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.
183 IntegerAdd, /// Sum of numbers.
184 IntegerMult, /// Product of numbers.
185 NoReduction /// Not a reduction.
188 // Holds a pairing of reduction instruction and the reduction kind.
189 typedef std::pair<Instruction*, ReductionKind> ReductionPair;
191 /// ReductionList contains the reduction variables
192 /// as well as a single EXIT (from the block) value and the kind of
193 /// reduction variable..
194 /// Notice that the EXIT instruction can also be the PHI itself.
195 typedef DenseMap<PHINode*, ReductionPair> ReductionList;
197 /// Returns the maximum vectorization factor that we *can* use to vectorize
198 /// this loop. This does not mean that it is profitable to vectorize this
199 /// loop, only that it is legal to do so. This may be a large number. We
200 /// can vectorize to any SIMD width below this number.
201 unsigned getLoopMaxVF();
203 /// Returns the Induction variable.
204 PHINode *getInduction() {return Induction;}
206 /// Returns the reduction variables found in the loop.
207 ReductionList *getReductionVars() { return &Reductions; }
209 /// Check that the GEP operands are all uniform except for the last index
210 /// which has to be the induction variable.
211 bool isConsecutiveGep(Value *Ptr);
214 /// Check if a single basic block loop is vectorizable.
215 /// At this point we know that this is a loop with a constant trip count
216 /// and we only need to check individual instructions.
217 bool canVectorizeBlock(BasicBlock &BB);
219 /// When we vectorize loops we may change the order in which
220 /// we read and write from memory. This method checks if it is
221 /// legal to vectorize the code, considering only memory constrains.
222 /// Returns true if BB is vectorizable
223 bool canVectorizeMemory(BasicBlock &BB);
225 // Check if a pointer value is known to be disjoint.
226 // Example: Alloca, Global, NoAlias.
227 bool isIdentifiedSafeObject(Value* Val);
229 /// Returns True, if 'Phi' is the kind of reduction variable for type
230 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
231 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
232 /// Checks if a constant matches the reduction kind.
233 /// Sums starts with zero. Products start at one.
234 bool isReductionConstant(Value *V, ReductionKind Kind);
235 /// Returns true if the instruction I can be a reduction variable of type
237 bool isReductionInstr(Instruction *I, ReductionKind Kind);
238 /// Returns True, if 'Phi' is an induction variable.
239 bool isInductionVariable(PHINode *Phi);
241 /// The loop that we evaluate.
245 /// DataLayout analysis.
248 // --- vectorization state --- //
250 /// Holds the induction variable.
252 /// Holds the reduction variables.
253 ReductionList Reductions;
254 /// Allowed outside users. This holds the reduction
255 /// vars which can be accessed from outside the loop.
256 SmallPtrSet<Value*, 4> AllowedExit;
259 struct LoopVectorize : public LoopPass {
260 static char ID; // Pass identification, replacement for typeid
262 LoopVectorize() : LoopPass(ID) {
263 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
270 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
272 // Only vectorize innermost loops.
276 SE = &getAnalysis<ScalarEvolution>();
277 DL = getAnalysisIfAvailable<DataLayout>();
278 LI = &getAnalysis<LoopInfo>();
280 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
281 L->getHeader()->getParent()->getName() << "\"\n");
283 // Check if it is legal to vectorize the loop.
284 LoopVectorizationLegality LVL(L, SE, DL);
285 unsigned MaxVF = LVL.getLoopMaxVF();
287 // Check that we can vectorize using the chosen vectorization width.
288 if (MaxVF < DefaultVectorizationFactor) {
289 DEBUG(dbgs() << "LV: non-vectorizable MaxVF ("<< MaxVF << ").\n");
293 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< MaxVF << ").\n");
295 // If we decided that is is *legal* to vectorizer the loop. Do it.
296 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, DefaultVectorizationFactor);
299 DEBUG(verifyFunction(*L->getHeader()->getParent()));
303 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
304 LoopPass::getAnalysisUsage(AU);
305 AU.addRequiredID(LoopSimplifyID);
306 AU.addRequiredID(LCSSAID);
307 AU.addRequired<LoopInfo>();
308 AU.addRequired<ScalarEvolution>();
313 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
314 // Instructions that access the old induction variable
315 // actually want to get the new one.
316 if (V == OldInduction)
319 LLVMContext &C = V->getContext();
320 Type *VTy = VectorType::get(V->getType(), VF);
321 Type *I32 = IntegerType::getInt32Ty(C);
322 Constant *Zero = ConstantInt::get(I32, 0);
323 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
324 Value *UndefVal = UndefValue::get(VTy);
325 // Insert the value into a new vector.
326 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
327 // Broadcast the scalar into all locations in the vector.
328 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
330 // We are accessing the induction variable. Make sure to promote the
331 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
333 return getConsecutiveVector(Shuf);
337 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
338 assert(Val->getType()->isVectorTy() && "Must be a vector");
339 assert(Val->getType()->getScalarType()->isIntegerTy() &&
340 "Elem must be an integer");
342 Type *ITy = Val->getType()->getScalarType();
343 VectorType *Ty = cast<VectorType>(Val->getType());
344 unsigned VLen = Ty->getNumElements();
345 SmallVector<Constant*, 8> Indices;
347 // Create a vector of consecutive numbers from zero to VF.
348 for (unsigned i = 0; i < VLen; ++i)
349 Indices.push_back(ConstantInt::get(ITy, i));
351 // Add the consecutive indices to the vector value.
352 Constant *Cv = ConstantVector::get(Indices);
353 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
354 return Builder.CreateAdd(Val, Cv, "induction");
357 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
358 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
362 unsigned NumOperands = Gep->getNumOperands();
363 Value *LastIndex = Gep->getOperand(NumOperands - 1);
365 // Check that all of the gep indices are uniform except for the last.
366 for (unsigned i = 0; i < NumOperands - 1; ++i)
367 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
370 // We can emit wide load/stores only of the last index is the induction
372 const SCEV *Last = SE->getSCEV(LastIndex);
373 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
374 const SCEV *Step = AR->getStepRecurrence(*SE);
376 // The memory is consecutive because the last index is consecutive
377 // and all other indices are loop invariant.
385 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
386 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
387 // If we saved a vectorized copy of V, use it.
388 ValueMap::iterator it = WidenMap.find(V);
389 if (it != WidenMap.end())
392 // Broadcast V and save the value for future uses.
393 Value *B = getBroadcastInstrs(V);
399 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
400 SmallVector<Constant*, 8> Indices;
401 // Create a vector of consecutive numbers from zero to VF.
402 for (unsigned i = 0; i < VF; ++i)
403 Indices.push_back(ConstantInt::get(ScalarTy, Val));
405 // Add the consecutive indices to the vector value.
406 return ConstantVector::get(Indices);
409 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
410 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
411 // Holds vector parameters or scalars, in case of uniform vals.
412 SmallVector<Value*, 8> Params;
414 // Find all of the vectorized parameters.
415 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
416 Value *SrcOp = Instr->getOperand(op);
418 // If we are accessing the old induction variable, use the new one.
419 if (SrcOp == OldInduction) {
420 Params.push_back(getBroadcastInstrs(Induction));
424 // Try using previously calculated values.
425 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
427 // If the src is an instruction that appeared earlier in the basic block
428 // then it should already be vectorized.
429 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
430 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
431 // The parameter is a vector value from earlier.
432 Params.push_back(WidenMap[SrcInst]);
434 // The parameter is a scalar from outside the loop. Maybe even a constant.
435 Params.push_back(SrcOp);
439 assert(Params.size() == Instr->getNumOperands() &&
440 "Invalid number of operands");
442 // Does this instruction return a value ?
443 bool IsVoidRetTy = Instr->getType()->isVoidTy();
444 Value *VecResults = 0;
446 // If we have a return value, create an empty vector. We place the scalarized
447 // instructions in this vector.
449 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
451 // For each scalar that we create.
452 for (unsigned i = 0; i < VF; ++i) {
453 Instruction *Cloned = Instr->clone();
455 Cloned->setName(Instr->getName() + ".cloned");
456 // Replace the operands of the cloned instrucions with extracted scalars.
457 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
458 Value *Op = Params[op];
459 // Param is a vector. Need to extract the right lane.
460 if (Op->getType()->isVectorTy())
461 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
462 Cloned->setOperand(op, Op);
465 // Place the cloned scalar in the new loop.
466 Builder.Insert(Cloned);
468 // If the original scalar returns a value we need to place it in a vector
469 // so that future users will be able to use it.
471 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
472 Builder.getInt32(i));
476 WidenMap[Instr] = VecResults;
479 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
481 In this function we generate a new loop. The new loop will contain
482 the vectorized instructions while the old loop will continue to run the
485 [ ] <-- vector loop bypass.
488 | [ ] <-- vector pre header.
492 | [ ]_| <-- vector loop.
495 >[ ] <--- middle-block.
498 | [ ] <--- new preheader.
502 | [ ]_| <-- old scalar loop to handle remainder.
509 // This is the original scalar-loop preheader.
510 BasicBlock *BypassBlock = Orig->getLoopPreheader();
511 BasicBlock *ExitBlock = Orig->getExitBlock();
512 assert(ExitBlock && "Must have an exit block");
514 assert(Orig->getNumBlocks() == 1 && "Invalid loop");
515 assert(BypassBlock && "Invalid loop structure");
517 BasicBlock *VectorPH =
518 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
519 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
522 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
524 BasicBlock *ScalarPH =
525 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
527 // Find the induction variable.
528 BasicBlock *OldBasicBlock = Orig->getHeader();
529 OldInduction = Legal->getInduction();
530 assert(OldInduction && "We must have a single phi node.");
531 Type *IdxTy = OldInduction->getType();
533 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
535 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
537 // Generate the induction variable.
538 Induction = Builder.CreatePHI(IdxTy, 2, "index");
539 Constant *Zero = ConstantInt::get(IdxTy, 0);
540 Constant *Step = ConstantInt::get(IdxTy, VF);
542 // Find the loop boundaries.
543 const SCEV *ExitCount = SE->getExitCount(Orig, Orig->getHeader());
544 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
546 // Get the total trip count from the count by adding 1.
547 ExitCount = SE->getAddExpr(ExitCount,
548 SE->getConstant(ExitCount->getType(), 1));
550 // Expand the trip count and place the new instructions in the preheader.
551 // Notice that the pre-header does not change, only the loop body.
552 SCEVExpander Exp(*SE, "induction");
553 Instruction *Loc = BypassBlock->getTerminator();
555 // We may need to extend the index in case there is a type mismatch.
556 // We know that the count starts at zero and does not overflow.
557 // We are using Zext because it should be less expensive.
558 if (ExitCount->getType() != Induction->getType())
559 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
561 // Count holds the overall loop count (N).
562 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
563 // Now we need to generate the expression for N - (N % VF), which is
564 // the part that the vectorized body will execute.
565 Constant *CIVF = ConstantInt::get(IdxTy, VF);
566 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
567 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
569 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
570 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
571 CountRoundDown, ConstantInt::getNullValue(IdxTy),
573 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
574 // Remove the old terminator.
575 Loc->eraseFromParent();
577 // Add a check in the middle block to see if we have completed
578 // all of the iterations in the first vector loop.
579 // If (N - N%VF) == N, then we *don't* need to run the remainder.
580 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
581 CountRoundDown, "cmp.n",
582 MiddleBlock->getTerminator());
584 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
585 // Remove the old terminator.
586 MiddleBlock->getTerminator()->eraseFromParent();
588 // Create i+1 and fill the PHINode.
589 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
590 Induction->addIncoming(Zero, VectorPH);
591 Induction->addIncoming(NextIdx, VecBody);
592 // Create the compare.
593 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
594 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
596 // Now we have two terminators. Remove the old one from the block.
597 VecBody->getTerminator()->eraseFromParent();
599 // Fix the scalar body iteration count.
600 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
601 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
603 // Get ready to start creating new instructions into the vectorized body.
604 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
606 // Register the new loop.
607 Loop* Lp = new Loop();
608 LPM->insertLoop(Lp, Orig->getParentLoop());
610 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
612 Loop *ParentLoop = Orig->getParentLoop();
614 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
615 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
616 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
620 LoopMiddleBlock = MiddleBlock;
621 LoopExitBlock = ExitBlock;
622 LoopVectorBody = VecBody;
623 LoopScalarBody = OldBasicBlock;
624 LoopBypassBlock = BypassBlock;
628 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
629 typedef SmallVector<PHINode*, 4> PhiVector;
630 BasicBlock &BB = *Orig->getHeader();
632 // In order to support reduction variables we need to be able to vectorize
633 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
634 // steages. First, we create a new vector PHI node with no incoming edges.
635 // We use this value when we vectorize all of the instructions that use the
636 // PHI. Next, after all of the instructions in the block are complete we
637 // add the new incoming edges to the PHI. At this point all of the
638 // instructions in the basic block are vectorized, so we can use them to
639 // construct the PHI.
642 // For each instruction in the old loop.
643 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
644 Instruction *Inst = it;
646 switch (Inst->getOpcode()) {
647 case Instruction::Br:
648 // Nothing to do for PHIs and BR, since we already took care of the
649 // loop control flow instructions.
651 case Instruction::PHI:{
652 PHINode* P = cast<PHINode>(Inst);
653 // Special handling for the induction var.
654 if (OldInduction == Inst)
656 // This is phase one of vectorizing PHIs.
657 // This has to be a reduction variable.
658 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
659 Type *VecTy = VectorType::get(Inst->getType(), VF);
660 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
661 PHIsToFix.push_back(P);
664 case Instruction::Add:
665 case Instruction::FAdd:
666 case Instruction::Sub:
667 case Instruction::FSub:
668 case Instruction::Mul:
669 case Instruction::FMul:
670 case Instruction::UDiv:
671 case Instruction::SDiv:
672 case Instruction::FDiv:
673 case Instruction::URem:
674 case Instruction::SRem:
675 case Instruction::FRem:
676 case Instruction::Shl:
677 case Instruction::LShr:
678 case Instruction::AShr:
679 case Instruction::And:
680 case Instruction::Or:
681 case Instruction::Xor: {
682 // Just widen binops.
683 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
684 Value *A = getVectorValue(Inst->getOperand(0));
685 Value *B = getVectorValue(Inst->getOperand(1));
686 // Use this vector value for all users of the original instruction.
687 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
690 case Instruction::Select: {
692 // TODO: If the selector is loop invariant we can issue a select
693 // instruction with a scalar condition.
694 Value *A = getVectorValue(Inst->getOperand(0));
695 Value *B = getVectorValue(Inst->getOperand(1));
696 Value *C = getVectorValue(Inst->getOperand(2));
697 WidenMap[Inst] = Builder.CreateSelect(A, B, C);
701 case Instruction::ICmp:
702 case Instruction::FCmp: {
703 // Widen compares. Generate vector compares.
704 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
705 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
706 Value *A = getVectorValue(Inst->getOperand(0));
707 Value *B = getVectorValue(Inst->getOperand(1));
709 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
711 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
715 case Instruction::Store: {
716 // Attempt to issue a wide store.
717 StoreInst *SI = dyn_cast<StoreInst>(Inst);
718 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
719 Value *Ptr = SI->getPointerOperand();
720 unsigned Alignment = SI->getAlignment();
721 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
722 // This store does not use GEPs.
723 if (!Legal->isConsecutiveGep(Gep)) {
724 scalarizeInstruction(Inst);
728 // Create the new GEP with the new induction variable.
729 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
730 unsigned NumOperands = Gep->getNumOperands();
731 Gep2->setOperand(NumOperands - 1, Induction);
732 Ptr = Builder.Insert(Gep2);
733 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
734 Value *Val = getVectorValue(SI->getValueOperand());
735 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
738 case Instruction::Load: {
739 // Attempt to issue a wide load.
740 LoadInst *LI = dyn_cast<LoadInst>(Inst);
741 Type *RetTy = VectorType::get(LI->getType(), VF);
742 Value *Ptr = LI->getPointerOperand();
743 unsigned Alignment = LI->getAlignment();
744 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
746 // We don't have a gep. Scalarize the load.
747 if (!Legal->isConsecutiveGep(Gep)) {
748 scalarizeInstruction(Inst);
752 // Create the new GEP with the new induction variable.
753 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
754 unsigned NumOperands = Gep->getNumOperands();
755 Gep2->setOperand(NumOperands - 1, Induction);
756 Ptr = Builder.Insert(Gep2);
757 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
758 LI = Builder.CreateLoad(Ptr);
759 LI->setAlignment(Alignment);
760 // Use this vector value for all users of the load.
764 case Instruction::ZExt:
765 case Instruction::SExt:
766 case Instruction::FPToUI:
767 case Instruction::FPToSI:
768 case Instruction::FPExt:
769 case Instruction::PtrToInt:
770 case Instruction::IntToPtr:
771 case Instruction::SIToFP:
772 case Instruction::UIToFP:
773 case Instruction::Trunc:
774 case Instruction::FPTrunc:
775 case Instruction::BitCast: {
776 /// Vectorize bitcasts.
777 CastInst *CI = dyn_cast<CastInst>(Inst);
778 Value *A = getVectorValue(Inst->getOperand(0));
779 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
780 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
785 /// All other instructions are unsupported. Scalarize them.
786 scalarizeInstruction(Inst);
789 }// end of for_each instr.
791 // At this point every instruction in the original loop is widended to
792 // a vector form. We are almost done. Now, we need to fix the PHI nodes
793 // that we vectorized. The PHI nodes are currently empty because we did
794 // not want to introduce cycles. Notice that the remaining PHI nodes
795 // that we need to fix are reduction variables.
797 // Create the 'reduced' values for each of the induction vars.
798 // The reduced values are the vector values that we scalarize and combine
799 // after the loop is finished.
800 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
802 PHINode *RdxPhi = *it;
803 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
804 assert(RdxPhi && "Unable to recover vectorized PHI");
806 // Find the reduction variable.
807 assert(Legal->getReductionVars()->count(RdxPhi) &&
808 "Unable to find the reduction variable");
809 LoopVectorizationLegality::ReductionPair ReductionVar =
810 (*Legal->getReductionVars())[RdxPhi];
812 // This is the vector-clone of the value that leaves the loop.
813 Value *VectorExit = getVectorValue(ReductionVar.first);
814 Type *VecTy = VectorExit->getType();
816 // This is the kind of reduction.
817 LoopVectorizationLegality::ReductionKind RdxKind = ReductionVar.second;
818 // Find the reduction identity variable.
819 // Zero for addition. One for Multiplication.
820 unsigned IdentitySclr =
821 (RdxKind == LoopVectorizationLegality::IntegerAdd ? 0 : 1);
822 Constant *Identity = getUniformVector(IdentitySclr, VecTy->getScalarType());
824 // Fix the vector-loop phi.
825 // We created the induction variable so we know that the
826 // preheader is the first entry.
827 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
828 VecRdxPhi->addIncoming(Identity, VecPreheader);
829 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
830 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
831 VecRdxPhi->addIncoming(Val, LoopVectorBody);
833 // Before each round, move the insertion point right between
834 // the PHIs and the values we are going to write.
835 // This allows us to write both PHINodes and the extractelement
837 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
839 // This PHINode contains the vectorized reduction variable, or
840 // the identity vector, if we bypass the vector loop.
841 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
842 NewPhi->addIncoming(Identity, LoopBypassBlock);
843 NewPhi->addIncoming(getVectorValue(ReductionVar.first), LoopVectorBody);
845 // Extract the first scalar.
847 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
848 // Extract and sum the remaining vector elements.
849 for (unsigned i=1; i < VF; ++i) {
851 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
852 if (RdxKind == LoopVectorizationLegality::IntegerAdd) {
853 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
855 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
859 // Now, we need to fix the users of the reduction variable
860 // inside and outside of the scalar remainder loop.
861 // We know that the loop is in LCSSA form. We need to update the
862 // PHI nodes in the exit blocks.
863 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
864 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
865 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
866 if (!LCSSAPhi) continue;
868 // All PHINodes need to have a single entry edge, or two if we already fixed them.
869 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
871 // We found our reduction value exit-PHI. Update it with the incoming bypass edge.
872 if (LCSSAPhi->getIncomingValue(0) == ReductionVar.first) {
873 // Add an edge coming from the bypass.
874 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
877 }// end of the LCSSA phi scan.
879 // Fix the scalar loop reduction variable with the incoming reduction sum
880 // from the vector body and from the backedge value.
881 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
882 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
883 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
884 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, ReductionVar.first);
885 }// end of for each redux variable.
888 void SingleBlockLoopVectorizer::cleanup() {
889 // The original basic block.
890 SE->forgetLoop(Orig);
893 unsigned LoopVectorizationLegality::getLoopMaxVF() {
894 if (!TheLoop->getLoopPreheader()) {
895 assert(false && "No preheader!!");
896 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
900 // We can only vectorize single basic block loops.
901 unsigned NumBlocks = TheLoop->getNumBlocks();
902 if (NumBlocks != 1) {
903 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
907 // We need to have a loop header.
908 BasicBlock *BB = TheLoop->getHeader();
909 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
911 // Go over each instruction and look at memory deps.
912 if (!canVectorizeBlock(*BB)) {
913 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
917 // ScalarEvolution needs to be able to find the exit count.
918 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
919 if (ExitCount == SE->getCouldNotCompute()) {
920 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
924 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
926 // Okay! We can vectorize. At this point we don't have any other mem analysis
927 // which may limit our maximum vectorization factor, so just return the
928 // maximum SIMD size.
929 return DefaultVectorizationFactor;
932 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
933 // Scan the instructions in the block and look for hazards.
934 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
937 PHINode *Phi = dyn_cast<PHINode>(I);
939 // This should not happen because the loop should be normalized.
940 if (Phi->getNumIncomingValues() != 2) {
941 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
944 // We only look at integer phi nodes.
945 if (!Phi->getType()->isIntegerTy()) {
946 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
950 if (isInductionVariable(Phi)) {
952 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
955 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
959 if (AddReductionVar(Phi, IntegerAdd)) {
960 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
963 if (AddReductionVar(Phi, IntegerMult)) {
964 DEBUG(dbgs() << "LV: Found an Mult reduction PHI."<< *Phi <<"\n");
968 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
970 }// end of PHI handling
972 // We still don't handle functions.
973 CallInst *CI = dyn_cast<CallInst>(I);
975 DEBUG(dbgs() << "LV: Found a call site:"<<
976 CI->getCalledFunction()->getName() << "\n");
980 // We do not re-vectorize vectors.
981 if (!VectorType::isValidElementType(I->getType()) &&
982 !I->getType()->isVoidTy()) {
983 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
987 // Reduction instructions are allowed to have exit users.
988 // All other instructions must not have external users.
989 if (!AllowedExit.count(I))
990 //Check that all of the users of the loop are inside the BB.
991 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
993 Instruction *U = cast<Instruction>(*it);
994 // This user may be a reduction exit value.
995 BasicBlock *Parent = U->getParent();
997 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1004 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1008 // If the memory dependencies do not prevent us from
1009 // vectorizing, then vectorize.
1010 return canVectorizeMemory(BB);
1013 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1014 typedef SmallVector<Value*, 16> ValueVector;
1015 typedef SmallPtrSet<Value*, 16> ValueSet;
1016 // Holds the Load and Store *instructions*.
1020 // Scan the BB and collect legal loads and stores.
1021 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1022 Instruction *I = it;
1024 // If this is a load, save it. If this instruction can read from memory
1025 // but is not a load, then we quit. Notice that we don't handle function
1026 // calls that read or write.
1027 if (I->mayReadFromMemory()) {
1028 LoadInst *Ld = dyn_cast<LoadInst>(I);
1029 if (!Ld) return false;
1030 if (!Ld->isSimple()) {
1031 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1034 Loads.push_back(Ld);
1038 // Save store instructions. Abort if other instructions write to memory.
1039 if (I->mayWriteToMemory()) {
1040 StoreInst *St = dyn_cast<StoreInst>(I);
1041 if (!St) return false;
1042 if (!St->isSimple()) {
1043 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1046 Stores.push_back(St);
1050 // Now we have two lists that hold the loads and the stores.
1051 // Next, we find the pointers that they use.
1053 // Check if we see any stores. If there are no stores, then we don't
1054 // care if the pointers are *restrict*.
1055 if (!Stores.size()) {
1056 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1060 // Holds the read and read-write *pointers* that we find.
1062 ValueVector ReadWrites;
1064 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1065 // multiple times on the same object. If the ptr is accessed twice, once
1066 // for read and once for write, it will only appear once (on the write
1067 // list). This is okay, since we are going to check for conflicts between
1068 // writes and between reads and writes, but not between reads and reads.
1071 ValueVector::iterator I, IE;
1072 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1073 StoreInst *ST = dyn_cast<StoreInst>(*I);
1074 assert(ST && "Bad StoreInst");
1075 Value* Ptr = ST->getPointerOperand();
1076 // If we did *not* see this pointer before, insert it to
1077 // the read-write list. At this phase it is only a 'write' list.
1078 if (Seen.insert(Ptr))
1079 ReadWrites.push_back(Ptr);
1082 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1083 LoadInst *LD = dyn_cast<LoadInst>(*I);
1084 assert(LD && "Bad LoadInst");
1085 Value* Ptr = LD->getPointerOperand();
1086 // If we did *not* see this pointer before, insert it to the
1087 // read list. If we *did* see it before, then it is already in
1088 // the read-write list. This allows us to vectorize expressions
1089 // such as A[i] += x; Because the address of A[i] is a read-write
1090 // pointer. This only works if the index of A[i] is consecutive.
1091 // If the address of i is unknown (for example A[B[i]]) then we may
1092 // read a few words, modify, and write a few words, and some of the
1093 // words may be written to the same address.
1094 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1095 Reads.push_back(Ptr);
1098 // Now that the pointers are in two lists (Reads and ReadWrites), we
1099 // can check that there are no conflicts between each of the writes and
1100 // between the writes to the reads.
1101 ValueSet WriteObjects;
1102 ValueVector TempObjects;
1104 // Check that the read-writes do not conflict with other read-write
1106 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1107 GetUnderlyingObjects(*I, TempObjects, DL);
1108 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1110 if (!isIdentifiedSafeObject(*it)) {
1111 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1114 if (!WriteObjects.insert(*it)) {
1115 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1120 TempObjects.clear();
1123 /// Check that the reads don't conflict with the read-writes.
1124 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1125 GetUnderlyingObjects(*I, TempObjects, DL);
1126 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1128 if (!isIdentifiedSafeObject(*it)) {
1129 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1132 if (WriteObjects.count(*it)) {
1133 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1138 TempObjects.clear();
1145 /// Checks if the value is a Global variable or if it is an Arguments
1146 /// marked with the NoAlias attribute.
1147 bool LoopVectorizationLegality::isIdentifiedSafeObject(Value* Val) {
1148 assert(Val && "Invalid value");
1149 if (dyn_cast<GlobalValue>(Val))
1151 if (dyn_cast<AllocaInst>(Val))
1153 Argument *A = dyn_cast<Argument>(Val);
1156 return A->hasNoAliasAttr();
1159 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1160 ReductionKind Kind) {
1161 if (Phi->getNumIncomingValues() != 2)
1164 // Find the possible incoming reduction variable.
1165 BasicBlock *BB = Phi->getParent();
1166 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1167 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1168 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1170 // We must have a constant that starts the reduction.
1171 if (!isReductionConstant(RdxStart, Kind))
1174 // ExitInstruction is the single value which is used outside the loop.
1175 // We only allow for a single reduction value to be used outside the loop.
1176 // This includes users of the reduction, variables (which form a cycle
1177 // which ends in the phi node).
1178 Instruction *ExitInstruction = 0;
1180 // Iter is our iterator. We start with the PHI node and scan for all of the
1181 // users of this instruction. All users must be instructions which can be
1182 // used as reduction variables (such as ADD). We may have a single
1183 // out-of-block user. They cycle must end with the original PHI.
1184 // Also, we can't have multiple block-local users.
1185 Instruction *Iter = Phi;
1187 // Any reduction instr must be of one of the allowed kinds.
1188 if (!isReductionInstr(Iter, Kind))
1191 // Did we found a user inside this block ?
1192 bool FoundInBlockUser = false;
1193 // Did we reach the initial PHI node ?
1194 bool FoundStartPHI = false;
1196 // If the instruction has no users then this is a broken
1197 // chain and can't be a reduction variable.
1198 if (Iter->use_begin() == Iter->use_end())
1201 // For each of the *users* of iter.
1202 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1204 Instruction *U = cast<Instruction>(*it);
1205 // We already know that the PHI is a user.
1207 FoundStartPHI = true;
1210 // Check if we found the exit user.
1211 BasicBlock *Parent = U->getParent();
1213 // We must have a single exit instruction.
1214 if (ExitInstruction != 0)
1216 ExitInstruction = Iter;
1218 // We can't have multiple inside users.
1219 if (FoundInBlockUser)
1221 FoundInBlockUser = true;
1225 // We found a reduction var if we have reached the original
1226 // phi node and we only have a single instruction with out-of-loop
1228 if (FoundStartPHI && ExitInstruction) {
1229 // This instruction is allowed to have out-of-loop users.
1230 AllowedExit.insert(ExitInstruction);
1231 // Mark this as a reduction var.
1232 Reductions[Phi] = std::make_pair(ExitInstruction, Kind);
1239 LoopVectorizationLegality::isReductionConstant(Value *V, ReductionKind Kind) {
1240 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1243 if (Kind == IntegerMult && CI->isOne())
1245 if (Kind == IntegerAdd && CI->isZero())
1251 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1252 ReductionKind Kind) {
1253 switch (I->getOpcode()) {
1256 case Instruction::PHI:
1259 case Instruction::Add:
1260 case Instruction::Sub:
1261 return Kind == IntegerAdd;
1262 case Instruction::Mul:
1263 case Instruction::UDiv:
1264 case Instruction::SDiv:
1265 return Kind == IntegerMult;
1269 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1270 // Check that the PHI is consecutive and starts at zero.
1271 const SCEV *PhiScev = SE->getSCEV(Phi);
1272 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1274 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1277 const SCEV *Step = AR->getStepRecurrence(*SE);
1278 const SCEV *Start = AR->getStart();
1280 if (!Step->isOne() || !Start->isZero()) {
1281 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1289 char LoopVectorize::ID = 0;
1290 static const char lv_name[] = "Loop Vectorization";
1291 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1292 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1293 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1294 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1295 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1298 Pass *createLoopVectorizePass() {
1299 return new LoopVectorize();