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 (Value::use_iterator User = Scalar->use_begin(),
565 UE = Scalar->use_end(); User != UE; ++User) {
566 DEBUG(dbgs() << "SLP: Checking user:" << **User << ".\n");
568 // Skip in-tree scalars that become vectors.
569 if (ScalarToTreeEntry.count(*User)) {
570 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
572 int Idx = ScalarToTreeEntry[*User]; (void) Idx;
573 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
576 Instruction *UserInst = dyn_cast<Instruction>(*User);
580 // Ignore uses that are part of the reduction.
581 if (Rdx && std::find(Rdx->begin(), Rdx->end(), UserInst) != Rdx->end())
584 DEBUG(dbgs() << "SLP: Need to extract:" << **User << " from lane " <<
585 Lane << " from " << *Scalar << ".\n");
586 ExternalUses.push_back(ExternalUser(Scalar, *User, Lane));
593 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
594 bool SameTy = getSameType(VL); (void)SameTy;
595 assert(SameTy && "Invalid types!");
597 if (Depth == RecursionMaxDepth) {
598 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
599 newTreeEntry(VL, false);
603 // Don't handle vectors.
604 if (VL[0]->getType()->isVectorTy()) {
605 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
606 newTreeEntry(VL, false);
610 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
611 if (SI->getValueOperand()->getType()->isVectorTy()) {
612 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
613 newTreeEntry(VL, false);
617 // If all of the operands are identical or constant we have a simple solution.
618 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
619 !getSameOpcode(VL)) {
620 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
621 newTreeEntry(VL, false);
625 // We now know that this is a vector of instructions of the same type from
628 // Check if this is a duplicate of another entry.
629 if (ScalarToTreeEntry.count(VL[0])) {
630 int Idx = ScalarToTreeEntry[VL[0]];
631 TreeEntry *E = &VectorizableTree[Idx];
632 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
633 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
634 if (E->Scalars[i] != VL[i]) {
635 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
636 newTreeEntry(VL, false);
640 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
644 // Check that none of the instructions in the bundle are already in the tree.
645 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
646 if (ScalarToTreeEntry.count(VL[i])) {
647 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
648 ") is already in tree.\n");
649 newTreeEntry(VL, false);
654 // If any of the scalars appears in the table OR it is marked as a value that
655 // needs to stat scalar then we need to gather the scalars.
656 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
657 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
658 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
659 newTreeEntry(VL, false);
664 // Check that all of the users of the scalars that we want to vectorize are
666 Instruction *VL0 = cast<Instruction>(VL[0]);
667 int MyLastIndex = getLastIndex(VL);
668 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
670 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
671 Instruction *Scalar = cast<Instruction>(VL[i]);
672 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
673 for (Value::use_iterator U = Scalar->use_begin(), UE = Scalar->use_end();
675 DEBUG(dbgs() << "SLP: \tUser " << **U << ". \n");
676 Instruction *User = dyn_cast<Instruction>(*U);
678 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
679 newTreeEntry(VL, false);
683 // We don't care if the user is in a different basic block.
684 BasicBlock *UserBlock = User->getParent();
685 if (UserBlock != BB) {
686 DEBUG(dbgs() << "SLP: User from a different basic block "
691 // If this is a PHINode within this basic block then we can place the
692 // extract wherever we want.
693 if (isa<PHINode>(*User)) {
694 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *User << ". \n");
698 // Check if this is a safe in-tree user.
699 if (ScalarToTreeEntry.count(User)) {
700 int Idx = ScalarToTreeEntry[User];
701 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
702 if (VecLocation <= MyLastIndex) {
703 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
704 newTreeEntry(VL, false);
707 DEBUG(dbgs() << "SLP: In-tree user (" << *User << ") at #" <<
708 VecLocation << " vector value (" << *Scalar << ") at #"
709 << MyLastIndex << ".\n");
713 // This user is part of the reduction.
714 if (RdxOps && RdxOps->count(User))
717 // Make sure that we can schedule this unknown user.
718 BlockNumbering &BN = BlocksNumbers[BB];
719 int UserIndex = BN.getIndex(User);
720 if (UserIndex < MyLastIndex) {
722 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
724 newTreeEntry(VL, false);
730 // Check that every instructions appears once in this bundle.
731 for (unsigned i = 0, e = VL.size(); i < e; ++i)
732 for (unsigned j = i+1; j < e; ++j)
733 if (VL[i] == VL[j]) {
734 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
735 newTreeEntry(VL, false);
739 // Check that instructions in this bundle don't reference other instructions.
740 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
741 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
742 for (Value::use_iterator U = VL[i]->use_begin(), UE = VL[i]->use_end();
744 for (unsigned j = 0; j < e; ++j) {
745 if (i != j && *U == VL[j]) {
746 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << **U << ". \n");
747 newTreeEntry(VL, false);
754 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
756 unsigned Opcode = getSameOpcode(VL);
758 // Check if it is safe to sink the loads or the stores.
759 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
760 Instruction *Last = getLastInstruction(VL);
762 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
765 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
767 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
768 << "\n because of " << *Barrier << ". Gathering.\n");
769 newTreeEntry(VL, false);
776 case Instruction::PHI: {
777 PHINode *PH = dyn_cast<PHINode>(VL0);
779 // Check for terminator values (e.g. invoke).
780 for (unsigned j = 0; j < VL.size(); ++j)
781 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
782 TerminatorInst *Term = dyn_cast<TerminatorInst>(
783 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
785 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
786 newTreeEntry(VL, false);
791 newTreeEntry(VL, true);
792 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
794 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
796 // Prepare the operand vector.
797 for (unsigned j = 0; j < VL.size(); ++j)
798 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
799 PH->getIncomingBlock(i)));
801 buildTree_rec(Operands, Depth + 1);
805 case Instruction::ExtractElement: {
806 bool Reuse = CanReuseExtract(VL);
808 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
810 newTreeEntry(VL, Reuse);
813 case Instruction::Load: {
814 // Check if the loads are consecutive or of we need to swizzle them.
815 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
816 LoadInst *L = cast<LoadInst>(VL[i]);
817 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
818 newTreeEntry(VL, false);
819 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
823 newTreeEntry(VL, true);
824 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
827 case Instruction::ZExt:
828 case Instruction::SExt:
829 case Instruction::FPToUI:
830 case Instruction::FPToSI:
831 case Instruction::FPExt:
832 case Instruction::PtrToInt:
833 case Instruction::IntToPtr:
834 case Instruction::SIToFP:
835 case Instruction::UIToFP:
836 case Instruction::Trunc:
837 case Instruction::FPTrunc:
838 case Instruction::BitCast: {
839 Type *SrcTy = VL0->getOperand(0)->getType();
840 for (unsigned i = 0; i < VL.size(); ++i) {
841 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
842 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
843 newTreeEntry(VL, false);
844 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
848 newTreeEntry(VL, true);
849 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
851 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
853 // Prepare the operand vector.
854 for (unsigned j = 0; j < VL.size(); ++j)
855 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
857 buildTree_rec(Operands, Depth+1);
861 case Instruction::ICmp:
862 case Instruction::FCmp: {
863 // Check that all of the compares have the same predicate.
864 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
865 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
866 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
867 CmpInst *Cmp = cast<CmpInst>(VL[i]);
868 if (Cmp->getPredicate() != P0 ||
869 Cmp->getOperand(0)->getType() != ComparedTy) {
870 newTreeEntry(VL, false);
871 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
876 newTreeEntry(VL, true);
877 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
879 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
881 // Prepare the operand vector.
882 for (unsigned j = 0; j < VL.size(); ++j)
883 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
885 buildTree_rec(Operands, Depth+1);
889 case Instruction::Select:
890 case Instruction::Add:
891 case Instruction::FAdd:
892 case Instruction::Sub:
893 case Instruction::FSub:
894 case Instruction::Mul:
895 case Instruction::FMul:
896 case Instruction::UDiv:
897 case Instruction::SDiv:
898 case Instruction::FDiv:
899 case Instruction::URem:
900 case Instruction::SRem:
901 case Instruction::FRem:
902 case Instruction::Shl:
903 case Instruction::LShr:
904 case Instruction::AShr:
905 case Instruction::And:
906 case Instruction::Or:
907 case Instruction::Xor: {
908 newTreeEntry(VL, true);
909 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
911 // Sort operands of the instructions so that each side is more likely to
912 // have the same opcode.
913 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
914 ValueList Left, Right;
915 reorderInputsAccordingToOpcode(VL, Left, Right);
916 buildTree_rec(Left, Depth + 1);
917 buildTree_rec(Right, Depth + 1);
921 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
923 // Prepare the operand vector.
924 for (unsigned j = 0; j < VL.size(); ++j)
925 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
927 buildTree_rec(Operands, Depth+1);
931 case Instruction::Store: {
932 // Check if the stores are consecutive or of we need to swizzle them.
933 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
934 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
935 newTreeEntry(VL, false);
936 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
940 newTreeEntry(VL, true);
941 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
944 for (unsigned j = 0; j < VL.size(); ++j)
945 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
947 // We can ignore these values because we are sinking them down.
948 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
949 buildTree_rec(Operands, Depth + 1);
953 newTreeEntry(VL, false);
954 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
959 int BoUpSLP::getEntryCost(TreeEntry *E) {
960 ArrayRef<Value*> VL = E->Scalars;
962 Type *ScalarTy = VL[0]->getType();
963 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
964 ScalarTy = SI->getValueOperand()->getType();
965 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
967 if (E->NeedToGather) {
971 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
973 return getGatherCost(E->Scalars);
976 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
978 Instruction *VL0 = cast<Instruction>(VL[0]);
979 unsigned Opcode = VL0->getOpcode();
981 case Instruction::PHI: {
984 case Instruction::ExtractElement: {
985 if (CanReuseExtract(VL))
987 return getGatherCost(VecTy);
989 case Instruction::ZExt:
990 case Instruction::SExt:
991 case Instruction::FPToUI:
992 case Instruction::FPToSI:
993 case Instruction::FPExt:
994 case Instruction::PtrToInt:
995 case Instruction::IntToPtr:
996 case Instruction::SIToFP:
997 case Instruction::UIToFP:
998 case Instruction::Trunc:
999 case Instruction::FPTrunc:
1000 case Instruction::BitCast: {
1001 Type *SrcTy = VL0->getOperand(0)->getType();
1003 // Calculate the cost of this instruction.
1004 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1005 VL0->getType(), SrcTy);
1007 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1008 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1009 return VecCost - ScalarCost;
1011 case Instruction::FCmp:
1012 case Instruction::ICmp:
1013 case Instruction::Select:
1014 case Instruction::Add:
1015 case Instruction::FAdd:
1016 case Instruction::Sub:
1017 case Instruction::FSub:
1018 case Instruction::Mul:
1019 case Instruction::FMul:
1020 case Instruction::UDiv:
1021 case Instruction::SDiv:
1022 case Instruction::FDiv:
1023 case Instruction::URem:
1024 case Instruction::SRem:
1025 case Instruction::FRem:
1026 case Instruction::Shl:
1027 case Instruction::LShr:
1028 case Instruction::AShr:
1029 case Instruction::And:
1030 case Instruction::Or:
1031 case Instruction::Xor: {
1032 // Calculate the cost of this instruction.
1035 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1036 Opcode == Instruction::Select) {
1037 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1038 ScalarCost = VecTy->getNumElements() *
1039 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1040 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1042 // Certain instructions can be cheaper to vectorize if they have a
1043 // constant second vector operand.
1044 TargetTransformInfo::OperandValueKind Op1VK =
1045 TargetTransformInfo::OK_AnyValue;
1046 TargetTransformInfo::OperandValueKind Op2VK =
1047 TargetTransformInfo::OK_UniformConstantValue;
1049 // If all operands are exactly the same ConstantInt then set the
1050 // operand kind to OK_UniformConstantValue.
1051 // If instead not all operands are constants, then set the operand kind
1052 // to OK_AnyValue. If all operands are constants but not the same,
1053 // then set the operand kind to OK_NonUniformConstantValue.
1054 ConstantInt *CInt = NULL;
1055 for (unsigned i = 0; i < VL.size(); ++i) {
1056 const Instruction *I = cast<Instruction>(VL[i]);
1057 if (!isa<ConstantInt>(I->getOperand(1))) {
1058 Op2VK = TargetTransformInfo::OK_AnyValue;
1062 CInt = cast<ConstantInt>(I->getOperand(1));
1065 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1066 CInt != cast<ConstantInt>(I->getOperand(1)))
1067 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1071 VecTy->getNumElements() *
1072 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1073 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1075 return VecCost - ScalarCost;
1077 case Instruction::Load: {
1078 // Cost of wide load - cost of scalar loads.
1079 int ScalarLdCost = VecTy->getNumElements() *
1080 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1081 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1082 return VecLdCost - ScalarLdCost;
1084 case Instruction::Store: {
1085 // We know that we can merge the stores. Calculate the cost.
1086 int ScalarStCost = VecTy->getNumElements() *
1087 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1088 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1089 return VecStCost - ScalarStCost;
1092 llvm_unreachable("Unknown instruction");
1096 bool BoUpSLP::isFullyVectorizableTinyTree() {
1097 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1098 VectorizableTree.size() << " is fully vectorizable .\n");
1100 // We only handle trees of height 2.
1101 if (VectorizableTree.size() != 2)
1104 // Handle splat stores.
1105 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1108 // Gathering cost would be too much for tiny trees.
1109 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1115 int BoUpSLP::getTreeCost() {
1117 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1118 VectorizableTree.size() << ".\n");
1120 // We only vectorize tiny trees if it is fully vectorizable.
1121 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1122 if (!VectorizableTree.size()) {
1123 assert(!ExternalUses.size() && "We should not have any external users");
1128 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1130 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1131 int C = getEntryCost(&VectorizableTree[i]);
1132 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1133 << *VectorizableTree[i].Scalars[0] << " .\n");
1137 SmallSet<Value *, 16> ExtractCostCalculated;
1138 int ExtractCost = 0;
1139 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1141 // We only add extract cost once for the same scalar.
1142 if (!ExtractCostCalculated.insert(I->Scalar))
1145 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1146 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1150 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1151 return Cost + ExtractCost;
1154 int BoUpSLP::getGatherCost(Type *Ty) {
1156 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1157 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1161 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1162 // Find the type of the operands in VL.
1163 Type *ScalarTy = VL[0]->getType();
1164 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1165 ScalarTy = SI->getValueOperand()->getType();
1166 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1167 // Find the cost of inserting/extracting values from the vector.
1168 return getGatherCost(VecTy);
1171 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1172 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1173 return AA->getLocation(SI);
1174 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1175 return AA->getLocation(LI);
1176 return AliasAnalysis::Location();
1179 Value *BoUpSLP::getPointerOperand(Value *I) {
1180 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1181 return LI->getPointerOperand();
1182 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1183 return SI->getPointerOperand();
1187 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1188 if (LoadInst *L = dyn_cast<LoadInst>(I))
1189 return L->getPointerAddressSpace();
1190 if (StoreInst *S = dyn_cast<StoreInst>(I))
1191 return S->getPointerAddressSpace();
1195 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1196 Value *PtrA = getPointerOperand(A);
1197 Value *PtrB = getPointerOperand(B);
1198 unsigned ASA = getAddressSpaceOperand(A);
1199 unsigned ASB = getAddressSpaceOperand(B);
1201 // Check that the address spaces match and that the pointers are valid.
1202 if (!PtrA || !PtrB || (ASA != ASB))
1205 // Make sure that A and B are different pointers of the same type.
1206 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1209 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1210 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1211 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1213 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1214 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1215 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1217 APInt OffsetDelta = OffsetB - OffsetA;
1219 // Check if they are based on the same pointer. That makes the offsets
1222 return OffsetDelta == Size;
1224 // Compute the necessary base pointer delta to have the necessary final delta
1225 // equal to the size.
1226 APInt BaseDelta = Size - OffsetDelta;
1228 // Otherwise compute the distance with SCEV between the base pointers.
1229 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1230 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1231 const SCEV *C = SE->getConstant(BaseDelta);
1232 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1233 return X == PtrSCEVB;
1236 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1237 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1238 BasicBlock::iterator I = Src, E = Dst;
1239 /// Scan all of the instruction from SRC to DST and check if
1240 /// the source may alias.
1241 for (++I; I != E; ++I) {
1242 // Ignore store instructions that are marked as 'ignore'.
1243 if (MemBarrierIgnoreList.count(I))
1245 if (Src->mayWriteToMemory()) /* Write */ {
1246 if (!I->mayReadOrWriteMemory())
1249 if (!I->mayWriteToMemory())
1252 AliasAnalysis::Location A = getLocation(&*I);
1253 AliasAnalysis::Location B = getLocation(Src);
1255 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1261 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1262 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1263 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1264 BlockNumbering &BN = BlocksNumbers[BB];
1266 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1267 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1268 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1272 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1273 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1274 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1275 BlockNumbering &BN = BlocksNumbers[BB];
1277 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1278 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1279 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1280 Instruction *I = BN.getInstruction(MaxIdx);
1281 assert(I && "bad location");
1285 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1286 Instruction *VL0 = cast<Instruction>(VL[0]);
1287 Instruction *LastInst = getLastInstruction(VL);
1288 BasicBlock::iterator NextInst = LastInst;
1290 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1291 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1294 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1295 Value *Vec = UndefValue::get(Ty);
1296 // Generate the 'InsertElement' instruction.
1297 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1298 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1299 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1300 GatherSeq.insert(Insrt);
1301 CSEBlocks.insert(Insrt->getParent());
1303 // Add to our 'need-to-extract' list.
1304 if (ScalarToTreeEntry.count(VL[i])) {
1305 int Idx = ScalarToTreeEntry[VL[i]];
1306 TreeEntry *E = &VectorizableTree[Idx];
1307 // Find which lane we need to extract.
1309 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1310 // Is this the lane of the scalar that we are looking for ?
1311 if (E->Scalars[Lane] == VL[i]) {
1316 assert(FoundLane >= 0 && "Could not find the correct lane");
1317 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1325 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1326 SmallDenseMap<Value*, int>::const_iterator Entry
1327 = ScalarToTreeEntry.find(VL[0]);
1328 if (Entry != ScalarToTreeEntry.end()) {
1329 int Idx = Entry->second;
1330 const TreeEntry *En = &VectorizableTree[Idx];
1331 if (En->isSame(VL) && En->VectorizedValue)
1332 return En->VectorizedValue;
1337 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1338 if (ScalarToTreeEntry.count(VL[0])) {
1339 int Idx = ScalarToTreeEntry[VL[0]];
1340 TreeEntry *E = &VectorizableTree[Idx];
1342 return vectorizeTree(E);
1345 Type *ScalarTy = VL[0]->getType();
1346 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1347 ScalarTy = SI->getValueOperand()->getType();
1348 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1350 return Gather(VL, VecTy);
1353 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1354 IRBuilder<>::InsertPointGuard Guard(Builder);
1356 if (E->VectorizedValue) {
1357 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1358 return E->VectorizedValue;
1361 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1362 Type *ScalarTy = VL0->getType();
1363 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1364 ScalarTy = SI->getValueOperand()->getType();
1365 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1367 if (E->NeedToGather) {
1368 setInsertPointAfterBundle(E->Scalars);
1369 return Gather(E->Scalars, VecTy);
1372 unsigned Opcode = VL0->getOpcode();
1373 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1376 case Instruction::PHI: {
1377 PHINode *PH = dyn_cast<PHINode>(VL0);
1378 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1379 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1380 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1381 E->VectorizedValue = NewPhi;
1383 // PHINodes may have multiple entries from the same block. We want to
1384 // visit every block once.
1385 SmallSet<BasicBlock*, 4> VisitedBBs;
1387 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1389 BasicBlock *IBB = PH->getIncomingBlock(i);
1391 if (!VisitedBBs.insert(IBB)) {
1392 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1396 // Prepare the operand vector.
1397 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1398 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1399 getIncomingValueForBlock(IBB));
1401 Builder.SetInsertPoint(IBB->getTerminator());
1402 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1403 Value *Vec = vectorizeTree(Operands);
1404 NewPhi->addIncoming(Vec, IBB);
1407 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1408 "Invalid number of incoming values");
1412 case Instruction::ExtractElement: {
1413 if (CanReuseExtract(E->Scalars)) {
1414 Value *V = VL0->getOperand(0);
1415 E->VectorizedValue = V;
1418 return Gather(E->Scalars, VecTy);
1420 case Instruction::ZExt:
1421 case Instruction::SExt:
1422 case Instruction::FPToUI:
1423 case Instruction::FPToSI:
1424 case Instruction::FPExt:
1425 case Instruction::PtrToInt:
1426 case Instruction::IntToPtr:
1427 case Instruction::SIToFP:
1428 case Instruction::UIToFP:
1429 case Instruction::Trunc:
1430 case Instruction::FPTrunc:
1431 case Instruction::BitCast: {
1433 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1434 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1436 setInsertPointAfterBundle(E->Scalars);
1438 Value *InVec = vectorizeTree(INVL);
1440 if (Value *V = alreadyVectorized(E->Scalars))
1443 CastInst *CI = dyn_cast<CastInst>(VL0);
1444 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1445 E->VectorizedValue = V;
1448 case Instruction::FCmp:
1449 case Instruction::ICmp: {
1450 ValueList LHSV, RHSV;
1451 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1452 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1453 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1456 setInsertPointAfterBundle(E->Scalars);
1458 Value *L = vectorizeTree(LHSV);
1459 Value *R = vectorizeTree(RHSV);
1461 if (Value *V = alreadyVectorized(E->Scalars))
1464 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1466 if (Opcode == Instruction::FCmp)
1467 V = Builder.CreateFCmp(P0, L, R);
1469 V = Builder.CreateICmp(P0, L, R);
1471 E->VectorizedValue = V;
1474 case Instruction::Select: {
1475 ValueList TrueVec, FalseVec, CondVec;
1476 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1477 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1478 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1479 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1482 setInsertPointAfterBundle(E->Scalars);
1484 Value *Cond = vectorizeTree(CondVec);
1485 Value *True = vectorizeTree(TrueVec);
1486 Value *False = vectorizeTree(FalseVec);
1488 if (Value *V = alreadyVectorized(E->Scalars))
1491 Value *V = Builder.CreateSelect(Cond, True, False);
1492 E->VectorizedValue = V;
1495 case Instruction::Add:
1496 case Instruction::FAdd:
1497 case Instruction::Sub:
1498 case Instruction::FSub:
1499 case Instruction::Mul:
1500 case Instruction::FMul:
1501 case Instruction::UDiv:
1502 case Instruction::SDiv:
1503 case Instruction::FDiv:
1504 case Instruction::URem:
1505 case Instruction::SRem:
1506 case Instruction::FRem:
1507 case Instruction::Shl:
1508 case Instruction::LShr:
1509 case Instruction::AShr:
1510 case Instruction::And:
1511 case Instruction::Or:
1512 case Instruction::Xor: {
1513 ValueList LHSVL, RHSVL;
1514 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1515 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1517 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1518 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1519 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1522 setInsertPointAfterBundle(E->Scalars);
1524 Value *LHS = vectorizeTree(LHSVL);
1525 Value *RHS = vectorizeTree(RHSVL);
1527 if (LHS == RHS && isa<Instruction>(LHS)) {
1528 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1531 if (Value *V = alreadyVectorized(E->Scalars))
1534 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1535 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1536 E->VectorizedValue = V;
1538 if (Instruction *I = dyn_cast<Instruction>(V))
1539 return propagateMetadata(I, E->Scalars);
1543 case Instruction::Load: {
1544 // Loads are inserted at the head of the tree because we don't want to
1545 // sink them all the way down past store instructions.
1546 setInsertPointAfterBundle(E->Scalars);
1548 LoadInst *LI = cast<LoadInst>(VL0);
1549 unsigned AS = LI->getPointerAddressSpace();
1551 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1552 VecTy->getPointerTo(AS));
1553 unsigned Alignment = LI->getAlignment();
1554 LI = Builder.CreateLoad(VecPtr);
1555 LI->setAlignment(Alignment);
1556 E->VectorizedValue = LI;
1557 return propagateMetadata(LI, E->Scalars);
1559 case Instruction::Store: {
1560 StoreInst *SI = cast<StoreInst>(VL0);
1561 unsigned Alignment = SI->getAlignment();
1562 unsigned AS = SI->getPointerAddressSpace();
1565 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1566 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1568 setInsertPointAfterBundle(E->Scalars);
1570 Value *VecValue = vectorizeTree(ValueOp);
1571 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1572 VecTy->getPointerTo(AS));
1573 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1574 S->setAlignment(Alignment);
1575 E->VectorizedValue = S;
1576 return propagateMetadata(S, E->Scalars);
1579 llvm_unreachable("unknown inst");
1584 Value *BoUpSLP::vectorizeTree() {
1585 Builder.SetInsertPoint(F->getEntryBlock().begin());
1586 vectorizeTree(&VectorizableTree[0]);
1588 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1590 // Extract all of the elements with the external uses.
1591 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1593 Value *Scalar = it->Scalar;
1594 llvm::User *User = it->User;
1596 // Skip users that we already RAUW. This happens when one instruction
1597 // has multiple uses of the same value.
1598 if (std::find(Scalar->use_begin(), Scalar->use_end(), User) ==
1601 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1603 int Idx = ScalarToTreeEntry[Scalar];
1604 TreeEntry *E = &VectorizableTree[Idx];
1605 assert(!E->NeedToGather && "Extracting from a gather list");
1607 Value *Vec = E->VectorizedValue;
1608 assert(Vec && "Can't find vectorizable value");
1610 Value *Lane = Builder.getInt32(it->Lane);
1611 // Generate extracts for out-of-tree users.
1612 // Find the insertion point for the extractelement lane.
1613 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1614 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt());
1615 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1616 CSEBlocks.insert(PN->getParent());
1617 User->replaceUsesOfWith(Scalar, Ex);
1618 } else if (isa<Instruction>(Vec)){
1619 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1620 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1621 if (PH->getIncomingValue(i) == Scalar) {
1622 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1623 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1624 CSEBlocks.insert(PH->getIncomingBlock(i));
1625 PH->setOperand(i, Ex);
1629 Builder.SetInsertPoint(cast<Instruction>(User));
1630 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1631 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1632 User->replaceUsesOfWith(Scalar, Ex);
1635 Builder.SetInsertPoint(F->getEntryBlock().begin());
1636 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1637 CSEBlocks.insert(&F->getEntryBlock());
1638 User->replaceUsesOfWith(Scalar, Ex);
1641 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1644 // For each vectorized value:
1645 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1646 TreeEntry *Entry = &VectorizableTree[EIdx];
1649 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1650 Value *Scalar = Entry->Scalars[Lane];
1652 // No need to handle users of gathered values.
1653 if (Entry->NeedToGather)
1656 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1658 Type *Ty = Scalar->getType();
1659 if (!Ty->isVoidTy()) {
1660 for (Value::use_iterator User = Scalar->use_begin(),
1661 UE = Scalar->use_end(); User != UE; ++User) {
1662 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n");
1664 assert((ScalarToTreeEntry.count(*User) ||
1665 // It is legal to replace the reduction users by undef.
1666 (RdxOps && RdxOps->count(*User))) &&
1667 "Replacing out-of-tree value with undef");
1669 Value *Undef = UndefValue::get(Ty);
1670 Scalar->replaceAllUsesWith(Undef);
1672 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1673 cast<Instruction>(Scalar)->eraseFromParent();
1677 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1678 BlocksNumbers[it].forget();
1680 Builder.ClearInsertionPoint();
1682 return VectorizableTree[0].VectorizedValue;
1685 void BoUpSLP::optimizeGatherSequence() {
1686 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1687 << " gather sequences instructions.\n");
1688 // LICM InsertElementInst sequences.
1689 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1690 e = GatherSeq.end(); it != e; ++it) {
1691 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1696 // Check if this block is inside a loop.
1697 Loop *L = LI->getLoopFor(Insert->getParent());
1701 // Check if it has a preheader.
1702 BasicBlock *PreHeader = L->getLoopPreheader();
1706 // If the vector or the element that we insert into it are
1707 // instructions that are defined in this basic block then we can't
1708 // hoist this instruction.
1709 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1710 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1711 if (CurrVec && L->contains(CurrVec))
1713 if (NewElem && L->contains(NewElem))
1716 // We can hoist this instruction. Move it to the pre-header.
1717 Insert->moveBefore(PreHeader->getTerminator());
1720 // Sort blocks by domination. This ensures we visit a block after all blocks
1721 // dominating it are visited.
1722 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1723 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1724 [this](const BasicBlock *A, const BasicBlock *B) {
1725 return DT->properlyDominates(A, B);
1728 // Perform O(N^2) search over the gather sequences and merge identical
1729 // instructions. TODO: We can further optimize this scan if we split the
1730 // instructions into different buckets based on the insert lane.
1731 SmallVector<Instruction *, 16> Visited;
1732 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1733 E = CSEWorkList.end();
1735 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1736 "Worklist not sorted properly!");
1737 BasicBlock *BB = *I;
1738 // For all instructions in blocks containing gather sequences:
1739 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1740 Instruction *In = it++;
1741 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1744 // Check if we can replace this instruction with any of the
1745 // visited instructions.
1746 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1749 if (In->isIdenticalTo(*v) &&
1750 DT->dominates((*v)->getParent(), In->getParent())) {
1751 In->replaceAllUsesWith(*v);
1752 In->eraseFromParent();
1758 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1759 Visited.push_back(In);
1767 /// The SLPVectorizer Pass.
1768 struct SLPVectorizer : public FunctionPass {
1769 typedef SmallVector<StoreInst *, 8> StoreList;
1770 typedef MapVector<Value *, StoreList> StoreListMap;
1772 /// Pass identification, replacement for typeid
1775 explicit SLPVectorizer() : FunctionPass(ID) {
1776 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1779 ScalarEvolution *SE;
1780 const DataLayout *DL;
1781 TargetTransformInfo *TTI;
1786 bool runOnFunction(Function &F) override {
1787 if (skipOptnoneFunction(F))
1790 SE = &getAnalysis<ScalarEvolution>();
1791 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1792 DL = DLP ? &DLP->getDataLayout() : 0;
1793 TTI = &getAnalysis<TargetTransformInfo>();
1794 AA = &getAnalysis<AliasAnalysis>();
1795 LI = &getAnalysis<LoopInfo>();
1796 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1799 bool Changed = false;
1801 // If the target claims to have no vector registers don't attempt
1803 if (!TTI->getNumberOfRegisters(true))
1806 // Must have DataLayout. We can't require it because some tests run w/o
1811 // Don't vectorize when the attribute NoImplicitFloat is used.
1812 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1815 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1817 // Use the bottom up slp vectorizer to construct chains that start with
1818 // he store instructions.
1819 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1821 // Scan the blocks in the function in post order.
1822 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1823 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1824 BasicBlock *BB = *it;
1826 // Vectorize trees that end at stores.
1827 if (unsigned count = collectStores(BB, R)) {
1829 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1830 Changed |= vectorizeStoreChains(R);
1833 // Vectorize trees that end at reductions.
1834 Changed |= vectorizeChainsInBlock(BB, R);
1838 R.optimizeGatherSequence();
1839 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1840 DEBUG(verifyFunction(F));
1845 void getAnalysisUsage(AnalysisUsage &AU) const override {
1846 FunctionPass::getAnalysisUsage(AU);
1847 AU.addRequired<ScalarEvolution>();
1848 AU.addRequired<AliasAnalysis>();
1849 AU.addRequired<TargetTransformInfo>();
1850 AU.addRequired<LoopInfo>();
1851 AU.addRequired<DominatorTreeWrapperPass>();
1852 AU.addPreserved<LoopInfo>();
1853 AU.addPreserved<DominatorTreeWrapperPass>();
1854 AU.setPreservesCFG();
1859 /// \brief Collect memory references and sort them according to their base
1860 /// object. We sort the stores to their base objects to reduce the cost of the
1861 /// quadratic search on the stores. TODO: We can further reduce this cost
1862 /// if we flush the chain creation every time we run into a memory barrier.
1863 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1865 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1866 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1868 /// \brief Try to vectorize a list of operands.
1869 /// \returns true if a value was vectorized.
1870 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1872 /// \brief Try to vectorize a chain that may start at the operands of \V;
1873 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1875 /// \brief Vectorize the stores that were collected in StoreRefs.
1876 bool vectorizeStoreChains(BoUpSLP &R);
1878 /// \brief Scan the basic block and look for patterns that are likely to start
1879 /// a vectorization chain.
1880 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1882 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1885 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1888 StoreListMap StoreRefs;
1891 /// \brief Check that the Values in the slice in VL array are still existent in
1892 /// the WeakVH array.
1893 /// Vectorization of part of the VL array may cause later values in the VL array
1894 /// to become invalid. We track when this has happened in the WeakVH array.
1895 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1896 SmallVectorImpl<WeakVH> &VH,
1897 unsigned SliceBegin,
1898 unsigned SliceSize) {
1899 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1906 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1907 int CostThreshold, BoUpSLP &R) {
1908 unsigned ChainLen = Chain.size();
1909 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1911 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1912 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1913 unsigned VF = MinVecRegSize / Sz;
1915 if (!isPowerOf2_32(Sz) || VF < 2)
1918 // Keep track of values that were delete by vectorizing in the loop below.
1919 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
1921 bool Changed = false;
1922 // Look for profitable vectorizable trees at all offsets, starting at zero.
1923 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1927 // Check that a previous iteration of this loop did not delete the Value.
1928 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
1931 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1933 ArrayRef<Value *> Operands = Chain.slice(i, VF);
1935 R.buildTree(Operands);
1937 int Cost = R.getTreeCost();
1939 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
1940 if (Cost < CostThreshold) {
1941 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
1944 // Move to the next bundle.
1953 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
1954 int costThreshold, BoUpSLP &R) {
1955 SetVector<Value *> Heads, Tails;
1956 SmallDenseMap<Value *, Value *> ConsecutiveChain;
1958 // We may run into multiple chains that merge into a single chain. We mark the
1959 // stores that we vectorized so that we don't visit the same store twice.
1960 BoUpSLP::ValueSet VectorizedStores;
1961 bool Changed = false;
1963 // Do a quadratic search on all of the given stores and find
1964 // all of the pairs of stores that follow each other.
1965 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
1966 for (unsigned j = 0; j < e; ++j) {
1970 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
1971 Tails.insert(Stores[j]);
1972 Heads.insert(Stores[i]);
1973 ConsecutiveChain[Stores[i]] = Stores[j];
1978 // For stores that start but don't end a link in the chain:
1979 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
1981 if (Tails.count(*it))
1984 // We found a store instr that starts a chain. Now follow the chain and try
1986 BoUpSLP::ValueList Operands;
1988 // Collect the chain into a list.
1989 while (Tails.count(I) || Heads.count(I)) {
1990 if (VectorizedStores.count(I))
1992 Operands.push_back(I);
1993 // Move to the next value in the chain.
1994 I = ConsecutiveChain[I];
1997 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
1999 // Mark the vectorized stores so that we don't vectorize them again.
2001 VectorizedStores.insert(Operands.begin(), Operands.end());
2002 Changed |= Vectorized;
2009 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2012 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2013 StoreInst *SI = dyn_cast<StoreInst>(it);
2017 // Don't touch volatile stores.
2018 if (!SI->isSimple())
2021 // Check that the pointer points to scalars.
2022 Type *Ty = SI->getValueOperand()->getType();
2023 if (Ty->isAggregateType() || Ty->isVectorTy())
2026 // Find the base pointer.
2027 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2029 // Save the store locations.
2030 StoreRefs[Ptr].push_back(SI);
2036 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2039 Value *VL[] = { A, B };
2040 return tryToVectorizeList(VL, R);
2043 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
2047 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2049 // Check that all of the parts are scalar instructions of the same type.
2050 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2054 unsigned Opcode0 = I0->getOpcode();
2056 Type *Ty0 = I0->getType();
2057 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2058 unsigned VF = MinVecRegSize / Sz;
2060 for (int i = 0, e = VL.size(); i < e; ++i) {
2061 Type *Ty = VL[i]->getType();
2062 if (Ty->isAggregateType() || Ty->isVectorTy())
2064 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2065 if (!Inst || Inst->getOpcode() != Opcode0)
2069 bool Changed = false;
2071 // Keep track of values that were delete by vectorizing in the loop below.
2072 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2074 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2075 unsigned OpsWidth = 0;
2082 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2085 // Check that a previous iteration of this loop did not delete the Value.
2086 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2089 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2091 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2094 int Cost = R.getTreeCost();
2096 if (Cost < -SLPCostThreshold) {
2097 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n");
2100 // Move to the next bundle.
2109 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2113 // Try to vectorize V.
2114 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2117 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2118 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2120 if (B && B->hasOneUse()) {
2121 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2122 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2123 if (tryToVectorizePair(A, B0, R)) {
2127 if (tryToVectorizePair(A, B1, R)) {
2134 if (A && A->hasOneUse()) {
2135 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2136 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2137 if (tryToVectorizePair(A0, B, R)) {
2141 if (tryToVectorizePair(A1, B, R)) {
2149 /// \brief Generate a shuffle mask to be used in a reduction tree.
2151 /// \param VecLen The length of the vector to be reduced.
2152 /// \param NumEltsToRdx The number of elements that should be reduced in the
2154 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2155 /// reduction. A pairwise reduction will generate a mask of
2156 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2157 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2158 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2159 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2160 bool IsPairwise, bool IsLeft,
2161 IRBuilder<> &Builder) {
2162 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2164 SmallVector<Constant *, 32> ShuffleMask(
2165 VecLen, UndefValue::get(Builder.getInt32Ty()));
2168 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2169 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2170 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2172 // Move the upper half of the vector to the lower half.
2173 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2174 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2176 return ConstantVector::get(ShuffleMask);
2180 /// Model horizontal reductions.
2182 /// A horizontal reduction is a tree of reduction operations (currently add and
2183 /// fadd) that has operations that can be put into a vector as its leaf.
2184 /// For example, this tree:
2191 /// This tree has "mul" as its reduced values and "+" as its reduction
2192 /// operations. A reduction might be feeding into a store or a binary operation
2207 class HorizontalReduction {
2208 SmallPtrSet<Value *, 16> ReductionOps;
2209 SmallVector<Value *, 32> ReducedVals;
2211 BinaryOperator *ReductionRoot;
2212 PHINode *ReductionPHI;
2214 /// The opcode of the reduction.
2215 unsigned ReductionOpcode;
2216 /// The opcode of the values we perform a reduction on.
2217 unsigned ReducedValueOpcode;
2218 /// The width of one full horizontal reduction operation.
2219 unsigned ReduxWidth;
2220 /// Should we model this reduction as a pairwise reduction tree or a tree that
2221 /// splits the vector in halves and adds those halves.
2222 bool IsPairwiseReduction;
2225 HorizontalReduction()
2226 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0),
2227 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2229 /// \brief Try to find a reduction tree.
2230 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2231 const DataLayout *DL) {
2233 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2234 "Thi phi needs to use the binary operator");
2236 // We could have a initial reductions that is not an add.
2237 // r *= v1 + v2 + v3 + v4
2238 // In such a case start looking for a tree rooted in the first '+'.
2240 if (B->getOperand(0) == Phi) {
2242 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2243 } else if (B->getOperand(1) == Phi) {
2245 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2252 Type *Ty = B->getType();
2253 if (Ty->isVectorTy())
2256 ReductionOpcode = B->getOpcode();
2257 ReducedValueOpcode = 0;
2258 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2265 // We currently only support adds.
2266 if (ReductionOpcode != Instruction::Add &&
2267 ReductionOpcode != Instruction::FAdd)
2270 // Post order traverse the reduction tree starting at B. We only handle true
2271 // trees containing only binary operators.
2272 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2273 Stack.push_back(std::make_pair(B, 0));
2274 while (!Stack.empty()) {
2275 BinaryOperator *TreeN = Stack.back().first;
2276 unsigned EdgeToVist = Stack.back().second++;
2277 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2279 // Only handle trees in the current basic block.
2280 if (TreeN->getParent() != B->getParent())
2283 // Each tree node needs to have one user except for the ultimate
2285 if (!TreeN->hasOneUse() && TreeN != B)
2289 if (EdgeToVist == 2 || IsReducedValue) {
2290 if (IsReducedValue) {
2291 // Make sure that the opcodes of the operations that we are going to
2293 if (!ReducedValueOpcode)
2294 ReducedValueOpcode = TreeN->getOpcode();
2295 else if (ReducedValueOpcode != TreeN->getOpcode())
2297 ReducedVals.push_back(TreeN);
2299 // We need to be able to reassociate the adds.
2300 if (!TreeN->isAssociative())
2302 ReductionOps.insert(TreeN);
2309 // Visit left or right.
2310 Value *NextV = TreeN->getOperand(EdgeToVist);
2311 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2313 Stack.push_back(std::make_pair(Next, 0));
2314 else if (NextV != Phi)
2320 /// \brief Attempt to vectorize the tree found by
2321 /// matchAssociativeReduction.
2322 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2323 if (ReducedVals.empty())
2326 unsigned NumReducedVals = ReducedVals.size();
2327 if (NumReducedVals < ReduxWidth)
2330 Value *VectorizedTree = 0;
2331 IRBuilder<> Builder(ReductionRoot);
2332 FastMathFlags Unsafe;
2333 Unsafe.setUnsafeAlgebra();
2334 Builder.SetFastMathFlags(Unsafe);
2337 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2338 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2339 V.buildTree(ValsToReduce, &ReductionOps);
2342 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2343 if (Cost >= -SLPCostThreshold)
2346 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2349 // Vectorize a tree.
2350 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2351 Value *VectorizedRoot = V.vectorizeTree();
2353 // Emit a reduction.
2354 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2355 if (VectorizedTree) {
2356 Builder.SetCurrentDebugLocation(Loc);
2357 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2358 ReducedSubTree, "bin.rdx");
2360 VectorizedTree = ReducedSubTree;
2363 if (VectorizedTree) {
2364 // Finish the reduction.
2365 for (; i < NumReducedVals; ++i) {
2366 Builder.SetCurrentDebugLocation(
2367 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2368 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2373 assert(ReductionRoot != NULL && "Need a reduction operation");
2374 ReductionRoot->setOperand(0, VectorizedTree);
2375 ReductionRoot->setOperand(1, ReductionPHI);
2377 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2379 return VectorizedTree != 0;
2384 /// \brief Calcuate the cost of a reduction.
2385 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2386 Type *ScalarTy = FirstReducedVal->getType();
2387 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2389 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2390 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2392 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2393 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2395 int ScalarReduxCost =
2396 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2398 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2399 << " for reduction that starts with " << *FirstReducedVal
2401 << (IsPairwiseReduction ? "pairwise" : "splitting")
2402 << " reduction)\n");
2404 return VecReduxCost - ScalarReduxCost;
2407 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2408 Value *R, const Twine &Name = "") {
2409 if (Opcode == Instruction::FAdd)
2410 return Builder.CreateFAdd(L, R, Name);
2411 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2414 /// \brief Emit a horizontal reduction of the vectorized value.
2415 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2416 assert(VectorizedValue && "Need to have a vectorized tree node");
2417 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2418 assert(isPowerOf2_32(ReduxWidth) &&
2419 "We only handle power-of-two reductions for now");
2421 Value *TmpVec = ValToReduce;
2422 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2423 if (IsPairwiseReduction) {
2425 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2427 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2429 Value *LeftShuf = Builder.CreateShuffleVector(
2430 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2431 Value *RightShuf = Builder.CreateShuffleVector(
2432 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2434 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2438 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2439 Value *Shuf = Builder.CreateShuffleVector(
2440 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2441 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2445 // The result is in the first element of the vector.
2446 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2450 /// \brief Recognize construction of vectors like
2451 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2452 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2453 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2454 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2456 /// Returns true if it matches
2458 static bool findBuildVector(InsertElementInst *IE,
2459 SmallVectorImpl<Value *> &Ops) {
2460 if (!isa<UndefValue>(IE->getOperand(0)))
2464 Ops.push_back(IE->getOperand(1));
2466 if (IE->use_empty())
2469 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->use_back());
2473 // If this isn't the final use, make sure the next insertelement is the only
2474 // use. It's OK if the final constructed vector is used multiple times
2475 if (!IE->hasOneUse())
2484 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2485 return V->getType() < V2->getType();
2488 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2489 bool Changed = false;
2490 SmallVector<Value *, 4> Incoming;
2491 SmallSet<Value *, 16> VisitedInstrs;
2493 bool HaveVectorizedPhiNodes = true;
2494 while (HaveVectorizedPhiNodes) {
2495 HaveVectorizedPhiNodes = false;
2497 // Collect the incoming values from the PHIs.
2499 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2501 PHINode *P = dyn_cast<PHINode>(instr);
2505 if (!VisitedInstrs.count(P))
2506 Incoming.push_back(P);
2510 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2512 // Try to vectorize elements base on their type.
2513 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2517 // Look for the next elements with the same type.
2518 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2519 while (SameTypeIt != E &&
2520 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2521 VisitedInstrs.insert(*SameTypeIt);
2525 // Try to vectorize them.
2526 unsigned NumElts = (SameTypeIt - IncIt);
2527 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2529 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2530 // Success start over because instructions might have been changed.
2531 HaveVectorizedPhiNodes = true;
2536 // Start over at the next instruction of a different type (or the end).
2541 VisitedInstrs.clear();
2543 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2544 // We may go through BB multiple times so skip the one we have checked.
2545 if (!VisitedInstrs.insert(it))
2548 if (isa<DbgInfoIntrinsic>(it))
2551 // Try to vectorize reductions that use PHINodes.
2552 if (PHINode *P = dyn_cast<PHINode>(it)) {
2553 // Check that the PHI is a reduction PHI.
2554 if (P->getNumIncomingValues() != 2)
2557 (P->getIncomingBlock(0) == BB
2558 ? (P->getIncomingValue(0))
2559 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
2560 // Check if this is a Binary Operator.
2561 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2565 // Try to match and vectorize a horizontal reduction.
2566 HorizontalReduction HorRdx;
2567 if (ShouldVectorizeHor &&
2568 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2569 HorRdx.tryToReduce(R, TTI)) {
2576 Value *Inst = BI->getOperand(0);
2578 Inst = BI->getOperand(1);
2580 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2581 // We would like to start over since some instructions are deleted
2582 // and the iterator may become invalid value.
2592 // Try to vectorize horizontal reductions feeding into a store.
2593 if (ShouldStartVectorizeHorAtStore)
2594 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2595 if (BinaryOperator *BinOp =
2596 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2597 HorizontalReduction HorRdx;
2598 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) &&
2599 HorRdx.tryToReduce(R, TTI)) ||
2600 tryToVectorize(BinOp, R))) {
2608 // Try to vectorize trees that start at compare instructions.
2609 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2610 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2612 // We would like to start over since some instructions are deleted
2613 // and the iterator may become invalid value.
2619 for (int i = 0; i < 2; ++i) {
2620 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2621 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2623 // We would like to start over since some instructions are deleted
2624 // and the iterator may become invalid value.
2633 // Try to vectorize trees that start at insertelement instructions.
2634 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2635 SmallVector<Value *, 8> Ops;
2636 if (!findBuildVector(IE, Ops))
2639 if (tryToVectorizeList(Ops, R)) {
2652 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2653 bool Changed = false;
2654 // Attempt to sort and vectorize each of the store-groups.
2655 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2657 if (it->second.size() < 2)
2660 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2661 << it->second.size() << ".\n");
2663 // Process the stores in chunks of 16.
2664 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2665 unsigned Len = std::min<unsigned>(CE - CI, 16);
2666 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2667 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2673 } // end anonymous namespace
2675 char SLPVectorizer::ID = 0;
2676 static const char lv_name[] = "SLP Vectorizer";
2677 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2678 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2679 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2680 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2681 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2682 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2685 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }