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 ~BuilderLocGuard() { if (Loc) Builder.SetInsertPoint(Loc); }
66 BuilderLocGuard(const BuilderLocGuard &);
67 BuilderLocGuard &operator=(const BuilderLocGuard &);
69 AssertingVH<Instruction> Loc;
72 /// A helper class for numbering instructions in multible blocks.
73 /// Numbers starts at zero for each basic block.
74 struct BlockNumbering {
76 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
78 BlockNumbering() : BB(0), Valid(false) {}
80 void numberInstructions() {
84 // Number the instructions in the block.
85 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
87 InstrVec.push_back(it);
88 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
93 int getIndex(Instruction *I) {
94 assert(I->getParent() == BB && "Invalid instruction");
97 assert(InstrIdx.count(I) && "Unknown instruction");
101 Instruction *getInstruction(unsigned loc) {
103 numberInstructions();
104 assert(InstrVec.size() > loc && "Invalid Index");
105 return InstrVec[loc];
108 void forget() { Valid = false; }
111 /// The block we are numbering.
113 /// Is the block numbered.
115 /// Maps instructions to numbers and back.
116 SmallDenseMap<Instruction *, int> InstrIdx;
117 /// Maps integers to Instructions.
118 std::vector<Instruction *> InstrVec;
121 /// \returns the parent basic block if all of the instructions in \p VL
122 /// are in the same block or null otherwise.
123 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
124 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
127 BasicBlock *BB = I0->getParent();
128 for (int i = 1, e = VL.size(); i < e; i++) {
129 Instruction *I = dyn_cast<Instruction>(VL[i]);
133 if (BB != I->getParent())
139 /// \returns True if all of the values in \p VL are constants.
140 static bool allConstant(ArrayRef<Value *> VL) {
141 for (unsigned i = 0, e = VL.size(); i < e; ++i)
142 if (!isa<Constant>(VL[i]))
147 /// \returns True if all of the values in \p VL are identical.
148 static bool isSplat(ArrayRef<Value *> VL) {
149 for (unsigned i = 1, e = VL.size(); i < e; ++i)
155 /// \returns The opcode if all of the Instructions in \p VL have the same
157 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
158 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
161 unsigned Opcode = I0->getOpcode();
162 for (int i = 1, e = VL.size(); i < e; i++) {
163 Instruction *I = dyn_cast<Instruction>(VL[i]);
164 if (!I || Opcode != I->getOpcode())
170 /// \returns The type that all of the values in \p VL have or null if there
171 /// are different types.
172 static Type* getSameType(ArrayRef<Value *> VL) {
173 Type *Ty = VL[0]->getType();
174 for (int i = 1, e = VL.size(); i < e; i++)
175 if (VL[0]->getType() != Ty)
181 /// \returns True if the ExtractElement instructions in VL can be vectorized
182 /// to use the original vector.
183 static bool CanReuseExtract(ArrayRef<Value *> VL) {
184 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
185 // Check if all of the extracts come from the same vector and from the
188 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
189 Value *Vec = E0->getOperand(0);
191 // We have to extract from the same vector type.
192 unsigned NElts = Vec->getType()->getVectorNumElements();
194 if (NElts != VL.size())
197 // Check that all of the indices extract from the correct offset.
198 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
199 if (!CI || CI->getZExtValue())
202 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
203 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
204 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
206 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
213 /// Bottom Up SLP Vectorizer.
216 typedef SmallVector<Value *, 8> ValueList;
217 typedef SmallVector<Instruction *, 16> InstrList;
218 typedef SmallPtrSet<Value *, 16> ValueSet;
219 typedef SmallVector<StoreInst *, 8> StoreList;
221 BoUpSLP(Function *Func, ScalarEvolution *Se, DataLayout *Dl,
222 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li,
224 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt),
225 Builder(Se->getContext()) {
226 // Setup the block numbering utility for all of the blocks in the
228 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
230 BlocksNumbers[BB] = BlockNumbering(BB);
234 /// \brief Vectorize the tree that starts with the elements in \p VL.
235 void vectorizeTree();
237 /// \returns the vectorization cost of the subtree that starts at \p VL.
238 /// A negative number means that this is profitable.
241 /// Construct a vectorizable tree that starts at \p Roots.
242 void buildTree(ArrayRef<Value *> Roots);
244 /// Clear the internal data structures that are created by 'buildTree'.
246 VectorizableTree.clear();
247 ScalarToTreeEntry.clear();
249 MemBarrierIgnoreList.clear();
252 /// \returns the scalarization cost for this list of values. Assuming that
253 /// this subtree gets vectorized, we may need to extract the values from the
254 /// roots. This method calculates the cost of extracting the values.
255 int getGatherCost(ArrayRef<Value *> VL);
257 /// \returns true if the memory operations A and B are consecutive.
258 bool isConsecutiveAccess(Value *A, Value *B);
260 /// \brief Perform LICM and CSE on the newly generated gather sequences.
261 void optimizeGatherSequence();
265 /// \returns the cost of the vectorizable entry.
266 int getEntryCost(TreeEntry *E);
268 /// This is the recursive part of buildTree.
269 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
271 /// Vectorizer a single entry in the tree.
272 Value *vectorizeTree(TreeEntry *E);
274 /// Vectorizer a single entry in the tree, starting in \p VL.
275 Value *vectorizeTree(ArrayRef<Value *> VL);
277 /// \brief Take the pointer operand from the Load/Store instruction.
278 /// \returns NULL if this is not a valid Load/Store instruction.
279 static Value *getPointerOperand(Value *I);
281 /// \brief Take the address space operand from the Load/Store instruction.
282 /// \returns -1 if this is not a valid Load/Store instruction.
283 static unsigned getAddressSpaceOperand(Value *I);
285 /// \returns the scalarization cost for this type. Scalarization in this
286 /// context means the creation of vectors from a group of scalars.
287 int getGatherCost(Type *Ty);
289 /// \returns the AA location that is being access by the instruction.
290 AliasAnalysis::Location getLocation(Instruction *I);
292 /// \brief Checks if it is possible to sink an instruction from
293 /// \p Src to \p Dst.
294 /// \returns the pointer to the barrier instruction if we can't sink.
295 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
297 /// \returns the index of the last instrucion in the BB from \p VL.
298 int getLastIndex(ArrayRef<Value *> VL);
300 /// \returns the Instrucion in the bundle \p VL.
301 Instruction *getLastInstruction(ArrayRef<Value *> VL);
303 /// \returns the Instruction at index \p Index which is in Block \p BB.
304 Instruction *getInstructionForIndex(unsigned Index, BasicBlock *BB);
306 /// \returns the index of the first User of \p VL.
307 int getFirstUserIndex(ArrayRef<Value *> VL);
309 /// \returns a vector from a collection of scalars in \p VL.
310 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
313 TreeEntry() : Scalars(), VectorizedValue(0), LastScalarIndex(0),
316 /// \returns true if the scalars in VL are equal to this entry.
317 bool isSame(ArrayRef<Value *> VL) {
318 assert(VL.size() == Scalars.size() && "Invalid size");
319 for (int i = 0, e = VL.size(); i != e; ++i)
320 if (VL[i] != Scalars[i])
325 /// A vector of scalars.
328 /// The Scalars are vectorized into this value. It is initialized to Null.
329 Value *VectorizedValue;
331 /// The index in the basic block of the last scalar.
334 /// Do we need to gather this sequence ?
338 /// Create a new VectorizableTree entry.
339 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
340 VectorizableTree.push_back(TreeEntry());
341 int idx = VectorizableTree.size() - 1;
342 TreeEntry *Last = &VectorizableTree[idx];
343 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
344 Last->NeedToGather = !Vectorized;
346 Last->LastScalarIndex = getLastIndex(VL);
347 for (int i = 0, e = VL.size(); i != e; ++i) {
348 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
349 ScalarToTreeEntry[VL[i]] = idx;
352 Last->LastScalarIndex = 0;
353 MustGather.insert(VL.begin(), VL.end());
358 /// -- Vectorization State --
359 /// Holds all of the tree entries.
360 std::vector<TreeEntry> VectorizableTree;
362 /// Maps a specific scalar to its tree entry.
363 SmallDenseMap<Value*, int> ScalarToTreeEntry;
365 /// A list of scalars that we found that we need to keep as scalars.
368 /// A list of instructions to ignore while sinking
369 /// memory instructions. This map must be reset between runs of getCost.
370 ValueSet MemBarrierIgnoreList;
372 /// Holds all of the instructions that we gathered.
373 SetVector<Instruction *> GatherSeq;
375 /// Numbers instructions in different blocks.
376 std::map<BasicBlock *, BlockNumbering> BlocksNumbers;
378 // Analysis and block reference.
382 TargetTransformInfo *TTI;
386 /// Instruction builder to construct the vectorized tree.
390 void BoUpSLP::buildTree(ArrayRef<Value *> Roots) {
392 buildTree_rec(Roots, 0);
396 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
397 bool SameTy = getSameType(VL); (void)SameTy;
398 assert(SameTy && "Invalid types!");
400 if (Depth == RecursionMaxDepth) {
401 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
402 newTreeEntry(VL, false);
406 // Don't handle vectors.
407 if (VL[0]->getType()->isVectorTy()) {
408 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
409 newTreeEntry(VL, false);
413 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
414 if (SI->getValueOperand()->getType()->isVectorTy()) {
415 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
416 newTreeEntry(VL, false);
420 // If all of the operands are identical or constant we have a simple solution.
421 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
422 !getSameOpcode(VL)) {
423 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
424 newTreeEntry(VL, false);
428 // We now know that this is a vector of instructions of the same type from
431 // Check if this is a duplicate of another entry.
432 if (ScalarToTreeEntry.count(VL[0])) {
433 int Idx = ScalarToTreeEntry[VL[0]];
434 TreeEntry *E = &VectorizableTree[Idx];
435 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
436 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
437 if (E->Scalars[i] != VL[i]) {
438 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
439 newTreeEntry(VL, false);
443 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
447 // Check that none of the instructions in the bundle are already in the tree.
448 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
449 if (ScalarToTreeEntry.count(VL[i])) {
450 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
451 ") is already in tree.\n");
452 newTreeEntry(VL, false);
457 // If any of the scalars appears in the table OR it is marked as a value that
458 // needs to stat scalar then we need to gather the scalars.
459 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
460 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
461 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
462 newTreeEntry(VL, false);
467 // Check that all of the users of the scalars that we want to vectorize are
469 Instruction *VL0 = cast<Instruction>(VL[0]);
470 int MyLastIndex = getLastIndex(VL);
471 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
473 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
474 Instruction *Scalar = cast<Instruction>(VL[i]);
475 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
476 for (Value::use_iterator U = Scalar->use_begin(), UE = Scalar->use_end();
478 DEBUG(dbgs() << "SLP: \tUser " << **U << ". \n");
479 Instruction *User = dyn_cast<Instruction>(*U);
481 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
482 newTreeEntry(VL, false);
486 // We don't care if the user is in a different basic block.
487 BasicBlock *UserBlock = User->getParent();
488 if (UserBlock != BB) {
489 DEBUG(dbgs() << "SLP: User from a different basic block "
494 // If this is a PHINode within this basic block then we can place the
495 // extract wherever we want.
496 if (isa<PHINode>(*User)) {
497 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *User << ". \n");
501 // Check if this is a safe in-tree user.
502 if (ScalarToTreeEntry.count(User)) {
503 int Idx = ScalarToTreeEntry[User];
504 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
505 if (VecLocation <= MyLastIndex) {
506 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
507 newTreeEntry(VL, false);
510 DEBUG(dbgs() << "SLP: In-tree user (" << *User << ") at #" <<
511 VecLocation << " vector value (" << *Scalar << ") at #"
512 << MyLastIndex << ".\n");
516 // Make sure that we can schedule this unknown user.
517 BlockNumbering &BN = BlocksNumbers[BB];
518 int UserIndex = BN.getIndex(User);
519 if (UserIndex < MyLastIndex) {
521 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
523 newTreeEntry(VL, false);
529 // Check that every instructions appears once in this bundle.
530 for (unsigned i = 0, e = VL.size(); i < e; ++i)
531 for (unsigned j = i+1; j < e; ++j)
532 if (VL[i] == VL[j]) {
533 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
534 newTreeEntry(VL, false);
538 // Check that instructions in this bundle don't reference other instructions.
539 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
540 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
541 for (Value::use_iterator U = VL[i]->use_begin(), UE = VL[i]->use_end();
543 for (unsigned j = 0; j < e; ++j) {
544 if (i != j && *U == VL[j]) {
545 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << **U << ". \n");
546 newTreeEntry(VL, false);
553 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
555 unsigned Opcode = getSameOpcode(VL);
557 // Check if it is safe to sink the loads or the stores.
558 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
559 Instruction *Last = getLastInstruction(VL);
561 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
564 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
566 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
567 << "\n because of " << *Barrier << ". Gathering.\n");
568 newTreeEntry(VL, false);
575 case Instruction::PHI: {
576 PHINode *PH = dyn_cast<PHINode>(VL0);
577 newTreeEntry(VL, true);
578 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
580 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
582 // Prepare the operand vector.
583 for (unsigned j = 0; j < VL.size(); ++j)
584 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValue(i));
586 buildTree_rec(Operands, Depth + 1);
590 case Instruction::ExtractElement: {
591 bool Reuse = CanReuseExtract(VL);
593 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
595 newTreeEntry(VL, Reuse);
598 case Instruction::Load: {
599 // Check if the loads are consecutive or of we need to swizzle them.
600 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
601 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
602 newTreeEntry(VL, false);
603 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
607 newTreeEntry(VL, true);
608 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
611 case Instruction::ZExt:
612 case Instruction::SExt:
613 case Instruction::FPToUI:
614 case Instruction::FPToSI:
615 case Instruction::FPExt:
616 case Instruction::PtrToInt:
617 case Instruction::IntToPtr:
618 case Instruction::SIToFP:
619 case Instruction::UIToFP:
620 case Instruction::Trunc:
621 case Instruction::FPTrunc:
622 case Instruction::BitCast: {
623 Type *SrcTy = VL0->getOperand(0)->getType();
624 for (unsigned i = 0; i < VL.size(); ++i) {
625 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
626 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
627 newTreeEntry(VL, false);
628 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
632 newTreeEntry(VL, true);
633 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
635 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
637 // Prepare the operand vector.
638 for (unsigned j = 0; j < VL.size(); ++j)
639 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
641 buildTree_rec(Operands, Depth+1);
645 case Instruction::ICmp:
646 case Instruction::FCmp: {
647 // Check that all of the compares have the same predicate.
648 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
649 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
650 CmpInst *Cmp = cast<CmpInst>(VL[i]);
651 if (Cmp->getPredicate() != P0) {
652 newTreeEntry(VL, false);
653 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
658 newTreeEntry(VL, true);
659 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
661 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
663 // Prepare the operand vector.
664 for (unsigned j = 0; j < VL.size(); ++j)
665 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
667 buildTree_rec(Operands, Depth+1);
671 case Instruction::Select:
672 case Instruction::Add:
673 case Instruction::FAdd:
674 case Instruction::Sub:
675 case Instruction::FSub:
676 case Instruction::Mul:
677 case Instruction::FMul:
678 case Instruction::UDiv:
679 case Instruction::SDiv:
680 case Instruction::FDiv:
681 case Instruction::URem:
682 case Instruction::SRem:
683 case Instruction::FRem:
684 case Instruction::Shl:
685 case Instruction::LShr:
686 case Instruction::AShr:
687 case Instruction::And:
688 case Instruction::Or:
689 case Instruction::Xor: {
690 newTreeEntry(VL, true);
691 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
693 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
695 // Prepare the operand vector.
696 for (unsigned j = 0; j < VL.size(); ++j)
697 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
699 buildTree_rec(Operands, Depth+1);
703 case Instruction::Store: {
704 // Check if the stores are consecutive or of we need to swizzle them.
705 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
706 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
707 newTreeEntry(VL, false);
708 DEBUG(dbgs() << "SLP: Non consecutive store.\n");
712 newTreeEntry(VL, true);
713 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
716 for (unsigned j = 0; j < VL.size(); ++j)
717 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
719 // We can ignore these values because we are sinking them down.
720 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
721 buildTree_rec(Operands, Depth + 1);
725 newTreeEntry(VL, false);
726 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
731 int BoUpSLP::getEntryCost(TreeEntry *E) {
732 ArrayRef<Value*> VL = E->Scalars;
734 Type *ScalarTy = VL[0]->getType();
735 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
736 ScalarTy = SI->getValueOperand()->getType();
737 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
739 if (E->NeedToGather) {
743 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
745 return getGatherCost(E->Scalars);
748 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
750 Instruction *VL0 = cast<Instruction>(VL[0]);
751 unsigned Opcode = VL0->getOpcode();
753 case Instruction::PHI: {
756 case Instruction::ExtractElement: {
757 if (CanReuseExtract(VL))
759 return getGatherCost(VecTy);
761 case Instruction::ZExt:
762 case Instruction::SExt:
763 case Instruction::FPToUI:
764 case Instruction::FPToSI:
765 case Instruction::FPExt:
766 case Instruction::PtrToInt:
767 case Instruction::IntToPtr:
768 case Instruction::SIToFP:
769 case Instruction::UIToFP:
770 case Instruction::Trunc:
771 case Instruction::FPTrunc:
772 case Instruction::BitCast: {
773 Type *SrcTy = VL0->getOperand(0)->getType();
775 // Calculate the cost of this instruction.
776 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
777 VL0->getType(), SrcTy);
779 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
780 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
781 return VecCost - ScalarCost;
783 case Instruction::FCmp:
784 case Instruction::ICmp:
785 case Instruction::Select:
786 case Instruction::Add:
787 case Instruction::FAdd:
788 case Instruction::Sub:
789 case Instruction::FSub:
790 case Instruction::Mul:
791 case Instruction::FMul:
792 case Instruction::UDiv:
793 case Instruction::SDiv:
794 case Instruction::FDiv:
795 case Instruction::URem:
796 case Instruction::SRem:
797 case Instruction::FRem:
798 case Instruction::Shl:
799 case Instruction::LShr:
800 case Instruction::AShr:
801 case Instruction::And:
802 case Instruction::Or:
803 case Instruction::Xor: {
804 // Calculate the cost of this instruction.
807 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
808 Opcode == Instruction::Select) {
809 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
810 ScalarCost = VecTy->getNumElements() *
811 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
812 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
814 ScalarCost = VecTy->getNumElements() *
815 TTI->getArithmeticInstrCost(Opcode, ScalarTy);
816 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
818 return VecCost - ScalarCost;
820 case Instruction::Load: {
821 // Cost of wide load - cost of scalar loads.
822 int ScalarLdCost = VecTy->getNumElements() *
823 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
824 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
825 return VecLdCost - ScalarLdCost;
827 case Instruction::Store: {
828 // We know that we can merge the stores. Calculate the cost.
829 int ScalarStCost = VecTy->getNumElements() *
830 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
831 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
832 return VecStCost - ScalarStCost;
835 llvm_unreachable("Unknown instruction");
839 int BoUpSLP::getTreeCost() {
841 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
842 VectorizableTree.size() << ".\n");
844 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
845 int C = getEntryCost(&VectorizableTree[i]);
846 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
847 << *VectorizableTree[i].Scalars[0] << " .\n");
850 DEBUG(dbgs() << "SLP: Total Cost " << Cost << ".\n");
854 int BoUpSLP::getGatherCost(Type *Ty) {
856 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
857 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
861 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
862 // Find the type of the operands in VL.
863 Type *ScalarTy = VL[0]->getType();
864 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
865 ScalarTy = SI->getValueOperand()->getType();
866 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
867 // Find the cost of inserting/extracting values from the vector.
868 return getGatherCost(VecTy);
871 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
872 if (StoreInst *SI = dyn_cast<StoreInst>(I))
873 return AA->getLocation(SI);
874 if (LoadInst *LI = dyn_cast<LoadInst>(I))
875 return AA->getLocation(LI);
876 return AliasAnalysis::Location();
879 Value *BoUpSLP::getPointerOperand(Value *I) {
880 if (LoadInst *LI = dyn_cast<LoadInst>(I))
881 return LI->getPointerOperand();
882 if (StoreInst *SI = dyn_cast<StoreInst>(I))
883 return SI->getPointerOperand();
887 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
888 if (LoadInst *L = dyn_cast<LoadInst>(I))
889 return L->getPointerAddressSpace();
890 if (StoreInst *S = dyn_cast<StoreInst>(I))
891 return S->getPointerAddressSpace();
895 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
896 Value *PtrA = getPointerOperand(A);
897 Value *PtrB = getPointerOperand(B);
898 unsigned ASA = getAddressSpaceOperand(A);
899 unsigned ASB = getAddressSpaceOperand(B);
901 // Check that the address spaces match and that the pointers are valid.
902 if (!PtrA || !PtrB || (ASA != ASB))
905 // Check that A and B are of the same type.
906 if (PtrA->getType() != PtrB->getType())
909 // Calculate the distance.
910 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
911 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
912 const SCEV *OffsetSCEV = SE->getMinusSCEV(PtrSCEVA, PtrSCEVB);
913 const SCEVConstant *ConstOffSCEV = dyn_cast<SCEVConstant>(OffsetSCEV);
915 // Non constant distance.
919 int64_t Offset = ConstOffSCEV->getValue()->getSExtValue();
920 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
921 // The Instructions are connsecutive if the size of the first load/store is
922 // the same as the offset.
923 int64_t Sz = DL->getTypeStoreSize(Ty);
924 return ((-Offset) == Sz);
927 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
928 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
929 BasicBlock::iterator I = Src, E = Dst;
930 /// Scan all of the instruction from SRC to DST and check if
931 /// the source may alias.
932 for (++I; I != E; ++I) {
933 // Ignore store instructions that are marked as 'ignore'.
934 if (MemBarrierIgnoreList.count(I))
936 if (Src->mayWriteToMemory()) /* Write */ {
937 if (!I->mayReadOrWriteMemory())
940 if (!I->mayWriteToMemory())
943 AliasAnalysis::Location A = getLocation(&*I);
944 AliasAnalysis::Location B = getLocation(Src);
946 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
952 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
953 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
954 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
955 BlockNumbering &BN = BlocksNumbers[BB];
957 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
958 for (unsigned i = 0, e = VL.size(); i < e; ++i)
959 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
963 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
964 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
965 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
966 BlockNumbering &BN = BlocksNumbers[BB];
968 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
969 for (unsigned i = 1, e = VL.size(); i < e; ++i)
970 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
971 Instruction *I = BN.getInstruction(MaxIdx);
972 assert(I && "bad location");
976 Instruction *BoUpSLP::getInstructionForIndex(unsigned Index, BasicBlock *BB) {
977 BlockNumbering &BN = BlocksNumbers[BB];
978 return BN.getInstruction(Index);
981 int BoUpSLP::getFirstUserIndex(ArrayRef<Value *> VL) {
982 BasicBlock *BB = getSameBlock(VL);
983 assert(BB && "All instructions must come from the same block");
984 BlockNumbering &BN = BlocksNumbers[BB];
986 // Find the first user of the values.
987 int FirstUser = BN.getIndex(BB->getTerminator());
988 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
989 for (Value::use_iterator U = VL[i]->use_begin(), UE = VL[i]->use_end();
991 Instruction *Instr = dyn_cast<Instruction>(*U);
993 if (!Instr || Instr->getParent() != BB)
996 FirstUser = std::min(FirstUser, BN.getIndex(Instr));
1002 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1003 Value *Vec = UndefValue::get(Ty);
1004 // Generate the 'InsertElement' instruction.
1005 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1006 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1007 if (Instruction *I = dyn_cast<Instruction>(Vec))
1008 GatherSeq.insert(I);
1014 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1015 if (ScalarToTreeEntry.count(VL[0])) {
1016 int Idx = ScalarToTreeEntry[VL[0]];
1017 TreeEntry *E = &VectorizableTree[Idx];
1019 return vectorizeTree(E);
1022 Type *ScalarTy = VL[0]->getType();
1023 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1024 ScalarTy = SI->getValueOperand()->getType();
1025 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1027 return Gather(VL, VecTy);
1030 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1031 BuilderLocGuard Guard(Builder);
1033 if (E->VectorizedValue) {
1034 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1035 return E->VectorizedValue;
1038 Type *ScalarTy = E->Scalars[0]->getType();
1039 if (StoreInst *SI = dyn_cast<StoreInst>(E->Scalars[0]))
1040 ScalarTy = SI->getValueOperand()->getType();
1041 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1043 if (E->NeedToGather) {
1044 return Gather(E->Scalars, VecTy);
1047 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1048 unsigned Opcode = VL0->getOpcode();
1049 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1052 case Instruction::PHI: {
1053 PHINode *PH = dyn_cast<PHINode>(VL0);
1054 Builder.SetInsertPoint(PH->getParent()->getFirstInsertionPt());
1055 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1056 E->VectorizedValue = NewPhi;
1058 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1060 BasicBlock *IBB = PH->getIncomingBlock(i);
1062 // Prepare the operand vector.
1063 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1064 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1065 getIncomingValueForBlock(IBB));
1067 Builder.SetInsertPoint(IBB->getTerminator());
1068 Value *Vec = vectorizeTree(Operands);
1069 NewPhi->addIncoming(Vec, IBB);
1072 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1073 "Invalid number of incoming values");
1077 case Instruction::ExtractElement: {
1078 if (CanReuseExtract(E->Scalars)) {
1079 Value *V = VL0->getOperand(0);
1080 E->VectorizedValue = V;
1083 return Gather(E->Scalars, VecTy);
1085 case Instruction::ZExt:
1086 case Instruction::SExt:
1087 case Instruction::FPToUI:
1088 case Instruction::FPToSI:
1089 case Instruction::FPExt:
1090 case Instruction::PtrToInt:
1091 case Instruction::IntToPtr:
1092 case Instruction::SIToFP:
1093 case Instruction::UIToFP:
1094 case Instruction::Trunc:
1095 case Instruction::FPTrunc:
1096 case Instruction::BitCast: {
1098 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1099 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1101 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1102 Value *InVec = vectorizeTree(INVL);
1103 CastInst *CI = dyn_cast<CastInst>(VL0);
1104 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1105 E->VectorizedValue = V;
1108 case Instruction::FCmp:
1109 case Instruction::ICmp: {
1110 ValueList LHSV, RHSV;
1111 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1112 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1113 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1116 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1117 Value *L = vectorizeTree(LHSV);
1118 Value *R = vectorizeTree(RHSV);
1121 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1122 if (Opcode == Instruction::FCmp)
1123 V = Builder.CreateFCmp(P0, L, R);
1125 V = Builder.CreateICmp(P0, L, R);
1127 E->VectorizedValue = V;
1130 case Instruction::Select: {
1131 ValueList TrueVec, FalseVec, CondVec;
1132 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1133 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1134 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1135 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1138 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1139 Value *Cond = vectorizeTree(CondVec);
1140 Value *True = vectorizeTree(TrueVec);
1141 Value *False = vectorizeTree(FalseVec);
1142 Value *V = Builder.CreateSelect(Cond, True, False);
1143 E->VectorizedValue = V;
1146 case Instruction::Add:
1147 case Instruction::FAdd:
1148 case Instruction::Sub:
1149 case Instruction::FSub:
1150 case Instruction::Mul:
1151 case Instruction::FMul:
1152 case Instruction::UDiv:
1153 case Instruction::SDiv:
1154 case Instruction::FDiv:
1155 case Instruction::URem:
1156 case Instruction::SRem:
1157 case Instruction::FRem:
1158 case Instruction::Shl:
1159 case Instruction::LShr:
1160 case Instruction::AShr:
1161 case Instruction::And:
1162 case Instruction::Or:
1163 case Instruction::Xor: {
1164 ValueList LHSVL, RHSVL;
1165 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1166 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1167 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1170 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1171 Value *LHS = vectorizeTree(LHSVL);
1172 Value *RHS = vectorizeTree(RHSVL);
1174 if (LHS == RHS && isa<Instruction>(LHS)) {
1175 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1178 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1179 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1180 E->VectorizedValue = V;
1183 case Instruction::Load: {
1184 // Loads are inserted at the head of the tree because we don't want to
1185 // sink them all the way down past store instructions.
1186 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1187 LoadInst *LI = cast<LoadInst>(VL0);
1189 Builder.CreateBitCast(LI->getPointerOperand(), VecTy->getPointerTo());
1190 unsigned Alignment = LI->getAlignment();
1191 LI = Builder.CreateLoad(VecPtr);
1192 LI->setAlignment(Alignment);
1193 E->VectorizedValue = LI;
1196 case Instruction::Store: {
1197 StoreInst *SI = cast<StoreInst>(VL0);
1198 unsigned Alignment = SI->getAlignment();
1201 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1202 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1204 Builder.SetInsertPoint(getLastInstruction(E->Scalars));
1205 Value *VecValue = vectorizeTree(ValueOp);
1207 Builder.CreateBitCast(SI->getPointerOperand(), VecTy->getPointerTo());
1208 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1209 S->setAlignment(Alignment);
1210 E->VectorizedValue = S;
1214 llvm_unreachable("unknown inst");
1219 void BoUpSLP::vectorizeTree() {
1220 Builder.SetInsertPoint(F->getEntryBlock().begin());
1221 vectorizeTree(&VectorizableTree[0]);
1223 // For each vectorized value:
1224 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1225 TreeEntry *Entry = &VectorizableTree[EIdx];
1228 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1229 Value *Scalar = Entry->Scalars[Lane];
1231 // No need to handle users of gathered values.
1232 if (Entry->NeedToGather)
1235 Value *Vec = Entry->VectorizedValue;
1236 assert(Vec && "Can't find vectorizable value");
1238 SmallVector<User*, 16> Users(Scalar->use_begin(), Scalar->use_end());
1240 for (SmallVector<User*, 16>::iterator User = Users.begin(),
1241 UE = Users.end(); User != UE; ++User) {
1242 DEBUG(dbgs() << "SLP: \tupdating user " << **User << ".\n");
1244 bool Gathered = MustGather.count(*User);
1246 // Skip in-tree scalars that become vectors.
1247 if (ScalarToTreeEntry.count(*User) && !Gathered) {
1248 DEBUG(dbgs() << "SLP: \tUser will be removed soon:" <<
1250 int Idx = ScalarToTreeEntry[*User]; (void) Idx;
1251 assert(!VectorizableTree[Idx].NeedToGather && "bad state ?");
1255 if (!isa<Instruction>(*User))
1258 // Generate extracts for out-of-tree users.
1259 // Find the insertion point for the extractelement lane.
1260 Instruction *Loc = 0;
1261 if (PHINode *PN = dyn_cast<PHINode>(Vec)) {
1262 Loc = PN->getParent()->getFirstInsertionPt();
1263 } else if (Instruction *Iv = dyn_cast<Instruction>(Vec)){
1264 Loc = ++((BasicBlock::iterator)*Iv);
1266 Loc = F->getEntryBlock().begin();
1269 Builder.SetInsertPoint(Loc);
1270 Value *Ex = Builder.CreateExtractElement(Vec, Builder.getInt32(Lane));
1271 (*User)->replaceUsesOfWith(Scalar, Ex);
1272 DEBUG(dbgs() << "SLP: \tupdated user:" << **User << ".\n");
1275 Type *Ty = Scalar->getType();
1276 if (!Ty->isVoidTy()) {
1277 for (Value::use_iterator User = Scalar->use_begin(), UE = Scalar->use_end();
1278 User != UE; ++User) {
1279 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n");
1280 assert(!MustGather.count(*User) &&
1281 "Replacing gathered value with undef");
1282 assert(ScalarToTreeEntry.count(*User) &&
1283 "Replacing out-of-tree value with undef");
1285 Value *Undef = UndefValue::get(Ty);
1286 Scalar->replaceAllUsesWith(Undef);
1288 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1289 cast<Instruction>(Scalar)->eraseFromParent();
1293 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1294 BlocksNumbers[it].forget();
1296 Builder.ClearInsertionPoint();
1299 void BoUpSLP::optimizeGatherSequence() {
1300 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1301 << " gather sequences instructions.\n");
1302 // LICM InsertElementInst sequences.
1303 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1304 e = GatherSeq.end(); it != e; ++it) {
1305 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1310 // Check if this block is inside a loop.
1311 Loop *L = LI->getLoopFor(Insert->getParent());
1315 // Check if it has a preheader.
1316 BasicBlock *PreHeader = L->getLoopPreheader();
1320 // If the vector or the element that we insert into it are
1321 // instructions that are defined in this basic block then we can't
1322 // hoist this instruction.
1323 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1324 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1325 if (CurrVec && L->contains(CurrVec))
1327 if (NewElem && L->contains(NewElem))
1330 // We can hoist this instruction. Move it to the pre-header.
1331 Insert->moveBefore(PreHeader->getTerminator());
1334 // Perform O(N^2) search over the gather sequences and merge identical
1335 // instructions. TODO: We can further optimize this scan if we split the
1336 // instructions into different buckets based on the insert lane.
1337 SmallPtrSet<Instruction*, 16> Visited;
1338 SmallVector<Instruction*, 16> ToRemove;
1339 ReversePostOrderTraversal<Function*> RPOT(F);
1340 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
1341 E = RPOT.end(); I != E; ++I) {
1342 BasicBlock *BB = *I;
1343 // For all instructions in the function:
1344 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1345 InsertElementInst *Insert = dyn_cast<InsertElementInst>(it);
1346 if (!Insert || !GatherSeq.count(Insert))
1349 // Check if we can replace this instruction with any of the
1350 // visited instructions.
1351 for (SmallPtrSet<Instruction*, 16>::iterator v = Visited.begin(),
1352 ve = Visited.end(); v != ve; ++v) {
1353 if (Insert->isIdenticalTo(*v) &&
1354 DT->dominates((*v)->getParent(), Insert->getParent())) {
1355 Insert->replaceAllUsesWith(*v);
1356 ToRemove.push_back(Insert);
1362 Visited.insert(Insert);
1366 // Erase all of the instructions that we RAUWed.
1367 for (SmallVectorImpl<Instruction *>::iterator v = ToRemove.begin(),
1368 ve = ToRemove.end(); v != ve; ++v) {
1369 assert((*v)->getNumUses() == 0 && "Can't remove instructions with uses");
1370 (*v)->eraseFromParent();
1374 /// The SLPVectorizer Pass.
1375 struct SLPVectorizer : public FunctionPass {
1376 typedef SmallVector<StoreInst *, 8> StoreList;
1377 typedef MapVector<Value *, StoreList> StoreListMap;
1379 /// Pass identification, replacement for typeid
1382 explicit SLPVectorizer() : FunctionPass(ID) {
1383 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1386 ScalarEvolution *SE;
1388 TargetTransformInfo *TTI;
1393 virtual bool runOnFunction(Function &F) {
1394 SE = &getAnalysis<ScalarEvolution>();
1395 DL = getAnalysisIfAvailable<DataLayout>();
1396 TTI = &getAnalysis<TargetTransformInfo>();
1397 AA = &getAnalysis<AliasAnalysis>();
1398 LI = &getAnalysis<LoopInfo>();
1399 DT = &getAnalysis<DominatorTree>();
1402 bool Changed = false;
1404 // Must have DataLayout. We can't require it because some tests run w/o
1409 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1411 // Use the bollom up slp vectorizer to construct chains that start with
1412 // he store instructions.
1413 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT);
1415 // Scan the blocks in the function in post order.
1416 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1417 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1418 BasicBlock *BB = *it;
1420 // Vectorize trees that end at reductions.
1421 Changed |= vectorizeChainsInBlock(BB, R);
1423 // Vectorize trees that end at stores.
1424 if (unsigned count = collectStores(BB, R)) {
1426 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1427 Changed |= vectorizeStoreChains(R);
1432 R.optimizeGatherSequence();
1433 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1434 DEBUG(verifyFunction(F));
1439 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1440 FunctionPass::getAnalysisUsage(AU);
1441 AU.addRequired<ScalarEvolution>();
1442 AU.addRequired<AliasAnalysis>();
1443 AU.addRequired<TargetTransformInfo>();
1444 AU.addRequired<LoopInfo>();
1445 AU.addRequired<DominatorTree>();
1446 AU.addPreserved<LoopInfo>();
1447 AU.addPreserved<DominatorTree>();
1448 AU.setPreservesCFG();
1453 /// \brief Collect memory references and sort them according to their base
1454 /// object. We sort the stores to their base objects to reduce the cost of the
1455 /// quadratic search on the stores. TODO: We can further reduce this cost
1456 /// if we flush the chain creation every time we run into a memory barrier.
1457 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1459 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1460 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1462 /// \brief Try to vectorize a list of operands. If \p NeedExtracts is true
1463 /// then we calculate the cost of extracting the scalars from the vector.
1464 /// \returns true if a value was vectorized.
1465 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, bool NeedExtracts);
1467 /// \brief Try to vectorize a chain that may start at the operands of \V;
1468 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1470 /// \brief Vectorize the stores that were collected in StoreRefs.
1471 bool vectorizeStoreChains(BoUpSLP &R);
1473 /// \brief Scan the basic block and look for patterns that are likely to start
1474 /// a vectorization chain.
1475 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1477 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1480 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1483 StoreListMap StoreRefs;
1486 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1487 int CostThreshold, BoUpSLP &R) {
1488 unsigned ChainLen = Chain.size();
1489 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
1491 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
1492 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
1493 unsigned VF = MinVecRegSize / Sz;
1495 if (!isPowerOf2_32(Sz) || VF < 2)
1498 bool Changed = false;
1499 // Look for profitable vectorizable trees at all offsets, starting at zero.
1500 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
1503 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
1505 ArrayRef<Value *> Operands = Chain.slice(i, VF);
1507 R.buildTree(Operands);
1509 int Cost = R.getTreeCost();
1511 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
1512 if (Cost < CostThreshold) {
1513 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
1516 // Move to the next bundle.
1522 if (Changed || ChainLen > VF)
1525 // Handle short chains. This helps us catch types such as <3 x float> that
1526 // are smaller than vector size.
1529 int Cost = R.getTreeCost();
1531 if (Cost < CostThreshold) {
1532 DEBUG(dbgs() << "SLP: Found store chain cost = " << Cost
1533 << " for size = " << ChainLen << "\n");
1541 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
1542 int costThreshold, BoUpSLP &R) {
1543 SetVector<Value *> Heads, Tails;
1544 SmallDenseMap<Value *, Value *> ConsecutiveChain;
1546 // We may run into multiple chains that merge into a single chain. We mark the
1547 // stores that we vectorized so that we don't visit the same store twice.
1548 BoUpSLP::ValueSet VectorizedStores;
1549 bool Changed = false;
1551 // Do a quadratic search on all of the given stores and find
1552 // all of the pairs of loads that follow each other.
1553 for (unsigned i = 0, e = Stores.size(); i < e; ++i)
1554 for (unsigned j = 0; j < e; ++j) {
1558 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
1559 Tails.insert(Stores[j]);
1560 Heads.insert(Stores[i]);
1561 ConsecutiveChain[Stores[i]] = Stores[j];
1565 // For stores that start but don't end a link in the chain:
1566 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
1568 if (Tails.count(*it))
1571 // We found a store instr that starts a chain. Now follow the chain and try
1573 BoUpSLP::ValueList Operands;
1575 // Collect the chain into a list.
1576 while (Tails.count(I) || Heads.count(I)) {
1577 if (VectorizedStores.count(I))
1579 Operands.push_back(I);
1580 // Move to the next value in the chain.
1581 I = ConsecutiveChain[I];
1584 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
1586 // Mark the vectorized stores so that we don't vectorize them again.
1588 VectorizedStores.insert(Operands.begin(), Operands.end());
1589 Changed |= Vectorized;
1596 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
1599 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1600 StoreInst *SI = dyn_cast<StoreInst>(it);
1604 // Check that the pointer points to scalars.
1605 Type *Ty = SI->getValueOperand()->getType();
1606 if (Ty->isAggregateType() || Ty->isVectorTy())
1609 // Find the base of the GEP.
1610 Value *Ptr = SI->getPointerOperand();
1611 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
1612 Ptr = GEP->getPointerOperand();
1614 // Save the store locations.
1615 StoreRefs[Ptr].push_back(SI);
1621 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
1624 Value *VL[] = { A, B };
1625 return tryToVectorizeList(VL, R, true);
1628 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
1629 bool NeedExtracts) {
1633 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
1635 // Check that all of the parts are scalar instructions of the same type.
1636 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
1640 unsigned Opcode0 = I0->getOpcode();
1642 for (int i = 0, e = VL.size(); i < e; ++i) {
1643 Type *Ty = VL[i]->getType();
1644 if (Ty->isAggregateType() || Ty->isVectorTy())
1646 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
1647 if (!Inst || Inst->getOpcode() != Opcode0)
1652 int Cost = R.getTreeCost();
1654 int ExtrCost = NeedExtracts ? R.getGatherCost(VL) : 0;
1655 DEBUG(dbgs() << "SLP: Cost of pair:" << Cost
1656 << " Cost of extract:" << ExtrCost << ".\n");
1657 if ((Cost + ExtrCost) >= -SLPCostThreshold)
1659 DEBUG(dbgs() << "SLP: Vectorizing pair.\n");
1664 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
1668 // Try to vectorize V.
1669 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
1672 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
1673 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
1675 if (B && B->hasOneUse()) {
1676 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
1677 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
1678 if (tryToVectorizePair(A, B0, R)) {
1682 if (tryToVectorizePair(A, B1, R)) {
1689 if (A && A->hasOneUse()) {
1690 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
1691 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
1692 if (tryToVectorizePair(A0, B, R)) {
1696 if (tryToVectorizePair(A1, B, R)) {
1704 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
1705 bool Changed = false;
1706 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1707 if (isa<DbgInfoIntrinsic>(it))
1710 // Try to vectorize reductions that use PHINodes.
1711 if (PHINode *P = dyn_cast<PHINode>(it)) {
1712 // Check that the PHI is a reduction PHI.
1713 if (P->getNumIncomingValues() != 2)
1716 (P->getIncomingBlock(0) == BB
1717 ? (P->getIncomingValue(0))
1718 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0));
1719 // Check if this is a Binary Operator.
1720 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
1724 Value *Inst = BI->getOperand(0);
1726 Inst = BI->getOperand(1);
1728 Changed |= tryToVectorize(dyn_cast<BinaryOperator>(Inst), R);
1732 // Try to vectorize trees that start at compare instructions.
1733 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
1734 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
1738 for (int i = 0; i < 2; ++i)
1739 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i)))
1741 tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R);
1746 // Scan the PHINodes in our successors in search for pairing hints.
1747 for (succ_iterator it = succ_begin(BB), e = succ_end(BB); it != e; ++it) {
1748 BasicBlock *Succ = *it;
1749 SmallVector<Value *, 4> Incoming;
1751 // Collect the incoming values from the PHIs.
1752 for (BasicBlock::iterator instr = Succ->begin(), ie = Succ->end();
1753 instr != ie; ++instr) {
1754 PHINode *P = dyn_cast<PHINode>(instr);
1759 Value *V = P->getIncomingValueForBlock(BB);
1760 if (Instruction *I = dyn_cast<Instruction>(V))
1761 if (I->getParent() == BB)
1762 Incoming.push_back(I);
1765 if (Incoming.size() > 1)
1766 Changed |= tryToVectorizeList(Incoming, R, true);
1772 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
1773 bool Changed = false;
1774 // Attempt to sort and vectorize each of the store-groups.
1775 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
1777 if (it->second.size() < 2)
1780 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
1781 << it->second.size() << ".\n");
1783 Changed |= vectorizeStores(it->second, -SLPCostThreshold, R);
1788 } // end anonymous namespace
1790 char SLPVectorizer::ID = 0;
1791 static const char lv_name[] = "SLP Vectorizer";
1792 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
1793 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1794 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
1795 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1796 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1797 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
1800 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }