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>(cast<PHINode>(VL[j])->getIncomingValue(i));
784 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
785 newTreeEntry(VL, false);
790 newTreeEntry(VL, true);
791 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
793 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
795 // Prepare the operand vector.
796 for (unsigned j = 0; j < VL.size(); ++j)
797 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValue(i));
799 buildTree_rec(Operands, Depth + 1);
803 case Instruction::ExtractElement: {
804 bool Reuse = CanReuseExtract(VL);
806 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
808 newTreeEntry(VL, Reuse);
811 case Instruction::Load: {
812 // Check if the loads are consecutive or of we need to swizzle them.
813 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
814 LoadInst *L = cast<LoadInst>(VL[i]);
815 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
816 newTreeEntry(VL, false);
817 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
821 newTreeEntry(VL, true);
822 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
825 case Instruction::ZExt:
826 case Instruction::SExt:
827 case Instruction::FPToUI:
828 case Instruction::FPToSI:
829 case Instruction::FPExt:
830 case Instruction::PtrToInt:
831 case Instruction::IntToPtr:
832 case Instruction::SIToFP:
833 case Instruction::UIToFP:
834 case Instruction::Trunc:
835 case Instruction::FPTrunc:
836 case Instruction::BitCast: {
837 Type *SrcTy = VL0->getOperand(0)->getType();
838 for (unsigned i = 0; i < VL.size(); ++i) {
839 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
840 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
841 newTreeEntry(VL, false);
842 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
846 newTreeEntry(VL, true);
847 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
849 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
851 // Prepare the operand vector.
852 for (unsigned j = 0; j < VL.size(); ++j)
853 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
855 buildTree_rec(Operands, Depth+1);
859 case Instruction::ICmp:
860 case Instruction::FCmp: {
861 // Check that all of the compares have the same predicate.
862 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
863 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
864 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
865 CmpInst *Cmp = cast<CmpInst>(VL[i]);
866 if (Cmp->getPredicate() != P0 ||
867 Cmp->getOperand(0)->getType() != ComparedTy) {
868 newTreeEntry(VL, false);
869 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
874 newTreeEntry(VL, true);
875 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
877 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
879 // Prepare the operand vector.
880 for (unsigned j = 0; j < VL.size(); ++j)
881 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
883 buildTree_rec(Operands, Depth+1);
887 case Instruction::Select:
888 case Instruction::Add:
889 case Instruction::FAdd:
890 case Instruction::Sub:
891 case Instruction::FSub:
892 case Instruction::Mul:
893 case Instruction::FMul:
894 case Instruction::UDiv:
895 case Instruction::SDiv:
896 case Instruction::FDiv:
897 case Instruction::URem:
898 case Instruction::SRem:
899 case Instruction::FRem:
900 case Instruction::Shl:
901 case Instruction::LShr:
902 case Instruction::AShr:
903 case Instruction::And:
904 case Instruction::Or:
905 case Instruction::Xor: {
906 newTreeEntry(VL, true);
907 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
909 // Sort operands of the instructions so that each side is more likely to
910 // have the same opcode.
911 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
912 ValueList Left, Right;
913 reorderInputsAccordingToOpcode(VL, Left, Right);
914 buildTree_rec(Left, Depth + 1);
915 buildTree_rec(Right, Depth + 1);
919 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
921 // Prepare the operand vector.
922 for (unsigned j = 0; j < VL.size(); ++j)
923 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
925 buildTree_rec(Operands, Depth+1);
929 case Instruction::Store: {
930 // Check if the stores are consecutive or of we need to swizzle them.
931 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
932 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
933 newTreeEntry(VL, false);
934 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
938 newTreeEntry(VL, true);
939 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
942 for (unsigned j = 0; j < VL.size(); ++j)
943 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
945 // We can ignore these values because we are sinking them down.
946 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
947 buildTree_rec(Operands, Depth + 1);
950 case Instruction::Call: {
951 // Check if the calls are all to the same vectorizable intrinsic.
952 IntrinsicInst *II = dyn_cast<IntrinsicInst>(VL[0]);
954 newTreeEntry(VL, false);
955 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
959 Intrinsic::ID ID = II->getIntrinsicID();
961 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
962 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[i]);
963 if (!II2 || II2->getIntrinsicID() != ID) {
964 newTreeEntry(VL, false);
965 DEBUG(dbgs() << "SLP: mismatched calls:" << *II << "!=" << *VL[i]
971 newTreeEntry(VL, true);
972 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i) {
974 // Prepare the operand vector.
975 for (unsigned j = 0; j < VL.size(); ++j) {
976 IntrinsicInst *II2 = dyn_cast<IntrinsicInst>(VL[j]);
977 Operands.push_back(II2->getArgOperand(i));
979 buildTree_rec(Operands, Depth + 1);
984 newTreeEntry(VL, false);
985 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
990 int BoUpSLP::getEntryCost(TreeEntry *E) {
991 ArrayRef<Value*> VL = E->Scalars;
993 Type *ScalarTy = VL[0]->getType();
994 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
995 ScalarTy = SI->getValueOperand()->getType();
996 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
998 if (E->NeedToGather) {
1002 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1004 return getGatherCost(E->Scalars);
1007 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1009 Instruction *VL0 = cast<Instruction>(VL[0]);
1010 unsigned Opcode = VL0->getOpcode();
1012 case Instruction::PHI: {
1015 case Instruction::ExtractElement: {
1016 if (CanReuseExtract(VL))
1018 return getGatherCost(VecTy);
1020 case Instruction::ZExt:
1021 case Instruction::SExt:
1022 case Instruction::FPToUI:
1023 case Instruction::FPToSI:
1024 case Instruction::FPExt:
1025 case Instruction::PtrToInt:
1026 case Instruction::IntToPtr:
1027 case Instruction::SIToFP:
1028 case Instruction::UIToFP:
1029 case Instruction::Trunc:
1030 case Instruction::FPTrunc:
1031 case Instruction::BitCast: {
1032 Type *SrcTy = VL0->getOperand(0)->getType();
1034 // Calculate the cost of this instruction.
1035 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1036 VL0->getType(), SrcTy);
1038 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1039 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1040 return VecCost - ScalarCost;
1042 case Instruction::FCmp:
1043 case Instruction::ICmp:
1044 case Instruction::Select:
1045 case Instruction::Add:
1046 case Instruction::FAdd:
1047 case Instruction::Sub:
1048 case Instruction::FSub:
1049 case Instruction::Mul:
1050 case Instruction::FMul:
1051 case Instruction::UDiv:
1052 case Instruction::SDiv:
1053 case Instruction::FDiv:
1054 case Instruction::URem:
1055 case Instruction::SRem:
1056 case Instruction::FRem:
1057 case Instruction::Shl:
1058 case Instruction::LShr:
1059 case Instruction::AShr:
1060 case Instruction::And:
1061 case Instruction::Or:
1062 case Instruction::Xor: {
1063 // Calculate the cost of this instruction.
1066 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1067 Opcode == Instruction::Select) {
1068 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1069 ScalarCost = VecTy->getNumElements() *
1070 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1071 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1073 // Certain instructions can be cheaper to vectorize if they have a
1074 // constant second vector operand.
1075 TargetTransformInfo::OperandValueKind Op1VK =
1076 TargetTransformInfo::OK_AnyValue;
1077 TargetTransformInfo::OperandValueKind Op2VK =
1078 TargetTransformInfo::OK_UniformConstantValue;
1080 // Check whether all second operands are constant.
1081 for (unsigned i = 0; i < VL.size(); ++i)
1082 if (!isa<ConstantInt>(cast<Instruction>(VL[i])->getOperand(1))) {
1083 Op2VK = TargetTransformInfo::OK_AnyValue;
1088 VecTy->getNumElements() *
1089 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1090 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1092 return VecCost - ScalarCost;
1094 case Instruction::Load: {
1095 // Cost of wide load - cost of scalar loads.
1096 int ScalarLdCost = VecTy->getNumElements() *
1097 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1098 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1099 return VecLdCost - ScalarLdCost;
1101 case Instruction::Store: {
1102 // We know that we can merge the stores. Calculate the cost.
1103 int ScalarStCost = VecTy->getNumElements() *
1104 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1105 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1106 return VecStCost - ScalarStCost;
1108 case Instruction::Call: {
1109 CallInst *CI = cast<CallInst>(VL0);
1110 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1111 Intrinsic::ID ID = II->getIntrinsicID();
1113 // Calculate the cost of the scalar and vector calls.
1114 SmallVector<Type*, 4> ScalarTys, VecTys;
1115 for (unsigned op = 0, opc = II->getNumArgOperands(); op!= opc; ++op) {
1116 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1117 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1118 VecTy->getNumElements()));
1121 int ScalarCallCost = VecTy->getNumElements() *
1122 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1124 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1126 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1127 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1128 << " for " << *II << "\n");
1130 return VecCallCost - ScalarCallCost;
1133 llvm_unreachable("Unknown instruction");
1137 bool BoUpSLP::isFullyVectorizableTinyTree() {
1138 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1139 VectorizableTree.size() << " is fully vectorizable .\n");
1141 // We only handle trees of height 2.
1142 if (VectorizableTree.size() != 2)
1145 // Gathering cost would be too much for tiny trees.
1146 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1152 int BoUpSLP::getTreeCost() {
1154 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1155 VectorizableTree.size() << ".\n");
1157 // We only vectorize tiny trees if it is fully vectorizable.
1158 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1159 if (!VectorizableTree.size()) {
1160 assert(!ExternalUses.size() && "We should not have any external users");
1165 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1167 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1168 int C = getEntryCost(&VectorizableTree[i]);
1169 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1170 << *VectorizableTree[i].Scalars[0] << " .\n");
1174 SmallSet<Value *, 16> ExtractCostCalculated;
1175 int ExtractCost = 0;
1176 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1178 // We only add extract cost once for the same scalar.
1179 if (!ExtractCostCalculated.insert(I->Scalar))
1182 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1183 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1187 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1188 return Cost + ExtractCost;
1191 int BoUpSLP::getGatherCost(Type *Ty) {
1193 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1194 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1198 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1199 // Find the type of the operands in VL.
1200 Type *ScalarTy = VL[0]->getType();
1201 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1202 ScalarTy = SI->getValueOperand()->getType();
1203 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1204 // Find the cost of inserting/extracting values from the vector.
1205 return getGatherCost(VecTy);
1208 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1209 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1210 return AA->getLocation(SI);
1211 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1212 return AA->getLocation(LI);
1213 return AliasAnalysis::Location();
1216 Value *BoUpSLP::getPointerOperand(Value *I) {
1217 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1218 return LI->getPointerOperand();
1219 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1220 return SI->getPointerOperand();
1224 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1225 if (LoadInst *L = dyn_cast<LoadInst>(I))
1226 return L->getPointerAddressSpace();
1227 if (StoreInst *S = dyn_cast<StoreInst>(I))
1228 return S->getPointerAddressSpace();
1232 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1233 Value *PtrA = getPointerOperand(A);
1234 Value *PtrB = getPointerOperand(B);
1235 unsigned ASA = getAddressSpaceOperand(A);
1236 unsigned ASB = getAddressSpaceOperand(B);
1238 // Check that the address spaces match and that the pointers are valid.
1239 if (!PtrA || !PtrB || (ASA != ASB))
1242 // Make sure that A and B are different pointers of the same type.
1243 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1246 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1247 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1248 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1250 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1251 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1252 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1254 APInt OffsetDelta = OffsetB - OffsetA;
1256 // Check if they are based on the same pointer. That makes the offsets
1259 return OffsetDelta == Size;
1261 // Compute the necessary base pointer delta to have the necessary final delta
1262 // equal to the size.
1263 APInt BaseDelta = Size - OffsetDelta;
1265 // Otherwise compute the distance with SCEV between the base pointers.
1266 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1267 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1268 const SCEV *C = SE->getConstant(BaseDelta);
1269 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1270 return X == PtrSCEVB;
1273 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1274 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1275 BasicBlock::iterator I = Src, E = Dst;
1276 /// Scan all of the instruction from SRC to DST and check if
1277 /// the source may alias.
1278 for (++I; I != E; ++I) {
1279 // Ignore store instructions that are marked as 'ignore'.
1280 if (MemBarrierIgnoreList.count(I))
1282 if (Src->mayWriteToMemory()) /* Write */ {
1283 if (!I->mayReadOrWriteMemory())
1286 if (!I->mayWriteToMemory())
1289 AliasAnalysis::Location A = getLocation(&*I);
1290 AliasAnalysis::Location B = getLocation(Src);
1292 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1298 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1299 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1300 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1301 BlockNumbering &BN = BlocksNumbers[BB];
1303 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1304 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1305 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1309 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1310 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1311 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1312 BlockNumbering &BN = BlocksNumbers[BB];
1314 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1315 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1316 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1317 Instruction *I = BN.getInstruction(MaxIdx);
1318 assert(I && "bad location");
1322 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1323 Instruction *VL0 = cast<Instruction>(VL[0]);
1324 Instruction *LastInst = getLastInstruction(VL);
1325 BasicBlock::iterator NextInst = LastInst;
1327 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1328 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1331 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1332 Value *Vec = UndefValue::get(Ty);
1333 // Generate the 'InsertElement' instruction.
1334 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1335 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1336 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1337 GatherSeq.insert(Insrt);
1338 CSEBlocks.insert(Insrt->getParent());
1340 // Add to our 'need-to-extract' list.
1341 if (ScalarToTreeEntry.count(VL[i])) {
1342 int Idx = ScalarToTreeEntry[VL[i]];
1343 TreeEntry *E = &VectorizableTree[Idx];
1344 // Find which lane we need to extract.
1346 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1347 // Is this the lane of the scalar that we are looking for ?
1348 if (E->Scalars[Lane] == VL[i]) {
1353 assert(FoundLane >= 0 && "Could not find the correct lane");
1354 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1362 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1363 SmallDenseMap<Value*, int>::const_iterator Entry
1364 = ScalarToTreeEntry.find(VL[0]);
1365 if (Entry != ScalarToTreeEntry.end()) {
1366 int Idx = Entry->second;
1367 const TreeEntry *En = &VectorizableTree[Idx];
1368 if (En->isSame(VL) && En->VectorizedValue)
1369 return En->VectorizedValue;
1374 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1375 if (ScalarToTreeEntry.count(VL[0])) {
1376 int Idx = ScalarToTreeEntry[VL[0]];
1377 TreeEntry *E = &VectorizableTree[Idx];
1379 return vectorizeTree(E);
1382 Type *ScalarTy = VL[0]->getType();
1383 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1384 ScalarTy = SI->getValueOperand()->getType();
1385 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1387 return Gather(VL, VecTy);
1390 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1391 IRBuilder<>::InsertPointGuard Guard(Builder);
1393 if (E->VectorizedValue) {
1394 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1395 return E->VectorizedValue;
1398 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1399 Type *ScalarTy = VL0->getType();
1400 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1401 ScalarTy = SI->getValueOperand()->getType();
1402 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1404 if (E->NeedToGather) {
1405 setInsertPointAfterBundle(E->Scalars);
1406 return Gather(E->Scalars, VecTy);
1409 unsigned Opcode = VL0->getOpcode();
1410 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1413 case Instruction::PHI: {
1414 PHINode *PH = dyn_cast<PHINode>(VL0);
1415 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1416 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1417 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1418 E->VectorizedValue = NewPhi;
1420 // PHINodes may have multiple entries from the same block. We want to
1421 // visit every block once.
1422 SmallSet<BasicBlock*, 4> VisitedBBs;
1424 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1426 BasicBlock *IBB = PH->getIncomingBlock(i);
1428 if (!VisitedBBs.insert(IBB)) {
1429 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1433 // Prepare the operand vector.
1434 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1435 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1436 getIncomingValueForBlock(IBB));
1438 Builder.SetInsertPoint(IBB->getTerminator());
1439 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1440 Value *Vec = vectorizeTree(Operands);
1441 NewPhi->addIncoming(Vec, IBB);
1444 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1445 "Invalid number of incoming values");
1449 case Instruction::ExtractElement: {
1450 if (CanReuseExtract(E->Scalars)) {
1451 Value *V = VL0->getOperand(0);
1452 E->VectorizedValue = V;
1455 return Gather(E->Scalars, VecTy);
1457 case Instruction::ZExt:
1458 case Instruction::SExt:
1459 case Instruction::FPToUI:
1460 case Instruction::FPToSI:
1461 case Instruction::FPExt:
1462 case Instruction::PtrToInt:
1463 case Instruction::IntToPtr:
1464 case Instruction::SIToFP:
1465 case Instruction::UIToFP:
1466 case Instruction::Trunc:
1467 case Instruction::FPTrunc:
1468 case Instruction::BitCast: {
1470 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1471 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1473 setInsertPointAfterBundle(E->Scalars);
1475 Value *InVec = vectorizeTree(INVL);
1477 if (Value *V = alreadyVectorized(E->Scalars))
1480 CastInst *CI = dyn_cast<CastInst>(VL0);
1481 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1482 E->VectorizedValue = V;
1485 case Instruction::FCmp:
1486 case Instruction::ICmp: {
1487 ValueList LHSV, RHSV;
1488 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1489 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1490 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1493 setInsertPointAfterBundle(E->Scalars);
1495 Value *L = vectorizeTree(LHSV);
1496 Value *R = vectorizeTree(RHSV);
1498 if (Value *V = alreadyVectorized(E->Scalars))
1501 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1503 if (Opcode == Instruction::FCmp)
1504 V = Builder.CreateFCmp(P0, L, R);
1506 V = Builder.CreateICmp(P0, L, R);
1508 E->VectorizedValue = V;
1511 case Instruction::Select: {
1512 ValueList TrueVec, FalseVec, CondVec;
1513 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1514 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1515 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1516 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1519 setInsertPointAfterBundle(E->Scalars);
1521 Value *Cond = vectorizeTree(CondVec);
1522 Value *True = vectorizeTree(TrueVec);
1523 Value *False = vectorizeTree(FalseVec);
1525 if (Value *V = alreadyVectorized(E->Scalars))
1528 Value *V = Builder.CreateSelect(Cond, True, False);
1529 E->VectorizedValue = V;
1532 case Instruction::Add:
1533 case Instruction::FAdd:
1534 case Instruction::Sub:
1535 case Instruction::FSub:
1536 case Instruction::Mul:
1537 case Instruction::FMul:
1538 case Instruction::UDiv:
1539 case Instruction::SDiv:
1540 case Instruction::FDiv:
1541 case Instruction::URem:
1542 case Instruction::SRem:
1543 case Instruction::FRem:
1544 case Instruction::Shl:
1545 case Instruction::LShr:
1546 case Instruction::AShr:
1547 case Instruction::And:
1548 case Instruction::Or:
1549 case Instruction::Xor: {
1550 ValueList LHSVL, RHSVL;
1551 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1552 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1554 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1555 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1556 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1559 setInsertPointAfterBundle(E->Scalars);
1561 Value *LHS = vectorizeTree(LHSVL);
1562 Value *RHS = vectorizeTree(RHSVL);
1564 if (LHS == RHS && isa<Instruction>(LHS)) {
1565 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1568 if (Value *V = alreadyVectorized(E->Scalars))
1571 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1572 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1573 E->VectorizedValue = V;
1575 if (Instruction *I = dyn_cast<Instruction>(V))
1576 return propagateMetadata(I, E->Scalars);
1580 case Instruction::Load: {
1581 // Loads are inserted at the head of the tree because we don't want to
1582 // sink them all the way down past store instructions.
1583 setInsertPointAfterBundle(E->Scalars);
1585 LoadInst *LI = cast<LoadInst>(VL0);
1586 unsigned AS = LI->getPointerAddressSpace();
1588 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1589 VecTy->getPointerTo(AS));
1590 unsigned Alignment = LI->getAlignment();
1591 LI = Builder.CreateLoad(VecPtr);
1592 LI->setAlignment(Alignment);
1593 E->VectorizedValue = LI;
1594 return propagateMetadata(LI, E->Scalars);
1596 case Instruction::Store: {
1597 StoreInst *SI = cast<StoreInst>(VL0);
1598 unsigned Alignment = SI->getAlignment();
1599 unsigned AS = SI->getPointerAddressSpace();
1602 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1603 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1605 setInsertPointAfterBundle(E->Scalars);
1607 Value *VecValue = vectorizeTree(ValueOp);
1608 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1609 VecTy->getPointerTo(AS));
1610 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1611 S->setAlignment(Alignment);
1612 E->VectorizedValue = S;
1613 return propagateMetadata(S, E->Scalars);
1615 case Instruction::Call: {
1616 CallInst *CI = cast<CallInst>(VL0);
1618 setInsertPointAfterBundle(E->Scalars);
1619 std::vector<Value *> OpVecs;
1620 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1622 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1623 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1624 OpVL.push_back(CEI->getArgOperand(j));
1627 Value *OpVec = vectorizeTree(OpVL);
1628 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1629 OpVecs.push_back(OpVec);
1632 Module *M = F->getParent();
1633 IntrinsicInst *II = cast<IntrinsicInst>(CI);
1634 Intrinsic::ID ID = II->getIntrinsicID();
1635 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
1636 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
1637 Value *V = Builder.CreateCall(CF, OpVecs);
1638 E->VectorizedValue = V;
1642 llvm_unreachable("unknown inst");
1647 Value *BoUpSLP::vectorizeTree() {
1648 Builder.SetInsertPoint(F->getEntryBlock().begin());
1649 vectorizeTree(&VectorizableTree[0]);
1651 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1653 // Extract all of the elements with the external uses.
1654 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1656 Value *Scalar = it->Scalar;
1657 llvm::User *User = it->User;
1659 // Skip users that we already RAUW. This happens when one instruction
1660 // has multiple uses of the same value.
1661 if (std::find(Scalar->use_begin(), Scalar->use_end(), User) ==
1664 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1666 int Idx = ScalarToTreeEntry[Scalar];
1667 TreeEntry *E = &VectorizableTree[Idx];
1668 assert(!E->NeedToGather && "Extracting from a gather list");
1670 Value *Vec = E->VectorizedValue;
1671 assert(Vec && "Can't find vectorizable value");
1673 Value *Lane = Builder.getInt32(it->Lane);
1674 // Generate extracts for out-of-tree users.
1675 // Find the insertion point for the extractelement lane.
1676 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1677 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt());
1678 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1679 CSEBlocks.insert(PN->getParent());
1680 User->replaceUsesOfWith(Scalar, Ex);
1681 } else if (isa<Instruction>(Vec)){
1682 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1683 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1684 if (PH->getIncomingValue(i) == Scalar) {
1685 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1686 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1687 CSEBlocks.insert(PH->getIncomingBlock(i));
1688 PH->setOperand(i, Ex);
1692 Builder.SetInsertPoint(cast<Instruction>(User));
1693 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1694 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1695 User->replaceUsesOfWith(Scalar, Ex);
1698 Builder.SetInsertPoint(F->getEntryBlock().begin());
1699 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1700 CSEBlocks.insert(&F->getEntryBlock());
1701 User->replaceUsesOfWith(Scalar, Ex);
1704 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1707 // For each vectorized value:
1708 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1709 TreeEntry *Entry = &VectorizableTree[EIdx];
1712 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1713 Value *Scalar = Entry->Scalars[Lane];
1715 // No need to handle users of gathered values.
1716 if (Entry->NeedToGather)
1719 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1721 Type *Ty = Scalar->getType();
1722 if (!Ty->isVoidTy()) {
1723 for (Value::use_iterator User = Scalar->use_begin(),
1724 UE = Scalar->use_end(); User != UE; ++User) {
1725 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n");
1727 assert((ScalarToTreeEntry.count(*User) ||
1728 // It is legal to replace the reduction users by undef.
1729 (RdxOps && RdxOps->count(*User))) &&
1730 "Replacing out-of-tree value with undef");
1732 Value *Undef = UndefValue::get(Ty);
1733 Scalar->replaceAllUsesWith(Undef);
1735 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1736 cast<Instruction>(Scalar)->eraseFromParent();
1740 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1741 BlocksNumbers[it].forget();
1743 Builder.ClearInsertionPoint();
1745 return VectorizableTree[0].VectorizedValue;
1749 const DominatorTree *DT;
1752 DTCmp(const DominatorTree *DT) : DT(DT) {}
1753 bool operator()(const BasicBlock *A, const BasicBlock *B) const {
1754 return DT->properlyDominates(A, B);
1758 void BoUpSLP::optimizeGatherSequence() {
1759 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1760 << " gather sequences instructions.\n");
1761 // LICM InsertElementInst sequences.
1762 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1763 e = GatherSeq.end(); it != e; ++it) {
1764 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1769 // Check if this block is inside a loop.
1770 Loop *L = LI->getLoopFor(Insert->getParent());
1774 // Check if it has a preheader.
1775 BasicBlock *PreHeader = L->getLoopPreheader();
1779 // If the vector or the element that we insert into it are
1780 // instructions that are defined in this basic block then we can't
1781 // hoist this instruction.
1782 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1783 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1784 if (CurrVec && L->contains(CurrVec))
1786 if (NewElem && L->contains(NewElem))
1789 // We can hoist this instruction. Move it to the pre-header.
1790 Insert->moveBefore(PreHeader->getTerminator());
1793 // Sort blocks by domination. This ensures we visit a block after all blocks
1794 // dominating it are visited.
1795 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1796 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), DTCmp(DT));
1798 // Perform O(N^2) search over the gather sequences and merge identical
1799 // instructions. TODO: We can further optimize this scan if we split the
1800 // instructions into different buckets based on the insert lane.
1801 SmallVector<Instruction *, 16> Visited;
1802 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1803 E = CSEWorkList.end();
1805 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *llvm::prior(I))) &&
1806 "Worklist not sorted properly!");
1807 BasicBlock *BB = *I;
1808 // For all instructions in blocks containing gather sequences:
1809 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1810 Instruction *In = it++;
1811 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1814 // Check if we can replace this instruction with any of the
1815 // visited instructions.
1816 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1819 if (In->isIdenticalTo(*v) &&
1820 DT->dominates((*v)->getParent(), In->getParent())) {
1821 In->replaceAllUsesWith(*v);
1822 In->eraseFromParent();
1828 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1829 Visited.push_back(In);
1837 /// The SLPVectorizer Pass.
1838 struct SLPVectorizer : public FunctionPass {
1839 typedef SmallVector<StoreInst *, 8> StoreList;
1840 typedef MapVector<Value *, StoreList> StoreListMap;
1842 /// Pass identification, replacement for typeid
1845 explicit SLPVectorizer() : FunctionPass(ID) {
1846 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1849 ScalarEvolution *SE;
1851 TargetTransformInfo *TTI;
1856 virtual bool runOnFunction(Function &F) {
1857 SE = &getAnalysis<ScalarEvolution>();
1858 DL = getAnalysisIfAvailable<DataLayout>();
1859 TTI = &getAnalysis<TargetTransformInfo>();
1860 AA = &getAnalysis<AliasAnalysis>();
1861 LI = &getAnalysis<LoopInfo>();
1862 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1865 bool Changed = false;
1867 // If the target claims to have no vector registers don't attempt
1869 if (!TTI->getNumberOfRegisters(true))
1872 // Must have DataLayout. We can't require it because some tests run w/o
1877 // Don't vectorize when the attribute NoImplicitFloat is used.
1878 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1881 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1883 // Use the bollom up slp vectorizer to construct chains that start with
1884 // he store instructions.
1885 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1887 // Scan the blocks in the function in post order.
1888 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1889 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1890 BasicBlock *BB = *it;
1892 // Vectorize trees that end at stores.
1893 if (unsigned count = collectStores(BB, R)) {
1895 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1896 Changed |= vectorizeStoreChains(R);
1899 // Vectorize trees that end at reductions.
1900 Changed |= vectorizeChainsInBlock(BB, R);
1904 R.optimizeGatherSequence();
1905 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1906 DEBUG(verifyFunction(F));
1911 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1912 FunctionPass::getAnalysisUsage(AU);
1913 AU.addRequired<ScalarEvolution>();
1914 AU.addRequired<AliasAnalysis>();
1915 AU.addRequired<TargetTransformInfo>();
1916 AU.addRequired<LoopInfo>();
1917 AU.addRequired<DominatorTreeWrapperPass>();
1918 AU.addPreserved<LoopInfo>();
1919 AU.addPreserved<DominatorTreeWrapperPass>();
1920 AU.setPreservesCFG();
1925 /// \brief Collect memory references and sort them according to their base
1926 /// object. We sort the stores to their base objects to reduce the cost of the
1927 /// quadratic search on the stores. TODO: We can further reduce this cost
1928 /// if we flush the chain creation every time we run into a memory barrier.
1929 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1931 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1932 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1934 /// \brief Try to vectorize a list of operands.
1935 /// \returns true if a value was vectorized.
1936 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1938 /// \brief Try to vectorize a chain that may start at the operands of \V;
1939 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1941 /// \brief Vectorize the stores that were collected in StoreRefs.
1942 bool vectorizeStoreChains(BoUpSLP &R);
1944 /// \brief Scan the basic block and look for patterns that are likely to start
1945 /// a vectorization chain.
1946 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1948 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1951 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1954 StoreListMap StoreRefs;
1957 /// \brief Check that the Values in the slice in VL array are still existent in
1958 /// the WeakVH array.
1959 /// Vectorization of part of the VL array may cause later values in the VL array
1960 /// to become invalid. We track when this has happened in the WeakVH array.
1961 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1962 SmallVectorImpl<WeakVH> &VH,
1963 unsigned SliceBegin,
1964 unsigned SliceSize) {
1965 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1972 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1973 int CostThreshold, BoUpSLP &R) {
1974 unsigned ChainLen = Chain.size();
1975 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1977 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1978 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1979 unsigned VF = MinVecRegSize / Sz;
1981 if (!isPowerOf2_32(Sz) || VF < 2)
1984 // Keep track of values that were delete by vectorizing in the loop below.
1985 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
1987 bool Changed = false;
1988 // Look for profitable vectorizable trees at all offsets, starting at zero.
1989 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1993 // Check that a previous iteration of this loop did not delete the Value.
1994 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
1997 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1999 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2001 R.buildTree(Operands);
2003 int Cost = R.getTreeCost();
2005 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2006 if (Cost < CostThreshold) {
2007 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2010 // Move to the next bundle.
2019 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2020 int costThreshold, BoUpSLP &R) {
2021 SetVector<Value *> Heads, Tails;
2022 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2024 // We may run into multiple chains that merge into a single chain. We mark the
2025 // stores that we vectorized so that we don't visit the same store twice.
2026 BoUpSLP::ValueSet VectorizedStores;
2027 bool Changed = false;
2029 // Do a quadratic search on all of the given stores and find
2030 // all of the pairs of stores that follow each other.
2031 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2032 for (unsigned j = 0; j < e; ++j) {
2036 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2037 Tails.insert(Stores[j]);
2038 Heads.insert(Stores[i]);
2039 ConsecutiveChain[Stores[i]] = Stores[j];
2044 // For stores that start but don't end a link in the chain:
2045 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2047 if (Tails.count(*it))
2050 // We found a store instr that starts a chain. Now follow the chain and try
2052 BoUpSLP::ValueList Operands;
2054 // Collect the chain into a list.
2055 while (Tails.count(I) || Heads.count(I)) {
2056 if (VectorizedStores.count(I))
2058 Operands.push_back(I);
2059 // Move to the next value in the chain.
2060 I = ConsecutiveChain[I];
2063 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2065 // Mark the vectorized stores so that we don't vectorize them again.
2067 VectorizedStores.insert(Operands.begin(), Operands.end());
2068 Changed |= Vectorized;
2075 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2078 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2079 StoreInst *SI = dyn_cast<StoreInst>(it);
2083 // Don't touch volatile stores.
2084 if (!SI->isSimple())
2087 // Check that the pointer points to scalars.
2088 Type *Ty = SI->getValueOperand()->getType();
2089 if (Ty->isAggregateType() || Ty->isVectorTy())
2092 // Find the base pointer.
2093 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2095 // Save the store locations.
2096 StoreRefs[Ptr].push_back(SI);
2102 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2105 Value *VL[] = { A, B };
2106 return tryToVectorizeList(VL, R);
2109 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
2113 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2115 // Check that all of the parts are scalar instructions of the same type.
2116 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2120 unsigned Opcode0 = I0->getOpcode();
2122 Type *Ty0 = I0->getType();
2123 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2124 unsigned VF = MinVecRegSize / Sz;
2126 for (int i = 0, e = VL.size(); i < e; ++i) {
2127 Type *Ty = VL[i]->getType();
2128 if (Ty->isAggregateType() || Ty->isVectorTy())
2130 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2131 if (!Inst || Inst->getOpcode() != Opcode0)
2135 bool Changed = false;
2137 // Keep track of values that were delete by vectorizing in the loop below.
2138 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2140 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2141 unsigned OpsWidth = 0;
2148 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2151 // Check that a previous iteration of this loop did not delete the Value.
2152 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2155 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2157 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2160 int Cost = R.getTreeCost();
2162 if (Cost < -SLPCostThreshold) {
2163 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n");
2166 // Move to the next bundle.
2175 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2179 // Try to vectorize V.
2180 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2183 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2184 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2186 if (B && B->hasOneUse()) {
2187 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2188 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2189 if (tryToVectorizePair(A, B0, R)) {
2193 if (tryToVectorizePair(A, B1, R)) {
2200 if (A && A->hasOneUse()) {
2201 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2202 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2203 if (tryToVectorizePair(A0, B, R)) {
2207 if (tryToVectorizePair(A1, B, R)) {
2215 /// \brief Generate a shuffle mask to be used in a reduction tree.
2217 /// \param VecLen The length of the vector to be reduced.
2218 /// \param NumEltsToRdx The number of elements that should be reduced in the
2220 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2221 /// reduction. A pairwise reduction will generate a mask of
2222 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2223 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2224 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2225 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2226 bool IsPairwise, bool IsLeft,
2227 IRBuilder<> &Builder) {
2228 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2230 SmallVector<Constant *, 32> ShuffleMask(
2231 VecLen, UndefValue::get(Builder.getInt32Ty()));
2234 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2235 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2236 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2238 // Move the upper half of the vector to the lower half.
2239 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2240 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2242 return ConstantVector::get(ShuffleMask);
2246 /// Model horizontal reductions.
2248 /// A horizontal reduction is a tree of reduction operations (currently add and
2249 /// fadd) that has operations that can be put into a vector as its leaf.
2250 /// For example, this tree:
2257 /// This tree has "mul" as its reduced values and "+" as its reduction
2258 /// operations. A reduction might be feeding into a store or a binary operation
2273 class HorizontalReduction {
2274 SmallPtrSet<Value *, 16> ReductionOps;
2275 SmallVector<Value *, 32> ReducedVals;
2277 BinaryOperator *ReductionRoot;
2278 PHINode *ReductionPHI;
2280 /// The opcode of the reduction.
2281 unsigned ReductionOpcode;
2282 /// The opcode of the values we perform a reduction on.
2283 unsigned ReducedValueOpcode;
2284 /// The width of one full horizontal reduction operation.
2285 unsigned ReduxWidth;
2286 /// Should we model this reduction as a pairwise reduction tree or a tree that
2287 /// splits the vector in halves and adds those halves.
2288 bool IsPairwiseReduction;
2291 HorizontalReduction()
2292 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0),
2293 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2295 /// \brief Try to find a reduction tree.
2296 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2299 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2300 "Thi phi needs to use the binary operator");
2302 // We could have a initial reductions that is not an add.
2303 // r *= v1 + v2 + v3 + v4
2304 // In such a case start looking for a tree rooted in the first '+'.
2306 if (B->getOperand(0) == Phi) {
2308 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2309 } else if (B->getOperand(1) == Phi) {
2311 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2318 Type *Ty = B->getType();
2319 if (Ty->isVectorTy())
2322 ReductionOpcode = B->getOpcode();
2323 ReducedValueOpcode = 0;
2324 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2331 // We currently only support adds.
2332 if (ReductionOpcode != Instruction::Add &&
2333 ReductionOpcode != Instruction::FAdd)
2336 // Post order traverse the reduction tree starting at B. We only handle true
2337 // trees containing only binary operators.
2338 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2339 Stack.push_back(std::make_pair(B, 0));
2340 while (!Stack.empty()) {
2341 BinaryOperator *TreeN = Stack.back().first;
2342 unsigned EdgeToVist = Stack.back().second++;
2343 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2345 // Only handle trees in the current basic block.
2346 if (TreeN->getParent() != B->getParent())
2349 // Each tree node needs to have one user except for the ultimate
2351 if (!TreeN->hasOneUse() && TreeN != B)
2355 if (EdgeToVist == 2 || IsReducedValue) {
2356 if (IsReducedValue) {
2357 // Make sure that the opcodes of the operations that we are going to
2359 if (!ReducedValueOpcode)
2360 ReducedValueOpcode = TreeN->getOpcode();
2361 else if (ReducedValueOpcode != TreeN->getOpcode())
2363 ReducedVals.push_back(TreeN);
2365 // We need to be able to reassociate the adds.
2366 if (!TreeN->isAssociative())
2368 ReductionOps.insert(TreeN);
2375 // Visit left or right.
2376 Value *NextV = TreeN->getOperand(EdgeToVist);
2377 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2379 Stack.push_back(std::make_pair(Next, 0));
2380 else if (NextV != Phi)
2386 /// \brief Attempt to vectorize the tree found by
2387 /// matchAssociativeReduction.
2388 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2389 if (ReducedVals.empty())
2392 unsigned NumReducedVals = ReducedVals.size();
2393 if (NumReducedVals < ReduxWidth)
2396 Value *VectorizedTree = 0;
2397 IRBuilder<> Builder(ReductionRoot);
2398 FastMathFlags Unsafe;
2399 Unsafe.setUnsafeAlgebra();
2400 Builder.SetFastMathFlags(Unsafe);
2403 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2404 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2405 V.buildTree(ValsToReduce, &ReductionOps);
2408 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2409 if (Cost >= -SLPCostThreshold)
2412 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2415 // Vectorize a tree.
2416 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2417 Value *VectorizedRoot = V.vectorizeTree();
2419 // Emit a reduction.
2420 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2421 if (VectorizedTree) {
2422 Builder.SetCurrentDebugLocation(Loc);
2423 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2424 ReducedSubTree, "bin.rdx");
2426 VectorizedTree = ReducedSubTree;
2429 if (VectorizedTree) {
2430 // Finish the reduction.
2431 for (; i < NumReducedVals; ++i) {
2432 Builder.SetCurrentDebugLocation(
2433 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2434 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2439 assert(ReductionRoot != NULL && "Need a reduction operation");
2440 ReductionRoot->setOperand(0, VectorizedTree);
2441 ReductionRoot->setOperand(1, ReductionPHI);
2443 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2445 return VectorizedTree != 0;
2450 /// \brief Calcuate the cost of a reduction.
2451 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2452 Type *ScalarTy = FirstReducedVal->getType();
2453 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2455 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2456 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2458 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2459 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2461 int ScalarReduxCost =
2462 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2464 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2465 << " for reduction that starts with " << *FirstReducedVal
2467 << (IsPairwiseReduction ? "pairwise" : "splitting")
2468 << " reduction)\n");
2470 return VecReduxCost - ScalarReduxCost;
2473 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2474 Value *R, const Twine &Name = "") {
2475 if (Opcode == Instruction::FAdd)
2476 return Builder.CreateFAdd(L, R, Name);
2477 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2480 /// \brief Emit a horizontal reduction of the vectorized value.
2481 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2482 assert(VectorizedValue && "Need to have a vectorized tree node");
2483 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2484 assert(isPowerOf2_32(ReduxWidth) &&
2485 "We only handle power-of-two reductions for now");
2487 Value *TmpVec = ValToReduce;
2488 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2489 if (IsPairwiseReduction) {
2491 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2493 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2495 Value *LeftShuf = Builder.CreateShuffleVector(
2496 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2497 Value *RightShuf = Builder.CreateShuffleVector(
2498 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2500 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2504 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2505 Value *Shuf = Builder.CreateShuffleVector(
2506 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2507 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2511 // The result is in the first element of the vector.
2512 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2516 /// \brief Recognize construction of vectors like
2517 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2518 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2519 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2520 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2522 /// Returns true if it matches
2524 static bool findBuildVector(InsertElementInst *IE,
2525 SmallVectorImpl<Value *> &Ops) {
2526 if (!isa<UndefValue>(IE->getOperand(0)))
2530 Ops.push_back(IE->getOperand(1));
2532 if (IE->use_empty())
2535 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->use_back());
2539 // If this isn't the final use, make sure the next insertelement is the only
2540 // use. It's OK if the final constructed vector is used multiple times
2541 if (!IE->hasOneUse())
2550 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2551 return V->getType() < V2->getType();
2554 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2555 bool Changed = false;
2556 SmallVector<Value *, 4> Incoming;
2557 SmallSet<Value *, 16> VisitedInstrs;
2559 bool HaveVectorizedPhiNodes = true;
2560 while (HaveVectorizedPhiNodes) {
2561 HaveVectorizedPhiNodes = false;
2563 // Collect the incoming values from the PHIs.
2565 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2567 PHINode *P = dyn_cast<PHINode>(instr);
2571 if (!VisitedInstrs.count(P))
2572 Incoming.push_back(P);
2576 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2578 // Try to vectorize elements base on their type.
2579 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2583 // Look for the next elements with the same type.
2584 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2585 while (SameTypeIt != E &&
2586 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2587 VisitedInstrs.insert(*SameTypeIt);
2591 // Try to vectorize them.
2592 unsigned NumElts = (SameTypeIt - IncIt);
2593 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2595 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2596 // Success start over because instructions might have been changed.
2597 HaveVectorizedPhiNodes = true;
2602 // Start over at the next instruction of a different type (or the end).
2607 VisitedInstrs.clear();
2609 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2610 // We may go through BB multiple times so skip the one we have checked.
2611 if (!VisitedInstrs.insert(it))
2614 if (isa<DbgInfoIntrinsic>(it))
2617 // Try to vectorize reductions that use PHINodes.
2618 if (PHINode *P = dyn_cast<PHINode>(it)) {
2619 // Check that the PHI is a reduction PHI.
2620 if (P->getNumIncomingValues() != 2)
2623 (P->getIncomingBlock(0) == BB
2624 ? (P->getIncomingValue(0))
2625 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
2626 // Check if this is a Binary Operator.
2627 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2631 // Try to match and vectorize a horizontal reduction.
2632 HorizontalReduction HorRdx;
2633 if (ShouldVectorizeHor &&
2634 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2635 HorRdx.tryToReduce(R, TTI)) {
2642 Value *Inst = BI->getOperand(0);
2644 Inst = BI->getOperand(1);
2646 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2647 // We would like to start over since some instructions are deleted
2648 // and the iterator may become invalid value.
2658 // Try to vectorize horizontal reductions feeding into a store.
2659 if (ShouldStartVectorizeHorAtStore)
2660 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2661 if (BinaryOperator *BinOp =
2662 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2663 HorizontalReduction HorRdx;
2664 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) &&
2665 HorRdx.tryToReduce(R, TTI)) ||
2666 tryToVectorize(BinOp, R))) {
2674 // Try to vectorize trees that start at compare instructions.
2675 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2676 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2678 // We would like to start over since some instructions are deleted
2679 // and the iterator may become invalid value.
2685 for (int i = 0; i < 2; ++i) {
2686 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2687 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2689 // We would like to start over since some instructions are deleted
2690 // and the iterator may become invalid value.
2699 // Try to vectorize trees that start at insertelement instructions.
2700 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2701 SmallVector<Value *, 8> Ops;
2702 if (!findBuildVector(IE, Ops))
2705 if (tryToVectorizeList(Ops, R)) {
2718 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2719 bool Changed = false;
2720 // Attempt to sort and vectorize each of the store-groups.
2721 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2723 if (it->second.size() < 2)
2726 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2727 << it->second.size() << ".\n");
2729 // Process the stores in chunks of 16.
2730 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2731 unsigned Len = std::min<unsigned>(CE - CI, 16);
2732 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2733 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2739 } // end anonymous namespace
2741 char SLPVectorizer::ID = 0;
2742 static const char lv_name[] = "SLP Vectorizer";
2743 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2744 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2745 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2746 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2747 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2748 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2751 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }