case Instruction::FDiv:
1020 case Instruction::URem:
1021 case Instruction::SRem:
1022 case Instruction::FRem:
1023 case Instruction::Shl:
1024 case Instruction::LShr:
1025 case Instruction::AShr:
1026 case Instruction::And:
1027 case Instruction::Or:
1028 case Instruction::Xor: {
1029 // Calculate the cost of this instruction.
1032 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1033 Opcode == Instruction::Select) {
1034 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1035 ScalarCost = VecTy->getNumElements() *
1036 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1037 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1039 // Certain instructions can be cheaper to vectorize if they have a
1040 // constant second vector operand.
1041 TargetTransformInfo::OperandValueKind Op1VK =
1042 TargetTransformInfo::OK_AnyValue;
1043 TargetTransformInfo::OperandValueKind Op2VK =
1044 TargetTransformInfo::OK_UniformConstantValue;
1046 // If all operands are exactly the same ConstantInt then set the
1047 // operand kind to OK_UniformConstantValue.
1048 // If instead not all operands are constants, then set the operand kind
1049 // to OK_AnyValue. If all operands are constants but not the same,
1050 // then set the operand kind to OK_NonUniformConstantValue.
1051 ConstantInt *CInt = NULL;
1052 for (unsigned i = 0; i < VL.size(); ++i) {
1053 const Instruction *I = cast<Instruction>(VL[i]);
1054 if (!isa<ConstantInt>(I->getOperand(1))) {
1055 Op2VK = TargetTransformInfo::OK_AnyValue;
1059 CInt = cast<ConstantInt>(I->getOperand(1));
1062 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1063 CInt != cast<ConstantInt>(I->getOperand(1)))
1064 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1068 VecTy->getNumElements() *
1069 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1070 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1072 return VecCost - ScalarCost;
1074 case Instruction::Load: {
1075 // Cost of wide load - cost of scalar loads.
1076 int ScalarLdCost = VecTy->getNumElements() *
1077 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1078 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1079 return VecLdCost - ScalarLdCost;
1081 case Instruction::Store: {
1082 // We know that we can merge the stores. Calculate the cost.
1083 int ScalarStCost = VecTy->getNumElements() *
1084 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1085 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1086 return VecStCost - ScalarStCost;
1089 llvm_unreachable("Unknown instruction");
1093 bool BoUpSLP::isFullyVectorizableTinyTree() {
1094 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1095 VectorizableTree.size() << " is fully vectorizable .\n");
1097 // We only handle trees of height 2.
1098 if (VectorizableTree.size() != 2)
1101 // Handle splat stores.
1102 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1105 // Gathering cost would be too much for tiny trees.
1106 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1112 int BoUpSLP::getTreeCost() {
1114 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1115 VectorizableTree.size() << ".\n");
1117 // We only vectorize tiny trees if it is fully vectorizable.
1118 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1119 if (!VectorizableTree.size()) {
1120 assert(!ExternalUses.size() && "We should not have any external users");
1125 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1127 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1128 int C = getEntryCost(&VectorizableTree[i]);
1129 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1130 << *VectorizableTree[i].Scalars[0] << " .\n");
1134 SmallSet<Value *, 16> ExtractCostCalculated;
1135 int ExtractCost = 0;
1136 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1138 // We only add extract cost once for the same scalar.
1139 if (!ExtractCostCalculated.insert(I->Scalar))
1142 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1143 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1147 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1148 return Cost + ExtractCost;
1151 int BoUpSLP::getGatherCost(Type *Ty) {
1153 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1154 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1158 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1159 // Find the type of the operands in VL.
1160 Type *ScalarTy = VL[0]->getType();
1161 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1162 ScalarTy = SI->getValueOperand()->getType();
1163 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1164 // Find the cost of inserting/extracting values from the vector.
1165 return getGatherCost(VecTy);
1168 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1169 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1170 return AA->getLocation(SI);
1171 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1172 return AA->getLocation(LI);
1173 return AliasAnalysis::Location();
1176 Value *BoUpSLP::getPointerOperand(Value *I) {
1177 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1178 return LI->getPointerOperand();
1179 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1180 return SI->getPointerOperand();
1184 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1185 if (LoadInst *L = dyn_cast<LoadInst>(I))
1186 return L->getPointerAddressSpace();
1187 if (StoreInst *S = dyn_cast<StoreInst>(I))
1188 return S->getPointerAddressSpace();
1192 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1193 Value *PtrA = getPointerOperand(A);
1194 Value *PtrB = getPointerOperand(B);
1195 unsigned ASA = getAddressSpaceOperand(A);
1196 unsigned ASB = getAddressSpaceOperand(B);
1198 // Check that the address spaces match and that the pointers are valid.
1199 if (!PtrA || !PtrB || (ASA != ASB))
1202 // Make sure that A and B are different pointers of the same type.
1203 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1206 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1207 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1208 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1210 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1211 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1212 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1214 APInt OffsetDelta = OffsetB - OffsetA;
1216 // Check if they are based on the same pointer. That makes the offsets
1219 return OffsetDelta == Size;
1221 // Compute the necessary base pointer delta to have the necessary final delta
1222 // equal to the size.
1223 APInt BaseDelta = Size - OffsetDelta;
1225 // Otherwise compute the distance with SCEV between the base pointers.
1226 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1227 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1228 const SCEV *C = SE->getConstant(BaseDelta);
1229 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1230 return X == PtrSCEVB;
1233 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1234 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1235 BasicBlock::iterator I = Src, E = Dst;
1236 /// Scan all of the instruction from SRC to DST and check if
1237 /// the source may alias.
1238 for (++I; I != E; ++I) {
1239 // Ignore store instructions that are marked as 'ignore'.
1240 if (MemBarrierIgnoreList.count(I))
1242 if (Src->mayWriteToMemory()) /* Write */ {
1243 if (!I->mayReadOrWriteMemory())
1246 if (!I->mayWriteToMemory())
1249 AliasAnalysis::Location A = getLocation(&*I);
1250 AliasAnalysis::Location B = getLocation(Src);
1252 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1258 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1259 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1260 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1261 BlockNumbering &BN = BlocksNumbers[BB];
1263 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1264 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1265 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1269 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1270 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1271 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1272 BlockNumbering &BN = BlocksNumbers[BB];
1274 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1275 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1276 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1277 Instruction *I = BN.getInstruction(MaxIdx);
1278 assert(I && "bad location");
1282 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1283 Instruction *VL0 = cast<Instruction>(VL[0]);
1284 Instruction *LastInst = getLastInstruction(VL);
1285 BasicBlock::iterator NextInst = LastInst;
1287 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1288 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1291 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1292 Value *Vec = UndefValue::get(Ty);
1293 // Generate the 'InsertElement' instruction.
1294 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1295 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1296 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1297 GatherSeq.insert(Insrt);
1298 CSEBlocks.insert(Insrt->getParent());
1300 // Add to our 'need-to-extract' list.
1301 if (ScalarToTreeEntry.count(VL[i])) {
1302 int Idx = ScalarToTreeEntry[VL[i]];
1303 TreeEntry *E = &VectorizableTree[Idx];
1304 // Find which lane we need to extract.
1306 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1307 // Is this the lane of the scalar that we are looking for ?
1308 if (E->Scalars[Lane] == VL[i]) {
1313 assert(FoundLane >= 0 && "Could not find the correct lane");
1314 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1322 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1323 SmallDenseMap<Value*, int>::const_iterator Entry
1324 = ScalarToTreeEntry.find(VL[0]);
1325 if (Entry != ScalarToTreeEntry.end()) {
1326 int Idx = Entry->second;
1327 const TreeEntry *En = &VectorizableTree[Idx];
1328 if (En->isSame(VL) && En->VectorizedValue)
1329 return En->VectorizedValue;
1334 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1335 if (ScalarToTreeEntry.count(VL[0])) {
1336 int Idx = ScalarToTreeEntry[VL[0]];
1337 TreeEntry *E = &VectorizableTree[Idx];
1339 return vectorizeTree(E);
1342 Type *ScalarTy = VL[0]->getType();
1343 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1344 ScalarTy = SI->getValueOperand()->getType();
1345 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1347 return Gather(VL, VecTy);
1350 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1351 IRBuilder<>::InsertPointGuard Guard(Builder);
1353 if (E->VectorizedValue) {
1354 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1355 return E->VectorizedValue;
1358 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1359 Type *ScalarTy = VL0->getType();
1360 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1361 ScalarTy = SI->getValueOperand()->getType();
1362 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1364 if (E->NeedToGather) {
1365 setInsertPointAfterBundle(E->Scalars);
1366 return Gather(E->Scalars, VecTy);
1369 unsigned Opcode = VL0->getOpcode();
1370 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1373 case Instruction::PHI: {
1374 PHINode *PH = dyn_cast<PHINode>(VL0);
1375 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1376 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1377 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1378 E->VectorizedValue = NewPhi;
1380 // PHINodes may have multiple entries from the same block. We want to
1381 // visit every block once.
1382 SmallSet<BasicBlock*, 4> VisitedBBs;
1384 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1386 BasicBlock *IBB = PH->getIncomingBlock(i);
1388 if (!VisitedBBs.insert(IBB)) {
1389 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1393 // Prepare the operand vector.
1394 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1395 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1396 getIncomingValueForBlock(IBB));
1398 Builder.SetInsertPoint(IBB->getTerminator());
1399 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1400 Value *Vec = vectorizeTree(Operands);
1401 NewPhi->addIncoming(Vec, IBB);
1404 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1405 "Invalid number of incoming values");
1409 case Instruction::ExtractElement: {
1410 if (CanReuseExtract(E->Scalars)) {
1411 Value *V = VL0->getOperand(0);
1412 E->VectorizedValue = V;
1415 return Gather(E->Scalars, VecTy);
1417 case Instruction::ZExt:
1418 case Instruction::SExt:
1419 case Instruction::FPToUI:
1420 case Instruction::FPToSI:
1421 case Instruction::FPExt:
1422 case Instruction::PtrToInt:
1423 case Instruction::IntToPtr:
1424 case Instruction::SIToFP:
1425 case Instruction::UIToFP:
1426 case Instruction::Trunc:
1427 case Instruction::FPTrunc:
1428 case Instruction::BitCast: {
1430 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1431 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1433 setInsertPointAfterBundle(E->Scalars);
1435 Value *InVec = vectorizeTree(INVL);
1437 if (Value *V = alreadyVectorized(E->Scalars))
1440 CastInst *CI = dyn_cast<CastInst>(VL0);
1441 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1442 E->VectorizedValue = V;
1445 case Instruction::FCmp:
1446 case Instruction::ICmp: {
1447 ValueList LHSV, RHSV;
1448 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1449 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1450 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1453 setInsertPointAfterBundle(E->Scalars);
1455 Value *L = vectorizeTree(LHSV);
1456 Value *R = vectorizeTree(RHSV);
1458 if (Value *V = alreadyVectorized(E->Scalars))
1461 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1463 if (Opcode == Instruction::FCmp)
1464 V = Builder.CreateFCmp(P0, L, R);
1466 V = Builder.CreateICmp(P0, L, R);
1468 E->VectorizedValue = V;
1471 case Instruction::Select: {
1472 ValueList TrueVec, FalseVec, CondVec;
1473 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1474 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1475 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1476 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1479 setInsertPointAfterBundle(E->Scalars);
1481 Value *Cond = vectorizeTree(CondVec);
1482 Value *True = vectorizeTree(TrueVec);
1483 Value *False = vectorizeTree(FalseVec);
1485 if (Value *V = alreadyVectorized(E->Scalars))
1488 Value *V = Builder.CreateSelect(Cond, True, False);
1489 E->VectorizedValue = V;
1492 case Instruction::Add:
1493 case Instruction::FAdd:
1494 case Instruction::Sub:
1495 case Instruction::FSub:
1496 case Instruction::Mul:
1497 case Instruction::FMul:
1498 case Instruction::UDiv:
1499 case Instruction::SDiv:
1500 case Instruction::FDiv:
1501 case Instruction::URem:
1502 case Instruction::SRem:
1503 case Instruction::FRem:
1504 case Instruction::Shl:
1505 case Instruction::LShr:
1506 case Instruction::AShr:
1507 case Instruction::And:
1508 case Instruction::Or:
1509 case Instruction::Xor: {
1510 ValueList LHSVL, RHSVL;
1511 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1512 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1514 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1515 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1516 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1519 setInsertPointAfterBundle(E->Scalars);
1521 Value *LHS = vectorizeTree(LHSVL);
1522 Value *RHS = vectorizeTree(RHSVL);
1524 if (LHS == RHS && isa<Instruction>(LHS)) {
1525 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1528 if (Value *V = alreadyVectorized(E->Scalars))
1531 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1532 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1533 E->VectorizedValue = V;
1535 if (Instruction *I = dyn_cast<Instruction>(V))
1536 return propagateMetadata(I, E->Scalars);
1540 case Instruction::Load: {
1541 // Loads are inserted at the head of the tree because we don't want to
1542 // sink them all the way down past store instructions.
1543 setInsertPointAfterBundle(E->Scalars);
1545 LoadInst *LI = cast<LoadInst>(VL0);
1546 unsigned AS = LI->getPointerAddressSpace();
1548 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1549 VecTy->getPointerTo(AS));
1550 unsigned Alignment = LI->getAlignment();
1551 LI = Builder.CreateLoad(VecPtr);
1552 LI->setAlignment(Alignment);
1553 E->VectorizedValue = LI;
1554 return propagateMetadata(LI, E->Scalars);
1556 case Instruction::Store: {
1557 StoreInst *SI = cast<StoreInst>(VL0);
1558 unsigned Alignment = SI->getAlignment();
1559 unsigned AS = SI->getPointerAddressSpace();
1562 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1563 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1565 setInsertPointAfterBundle(E->Scalars);
1567 Value *VecValue = vectorizeTree(ValueOp);
1568 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1569 VecTy->getPointerTo(AS));
1570 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1571 S->setAlignment(Alignment);
1572 E->VectorizedValue = S;
1573 return propagateMetadata(S, E->Scalars);
1576 llvm_unreachable("unknown inst");
1581 Value *BoUpSLP::vectorizeTree() {
1582 Builder.SetInsertPoint(F->getEntryBlock().begin());
1583 vectorizeTree(&VectorizableTree[0]);
1585 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1587 // Extract all of the elements with the external uses.
1588 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1590 Value *Scalar = it->Scalar;
1591 llvm::User *User = it->User;
1593 // Skip users that we already RAUW. This happens when one instruction
1594 // has multiple uses of the same value.
1595 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
1598 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1600 int Idx = ScalarToTreeEntry[Scalar];
1601 TreeEntry *E = &VectorizableTree[Idx];
1602 assert(!E->NeedToGather && "Extracting from a gather list");
1604 Value *Vec = E->VectorizedValue;
1605 assert(Vec && "Can't find vectorizable value");
1607 Value *Lane = Builder.getInt32(it->Lane);
1608 // Generate extracts for out-of-tree users.
1609 // Find the insertion point for the extractelement lane.
1610 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1611 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt());
1612 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1613 CSEBlocks.insert(PN->getParent());
1614 User->replaceUsesOfWith(Scalar, Ex);
1615 } else if (isa<Instruction>(Vec)){
1616 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1617 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1618 if (PH->getIncomingValue(i) == Scalar) {
1619 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1620 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1621 CSEBlocks.insert(PH->getIncomingBlock(i));
1622 PH->setOperand(i, Ex);
1626 Builder.SetInsertPoint(cast<Instruction>(User));
1627 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1628 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1629 User->replaceUsesOfWith(Scalar, Ex);
1632 Builder.SetInsertPoint(F->getEntryBlock().begin());
1633 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1634 CSEBlocks.insert(&F->getEntryBlock());
1635 User->replaceUsesOfWith(Scalar, Ex);
1638 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1641 // For each vectorized value:
1642 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1643 TreeEntry *Entry = &VectorizableTree[EIdx];
1646 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1647 Value *Scalar = Entry->Scalars[Lane];
1649 // No need to handle users of gathered values.
1650 if (Entry->NeedToGather)
1653 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1655 Type *Ty = Scalar->getType();
1656 if (!Ty->isVoidTy()) {
1657 for (User *U : Scalar->users()) {
1658 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
1660 assert((ScalarToTreeEntry.count(U) ||
1661 // It is legal to replace the reduction users by undef.
1662 (RdxOps && RdxOps->count(U))) &&
1663 "Replacing out-of-tree value with undef");
1665 Value *Undef = UndefValue::get(Ty);
1666 Scalar->replaceAllUsesWith(Undef);
1668 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1669 cast<Instruction>(Scalar)->eraseFromParent();
1673 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1674 BlocksNumbers[it].forget();
1676 Builder.ClearInsertionPoint();
1678 return VectorizableTree[0].VectorizedValue;
1681 void BoUpSLP::optimizeGatherSequence() {
1682 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1683 << " gather sequences instructions.\n");
1684 // LICM InsertElementInst sequences.
1685 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1686 e = GatherSeq.end(); it != e; ++it) {
1687 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1692 // Check if this block is inside a loop.
1693 Loop *L = LI->getLoopFor(Insert->getParent());
1697 // Check if it has a preheader.
1698 BasicBlock *PreHeader = L->getLoopPreheader();
1702 // If the vector or the element that we insert into it are
1703 // instructions that are defined in this basic block then we can't
1704 // hoist this instruction.
1705 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1706 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1707 if (CurrVec && L->contains(CurrVec))
1709 if (NewElem && L->contains(NewElem))
1712 // We can hoist this instruction. Move it to the pre-header.
1713 Insert->moveBefore(PreHeader->getTerminator());
1716 // Sort blocks by domination. This ensures we visit a block after all blocks
1717 // dominating it are visited.
1718 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1719 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1720 [this](const BasicBlock *A, const BasicBlock *B) {
1721 return DT->properlyDominates(A, B);
1724 // Perform O(N^2) search over the gather sequences and merge identical
1725 // instructions. TODO: We can further optimize this scan if we split the
1726 // instructions into different buckets based on the insert lane.
1727 SmallVector<Instruction *, 16> Visited;
1728 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1729 E = CSEWorkList.end();
1731 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1732 "Worklist not sorted properly!");
1733 BasicBlock *BB = *I;
1734 // For all instructions in blocks containing gather sequences:
1735 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1736 Instruction *In = it++;
1737 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1740 // Check if we can replace this instruction with any of the
1741 // visited instructions.
1742 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1745 if (In->isIdenticalTo(*v) &&
1746 DT->dominates((*v)->getParent(), In->getParent())) {
1747 In->replaceAllUsesWith(*v);
1748 In->eraseFromParent();
1754 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1755 Visited.push_back(In);
1763 /// The SLPVectorizer Pass.
1764 struct SLPVectorizer : public FunctionPass {
1765 typedef SmallVector<StoreInst *, 8> StoreList;
1766 typedef MapVector<Value *, StoreList> StoreListMap;
1768 /// Pass identification, replacement for typeid
1771 explicit SLPVectorizer() : FunctionPass(ID) {
1772 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1775 ScalarEvolution *SE;
1776 const DataLayout *DL;
1777 TargetTransformInfo *TTI;
1782 bool runOnFunction(Function &F) override {
1783 if (skipOptnoneFunction(F))
1786 SE = &getAnalysis<ScalarEvolution>();
1787 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1788 DL = DLP ? &DLP->getDataLayout() : 0;
1789 TTI = &getAnalysis<TargetTransformInfo>();
1790 AA = &getAnalysis<AliasAnalysis>();
1791 LI = &getAnalysis<LoopInfo>();
1792 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1795 bool Changed = false;
1797 // If the target claims to have no vector registers don't attempt
1799 if (!TTI->getNumberOfRegisters(true))
1802 // Must have DataLayout. We can't require it because some tests run w/o
1807 // Don't vectorize when the attribute NoImplicitFloat is used.
1808 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1811 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1813 // Use the bottom up slp vectorizer to construct chains that start with
1814 // he store instructions.
1815 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1817 // Scan the blocks in the function in post order.
1818 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1819 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1820 BasicBlock *BB = *it;
1822 // Vectorize trees that end at stores.
1823 if (unsigned count = collectStores(BB, R)) {
1825 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1826 Changed |= vectorizeStoreChains(R);
1829 // Vectorize trees that end at reductions.
1830 Changed |= vectorizeChainsInBlock(BB, R);
1834 R.optimizeGatherSequence();
1835 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1836 DEBUG(verifyFunction(F));
1841 void getAnalysisUsage(AnalysisUsage &AU) const override {
1842 FunctionPass::getAnalysisUsage(AU);
1843 AU.addRequired<ScalarEvolution>();
1844 AU.addRequired<AliasAnalysis>();
1845 AU.addRequired<TargetTransformInfo>();
1846 AU.addRequired<LoopInfo>();
1847 AU.addRequired<DominatorTreeWrapperPass>();
1848 AU.addPreserved<LoopInfo>();
1849 AU.addPreserved<DominatorTreeWrapperPass>();
1850 AU.setPreservesCFG();
1855 /// \brief Collect memory references and sort them according to their base
1856 /// object. We sort the stores to their base objects to reduce the cost of the
1857 /// quadratic search on the stores. TODO: We can further reduce this cost
1858 /// if we flush the chain creation every time we run into a memory barrier.
1859 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1861 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1862 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1864 /// \brief Try to vectorize a list of operands.
1865 /// \returns true if a value was vectorized.
1866 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1868 /// \brief Try to vectorize a chain that may start at the operands of \V;
1869 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1871 /// \brief Vectorize the stores that were collected in StoreRefs.
1872 bool vectorizeStoreChains(BoUpSLP &R);
1874 /// \brief Scan the basic block and look for patterns that are likely to start
1875 /// a vectorization chain.
1876 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1878 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1881 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1884 StoreListMap StoreRefs;
1887 /// \brief Check that the Values in the slice in VL array are still existent in
1888 /// the WeakVH array.
1889 /// Vectorization of part of the VL array may cause later values in the VL array
1890 /// to become invalid. We track when this has happened in the WeakVH array.
1891 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1892 SmallVectorImpl<WeakVH> &VH,
1893 unsigned SliceBegin,
1894 unsigned SliceSize) {
1895 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1902 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1903 int CostThreshold, BoUpSLP &R) {
1904 unsigned ChainLen = Chain.size();
1905 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1907 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1908 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1909 unsigned VF = MinVecRegSize / Sz;
1911 if (!isPowerOf2_32(Sz) || VF < 2)
1914 // Keep track of values that were delete by vectorizing in the loop below.
1915 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
1917 bool Changed = false;
1918 // Look for profitable vectorizable trees at all offsets, starting at zero.
1919 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1923 // Check that a previous iteration of this loop did not delete the Value.
1924 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
1927 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1929 ArrayRef<Value *> Operands = Chain.slice(i, VF);
1931 R.buildTree(Operands);
1933 int Cost = R.getTreeCost();
1935 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
1936 if (Cost < CostThreshold) {
1937 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
1940 // Move to the next bundle.
1949 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
1950 int costThreshold, BoUpSLP &R) {
1951 SetVector<Value *> Heads, Tails;
1952 SmallDenseMap<Value *, Value *> ConsecutiveChain;
1954 // We may run into multiple chains that merge into a single chain. We mark the
1955 // stores that we vectorized so that we don't visit the same store twice.
1956 BoUpSLP::ValueSet VectorizedStores;
1957 bool Changed = false;
1959 // Do a quadratic search on all of the given stores and find
1960 // all of the pairs of stores that follow each other.
1961 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
1962 for (unsigned j = 0; j < e; ++j) {
1966 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
1967 Tails.insert(Stores[j]);
1968 Heads.insert(Stores[i]);
1969 ConsecutiveChain[Stores[i]] = Stores[j];
1974 // For stores that start but don't end a link in the chain:
1975 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
1977 if (Tails.count(*it))
1980 // We found a store instr that starts a chain. Now follow the chain and try
1982 BoUpSLP::ValueList Operands;
1984 // Collect the chain into a list.
1985 while (Tails.count(I) || Heads.count(I)) {
1986 if (VectorizedStores.count(I))
1988 Operands.push_back(I);
1989 // Move to the next value in the chain.
1990 I = ConsecutiveChain[I];
1993 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
1995 // Mark the vectorized stores so that we don't vectorize them again.
1997 VectorizedStores.insert(Operands.begin(), Operands.end());
1998 Changed |= Vectorized;
2005 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2008 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2009 StoreInst *SI = dyn_cast<StoreInst>(it);
2013 // Don't touch volatile stores.
2014 if (!SI->isSimple())
2017 // Check that the pointer points to scalars.
2018 Type *Ty = SI->getValueOperand()->getType();
2019 if (Ty->isAggregateType() || Ty->isVectorTy())
2022 // Find the base pointer.
2023 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2025 // Save the store locations.
2026 StoreRefs[Ptr].push_back(SI);
2032 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2035 Value *VL[] = { A, B };
2036 return tryToVectorizeList(VL, R);
2039 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
2043 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2045 // Check that all of the parts are scalar instructions of the same type.
2046 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2050 unsigned Opcode0 = I0->getOpcode();
2052 Type *Ty0 = I0->getType();
2053 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2054 unsigned VF = MinVecRegSize / Sz;
2056 for (int i = 0, e = VL.size(); i < e; ++i) {
2057 Type *Ty = VL[i]->getType();
2058 if (Ty->isAggregateType() || Ty->isVectorTy())
2060 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2061 if (!Inst || Inst->getOpcode() != Opcode0)
2065 bool Changed = false;
2067 // Keep track of values that were delete by vectorizing in the loop below.
2068 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2070 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2071 unsigned OpsWidth = 0;
2078 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2081 // Check that a previous iteration of this loop did not delete the Value.
2082 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2085 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2087 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2090 int Cost = R.getTreeCost();
2092 if (Cost < -SLPCostThreshold) {
2093 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n");
2096 // Move to the next bundle.
2105 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2109 // Try to vectorize V.
2110 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2113 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2114 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2116 if (B && B->hasOneUse()) {
2117 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2118 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2119 if (tryToVectorizePair(A, B0, R)) {
2123 if (tryToVectorizePair(A, B1, R)) {
2130 if (A && A->hasOneUse()) {
2131 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2132 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2133 if (tryToVectorizePair(A0, B, R)) {
2137 if (tryToVectorizePair(A1, B, R)) {
2145 /// \brief Generate a shuffle mask to be used in a reduction tree.
2147 /// \param VecLen The length of the vector to be reduced.
2148 /// \param NumEltsToRdx The number of elements that should be reduced in the
2150 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2151 /// reduction. A pairwise reduction will generate a mask of
2152 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2153 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2154 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2155 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2156 bool IsPairwise, bool IsLeft,
2157 IRBuilder<> &Builder) {
2158 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2160 SmallVector<Constant *, 32> ShuffleMask(
2161 VecLen, UndefValue::get(Builder.getInt32Ty()));
2164 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2165 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2166 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2168 // Move the upper half of the vector to the lower half.
2169 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2170 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2172 return ConstantVector::get(ShuffleMask);
2176 /// Model horizontal reductions.
2178 /// A horizontal reduction is a tree of reduction operations (currently add and
2179 /// fadd) that has operations that can be put into a vector as its leaf.
2180 /// For example, this tree:
2187 /// This tree has "mul" as its reduced values and "+" as its reduction
2188 /// operations. A reduction might be feeding into a store or a binary operation
2203 class HorizontalReduction {
2204 SmallPtrSet<Value *, 16> ReductionOps;
2205 SmallVector<Value *, 32> ReducedVals;
2207 BinaryOperator *ReductionRoot;
2208 PHINode *ReductionPHI;
2210 /// The opcode of the reduction.
2211 unsigned ReductionOpcode;
2212 /// The opcode of the values we perform a reduction on.
2213 unsigned ReducedValueOpcode;
2214 /// The width of one full horizontal reduction operation.
2215 unsigned ReduxWidth;
2216 /// Should we model this reduction as a pairwise reduction tree or a tree that
2217 /// splits the vector in halves and adds those halves.
2218 bool IsPairwiseReduction;
2221 HorizontalReduction()
2222 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0),
2223 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2225 /// \brief Try to find a reduction tree.
2226 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2227 const DataLayout *DL) {
2229 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2230 "Thi phi needs to use the binary operator");
2232 // We could have a initial reductions that is not an add.
2233 // r *= v1 + v2 + v3 + v4
2234 // In such a case start looking for a tree rooted in the first '+'.
2236 if (B->getOperand(0) == Phi) {
2238 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2239 } else if (B->getOperand(1) == Phi) {
2241 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2248 Type *Ty = B->getType();
2249 if (Ty->isVectorTy())
2252 ReductionOpcode = B->getOpcode();
2253 ReducedValueOpcode = 0;
2254 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2261 // We currently only support adds.
2262 if (ReductionOpcode != Instruction::Add &&
2263 ReductionOpcode != Instruction::FAdd)
2266 // Post order traverse the reduction tree starting at B. We only handle true
2267 // trees containing only binary operators.
2268 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2269 Stack.push_back(std::make_pair(B, 0));
2270 while (!Stack.empty()) {
2271 BinaryOperator *TreeN = Stack.back().first;
2272 unsigned EdgeToVist = Stack.back().second++;
2273 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2275 // Only handle trees in the current basic block.
2276 if (TreeN->getParent() != B->getParent())
2279 // Each tree node needs to have one user except for the ultimate
2281 if (!TreeN->hasOneUse() && TreeN != B)
2285 if (EdgeToVist == 2 || IsReducedValue) {
2286 if (IsReducedValue) {
2287 // Make sure that the opcodes of the operations that we are going to
2289 if (!ReducedValueOpcode)
2290 ReducedValueOpcode = TreeN->getOpcode();
2291 else if (ReducedValueOpcode != TreeN->getOpcode())
2293 ReducedVals.push_back(TreeN);
2295 // We need to be able to reassociate the adds.
2296 if (!TreeN->isAssociative())
2298 ReductionOps.insert(TreeN);
2305 // Visit left or right.
2306 Value *NextV = TreeN->getOperand(EdgeToVist);
2307 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2309 Stack.push_back(std::make_pair(Next, 0));
2310 else if (NextV != Phi)
2316 /// \brief Attempt to vectorize the tree found by
2317 /// matchAssociativeReduction.
2318 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2319 if (ReducedVals.empty())
2322 unsigned NumReducedVals = ReducedVals.size();
2323 if (NumReducedVals < ReduxWidth)
2326 Value *VectorizedTree = 0;
2327 IRBuilder<> Builder(ReductionRoot);
2328 FastMathFlags Unsafe;
2329 Unsafe.setUnsafeAlgebra();
2330 Builder.SetFastMathFlags(Unsafe);
2333 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2334 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2335 V.buildTree(ValsToReduce, &ReductionOps);
2338 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2339 if (Cost >= -SLPCostThreshold)
2342 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2345 // Vectorize a tree.
2346 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2347 Value *VectorizedRoot = V.vectorizeTree();
2349 // Emit a reduction.
2350 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2351 if (VectorizedTree) {
2352 Builder.SetCurrentDebugLocation(Loc);
2353 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2354 ReducedSubTree, "bin.rdx");
2356 VectorizedTree = ReducedSubTree;
2359 if (VectorizedTree) {
2360 // Finish the reduction.
2361 for (; i < NumReducedVals; ++i) {
2362 Builder.SetCurrentDebugLocation(
2363 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2364 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2369 assert(ReductionRoot != NULL && "Need a reduction operation");
2370 ReductionRoot->setOperand(0, VectorizedTree);
2371 ReductionRoot->setOperand(1, ReductionPHI);
2373 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2375 return VectorizedTree != 0;
2380 /// \brief Calcuate the cost of a reduction.
2381 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2382 Type *ScalarTy = FirstReducedVal->getType();
2383 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2385 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2386 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2388 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2389 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2391 int ScalarReduxCost =
2392 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2394 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2395 << " for reduction that starts with " << *FirstReducedVal
2397 << (IsPairwiseReduction ? "pairwise" : "splitting")
2398 << " reduction)\n");
2400 return VecReduxCost - ScalarReduxCost;
2403 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2404 Value *R, const Twine &Name = "") {
2405 if (Opcode == Instruction::FAdd)
2406 return Builder.CreateFAdd(L, R, Name);
2407 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2410 /// \brief Emit a horizontal reduction of the vectorized value.
2411 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2412 assert(VectorizedValue && "Need to have a vectorized tree node");
2413 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2414 assert(isPowerOf2_32(ReduxWidth) &&
2415 "We only handle power-of-two reductions for now");
2417 Value *TmpVec = ValToReduce;
2418 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2419 if (IsPairwiseReduction) {
2421 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2423 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2425 Value *LeftShuf = Builder.CreateShuffleVector(
2426 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2427 Value *RightShuf = Builder.CreateShuffleVector(
2428 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2430 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2434 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2435 Value *Shuf = Builder.CreateShuffleVector(
2436 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2437 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2441 // The result is in the first element of the vector.
2442 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2446 /// \brief Recognize construction of vectors like
2447 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2448 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2449 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2450 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2452 /// Returns true if it matches
2454 static bool findBuildVector(InsertElementInst *IE,
2455 SmallVectorImpl<Value *> &Ops) {
2456 if (!isa<UndefValue>(IE->getOperand(0)))
2460 Ops.push_back(IE->getOperand(1));
2462 if (IE->use_empty())
2465 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2469 // If this isn't the final use, make sure the next insertelement is the only
2470 // use. It's OK if the final constructed vector is used multiple times
2471 if (!IE->hasOneUse())
2480 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2481 return V->getType() < V2->getType();
2484 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2485 bool Changed = false;
2486 SmallVector<Value *, 4> Incoming;
2487 SmallSet<Value *, 16> VisitedInstrs;
2489 bool HaveVectorizedPhiNodes = true;
2490 while (HaveVectorizedPhiNodes) {
2491 HaveVectorizedPhiNodes = false;
2493 // Collect the incoming values from the PHIs.
2495 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2497 PHINode *P = dyn_cast<PHINode>(instr);
2501 if (!VisitedInstrs.count(P))
2502 Incoming.push_back(P);
2506 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2508 // Try to vectorize elements base on their type.
2509 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2513 // Look for the next elements with the same type.
2514 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2515 while (SameTypeIt != E &&
2516 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2517 VisitedInstrs.insert(*SameTypeIt);
2521 // Try to vectorize them.
2522 unsigned NumElts = (SameTypeIt - IncIt);
2523 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2525 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2526 // Success start over because instructions might have been changed.
2527 HaveVectorizedPhiNodes = true;
2532 // Start over at the next instruction of a different type (or the end).
2537 VisitedInstrs.clear();
2539 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2540 // We may go through BB multiple times so skip the one we have checked.
2541 if (!VisitedInstrs.insert(it))
2544 if (isa<DbgInfoIntrinsic>(it))
2547 // Try to vectorize reductions that use PHINodes.
2548 if (PHINode *P = dyn_cast<PHINode>(it)) {
2549 // Check that the PHI is a reduction PHI.
2550 if (P->getNumIncomingValues() != 2)
2553 (P->getIncomingBlock(0) == BB
2554 ? (P->getIncomingValue(0))
2555 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
2556 // Check if this is a Binary Operator.
2557 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2561 // Try to match and vectorize a horizontal reduction.
2562 HorizontalReduction HorRdx;
2563 if (ShouldVectorizeHor &&
2564 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2565 HorRdx.tryToReduce(R, TTI)) {
2572 Value *Inst = BI->getOperand(0);
2574 Inst = BI->getOperand(1);
2576 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2577 // We would like to start over since some instructions are deleted
2578 // and the iterator may become invalid value.
2588 // Try to vectorize horizontal reductions feeding into a store.
2589 if (ShouldStartVectorizeHorAtStore)
2590 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2591 if (BinaryOperator *BinOp =
2592 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2593 HorizontalReduction HorRdx;
2594 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) &&
2595 HorRdx.tryToReduce(R, TTI)) ||
2596 tryToVectorize(BinOp, R))) {
2604 // Try to vectorize trees that start at compare instructions.
2605 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2606 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2608 // We would like to start over since some instructions are deleted
2609 // and the iterator may become invalid value.
2615 for (int i = 0; i < 2; ++i) {
2616 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2617 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2619 // We would like to start over since some instructions are deleted
2620 // and the iterator may become invalid value.
2629 // Try to vectorize trees that start at insertelement instructions.
2630 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2631 SmallVector<Value *, 8> Ops;
2632 if (!findBuildVector(IE, Ops))
2635 if (tryToVectorizeList(Ops, R)) {
2648 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2649 bool Changed = false;
2650 // Attempt to sort and vectorize each of the store-groups.
2651 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2653 if (it->second.size() < 2)
2656 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2657 << it->second.size() << ".\n");
2659 // Process the stores in chunks of 16.
2660 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2661 unsigned Len = std::min<unsigned>(CE - CI, 16);
2662 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2663 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2669 } // end anonymous namespace
2671 char SLPVectorizer::ID = 0;
2672 static const char lv_name[] = "SLP Vectorizer";
2673 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2674 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2675 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2676 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2677 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2678 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2681 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }