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 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
79 /// block to a specified vectorization factor (VF).
80 /// This class performs the widening of scalars into vectors, or multiple
81 /// scalars. This class also implements the following features:
82 /// * It inserts an epilogue loop for handling loops that don't have iteration
83 /// counts that are known to be a multiple of the vectorization factor.
84 /// * It handles the code generation for reduction variables.
85 /// * Scalarization (implementation using scalars) of un-vectorizable
87 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
88 /// checks, and relies on the caller to check for the different legality
89 /// aspects. The SingleBlockLoopVectorizer relies on the
90 /// LoopVectorizationLegality class to provide information about the induction
91 /// and reduction variables that were found to a given vectorization factor.
92 class SingleBlockLoopVectorizer {
95 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
96 LPPassManager *Lpm, unsigned VecWidth):
97 OrigLoop(Orig), SE(Se), LI(Li), LPM(Lpm), VF(VecWidth),
98 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
100 // Perform the actual loop widening (vectorization).
101 void vectorize(LoopVectorizationLegality *Legal) {
102 ///Create a new empty loop. Unlink the old loop and connect the new one.
103 createEmptyLoop(Legal);
104 /// Widen each instruction in the old loop to a new one in the new loop.
105 /// Use the Legality module to find the induction and reduction variables.
106 vectorizeLoop(Legal);
107 // register the new loop.
112 /// Create an empty loop, based on the loop ranges of the old loop.
113 void createEmptyLoop(LoopVectorizationLegality *Legal);
114 /// Copy and widen the instructions from the old loop.
115 void vectorizeLoop(LoopVectorizationLegality *Legal);
116 /// Insert the new loop to the loop hierarchy and pass manager.
119 /// This instruction is un-vectorizable. Implement it as a sequence
121 void scalarizeInstruction(Instruction *Instr);
123 /// Create a broadcast instruction. This method generates a broadcast
124 /// instruction (shuffle) for loop invariant values and for the induction
125 /// value. If this is the induction variable then we extend it to N, N+1, ...
126 /// this is needed because each iteration in the loop corresponds to a SIMD
128 Value *getBroadcastInstrs(Value *V);
130 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
131 /// for each element in the vector. Starting from zero.
132 Value *getConsecutiveVector(Value* Val);
134 /// When we go over instructions in the basic block we rely on previous
135 /// values within the current basic block or on loop invariant values.
136 /// When we widen (vectorize) values we place them in the map. If the values
137 /// are not within the map, they have to be loop invariant, so we simply
138 /// broadcast them into a vector.
139 Value *getVectorValue(Value *V);
141 /// Get a uniform vector of constant integers. We use this to get
142 /// vectors of ones and zeros for the reduction code.
143 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
145 typedef DenseMap<Value*, Value*> ValueMap;
147 /// The original loop.
149 // Scev analysis to use.
153 // Loop Pass Manager;
155 // The vectorization factor to use.
158 // The builder that we use
161 // --- Vectorization state ---
163 /// Middle Block between the vector and the scalar.
164 BasicBlock *LoopMiddleBlock;
165 ///The ExitBlock of the scalar loop.
166 BasicBlock *LoopExitBlock;
167 ///The vector loop body.
168 BasicBlock *LoopVectorBody;
169 ///The scalar loop body.
170 BasicBlock *LoopScalarBody;
171 ///The first bypass block.
172 BasicBlock *LoopBypassBlock;
174 /// The new Induction variable which was added to the new block.
176 /// The induction variable of the old basic block.
177 PHINode *OldInduction;
178 // Maps scalars to widened vectors.
182 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
183 /// to what vectorization factor.
184 /// This class does not look at the profitability of vectorization, only the
185 /// legality. This class has two main kinds of checks:
186 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
187 /// will change the order of memory accesses in a way that will change the
188 /// correctness of the program.
189 /// * Scalars checks - The code in canVectorizeBlock checks for a number
190 /// of different conditions, such as the availability of a single induction
191 /// variable, that all types are supported and vectorize-able, etc.
192 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
193 /// This class is also used by SingleBlockLoopVectorizer for identifying
194 /// induction variable and the different reduction variables.
195 class LoopVectorizationLegality {
197 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
198 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
200 /// This represents the kinds of reductions that we support.
201 /// We use the enum values to hold the 'identity' value for
202 /// each operand. This value does not change the result if applied.
204 NoReduction = -1, /// Not a reduction.
205 IntegerAdd = 0, /// Sum of numbers.
206 IntegerMult = 1 /// Product of numbers.
209 /// This POD struct holds information about reduction variables.
210 struct ReductionDescriptor {
212 ReductionDescriptor():
213 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
216 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
217 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
219 // The starting value of the reduction.
220 // It does not have to be zero!
222 // The instruction who's value is used outside the loop.
223 Instruction *LoopExitInstr;
224 // The kind of the reduction.
228 /// ReductionList contains the reduction descriptors for all
229 /// of the reductions that were found in the loop.
230 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
232 /// Returns the maximum vectorization factor that we *can* use to vectorize
233 /// this loop. This does not mean that it is profitable to vectorize this
234 /// loop, only that it is legal to do so. This may be a large number. We
235 /// can vectorize to any SIMD width below this number.
236 unsigned getLoopMaxVF();
238 /// Returns the Induction variable.
239 PHINode *getInduction() {return Induction;}
241 /// Returns the reduction variables found in the loop.
242 ReductionList *getReductionVars() { return &Reductions; }
244 /// Check if the pointer returned by this GEP is consecutive
245 /// when the index is vectorized. This happens when the last
246 /// index of the GEP is consecutive, like the induction variable.
247 /// This check allows us to vectorize A[idx] into a wide load/store.
248 bool isConsecutiveGep(Value *Ptr);
251 /// Check if a single basic block loop is vectorizable.
252 /// At this point we know that this is a loop with a constant trip count
253 /// and we only need to check individual instructions.
254 bool canVectorizeBlock(BasicBlock &BB);
256 /// When we vectorize loops we may change the order in which
257 /// we read and write from memory. This method checks if it is
258 /// legal to vectorize the code, considering only memory constrains.
259 /// Returns true if BB is vectorizable
260 bool canVectorizeMemory(BasicBlock &BB);
262 /// Returns True, if 'Phi' is the kind of reduction variable for type
263 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
264 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
265 /// Returns true if the instruction I can be a reduction variable of type
267 bool isReductionInstr(Instruction *I, ReductionKind Kind);
268 /// Returns True, if 'Phi' is an induction variable.
269 bool isInductionVariable(PHINode *Phi);
271 /// The loop that we evaluate.
275 /// DataLayout analysis.
278 // --- vectorization state --- //
280 /// Holds the induction variable.
282 /// Holds the reduction variables.
283 ReductionList Reductions;
284 /// Allowed outside users. This holds the reduction
285 /// vars which can be accessed from outside the loop.
286 SmallPtrSet<Value*, 4> AllowedExit;
289 struct LoopVectorize : public LoopPass {
290 static char ID; // Pass identification, replacement for typeid
292 LoopVectorize() : LoopPass(ID) {
293 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
300 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
301 // We only vectorize innermost loops.
305 SE = &getAnalysis<ScalarEvolution>();
306 DL = getAnalysisIfAvailable<DataLayout>();
307 LI = &getAnalysis<LoopInfo>();
309 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
310 L->getHeader()->getParent()->getName() << "\"\n");
312 // Check if it is legal to vectorize the loop.
313 LoopVectorizationLegality LVL(L, SE, DL);
314 unsigned MaxVF = LVL.getLoopMaxVF();
316 // Check that we can vectorize this loop using the chosen vectorization
318 if (MaxVF < DefaultVectorizationFactor) {
319 DEBUG(dbgs() << "LV: non-vectorizable MaxVF ("<< MaxVF << ").\n");
323 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< MaxVF << ").\n");
325 // If we decided that it is *legal* to vectorizer the loop then do it.
326 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, DefaultVectorizationFactor);
329 DEBUG(verifyFunction(*L->getHeader()->getParent()));
333 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
334 LoopPass::getAnalysisUsage(AU);
335 AU.addRequiredID(LoopSimplifyID);
336 AU.addRequiredID(LCSSAID);
337 AU.addRequired<LoopInfo>();
338 AU.addRequired<ScalarEvolution>();
343 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
344 // Instructions that access the old induction variable
345 // actually want to get the new one.
346 if (V == OldInduction)
349 LLVMContext &C = V->getContext();
350 Type *VTy = VectorType::get(V->getType(), VF);
351 Type *I32 = IntegerType::getInt32Ty(C);
352 Constant *Zero = ConstantInt::get(I32, 0);
353 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
354 Value *UndefVal = UndefValue::get(VTy);
355 // Insert the value into a new vector.
356 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
357 // Broadcast the scalar into all locations in the vector.
358 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
360 // We are accessing the induction variable. Make sure to promote the
361 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
363 return getConsecutiveVector(Shuf);
367 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
368 assert(Val->getType()->isVectorTy() && "Must be a vector");
369 assert(Val->getType()->getScalarType()->isIntegerTy() &&
370 "Elem must be an integer");
372 Type *ITy = Val->getType()->getScalarType();
373 VectorType *Ty = cast<VectorType>(Val->getType());
374 unsigned VLen = Ty->getNumElements();
375 SmallVector<Constant*, 8> Indices;
377 // Create a vector of consecutive numbers from zero to VF.
378 for (unsigned i = 0; i < VLen; ++i)
379 Indices.push_back(ConstantInt::get(ITy, i));
381 // Add the consecutive indices to the vector value.
382 Constant *Cv = ConstantVector::get(Indices);
383 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
384 return Builder.CreateAdd(Val, Cv, "induction");
387 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
388 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
392 unsigned NumOperands = Gep->getNumOperands();
393 Value *LastIndex = Gep->getOperand(NumOperands - 1);
395 // Check that all of the gep indices are uniform except for the last.
396 for (unsigned i = 0; i < NumOperands - 1; ++i)
397 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
400 // We can emit wide load/stores only of the last index is the induction
402 const SCEV *Last = SE->getSCEV(LastIndex);
403 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
404 const SCEV *Step = AR->getStepRecurrence(*SE);
406 // The memory is consecutive because the last index is consecutive
407 // and all other indices are loop invariant.
415 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
416 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
417 // If we saved a vectorized copy of V, use it.
418 Value *&MapEntry = WidenMap[V];
422 // Broadcast V and save the value for future uses.
423 Value *B = getBroadcastInstrs(V);
429 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
430 SmallVector<Constant*, 8> Indices;
431 // Create a vector of consecutive numbers from zero to VF.
432 for (unsigned i = 0; i < VF; ++i)
433 Indices.push_back(ConstantInt::get(ScalarTy, Val));
435 // Add the consecutive indices to the vector value.
436 return ConstantVector::get(Indices);
439 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
440 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
441 // Holds vector parameters or scalars, in case of uniform vals.
442 SmallVector<Value*, 8> Params;
444 // Find all of the vectorized parameters.
445 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
446 Value *SrcOp = Instr->getOperand(op);
448 // If we are accessing the old induction variable, use the new one.
449 if (SrcOp == OldInduction) {
450 Params.push_back(getBroadcastInstrs(Induction));
454 // Try using previously calculated values.
455 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
457 // If the src is an instruction that appeared earlier in the basic block
458 // then it should already be vectorized.
459 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
460 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
461 // The parameter is a vector value from earlier.
462 Params.push_back(WidenMap[SrcInst]);
464 // The parameter is a scalar from outside the loop. Maybe even a constant.
465 Params.push_back(SrcOp);
469 assert(Params.size() == Instr->getNumOperands() &&
470 "Invalid number of operands");
472 // Does this instruction return a value ?
473 bool IsVoidRetTy = Instr->getType()->isVoidTy();
474 Value *VecResults = 0;
476 // If we have a return value, create an empty vector. We place the scalarized
477 // instructions in this vector.
479 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
481 // For each scalar that we create:
482 for (unsigned i = 0; i < VF; ++i) {
483 Instruction *Cloned = Instr->clone();
485 Cloned->setName(Instr->getName() + ".cloned");
486 // Replace the operands of the cloned instrucions with extracted scalars.
487 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
488 Value *Op = Params[op];
489 // Param is a vector. Need to extract the right lane.
490 if (Op->getType()->isVectorTy())
491 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
492 Cloned->setOperand(op, Op);
495 // Place the cloned scalar in the new loop.
496 Builder.Insert(Cloned);
498 // If the original scalar returns a value we need to place it in a vector
499 // so that future users will be able to use it.
501 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
502 Builder.getInt32(i));
506 WidenMap[Instr] = VecResults;
509 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
511 In this function we generate a new loop. The new loop will contain
512 the vectorized instructions while the old loop will continue to run the
515 [ ] <-- vector loop bypass.
518 | [ ] <-- vector pre header.
522 | [ ]_| <-- vector loop.
525 >[ ] <--- middle-block.
528 | [ ] <--- new preheader.
532 | [ ]_| <-- old scalar loop to handle remainder.
539 // This is the original scalar-loop preheader.
540 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
541 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
542 assert(ExitBlock && "Must have an exit block");
544 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
545 assert(BypassBlock && "Invalid loop structure");
547 BasicBlock *VectorPH =
548 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
549 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
552 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
554 BasicBlock *ScalarPH =
555 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
557 // Find the induction variable.
558 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
559 OldInduction = Legal->getInduction();
560 assert(OldInduction && "We must have a single phi node.");
561 Type *IdxTy = OldInduction->getType();
563 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
565 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
567 // Generate the induction variable.
568 Induction = Builder.CreatePHI(IdxTy, 2, "index");
569 Constant *Zero = ConstantInt::get(IdxTy, 0);
570 Constant *Step = ConstantInt::get(IdxTy, VF);
572 // Find the loop boundaries.
573 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
574 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
576 // Get the total trip count from the count by adding 1.
577 ExitCount = SE->getAddExpr(ExitCount,
578 SE->getConstant(ExitCount->getType(), 1));
580 // Expand the trip count and place the new instructions in the preheader.
581 // Notice that the pre-header does not change, only the loop body.
582 SCEVExpander Exp(*SE, "induction");
583 Instruction *Loc = BypassBlock->getTerminator();
585 // We may need to extend the index in case there is a type mismatch.
586 // We know that the count starts at zero and does not overflow.
587 // We are using Zext because it should be less expensive.
588 if (ExitCount->getType() != Induction->getType())
589 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
591 // Count holds the overall loop count (N).
592 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
593 // Now we need to generate the expression for N - (N % VF), which is
594 // the part that the vectorized body will execute.
595 Constant *CIVF = ConstantInt::get(IdxTy, VF);
596 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
597 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
599 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
600 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
601 CountRoundDown, ConstantInt::getNullValue(IdxTy),
603 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
604 // Remove the old terminator.
605 Loc->eraseFromParent();
607 // Add a check in the middle block to see if we have completed
608 // all of the iterations in the first vector loop.
609 // If (N - N%VF) == N, then we *don't* need to run the remainder.
610 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
611 CountRoundDown, "cmp.n",
612 MiddleBlock->getTerminator());
614 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
615 // Remove the old terminator.
616 MiddleBlock->getTerminator()->eraseFromParent();
618 // Create i+1 and fill the PHINode.
619 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
620 Induction->addIncoming(Zero, VectorPH);
621 Induction->addIncoming(NextIdx, VecBody);
622 // Create the compare.
623 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
624 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
626 // Now we have two terminators. Remove the old one from the block.
627 VecBody->getTerminator()->eraseFromParent();
629 // Fix the scalar body iteration count.
630 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
631 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
633 // Get ready to start creating new instructions into the vectorized body.
634 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
636 // Register the new loop.
637 Loop* Lp = new Loop();
638 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
640 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
642 Loop *ParentLoop = OrigLoop->getParentLoop();
644 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
645 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
646 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
650 LoopMiddleBlock = MiddleBlock;
651 LoopExitBlock = ExitBlock;
652 LoopVectorBody = VecBody;
653 LoopScalarBody = OldBasicBlock;
654 LoopBypassBlock = BypassBlock;
658 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
659 typedef SmallVector<PHINode*, 4> PhiVector;
660 BasicBlock &BB = *OrigLoop->getHeader();
661 Constant *Zero = ConstantInt::get(
662 IntegerType::getInt32Ty(BB.getContext()), 0);
664 // In order to support reduction variables we need to be able to vectorize
665 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
666 // steages. First, we create a new vector PHI node with no incoming edges.
667 // We use this value when we vectorize all of the instructions that use the
668 // PHI. Next, after all of the instructions in the block are complete we
669 // add the new incoming edges to the PHI. At this point all of the
670 // instructions in the basic block are vectorized, so we can use them to
671 // construct the PHI.
674 // For each instruction in the old loop.
675 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
676 Instruction *Inst = it;
678 switch (Inst->getOpcode()) {
679 case Instruction::Br:
680 // Nothing to do for PHIs and BR, since we already took care of the
681 // loop control flow instructions.
683 case Instruction::PHI:{
684 PHINode* P = cast<PHINode>(Inst);
685 // Special handling for the induction var.
686 if (OldInduction == Inst)
688 // This is phase one of vectorizing PHIs.
689 // This has to be a reduction variable.
690 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
691 Type *VecTy = VectorType::get(Inst->getType(), VF);
692 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
693 PHIsToFix.push_back(P);
696 case Instruction::Add:
697 case Instruction::FAdd:
698 case Instruction::Sub:
699 case Instruction::FSub:
700 case Instruction::Mul:
701 case Instruction::FMul:
702 case Instruction::UDiv:
703 case Instruction::SDiv:
704 case Instruction::FDiv:
705 case Instruction::URem:
706 case Instruction::SRem:
707 case Instruction::FRem:
708 case Instruction::Shl:
709 case Instruction::LShr:
710 case Instruction::AShr:
711 case Instruction::And:
712 case Instruction::Or:
713 case Instruction::Xor: {
714 // Just widen binops.
715 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
716 Value *A = getVectorValue(Inst->getOperand(0));
717 Value *B = getVectorValue(Inst->getOperand(1));
718 // Use this vector value for all users of the original instruction.
719 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
722 case Instruction::Select: {
724 // If the selector is loop invariant we can create a select
725 // instruction with a scalar condition. Otherwise, use vector-select.
726 Value *Cond = Inst->getOperand(0);
727 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
729 // The condition can be loop invariant but still defined inside the
730 // loop. This means that we can't just use the original 'cond' value.
731 // We have to take the 'vectorized' value and pick the first lane.
732 // Instcombine will make this a no-op.
733 Cond = getVectorValue(Cond);
735 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
737 Value *Op0 = getVectorValue(Inst->getOperand(1));
738 Value *Op1 = getVectorValue(Inst->getOperand(2));
739 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
743 case Instruction::ICmp:
744 case Instruction::FCmp: {
745 // Widen compares. Generate vector compares.
746 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
747 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
748 Value *A = getVectorValue(Inst->getOperand(0));
749 Value *B = getVectorValue(Inst->getOperand(1));
751 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
753 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
757 case Instruction::Store: {
758 // Attempt to issue a wide store.
759 StoreInst *SI = dyn_cast<StoreInst>(Inst);
760 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
761 Value *Ptr = SI->getPointerOperand();
762 unsigned Alignment = SI->getAlignment();
763 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
764 // This store does not use GEPs.
765 if (!Legal->isConsecutiveGep(Gep)) {
766 scalarizeInstruction(Inst);
770 // The last index does not have to be the induction. It can be
771 // consecutive and be a function of the index. For example A[I+1];
772 unsigned NumOperands = Gep->getNumOperands();
773 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
774 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
776 // Create the new GEP with the new induction variable.
777 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
778 Gep2->setOperand(NumOperands - 1, LastIndex);
779 Ptr = Builder.Insert(Gep2);
780 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
781 Value *Val = getVectorValue(SI->getValueOperand());
782 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
785 case Instruction::Load: {
786 // Attempt to issue a wide load.
787 LoadInst *LI = dyn_cast<LoadInst>(Inst);
788 Type *RetTy = VectorType::get(LI->getType(), VF);
789 Value *Ptr = LI->getPointerOperand();
790 unsigned Alignment = LI->getAlignment();
791 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
793 // We don't have a gep. Scalarize the load.
794 if (!Legal->isConsecutiveGep(Gep)) {
795 scalarizeInstruction(Inst);
799 // The last index does not have to be the induction. It can be
800 // consecutive and be a function of the index. For example A[I+1];
801 unsigned NumOperands = Gep->getNumOperands();
802 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
803 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
805 // Create the new GEP with the new induction variable.
806 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
807 Gep2->setOperand(NumOperands - 1, LastIndex);
808 Ptr = Builder.Insert(Gep2);
809 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
810 LI = Builder.CreateLoad(Ptr);
811 LI->setAlignment(Alignment);
812 // Use this vector value for all users of the load.
816 case Instruction::ZExt:
817 case Instruction::SExt:
818 case Instruction::FPToUI:
819 case Instruction::FPToSI:
820 case Instruction::FPExt:
821 case Instruction::PtrToInt:
822 case Instruction::IntToPtr:
823 case Instruction::SIToFP:
824 case Instruction::UIToFP:
825 case Instruction::Trunc:
826 case Instruction::FPTrunc:
827 case Instruction::BitCast: {
828 /// Vectorize bitcasts.
829 CastInst *CI = dyn_cast<CastInst>(Inst);
830 Value *A = getVectorValue(Inst->getOperand(0));
831 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
832 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
837 /// All other instructions are unsupported. Scalarize them.
838 scalarizeInstruction(Inst);
841 }// end of for_each instr.
843 // At this point every instruction in the original loop is widended to
844 // a vector form. We are almost done. Now, we need to fix the PHI nodes
845 // that we vectorized. The PHI nodes are currently empty because we did
846 // not want to introduce cycles. Notice that the remaining PHI nodes
847 // that we need to fix are reduction variables.
849 // Create the 'reduced' values for each of the induction vars.
850 // The reduced values are the vector values that we scalarize and combine
851 // after the loop is finished.
852 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
854 PHINode *RdxPhi = *it;
855 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
856 assert(RdxPhi && "Unable to recover vectorized PHI");
858 // Find the reduction variable descriptor.
859 assert(Legal->getReductionVars()->count(RdxPhi) &&
860 "Unable to find the reduction variable");
861 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
862 (*Legal->getReductionVars())[RdxPhi];
864 // We need to generate a reduction vector from the incoming scalar.
865 // To do so, we need to generate the 'identity' vector and overide
866 // one of the elements with the incoming scalar reduction. We need
867 // to do it in the vector-loop preheader.
868 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
870 // This is the vector-clone of the value that leaves the loop.
871 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
872 Type *VecTy = VectorExit->getType();
874 // Find the reduction identity variable. The value of the enum is the
875 // identity. Zero for addition. One for Multiplication.
876 unsigned IdentitySclr = RdxDesc.Kind;
877 Constant *Identity = getUniformVector(IdentitySclr,
878 VecTy->getScalarType());
880 // This vector is the Identity vector where the first element is the
881 // incoming scalar reduction.
882 Value *VectorStart = Builder.CreateInsertElement(Identity,
883 RdxDesc.StartValue, Zero);
886 // Fix the vector-loop phi.
887 // We created the induction variable so we know that the
888 // preheader is the first entry.
889 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
891 // Reductions do not have to start at zero. They can start with
892 // any loop invariant values.
893 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
894 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
895 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
896 VecRdxPhi->addIncoming(Val, LoopVectorBody);
898 // Before each round, move the insertion point right between
899 // the PHIs and the values we are going to write.
900 // This allows us to write both PHINodes and the extractelement
902 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
904 // This PHINode contains the vectorized reduction variable, or
905 // the initial value vector, if we bypass the vector loop.
906 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
907 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
908 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
910 // Extract the first scalar.
912 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
913 // Extract and sum the remaining vector elements.
914 for (unsigned i=1; i < VF; ++i) {
916 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
917 if (RdxDesc.Kind == LoopVectorizationLegality::IntegerAdd) {
918 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
920 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
924 // Now, we need to fix the users of the reduction variable
925 // inside and outside of the scalar remainder loop.
926 // We know that the loop is in LCSSA form. We need to update the
927 // PHI nodes in the exit blocks.
928 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
929 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
930 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
931 if (!LCSSAPhi) continue;
933 // All PHINodes need to have a single entry edge, or two if
934 // we already fixed them.
935 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
937 // We found our reduction value exit-PHI. Update it with the
938 // incoming bypass edge.
939 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
940 // Add an edge coming from the bypass.
941 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
944 }// end of the LCSSA phi scan.
946 // Fix the scalar loop reduction variable with the incoming reduction sum
947 // from the vector body and from the backedge value.
948 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
949 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
950 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
951 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
952 }// end of for each redux variable.
955 void SingleBlockLoopVectorizer::cleanup() {
956 // The original basic block.
957 SE->forgetLoop(OrigLoop);
960 unsigned LoopVectorizationLegality::getLoopMaxVF() {
961 if (!TheLoop->getLoopPreheader()) {
962 assert(false && "No preheader!!");
963 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
967 // We can only vectorize single basic block loops.
968 unsigned NumBlocks = TheLoop->getNumBlocks();
969 if (NumBlocks != 1) {
970 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
974 // We need to have a loop header.
975 BasicBlock *BB = TheLoop->getHeader();
976 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
978 // Go over each instruction and look at memory deps.
979 if (!canVectorizeBlock(*BB)) {
980 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
984 // ScalarEvolution needs to be able to find the exit count.
985 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
986 if (ExitCount == SE->getCouldNotCompute()) {
987 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
991 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
993 // Okay! We can vectorize. At this point we don't have any other mem analysis
994 // which may limit our maximum vectorization factor, so just return the
995 // maximum SIMD size.
996 return DefaultVectorizationFactor;
999 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1000 // Scan the instructions in the block and look for hazards.
1001 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1002 Instruction *I = it;
1004 PHINode *Phi = dyn_cast<PHINode>(I);
1006 // This should not happen because the loop should be normalized.
1007 if (Phi->getNumIncomingValues() != 2) {
1008 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1011 // We only look at integer phi nodes.
1012 if (!Phi->getType()->isIntegerTy()) {
1013 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
1017 if (isInductionVariable(Phi)) {
1019 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1022 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1026 if (AddReductionVar(Phi, IntegerAdd)) {
1027 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1030 if (AddReductionVar(Phi, IntegerMult)) {
1031 DEBUG(dbgs() << "LV: Found an Mult reduction PHI."<< *Phi <<"\n");
1035 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1037 }// end of PHI handling
1039 // We still don't handle functions.
1040 CallInst *CI = dyn_cast<CallInst>(I);
1042 DEBUG(dbgs() << "LV: Found a call site:"<<
1043 CI->getCalledFunction()->getName() << "\n");
1047 // We do not re-vectorize vectors.
1048 if (!VectorType::isValidElementType(I->getType()) &&
1049 !I->getType()->isVoidTy()) {
1050 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1054 // Reduction instructions are allowed to have exit users.
1055 // All other instructions must not have external users.
1056 if (!AllowedExit.count(I))
1057 //Check that all of the users of the loop are inside the BB.
1058 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1060 Instruction *U = cast<Instruction>(*it);
1061 // This user may be a reduction exit value.
1062 BasicBlock *Parent = U->getParent();
1063 if (Parent != &BB) {
1064 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1071 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1075 // If the memory dependencies do not prevent us from
1076 // vectorizing, then vectorize.
1077 return canVectorizeMemory(BB);
1080 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1081 typedef SmallVector<Value*, 16> ValueVector;
1082 typedef SmallPtrSet<Value*, 16> ValueSet;
1083 // Holds the Load and Store *instructions*.
1087 // Scan the BB and collect legal loads and stores.
1088 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1089 Instruction *I = it;
1091 // If this is a load, save it. If this instruction can read from memory
1092 // but is not a load, then we quit. Notice that we don't handle function
1093 // calls that read or write.
1094 if (I->mayReadFromMemory()) {
1095 LoadInst *Ld = dyn_cast<LoadInst>(I);
1096 if (!Ld) return false;
1097 if (!Ld->isSimple()) {
1098 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1101 Loads.push_back(Ld);
1105 // Save store instructions. Abort if other instructions write to memory.
1106 if (I->mayWriteToMemory()) {
1107 StoreInst *St = dyn_cast<StoreInst>(I);
1108 if (!St) return false;
1109 if (!St->isSimple()) {
1110 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1113 Stores.push_back(St);
1117 // Now we have two lists that hold the loads and the stores.
1118 // Next, we find the pointers that they use.
1120 // Check if we see any stores. If there are no stores, then we don't
1121 // care if the pointers are *restrict*.
1122 if (!Stores.size()) {
1123 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1127 // Holds the read and read-write *pointers* that we find.
1129 ValueVector ReadWrites;
1131 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1132 // multiple times on the same object. If the ptr is accessed twice, once
1133 // for read and once for write, it will only appear once (on the write
1134 // list). This is okay, since we are going to check for conflicts between
1135 // writes and between reads and writes, but not between reads and reads.
1138 ValueVector::iterator I, IE;
1139 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1140 StoreInst *ST = dyn_cast<StoreInst>(*I);
1141 assert(ST && "Bad StoreInst");
1142 Value* Ptr = ST->getPointerOperand();
1143 // If we did *not* see this pointer before, insert it to
1144 // the read-write list. At this phase it is only a 'write' list.
1145 if (Seen.insert(Ptr))
1146 ReadWrites.push_back(Ptr);
1149 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1150 LoadInst *LD = dyn_cast<LoadInst>(*I);
1151 assert(LD && "Bad LoadInst");
1152 Value* Ptr = LD->getPointerOperand();
1153 // If we did *not* see this pointer before, insert it to the
1154 // read list. If we *did* see it before, then it is already in
1155 // the read-write list. This allows us to vectorize expressions
1156 // such as A[i] += x; Because the address of A[i] is a read-write
1157 // pointer. This only works if the index of A[i] is consecutive.
1158 // If the address of i is unknown (for example A[B[i]]) then we may
1159 // read a few words, modify, and write a few words, and some of the
1160 // words may be written to the same address.
1161 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1162 Reads.push_back(Ptr);
1165 // Now that the pointers are in two lists (Reads and ReadWrites), we
1166 // can check that there are no conflicts between each of the writes and
1167 // between the writes to the reads.
1168 ValueSet WriteObjects;
1169 ValueVector TempObjects;
1171 // Check that the read-writes do not conflict with other read-write
1173 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1174 GetUnderlyingObjects(*I, TempObjects, DL);
1175 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1177 if (!isIdentifiedObject(*it)) {
1178 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1181 if (!WriteObjects.insert(*it)) {
1182 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1187 TempObjects.clear();
1190 /// Check that the reads don't conflict with the read-writes.
1191 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1192 GetUnderlyingObjects(*I, TempObjects, DL);
1193 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1195 if (!isIdentifiedObject(*it)) {
1196 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1199 if (WriteObjects.count(*it)) {
1200 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1205 TempObjects.clear();
1212 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1213 ReductionKind Kind) {
1214 if (Phi->getNumIncomingValues() != 2)
1217 // Find the possible incoming reduction variable.
1218 BasicBlock *BB = Phi->getParent();
1219 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1220 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1221 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1223 // ExitInstruction is the single value which is used outside the loop.
1224 // We only allow for a single reduction value to be used outside the loop.
1225 // This includes users of the reduction, variables (which form a cycle
1226 // which ends in the phi node).
1227 Instruction *ExitInstruction = 0;
1229 // Iter is our iterator. We start with the PHI node and scan for all of the
1230 // users of this instruction. All users must be instructions which can be
1231 // used as reduction variables (such as ADD). We may have a single
1232 // out-of-block user. They cycle must end with the original PHI.
1233 // Also, we can't have multiple block-local users.
1234 Instruction *Iter = Phi;
1236 // Any reduction instr must be of one of the allowed kinds.
1237 if (!isReductionInstr(Iter, Kind))
1240 // Did we found a user inside this block ?
1241 bool FoundInBlockUser = false;
1242 // Did we reach the initial PHI node ?
1243 bool FoundStartPHI = false;
1245 // If the instruction has no users then this is a broken
1246 // chain and can't be a reduction variable.
1247 if (Iter->use_empty())
1250 // For each of the *users* of iter.
1251 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1253 Instruction *U = cast<Instruction>(*it);
1254 // We already know that the PHI is a user.
1256 FoundStartPHI = true;
1259 // Check if we found the exit user.
1260 BasicBlock *Parent = U->getParent();
1262 // We must have a single exit instruction.
1263 if (ExitInstruction != 0)
1265 ExitInstruction = Iter;
1267 // We can't have multiple inside users.
1268 if (FoundInBlockUser)
1270 FoundInBlockUser = true;
1274 // We found a reduction var if we have reached the original
1275 // phi node and we only have a single instruction with out-of-loop
1277 if (FoundStartPHI && ExitInstruction) {
1278 // This instruction is allowed to have out-of-loop users.
1279 AllowedExit.insert(ExitInstruction);
1281 // Save the description of this reduction variable.
1282 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1283 Reductions[Phi] = RD;
1290 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1291 ReductionKind Kind) {
1292 switch (I->getOpcode()) {
1295 case Instruction::PHI:
1298 case Instruction::Add:
1299 case Instruction::Sub:
1300 return Kind == IntegerAdd;
1301 case Instruction::Mul:
1302 case Instruction::UDiv:
1303 case Instruction::SDiv:
1304 return Kind == IntegerMult;
1308 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1309 // Check that the PHI is consecutive and starts at zero.
1310 const SCEV *PhiScev = SE->getSCEV(Phi);
1311 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1313 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1316 const SCEV *Step = AR->getStepRecurrence(*SE);
1317 const SCEV *Start = AR->getStart();
1319 if (!Step->isOne() || !Start->isZero()) {
1320 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1328 char LoopVectorize::ID = 0;
1329 static const char lv_name[] = "Loop Vectorization";
1330 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1331 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1332 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1333 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1334 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1337 Pass *createLoopVectorizePass() {
1338 return new LoopVectorize();