1 //===- LoopVectorize.cpp - A Loop 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 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR. Legalization of the IR is done
12 // in the codegen. However, the vectorizes uses (will use) the codegen
13 // interfaces to generate IR that is likely to result in an optimal binary.
15 // The loop vectorizer combines consecutive loop iteration into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A helper class that checks for the legality
22 // of the vectorization.
23 // 3. SingleBlockLoopVectorizer - A helper class that performs the actual
24 // widening of instructions.
25 //===----------------------------------------------------------------------===//
27 // The reduction-variable vectorization is based on the paper:
28 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
30 // Variable uniformity checks are inspired by:
31 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
33 // Other ideas/concepts are from:
34 // A. Zaks and D. Nuzman. Autovectorization in GCC—two years later.
36 //===----------------------------------------------------------------------===//
37 #define LV_NAME "loop-vectorize"
38 #define DEBUG_TYPE LV_NAME
39 #include "llvm/Constants.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/Instructions.h"
42 #include "llvm/LLVMContext.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Value.h"
46 #include "llvm/Function.h"
47 #include "llvm/Analysis/Verifier.h"
48 #include "llvm/Module.h"
49 #include "llvm/Type.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/StringExtras.h"
52 #include "llvm/Analysis/AliasAnalysis.h"
53 #include "llvm/Analysis/AliasSetTracker.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Analysis/ScalarEvolution.h"
56 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
57 #include "llvm/Analysis/ScalarEvolutionExpander.h"
58 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
59 #include "llvm/Analysis/ValueTracking.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Support/CommandLine.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/DataLayout.h"
65 #include "llvm/Transforms/Utils/Local.h"
69 static cl::opt<unsigned>
70 DefaultVectorizationFactor("default-loop-vectorize-width",
71 cl::init(4), cl::Hidden,
72 cl::desc("Set the default loop vectorization width"));
75 // Forward declaration.
76 class LoopVectorizationLegality;
78 /// Vectorize a simple loop. This class performs the widening of simple single
79 /// basic block loops into vectors. It does not perform any
80 /// vectorization-legality checks, and just does it. It widens the vectors
81 /// to a given vectorization factor (VF).
82 class SingleBlockLoopVectorizer {
85 SingleBlockLoopVectorizer(Loop *OrigLoop, ScalarEvolution *Se, LoopInfo *Li,
86 LPPassManager *Lpm, unsigned VecWidth):
87 Orig(OrigLoop), SE(Se), LI(Li), LPM(Lpm), VF(VecWidth),
88 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
90 // Perform the actual loop widening (vectorization).
91 void vectorize(LoopVectorizationLegality *Legal) {
92 ///Create a new empty loop. Unlink the old loop and connect the new one.
93 createEmptyLoop(Legal);
94 /// Widen each instruction in the old loop to a new one in the new loop.
95 /// Use the Legality module to find the induction and reduction variables.
97 // register the new loop.
102 /// Create an empty loop, based on the loop ranges of the old loop.
103 void createEmptyLoop(LoopVectorizationLegality *Legal);
104 /// Copy and widen the instructions from the old loop.
105 void vectorizeLoop(LoopVectorizationLegality *Legal);
106 /// Insert the new loop to the loop hierarchy and pass manager.
109 /// This instruction is un-vectorizable. Implement it as a sequence
111 void scalarizeInstruction(Instruction *Instr);
113 /// Create a broadcast instruction. This method generates a broadcast
114 /// instruction (shuffle) for loop invariant values and for the induction
115 /// value. If this is the induction variable then we extend it to N, N+1, ...
116 /// this is needed because each iteration in the loop corresponds to a SIMD
118 Value *getBroadcastInstrs(Value *V);
120 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
121 /// for each element in the vector. Starting from zero.
122 Value *getConsecutiveVector(Value* Val);
124 /// When we go over instructions in the basic block we rely on previous
125 /// values within the current basic block or on loop invariant values.
126 /// When we widen (vectorize) values we place them in the map. If the values
127 /// are not within the map, they have to be loop invariant, so we simply
128 /// broadcast them into a vector.
129 Value *getVectorValue(Value *V);
131 /// Get a uniform vector of constant integers. We use this to get
132 /// vectors of ones and zeros for the reduction code.
133 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
135 typedef DenseMap<Value*, Value*> ValueMap;
137 /// The original loop.
139 // Scev analysis to use.
143 // Loop Pass Manager;
145 // The vectorization factor to use.
148 // The builder that we use
151 // --- Vectorization state ---
153 /// Middle Block between the vector and the scalar.
154 BasicBlock *LoopMiddleBlock;
155 ///The ExitBlock of the scalar loop.
156 BasicBlock *LoopExitBlock;
157 ///The vector loop body.
158 BasicBlock *LoopVectorBody;
159 ///The scalar loop body.
160 BasicBlock *LoopScalarBody;
161 ///The first bypass block.
162 BasicBlock *LoopBypassBlock;
164 /// The new Induction variable which was added to the new block.
166 /// The induction variable of the old basic block.
167 PHINode *OldInduction;
168 // Maps scalars to widened vectors.
172 /// Perform the vectorization legality check. This class does not look at the
173 /// profitability of vectorization, only the legality. At the moment the checks
174 /// are very simple and focus on single basic block loops with a constant
175 /// iteration count and no reductions.
176 class LoopVectorizationLegality {
178 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
179 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
181 /// This represents the kinds of reductions that we support.
182 /// We use the enum values to hold the 'identity' value for
183 /// each operand. This value does not change the result if applied.
185 NoReduction = -1, /// Not a reduction.
186 IntegerAdd = 0, /// Sum of numbers.
187 IntegerMult = 1 /// Product of numbers.
190 /// This POD struct holds information about reduction variables.
191 struct ReductionDescriptor {
193 ReductionDescriptor():
194 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
197 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
198 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
200 // The starting value of the reduction.
201 // It does not have to be zero!
203 // The instruction who's value is used outside the loop.
204 Instruction *LoopExitInstr;
205 // The kind of the reduction.
209 /// ReductionList contains the reduction descriptors for all
210 /// of the reductions that were found in the loop.
211 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
213 /// Returns the maximum vectorization factor that we *can* use to vectorize
214 /// this loop. This does not mean that it is profitable to vectorize this
215 /// loop, only that it is legal to do so. This may be a large number. We
216 /// can vectorize to any SIMD width below this number.
217 unsigned getLoopMaxVF();
219 /// Returns the Induction variable.
220 PHINode *getInduction() {return Induction;}
222 /// Returns the reduction variables found in the loop.
223 ReductionList *getReductionVars() { return &Reductions; }
225 /// Check that the GEP operands are all uniform except for the last index
226 /// which has to be the induction variable.
227 bool isConsecutiveGep(Value *Ptr);
230 /// Check if a single basic block loop is vectorizable.
231 /// At this point we know that this is a loop with a constant trip count
232 /// and we only need to check individual instructions.
233 bool canVectorizeBlock(BasicBlock &BB);
235 /// When we vectorize loops we may change the order in which
236 /// we read and write from memory. This method checks if it is
237 /// legal to vectorize the code, considering only memory constrains.
238 /// Returns true if BB is vectorizable
239 bool canVectorizeMemory(BasicBlock &BB);
241 // Check if a pointer value is known to be disjoint.
242 // Example: Alloca, Global, NoAlias.
243 bool isIdentifiedSafeObject(Value* Val);
245 /// Returns True, if 'Phi' is the kind of reduction variable for type
246 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
247 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
248 /// Returns true if the instruction I can be a reduction variable of type
250 bool isReductionInstr(Instruction *I, ReductionKind Kind);
251 /// Returns True, if 'Phi' is an induction variable.
252 bool isInductionVariable(PHINode *Phi);
254 /// The loop that we evaluate.
258 /// DataLayout analysis.
261 // --- vectorization state --- //
263 /// Holds the induction variable.
265 /// Holds the reduction variables.
266 ReductionList Reductions;
267 /// Allowed outside users. This holds the reduction
268 /// vars which can be accessed from outside the loop.
269 SmallPtrSet<Value*, 4> AllowedExit;
272 struct LoopVectorize : public LoopPass {
273 static char ID; // Pass identification, replacement for typeid
275 LoopVectorize() : LoopPass(ID) {
276 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
283 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
285 // Only vectorize innermost loops.
289 SE = &getAnalysis<ScalarEvolution>();
290 DL = getAnalysisIfAvailable<DataLayout>();
291 LI = &getAnalysis<LoopInfo>();
293 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
294 L->getHeader()->getParent()->getName() << "\"\n");
296 // Check if it is legal to vectorize the loop.
297 LoopVectorizationLegality LVL(L, SE, DL);
298 unsigned MaxVF = LVL.getLoopMaxVF();
300 // Check that we can vectorize using the chosen vectorization width.
301 if (MaxVF < DefaultVectorizationFactor) {
302 DEBUG(dbgs() << "LV: non-vectorizable MaxVF ("<< MaxVF << ").\n");
306 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< MaxVF << ").\n");
308 // If we decided that is is *legal* to vectorizer the loop. Do it.
309 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, DefaultVectorizationFactor);
312 DEBUG(verifyFunction(*L->getHeader()->getParent()));
316 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
317 LoopPass::getAnalysisUsage(AU);
318 AU.addRequiredID(LoopSimplifyID);
319 AU.addRequiredID(LCSSAID);
320 AU.addRequired<LoopInfo>();
321 AU.addRequired<ScalarEvolution>();
326 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
327 // Instructions that access the old induction variable
328 // actually want to get the new one.
329 if (V == OldInduction)
332 LLVMContext &C = V->getContext();
333 Type *VTy = VectorType::get(V->getType(), VF);
334 Type *I32 = IntegerType::getInt32Ty(C);
335 Constant *Zero = ConstantInt::get(I32, 0);
336 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
337 Value *UndefVal = UndefValue::get(VTy);
338 // Insert the value into a new vector.
339 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
340 // Broadcast the scalar into all locations in the vector.
341 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
343 // We are accessing the induction variable. Make sure to promote the
344 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
346 return getConsecutiveVector(Shuf);
350 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
351 assert(Val->getType()->isVectorTy() && "Must be a vector");
352 assert(Val->getType()->getScalarType()->isIntegerTy() &&
353 "Elem must be an integer");
355 Type *ITy = Val->getType()->getScalarType();
356 VectorType *Ty = cast<VectorType>(Val->getType());
357 unsigned VLen = Ty->getNumElements();
358 SmallVector<Constant*, 8> Indices;
360 // Create a vector of consecutive numbers from zero to VF.
361 for (unsigned i = 0; i < VLen; ++i)
362 Indices.push_back(ConstantInt::get(ITy, i));
364 // Add the consecutive indices to the vector value.
365 Constant *Cv = ConstantVector::get(Indices);
366 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
367 return Builder.CreateAdd(Val, Cv, "induction");
370 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
371 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
375 unsigned NumOperands = Gep->getNumOperands();
376 Value *LastIndex = Gep->getOperand(NumOperands - 1);
378 // Check that all of the gep indices are uniform except for the last.
379 for (unsigned i = 0; i < NumOperands - 1; ++i)
380 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
383 // We can emit wide load/stores only of the last index is the induction
385 const SCEV *Last = SE->getSCEV(LastIndex);
386 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
387 const SCEV *Step = AR->getStepRecurrence(*SE);
389 // The memory is consecutive because the last index is consecutive
390 // and all other indices are loop invariant.
398 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
399 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
400 // If we saved a vectorized copy of V, use it.
401 ValueMap::iterator it = WidenMap.find(V);
402 if (it != WidenMap.end())
405 // Broadcast V and save the value for future uses.
406 Value *B = getBroadcastInstrs(V);
412 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
413 SmallVector<Constant*, 8> Indices;
414 // Create a vector of consecutive numbers from zero to VF.
415 for (unsigned i = 0; i < VF; ++i)
416 Indices.push_back(ConstantInt::get(ScalarTy, Val));
418 // Add the consecutive indices to the vector value.
419 return ConstantVector::get(Indices);
422 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
423 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
424 // Holds vector parameters or scalars, in case of uniform vals.
425 SmallVector<Value*, 8> Params;
427 // Find all of the vectorized parameters.
428 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
429 Value *SrcOp = Instr->getOperand(op);
431 // If we are accessing the old induction variable, use the new one.
432 if (SrcOp == OldInduction) {
433 Params.push_back(getBroadcastInstrs(Induction));
437 // Try using previously calculated values.
438 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
440 // If the src is an instruction that appeared earlier in the basic block
441 // then it should already be vectorized.
442 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
443 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
444 // The parameter is a vector value from earlier.
445 Params.push_back(WidenMap[SrcInst]);
447 // The parameter is a scalar from outside the loop. Maybe even a constant.
448 Params.push_back(SrcOp);
452 assert(Params.size() == Instr->getNumOperands() &&
453 "Invalid number of operands");
455 // Does this instruction return a value ?
456 bool IsVoidRetTy = Instr->getType()->isVoidTy();
457 Value *VecResults = 0;
459 // If we have a return value, create an empty vector. We place the scalarized
460 // instructions in this vector.
462 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
464 // For each scalar that we create.
465 for (unsigned i = 0; i < VF; ++i) {
466 Instruction *Cloned = Instr->clone();
468 Cloned->setName(Instr->getName() + ".cloned");
469 // Replace the operands of the cloned instrucions with extracted scalars.
470 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
471 Value *Op = Params[op];
472 // Param is a vector. Need to extract the right lane.
473 if (Op->getType()->isVectorTy())
474 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
475 Cloned->setOperand(op, Op);
478 // Place the cloned scalar in the new loop.
479 Builder.Insert(Cloned);
481 // If the original scalar returns a value we need to place it in a vector
482 // so that future users will be able to use it.
484 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
485 Builder.getInt32(i));
489 WidenMap[Instr] = VecResults;
492 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
494 In this function we generate a new loop. The new loop will contain
495 the vectorized instructions while the old loop will continue to run the
498 [ ] <-- vector loop bypass.
501 | [ ] <-- vector pre header.
505 | [ ]_| <-- vector loop.
508 >[ ] <--- middle-block.
511 | [ ] <--- new preheader.
515 | [ ]_| <-- old scalar loop to handle remainder.
522 // This is the original scalar-loop preheader.
523 BasicBlock *BypassBlock = Orig->getLoopPreheader();
524 BasicBlock *ExitBlock = Orig->getExitBlock();
525 assert(ExitBlock && "Must have an exit block");
527 assert(Orig->getNumBlocks() == 1 && "Invalid loop");
528 assert(BypassBlock && "Invalid loop structure");
530 BasicBlock *VectorPH =
531 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
532 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
535 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
537 BasicBlock *ScalarPH =
538 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
540 // Find the induction variable.
541 BasicBlock *OldBasicBlock = Orig->getHeader();
542 OldInduction = Legal->getInduction();
543 assert(OldInduction && "We must have a single phi node.");
544 Type *IdxTy = OldInduction->getType();
546 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
548 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
550 // Generate the induction variable.
551 Induction = Builder.CreatePHI(IdxTy, 2, "index");
552 Constant *Zero = ConstantInt::get(IdxTy, 0);
553 Constant *Step = ConstantInt::get(IdxTy, VF);
555 // Find the loop boundaries.
556 const SCEV *ExitCount = SE->getExitCount(Orig, Orig->getHeader());
557 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
559 // Get the total trip count from the count by adding 1.
560 ExitCount = SE->getAddExpr(ExitCount,
561 SE->getConstant(ExitCount->getType(), 1));
563 // Expand the trip count and place the new instructions in the preheader.
564 // Notice that the pre-header does not change, only the loop body.
565 SCEVExpander Exp(*SE, "induction");
566 Instruction *Loc = BypassBlock->getTerminator();
568 // We may need to extend the index in case there is a type mismatch.
569 // We know that the count starts at zero and does not overflow.
570 // We are using Zext because it should be less expensive.
571 if (ExitCount->getType() != Induction->getType())
572 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
574 // Count holds the overall loop count (N).
575 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
576 // Now we need to generate the expression for N - (N % VF), which is
577 // the part that the vectorized body will execute.
578 Constant *CIVF = ConstantInt::get(IdxTy, VF);
579 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
580 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
582 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
583 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
584 CountRoundDown, ConstantInt::getNullValue(IdxTy),
586 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
587 // Remove the old terminator.
588 Loc->eraseFromParent();
590 // Add a check in the middle block to see if we have completed
591 // all of the iterations in the first vector loop.
592 // If (N - N%VF) == N, then we *don't* need to run the remainder.
593 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
594 CountRoundDown, "cmp.n",
595 MiddleBlock->getTerminator());
597 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
598 // Remove the old terminator.
599 MiddleBlock->getTerminator()->eraseFromParent();
601 // Create i+1 and fill the PHINode.
602 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
603 Induction->addIncoming(Zero, VectorPH);
604 Induction->addIncoming(NextIdx, VecBody);
605 // Create the compare.
606 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
607 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
609 // Now we have two terminators. Remove the old one from the block.
610 VecBody->getTerminator()->eraseFromParent();
612 // Fix the scalar body iteration count.
613 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
614 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
616 // Get ready to start creating new instructions into the vectorized body.
617 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
619 // Register the new loop.
620 Loop* Lp = new Loop();
621 LPM->insertLoop(Lp, Orig->getParentLoop());
623 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
625 Loop *ParentLoop = Orig->getParentLoop();
627 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
628 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
629 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
633 LoopMiddleBlock = MiddleBlock;
634 LoopExitBlock = ExitBlock;
635 LoopVectorBody = VecBody;
636 LoopScalarBody = OldBasicBlock;
637 LoopBypassBlock = BypassBlock;
641 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
642 typedef SmallVector<PHINode*, 4> PhiVector;
643 BasicBlock &BB = *Orig->getHeader();
644 Constant *Zero = ConstantInt::get(
645 IntegerType::getInt32Ty(BB.getContext()), 0);
647 // In order to support reduction variables we need to be able to vectorize
648 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
649 // steages. First, we create a new vector PHI node with no incoming edges.
650 // We use this value when we vectorize all of the instructions that use the
651 // PHI. Next, after all of the instructions in the block are complete we
652 // add the new incoming edges to the PHI. At this point all of the
653 // instructions in the basic block are vectorized, so we can use them to
654 // construct the PHI.
657 // For each instruction in the old loop.
658 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
659 Instruction *Inst = it;
661 switch (Inst->getOpcode()) {
662 case Instruction::Br:
663 // Nothing to do for PHIs and BR, since we already took care of the
664 // loop control flow instructions.
666 case Instruction::PHI:{
667 PHINode* P = cast<PHINode>(Inst);
668 // Special handling for the induction var.
669 if (OldInduction == Inst)
671 // This is phase one of vectorizing PHIs.
672 // This has to be a reduction variable.
673 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
674 Type *VecTy = VectorType::get(Inst->getType(), VF);
675 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
676 PHIsToFix.push_back(P);
679 case Instruction::Add:
680 case Instruction::FAdd:
681 case Instruction::Sub:
682 case Instruction::FSub:
683 case Instruction::Mul:
684 case Instruction::FMul:
685 case Instruction::UDiv:
686 case Instruction::SDiv:
687 case Instruction::FDiv:
688 case Instruction::URem:
689 case Instruction::SRem:
690 case Instruction::FRem:
691 case Instruction::Shl:
692 case Instruction::LShr:
693 case Instruction::AShr:
694 case Instruction::And:
695 case Instruction::Or:
696 case Instruction::Xor: {
697 // Just widen binops.
698 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
699 Value *A = getVectorValue(Inst->getOperand(0));
700 Value *B = getVectorValue(Inst->getOperand(1));
701 // Use this vector value for all users of the original instruction.
702 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
705 case Instruction::Select: {
707 // TODO: If the selector is loop invariant we can issue a select
708 // instruction with a scalar condition.
709 Value *A = getVectorValue(Inst->getOperand(0));
710 Value *B = getVectorValue(Inst->getOperand(1));
711 Value *C = getVectorValue(Inst->getOperand(2));
712 WidenMap[Inst] = Builder.CreateSelect(A, B, C);
716 case Instruction::ICmp:
717 case Instruction::FCmp: {
718 // Widen compares. Generate vector compares.
719 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
720 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
721 Value *A = getVectorValue(Inst->getOperand(0));
722 Value *B = getVectorValue(Inst->getOperand(1));
724 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
726 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
730 case Instruction::Store: {
731 // Attempt to issue a wide store.
732 StoreInst *SI = dyn_cast<StoreInst>(Inst);
733 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
734 Value *Ptr = SI->getPointerOperand();
735 unsigned Alignment = SI->getAlignment();
736 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
737 // This store does not use GEPs.
738 if (!Legal->isConsecutiveGep(Gep)) {
739 scalarizeInstruction(Inst);
743 // The last index does not have to be the induction. It can be
744 // consecutive and be a function of the index. For example A[I+1];
745 unsigned NumOperands = Gep->getNumOperands();
746 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
747 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
749 // Create the new GEP with the new induction variable.
750 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
751 Gep2->setOperand(NumOperands - 1, LastIndex);
752 Ptr = Builder.Insert(Gep2);
753 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
754 Value *Val = getVectorValue(SI->getValueOperand());
755 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
758 case Instruction::Load: {
759 // Attempt to issue a wide load.
760 LoadInst *LI = dyn_cast<LoadInst>(Inst);
761 Type *RetTy = VectorType::get(LI->getType(), VF);
762 Value *Ptr = LI->getPointerOperand();
763 unsigned Alignment = LI->getAlignment();
764 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
766 // We don't have a gep. Scalarize the load.
767 if (!Legal->isConsecutiveGep(Gep)) {
768 scalarizeInstruction(Inst);
772 // The last index does not have to be the induction. It can be
773 // consecutive and be a function of the index. For example A[I+1];
774 unsigned NumOperands = Gep->getNumOperands();
775 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
776 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
778 // Create the new GEP with the new induction variable.
779 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
780 Gep2->setOperand(NumOperands - 1, LastIndex);
781 Ptr = Builder.Insert(Gep2);
782 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
783 LI = Builder.CreateLoad(Ptr);
784 LI->setAlignment(Alignment);
785 // Use this vector value for all users of the load.
789 case Instruction::ZExt:
790 case Instruction::SExt:
791 case Instruction::FPToUI:
792 case Instruction::FPToSI:
793 case Instruction::FPExt:
794 case Instruction::PtrToInt:
795 case Instruction::IntToPtr:
796 case Instruction::SIToFP:
797 case Instruction::UIToFP:
798 case Instruction::Trunc:
799 case Instruction::FPTrunc:
800 case Instruction::BitCast: {
801 /// Vectorize bitcasts.
802 CastInst *CI = dyn_cast<CastInst>(Inst);
803 Value *A = getVectorValue(Inst->getOperand(0));
804 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
805 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
810 /// All other instructions are unsupported. Scalarize them.
811 scalarizeInstruction(Inst);
814 }// end of for_each instr.
816 // At this point every instruction in the original loop is widended to
817 // a vector form. We are almost done. Now, we need to fix the PHI nodes
818 // that we vectorized. The PHI nodes are currently empty because we did
819 // not want to introduce cycles. Notice that the remaining PHI nodes
820 // that we need to fix are reduction variables.
822 // Create the 'reduced' values for each of the induction vars.
823 // The reduced values are the vector values that we scalarize and combine
824 // after the loop is finished.
825 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
827 PHINode *RdxPhi = *it;
828 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
829 assert(RdxPhi && "Unable to recover vectorized PHI");
831 // Find the reduction variable descriptor.
832 assert(Legal->getReductionVars()->count(RdxPhi) &&
833 "Unable to find the reduction variable");
834 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
835 (*Legal->getReductionVars())[RdxPhi];
837 // We need to generate a reduction vector from the incoming scalar.
838 // To do so, we need to generate the 'identity' vector and overide
839 // one of the elements with the incoming scalar reduction. We need
840 // to do it in the vector-loop preheader.
841 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
843 // This is the vector-clone of the value that leaves the loop.
844 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
845 Type *VecTy = VectorExit->getType();
847 // Find the reduction identity variable. The value of the enum is the
848 // identity. Zero for addition. One for Multiplication.
849 unsigned IdentitySclr = RdxDesc.Kind;
850 Constant *Identity = getUniformVector(IdentitySclr,
851 VecTy->getScalarType());
853 // This vector is the Identity vector where the first element is the
854 // incoming scalar reduction.
855 Value *VectorStart = Builder.CreateInsertElement(Identity,
856 RdxDesc.StartValue, Zero);
859 // Fix the vector-loop phi.
860 // We created the induction variable so we know that the
861 // preheader is the first entry.
862 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
864 // Reductions do not have to start at zero. They can start with
865 // any loop invariant values.
866 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
867 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
868 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
869 VecRdxPhi->addIncoming(Val, LoopVectorBody);
871 // Before each round, move the insertion point right between
872 // the PHIs and the values we are going to write.
873 // This allows us to write both PHINodes and the extractelement
875 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
877 // This PHINode contains the vectorized reduction variable, or
878 // the initial value vector, if we bypass the vector loop.
879 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
880 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
881 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
883 // Extract the first scalar.
885 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
886 // Extract and sum the remaining vector elements.
887 for (unsigned i=1; i < VF; ++i) {
889 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
890 if (RdxDesc.Kind == LoopVectorizationLegality::IntegerAdd) {
891 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
893 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
897 // Now, we need to fix the users of the reduction variable
898 // inside and outside of the scalar remainder loop.
899 // We know that the loop is in LCSSA form. We need to update the
900 // PHI nodes in the exit blocks.
901 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
902 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
903 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
904 if (!LCSSAPhi) continue;
906 // All PHINodes need to have a single entry edge, or two if
907 // we already fixed them.
908 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
910 // We found our reduction value exit-PHI. Update it with the
911 // incoming bypass edge.
912 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
913 // Add an edge coming from the bypass.
914 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
917 }// end of the LCSSA phi scan.
919 // Fix the scalar loop reduction variable with the incoming reduction sum
920 // from the vector body and from the backedge value.
921 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
922 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
923 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
924 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
925 }// end of for each redux variable.
928 void SingleBlockLoopVectorizer::cleanup() {
929 // The original basic block.
930 SE->forgetLoop(Orig);
933 unsigned LoopVectorizationLegality::getLoopMaxVF() {
934 if (!TheLoop->getLoopPreheader()) {
935 assert(false && "No preheader!!");
936 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
940 // We can only vectorize single basic block loops.
941 unsigned NumBlocks = TheLoop->getNumBlocks();
942 if (NumBlocks != 1) {
943 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
947 // We need to have a loop header.
948 BasicBlock *BB = TheLoop->getHeader();
949 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
951 // Go over each instruction and look at memory deps.
952 if (!canVectorizeBlock(*BB)) {
953 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
957 // ScalarEvolution needs to be able to find the exit count.
958 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
959 if (ExitCount == SE->getCouldNotCompute()) {
960 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
964 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
966 // Okay! We can vectorize. At this point we don't have any other mem analysis
967 // which may limit our maximum vectorization factor, so just return the
968 // maximum SIMD size.
969 return DefaultVectorizationFactor;
972 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
973 // Scan the instructions in the block and look for hazards.
974 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
977 PHINode *Phi = dyn_cast<PHINode>(I);
979 // This should not happen because the loop should be normalized.
980 if (Phi->getNumIncomingValues() != 2) {
981 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
984 // We only look at integer phi nodes.
985 if (!Phi->getType()->isIntegerTy()) {
986 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
990 if (isInductionVariable(Phi)) {
992 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
995 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
999 if (AddReductionVar(Phi, IntegerAdd)) {
1000 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1003 if (AddReductionVar(Phi, IntegerMult)) {
1004 DEBUG(dbgs() << "LV: Found an Mult reduction PHI."<< *Phi <<"\n");
1008 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1010 }// end of PHI handling
1012 // We still don't handle functions.
1013 CallInst *CI = dyn_cast<CallInst>(I);
1015 DEBUG(dbgs() << "LV: Found a call site:"<<
1016 CI->getCalledFunction()->getName() << "\n");
1020 // We do not re-vectorize vectors.
1021 if (!VectorType::isValidElementType(I->getType()) &&
1022 !I->getType()->isVoidTy()) {
1023 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1027 // Reduction instructions are allowed to have exit users.
1028 // All other instructions must not have external users.
1029 if (!AllowedExit.count(I))
1030 //Check that all of the users of the loop are inside the BB.
1031 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1033 Instruction *U = cast<Instruction>(*it);
1034 // This user may be a reduction exit value.
1035 BasicBlock *Parent = U->getParent();
1036 if (Parent != &BB) {
1037 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1044 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1048 // If the memory dependencies do not prevent us from
1049 // vectorizing, then vectorize.
1050 return canVectorizeMemory(BB);
1053 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1054 typedef SmallVector<Value*, 16> ValueVector;
1055 typedef SmallPtrSet<Value*, 16> ValueSet;
1056 // Holds the Load and Store *instructions*.
1060 // Scan the BB and collect legal loads and stores.
1061 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1062 Instruction *I = it;
1064 // If this is a load, save it. If this instruction can read from memory
1065 // but is not a load, then we quit. Notice that we don't handle function
1066 // calls that read or write.
1067 if (I->mayReadFromMemory()) {
1068 LoadInst *Ld = dyn_cast<LoadInst>(I);
1069 if (!Ld) return false;
1070 if (!Ld->isSimple()) {
1071 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1074 Loads.push_back(Ld);
1078 // Save store instructions. Abort if other instructions write to memory.
1079 if (I->mayWriteToMemory()) {
1080 StoreInst *St = dyn_cast<StoreInst>(I);
1081 if (!St) return false;
1082 if (!St->isSimple()) {
1083 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1086 Stores.push_back(St);
1090 // Now we have two lists that hold the loads and the stores.
1091 // Next, we find the pointers that they use.
1093 // Check if we see any stores. If there are no stores, then we don't
1094 // care if the pointers are *restrict*.
1095 if (!Stores.size()) {
1096 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1100 // Holds the read and read-write *pointers* that we find.
1102 ValueVector ReadWrites;
1104 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1105 // multiple times on the same object. If the ptr is accessed twice, once
1106 // for read and once for write, it will only appear once (on the write
1107 // list). This is okay, since we are going to check for conflicts between
1108 // writes and between reads and writes, but not between reads and reads.
1111 ValueVector::iterator I, IE;
1112 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1113 StoreInst *ST = dyn_cast<StoreInst>(*I);
1114 assert(ST && "Bad StoreInst");
1115 Value* Ptr = ST->getPointerOperand();
1116 // If we did *not* see this pointer before, insert it to
1117 // the read-write list. At this phase it is only a 'write' list.
1118 if (Seen.insert(Ptr))
1119 ReadWrites.push_back(Ptr);
1122 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1123 LoadInst *LD = dyn_cast<LoadInst>(*I);
1124 assert(LD && "Bad LoadInst");
1125 Value* Ptr = LD->getPointerOperand();
1126 // If we did *not* see this pointer before, insert it to the
1127 // read list. If we *did* see it before, then it is already in
1128 // the read-write list. This allows us to vectorize expressions
1129 // such as A[i] += x; Because the address of A[i] is a read-write
1130 // pointer. This only works if the index of A[i] is consecutive.
1131 // If the address of i is unknown (for example A[B[i]]) then we may
1132 // read a few words, modify, and write a few words, and some of the
1133 // words may be written to the same address.
1134 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1135 Reads.push_back(Ptr);
1138 // Now that the pointers are in two lists (Reads and ReadWrites), we
1139 // can check that there are no conflicts between each of the writes and
1140 // between the writes to the reads.
1141 ValueSet WriteObjects;
1142 ValueVector TempObjects;
1144 // Check that the read-writes do not conflict with other read-write
1146 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1147 GetUnderlyingObjects(*I, TempObjects, DL);
1148 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1150 if (!isIdentifiedSafeObject(*it)) {
1151 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1154 if (!WriteObjects.insert(*it)) {
1155 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1160 TempObjects.clear();
1163 /// Check that the reads don't conflict with the read-writes.
1164 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1165 GetUnderlyingObjects(*I, TempObjects, DL);
1166 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1168 if (!isIdentifiedSafeObject(*it)) {
1169 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1172 if (WriteObjects.count(*it)) {
1173 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1178 TempObjects.clear();
1185 /// Checks if the value is a Global variable or if it is an Arguments
1186 /// marked with the NoAlias attribute.
1187 bool LoopVectorizationLegality::isIdentifiedSafeObject(Value* Val) {
1188 assert(Val && "Invalid value");
1189 if (isa<GlobalValue>(Val))
1191 if (isa<AllocaInst>(Val))
1193 if (Argument *A = dyn_cast<Argument>(Val))
1194 return A->hasNoAliasAttr();
1198 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1199 ReductionKind Kind) {
1200 if (Phi->getNumIncomingValues() != 2)
1203 // Find the possible incoming reduction variable.
1204 BasicBlock *BB = Phi->getParent();
1205 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1206 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1207 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1209 // ExitInstruction is the single value which is used outside the loop.
1210 // We only allow for a single reduction value to be used outside the loop.
1211 // This includes users of the reduction, variables (which form a cycle
1212 // which ends in the phi node).
1213 Instruction *ExitInstruction = 0;
1215 // Iter is our iterator. We start with the PHI node and scan for all of the
1216 // users of this instruction. All users must be instructions which can be
1217 // used as reduction variables (such as ADD). We may have a single
1218 // out-of-block user. They cycle must end with the original PHI.
1219 // Also, we can't have multiple block-local users.
1220 Instruction *Iter = Phi;
1222 // Any reduction instr must be of one of the allowed kinds.
1223 if (!isReductionInstr(Iter, Kind))
1226 // Did we found a user inside this block ?
1227 bool FoundInBlockUser = false;
1228 // Did we reach the initial PHI node ?
1229 bool FoundStartPHI = false;
1231 // If the instruction has no users then this is a broken
1232 // chain and can't be a reduction variable.
1233 if (Iter->use_empty())
1236 // For each of the *users* of iter.
1237 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1239 Instruction *U = cast<Instruction>(*it);
1240 // We already know that the PHI is a user.
1242 FoundStartPHI = true;
1245 // Check if we found the exit user.
1246 BasicBlock *Parent = U->getParent();
1248 // We must have a single exit instruction.
1249 if (ExitInstruction != 0)
1251 ExitInstruction = Iter;
1253 // We can't have multiple inside users.
1254 if (FoundInBlockUser)
1256 FoundInBlockUser = true;
1260 // We found a reduction var if we have reached the original
1261 // phi node and we only have a single instruction with out-of-loop
1263 if (FoundStartPHI && ExitInstruction) {
1264 // This instruction is allowed to have out-of-loop users.
1265 AllowedExit.insert(ExitInstruction);
1267 // Save the description of this reduction variable.
1268 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1269 Reductions[Phi] = RD;
1276 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1277 ReductionKind Kind) {
1278 switch (I->getOpcode()) {
1281 case Instruction::PHI:
1284 case Instruction::Add:
1285 case Instruction::Sub:
1286 return Kind == IntegerAdd;
1287 case Instruction::Mul:
1288 case Instruction::UDiv:
1289 case Instruction::SDiv:
1290 return Kind == IntegerMult;
1294 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1295 // Check that the PHI is consecutive and starts at zero.
1296 const SCEV *PhiScev = SE->getSCEV(Phi);
1297 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1299 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1302 const SCEV *Step = AR->getStepRecurrence(*SE);
1303 const SCEV *Start = AR->getStart();
1305 if (!Step->isOne() || !Start->isZero()) {
1306 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1314 char LoopVectorize::ID = 0;
1315 static const char lv_name[] = "Loop Vectorization";
1316 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1317 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1318 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1319 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1320 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1323 Pass *createLoopVectorizePass() {
1324 return new LoopVectorize();