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.
212 IntegerOr = 2, /// Bitwise or logical OR of numbers.
213 IntegerAnd = 3, /// Bitwise or logical AND of numbers.
214 IntegerXor = 4 /// Bitwise or logical XOR of numbers.
217 /// This POD struct holds information about reduction variables.
218 struct ReductionDescriptor {
220 ReductionDescriptor():
221 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
224 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
225 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
227 // The starting value of the reduction.
228 // It does not have to be zero!
230 // The instruction who's value is used outside the loop.
231 Instruction *LoopExitInstr;
232 // The kind of the reduction.
236 /// ReductionList contains the reduction descriptors for all
237 /// of the reductions that were found in the loop.
238 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
240 /// Returns true if it is legal to vectorize this loop.
241 /// This does not mean that it is profitable to vectorize this
242 /// loop, only that it is legal to do so.
245 /// Returns the Induction variable.
246 PHINode *getInduction() {return Induction;}
248 /// Returns the reduction variables found in the loop.
249 ReductionList *getReductionVars() { return &Reductions; }
251 /// Check if the pointer returned by this GEP is consecutive
252 /// when the index is vectorized. This happens when the last
253 /// index of the GEP is consecutive, like the induction variable.
254 /// This check allows us to vectorize A[idx] into a wide load/store.
255 bool isConsecutiveGep(Value *Ptr);
258 /// Check if a single basic block loop is vectorizable.
259 /// At this point we know that this is a loop with a constant trip count
260 /// and we only need to check individual instructions.
261 bool canVectorizeBlock(BasicBlock &BB);
263 /// When we vectorize loops we may change the order in which
264 /// we read and write from memory. This method checks if it is
265 /// legal to vectorize the code, considering only memory constrains.
266 /// Returns true if BB is vectorizable
267 bool canVectorizeMemory(BasicBlock &BB);
269 /// Returns True, if 'Phi' is the kind of reduction variable for type
270 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
271 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
272 /// Returns true if the instruction I can be a reduction variable of type
274 bool isReductionInstr(Instruction *I, ReductionKind Kind);
275 /// Returns True, if 'Phi' is an induction variable.
276 bool isInductionVariable(PHINode *Phi);
278 /// The loop that we evaluate.
282 /// DataLayout analysis.
285 // --- vectorization state --- //
287 /// Holds the induction variable.
289 /// Holds the reduction variables.
290 ReductionList Reductions;
291 /// Allowed outside users. This holds the reduction
292 /// vars which can be accessed from outside the loop.
293 SmallPtrSet<Value*, 4> AllowedExit;
296 /// LoopVectorizationCostModel - estimates the expected speedups due to
298 /// In many cases vectorization is not profitable. This can happen because
299 /// of a number of reasons. In this class we mainly attempt to predict
300 /// the expected speedup/slowdowns due to the supported instruction set.
301 /// We use the VectorTargetTransformInfo to query the different backends
302 /// for the cost of different operations.
303 class LoopVectorizationCostModel {
306 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
307 LoopVectorizationLegality *Leg,
308 const VectorTargetTransformInfo *Vtti):
309 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
311 /// Returns the most profitable vectorization factor for the loop that is
312 /// smaller or equal to the VF argument. This method checks every power
314 unsigned findBestVectorizationFactor(unsigned VF = 4);
317 /// Returns the expected execution cost. The unit of the cost does
318 /// not matter because we use the 'cost' units to compare different
319 /// vector widths. The cost that is returned is *not* normalized by
320 /// the factor width.
321 unsigned expectedCost(unsigned VF);
323 /// Returns the execution time cost of an instruction for a given vector
324 /// width. Vector width of one means scalar.
325 unsigned getInstructionCost(Instruction *I, unsigned VF);
327 /// A helper function for converting Scalar types to vector types.
328 /// If the incoming type is void, we return void. If the VF is 1, we return
330 static Type* ToVectorTy(Type *Scalar, unsigned VF);
332 /// The loop that we evaluate.
337 /// Vectorization legality.
338 LoopVectorizationLegality *Legal;
339 /// Vector target information.
340 const VectorTargetTransformInfo *VTTI;
343 struct LoopVectorize : public LoopPass {
344 static char ID; // Pass identification, replacement for typeid
346 LoopVectorize() : LoopPass(ID) {
347 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
353 TargetTransformInfo *TTI;
355 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
356 // We only vectorize innermost loops.
360 SE = &getAnalysis<ScalarEvolution>();
361 DL = getAnalysisIfAvailable<DataLayout>();
362 LI = &getAnalysis<LoopInfo>();
363 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
365 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
366 L->getHeader()->getParent()->getName() << "\"\n");
368 // Check if it is legal to vectorize the loop.
369 LoopVectorizationLegality LVL(L, SE, DL);
370 if (!LVL.canVectorize()) {
371 DEBUG(dbgs() << "LV: Not vectorizing.\n");
375 // Select the preffered vectorization factor.
377 if (VectorizationFactor == 0) {
378 const VectorTargetTransformInfo *VTTI = 0;
380 VTTI = TTI->getVectorTargetTransformInfo();
381 // Use the cost model.
382 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
383 VF = CM.findBestVectorizationFactor();
386 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
391 // Use the user command flag.
392 VF = VectorizationFactor;
395 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ").\n");
397 // If we decided that it is *legal* to vectorizer the loop then do it.
398 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, VF);
401 DEBUG(verifyFunction(*L->getHeader()->getParent()));
405 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
406 LoopPass::getAnalysisUsage(AU);
407 AU.addRequiredID(LoopSimplifyID);
408 AU.addRequiredID(LCSSAID);
409 AU.addRequired<LoopInfo>();
410 AU.addRequired<ScalarEvolution>();
415 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
416 // Instructions that access the old induction variable
417 // actually want to get the new one.
418 if (V == OldInduction)
421 LLVMContext &C = V->getContext();
422 Type *VTy = VectorType::get(V->getType(), VF);
423 Type *I32 = IntegerType::getInt32Ty(C);
424 Constant *Zero = ConstantInt::get(I32, 0);
425 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
426 Value *UndefVal = UndefValue::get(VTy);
427 // Insert the value into a new vector.
428 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
429 // Broadcast the scalar into all locations in the vector.
430 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
432 // We are accessing the induction variable. Make sure to promote the
433 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
435 return getConsecutiveVector(Shuf);
439 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
440 assert(Val->getType()->isVectorTy() && "Must be a vector");
441 assert(Val->getType()->getScalarType()->isIntegerTy() &&
442 "Elem must be an integer");
444 Type *ITy = Val->getType()->getScalarType();
445 VectorType *Ty = cast<VectorType>(Val->getType());
446 unsigned VLen = Ty->getNumElements();
447 SmallVector<Constant*, 8> Indices;
449 // Create a vector of consecutive numbers from zero to VF.
450 for (unsigned i = 0; i < VLen; ++i)
451 Indices.push_back(ConstantInt::get(ITy, i));
453 // Add the consecutive indices to the vector value.
454 Constant *Cv = ConstantVector::get(Indices);
455 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
456 return Builder.CreateAdd(Val, Cv, "induction");
459 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
460 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
464 unsigned NumOperands = Gep->getNumOperands();
465 Value *LastIndex = Gep->getOperand(NumOperands - 1);
467 // Check that all of the gep indices are uniform except for the last.
468 for (unsigned i = 0; i < NumOperands - 1; ++i)
469 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
472 // We can emit wide load/stores only of the last index is the induction
474 const SCEV *Last = SE->getSCEV(LastIndex);
475 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
476 const SCEV *Step = AR->getStepRecurrence(*SE);
478 // The memory is consecutive because the last index is consecutive
479 // and all other indices are loop invariant.
487 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
488 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
489 // If we saved a vectorized copy of V, use it.
490 Value *&MapEntry = WidenMap[V];
494 // Broadcast V and save the value for future uses.
495 Value *B = getBroadcastInstrs(V);
501 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
502 SmallVector<Constant*, 8> Indices;
503 // Create a vector of consecutive numbers from zero to VF.
504 for (unsigned i = 0; i < VF; ++i)
505 Indices.push_back(ConstantInt::get(ScalarTy, Val));
507 // Add the consecutive indices to the vector value.
508 return ConstantVector::get(Indices);
511 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
512 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
513 // Holds vector parameters or scalars, in case of uniform vals.
514 SmallVector<Value*, 8> Params;
516 // Find all of the vectorized parameters.
517 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
518 Value *SrcOp = Instr->getOperand(op);
520 // If we are accessing the old induction variable, use the new one.
521 if (SrcOp == OldInduction) {
522 Params.push_back(getBroadcastInstrs(Induction));
526 // Try using previously calculated values.
527 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
529 // If the src is an instruction that appeared earlier in the basic block
530 // then it should already be vectorized.
531 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
532 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
533 // The parameter is a vector value from earlier.
534 Params.push_back(WidenMap[SrcInst]);
536 // The parameter is a scalar from outside the loop. Maybe even a constant.
537 Params.push_back(SrcOp);
541 assert(Params.size() == Instr->getNumOperands() &&
542 "Invalid number of operands");
544 // Does this instruction return a value ?
545 bool IsVoidRetTy = Instr->getType()->isVoidTy();
546 Value *VecResults = 0;
548 // If we have a return value, create an empty vector. We place the scalarized
549 // instructions in this vector.
551 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
553 // For each scalar that we create:
554 for (unsigned i = 0; i < VF; ++i) {
555 Instruction *Cloned = Instr->clone();
557 Cloned->setName(Instr->getName() + ".cloned");
558 // Replace the operands of the cloned instrucions with extracted scalars.
559 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
560 Value *Op = Params[op];
561 // Param is a vector. Need to extract the right lane.
562 if (Op->getType()->isVectorTy())
563 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
564 Cloned->setOperand(op, Op);
567 // Place the cloned scalar in the new loop.
568 Builder.Insert(Cloned);
570 // If the original scalar returns a value we need to place it in a vector
571 // so that future users will be able to use it.
573 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
574 Builder.getInt32(i));
578 WidenMap[Instr] = VecResults;
581 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
583 In this function we generate a new loop. The new loop will contain
584 the vectorized instructions while the old loop will continue to run the
587 [ ] <-- vector loop bypass.
590 | [ ] <-- vector pre header.
594 | [ ]_| <-- vector loop.
597 >[ ] <--- middle-block.
600 | [ ] <--- new preheader.
604 | [ ]_| <-- old scalar loop to handle remainder.
611 // This is the original scalar-loop preheader.
612 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
613 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
614 assert(ExitBlock && "Must have an exit block");
616 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
617 assert(BypassBlock && "Invalid loop structure");
619 BasicBlock *VectorPH =
620 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
621 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
624 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
626 BasicBlock *ScalarPH =
627 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
629 // Find the induction variable.
630 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
631 OldInduction = Legal->getInduction();
632 assert(OldInduction && "We must have a single phi node.");
633 Type *IdxTy = OldInduction->getType();
635 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
637 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
639 // Generate the induction variable.
640 Induction = Builder.CreatePHI(IdxTy, 2, "index");
641 Constant *Zero = ConstantInt::get(IdxTy, 0);
642 Constant *Step = ConstantInt::get(IdxTy, VF);
644 // Find the loop boundaries.
645 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
646 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
648 // Get the total trip count from the count by adding 1.
649 ExitCount = SE->getAddExpr(ExitCount,
650 SE->getConstant(ExitCount->getType(), 1));
652 // Expand the trip count and place the new instructions in the preheader.
653 // Notice that the pre-header does not change, only the loop body.
654 SCEVExpander Exp(*SE, "induction");
655 Instruction *Loc = BypassBlock->getTerminator();
657 // We may need to extend the index in case there is a type mismatch.
658 // We know that the count starts at zero and does not overflow.
659 // We are using Zext because it should be less expensive.
660 if (ExitCount->getType() != Induction->getType())
661 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
663 // Count holds the overall loop count (N).
664 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
665 // Now we need to generate the expression for N - (N % VF), which is
666 // the part that the vectorized body will execute.
667 Constant *CIVF = ConstantInt::get(IdxTy, VF);
668 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
669 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
671 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
672 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
673 CountRoundDown, ConstantInt::getNullValue(IdxTy),
675 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
676 // Remove the old terminator.
677 Loc->eraseFromParent();
679 // Add a check in the middle block to see if we have completed
680 // all of the iterations in the first vector loop.
681 // If (N - N%VF) == N, then we *don't* need to run the remainder.
682 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
683 CountRoundDown, "cmp.n",
684 MiddleBlock->getTerminator());
686 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
687 // Remove the old terminator.
688 MiddleBlock->getTerminator()->eraseFromParent();
690 // Create i+1 and fill the PHINode.
691 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
692 Induction->addIncoming(Zero, VectorPH);
693 Induction->addIncoming(NextIdx, VecBody);
694 // Create the compare.
695 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
696 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
698 // Now we have two terminators. Remove the old one from the block.
699 VecBody->getTerminator()->eraseFromParent();
701 // Fix the scalar body iteration count.
702 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
703 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
705 // Get ready to start creating new instructions into the vectorized body.
706 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
708 // Register the new loop.
709 Loop* Lp = new Loop();
710 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
712 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
714 Loop *ParentLoop = OrigLoop->getParentLoop();
716 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
717 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
718 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
722 LoopMiddleBlock = MiddleBlock;
723 LoopExitBlock = ExitBlock;
724 LoopVectorBody = VecBody;
725 LoopScalarBody = OldBasicBlock;
726 LoopBypassBlock = BypassBlock;
730 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
731 //===------------------------------------------------===//
733 // Notice: any optimization or new instruction that go
734 // into the code below should be also be implemented in
737 //===------------------------------------------------===//
738 typedef SmallVector<PHINode*, 4> PhiVector;
739 BasicBlock &BB = *OrigLoop->getHeader();
740 Constant *Zero = ConstantInt::get(
741 IntegerType::getInt32Ty(BB.getContext()), 0);
743 // In order to support reduction variables we need to be able to vectorize
744 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
745 // steages. First, we create a new vector PHI node with no incoming edges.
746 // We use this value when we vectorize all of the instructions that use the
747 // PHI. Next, after all of the instructions in the block are complete we
748 // add the new incoming edges to the PHI. At this point all of the
749 // instructions in the basic block are vectorized, so we can use them to
750 // construct the PHI.
753 // For each instruction in the old loop.
754 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
755 Instruction *Inst = it;
757 switch (Inst->getOpcode()) {
758 case Instruction::Br:
759 // Nothing to do for PHIs and BR, since we already took care of the
760 // loop control flow instructions.
762 case Instruction::PHI:{
763 PHINode* P = cast<PHINode>(Inst);
764 // Special handling for the induction var.
765 if (OldInduction == Inst)
767 // This is phase one of vectorizing PHIs.
768 // This has to be a reduction variable.
769 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
770 Type *VecTy = VectorType::get(Inst->getType(), VF);
771 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
772 PHIsToFix.push_back(P);
775 case Instruction::Add:
776 case Instruction::FAdd:
777 case Instruction::Sub:
778 case Instruction::FSub:
779 case Instruction::Mul:
780 case Instruction::FMul:
781 case Instruction::UDiv:
782 case Instruction::SDiv:
783 case Instruction::FDiv:
784 case Instruction::URem:
785 case Instruction::SRem:
786 case Instruction::FRem:
787 case Instruction::Shl:
788 case Instruction::LShr:
789 case Instruction::AShr:
790 case Instruction::And:
791 case Instruction::Or:
792 case Instruction::Xor: {
793 // Just widen binops.
794 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
795 Value *A = getVectorValue(Inst->getOperand(0));
796 Value *B = getVectorValue(Inst->getOperand(1));
797 // Use this vector value for all users of the original instruction.
798 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
801 case Instruction::Select: {
803 // If the selector is loop invariant we can create a select
804 // instruction with a scalar condition. Otherwise, use vector-select.
805 Value *Cond = Inst->getOperand(0);
806 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
808 // The condition can be loop invariant but still defined inside the
809 // loop. This means that we can't just use the original 'cond' value.
810 // We have to take the 'vectorized' value and pick the first lane.
811 // Instcombine will make this a no-op.
812 Cond = getVectorValue(Cond);
814 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
816 Value *Op0 = getVectorValue(Inst->getOperand(1));
817 Value *Op1 = getVectorValue(Inst->getOperand(2));
818 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
822 case Instruction::ICmp:
823 case Instruction::FCmp: {
824 // Widen compares. Generate vector compares.
825 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
826 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
827 Value *A = getVectorValue(Inst->getOperand(0));
828 Value *B = getVectorValue(Inst->getOperand(1));
830 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
832 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
836 case Instruction::Store: {
837 // Attempt to issue a wide store.
838 StoreInst *SI = dyn_cast<StoreInst>(Inst);
839 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
840 Value *Ptr = SI->getPointerOperand();
841 unsigned Alignment = SI->getAlignment();
842 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
843 // This store does not use GEPs.
844 if (!Legal->isConsecutiveGep(Gep)) {
845 scalarizeInstruction(Inst);
849 // The last index does not have to be the induction. It can be
850 // consecutive and be a function of the index. For example A[I+1];
851 unsigned NumOperands = Gep->getNumOperands();
852 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
853 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
855 // Create the new GEP with the new induction variable.
856 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
857 Gep2->setOperand(NumOperands - 1, LastIndex);
858 Ptr = Builder.Insert(Gep2);
859 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
860 Value *Val = getVectorValue(SI->getValueOperand());
861 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
864 case Instruction::Load: {
865 // Attempt to issue a wide load.
866 LoadInst *LI = dyn_cast<LoadInst>(Inst);
867 Type *RetTy = VectorType::get(LI->getType(), VF);
868 Value *Ptr = LI->getPointerOperand();
869 unsigned Alignment = LI->getAlignment();
870 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
872 // We don't have a gep. Scalarize the load.
873 if (!Legal->isConsecutiveGep(Gep)) {
874 scalarizeInstruction(Inst);
878 // The last index does not have to be the induction. It can be
879 // consecutive and be a function of the index. For example A[I+1];
880 unsigned NumOperands = Gep->getNumOperands();
881 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
882 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
884 // Create the new GEP with the new induction variable.
885 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
886 Gep2->setOperand(NumOperands - 1, LastIndex);
887 Ptr = Builder.Insert(Gep2);
888 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
889 LI = Builder.CreateLoad(Ptr);
890 LI->setAlignment(Alignment);
891 // Use this vector value for all users of the load.
895 case Instruction::ZExt:
896 case Instruction::SExt:
897 case Instruction::FPToUI:
898 case Instruction::FPToSI:
899 case Instruction::FPExt:
900 case Instruction::PtrToInt:
901 case Instruction::IntToPtr:
902 case Instruction::SIToFP:
903 case Instruction::UIToFP:
904 case Instruction::Trunc:
905 case Instruction::FPTrunc:
906 case Instruction::BitCast: {
907 /// Vectorize bitcasts.
908 CastInst *CI = dyn_cast<CastInst>(Inst);
909 Value *A = getVectorValue(Inst->getOperand(0));
910 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
911 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
916 /// All other instructions are unsupported. Scalarize them.
917 scalarizeInstruction(Inst);
920 }// end of for_each instr.
922 // At this point every instruction in the original loop is widended to
923 // a vector form. We are almost done. Now, we need to fix the PHI nodes
924 // that we vectorized. The PHI nodes are currently empty because we did
925 // not want to introduce cycles. Notice that the remaining PHI nodes
926 // that we need to fix are reduction variables.
928 // Create the 'reduced' values for each of the induction vars.
929 // The reduced values are the vector values that we scalarize and combine
930 // after the loop is finished.
931 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
933 PHINode *RdxPhi = *it;
934 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
935 assert(RdxPhi && "Unable to recover vectorized PHI");
937 // Find the reduction variable descriptor.
938 assert(Legal->getReductionVars()->count(RdxPhi) &&
939 "Unable to find the reduction variable");
940 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
941 (*Legal->getReductionVars())[RdxPhi];
943 // We need to generate a reduction vector from the incoming scalar.
944 // To do so, we need to generate the 'identity' vector and overide
945 // one of the elements with the incoming scalar reduction. We need
946 // to do it in the vector-loop preheader.
947 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
949 // This is the vector-clone of the value that leaves the loop.
950 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
951 Type *VecTy = VectorExit->getType();
953 // Find the reduction identity variable. The value of the enum is the
954 // identity. Zero for addition. One for Multiplication.
955 unsigned IdentitySclr = RdxDesc.Kind;
956 Constant *Identity = getUniformVector(IdentitySclr,
957 VecTy->getScalarType());
959 // This vector is the Identity vector where the first element is the
960 // incoming scalar reduction.
961 Value *VectorStart = Builder.CreateInsertElement(Identity,
962 RdxDesc.StartValue, Zero);
965 // Fix the vector-loop phi.
966 // We created the induction variable so we know that the
967 // preheader is the first entry.
968 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
970 // Reductions do not have to start at zero. They can start with
971 // any loop invariant values.
972 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
973 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
974 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
975 VecRdxPhi->addIncoming(Val, LoopVectorBody);
977 // Before each round, move the insertion point right between
978 // the PHIs and the values we are going to write.
979 // This allows us to write both PHINodes and the extractelement
981 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
983 // This PHINode contains the vectorized reduction variable, or
984 // the initial value vector, if we bypass the vector loop.
985 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
986 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
987 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
989 // Extract the first scalar.
991 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
992 // Extract and reduce the remaining vector elements.
993 for (unsigned i=1; i < VF; ++i) {
995 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
996 switch (RdxDesc.Kind) {
997 case LoopVectorizationLegality::IntegerAdd:
998 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1000 case LoopVectorizationLegality::IntegerMult:
1001 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1003 case LoopVectorizationLegality::IntegerOr:
1004 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1006 case LoopVectorizationLegality::IntegerAnd:
1007 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1009 case LoopVectorizationLegality::IntegerXor:
1010 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1013 llvm_unreachable("Unknown reduction operation");
1017 // Now, we need to fix the users of the reduction variable
1018 // inside and outside of the scalar remainder loop.
1019 // We know that the loop is in LCSSA form. We need to update the
1020 // PHI nodes in the exit blocks.
1021 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1022 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1023 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1024 if (!LCSSAPhi) continue;
1026 // All PHINodes need to have a single entry edge, or two if
1027 // we already fixed them.
1028 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1030 // We found our reduction value exit-PHI. Update it with the
1031 // incoming bypass edge.
1032 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1033 // Add an edge coming from the bypass.
1034 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1037 }// end of the LCSSA phi scan.
1039 // Fix the scalar loop reduction variable with the incoming reduction sum
1040 // from the vector body and from the backedge value.
1041 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1042 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1043 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1044 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1045 }// end of for each redux variable.
1048 void SingleBlockLoopVectorizer::cleanup() {
1049 // The original basic block.
1050 SE->forgetLoop(OrigLoop);
1053 bool LoopVectorizationLegality::canVectorize() {
1054 if (!TheLoop->getLoopPreheader()) {
1055 assert(false && "No preheader!!");
1056 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1060 // We can only vectorize single basic block loops.
1061 unsigned NumBlocks = TheLoop->getNumBlocks();
1062 if (NumBlocks != 1) {
1063 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1067 // We need to have a loop header.
1068 BasicBlock *BB = TheLoop->getHeader();
1069 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1071 // Go over each instruction and look at memory deps.
1072 if (!canVectorizeBlock(*BB)) {
1073 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1077 // ScalarEvolution needs to be able to find the exit count.
1078 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1079 if (ExitCount == SE->getCouldNotCompute()) {
1080 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1084 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
1086 // Okay! We can vectorize. At this point we don't have any other mem analysis
1087 // which may limit our maximum vectorization factor, so just return true with
1092 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1093 // Scan the instructions in the block and look for hazards.
1094 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1095 Instruction *I = it;
1097 PHINode *Phi = dyn_cast<PHINode>(I);
1099 // This should not happen because the loop should be normalized.
1100 if (Phi->getNumIncomingValues() != 2) {
1101 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1104 // We only look at integer phi nodes.
1105 if (!Phi->getType()->isIntegerTy()) {
1106 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
1110 if (isInductionVariable(Phi)) {
1112 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1115 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1119 if (AddReductionVar(Phi, IntegerAdd)) {
1120 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1123 if (AddReductionVar(Phi, IntegerMult)) {
1124 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1127 if (AddReductionVar(Phi, IntegerOr)) {
1128 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1131 if (AddReductionVar(Phi, IntegerAnd)) {
1132 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1135 if (AddReductionVar(Phi, IntegerXor)) {
1136 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1140 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1142 }// end of PHI handling
1144 // We still don't handle functions.
1145 CallInst *CI = dyn_cast<CallInst>(I);
1147 DEBUG(dbgs() << "LV: Found a call site:"<<
1148 CI->getCalledFunction()->getName() << "\n");
1152 // We do not re-vectorize vectors.
1153 if (!VectorType::isValidElementType(I->getType()) &&
1154 !I->getType()->isVoidTy()) {
1155 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1159 // Reduction instructions are allowed to have exit users.
1160 // All other instructions must not have external users.
1161 if (!AllowedExit.count(I))
1162 //Check that all of the users of the loop are inside the BB.
1163 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1165 Instruction *U = cast<Instruction>(*it);
1166 // This user may be a reduction exit value.
1167 BasicBlock *Parent = U->getParent();
1168 if (Parent != &BB) {
1169 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1176 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1180 // If the memory dependencies do not prevent us from
1181 // vectorizing, then vectorize.
1182 return canVectorizeMemory(BB);
1185 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1186 typedef SmallVector<Value*, 16> ValueVector;
1187 typedef SmallPtrSet<Value*, 16> ValueSet;
1188 // Holds the Load and Store *instructions*.
1192 // Scan the BB and collect legal loads and stores.
1193 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1194 Instruction *I = it;
1196 // If this is a load, save it. If this instruction can read from memory
1197 // but is not a load, then we quit. Notice that we don't handle function
1198 // calls that read or write.
1199 if (I->mayReadFromMemory()) {
1200 LoadInst *Ld = dyn_cast<LoadInst>(I);
1201 if (!Ld) return false;
1202 if (!Ld->isSimple()) {
1203 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1206 Loads.push_back(Ld);
1210 // Save store instructions. Abort if other instructions write to memory.
1211 if (I->mayWriteToMemory()) {
1212 StoreInst *St = dyn_cast<StoreInst>(I);
1213 if (!St) return false;
1214 if (!St->isSimple()) {
1215 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1218 Stores.push_back(St);
1222 // Now we have two lists that hold the loads and the stores.
1223 // Next, we find the pointers that they use.
1225 // Check if we see any stores. If there are no stores, then we don't
1226 // care if the pointers are *restrict*.
1227 if (!Stores.size()) {
1228 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1232 // Holds the read and read-write *pointers* that we find.
1234 ValueVector ReadWrites;
1236 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1237 // multiple times on the same object. If the ptr is accessed twice, once
1238 // for read and once for write, it will only appear once (on the write
1239 // list). This is okay, since we are going to check for conflicts between
1240 // writes and between reads and writes, but not between reads and reads.
1243 ValueVector::iterator I, IE;
1244 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1245 StoreInst *ST = dyn_cast<StoreInst>(*I);
1246 assert(ST && "Bad StoreInst");
1247 Value* Ptr = ST->getPointerOperand();
1248 // If we did *not* see this pointer before, insert it to
1249 // the read-write list. At this phase it is only a 'write' list.
1250 if (Seen.insert(Ptr))
1251 ReadWrites.push_back(Ptr);
1254 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1255 LoadInst *LD = dyn_cast<LoadInst>(*I);
1256 assert(LD && "Bad LoadInst");
1257 Value* Ptr = LD->getPointerOperand();
1258 // If we did *not* see this pointer before, insert it to the
1259 // read list. If we *did* see it before, then it is already in
1260 // the read-write list. This allows us to vectorize expressions
1261 // such as A[i] += x; Because the address of A[i] is a read-write
1262 // pointer. This only works if the index of A[i] is consecutive.
1263 // If the address of i is unknown (for example A[B[i]]) then we may
1264 // read a few words, modify, and write a few words, and some of the
1265 // words may be written to the same address.
1266 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1267 Reads.push_back(Ptr);
1270 // Now that the pointers are in two lists (Reads and ReadWrites), we
1271 // can check that there are no conflicts between each of the writes and
1272 // between the writes to the reads.
1273 ValueSet WriteObjects;
1274 ValueVector TempObjects;
1276 // Check that the read-writes do not conflict with other read-write
1278 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1279 GetUnderlyingObjects(*I, TempObjects, DL);
1280 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1282 if (!isIdentifiedObject(*it)) {
1283 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1286 if (!WriteObjects.insert(*it)) {
1287 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1292 TempObjects.clear();
1295 /// Check that the reads don't conflict with the read-writes.
1296 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1297 GetUnderlyingObjects(*I, TempObjects, DL);
1298 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1300 if (!isIdentifiedObject(*it)) {
1301 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1304 if (WriteObjects.count(*it)) {
1305 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1310 TempObjects.clear();
1317 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1318 ReductionKind Kind) {
1319 if (Phi->getNumIncomingValues() != 2)
1322 // Find the possible incoming reduction variable.
1323 BasicBlock *BB = Phi->getParent();
1324 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1325 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1326 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1328 // ExitInstruction is the single value which is used outside the loop.
1329 // We only allow for a single reduction value to be used outside the loop.
1330 // This includes users of the reduction, variables (which form a cycle
1331 // which ends in the phi node).
1332 Instruction *ExitInstruction = 0;
1334 // Iter is our iterator. We start with the PHI node and scan for all of the
1335 // users of this instruction. All users must be instructions which can be
1336 // used as reduction variables (such as ADD). We may have a single
1337 // out-of-block user. They cycle must end with the original PHI.
1338 // Also, we can't have multiple block-local users.
1339 Instruction *Iter = Phi;
1341 // Any reduction instr must be of one of the allowed kinds.
1342 if (!isReductionInstr(Iter, Kind))
1345 // Did we found a user inside this block ?
1346 bool FoundInBlockUser = false;
1347 // Did we reach the initial PHI node ?
1348 bool FoundStartPHI = false;
1350 // If the instruction has no users then this is a broken
1351 // chain and can't be a reduction variable.
1352 if (Iter->use_empty())
1355 // For each of the *users* of iter.
1356 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1358 Instruction *U = cast<Instruction>(*it);
1359 // We already know that the PHI is a user.
1361 FoundStartPHI = true;
1364 // Check if we found the exit user.
1365 BasicBlock *Parent = U->getParent();
1367 // We must have a single exit instruction.
1368 if (ExitInstruction != 0)
1370 ExitInstruction = Iter;
1372 // We can't have multiple inside users.
1373 if (FoundInBlockUser)
1375 FoundInBlockUser = true;
1379 // We found a reduction var if we have reached the original
1380 // phi node and we only have a single instruction with out-of-loop
1382 if (FoundStartPHI && ExitInstruction) {
1383 // This instruction is allowed to have out-of-loop users.
1384 AllowedExit.insert(ExitInstruction);
1386 // Save the description of this reduction variable.
1387 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1388 Reductions[Phi] = RD;
1395 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1396 ReductionKind Kind) {
1397 switch (I->getOpcode()) {
1400 case Instruction::PHI:
1403 case Instruction::Add:
1404 case Instruction::Sub:
1405 return Kind == IntegerAdd;
1406 case Instruction::Mul:
1407 case Instruction::UDiv:
1408 case Instruction::SDiv:
1409 return Kind == IntegerMult;
1410 case Instruction::And:
1411 return Kind == IntegerAnd;
1412 case Instruction::Or:
1413 return Kind == IntegerOr;
1414 case Instruction::Xor:
1415 return Kind == IntegerXor;
1419 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1420 // Check that the PHI is consecutive and starts at zero.
1421 const SCEV *PhiScev = SE->getSCEV(Phi);
1422 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1424 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1427 const SCEV *Step = AR->getStepRecurrence(*SE);
1428 const SCEV *Start = AR->getStart();
1430 if (!Step->isOne() || !Start->isZero()) {
1431 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1438 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1440 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1444 float Cost = expectedCost(1);
1446 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1447 for (unsigned i=2; i <= VF; i*=2) {
1448 // Notice that the vector loop needs to be executed less times, so
1449 // we need to divide the cost of the vector loops by the width of
1450 // the vector elements.
1451 float VectorCost = expectedCost(i) / (float)i;
1452 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1453 (int)VectorCost << ".\n");
1454 if (VectorCost < Cost) {
1460 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1464 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1465 // We can only estimate the cost of single basic block loops.
1466 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1468 BasicBlock *BB = TheLoop->getHeader();
1471 // For each instruction in the old loop.
1472 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1473 Instruction *Inst = it;
1474 unsigned C = getInstructionCost(Inst, VF);
1476 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1477 " For instruction: "<< *Inst << "\n");
1484 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1485 assert(VTTI && "Invalid vector target transformation info");
1487 Type *RetTy = I->getType();
1488 Type *VectorTy = ToVectorTy(RetTy, VF);
1490 // TODO: We need to estimate the cost of intrinsic calls.
1491 switch (I->getOpcode()) {
1492 case Instruction::GetElementPtr:
1493 // We mark this instruction as zero-cost because scalar GEPs are usually
1494 // lowered to the intruction addressing mode. At the moment we don't
1495 // generate vector geps.
1497 case Instruction::Br: {
1498 return VTTI->getInstrCost(I->getOpcode());
1500 case Instruction::PHI:
1502 case Instruction::Add:
1503 case Instruction::FAdd:
1504 case Instruction::Sub:
1505 case Instruction::FSub:
1506 case Instruction::Mul:
1507 case Instruction::FMul:
1508 case Instruction::UDiv:
1509 case Instruction::SDiv:
1510 case Instruction::FDiv:
1511 case Instruction::URem:
1512 case Instruction::SRem:
1513 case Instruction::FRem:
1514 case Instruction::Shl:
1515 case Instruction::LShr:
1516 case Instruction::AShr:
1517 case Instruction::And:
1518 case Instruction::Or:
1519 case Instruction::Xor: {
1520 return VTTI->getInstrCost(I->getOpcode(), VectorTy);
1522 case Instruction::Select: {
1523 SelectInst *SI = cast<SelectInst>(I);
1524 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1525 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1526 Type *CondTy = SI->getCondition()->getType();
1528 CondTy = VectorType::get(CondTy, VF);
1530 return VTTI->getInstrCost(I->getOpcode(), VectorTy, CondTy);
1532 case Instruction::ICmp:
1533 case Instruction::FCmp: {
1534 Type *ValTy = I->getOperand(0)->getType();
1535 VectorTy = ToVectorTy(ValTy, VF);
1536 return VTTI->getInstrCost(I->getOpcode(), VectorTy);
1538 case Instruction::Store: {
1539 StoreInst *SI = cast<StoreInst>(I);
1540 Type *ValTy = SI->getValueOperand()->getType();
1541 VectorTy = ToVectorTy(ValTy, VF);
1544 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
1545 SI->getAlignment(), SI->getPointerAddressSpace());
1547 // Scalarized stores.
1548 if (!Legal->isConsecutiveGep(SI->getPointerOperand())) {
1550 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1552 // The cost of extracting from the value vector.
1553 Cost += VF * (ExtCost);
1554 // The cost of the scalar stores.
1555 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1556 ValTy->getScalarType(),
1558 SI->getPointerAddressSpace());
1563 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
1564 SI->getPointerAddressSpace());
1566 case Instruction::Load: {
1567 LoadInst *LI = cast<LoadInst>(I);
1570 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
1572 LI->getPointerAddressSpace());
1574 // Scalarized loads.
1575 if (!Legal->isConsecutiveGep(LI->getPointerOperand())) {
1577 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
1578 // The cost of inserting the loaded value into the result vector.
1579 Cost += VF * (InCost);
1580 // The cost of the scalar stores.
1581 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1582 RetTy->getScalarType(),
1584 LI->getPointerAddressSpace());
1589 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
1590 LI->getPointerAddressSpace());
1592 case Instruction::ZExt:
1593 case Instruction::SExt:
1594 case Instruction::FPToUI:
1595 case Instruction::FPToSI:
1596 case Instruction::FPExt:
1597 case Instruction::PtrToInt:
1598 case Instruction::IntToPtr:
1599 case Instruction::SIToFP:
1600 case Instruction::UIToFP:
1601 case Instruction::Trunc:
1602 case Instruction::FPTrunc:
1603 case Instruction::BitCast: {
1604 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
1605 return VTTI->getInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
1608 // We are scalarizing the instruction. Return the cost of the scalar
1609 // instruction, plus the cost of insert and extract into vector
1610 // elements, times the vector width.
1613 bool IsVoid = RetTy->isVoidTy();
1615 unsigned InsCost = (IsVoid ? 0 :
1616 VTTI->getInstrCost(Instruction::InsertElement,
1619 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1622 // The cost of inserting the results plus extracting each one of the
1624 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
1626 // The cost of executing VF copies of the scalar instruction.
1627 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
1633 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
1634 if (Scalar->isVoidTy() || VF == 1)
1636 return VectorType::get(Scalar, VF);
1641 char LoopVectorize::ID = 0;
1642 static const char lv_name[] = "Loop Vectorization";
1643 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1644 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1645 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1646 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1647 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1650 Pass *createLoopVectorizePass() {
1651 return new LoopVectorize();