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, 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.
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 // Gathering cost would be too much for tiny trees.
1105 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1111 int BoUpSLP::getTreeCost() {
1113 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1114 VectorizableTree.size() << ".\n");
1116 // We only vectorize tiny trees if it is fully vectorizable.
1117 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1118 if (!VectorizableTree.size()) {
1119 assert(!ExternalUses.size() && "We should not have any external users");
1124 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1126 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1127 int C = getEntryCost(&VectorizableTree[i]);
1128 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1129 << *VectorizableTree[i].Scalars[0] << " .\n");
1133 SmallSet<Value *, 16> ExtractCostCalculated;
1134 int ExtractCost = 0;
1135 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1137 // We only add extract cost once for the same scalar.
1138 if (!ExtractCostCalculated.insert(I->Scalar))
1141 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1142 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1146 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1147 return Cost + ExtractCost;
1150 int BoUpSLP::getGatherCost(Type *Ty) {
1152 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1153 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1157 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1158 // Find the type of the operands in VL.
1159 Type *ScalarTy = VL[0]->getType();
1160 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1161 ScalarTy = SI->getValueOperand()->getType();
1162 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1163 // Find the cost of inserting/extracting values from the vector.
1164 return getGatherCost(VecTy);
1167 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1168 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1169 return AA->getLocation(SI);
1170 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1171 return AA->getLocation(LI);
1172 return AliasAnalysis::Location();
1175 Value *BoUpSLP::getPointerOperand(Value *I) {
1176 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1177 return LI->getPointerOperand();
1178 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1179 return SI->getPointerOperand();
1183 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1184 if (LoadInst *L = dyn_cast<LoadInst>(I))
1185 return L->getPointerAddressSpace();
1186 if (StoreInst *S = dyn_cast<StoreInst>(I))
1187 return S->getPointerAddressSpace();
1191 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1192 Value *PtrA = getPointerOperand(A);
1193 Value *PtrB = getPointerOperand(B);
1194 unsigned ASA = getAddressSpaceOperand(A);
1195 unsigned ASB = getAddressSpaceOperand(B);
1197 // Check that the address spaces match and that the pointers are valid.
1198 if (!PtrA || !PtrB || (ASA != ASB))
1201 // Make sure that A and B are different pointers of the same type.
1202 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1205 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1206 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1207 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1209 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1210 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1211 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1213 APInt OffsetDelta = OffsetB - OffsetA;
1215 // Check if they are based on the same pointer. That makes the offsets
1218 return OffsetDelta == Size;
1220 // Compute the necessary base pointer delta to have the necessary final delta
1221 // equal to the size.
1222 APInt BaseDelta = Size - OffsetDelta;
1224 // Otherwise compute the distance with SCEV between the base pointers.
1225 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1226 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1227 const SCEV *C = SE->getConstant(BaseDelta);
1228 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1229 return X == PtrSCEVB;
1232 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1233 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1234 BasicBlock::iterator I = Src, E = Dst;
1235 /// Scan all of the instruction from SRC to DST and check if
1236 /// the source may alias.
1237 for (++I; I != E; ++I) {
1238 // Ignore store instructions that are marked as 'ignore'.
1239 if (MemBarrierIgnoreList.count(I))
1241 if (Src->mayWriteToMemory()) /* Write */ {
1242 if (!I->mayReadOrWriteMemory())
1245 if (!I->mayWriteToMemory())
1248 AliasAnalysis::Location A = getLocation(&*I);
1249 AliasAnalysis::Location B = getLocation(Src);
1251 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1257 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1258 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1259 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1260 BlockNumbering &BN = BlocksNumbers[BB];
1262 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1263 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1264 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1268 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1269 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1270 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1271 BlockNumbering &BN = BlocksNumbers[BB];
1273 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1274 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1275 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1276 Instruction *I = BN.getInstruction(MaxIdx);
1277 assert(I && "bad location");
1281 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1282 Instruction *VL0 = cast<Instruction>(VL[0]);
1283 Instruction *LastInst = getLastInstruction(VL);
1284 BasicBlock::iterator NextInst = LastInst;
1286 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1287 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1290 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1291 Value *Vec = UndefValue::get(Ty);
1292 // Generate the 'InsertElement' instruction.
1293 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1294 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1295 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1296 GatherSeq.insert(Insrt);
1297 CSEBlocks.insert(Insrt->getParent());
1299 // Add to our 'need-to-extract' list.
1300 if (ScalarToTreeEntry.count(VL[i])) {
1301 int Idx = ScalarToTreeEntry[VL[i]];
1302 TreeEntry *E = &VectorizableTree[Idx];
1303 // Find which lane we need to extract.
1305 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1306 // Is this the lane of the scalar that we are looking for ?
1307 if (E->Scalars[Lane] == VL[i]) {
1312 assert(FoundLane >= 0 && "Could not find the correct lane");
1313 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1321 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1322 SmallDenseMap<Value*, int>::const_iterator Entry
1323 = ScalarToTreeEntry.find(VL[0]);
1324 if (Entry != ScalarToTreeEntry.end()) {
1325 int Idx = Entry->second;
1326 const TreeEntry *En = &VectorizableTree[Idx];
1327 if (En->isSame(VL) && En->VectorizedValue)
1328 return En->VectorizedValue;
1333 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1334 if (ScalarToTreeEntry.count(VL[0])) {
1335 int Idx = ScalarToTreeEntry[VL[0]];
1336 TreeEntry *E = &VectorizableTree[Idx];
1338 return vectorizeTree(E);
1341 Type *ScalarTy = VL[0]->getType();
1342 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1343 ScalarTy = SI->getValueOperand()->getType();
1344 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1346 return Gather(VL, VecTy);
1349 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1350 IRBuilder<>::InsertPointGuard Guard(Builder);
1352 if (E->VectorizedValue) {
1353 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1354 return E->VectorizedValue;
1357 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1358 Type *ScalarTy = VL0->getType();
1359 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1360 ScalarTy = SI->getValueOperand()->getType();
1361 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1363 if (E->NeedToGather) {
1364 setInsertPointAfterBundle(E->Scalars);
1365 return Gather(E->Scalars, VecTy);
1368 unsigned Opcode = VL0->getOpcode();
1369 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1372 case Instruction::PHI: {
1373 PHINode *PH = dyn_cast<PHINode>(VL0);
1374 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1375 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1376 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1377 E->VectorizedValue = NewPhi;
1379 // PHINodes may have multiple entries from the same block. We want to
1380 // visit every block once.
1381 SmallSet<BasicBlock*, 4> VisitedBBs;
1383 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1385 BasicBlock *IBB = PH->getIncomingBlock(i);
1387 if (!VisitedBBs.insert(IBB)) {
1388 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1392 // Prepare the operand vector.
1393 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1394 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1395 getIncomingValueForBlock(IBB));
1397 Builder.SetInsertPoint(IBB->getTerminator());
1398 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1399 Value *Vec = vectorizeTree(Operands);
1400 NewPhi->addIncoming(Vec, IBB);
1403 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1404 "Invalid number of incoming values");
1408 case Instruction::ExtractElement: {
1409 if (CanReuseExtract(E->Scalars)) {
1410 Value *V = VL0->getOperand(0);
1411 E->VectorizedValue = V;
1414 return Gather(E->Scalars, VecTy);
1416 case Instruction::ZExt:
1417 case Instruction::SExt:
1418 case Instruction::FPToUI:
1419 case Instruction::FPToSI:
1420 case Instruction::FPExt:
1421 case Instruction::PtrToInt:
1422 case Instruction::IntToPtr:
1423 case Instruction::SIToFP:
1424 case Instruction::UIToFP:
1425 case Instruction::Trunc:
1426 case Instruction::FPTrunc:
1427 case Instruction::BitCast: {
1429 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1430 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1432 setInsertPointAfterBundle(E->Scalars);
1434 Value *InVec = vectorizeTree(INVL);
1436 if (Value *V = alreadyVectorized(E->Scalars))
1439 CastInst *CI = dyn_cast<CastInst>(VL0);
1440 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1441 E->VectorizedValue = V;
1444 case Instruction::FCmp:
1445 case Instruction::ICmp: {
1446 ValueList LHSV, RHSV;
1447 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1448 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1449 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1452 setInsertPointAfterBundle(E->Scalars);
1454 Value *L = vectorizeTree(LHSV);
1455 Value *R = vectorizeTree(RHSV);
1457 if (Value *V = alreadyVectorized(E->Scalars))
1460 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1462 if (Opcode == Instruction::FCmp)
1463 V = Builder.CreateFCmp(P0, L, R);
1465 V = Builder.CreateICmp(P0, L, R);
1467 E->VectorizedValue = V;
1470 case Instruction::Select: {
1471 ValueList TrueVec, FalseVec, CondVec;
1472 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1473 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1474 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1475 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1478 setInsertPointAfterBundle(E->Scalars);
1480 Value *Cond = vectorizeTree(CondVec);
1481 Value *True = vectorizeTree(TrueVec);
1482 Value *False = vectorizeTree(FalseVec);
1484 if (Value *V = alreadyVectorized(E->Scalars))
1487 Value *V = Builder.CreateSelect(Cond, True, False);
1488 E->VectorizedValue = V;
1491 case Instruction::Add:
1492 case Instruction::FAdd:
1493 case Instruction::Sub:
1494 case Instruction::FSub:
1495 case Instruction::Mul:
1496 case Instruction::FMul:
1497 case Instruction::UDiv:
1498 case Instruction::SDiv:
1499 case Instruction::FDiv:
1500 case Instruction::URem:
1501 case Instruction::SRem:
1502 case Instruction::FRem:
1503 case Instruction::Shl:
1504 case Instruction::LShr:
1505 case Instruction::AShr:
1506 case Instruction::And:
1507 case Instruction::Or:
1508 case Instruction::Xor: {
1509 ValueList LHSVL, RHSVL;
1510 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1511 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1513 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1514 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1515 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1518 setInsertPointAfterBundle(E->Scalars);
1520 Value *LHS = vectorizeTree(LHSVL);
1521 Value *RHS = vectorizeTree(RHSVL);
1523 if (LHS == RHS && isa<Instruction>(LHS)) {
1524 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1527 if (Value *V = alreadyVectorized(E->Scalars))
1530 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1531 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1532 E->VectorizedValue = V;
1534 if (Instruction *I = dyn_cast<Instruction>(V))
1535 return propagateMetadata(I, E->Scalars);
1539 case Instruction::Load: {
1540 // Loads are inserted at the head of the tree because we don't want to
1541 // sink them all the way down past store instructions.
1542 setInsertPointAfterBundle(E->Scalars);
1544 LoadInst *LI = cast<LoadInst>(VL0);
1545 unsigned AS = LI->getPointerAddressSpace();
1547 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1548 VecTy->getPointerTo(AS));
1549 unsigned Alignment = LI->getAlignment();
1550 LI = Builder.CreateLoad(VecPtr);
1551 LI->setAlignment(Alignment);
1552 E->VectorizedValue = LI;
1553 return propagateMetadata(LI, E->Scalars);
1555 case Instruction::Store: {
1556 StoreInst *SI = cast<StoreInst>(VL0);
1557 unsigned Alignment = SI->getAlignment();
1558 unsigned AS = SI->getPointerAddressSpace();
1561 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1562 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1564 setInsertPointAfterBundle(E->Scalars);
1566 Value *VecValue = vectorizeTree(ValueOp);
1567 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1568 VecTy->getPointerTo(AS));
1569 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1570 S->setAlignment(Alignment);
1571 E->VectorizedValue = S;
1572 return propagateMetadata(S, E->Scalars);
1575 llvm_unreachable("unknown inst");
1580 Value *BoUpSLP::vectorizeTree() {
1581 Builder.SetInsertPoint(F->getEntryBlock().begin());
1582 vectorizeTree(&VectorizableTree[0]);
1584 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1586 // Extract all of the elements with the external uses.
1587 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1589 Value *Scalar = it->Scalar;
1590 llvm::User *User = it->User;
1592 // Skip users that we already RAUW. This happens when one instruction
1593 // has multiple uses of the same value.
1594 if (std::find(Scalar->use_begin(), Scalar->use_end(), User) ==
1597 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1599 int Idx = ScalarToTreeEntry[Scalar];
1600 TreeEntry *E = &VectorizableTree[Idx];
1601 assert(!E->NeedToGather && "Extracting from a gather list");
1603 Value *Vec = E->VectorizedValue;
1604 assert(Vec && "Can't find vectorizable value");
1606 Value *Lane = Builder.getInt32(it->Lane);
1607 // Generate extracts for out-of-tree users.
1608 // Find the insertion point for the extractelement lane.
1609 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1610 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt());
1611 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1612 CSEBlocks.insert(PN->getParent());
1613 User->replaceUsesOfWith(Scalar, Ex);
1614 } else if (isa<Instruction>(Vec)){
1615 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1616 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1617 if (PH->getIncomingValue(i) == Scalar) {
1618 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1619 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1620 CSEBlocks.insert(PH->getIncomingBlock(i));
1621 PH->setOperand(i, Ex);
1625 Builder.SetInsertPoint(cast<Instruction>(User));
1626 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1627 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1628 User->replaceUsesOfWith(Scalar, Ex);
1631 Builder.SetInsertPoint(F->getEntryBlock().begin());
1632 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1633 CSEBlocks.insert(&F->getEntryBlock());
1634 User->replaceUsesOfWith(Scalar, Ex);
1637 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1640 // For each vectorized value:
1641 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1642 TreeEntry *Entry = &VectorizableTree[EIdx];
1645 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1646 Value *Scalar = Entry->Scalars[Lane];
1648 // No need to handle users of gathered values.
1649 if (Entry->NeedToGather)
1652 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1654 Type *Ty = Scalar->getType();
1655 if (!Ty->isVoidTy()) {
1656 for (Value::use_iterator User = Scalar->use_begin(),
1657 UE = Scalar->use_end(); User != UE; ++User) {
1658 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n");
1660 assert((ScalarToTreeEntry.count(*User) ||
1661 // It is legal to replace the reduction users by undef.
1662 (RdxOps && RdxOps->count(*User))) &&
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;
1682 const DominatorTree *DT;
1685 DTCmp(const DominatorTree *DT) : DT(DT) {}
1686 bool operator()(const BasicBlock *A, const BasicBlock *B) const {
1687 return DT->properlyDominates(A, B);
1691 void BoUpSLP::optimizeGatherSequence() {
1692 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1693 << " gather sequences instructions.\n");
1694 // LICM InsertElementInst sequences.
1695 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1696 e = GatherSeq.end(); it != e; ++it) {
1697 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1702 // Check if this block is inside a loop.
1703 Loop *L = LI->getLoopFor(Insert->getParent());
1707 // Check if it has a preheader.
1708 BasicBlock *PreHeader = L->getLoopPreheader();
1712 // If the vector or the element that we insert into it are
1713 // instructions that are defined in this basic block then we can't
1714 // hoist this instruction.
1715 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1716 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1717 if (CurrVec && L->contains(CurrVec))
1719 if (NewElem && L->contains(NewElem))
1722 // We can hoist this instruction. Move it to the pre-header.
1723 Insert->moveBefore(PreHeader->getTerminator());
1726 // Sort blocks by domination. This ensures we visit a block after all blocks
1727 // dominating it are visited.
1728 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1729 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), DTCmp(DT));
1731 // Perform O(N^2) search over the gather sequences and merge identical
1732 // instructions. TODO: We can further optimize this scan if we split the
1733 // instructions into different buckets based on the insert lane.
1734 SmallVector<Instruction *, 16> Visited;
1735 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1736 E = CSEWorkList.end();
1738 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *llvm::prior(I))) &&
1739 "Worklist not sorted properly!");
1740 BasicBlock *BB = *I;
1741 // For all instructions in blocks containing gather sequences:
1742 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1743 Instruction *In = it++;
1744 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1747 // Check if we can replace this instruction with any of the
1748 // visited instructions.
1749 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1752 if (In->isIdenticalTo(*v) &&
1753 DT->dominates((*v)->getParent(), In->getParent())) {
1754 In->replaceAllUsesWith(*v);
1755 In->eraseFromParent();
1761 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1762 Visited.push_back(In);
1770 /// The SLPVectorizer Pass.
1771 struct SLPVectorizer : public FunctionPass {
1772 typedef SmallVector<StoreInst *, 8> StoreList;
1773 typedef MapVector<Value *, StoreList> StoreListMap;
1775 /// Pass identification, replacement for typeid
1778 explicit SLPVectorizer() : FunctionPass(ID) {
1779 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1782 ScalarEvolution *SE;
1784 TargetTransformInfo *TTI;
1789 virtual bool runOnFunction(Function &F) {
1790 if (skipOptnoneFunction(F))
1793 SE = &getAnalysis<ScalarEvolution>();
1794 DL = getAnalysisIfAvailable<DataLayout>();
1795 TTI = &getAnalysis<TargetTransformInfo>();
1796 AA = &getAnalysis<AliasAnalysis>();
1797 LI = &getAnalysis<LoopInfo>();
1798 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1801 bool Changed = false;
1803 // If the target claims to have no vector registers don't attempt
1805 if (!TTI->getNumberOfRegisters(true))
1808 // Must have DataLayout. We can't require it because some tests run w/o
1813 // Don't vectorize when the attribute NoImplicitFloat is used.
1814 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1817 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1819 // Use the bottom up slp vectorizer to construct chains that start with
1820 // he store instructions.
1821 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1823 // Scan the blocks in the function in post order.
1824 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1825 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1826 BasicBlock *BB = *it;
1828 // Vectorize trees that end at stores.
1829 if (unsigned count = collectStores(BB, R)) {
1831 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1832 Changed |= vectorizeStoreChains(R);
1835 // Vectorize trees that end at reductions.
1836 Changed |= vectorizeChainsInBlock(BB, R);
1840 R.optimizeGatherSequence();
1841 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1842 DEBUG(verifyFunction(F));
1847 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1848 FunctionPass::getAnalysisUsage(AU);
1849 AU.addRequired<ScalarEvolution>();
1850 AU.addRequired<AliasAnalysis>();
1851 AU.addRequired<TargetTransformInfo>();
1852 AU.addRequired<LoopInfo>();
1853 AU.addRequired<DominatorTreeWrapperPass>();
1854 AU.addPreserved<LoopInfo>();
1855 AU.addPreserved<DominatorTreeWrapperPass>();
1856 AU.setPreservesCFG();
1861 /// \brief Collect memory references and sort them according to their base
1862 /// object. We sort the stores to their base objects to reduce the cost of the
1863 /// quadratic search on the stores. TODO: We can further reduce this cost
1864 /// if we flush the chain creation every time we run into a memory barrier.
1865 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1867 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1868 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1870 /// \brief Try to vectorize a list of operands.
1871 /// \returns true if a value was vectorized.
1872 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1874 /// \brief Try to vectorize a chain that may start at the operands of \V;
1875 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1877 /// \brief Vectorize the stores that were collected in StoreRefs.
1878 bool vectorizeStoreChains(BoUpSLP &R);
1880 /// \brief Scan the basic block and look for patterns that are likely to start
1881 /// a vectorization chain.
1882 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1884 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1887 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1890 StoreListMap StoreRefs;
1893 /// \brief Check that the Values in the slice in VL array are still existent in
1894 /// the WeakVH array.
1895 /// Vectorization of part of the VL array may cause later values in the VL array
1896 /// to become invalid. We track when this has happened in the WeakVH array.
1897 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1898 SmallVectorImpl<WeakVH> &VH,
1899 unsigned SliceBegin,
1900 unsigned SliceSize) {
1901 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1908 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1909 int CostThreshold, BoUpSLP &R) {
1910 unsigned ChainLen = Chain.size();
1911 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1913 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1914 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1915 unsigned VF = MinVecRegSize / Sz;
1917 if (!isPowerOf2_32(Sz) || VF < 2)
1920 // Keep track of values that were delete by vectorizing in the loop below.
1921 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
1923 bool Changed = false;
1924 // Look for profitable vectorizable trees at all offsets, starting at zero.
1925 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1929 // Check that a previous iteration of this loop did not delete the Value.
1930 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
1933 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1935 ArrayRef<Value *> Operands = Chain.slice(i, VF);
1937 R.buildTree(Operands);
1939 int Cost = R.getTreeCost();
1941 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
1942 if (Cost < CostThreshold) {
1943 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
1946 // Move to the next bundle.
1955 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
1956 int costThreshold, BoUpSLP &R) {
1957 SetVector<Value *> Heads, Tails;
1958 SmallDenseMap<Value *, Value *> ConsecutiveChain;
1960 // We may run into multiple chains that merge into a single chain. We mark the
1961 // stores that we vectorized so that we don't visit the same store twice.
1962 BoUpSLP::ValueSet VectorizedStores;
1963 bool Changed = false;
1965 // Do a quadratic search on all of the given stores and find
1966 // all of the pairs of stores that follow each other.
1967 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
1968 for (unsigned j = 0; j < e; ++j) {
1972 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
1973 Tails.insert(Stores[j]);
1974 Heads.insert(Stores[i]);
1975 ConsecutiveChain[Stores[i]] = Stores[j];
1980 // For stores that start but don't end a link in the chain:
1981 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
1983 if (Tails.count(*it))
1986 // We found a store instr that starts a chain. Now follow the chain and try
1988 BoUpSLP::ValueList Operands;
1990 // Collect the chain into a list.
1991 while (Tails.count(I) || Heads.count(I)) {
1992 if (VectorizedStores.count(I))
1994 Operands.push_back(I);
1995 // Move to the next value in the chain.
1996 I = ConsecutiveChain[I];
1999 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2001 // Mark the vectorized stores so that we don't vectorize them again.
2003 VectorizedStores.insert(Operands.begin(), Operands.end());
2004 Changed |= Vectorized;
2011 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2014 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2015 StoreInst *SI = dyn_cast<StoreInst>(it);
2019 // Don't touch volatile stores.
2020 if (!SI->isSimple())
2023 // Check that the pointer points to scalars.
2024 Type *Ty = SI->getValueOperand()->getType();
2025 if (Ty->isAggregateType() || Ty->isVectorTy())
2028 // Find the base pointer.
2029 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2031 // Save the store locations.
2032 StoreRefs[Ptr].push_back(SI);
2038 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2041 Value *VL[] = { A, B };
2042 return tryToVectorizeList(VL, R);
2045 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
2049 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2051 // Check that all of the parts are scalar instructions of the same type.
2052 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2056 unsigned Opcode0 = I0->getOpcode();
2058 Type *Ty0 = I0->getType();
2059 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2060 unsigned VF = MinVecRegSize / Sz;
2062 for (int i = 0, e = VL.size(); i < e; ++i) {
2063 Type *Ty = VL[i]->getType();
2064 if (Ty->isAggregateType() || Ty->isVectorTy())
2066 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2067 if (!Inst || Inst->getOpcode() != Opcode0)
2071 bool Changed = false;
2073 // Keep track of values that were delete by vectorizing in the loop below.
2074 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2076 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2077 unsigned OpsWidth = 0;
2084 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2087 // Check that a previous iteration of this loop did not delete the Value.
2088 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2091 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2093 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2096 int Cost = R.getTreeCost();
2098 if (Cost < -SLPCostThreshold) {
2099 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n");
2102 // Move to the next bundle.
2111 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2115 // Try to vectorize V.
2116 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2119 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2120 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2122 if (B && B->hasOneUse()) {
2123 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2124 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2125 if (tryToVectorizePair(A, B0, R)) {
2129 if (tryToVectorizePair(A, B1, R)) {
2136 if (A && A->hasOneUse()) {
2137 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2138 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2139 if (tryToVectorizePair(A0, B, R)) {
2143 if (tryToVectorizePair(A1, B, R)) {
2151 /// \brief Generate a shuffle mask to be used in a reduction tree.
2153 /// \param VecLen The length of the vector to be reduced.
2154 /// \param NumEltsToRdx The number of elements that should be reduced in the
2156 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2157 /// reduction. A pairwise reduction will generate a mask of
2158 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2159 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2160 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2161 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2162 bool IsPairwise, bool IsLeft,
2163 IRBuilder<> &Builder) {
2164 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2166 SmallVector<Constant *, 32> ShuffleMask(
2167 VecLen, UndefValue::get(Builder.getInt32Ty()));
2170 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2171 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2172 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2174 // Move the upper half of the vector to the lower half.
2175 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2176 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2178 return ConstantVector::get(ShuffleMask);
2182 /// Model horizontal reductions.
2184 /// A horizontal reduction is a tree of reduction operations (currently add and
2185 /// fadd) that has operations that can be put into a vector as its leaf.
2186 /// For example, this tree:
2193 /// This tree has "mul" as its reduced values and "+" as its reduction
2194 /// operations. A reduction might be feeding into a store or a binary operation
2209 class HorizontalReduction {
2210 SmallPtrSet<Value *, 16> ReductionOps;
2211 SmallVector<Value *, 32> ReducedVals;
2213 BinaryOperator *ReductionRoot;
2214 PHINode *ReductionPHI;
2216 /// The opcode of the reduction.
2217 unsigned ReductionOpcode;
2218 /// The opcode of the values we perform a reduction on.
2219 unsigned ReducedValueOpcode;
2220 /// The width of one full horizontal reduction operation.
2221 unsigned ReduxWidth;
2222 /// Should we model this reduction as a pairwise reduction tree or a tree that
2223 /// splits the vector in halves and adds those halves.
2224 bool IsPairwiseReduction;
2227 HorizontalReduction()
2228 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0),
2229 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2231 /// \brief Try to find a reduction tree.
2232 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2235 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2236 "Thi phi needs to use the binary operator");
2238 // We could have a initial reductions that is not an add.
2239 // r *= v1 + v2 + v3 + v4
2240 // In such a case start looking for a tree rooted in the first '+'.
2242 if (B->getOperand(0) == Phi) {
2244 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2245 } else if (B->getOperand(1) == Phi) {
2247 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2254 Type *Ty = B->getType();
2255 if (Ty->isVectorTy())
2258 ReductionOpcode = B->getOpcode();
2259 ReducedValueOpcode = 0;
2260 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2267 // We currently only support adds.
2268 if (ReductionOpcode != Instruction::Add &&
2269 ReductionOpcode != Instruction::FAdd)
2272 // Post order traverse the reduction tree starting at B. We only handle true
2273 // trees containing only binary operators.
2274 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2275 Stack.push_back(std::make_pair(B, 0));
2276 while (!Stack.empty()) {
2277 BinaryOperator *TreeN = Stack.back().first;
2278 unsigned EdgeToVist = Stack.back().second++;
2279 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2281 // Only handle trees in the current basic block.
2282 if (TreeN->getParent() != B->getParent())
2285 // Each tree node needs to have one user except for the ultimate
2287 if (!TreeN->hasOneUse() && TreeN != B)
2291 if (EdgeToVist == 2 || IsReducedValue) {
2292 if (IsReducedValue) {
2293 // Make sure that the opcodes of the operations that we are going to
2295 if (!ReducedValueOpcode)
2296 ReducedValueOpcode = TreeN->getOpcode();
2297 else if (ReducedValueOpcode != TreeN->getOpcode())
2299 ReducedVals.push_back(TreeN);
2301 // We need to be able to reassociate the adds.
2302 if (!TreeN->isAssociative())
2304 ReductionOps.insert(TreeN);
2311 // Visit left or right.
2312 Value *NextV = TreeN->getOperand(EdgeToVist);
2313 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2315 Stack.push_back(std::make_pair(Next, 0));
2316 else if (NextV != Phi)
2322 /// \brief Attempt to vectorize the tree found by
2323 /// matchAssociativeReduction.
2324 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2325 if (ReducedVals.empty())
2328 unsigned NumReducedVals = ReducedVals.size();
2329 if (NumReducedVals < ReduxWidth)
2332 Value *VectorizedTree = 0;
2333 IRBuilder<> Builder(ReductionRoot);
2334 FastMathFlags Unsafe;
2335 Unsafe.setUnsafeAlgebra();
2336 Builder.SetFastMathFlags(Unsafe);
2339 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2340 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2341 V.buildTree(ValsToReduce, &ReductionOps);
2344 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2345 if (Cost >= -SLPCostThreshold)
2348 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2351 // Vectorize a tree.
2352 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2353 Value *VectorizedRoot = V.vectorizeTree();
2355 // Emit a reduction.
2356 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2357 if (VectorizedTree) {
2358 Builder.SetCurrentDebugLocation(Loc);
2359 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2360 ReducedSubTree, "bin.rdx");
2362 VectorizedTree = ReducedSubTree;
2365 if (VectorizedTree) {
2366 // Finish the reduction.
2367 for (; i < NumReducedVals; ++i) {
2368 Builder.SetCurrentDebugLocation(
2369 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2370 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2375 assert(ReductionRoot != NULL && "Need a reduction operation");
2376 ReductionRoot->setOperand(0, VectorizedTree);
2377 ReductionRoot->setOperand(1, ReductionPHI);
2379 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2381 return VectorizedTree != 0;
2386 /// \brief Calcuate the cost of a reduction.
2387 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2388 Type *ScalarTy = FirstReducedVal->getType();
2389 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2391 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2392 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2394 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2395 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2397 int ScalarReduxCost =
2398 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2400 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2401 << " for reduction that starts with " << *FirstReducedVal
2403 << (IsPairwiseReduction ? "pairwise" : "splitting")
2404 << " reduction)\n");
2406 return VecReduxCost - ScalarReduxCost;
2409 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2410 Value *R, const Twine &Name = "") {
2411 if (Opcode == Instruction::FAdd)
2412 return Builder.CreateFAdd(L, R, Name);
2413 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2416 /// \brief Emit a horizontal reduction of the vectorized value.
2417 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2418 assert(VectorizedValue && "Need to have a vectorized tree node");
2419 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2420 assert(isPowerOf2_32(ReduxWidth) &&
2421 "We only handle power-of-two reductions for now");
2423 Value *TmpVec = ValToReduce;
2424 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2425 if (IsPairwiseReduction) {
2427 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2429 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2431 Value *LeftShuf = Builder.CreateShuffleVector(
2432 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2433 Value *RightShuf = Builder.CreateShuffleVector(
2434 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2436 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2440 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2441 Value *Shuf = Builder.CreateShuffleVector(
2442 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2443 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2447 // The result is in the first element of the vector.
2448 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2452 /// \brief Recognize construction of vectors like
2453 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2454 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2455 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2456 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2458 /// Returns true if it matches
2460 static bool findBuildVector(InsertElementInst *IE,
2461 SmallVectorImpl<Value *> &Ops) {
2462 if (!isa<UndefValue>(IE->getOperand(0)))
2466 Ops.push_back(IE->getOperand(1));
2468 if (IE->use_empty())
2471 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->use_back());
2475 // If this isn't the final use, make sure the next insertelement is the only
2476 // use. It's OK if the final constructed vector is used multiple times
2477 if (!IE->hasOneUse())
2486 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2487 return V->getType() < V2->getType();
2490 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2491 bool Changed = false;
2492 SmallVector<Value *, 4> Incoming;
2493 SmallSet<Value *, 16> VisitedInstrs;
2495 bool HaveVectorizedPhiNodes = true;
2496 while (HaveVectorizedPhiNodes) {
2497 HaveVectorizedPhiNodes = false;
2499 // Collect the incoming values from the PHIs.
2501 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2503 PHINode *P = dyn_cast<PHINode>(instr);
2507 if (!VisitedInstrs.count(P))
2508 Incoming.push_back(P);
2512 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2514 // Try to vectorize elements base on their type.
2515 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2519 // Look for the next elements with the same type.
2520 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2521 while (SameTypeIt != E &&
2522 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2523 VisitedInstrs.insert(*SameTypeIt);
2527 // Try to vectorize them.
2528 unsigned NumElts = (SameTypeIt - IncIt);
2529 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2531 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2532 // Success start over because instructions might have been changed.
2533 HaveVectorizedPhiNodes = true;
2538 // Start over at the next instruction of a different type (or the end).
2543 VisitedInstrs.clear();
2545 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2546 // We may go through BB multiple times so skip the one we have checked.
2547 if (!VisitedInstrs.insert(it))
2550 if (isa<DbgInfoIntrinsic>(it))
2553 // Try to vectorize reductions that use PHINodes.
2554 if (PHINode *P = dyn_cast<PHINode>(it)) {
2555 // Check that the PHI is a reduction PHI.
2556 if (P->getNumIncomingValues() != 2)
2559 (P->getIncomingBlock(0) == BB
2560 ? (P->getIncomingValue(0))
2561 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
2562 // Check if this is a Binary Operator.
2563 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2567 // Try to match and vectorize a horizontal reduction.
2568 HorizontalReduction HorRdx;
2569 if (ShouldVectorizeHor &&
2570 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2571 HorRdx.tryToReduce(R, TTI)) {
2578 Value *Inst = BI->getOperand(0);
2580 Inst = BI->getOperand(1);
2582 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2583 // We would like to start over since some instructions are deleted
2584 // and the iterator may become invalid value.
2594 // Try to vectorize horizontal reductions feeding into a store.
2595 if (ShouldStartVectorizeHorAtStore)
2596 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2597 if (BinaryOperator *BinOp =
2598 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2599 HorizontalReduction HorRdx;
2600 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) &&
2601 HorRdx.tryToReduce(R, TTI)) ||
2602 tryToVectorize(BinOp, R))) {
2610 // Try to vectorize trees that start at compare instructions.
2611 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2612 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2614 // We would like to start over since some instructions are deleted
2615 // and the iterator may become invalid value.
2621 for (int i = 0; i < 2; ++i) {
2622 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2623 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2625 // We would like to start over since some instructions are deleted
2626 // and the iterator may become invalid value.
2635 // Try to vectorize trees that start at insertelement instructions.
2636 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2637 SmallVector<Value *, 8> Ops;
2638 if (!findBuildVector(IE, Ops))
2641 if (tryToVectorizeList(Ops, R)) {
2654 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2655 bool Changed = false;
2656 // Attempt to sort and vectorize each of the store-groups.
2657 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2659 if (it->second.size() < 2)
2662 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2663 << it->second.size() << ".\n");
2665 // Process the stores in chunks of 16.
2666 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2667 unsigned Len = std::min<unsigned>(CE - CI, 16);
2668 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2669 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2675 } // end anonymous namespace
2677 char SLPVectorizer::ID = 0;
2678 static const char lv_name[] = "SLP Vectorizer";
2679 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2680 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2681 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2682 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2683 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2684 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2687 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }