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 unit that checks for the legality
22 // of the vectorization.
23 // 3. SingleBlockLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
28 //===----------------------------------------------------------------------===//
30 // The reduction-variable vectorization is based on the paper:
31 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33 // Variable uniformity checks are inspired by:
34 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
36 // Other ideas/concepts are from:
37 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
39 //===----------------------------------------------------------------------===//
40 #define LV_NAME "loop-vectorize"
41 #define DEBUG_TYPE LV_NAME
42 #include "llvm/Constants.h"
43 #include "llvm/DerivedTypes.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/LLVMContext.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Analysis/LoopPass.h"
48 #include "llvm/Value.h"
49 #include "llvm/Function.h"
50 #include "llvm/Analysis/Verifier.h"
51 #include "llvm/Module.h"
52 #include "llvm/Type.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/StringExtras.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/AliasSetTracker.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
59 #include "llvm/Analysis/ScalarEvolutionExpander.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/ValueTracking.h"
62 #include "llvm/Transforms/Scalar.h"
63 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
64 #include "llvm/TargetTransformInfo.h"
65 #include "llvm/Support/CommandLine.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/raw_ostream.h"
68 #include "llvm/DataLayout.h"
69 #include "llvm/Transforms/Utils/Local.h"
73 static cl::opt<unsigned>
74 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
75 cl::desc("Set the default vectorization width. Zero is autoselect."));
79 // Forward declarations.
80 class LoopVectorizationLegality;
81 class LoopVectorizationCostModel;
83 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
84 /// block to a specified vectorization factor (VF).
85 /// This class performs the widening of scalars into vectors, or multiple
86 /// scalars. This class also implements the following features:
87 /// * It inserts an epilogue loop for handling loops that don't have iteration
88 /// counts that are known to be a multiple of the vectorization factor.
89 /// * It handles the code generation for reduction variables.
90 /// * Scalarization (implementation using scalars) of un-vectorizable
92 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
93 /// checks, and relies on the caller to check for the different legality
94 /// aspects. The SingleBlockLoopVectorizer relies on the
95 /// LoopVectorizationLegality class to provide information about the induction
96 /// and reduction variables that were found to a given vectorization factor.
97 class SingleBlockLoopVectorizer {
100 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
101 LPPassManager *Lpm, unsigned VecWidth):
102 OrigLoop(Orig), SE(Se), LI(Li), LPM(Lpm), VF(VecWidth),
103 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
105 // Perform the actual loop widening (vectorization).
106 void vectorize(LoopVectorizationLegality *Legal) {
107 ///Create a new empty loop. Unlink the old loop and connect the new one.
108 createEmptyLoop(Legal);
109 /// Widen each instruction in the old loop to a new one in the new loop.
110 /// Use the Legality module to find the induction and reduction variables.
111 vectorizeLoop(Legal);
112 // register the new loop.
117 /// Create an empty loop, based on the loop ranges of the old loop.
118 void createEmptyLoop(LoopVectorizationLegality *Legal);
119 /// Copy and widen the instructions from the old loop.
120 void vectorizeLoop(LoopVectorizationLegality *Legal);
121 /// Insert the new loop to the loop hierarchy and pass manager.
124 /// This instruction is un-vectorizable. Implement it as a sequence
126 void scalarizeInstruction(Instruction *Instr);
128 /// Create a broadcast instruction. This method generates a broadcast
129 /// instruction (shuffle) for loop invariant values and for the induction
130 /// value. If this is the induction variable then we extend it to N, N+1, ...
131 /// this is needed because each iteration in the loop corresponds to a SIMD
133 Value *getBroadcastInstrs(Value *V);
135 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
136 /// for each element in the vector. Starting from zero.
137 Value *getConsecutiveVector(Value* Val);
139 /// When we go over instructions in the basic block we rely on previous
140 /// values within the current basic block or on loop invariant values.
141 /// When we widen (vectorize) values we place them in the map. If the values
142 /// are not within the map, they have to be loop invariant, so we simply
143 /// broadcast them into a vector.
144 Value *getVectorValue(Value *V);
146 /// Get a uniform vector of constant integers. We use this to get
147 /// vectors of ones and zeros for the reduction code.
148 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
150 typedef DenseMap<Value*, Value*> ValueMap;
152 /// The original loop.
154 // Scev analysis to use.
158 // Loop Pass Manager;
160 // The vectorization factor to use.
163 // The builder that we use
166 // --- Vectorization state ---
168 /// Middle Block between the vector and the scalar.
169 BasicBlock *LoopMiddleBlock;
170 ///The ExitBlock of the scalar loop.
171 BasicBlock *LoopExitBlock;
172 ///The vector loop body.
173 BasicBlock *LoopVectorBody;
174 ///The scalar loop body.
175 BasicBlock *LoopScalarBody;
176 ///The first bypass block.
177 BasicBlock *LoopBypassBlock;
179 /// The new Induction variable which was added to the new block.
181 /// The induction variable of the old basic block.
182 PHINode *OldInduction;
183 // Maps scalars to widened vectors.
187 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
188 /// to what vectorization factor.
189 /// This class does not look at the profitability of vectorization, only the
190 /// legality. This class has two main kinds of checks:
191 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
192 /// will change the order of memory accesses in a way that will change the
193 /// correctness of the program.
194 /// * Scalars checks - The code in canVectorizeBlock checks for a number
195 /// of different conditions, such as the availability of a single induction
196 /// variable, that all types are supported and vectorize-able, etc.
197 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
198 /// This class is also used by SingleBlockLoopVectorizer for identifying
199 /// induction variable and the different reduction variables.
200 class LoopVectorizationLegality {
202 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
203 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
205 /// This represents the kinds of reductions that we support.
206 /// We use the enum values to hold the 'identity' value for
207 /// each operand. This value does not change the result if applied.
209 NoReduction = -1, /// Not a reduction.
210 IntegerAdd = 0, /// Sum of numbers.
211 IntegerMult = 1 /// Product of numbers.
214 /// This POD struct holds information about reduction variables.
215 struct ReductionDescriptor {
217 ReductionDescriptor():
218 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
221 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
222 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
224 // The starting value of the reduction.
225 // It does not have to be zero!
227 // The instruction who's value is used outside the loop.
228 Instruction *LoopExitInstr;
229 // The kind of the reduction.
233 /// ReductionList contains the reduction descriptors for all
234 /// of the reductions that were found in the loop.
235 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
237 /// Returns true if it is legal to vectorize this loop.
238 /// This does not mean that it is profitable to vectorize this
239 /// loop, only that it is legal to do so.
242 /// Returns the Induction variable.
243 PHINode *getInduction() {return Induction;}
245 /// Returns the reduction variables found in the loop.
246 ReductionList *getReductionVars() { return &Reductions; }
248 /// Check if the pointer returned by this GEP is consecutive
249 /// when the index is vectorized. This happens when the last
250 /// index of the GEP is consecutive, like the induction variable.
251 /// This check allows us to vectorize A[idx] into a wide load/store.
252 bool isConsecutiveGep(Value *Ptr);
255 /// Check if a single basic block loop is vectorizable.
256 /// At this point we know that this is a loop with a constant trip count
257 /// and we only need to check individual instructions.
258 bool canVectorizeBlock(BasicBlock &BB);
260 /// When we vectorize loops we may change the order in which
261 /// we read and write from memory. This method checks if it is
262 /// legal to vectorize the code, considering only memory constrains.
263 /// Returns true if BB is vectorizable
264 bool canVectorizeMemory(BasicBlock &BB);
266 /// Returns True, if 'Phi' is the kind of reduction variable for type
267 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
268 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
269 /// Returns true if the instruction I can be a reduction variable of type
271 bool isReductionInstr(Instruction *I, ReductionKind Kind);
272 /// Returns True, if 'Phi' is an induction variable.
273 bool isInductionVariable(PHINode *Phi);
275 /// The loop that we evaluate.
279 /// DataLayout analysis.
282 // --- vectorization state --- //
284 /// Holds the induction variable.
286 /// Holds the reduction variables.
287 ReductionList Reductions;
288 /// Allowed outside users. This holds the reduction
289 /// vars which can be accessed from outside the loop.
290 SmallPtrSet<Value*, 4> AllowedExit;
293 /// LoopVectorizationCostModel - estimates the expected speedups due to
295 /// In many cases vectorization is not profitable. This can happen because
296 /// of a number of reasons. In this class we mainly attempt to predict
297 /// the expected speedup/slowdowns due to the supported instruction set.
298 /// We use the VectorTargetTransformInfo to query the different backends
299 /// for the cost of different operations.
300 class LoopVectorizationCostModel {
303 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
304 LoopVectorizationLegality *Leg,
305 const VectorTargetTransformInfo *Vtti):
306 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
308 /// Returns the most profitable vectorization factor for the loop that is
309 /// smaller or equal to the VF argument. This method checks every power
311 unsigned findBestVectorizationFactor(unsigned VF = 4);
314 /// Returns the expected execution cost. The unit of the cost does
315 /// not matter because we use the 'cost' units to compare different
316 /// vector widths. The cost that is returned is *not* normalized by
317 /// the factor width.
318 unsigned expectedCost(unsigned VF);
320 /// Returns the execution time cost of an instruction for a given vector
321 /// width. Vector width of one means scalar.
322 unsigned getInstructionCost(Instruction *I, unsigned VF);
324 /// The loop that we evaluate.
329 /// Vectorization legality.
330 LoopVectorizationLegality *Legal;
331 /// Vector target information.
332 const VectorTargetTransformInfo *VTTI;
335 struct LoopVectorize : public LoopPass {
336 static char ID; // Pass identification, replacement for typeid
338 LoopVectorize() : LoopPass(ID) {
339 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
345 TargetTransformInfo *TTI;
347 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
348 // We only vectorize innermost loops.
352 SE = &getAnalysis<ScalarEvolution>();
353 DL = getAnalysisIfAvailable<DataLayout>();
354 LI = &getAnalysis<LoopInfo>();
355 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
357 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
358 L->getHeader()->getParent()->getName() << "\"\n");
360 // Check if it is legal to vectorize the loop.
361 LoopVectorizationLegality LVL(L, SE, DL);
362 if (!LVL.canVectorize()) {
363 DEBUG(dbgs() << "LV: Not vectorizing.\n");
367 // Select the preffered vectorization factor.
369 if (VectorizationFactor == 0) {
370 const VectorTargetTransformInfo *VTTI = 0;
372 VTTI = TTI->getVectorTargetTransformInfo();
373 // Use the cost model.
374 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
375 VF = CM.findBestVectorizationFactor();
378 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
383 // Use the user command flag.
384 VF = VectorizationFactor;
387 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ").\n");
389 // If we decided that it is *legal* to vectorizer the loop then do it.
390 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, VF);
393 DEBUG(verifyFunction(*L->getHeader()->getParent()));
397 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
398 LoopPass::getAnalysisUsage(AU);
399 AU.addRequiredID(LoopSimplifyID);
400 AU.addRequiredID(LCSSAID);
401 AU.addRequired<LoopInfo>();
402 AU.addRequired<ScalarEvolution>();
407 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
408 // Instructions that access the old induction variable
409 // actually want to get the new one.
410 if (V == OldInduction)
413 LLVMContext &C = V->getContext();
414 Type *VTy = VectorType::get(V->getType(), VF);
415 Type *I32 = IntegerType::getInt32Ty(C);
416 Constant *Zero = ConstantInt::get(I32, 0);
417 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
418 Value *UndefVal = UndefValue::get(VTy);
419 // Insert the value into a new vector.
420 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
421 // Broadcast the scalar into all locations in the vector.
422 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
424 // We are accessing the induction variable. Make sure to promote the
425 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
427 return getConsecutiveVector(Shuf);
431 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
432 assert(Val->getType()->isVectorTy() && "Must be a vector");
433 assert(Val->getType()->getScalarType()->isIntegerTy() &&
434 "Elem must be an integer");
436 Type *ITy = Val->getType()->getScalarType();
437 VectorType *Ty = cast<VectorType>(Val->getType());
438 unsigned VLen = Ty->getNumElements();
439 SmallVector<Constant*, 8> Indices;
441 // Create a vector of consecutive numbers from zero to VF.
442 for (unsigned i = 0; i < VLen; ++i)
443 Indices.push_back(ConstantInt::get(ITy, i));
445 // Add the consecutive indices to the vector value.
446 Constant *Cv = ConstantVector::get(Indices);
447 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
448 return Builder.CreateAdd(Val, Cv, "induction");
451 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
452 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
456 unsigned NumOperands = Gep->getNumOperands();
457 Value *LastIndex = Gep->getOperand(NumOperands - 1);
459 // Check that all of the gep indices are uniform except for the last.
460 for (unsigned i = 0; i < NumOperands - 1; ++i)
461 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
464 // We can emit wide load/stores only of the last index is the induction
466 const SCEV *Last = SE->getSCEV(LastIndex);
467 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
468 const SCEV *Step = AR->getStepRecurrence(*SE);
470 // The memory is consecutive because the last index is consecutive
471 // and all other indices are loop invariant.
479 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
480 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
481 // If we saved a vectorized copy of V, use it.
482 Value *&MapEntry = WidenMap[V];
486 // Broadcast V and save the value for future uses.
487 Value *B = getBroadcastInstrs(V);
493 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
494 SmallVector<Constant*, 8> Indices;
495 // Create a vector of consecutive numbers from zero to VF.
496 for (unsigned i = 0; i < VF; ++i)
497 Indices.push_back(ConstantInt::get(ScalarTy, Val));
499 // Add the consecutive indices to the vector value.
500 return ConstantVector::get(Indices);
503 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
504 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
505 // Holds vector parameters or scalars, in case of uniform vals.
506 SmallVector<Value*, 8> Params;
508 // Find all of the vectorized parameters.
509 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
510 Value *SrcOp = Instr->getOperand(op);
512 // If we are accessing the old induction variable, use the new one.
513 if (SrcOp == OldInduction) {
514 Params.push_back(getBroadcastInstrs(Induction));
518 // Try using previously calculated values.
519 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
521 // If the src is an instruction that appeared earlier in the basic block
522 // then it should already be vectorized.
523 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
524 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
525 // The parameter is a vector value from earlier.
526 Params.push_back(WidenMap[SrcInst]);
528 // The parameter is a scalar from outside the loop. Maybe even a constant.
529 Params.push_back(SrcOp);
533 assert(Params.size() == Instr->getNumOperands() &&
534 "Invalid number of operands");
536 // Does this instruction return a value ?
537 bool IsVoidRetTy = Instr->getType()->isVoidTy();
538 Value *VecResults = 0;
540 // If we have a return value, create an empty vector. We place the scalarized
541 // instructions in this vector.
543 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
545 // For each scalar that we create:
546 for (unsigned i = 0; i < VF; ++i) {
547 Instruction *Cloned = Instr->clone();
549 Cloned->setName(Instr->getName() + ".cloned");
550 // Replace the operands of the cloned instrucions with extracted scalars.
551 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
552 Value *Op = Params[op];
553 // Param is a vector. Need to extract the right lane.
554 if (Op->getType()->isVectorTy())
555 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
556 Cloned->setOperand(op, Op);
559 // Place the cloned scalar in the new loop.
560 Builder.Insert(Cloned);
562 // If the original scalar returns a value we need to place it in a vector
563 // so that future users will be able to use it.
565 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
566 Builder.getInt32(i));
570 WidenMap[Instr] = VecResults;
573 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
575 In this function we generate a new loop. The new loop will contain
576 the vectorized instructions while the old loop will continue to run the
579 [ ] <-- vector loop bypass.
582 | [ ] <-- vector pre header.
586 | [ ]_| <-- vector loop.
589 >[ ] <--- middle-block.
592 | [ ] <--- new preheader.
596 | [ ]_| <-- old scalar loop to handle remainder.
603 // This is the original scalar-loop preheader.
604 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
605 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
606 assert(ExitBlock && "Must have an exit block");
608 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
609 assert(BypassBlock && "Invalid loop structure");
611 BasicBlock *VectorPH =
612 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
613 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
616 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
618 BasicBlock *ScalarPH =
619 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
621 // Find the induction variable.
622 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
623 OldInduction = Legal->getInduction();
624 assert(OldInduction && "We must have a single phi node.");
625 Type *IdxTy = OldInduction->getType();
627 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
629 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
631 // Generate the induction variable.
632 Induction = Builder.CreatePHI(IdxTy, 2, "index");
633 Constant *Zero = ConstantInt::get(IdxTy, 0);
634 Constant *Step = ConstantInt::get(IdxTy, VF);
636 // Find the loop boundaries.
637 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
638 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
640 // Get the total trip count from the count by adding 1.
641 ExitCount = SE->getAddExpr(ExitCount,
642 SE->getConstant(ExitCount->getType(), 1));
644 // Expand the trip count and place the new instructions in the preheader.
645 // Notice that the pre-header does not change, only the loop body.
646 SCEVExpander Exp(*SE, "induction");
647 Instruction *Loc = BypassBlock->getTerminator();
649 // We may need to extend the index in case there is a type mismatch.
650 // We know that the count starts at zero and does not overflow.
651 // We are using Zext because it should be less expensive.
652 if (ExitCount->getType() != Induction->getType())
653 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
655 // Count holds the overall loop count (N).
656 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
657 // Now we need to generate the expression for N - (N % VF), which is
658 // the part that the vectorized body will execute.
659 Constant *CIVF = ConstantInt::get(IdxTy, VF);
660 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
661 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
663 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
664 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
665 CountRoundDown, ConstantInt::getNullValue(IdxTy),
667 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
668 // Remove the old terminator.
669 Loc->eraseFromParent();
671 // Add a check in the middle block to see if we have completed
672 // all of the iterations in the first vector loop.
673 // If (N - N%VF) == N, then we *don't* need to run the remainder.
674 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
675 CountRoundDown, "cmp.n",
676 MiddleBlock->getTerminator());
678 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
679 // Remove the old terminator.
680 MiddleBlock->getTerminator()->eraseFromParent();
682 // Create i+1 and fill the PHINode.
683 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
684 Induction->addIncoming(Zero, VectorPH);
685 Induction->addIncoming(NextIdx, VecBody);
686 // Create the compare.
687 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
688 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
690 // Now we have two terminators. Remove the old one from the block.
691 VecBody->getTerminator()->eraseFromParent();
693 // Fix the scalar body iteration count.
694 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
695 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
697 // Get ready to start creating new instructions into the vectorized body.
698 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
700 // Register the new loop.
701 Loop* Lp = new Loop();
702 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
704 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
706 Loop *ParentLoop = OrigLoop->getParentLoop();
708 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
709 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
710 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
714 LoopMiddleBlock = MiddleBlock;
715 LoopExitBlock = ExitBlock;
716 LoopVectorBody = VecBody;
717 LoopScalarBody = OldBasicBlock;
718 LoopBypassBlock = BypassBlock;
722 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
723 //===------------------------------------------------===//
725 // Notice: any optimization or new instruction that go
726 // into the code below should be also be implemented in
729 //===------------------------------------------------===//
730 typedef SmallVector<PHINode*, 4> PhiVector;
731 BasicBlock &BB = *OrigLoop->getHeader();
732 Constant *Zero = ConstantInt::get(
733 IntegerType::getInt32Ty(BB.getContext()), 0);
735 // In order to support reduction variables we need to be able to vectorize
736 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
737 // steages. First, we create a new vector PHI node with no incoming edges.
738 // We use this value when we vectorize all of the instructions that use the
739 // PHI. Next, after all of the instructions in the block are complete we
740 // add the new incoming edges to the PHI. At this point all of the
741 // instructions in the basic block are vectorized, so we can use them to
742 // construct the PHI.
745 // For each instruction in the old loop.
746 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
747 Instruction *Inst = it;
749 switch (Inst->getOpcode()) {
750 case Instruction::Br:
751 // Nothing to do for PHIs and BR, since we already took care of the
752 // loop control flow instructions.
754 case Instruction::PHI:{
755 PHINode* P = cast<PHINode>(Inst);
756 // Special handling for the induction var.
757 if (OldInduction == Inst)
759 // This is phase one of vectorizing PHIs.
760 // This has to be a reduction variable.
761 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
762 Type *VecTy = VectorType::get(Inst->getType(), VF);
763 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
764 PHIsToFix.push_back(P);
767 case Instruction::Add:
768 case Instruction::FAdd:
769 case Instruction::Sub:
770 case Instruction::FSub:
771 case Instruction::Mul:
772 case Instruction::FMul:
773 case Instruction::UDiv:
774 case Instruction::SDiv:
775 case Instruction::FDiv:
776 case Instruction::URem:
777 case Instruction::SRem:
778 case Instruction::FRem:
779 case Instruction::Shl:
780 case Instruction::LShr:
781 case Instruction::AShr:
782 case Instruction::And:
783 case Instruction::Or:
784 case Instruction::Xor: {
785 // Just widen binops.
786 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
787 Value *A = getVectorValue(Inst->getOperand(0));
788 Value *B = getVectorValue(Inst->getOperand(1));
789 // Use this vector value for all users of the original instruction.
790 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
793 case Instruction::Select: {
795 // If the selector is loop invariant we can create a select
796 // instruction with a scalar condition. Otherwise, use vector-select.
797 Value *Cond = Inst->getOperand(0);
798 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
800 // The condition can be loop invariant but still defined inside the
801 // loop. This means that we can't just use the original 'cond' value.
802 // We have to take the 'vectorized' value and pick the first lane.
803 // Instcombine will make this a no-op.
804 Cond = getVectorValue(Cond);
806 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
808 Value *Op0 = getVectorValue(Inst->getOperand(1));
809 Value *Op1 = getVectorValue(Inst->getOperand(2));
810 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
814 case Instruction::ICmp:
815 case Instruction::FCmp: {
816 // Widen compares. Generate vector compares.
817 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
818 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
819 Value *A = getVectorValue(Inst->getOperand(0));
820 Value *B = getVectorValue(Inst->getOperand(1));
822 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
824 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
828 case Instruction::Store: {
829 // Attempt to issue a wide store.
830 StoreInst *SI = dyn_cast<StoreInst>(Inst);
831 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
832 Value *Ptr = SI->getPointerOperand();
833 unsigned Alignment = SI->getAlignment();
834 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
835 // This store does not use GEPs.
836 if (!Legal->isConsecutiveGep(Gep)) {
837 scalarizeInstruction(Inst);
841 // The last index does not have to be the induction. It can be
842 // consecutive and be a function of the index. For example A[I+1];
843 unsigned NumOperands = Gep->getNumOperands();
844 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
845 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
847 // Create the new GEP with the new induction variable.
848 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
849 Gep2->setOperand(NumOperands - 1, LastIndex);
850 Ptr = Builder.Insert(Gep2);
851 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
852 Value *Val = getVectorValue(SI->getValueOperand());
853 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
856 case Instruction::Load: {
857 // Attempt to issue a wide load.
858 LoadInst *LI = dyn_cast<LoadInst>(Inst);
859 Type *RetTy = VectorType::get(LI->getType(), VF);
860 Value *Ptr = LI->getPointerOperand();
861 unsigned Alignment = LI->getAlignment();
862 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
864 // We don't have a gep. Scalarize the load.
865 if (!Legal->isConsecutiveGep(Gep)) {
866 scalarizeInstruction(Inst);
870 // The last index does not have to be the induction. It can be
871 // consecutive and be a function of the index. For example A[I+1];
872 unsigned NumOperands = Gep->getNumOperands();
873 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
874 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
876 // Create the new GEP with the new induction variable.
877 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
878 Gep2->setOperand(NumOperands - 1, LastIndex);
879 Ptr = Builder.Insert(Gep2);
880 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
881 LI = Builder.CreateLoad(Ptr);
882 LI->setAlignment(Alignment);
883 // Use this vector value for all users of the load.
887 case Instruction::ZExt:
888 case Instruction::SExt:
889 case Instruction::FPToUI:
890 case Instruction::FPToSI:
891 case Instruction::FPExt:
892 case Instruction::PtrToInt:
893 case Instruction::IntToPtr:
894 case Instruction::SIToFP:
895 case Instruction::UIToFP:
896 case Instruction::Trunc:
897 case Instruction::FPTrunc:
898 case Instruction::BitCast: {
899 /// Vectorize bitcasts.
900 CastInst *CI = dyn_cast<CastInst>(Inst);
901 Value *A = getVectorValue(Inst->getOperand(0));
902 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
903 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
908 /// All other instructions are unsupported. Scalarize them.
909 scalarizeInstruction(Inst);
912 }// end of for_each instr.
914 // At this point every instruction in the original loop is widended to
915 // a vector form. We are almost done. Now, we need to fix the PHI nodes
916 // that we vectorized. The PHI nodes are currently empty because we did
917 // not want to introduce cycles. Notice that the remaining PHI nodes
918 // that we need to fix are reduction variables.
920 // Create the 'reduced' values for each of the induction vars.
921 // The reduced values are the vector values that we scalarize and combine
922 // after the loop is finished.
923 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
925 PHINode *RdxPhi = *it;
926 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
927 assert(RdxPhi && "Unable to recover vectorized PHI");
929 // Find the reduction variable descriptor.
930 assert(Legal->getReductionVars()->count(RdxPhi) &&
931 "Unable to find the reduction variable");
932 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
933 (*Legal->getReductionVars())[RdxPhi];
935 // We need to generate a reduction vector from the incoming scalar.
936 // To do so, we need to generate the 'identity' vector and overide
937 // one of the elements with the incoming scalar reduction. We need
938 // to do it in the vector-loop preheader.
939 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
941 // This is the vector-clone of the value that leaves the loop.
942 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
943 Type *VecTy = VectorExit->getType();
945 // Find the reduction identity variable. The value of the enum is the
946 // identity. Zero for addition. One for Multiplication.
947 unsigned IdentitySclr = RdxDesc.Kind;
948 Constant *Identity = getUniformVector(IdentitySclr,
949 VecTy->getScalarType());
951 // This vector is the Identity vector where the first element is the
952 // incoming scalar reduction.
953 Value *VectorStart = Builder.CreateInsertElement(Identity,
954 RdxDesc.StartValue, Zero);
957 // Fix the vector-loop phi.
958 // We created the induction variable so we know that the
959 // preheader is the first entry.
960 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
962 // Reductions do not have to start at zero. They can start with
963 // any loop invariant values.
964 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
965 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
966 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
967 VecRdxPhi->addIncoming(Val, LoopVectorBody);
969 // Before each round, move the insertion point right between
970 // the PHIs and the values we are going to write.
971 // This allows us to write both PHINodes and the extractelement
973 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
975 // This PHINode contains the vectorized reduction variable, or
976 // the initial value vector, if we bypass the vector loop.
977 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
978 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
979 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
981 // Extract the first scalar.
983 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
984 // Extract and sum the remaining vector elements.
985 for (unsigned i=1; i < VF; ++i) {
987 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
988 if (RdxDesc.Kind == LoopVectorizationLegality::IntegerAdd) {
989 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
991 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
995 // Now, we need to fix the users of the reduction variable
996 // inside and outside of the scalar remainder loop.
997 // We know that the loop is in LCSSA form. We need to update the
998 // PHI nodes in the exit blocks.
999 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1000 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1001 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1002 if (!LCSSAPhi) continue;
1004 // All PHINodes need to have a single entry edge, or two if
1005 // we already fixed them.
1006 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1008 // We found our reduction value exit-PHI. Update it with the
1009 // incoming bypass edge.
1010 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1011 // Add an edge coming from the bypass.
1012 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1015 }// end of the LCSSA phi scan.
1017 // Fix the scalar loop reduction variable with the incoming reduction sum
1018 // from the vector body and from the backedge value.
1019 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1020 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1021 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1022 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1023 }// end of for each redux variable.
1026 void SingleBlockLoopVectorizer::cleanup() {
1027 // The original basic block.
1028 SE->forgetLoop(OrigLoop);
1031 bool LoopVectorizationLegality::canVectorize() {
1032 if (!TheLoop->getLoopPreheader()) {
1033 assert(false && "No preheader!!");
1034 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1038 // We can only vectorize single basic block loops.
1039 unsigned NumBlocks = TheLoop->getNumBlocks();
1040 if (NumBlocks != 1) {
1041 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1045 // We need to have a loop header.
1046 BasicBlock *BB = TheLoop->getHeader();
1047 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1049 // Go over each instruction and look at memory deps.
1050 if (!canVectorizeBlock(*BB)) {
1051 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1055 // ScalarEvolution needs to be able to find the exit count.
1056 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1057 if (ExitCount == SE->getCouldNotCompute()) {
1058 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1062 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
1064 // Okay! We can vectorize. At this point we don't have any other mem analysis
1065 // which may limit our maximum vectorization factor, so just return true with
1070 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1071 // Scan the instructions in the block and look for hazards.
1072 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1073 Instruction *I = it;
1075 PHINode *Phi = dyn_cast<PHINode>(I);
1077 // This should not happen because the loop should be normalized.
1078 if (Phi->getNumIncomingValues() != 2) {
1079 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1082 // We only look at integer phi nodes.
1083 if (!Phi->getType()->isIntegerTy()) {
1084 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
1088 if (isInductionVariable(Phi)) {
1090 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1093 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1097 if (AddReductionVar(Phi, IntegerAdd)) {
1098 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1101 if (AddReductionVar(Phi, IntegerMult)) {
1102 DEBUG(dbgs() << "LV: Found an Mult reduction PHI."<< *Phi <<"\n");
1106 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1108 }// end of PHI handling
1110 // We still don't handle functions.
1111 CallInst *CI = dyn_cast<CallInst>(I);
1113 DEBUG(dbgs() << "LV: Found a call site:"<<
1114 CI->getCalledFunction()->getName() << "\n");
1118 // We do not re-vectorize vectors.
1119 if (!VectorType::isValidElementType(I->getType()) &&
1120 !I->getType()->isVoidTy()) {
1121 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1125 // Reduction instructions are allowed to have exit users.
1126 // All other instructions must not have external users.
1127 if (!AllowedExit.count(I))
1128 //Check that all of the users of the loop are inside the BB.
1129 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1131 Instruction *U = cast<Instruction>(*it);
1132 // This user may be a reduction exit value.
1133 BasicBlock *Parent = U->getParent();
1134 if (Parent != &BB) {
1135 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1142 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1146 // If the memory dependencies do not prevent us from
1147 // vectorizing, then vectorize.
1148 return canVectorizeMemory(BB);
1151 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1152 typedef SmallVector<Value*, 16> ValueVector;
1153 typedef SmallPtrSet<Value*, 16> ValueSet;
1154 // Holds the Load and Store *instructions*.
1158 // Scan the BB and collect legal loads and stores.
1159 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1160 Instruction *I = it;
1162 // If this is a load, save it. If this instruction can read from memory
1163 // but is not a load, then we quit. Notice that we don't handle function
1164 // calls that read or write.
1165 if (I->mayReadFromMemory()) {
1166 LoadInst *Ld = dyn_cast<LoadInst>(I);
1167 if (!Ld) return false;
1168 if (!Ld->isSimple()) {
1169 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1172 Loads.push_back(Ld);
1176 // Save store instructions. Abort if other instructions write to memory.
1177 if (I->mayWriteToMemory()) {
1178 StoreInst *St = dyn_cast<StoreInst>(I);
1179 if (!St) return false;
1180 if (!St->isSimple()) {
1181 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1184 Stores.push_back(St);
1188 // Now we have two lists that hold the loads and the stores.
1189 // Next, we find the pointers that they use.
1191 // Check if we see any stores. If there are no stores, then we don't
1192 // care if the pointers are *restrict*.
1193 if (!Stores.size()) {
1194 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1198 // Holds the read and read-write *pointers* that we find.
1200 ValueVector ReadWrites;
1202 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1203 // multiple times on the same object. If the ptr is accessed twice, once
1204 // for read and once for write, it will only appear once (on the write
1205 // list). This is okay, since we are going to check for conflicts between
1206 // writes and between reads and writes, but not between reads and reads.
1209 ValueVector::iterator I, IE;
1210 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1211 StoreInst *ST = dyn_cast<StoreInst>(*I);
1212 assert(ST && "Bad StoreInst");
1213 Value* Ptr = ST->getPointerOperand();
1214 // If we did *not* see this pointer before, insert it to
1215 // the read-write list. At this phase it is only a 'write' list.
1216 if (Seen.insert(Ptr))
1217 ReadWrites.push_back(Ptr);
1220 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1221 LoadInst *LD = dyn_cast<LoadInst>(*I);
1222 assert(LD && "Bad LoadInst");
1223 Value* Ptr = LD->getPointerOperand();
1224 // If we did *not* see this pointer before, insert it to the
1225 // read list. If we *did* see it before, then it is already in
1226 // the read-write list. This allows us to vectorize expressions
1227 // such as A[i] += x; Because the address of A[i] is a read-write
1228 // pointer. This only works if the index of A[i] is consecutive.
1229 // If the address of i is unknown (for example A[B[i]]) then we may
1230 // read a few words, modify, and write a few words, and some of the
1231 // words may be written to the same address.
1232 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1233 Reads.push_back(Ptr);
1236 // Now that the pointers are in two lists (Reads and ReadWrites), we
1237 // can check that there are no conflicts between each of the writes and
1238 // between the writes to the reads.
1239 ValueSet WriteObjects;
1240 ValueVector TempObjects;
1242 // Check that the read-writes do not conflict with other read-write
1244 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1245 GetUnderlyingObjects(*I, TempObjects, DL);
1246 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1248 if (!isIdentifiedObject(*it)) {
1249 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1252 if (!WriteObjects.insert(*it)) {
1253 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1258 TempObjects.clear();
1261 /// Check that the reads don't conflict with the read-writes.
1262 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1263 GetUnderlyingObjects(*I, TempObjects, DL);
1264 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1266 if (!isIdentifiedObject(*it)) {
1267 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1270 if (WriteObjects.count(*it)) {
1271 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1276 TempObjects.clear();
1283 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1284 ReductionKind Kind) {
1285 if (Phi->getNumIncomingValues() != 2)
1288 // Find the possible incoming reduction variable.
1289 BasicBlock *BB = Phi->getParent();
1290 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1291 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1292 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1294 // ExitInstruction is the single value which is used outside the loop.
1295 // We only allow for a single reduction value to be used outside the loop.
1296 // This includes users of the reduction, variables (which form a cycle
1297 // which ends in the phi node).
1298 Instruction *ExitInstruction = 0;
1300 // Iter is our iterator. We start with the PHI node and scan for all of the
1301 // users of this instruction. All users must be instructions which can be
1302 // used as reduction variables (such as ADD). We may have a single
1303 // out-of-block user. They cycle must end with the original PHI.
1304 // Also, we can't have multiple block-local users.
1305 Instruction *Iter = Phi;
1307 // Any reduction instr must be of one of the allowed kinds.
1308 if (!isReductionInstr(Iter, Kind))
1311 // Did we found a user inside this block ?
1312 bool FoundInBlockUser = false;
1313 // Did we reach the initial PHI node ?
1314 bool FoundStartPHI = false;
1316 // If the instruction has no users then this is a broken
1317 // chain and can't be a reduction variable.
1318 if (Iter->use_empty())
1321 // For each of the *users* of iter.
1322 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1324 Instruction *U = cast<Instruction>(*it);
1325 // We already know that the PHI is a user.
1327 FoundStartPHI = true;
1330 // Check if we found the exit user.
1331 BasicBlock *Parent = U->getParent();
1333 // We must have a single exit instruction.
1334 if (ExitInstruction != 0)
1336 ExitInstruction = Iter;
1338 // We can't have multiple inside users.
1339 if (FoundInBlockUser)
1341 FoundInBlockUser = true;
1345 // We found a reduction var if we have reached the original
1346 // phi node and we only have a single instruction with out-of-loop
1348 if (FoundStartPHI && ExitInstruction) {
1349 // This instruction is allowed to have out-of-loop users.
1350 AllowedExit.insert(ExitInstruction);
1352 // Save the description of this reduction variable.
1353 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1354 Reductions[Phi] = RD;
1361 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1362 ReductionKind Kind) {
1363 switch (I->getOpcode()) {
1366 case Instruction::PHI:
1369 case Instruction::Add:
1370 case Instruction::Sub:
1371 return Kind == IntegerAdd;
1372 case Instruction::Mul:
1373 case Instruction::UDiv:
1374 case Instruction::SDiv:
1375 return Kind == IntegerMult;
1379 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1380 // Check that the PHI is consecutive and starts at zero.
1381 const SCEV *PhiScev = SE->getSCEV(Phi);
1382 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1384 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1387 const SCEV *Step = AR->getStepRecurrence(*SE);
1388 const SCEV *Start = AR->getStart();
1390 if (!Step->isOne() || !Start->isZero()) {
1391 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1398 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1400 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1404 float Cost = expectedCost(1);
1406 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1407 for (unsigned i=2; i <= VF; i*=2) {
1408 // Notice that the vector loop needs to be executed less times, so
1409 // we need to divide the cost of the vector loops by the width of
1410 // the vector elements.
1411 float VectorCost = expectedCost(i) / (float)i;
1412 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1413 (int)VectorCost << ".\n");
1414 if (VectorCost < Cost) {
1420 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1424 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1425 // We can only estimate the cost of single basic block loops.
1426 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1428 BasicBlock *BB = TheLoop->getHeader();
1431 // For each instruction in the old loop.
1432 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1433 Instruction *Inst = it;
1434 unsigned C = getInstructionCost(Inst, VF);
1436 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1437 " For instruction: "<< *Inst << "\n");
1444 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1445 assert(VTTI && "Invalid vector target transformation info");
1446 switch (I->getOpcode()) {
1447 case Instruction::GetElementPtr:
1449 case Instruction::Br: {
1450 return VTTI->getInstrCost(I->getOpcode());
1452 case Instruction::PHI:
1454 case Instruction::Add:
1455 case Instruction::FAdd:
1456 case Instruction::Sub:
1457 case Instruction::FSub:
1458 case Instruction::Mul:
1459 case Instruction::FMul:
1460 case Instruction::UDiv:
1461 case Instruction::SDiv:
1462 case Instruction::FDiv:
1463 case Instruction::URem:
1464 case Instruction::SRem:
1465 case Instruction::FRem:
1466 case Instruction::Shl:
1467 case Instruction::LShr:
1468 case Instruction::AShr:
1469 case Instruction::And:
1470 case Instruction::Or:
1471 case Instruction::Xor: {
1472 Type *VTy = VectorType::get(I->getType(), VF);
1473 return VTTI->getInstrCost(I->getOpcode(), VTy);
1475 case Instruction::Select: {
1476 SelectInst *SI = cast<SelectInst>(I);
1477 Type *VTy = VectorType::get(I->getType(), VF);
1478 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1479 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1480 Type *CondTy = SI->getCondition()->getType();
1482 CondTy = VectorType::get(CondTy, VF);
1484 return VTTI->getInstrCost(I->getOpcode(), VTy, CondTy);
1486 case Instruction::ICmp:
1487 case Instruction::FCmp: {
1488 Type *VTy = VectorType::get(I->getOperand(0)->getType(), VF);
1489 return VTTI->getInstrCost(I->getOpcode(), VTy);
1491 case Instruction::Store: {
1492 StoreInst *SI = cast<StoreInst>(I);
1493 Type *VTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1495 // Scalarized stores.
1496 if (!Legal->isConsecutiveGep(SI->getPointerOperand())) {
1499 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1501 // The cost of extracting from the value vector and pointer vector.
1502 Cost += VF * (ExtCost * 2);
1504 // The cost of the scalar stores.
1505 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1506 VTy->getScalarType(),
1508 SI->getPointerAddressSpace());
1513 return VTTI->getMemoryOpCost(I->getOpcode(), VTy, SI->getAlignment(),
1514 SI->getPointerAddressSpace());
1516 case Instruction::Load: {
1517 LoadInst *LI = cast<LoadInst>(I);
1518 Type *VTy = VectorType::get(I->getType(), VF);
1520 // Scalarized loads.
1521 if (!Legal->isConsecutiveGep(LI->getPointerOperand())) {
1524 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, VTy);
1525 unsigned ExCost = VTTI->getInstrCost(Instruction::ExtractValue, VTy);
1527 // The cost of inserting the loaded value into the result vector, and
1528 // extracting from a vector of pointers.
1529 Cost += VF * (InCost + ExCost);
1531 // The cost of the scalar stores.
1532 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(), VTy->getScalarType(),
1534 LI->getPointerAddressSpace());
1539 return VTTI->getMemoryOpCost(I->getOpcode(), VTy, LI->getAlignment(),
1540 LI->getPointerAddressSpace());
1542 case Instruction::ZExt:
1543 case Instruction::SExt:
1544 case Instruction::FPToUI:
1545 case Instruction::FPToSI:
1546 case Instruction::FPExt:
1547 case Instruction::PtrToInt:
1548 case Instruction::IntToPtr:
1549 case Instruction::SIToFP:
1550 case Instruction::UIToFP:
1551 case Instruction::Trunc:
1552 case Instruction::FPTrunc:
1553 case Instruction::BitCast: {
1554 Type *SrcTy = VectorType::get(I->getOperand(0)->getType(), VF);
1555 Type *DstTy = VectorType::get(I->getType(), VF);
1556 return VTTI->getInstrCost(I->getOpcode(), DstTy, SrcTy);
1559 // We are scalarizing the instruction. Return the cost of the scalar
1560 // instruction, plus the cost of insert and extract into vector
1561 // elements, times the vector width.
1563 Type *Ty = I->getType();
1565 if (!Ty->isVoidTy()) {
1566 Type *VTy = VectorType::get(Ty, VF);
1567 unsigned InsCost = VTTI->getInstrCost(Instruction::InsertElement, VTy);
1568 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement, VTy);
1569 Cost += VF * (InsCost + ExtCost);
1572 /// We don't have any information on the scalar instruction, but maybe
1574 /// TODO: This may be a target-specific intrinsic.
1575 /// Need to add API for that.
1576 Cost += VF * VTTI->getInstrCost(I->getOpcode(), Ty);
1586 char LoopVectorize::ID = 0;
1587 static const char lv_name[] = "Loop Vectorization";
1588 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1589 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1590 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1591 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1592 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1595 Pass *createLoopVectorizePass() {
1596 return new LoopVectorize();