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/ScalarEvolution.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/Analysis/AliasAnalysis.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/Analysis/Verifier.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
49 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
50 cl::desc("Only vectorize if you gain more than this "
54 static const unsigned MinVecRegSize = 128;
56 static const unsigned RecursionMaxDepth = 12;
58 /// RAII pattern to save the insertion point of the IR builder.
59 class BuilderLocGuard {
61 BuilderLocGuard(IRBuilder<> &B) : Builder(B), Loc(B.GetInsertPoint()),
62 DbgLoc(B.getCurrentDebugLocation()) {}
64 Builder.SetCurrentDebugLocation(DbgLoc);
66 Builder.SetInsertPoint(Loc);
71 BuilderLocGuard(const BuilderLocGuard &);
72 BuilderLocGuard &operator=(const BuilderLocGuard &);
74 AssertingVH<Instruction> Loc;
78 /// A helper class for numbering instructions in multiple blocks.
79 /// Numbers start at zero for each basic block.
80 struct BlockNumbering {
82 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
84 BlockNumbering() : BB(0), Valid(false) {}
86 void numberInstructions() {
90 // Number the instructions in the block.
91 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
93 InstrVec.push_back(it);
94 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
99 int getIndex(Instruction *I) {
100 assert(I->getParent() == BB && "Invalid instruction");
102 numberInstructions();
103 assert(InstrIdx.count(I) && "Unknown instruction");
107 Instruction *getInstruction(unsigned loc) {
109 numberInstructions();
110 assert(InstrVec.size() > loc && "Invalid Index");
111 return InstrVec[loc];
114 void forget() { Valid = false; }
117 /// The block we are numbering.
119 /// Is the block numbered.
121 /// Maps instructions to numbers and back.
122 SmallDenseMap<Instruction *, int> InstrIdx;
123 /// Maps integers to Instructions.
124 SmallVector<Instruction *, 32> InstrVec;
127 /// \returns the parent basic block if all of the instructions in \p VL
128 /// are in the same block or null otherwise.
129 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
130 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
133 BasicBlock *BB = I0->getParent();
134 for (int i = 1, e = VL.size(); i < e; i++) {
135 Instruction *I = dyn_cast<Instruction>(VL[i]);
139 if (BB != I->getParent())
145 /// \returns True if all of the values in \p VL are constants.
146 static bool allConstant(ArrayRef<Value *> VL) {
147 for (unsigned i = 0, e = VL.size(); i < e; ++i)
148 if (!isa<Constant>(VL[i]))
153 /// \returns True if all of the values in \p VL are identical.
154 static bool isSplat(ArrayRef<Value *> VL) {
155 for (unsigned i = 1, e = VL.size(); i < e; ++i)
161 /// \returns The opcode if all of the Instructions in \p VL have the same
163 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
164 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
167 unsigned Opcode = I0->getOpcode();
168 for (int i = 1, e = VL.size(); i < e; i++) {
169 Instruction *I = dyn_cast<Instruction>(VL[i]);
170 if (!I || Opcode != I->getOpcode())
176 /// \returns The type that all of the values in \p VL have or null if there
177 /// are different types.
178 static Type* getSameType(ArrayRef<Value *> VL) {
179 Type *Ty = VL[0]->getType();
180 for (int i = 1, e = VL.size(); i < e; i++)
181 if (VL[i]->getType() != Ty)
187 /// \returns True if the ExtractElement instructions in VL can be vectorized
188 /// to use the original vector.
189 static bool CanReuseExtract(ArrayRef<Value *> VL) {
190 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
191 // Check if all of the extracts come from the same vector and from the
194 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
195 Value *Vec = E0->getOperand(0);
197 // We have to extract from the same vector type.
198 unsigned NElts = Vec->getType()->getVectorNumElements();
200 if (NElts != VL.size())
203 // Check that all of the indices extract from the correct offset.
204 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
205 if (!CI || CI->getZExtValue())
208 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
209 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
210 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
212 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
219 /// Bottom Up SLP Vectorizer.
222 typedef SmallVector<Value *, 8> ValueList;
223 typedef SmallVector<Instruction *, 16> InstrList;
224 typedef SmallPtrSet<Value *, 16> ValueSet;
225 typedef SmallVector<StoreInst *, 8> StoreList;
227 BoUpSLP(Function *Func, ScalarEvolution *Se, DataLayout *Dl,
228 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li,
230 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt),
231 Builder(Se->getContext()) {
232 // Setup the block numbering utility for all of the blocks in the
234 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
236 BlocksNumbers[BB] = BlockNumbering(BB);
240 /// \brief Vectorize the tree that starts with the elements in \p VL.
241 void vectorizeTree();
243 /// \returns the vectorization cost of the subtree that starts at \p VL.
244 /// A negative number means that this is profitable.
247 /// Construct a vectorizable tree that starts at \p Roots.
248 void buildTree(ArrayRef<Value *> Roots);
250 /// Clear the internal data structures that are created by 'buildTree'.
252 VectorizableTree.clear();
253 ScalarToTreeEntry.clear();
255 ExternalUses.clear();
256 MemBarrierIgnoreList.clear();
259 /// \returns true if the memory operations A and B are consecutive.
260 bool isConsecutiveAccess(Value *A, Value *B);
262 /// \brief Perform LICM and CSE on the newly generated gather sequences.
263 void optimizeGatherSequence();
267 /// \returns the cost of the vectorizable entry.
268 int getEntryCost(TreeEntry *E);
270 /// This is the recursive part of buildTree.
271 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
273 /// Vectorize a single entry in the tree.
274 Value *vectorizeTree(TreeEntry *E);
276 /// Vectorize a single entry in the tree, starting in \p VL.
277 Value *vectorizeTree(ArrayRef<Value *> VL);
279 /// \returns the pointer to the vectorized value if \p VL is already
280 /// vectorized, or NULL. They may happen in cycles.
281 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
283 /// \brief Take the pointer operand from the Load/Store instruction.
284 /// \returns NULL if this is not a valid Load/Store instruction.
285 static Value *getPointerOperand(Value *I);
287 /// \brief Take the address space operand from the Load/Store instruction.
288 /// \returns -1 if this is not a valid Load/Store instruction.
289 static unsigned getAddressSpaceOperand(Value *I);
291 /// \returns the scalarization cost for this type. Scalarization in this
292 /// context means the creation of vectors from a group of scalars.
293 int getGatherCost(Type *Ty);
295 /// \returns the scalarization cost for this list of values. Assuming that
296 /// this subtree gets vectorized, we may need to extract the values from the
297 /// roots. This method calculates the cost of extracting the values.
298 int getGatherCost(ArrayRef<Value *> VL);
300 /// \returns the AA location that is being access by the instruction.
301 AliasAnalysis::Location getLocation(Instruction *I);
303 /// \brief Checks if it is possible to sink an instruction from
304 /// \p Src to \p Dst.
305 /// \returns the pointer to the barrier instruction if we can't sink.
306 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
308 /// \returns the index of the last instrucion in the BB from \p VL.
309 int getLastIndex(ArrayRef<Value *> VL);
311 /// \returns the Instruction in the bundle \p VL.
312 Instruction *getLastInstruction(ArrayRef<Value *> VL);
314 /// \brief Set the Builder insert point to one after the last instruction in
316 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
318 /// \returns a vector from a collection of scalars in \p VL.
319 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
322 TreeEntry() : Scalars(), VectorizedValue(0), LastScalarIndex(0),
325 /// \returns true if the scalars in VL are equal to this entry.
326 bool isSame(ArrayRef<Value *> VL) const {
327 assert(VL.size() == Scalars.size() && "Invalid size");
328 for (int i = 0, e = VL.size(); i != e; ++i)
329 if (VL[i] != Scalars[i])
334 /// A vector of scalars.
337 /// The Scalars are vectorized into this value. It is initialized to Null.
338 Value *VectorizedValue;
340 /// The index in the basic block of the last scalar.
343 /// Do we need to gather this sequence ?
347 /// Create a new VectorizableTree entry.
348 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
349 VectorizableTree.push_back(TreeEntry());
350 int idx = VectorizableTree.size() - 1;
351 TreeEntry *Last = &VectorizableTree[idx];
352 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
353 Last->NeedToGather = !Vectorized;
355 Last->LastScalarIndex = getLastIndex(VL);
356 for (int i = 0, e = VL.size(); i != e; ++i) {
357 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
358 ScalarToTreeEntry[VL[i]] = idx;
361 Last->LastScalarIndex = 0;
362 MustGather.insert(VL.begin(), VL.end());
367 /// -- Vectorization State --
368 /// Holds all of the tree entries.
369 std::vector<TreeEntry> VectorizableTree;
371 /// Maps a specific scalar to its tree entry.
372 SmallDenseMap<Value*, int> ScalarToTreeEntry;
374 /// A list of scalars that we found that we need to keep as scalars.
377 /// This POD struct describes one external user in the vectorized tree.
378 struct ExternalUser {
379 ExternalUser (Value *S, llvm::User *U, int L) :
380 Scalar(S), User(U), Lane(L){};
381 // Which scalar in our function.
383 // Which user that uses the scalar.
385 // Which lane does the scalar belong to.
388 typedef SmallVector<ExternalUser, 16> UserList;
390 /// A list of values that need to extracted out of the tree.
391 /// This list holds pairs of (Internal Scalar : External User).
392 UserList ExternalUses;
394 /// A list of instructions to ignore while sinking
395 /// memory instructions. This map must be reset between runs of getCost.
396 ValueSet MemBarrierIgnoreList;
398 /// Holds all of the instructions that we gathered.
399 SetVector<Instruction *> GatherSeq;
401 /// Numbers instructions in different blocks.
402 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
404 // Analysis and block reference.
408 TargetTransformInfo *TTI;
412 /// Instruction builder to construct the vectorized tree.
416 void BoUpSLP::buildTree(ArrayRef<Value *> Roots) {
418 if (!getSameType(Roots))
420 buildTree_rec(Roots, 0);
422 // Collect the values that we need to extract from the tree.
423 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
424 TreeEntry *Entry = &VectorizableTree[EIdx];
427 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
428 Value *Scalar = Entry->Scalars[Lane];
430 // No need to handle users of gathered values.
431 if (Entry->NeedToGather)
434 for (Value::use_iterator User = Scalar->use_begin(),
435 UE = Scalar->use_end(); User != UE; ++User) {
436 DEBUG(dbgs() << "SLP: Checking user:" << **User << ".\n");
438 bool Gathered = MustGather.count(*User);
440 // Skip in-tree scalars that become vectors.
441 if (ScalarToTreeEntry.count(*User) && !Gathered) {
442 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
444 int Idx = ScalarToTreeEntry[*User]; (void) Idx;
445 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
449 if (!isa<Instruction>(*User))
452 DEBUG(dbgs() << "SLP: Need to extract:" << **User << " from lane " <<
453 Lane << " from " << *Scalar << ".\n");
454 ExternalUses.push_back(ExternalUser(Scalar, *User, Lane));
461 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
462 bool SameTy = getSameType(VL); (void)SameTy;
463 assert(SameTy && "Invalid types!");
465 if (Depth == RecursionMaxDepth) {
466 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
467 newTreeEntry(VL, false);
471 // Don't handle vectors.
472 if (VL[0]->getType()->isVectorTy()) {
473 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
474 newTreeEntry(VL, false);
478 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
479 if (SI->getValueOperand()->getType()->isVectorTy()) {
480 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
481 newTreeEntry(VL, false);
485 // If all of the operands are identical or constant we have a simple solution.
486 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
487 !getSameOpcode(VL)) {
488 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
489 newTreeEntry(VL, false);
493 // We now know that this is a vector of instructions of the same type from
496 // Check if this is a duplicate of another entry.
497 if (ScalarToTreeEntry.count(VL[0])) {
498 int Idx = ScalarToTreeEntry[VL[0]];
499 TreeEntry *E = &VectorizableTree[Idx];
500 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
501 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
502 if (E->Scalars[i] != VL[i]) {
503 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
504 newTreeEntry(VL, false);
508 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
512 // Check that none of the instructions in the bundle are already in the tree.
513 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
514 if (ScalarToTreeEntry.count(VL[i])) {
515 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
516 ") is already in tree.\n");
517 newTreeEntry(VL, false);
522 // If any of the scalars appears in the table OR it is marked as a value that
523 // needs to stat scalar then we need to gather the scalars.
524 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
525 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
526 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
527 newTreeEntry(VL, false);
532 // Check that all of the users of the scalars that we want to vectorize are
534 Instruction *VL0 = cast<Instruction>(VL[0]);
535 int MyLastIndex = getLastIndex(VL);
536 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
538 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
539 Instruction *Scalar = cast<Instruction>(VL[i]);
540 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
541 for (Value::use_iterator U = Scalar->use_begin(), UE = Scalar->use_end();
543 DEBUG(dbgs() << "SLP: \tUser " << **U << ". \n");
544 Instruction *User = dyn_cast<Instruction>(*U);
546 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
547 newTreeEntry(VL, false);
551 // We don't care if the user is in a different basic block.
552 BasicBlock *UserBlock = User->getParent();
553 if (UserBlock != BB) {
554 DEBUG(dbgs() << "SLP: User from a different basic block "
559 // If this is a PHINode within this basic block then we can place the
560 // extract wherever we want.
561 if (isa<PHINode>(*User)) {
562 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *User << ". \n");
566 // Check if this is a safe in-tree user.
567 if (ScalarToTreeEntry.count(User)) {
568 int Idx = ScalarToTreeEntry[User];
569 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
570 if (VecLocation <= MyLastIndex) {
571 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
572 newTreeEntry(VL, false);
575 DEBUG(dbgs() << "SLP: In-tree user (" << *User << ") at #" <<
576 VecLocation << " vector value (" << *Scalar << ") at #"
577 << MyLastIndex << ".\n");
581 // Make sure that we can schedule this unknown user.
582 BlockNumbering &BN = BlocksNumbers[BB];
583 int UserIndex = BN.getIndex(User);
584 if (UserIndex < MyLastIndex) {
586 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
588 newTreeEntry(VL, false);
594 // Check that every instructions appears once in this bundle.
595 for (unsigned i = 0, e = VL.size(); i < e; ++i)
596 for (unsigned j = i+1; j < e; ++j)
597 if (VL[i] == VL[j]) {
598 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
599 newTreeEntry(VL, false);
603 // Check that instructions in this bundle don't reference other instructions.
604 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
605 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
606 for (Value::use_iterator U = VL[i]->use_begin(), UE = VL[i]->use_end();
608 for (unsigned j = 0; j < e; ++j) {
609 if (i != j && *U == VL[j]) {
610 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << **U << ". \n");
611 newTreeEntry(VL, false);
618 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
620 unsigned Opcode = getSameOpcode(VL);
622 // Check if it is safe to sink the loads or the stores.
623 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
624 Instruction *Last = getLastInstruction(VL);
626 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
629 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
631 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
632 << "\n because of " << *Barrier << ". Gathering.\n");
633 newTreeEntry(VL, false);
640 case Instruction::PHI: {
641 PHINode *PH = dyn_cast<PHINode>(VL0);
643 // Check for terminator values (e.g. invoke).
644 for (unsigned j = 0; j < VL.size(); ++j)
645 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
646 TerminatorInst *Term = dyn_cast<TerminatorInst>(cast<PHINode>(VL[j])->getIncomingValue(i));
648 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
649 newTreeEntry(VL, false);
654 newTreeEntry(VL, true);
655 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
657 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
659 // Prepare the operand vector.
660 for (unsigned j = 0; j < VL.size(); ++j)
661 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValue(i));
663 buildTree_rec(Operands, Depth + 1);
667 case Instruction::ExtractElement: {
668 bool Reuse = CanReuseExtract(VL);
670 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
672 newTreeEntry(VL, Reuse);
675 case Instruction::Load: {
676 // Check if the loads are consecutive or of we need to swizzle them.
677 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
678 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
679 newTreeEntry(VL, false);
680 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
684 newTreeEntry(VL, true);
685 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
688 case Instruction::ZExt:
689 case Instruction::SExt:
690 case Instruction::FPToUI:
691 case Instruction::FPToSI:
692 case Instruction::FPExt:
693 case Instruction::PtrToInt:
694 case Instruction::IntToPtr:
695 case Instruction::SIToFP:
696 case Instruction::UIToFP:
697 case Instruction::Trunc:
698 case Instruction::FPTrunc:
699 case Instruction::BitCast: {
700 Type *SrcTy = VL0->getOperand(0)->getType();
701 for (unsigned i = 0; i < VL.size(); ++i) {
702 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
703 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
704 newTreeEntry(VL, false);
705 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
709 newTreeEntry(VL, true);
710 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
712 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
714 // Prepare the operand vector.
715 for (unsigned j = 0; j < VL.size(); ++j)
716 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
718 buildTree_rec(Operands, Depth+1);
722 case Instruction::ICmp:
723 case Instruction::FCmp: {
724 // Check that all of the compares have the same predicate.
725 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
726 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
727 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
728 CmpInst *Cmp = cast<CmpInst>(VL[i]);
729 if (Cmp->getPredicate() != P0 ||
730 Cmp->getOperand(0)->getType() != ComparedTy) {
731 newTreeEntry(VL, false);
732 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
737 newTreeEntry(VL, true);
738 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
740 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
742 // Prepare the operand vector.
743 for (unsigned j = 0; j < VL.size(); ++j)
744 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
746 buildTree_rec(Operands, Depth+1);
750 case Instruction::Select:
751 case Instruction::Add:
752 case Instruction::FAdd:
753 case Instruction::Sub:
754 case Instruction::FSub:
755 case Instruction::Mul:
756 case Instruction::FMul:
757 case Instruction::UDiv:
758 case Instruction::SDiv:
759 case Instruction::FDiv:
760 case Instruction::URem:
761 case Instruction::SRem:
762 case Instruction::FRem:
763 case Instruction::Shl:
764 case Instruction::LShr:
765 case Instruction::AShr:
766 case Instruction::And:
767 case Instruction::Or:
768 case Instruction::Xor: {
769 newTreeEntry(VL, true);
770 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
772 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
774 // Prepare the operand vector.
775 for (unsigned j = 0; j < VL.size(); ++j)
776 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
778 buildTree_rec(Operands, Depth+1);
782 case Instruction::Store: {
783 // Check if the stores are consecutive or of we need to swizzle them.
784 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
785 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
786 newTreeEntry(VL, false);
787 DEBUG(dbgs() << "SLP: Non consecutive store.\n");
791 newTreeEntry(VL, true);
792 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
795 for (unsigned j = 0; j < VL.size(); ++j)
796 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
798 // We can ignore these values because we are sinking them down.
799 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
800 buildTree_rec(Operands, Depth + 1);
804 newTreeEntry(VL, false);
805 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
810 int BoUpSLP::getEntryCost(TreeEntry *E) {
811 ArrayRef<Value*> VL = E->Scalars;
813 Type *ScalarTy = VL[0]->getType();
814 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
815 ScalarTy = SI->getValueOperand()->getType();
816 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
818 if (E->NeedToGather) {
822 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
824 return getGatherCost(E->Scalars);
827 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
829 Instruction *VL0 = cast<Instruction>(VL[0]);
830 unsigned Opcode = VL0->getOpcode();
832 case Instruction::PHI: {
835 case Instruction::ExtractElement: {
836 if (CanReuseExtract(VL))
838 return getGatherCost(VecTy);
840 case Instruction::ZExt:
841 case Instruction::SExt:
842 case Instruction::FPToUI:
843 case Instruction::FPToSI:
844 case Instruction::FPExt:
845 case Instruction::PtrToInt:
846 case Instruction::IntToPtr:
847 case Instruction::SIToFP:
848 case Instruction::UIToFP:
849 case Instruction::Trunc:
850 case Instruction::FPTrunc:
851 case Instruction::BitCast: {
852 Type *SrcTy = VL0->getOperand(0)->getType();
854 // Calculate the cost of this instruction.
855 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
856 VL0->getType(), SrcTy);
858 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
859 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
860 return VecCost - ScalarCost;
862 case Instruction::FCmp:
863 case Instruction::ICmp:
864 case Instruction::Select:
865 case Instruction::Add:
866 case Instruction::FAdd:
867 case Instruction::Sub:
868 case Instruction::FSub:
869 case Instruction::Mul:
870 case Instruction::FMul:
871 case Instruction::UDiv:
872 case Instruction::SDiv:
873 case Instruction::FDiv:
874 case Instruction::URem:
875 case Instruction::SRem:
876 case Instruction::FRem:
877 case Instruction::Shl:
878 case Instruction::LShr:
879 case Instruction::AShr:
880 case Instruction::And:
881 case Instruction::Or:
882 case Instruction::Xor: {
883 // Calculate the cost of this instruction.
886 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
887 Opcode == Instruction::Select) {
888 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
889 ScalarCost = VecTy->getNumElements() *
890 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
891 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
893 ScalarCost = VecTy->getNumElements() *
894 TTI->getArithmeticInstrCost(Opcode, ScalarTy);
895 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
897 return VecCost - ScalarCost;
899 case Instruction::Load: {
900 // Cost of wide load - cost of scalar loads.
901 int ScalarLdCost = VecTy->getNumElements() *
902 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
903 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
904 return VecLdCost - ScalarLdCost;
906 case Instruction::Store: {
907 // We know that we can merge the stores. Calculate the cost.
908 int ScalarStCost = VecTy->getNumElements() *
909 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
910 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
911 return VecStCost - ScalarStCost;
914 llvm_unreachable("Unknown instruction");
918 int BoUpSLP::getTreeCost() {
920 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
921 VectorizableTree.size() << ".\n");
923 // Don't vectorize tiny trees. Small load/store chains or consecutive stores
924 // of constants will be vectoried in SelectionDAG in MergeConsecutiveStores.
925 // The SelectionDAG vectorizer can only handle pairs (trees of height = 2).
926 if (VectorizableTree.size() < 3) {
927 if (!VectorizableTree.size()) {
928 assert(!ExternalUses.size() && "We should not have any external users");
933 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
935 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
936 int C = getEntryCost(&VectorizableTree[i]);
937 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
938 << *VectorizableTree[i].Scalars[0] << " .\n");
943 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
946 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
947 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
952 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
953 return Cost + ExtractCost;
956 int BoUpSLP::getGatherCost(Type *Ty) {
958 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
959 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
963 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
964 // Find the type of the operands in VL.
965 Type *ScalarTy = VL[0]->getType();
966 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
967 ScalarTy = SI->getValueOperand()->getType();
968 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
969 // Find the cost of inserting/extracting values from the vector.
970 return getGatherCost(VecTy);
973 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
974 if (StoreInst *SI = dyn_cast<StoreInst>(I))
975 return AA->getLocation(SI);
976 if (LoadInst *LI = dyn_cast<LoadInst>(I))
977 return AA->getLocation(LI);
978 return AliasAnalysis::Location();
981 Value *BoUpSLP::getPointerOperand(Value *I) {
982 if (LoadInst *LI = dyn_cast<LoadInst>(I))
983 return LI->getPointerOperand();
984 if (StoreInst *SI = dyn_cast<StoreInst>(I))
985 return SI->getPointerOperand();
989 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
990 if (LoadInst *L = dyn_cast<LoadInst>(I))
991 return L->getPointerAddressSpace();
992 if (StoreInst *S = dyn_cast<StoreInst>(I))
993 return S->getPointerAddressSpace();
997 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
998 Value *PtrA = getPointerOperand(A);
999 Value *PtrB = getPointerOperand(B);
1000 unsigned ASA = getAddressSpaceOperand(A);
1001 unsigned ASB = getAddressSpaceOperand(B);
1003 // Check that the address spaces match and that the pointers are valid.
1004 if (!PtrA || !PtrB || (ASA != ASB))
1007 // Make sure that A and B are different pointers of the same type.
1008 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1011 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1012 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1013 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1015 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1016 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1017 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1019 APInt OffsetDelta = OffsetB - OffsetA;
1021 // Check if they are based on the same pointer. That makes the offsets
1024 return OffsetDelta == Size;
1026 // Compute the necessary base pointer delta to have the necessary final delta
1027 // equal to the size.
1028 APInt BaseDelta = Size - OffsetDelta;
1030 // Otherwise compute the distance with SCEV between the base pointers.
1031 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1032 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1033 const SCEV *C = SE->getConstant(BaseDelta);
1034 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1035 return X == PtrSCEVB;
1038 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1039 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1040 BasicBlock::iterator I = Src, E = Dst;
1041 /// Scan all of the instruction from SRC to DST and check if
1042 /// the source may alias.
1043 for (++I; I != E; ++I) {
1044 // Ignore store instructions that are marked as 'ignore'.
1045 if (MemBarrierIgnoreList.count(I))
1047 if (Src->mayWriteToMemory()) /* Write */ {
1048 if (!I->mayReadOrWriteMemory())
1051 if (!I->mayWriteToMemory())
1054 AliasAnalysis::Location A = getLocation(&*I);
1055 AliasAnalysis::Location B = getLocation(Src);
1057 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1063 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1064 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1065 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1066 BlockNumbering &BN = BlocksNumbers[BB];
1068 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1069 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1070 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1074 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1075 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1076 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1077 BlockNumbering &BN = BlocksNumbers[BB];
1079 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1080 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1081 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1082 Instruction *I = BN.getInstruction(MaxIdx);
1083 assert(I && "bad location");
1087 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1088 Instruction *VL0 = cast<Instruction>(VL[0]);
1089 Instruction *LastInst = getLastInstruction(VL);
1090 BasicBlock::iterator NextInst = LastInst;
1092 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1093 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1096 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1097 Value *Vec = UndefValue::get(Ty);
1098 // Generate the 'InsertElement' instruction.
1099 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1100 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1101 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1102 GatherSeq.insert(Insrt);
1104 // Add to our 'need-to-extract' list.
1105 if (ScalarToTreeEntry.count(VL[i])) {
1106 int Idx = ScalarToTreeEntry[VL[i]];
1107 TreeEntry *E = &VectorizableTree[Idx];
1108 // Find which lane we need to extract.
1110 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1111 // Is this the lane of the scalar that we are looking for ?
1112 if (E->Scalars[Lane] == VL[i]) {
1117 assert(FoundLane >= 0 && "Could not find the correct lane");
1118 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1126 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1127 SmallDenseMap<Value*, int>::const_iterator Entry
1128 = ScalarToTreeEntry.find(VL[0]);
1129 if (Entry != ScalarToTreeEntry.end()) {
1130 int Idx = Entry->second;
1131 const TreeEntry *En = &VectorizableTree[Idx];
1132 if (En->isSame(VL) && En->VectorizedValue)
1133 return En->VectorizedValue;
1138 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1139 if (ScalarToTreeEntry.count(VL[0])) {
1140 int Idx = ScalarToTreeEntry[VL[0]];
1141 TreeEntry *E = &VectorizableTree[Idx];
1143 return vectorizeTree(E);
1146 Type *ScalarTy = VL[0]->getType();
1147 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1148 ScalarTy = SI->getValueOperand()->getType();
1149 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1151 return Gather(VL, VecTy);
1154 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1155 BuilderLocGuard Guard(Builder);
1157 if (E->VectorizedValue) {
1158 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1159 return E->VectorizedValue;
1162 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1163 Type *ScalarTy = VL0->getType();
1164 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1165 ScalarTy = SI->getValueOperand()->getType();
1166 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1168 if (E->NeedToGather) {
1169 setInsertPointAfterBundle(E->Scalars);
1170 return Gather(E->Scalars, VecTy);
1173 unsigned Opcode = VL0->getOpcode();
1174 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1177 case Instruction::PHI: {
1178 PHINode *PH = dyn_cast<PHINode>(VL0);
1179 Builder.SetInsertPoint(PH->getParent()->getFirstInsertionPt());
1180 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1181 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1182 E->VectorizedValue = NewPhi;
1184 // PHINodes may have multiple entries from the same block. We want to
1185 // visit every block once.
1186 SmallSet<BasicBlock*, 4> VisitedBBs;
1188 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1190 BasicBlock *IBB = PH->getIncomingBlock(i);
1192 if (!VisitedBBs.insert(IBB)) {
1193 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1197 // Prepare the operand vector.
1198 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1199 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1200 getIncomingValueForBlock(IBB));
1202 Builder.SetInsertPoint(IBB->getTerminator());
1203 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1204 Value *Vec = vectorizeTree(Operands);
1205 NewPhi->addIncoming(Vec, IBB);
1208 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1209 "Invalid number of incoming values");
1213 case Instruction::ExtractElement: {
1214 if (CanReuseExtract(E->Scalars)) {
1215 Value *V = VL0->getOperand(0);
1216 E->VectorizedValue = V;
1219 return Gather(E->Scalars, VecTy);
1221 case Instruction::ZExt:
1222 case Instruction::SExt:
1223 case Instruction::FPToUI:
1224 case Instruction::FPToSI:
1225 case Instruction::FPExt:
1226 case Instruction::PtrToInt:
1227 case Instruction::IntToPtr:
1228 case Instruction::SIToFP:
1229 case Instruction::UIToFP:
1230 case Instruction::Trunc:
1231 case Instruction::FPTrunc:
1232 case Instruction::BitCast: {
1234 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1235 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1237 setInsertPointAfterBundle(E->Scalars);
1239 Value *InVec = vectorizeTree(INVL);
1241 if (Value *V = alreadyVectorized(E->Scalars))
1244 CastInst *CI = dyn_cast<CastInst>(VL0);
1245 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1246 E->VectorizedValue = V;
1249 case Instruction::FCmp:
1250 case Instruction::ICmp: {
1251 ValueList LHSV, RHSV;
1252 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1253 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1254 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1257 setInsertPointAfterBundle(E->Scalars);
1259 Value *L = vectorizeTree(LHSV);
1260 Value *R = vectorizeTree(RHSV);
1262 if (Value *V = alreadyVectorized(E->Scalars))
1265 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1267 if (Opcode == Instruction::FCmp)
1268 V = Builder.CreateFCmp(P0, L, R);
1270 V = Builder.CreateICmp(P0, L, R);
1272 E->VectorizedValue = V;
1275 case Instruction::Select: {
1276 ValueList TrueVec, FalseVec, CondVec;
1277 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1278 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1279 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1280 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1283 setInsertPointAfterBundle(E->Scalars);
1285 Value *Cond = vectorizeTree(CondVec);
1286 Value *True = vectorizeTree(TrueVec);
1287 Value *False = vectorizeTree(FalseVec);
1289 if (Value *V = alreadyVectorized(E->Scalars))
1292 Value *V = Builder.CreateSelect(Cond, True, False);
1293 E->VectorizedValue = V;
1296 case Instruction::Add:
1297 case Instruction::FAdd:
1298 case Instruction::Sub:
1299 case Instruction::FSub:
1300 case Instruction::Mul:
1301 case Instruction::FMul:
1302 case Instruction::UDiv:
1303 case Instruction::SDiv:
1304 case Instruction::FDiv:
1305 case Instruction::URem:
1306 case Instruction::SRem:
1307 case Instruction::FRem:
1308 case Instruction::Shl:
1309 case Instruction::LShr:
1310 case Instruction::AShr:
1311 case Instruction::And:
1312 case Instruction::Or:
1313 case Instruction::Xor: {
1314 ValueList LHSVL, RHSVL;
1315 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1316 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1317 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1320 setInsertPointAfterBundle(E->Scalars);
1322 Value *LHS = vectorizeTree(LHSVL);
1323 Value *RHS = vectorizeTree(RHSVL);
1325 if (LHS == RHS && isa<Instruction>(LHS)) {
1326 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1329 if (Value *V = alreadyVectorized(E->Scalars))
1332 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1333 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1334 E->VectorizedValue = V;
1337 case Instruction::Load: {
1338 // Loads are inserted at the head of the tree because we don't want to
1339 // sink them all the way down past store instructions.
1340 setInsertPointAfterBundle(E->Scalars);
1342 LoadInst *LI = cast<LoadInst>(VL0);
1344 Builder.CreateBitCast(LI->getPointerOperand(), VecTy->getPointerTo());
1345 unsigned Alignment = LI->getAlignment();
1346 LI = Builder.CreateLoad(VecPtr);
1347 LI->setAlignment(Alignment);
1348 E->VectorizedValue = LI;
1351 case Instruction::Store: {
1352 StoreInst *SI = cast<StoreInst>(VL0);
1353 unsigned Alignment = SI->getAlignment();
1356 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1357 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1359 setInsertPointAfterBundle(E->Scalars);
1361 Value *VecValue = vectorizeTree(ValueOp);
1363 Builder.CreateBitCast(SI->getPointerOperand(), VecTy->getPointerTo());
1364 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1365 S->setAlignment(Alignment);
1366 E->VectorizedValue = S;
1370 llvm_unreachable("unknown inst");
1375 void BoUpSLP::vectorizeTree() {
1376 Builder.SetInsertPoint(F->getEntryBlock().begin());
1377 vectorizeTree(&VectorizableTree[0]);
1379 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1381 // Extract all of the elements with the external uses.
1382 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1384 Value *Scalar = it->Scalar;
1385 llvm::User *User = it->User;
1387 // Skip users that we already RAUW. This happens when one instruction
1388 // has multiple uses of the same value.
1389 if (std::find(Scalar->use_begin(), Scalar->use_end(), User) ==
1392 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1394 int Idx = ScalarToTreeEntry[Scalar];
1395 TreeEntry *E = &VectorizableTree[Idx];
1396 assert(!E->NeedToGather && "Extracting from a gather list");
1398 Value *Vec = E->VectorizedValue;
1399 assert(Vec && "Can't find vectorizable value");
1401 Value *Lane = Builder.getInt32(it->Lane);
1402 // Generate extracts for out-of-tree users.
1403 // Find the insertion point for the extractelement lane.
1404 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1405 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt());
1406 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1407 User->replaceUsesOfWith(Scalar, Ex);
1408 } else if (isa<Instruction>(Vec)){
1409 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1410 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1411 if (PH->getIncomingValue(i) == Scalar) {
1412 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1413 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1414 PH->setOperand(i, Ex);
1418 Builder.SetInsertPoint(cast<Instruction>(User));
1419 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1420 User->replaceUsesOfWith(Scalar, Ex);
1423 Builder.SetInsertPoint(F->getEntryBlock().begin());
1424 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1425 User->replaceUsesOfWith(Scalar, Ex);
1428 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1431 // For each vectorized value:
1432 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1433 TreeEntry *Entry = &VectorizableTree[EIdx];
1436 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1437 Value *Scalar = Entry->Scalars[Lane];
1439 // No need to handle users of gathered values.
1440 if (Entry->NeedToGather)
1443 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1445 Type *Ty = Scalar->getType();
1446 if (!Ty->isVoidTy()) {
1447 for (Value::use_iterator User = Scalar->use_begin(),
1448 UE = Scalar->use_end(); User != UE; ++User) {
1449 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n");
1450 assert(!MustGather.count(*User) &&
1451 "Replacing gathered value with undef");
1452 assert(ScalarToTreeEntry.count(*User) &&
1453 "Replacing out-of-tree value with undef");
1455 Value *Undef = UndefValue::get(Ty);
1456 Scalar->replaceAllUsesWith(Undef);
1458 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1459 cast<Instruction>(Scalar)->eraseFromParent();
1463 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1464 BlocksNumbers[it].forget();
1466 Builder.ClearInsertionPoint();
1469 void BoUpSLP::optimizeGatherSequence() {
1470 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1471 << " gather sequences instructions.\n");
1472 // LICM InsertElementInst sequences.
1473 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1474 e = GatherSeq.end(); it != e; ++it) {
1475 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1480 // Check if this block is inside a loop.
1481 Loop *L = LI->getLoopFor(Insert->getParent());
1485 // Check if it has a preheader.
1486 BasicBlock *PreHeader = L->getLoopPreheader();
1490 // If the vector or the element that we insert into it are
1491 // instructions that are defined in this basic block then we can't
1492 // hoist this instruction.
1493 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1494 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1495 if (CurrVec && L->contains(CurrVec))
1497 if (NewElem && L->contains(NewElem))
1500 // We can hoist this instruction. Move it to the pre-header.
1501 Insert->moveBefore(PreHeader->getTerminator());
1504 // Perform O(N^2) search over the gather sequences and merge identical
1505 // instructions. TODO: We can further optimize this scan if we split the
1506 // instructions into different buckets based on the insert lane.
1507 SmallPtrSet<Instruction*, 16> Visited;
1508 SmallVector<Instruction*, 16> ToRemove;
1509 ReversePostOrderTraversal<Function*> RPOT(F);
1510 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
1511 E = RPOT.end(); I != E; ++I) {
1512 BasicBlock *BB = *I;
1513 // For all instructions in the function:
1514 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1515 Instruction *In = it;
1516 if ((!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) ||
1517 !GatherSeq.count(In))
1520 // Check if we can replace this instruction with any of the
1521 // visited instructions.
1522 for (SmallPtrSet<Instruction*, 16>::iterator v = Visited.begin(),
1523 ve = Visited.end(); v != ve; ++v) {
1524 if (In->isIdenticalTo(*v) &&
1525 DT->dominates((*v)->getParent(), In->getParent())) {
1526 In->replaceAllUsesWith(*v);
1527 ToRemove.push_back(In);
1537 // Erase all of the instructions that we RAUWed.
1538 for (SmallVectorImpl<Instruction *>::iterator v = ToRemove.begin(),
1539 ve = ToRemove.end(); v != ve; ++v) {
1540 assert((*v)->getNumUses() == 0 && "Can't remove instructions with uses");
1541 (*v)->eraseFromParent();
1545 /// The SLPVectorizer Pass.
1546 struct SLPVectorizer : public FunctionPass {
1547 typedef SmallVector<StoreInst *, 8> StoreList;
1548 typedef MapVector<Value *, StoreList> StoreListMap;
1550 /// Pass identification, replacement for typeid
1553 explicit SLPVectorizer() : FunctionPass(ID) {
1554 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1557 ScalarEvolution *SE;
1559 TargetTransformInfo *TTI;
1564 virtual bool runOnFunction(Function &F) {
1565 SE = &getAnalysis<ScalarEvolution>();
1566 DL = getAnalysisIfAvailable<DataLayout>();
1567 TTI = &getAnalysis<TargetTransformInfo>();
1568 AA = &getAnalysis<AliasAnalysis>();
1569 LI = &getAnalysis<LoopInfo>();
1570 DT = &getAnalysis<DominatorTree>();
1573 bool Changed = false;
1575 // Must have DataLayout. We can't require it because some tests run w/o
1580 // Don't vectorize when the attribute NoImplicitFloat is used.
1581 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1584 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1586 // Use the bollom up slp vectorizer to construct chains that start with
1587 // he store instructions.
1588 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1590 // Scan the blocks in the function in post order.
1591 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1592 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1593 BasicBlock *BB = *it;
1595 // Vectorize trees that end at stores.
1596 if (unsigned count = collectStores(BB, R)) {
1598 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1599 Changed |= vectorizeStoreChains(R);
1602 // Vectorize trees that end at reductions.
1603 Changed |= vectorizeChainsInBlock(BB, R);
1607 R.optimizeGatherSequence();
1608 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1609 DEBUG(verifyFunction(F));
1614 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1615 FunctionPass::getAnalysisUsage(AU);
1616 AU.addRequired<ScalarEvolution>();
1617 AU.addRequired<AliasAnalysis>();
1618 AU.addRequired<TargetTransformInfo>();
1619 AU.addRequired<LoopInfo>();
1620 AU.addRequired<DominatorTree>();
1621 AU.addPreserved<LoopInfo>();
1622 AU.addPreserved<DominatorTree>();
1623 AU.setPreservesCFG();
1628 /// \brief Collect memory references and sort them according to their base
1629 /// object. We sort the stores to their base objects to reduce the cost of the
1630 /// quadratic search on the stores. TODO: We can further reduce this cost
1631 /// if we flush the chain creation every time we run into a memory barrier.
1632 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1634 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1635 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1637 /// \brief Try to vectorize a list of operands.
1638 /// \returns true if a value was vectorized.
1639 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1641 /// \brief Try to vectorize a chain that may start at the operands of \V;
1642 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1644 /// \brief Vectorize the stores that were collected in StoreRefs.
1645 bool vectorizeStoreChains(BoUpSLP &R);
1647 /// \brief Scan the basic block and look for patterns that are likely to start
1648 /// a vectorization chain.
1649 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1651 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1654 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1657 StoreListMap StoreRefs;
1660 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1661 int CostThreshold, BoUpSLP &R) {
1662 unsigned ChainLen = Chain.size();
1663 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1665 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1666 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1667 unsigned VF = MinVecRegSize / Sz;
1669 if (!isPowerOf2_32(Sz) || VF < 2)
1672 bool Changed = false;
1673 // Look for profitable vectorizable trees at all offsets, starting at zero.
1674 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1677 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1679 ArrayRef<Value *> Operands = Chain.slice(i, VF);
1681 R.buildTree(Operands);
1683 int Cost = R.getTreeCost();
1685 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
1686 if (Cost < CostThreshold) {
1687 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
1690 // Move to the next bundle.
1699 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
1700 int costThreshold, BoUpSLP &R) {
1701 SetVector<Value *> Heads, Tails;
1702 SmallDenseMap<Value *, Value *> ConsecutiveChain;
1704 // We may run into multiple chains that merge into a single chain. We mark the
1705 // stores that we vectorized so that we don't visit the same store twice.
1706 BoUpSLP::ValueSet VectorizedStores;
1707 bool Changed = false;
1709 // Do a quadratic search on all of the given stores and find
1710 // all of the pairs of stores that follow each other.
1711 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
1712 for (unsigned j = 0; j < e; ++j) {
1716 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
1717 Tails.insert(Stores[j]);
1718 Heads.insert(Stores[i]);
1719 ConsecutiveChain[Stores[i]] = Stores[j];
1724 // For stores that start but don't end a link in the chain:
1725 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
1727 if (Tails.count(*it))
1730 // We found a store instr that starts a chain. Now follow the chain and try
1732 BoUpSLP::ValueList Operands;
1734 // Collect the chain into a list.
1735 while (Tails.count(I) || Heads.count(I)) {
1736 if (VectorizedStores.count(I))
1738 Operands.push_back(I);
1739 // Move to the next value in the chain.
1740 I = ConsecutiveChain[I];
1743 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
1745 // Mark the vectorized stores so that we don't vectorize them again.
1747 VectorizedStores.insert(Operands.begin(), Operands.end());
1748 Changed |= Vectorized;
1755 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
1758 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1759 StoreInst *SI = dyn_cast<StoreInst>(it);
1763 // Check that the pointer points to scalars.
1764 Type *Ty = SI->getValueOperand()->getType();
1765 if (Ty->isAggregateType() || Ty->isVectorTy())
1768 // Find the base of the GEP.
1769 Value *Ptr = SI->getPointerOperand();
1770 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
1771 Ptr = GEP->getPointerOperand();
1773 // Save the store locations.
1774 StoreRefs[Ptr].push_back(SI);
1780 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
1783 Value *VL[] = { A, B };
1784 return tryToVectorizeList(VL, R);
1787 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
1791 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
1793 // Check that all of the parts are scalar instructions of the same type.
1794 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
1798 unsigned Opcode0 = I0->getOpcode();
1800 Type *Ty0 = I0->getType();
1801 unsigned Sz = DL->getTypeSizeInBits(Ty0);
1802 unsigned VF = MinVecRegSize / Sz;
1804 for (int i = 0, e = VL.size(); i < e; ++i) {
1805 Type *Ty = VL[i]->getType();
1806 if (Ty->isAggregateType() || Ty->isVectorTy())
1808 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
1809 if (!Inst || Inst->getOpcode() != Opcode0)
1813 bool Changed = false;
1815 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1816 unsigned OpsWidth = 0;
1823 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
1826 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n");
1827 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
1830 int Cost = R.getTreeCost();
1832 if (Cost < -SLPCostThreshold) {
1833 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n");
1836 // Move to the next bundle.
1845 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
1849 // Try to vectorize V.
1850 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
1853 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
1854 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
1856 if (B && B->hasOneUse()) {
1857 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
1858 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
1859 if (tryToVectorizePair(A, B0, R)) {
1863 if (tryToVectorizePair(A, B1, R)) {
1870 if (A && A->hasOneUse()) {
1871 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
1872 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
1873 if (tryToVectorizePair(A0, B, R)) {
1877 if (tryToVectorizePair(A1, B, R)) {
1885 /// \brief Recognize construction of vectors like
1886 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
1887 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
1888 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
1889 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
1891 /// Returns true if it matches
1893 static bool findBuildVector(InsertElementInst *IE,
1894 SmallVectorImpl<Value *> &Ops) {
1895 if (!isa<UndefValue>(IE->getOperand(0)))
1899 Ops.push_back(IE->getOperand(1));
1901 if (IE->use_empty())
1904 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->use_back());
1908 // If this isn't the final use, make sure the next insertelement is the only
1909 // use. It's OK if the final constructed vector is used multiple times
1910 if (!IE->hasOneUse())
1919 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
1920 bool Changed = false;
1921 SmallVector<Value *, 4> Incoming;
1922 SmallSet<Instruction *, 16> VisitedInstrs;
1924 // Collect the incoming values from the PHIs.
1925 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
1927 PHINode *P = dyn_cast<PHINode>(instr);
1932 // We may go through BB multiple times so skip the one we have checked.
1933 if (!VisitedInstrs.insert(instr))
1936 // Stop constructing the list when you reach a different type.
1937 if (Incoming.size() && P->getType() != Incoming[0]->getType()) {
1938 if (tryToVectorizeList(Incoming, R)) {
1939 // We would like to start over since some instructions are deleted
1940 // and the iterator may become invalid value.
1942 instr = BB->begin();
1949 Incoming.push_back(P);
1952 if (Incoming.size() > 1)
1953 Changed |= tryToVectorizeList(Incoming, R);
1955 VisitedInstrs.clear();
1957 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
1958 // We may go through BB multiple times so skip the one we have checked.
1959 if (!VisitedInstrs.insert(it))
1962 if (isa<DbgInfoIntrinsic>(it))
1965 // Try to vectorize reductions that use PHINodes.
1966 if (PHINode *P = dyn_cast<PHINode>(it)) {
1967 // Check that the PHI is a reduction PHI.
1968 if (P->getNumIncomingValues() != 2)
1971 (P->getIncomingBlock(0) == BB
1972 ? (P->getIncomingValue(0))
1973 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
1974 // Check if this is a Binary Operator.
1975 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
1979 Value *Inst = BI->getOperand(0);
1981 Inst = BI->getOperand(1);
1983 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
1984 // We would like to start over since some instructions are deleted
1985 // and the iterator may become invalid value.
1993 // Try to vectorize trees that start at compare instructions.
1994 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
1995 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
1997 // We would like to start over since some instructions are deleted
1998 // and the iterator may become invalid value.
2004 for (int i = 0; i < 2; ++i) {
2005 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2006 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2008 // We would like to start over since some instructions are deleted
2009 // and the iterator may become invalid value.
2018 // Try to vectorize trees that start at insertelement instructions.
2019 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2020 SmallVector<Value *, 8> Ops;
2021 if (!findBuildVector(IE, Ops))
2024 if (tryToVectorizeList(Ops, R)) {
2037 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2038 bool Changed = false;
2039 // Attempt to sort and vectorize each of the store-groups.
2040 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2042 if (it->second.size() < 2)
2045 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2046 << it->second.size() << ".\n");
2048 // Process the stores in chunks of 16.
2049 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2050 unsigned Len = std::min<unsigned>(CE - CI, 16);
2051 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2052 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2058 } // end anonymous namespace
2060 char SLPVectorizer::ID = 0;
2061 static const char lv_name[] = "SLP Vectorizer";
2062 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2063 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2064 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2065 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2066 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2067 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2070 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }