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);
257 /// Returns true if this instruction will remain scalar after vectorization.
258 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
261 /// Check if a single basic block loop is vectorizable.
262 /// At this point we know that this is a loop with a constant trip count
263 /// and we only need to check individual instructions.
264 bool canVectorizeBlock(BasicBlock &BB);
266 /// When we vectorize loops we may change the order in which
267 /// we read and write from memory. This method checks if it is
268 /// legal to vectorize the code, considering only memory constrains.
269 /// Returns true if BB is vectorizable
270 bool canVectorizeMemory(BasicBlock &BB);
272 /// Returns True, if 'Phi' is the kind of reduction variable for type
273 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
274 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
275 /// Returns true if the instruction I can be a reduction variable of type
277 bool isReductionInstr(Instruction *I, ReductionKind Kind);
278 /// Returns True, if 'Phi' is an induction variable.
279 bool isInductionVariable(PHINode *Phi);
281 /// The loop that we evaluate.
285 /// DataLayout analysis.
288 // --- vectorization state --- //
290 /// Holds the induction variable.
292 /// Holds the reduction variables.
293 ReductionList Reductions;
294 /// Allowed outside users. This holds the reduction
295 /// vars which can be accessed from outside the loop.
296 SmallPtrSet<Value*, 4> AllowedExit;
297 /// This set holds the variables which are known to be uniform after
299 SmallPtrSet<Instruction*, 4> Uniforms;
302 /// LoopVectorizationCostModel - estimates the expected speedups due to
304 /// In many cases vectorization is not profitable. This can happen because
305 /// of a number of reasons. In this class we mainly attempt to predict
306 /// the expected speedup/slowdowns due to the supported instruction set.
307 /// We use the VectorTargetTransformInfo to query the different backends
308 /// for the cost of different operations.
309 class LoopVectorizationCostModel {
312 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
313 LoopVectorizationLegality *Leg,
314 const VectorTargetTransformInfo *Vtti):
315 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
317 /// Returns the most profitable vectorization factor for the loop that is
318 /// smaller or equal to the VF argument. This method checks every power
320 unsigned findBestVectorizationFactor(unsigned VF = 8);
323 /// Returns the expected execution cost. The unit of the cost does
324 /// not matter because we use the 'cost' units to compare different
325 /// vector widths. The cost that is returned is *not* normalized by
326 /// the factor width.
327 unsigned expectedCost(unsigned VF);
329 /// Returns the execution time cost of an instruction for a given vector
330 /// width. Vector width of one means scalar.
331 unsigned getInstructionCost(Instruction *I, unsigned VF);
333 /// A helper function for converting Scalar types to vector types.
334 /// If the incoming type is void, we return void. If the VF is 1, we return
336 static Type* ToVectorTy(Type *Scalar, unsigned VF);
338 /// The loop that we evaluate.
343 /// Vectorization legality.
344 LoopVectorizationLegality *Legal;
345 /// Vector target information.
346 const VectorTargetTransformInfo *VTTI;
349 struct LoopVectorize : public LoopPass {
350 static char ID; // Pass identification, replacement for typeid
352 LoopVectorize() : LoopPass(ID) {
353 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
359 TargetTransformInfo *TTI;
361 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
362 // We only vectorize innermost loops.
366 SE = &getAnalysis<ScalarEvolution>();
367 DL = getAnalysisIfAvailable<DataLayout>();
368 LI = &getAnalysis<LoopInfo>();
369 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
371 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
372 L->getHeader()->getParent()->getName() << "\"\n");
374 // Check if it is legal to vectorize the loop.
375 LoopVectorizationLegality LVL(L, SE, DL);
376 if (!LVL.canVectorize()) {
377 DEBUG(dbgs() << "LV: Not vectorizing.\n");
381 // Select the preffered vectorization factor.
383 if (VectorizationFactor == 0) {
384 const VectorTargetTransformInfo *VTTI = 0;
386 VTTI = TTI->getVectorTargetTransformInfo();
387 // Use the cost model.
388 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
389 VF = CM.findBestVectorizationFactor();
392 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
397 // Use the user command flag.
398 VF = VectorizationFactor;
401 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ").\n");
403 // If we decided that it is *legal* to vectorizer the loop then do it.
404 SingleBlockLoopVectorizer LB(L, SE, LI, &LPM, VF);
407 DEBUG(verifyFunction(*L->getHeader()->getParent()));
411 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
412 LoopPass::getAnalysisUsage(AU);
413 AU.addRequiredID(LoopSimplifyID);
414 AU.addRequiredID(LCSSAID);
415 AU.addRequired<LoopInfo>();
416 AU.addRequired<ScalarEvolution>();
421 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
422 // Instructions that access the old induction variable
423 // actually want to get the new one.
424 if (V == OldInduction)
427 LLVMContext &C = V->getContext();
428 Type *VTy = VectorType::get(V->getType(), VF);
429 Type *I32 = IntegerType::getInt32Ty(C);
430 Constant *Zero = ConstantInt::get(I32, 0);
431 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
432 Value *UndefVal = UndefValue::get(VTy);
433 // Insert the value into a new vector.
434 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
435 // Broadcast the scalar into all locations in the vector.
436 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
438 // We are accessing the induction variable. Make sure to promote the
439 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
441 return getConsecutiveVector(Shuf);
445 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
446 assert(Val->getType()->isVectorTy() && "Must be a vector");
447 assert(Val->getType()->getScalarType()->isIntegerTy() &&
448 "Elem must be an integer");
450 Type *ITy = Val->getType()->getScalarType();
451 VectorType *Ty = cast<VectorType>(Val->getType());
452 unsigned VLen = Ty->getNumElements();
453 SmallVector<Constant*, 8> Indices;
455 // Create a vector of consecutive numbers from zero to VF.
456 for (unsigned i = 0; i < VLen; ++i)
457 Indices.push_back(ConstantInt::get(ITy, i));
459 // Add the consecutive indices to the vector value.
460 Constant *Cv = ConstantVector::get(Indices);
461 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
462 return Builder.CreateAdd(Val, Cv, "induction");
465 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
466 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
470 unsigned NumOperands = Gep->getNumOperands();
471 Value *LastIndex = Gep->getOperand(NumOperands - 1);
473 // Check that all of the gep indices are uniform except for the last.
474 for (unsigned i = 0; i < NumOperands - 1; ++i)
475 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
478 // We can emit wide load/stores only of the last index is the induction
480 const SCEV *Last = SE->getSCEV(LastIndex);
481 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
482 const SCEV *Step = AR->getStepRecurrence(*SE);
484 // The memory is consecutive because the last index is consecutive
485 // and all other indices are loop invariant.
493 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
494 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
495 // If we saved a vectorized copy of V, use it.
496 Value *&MapEntry = WidenMap[V];
500 // Broadcast V and save the value for future uses.
501 Value *B = getBroadcastInstrs(V);
507 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
508 SmallVector<Constant*, 8> Indices;
509 // Create a vector of consecutive numbers from zero to VF.
510 for (unsigned i = 0; i < VF; ++i)
511 Indices.push_back(ConstantInt::get(ScalarTy, Val));
513 // Add the consecutive indices to the vector value.
514 return ConstantVector::get(Indices);
517 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
518 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
519 // Holds vector parameters or scalars, in case of uniform vals.
520 SmallVector<Value*, 8> Params;
522 // Find all of the vectorized parameters.
523 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
524 Value *SrcOp = Instr->getOperand(op);
526 // If we are accessing the old induction variable, use the new one.
527 if (SrcOp == OldInduction) {
528 Params.push_back(getBroadcastInstrs(Induction));
532 // Try using previously calculated values.
533 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
535 // If the src is an instruction that appeared earlier in the basic block
536 // then it should already be vectorized.
537 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
538 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
539 // The parameter is a vector value from earlier.
540 Params.push_back(WidenMap[SrcInst]);
542 // The parameter is a scalar from outside the loop. Maybe even a constant.
543 Params.push_back(SrcOp);
547 assert(Params.size() == Instr->getNumOperands() &&
548 "Invalid number of operands");
550 // Does this instruction return a value ?
551 bool IsVoidRetTy = Instr->getType()->isVoidTy();
552 Value *VecResults = 0;
554 // If we have a return value, create an empty vector. We place the scalarized
555 // instructions in this vector.
557 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
559 // For each scalar that we create:
560 for (unsigned i = 0; i < VF; ++i) {
561 Instruction *Cloned = Instr->clone();
563 Cloned->setName(Instr->getName() + ".cloned");
564 // Replace the operands of the cloned instrucions with extracted scalars.
565 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
566 Value *Op = Params[op];
567 // Param is a vector. Need to extract the right lane.
568 if (Op->getType()->isVectorTy())
569 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
570 Cloned->setOperand(op, Op);
573 // Place the cloned scalar in the new loop.
574 Builder.Insert(Cloned);
576 // If the original scalar returns a value we need to place it in a vector
577 // so that future users will be able to use it.
579 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
580 Builder.getInt32(i));
584 WidenMap[Instr] = VecResults;
587 void SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
589 In this function we generate a new loop. The new loop will contain
590 the vectorized instructions while the old loop will continue to run the
593 [ ] <-- vector loop bypass.
596 | [ ] <-- vector pre header.
600 | [ ]_| <-- vector loop.
603 >[ ] <--- middle-block.
606 | [ ] <--- new preheader.
610 | [ ]_| <-- old scalar loop to handle remainder.
617 // This is the original scalar-loop preheader.
618 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
619 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
620 assert(ExitBlock && "Must have an exit block");
622 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
623 assert(BypassBlock && "Invalid loop structure");
625 BasicBlock *VectorPH =
626 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
627 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
630 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
632 BasicBlock *ScalarPH =
633 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
635 // Find the induction variable.
636 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
637 OldInduction = Legal->getInduction();
638 assert(OldInduction && "We must have a single phi node.");
639 Type *IdxTy = OldInduction->getType();
641 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
643 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
645 // Generate the induction variable.
646 Induction = Builder.CreatePHI(IdxTy, 2, "index");
647 Constant *Zero = ConstantInt::get(IdxTy, 0);
648 Constant *Step = ConstantInt::get(IdxTy, VF);
650 // Find the loop boundaries.
651 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
652 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
654 // Get the total trip count from the count by adding 1.
655 ExitCount = SE->getAddExpr(ExitCount,
656 SE->getConstant(ExitCount->getType(), 1));
658 // Expand the trip count and place the new instructions in the preheader.
659 // Notice that the pre-header does not change, only the loop body.
660 SCEVExpander Exp(*SE, "induction");
661 Instruction *Loc = BypassBlock->getTerminator();
663 // We may need to extend the index in case there is a type mismatch.
664 // We know that the count starts at zero and does not overflow.
665 // We are using Zext because it should be less expensive.
666 if (ExitCount->getType() != Induction->getType())
667 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
669 // Count holds the overall loop count (N).
670 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
671 // Now we need to generate the expression for N - (N % VF), which is
672 // the part that the vectorized body will execute.
673 Constant *CIVF = ConstantInt::get(IdxTy, VF);
674 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
675 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
677 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
678 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
679 CountRoundDown, ConstantInt::getNullValue(IdxTy),
681 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
682 // Remove the old terminator.
683 Loc->eraseFromParent();
685 // Add a check in the middle block to see if we have completed
686 // all of the iterations in the first vector loop.
687 // If (N - N%VF) == N, then we *don't* need to run the remainder.
688 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
689 CountRoundDown, "cmp.n",
690 MiddleBlock->getTerminator());
692 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
693 // Remove the old terminator.
694 MiddleBlock->getTerminator()->eraseFromParent();
696 // Create i+1 and fill the PHINode.
697 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
698 Induction->addIncoming(Zero, VectorPH);
699 Induction->addIncoming(NextIdx, VecBody);
700 // Create the compare.
701 Value *ICmp = Builder.CreateICmpEQ(NextIdx, CountRoundDown);
702 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
704 // Now we have two terminators. Remove the old one from the block.
705 VecBody->getTerminator()->eraseFromParent();
707 // Fix the scalar body iteration count.
708 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
709 OldInduction->setIncomingValue(BlockIdx, CountRoundDown);
711 // Get ready to start creating new instructions into the vectorized body.
712 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
714 // Register the new loop.
715 Loop* Lp = new Loop();
716 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
718 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
720 Loop *ParentLoop = OrigLoop->getParentLoop();
722 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
723 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
724 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
728 LoopMiddleBlock = MiddleBlock;
729 LoopExitBlock = ExitBlock;
730 LoopVectorBody = VecBody;
731 LoopScalarBody = OldBasicBlock;
732 LoopBypassBlock = BypassBlock;
736 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
737 //===------------------------------------------------===//
739 // Notice: any optimization or new instruction that go
740 // into the code below should be also be implemented in
743 //===------------------------------------------------===//
744 typedef SmallVector<PHINode*, 4> PhiVector;
745 BasicBlock &BB = *OrigLoop->getHeader();
746 Constant *Zero = ConstantInt::get(
747 IntegerType::getInt32Ty(BB.getContext()), 0);
749 // In order to support reduction variables we need to be able to vectorize
750 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
751 // steages. First, we create a new vector PHI node with no incoming edges.
752 // We use this value when we vectorize all of the instructions that use the
753 // PHI. Next, after all of the instructions in the block are complete we
754 // add the new incoming edges to the PHI. At this point all of the
755 // instructions in the basic block are vectorized, so we can use them to
756 // construct the PHI.
759 // For each instruction in the old loop.
760 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
761 Instruction *Inst = it;
763 switch (Inst->getOpcode()) {
764 case Instruction::Br:
765 // Nothing to do for PHIs and BR, since we already took care of the
766 // loop control flow instructions.
768 case Instruction::PHI:{
769 PHINode* P = cast<PHINode>(Inst);
770 // Special handling for the induction var.
771 if (OldInduction == Inst)
773 // This is phase one of vectorizing PHIs.
774 // This has to be a reduction variable.
775 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
776 Type *VecTy = VectorType::get(Inst->getType(), VF);
777 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
778 PHIsToFix.push_back(P);
781 case Instruction::Add:
782 case Instruction::FAdd:
783 case Instruction::Sub:
784 case Instruction::FSub:
785 case Instruction::Mul:
786 case Instruction::FMul:
787 case Instruction::UDiv:
788 case Instruction::SDiv:
789 case Instruction::FDiv:
790 case Instruction::URem:
791 case Instruction::SRem:
792 case Instruction::FRem:
793 case Instruction::Shl:
794 case Instruction::LShr:
795 case Instruction::AShr:
796 case Instruction::And:
797 case Instruction::Or:
798 case Instruction::Xor: {
799 // Just widen binops.
800 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
801 Value *A = getVectorValue(Inst->getOperand(0));
802 Value *B = getVectorValue(Inst->getOperand(1));
803 // Use this vector value for all users of the original instruction.
804 WidenMap[Inst] = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
807 case Instruction::Select: {
809 // If the selector is loop invariant we can create a select
810 // instruction with a scalar condition. Otherwise, use vector-select.
811 Value *Cond = Inst->getOperand(0);
812 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
814 // The condition can be loop invariant but still defined inside the
815 // loop. This means that we can't just use the original 'cond' value.
816 // We have to take the 'vectorized' value and pick the first lane.
817 // Instcombine will make this a no-op.
818 Cond = getVectorValue(Cond);
820 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
822 Value *Op0 = getVectorValue(Inst->getOperand(1));
823 Value *Op1 = getVectorValue(Inst->getOperand(2));
824 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
828 case Instruction::ICmp:
829 case Instruction::FCmp: {
830 // Widen compares. Generate vector compares.
831 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
832 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
833 Value *A = getVectorValue(Inst->getOperand(0));
834 Value *B = getVectorValue(Inst->getOperand(1));
836 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
838 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
842 case Instruction::Store: {
843 // Attempt to issue a wide store.
844 StoreInst *SI = dyn_cast<StoreInst>(Inst);
845 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
846 Value *Ptr = SI->getPointerOperand();
847 unsigned Alignment = SI->getAlignment();
848 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
849 // This store does not use GEPs.
850 if (!Legal->isConsecutiveGep(Gep)) {
851 scalarizeInstruction(Inst);
855 // The last index does not have to be the induction. It can be
856 // consecutive and be a function of the index. For example A[I+1];
857 unsigned NumOperands = Gep->getNumOperands();
858 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
859 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
861 // Create the new GEP with the new induction variable.
862 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
863 Gep2->setOperand(NumOperands - 1, LastIndex);
864 Ptr = Builder.Insert(Gep2);
865 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
866 Value *Val = getVectorValue(SI->getValueOperand());
867 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
870 case Instruction::Load: {
871 // Attempt to issue a wide load.
872 LoadInst *LI = dyn_cast<LoadInst>(Inst);
873 Type *RetTy = VectorType::get(LI->getType(), VF);
874 Value *Ptr = LI->getPointerOperand();
875 unsigned Alignment = LI->getAlignment();
876 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
878 // We don't have a gep. Scalarize the load.
879 if (!Legal->isConsecutiveGep(Gep)) {
880 scalarizeInstruction(Inst);
884 // The last index does not have to be the induction. It can be
885 // consecutive and be a function of the index. For example A[I+1];
886 unsigned NumOperands = Gep->getNumOperands();
887 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
888 LastIndex = Builder.CreateExtractElement(LastIndex, Builder.getInt32(0));
890 // Create the new GEP with the new induction variable.
891 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
892 Gep2->setOperand(NumOperands - 1, LastIndex);
893 Ptr = Builder.Insert(Gep2);
894 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
895 LI = Builder.CreateLoad(Ptr);
896 LI->setAlignment(Alignment);
897 // Use this vector value for all users of the load.
901 case Instruction::ZExt:
902 case Instruction::SExt:
903 case Instruction::FPToUI:
904 case Instruction::FPToSI:
905 case Instruction::FPExt:
906 case Instruction::PtrToInt:
907 case Instruction::IntToPtr:
908 case Instruction::SIToFP:
909 case Instruction::UIToFP:
910 case Instruction::Trunc:
911 case Instruction::FPTrunc:
912 case Instruction::BitCast: {
913 /// Vectorize bitcasts.
914 CastInst *CI = dyn_cast<CastInst>(Inst);
915 Value *A = getVectorValue(Inst->getOperand(0));
916 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
917 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
922 /// All other instructions are unsupported. Scalarize them.
923 scalarizeInstruction(Inst);
926 }// end of for_each instr.
928 // At this point every instruction in the original loop is widended to
929 // a vector form. We are almost done. Now, we need to fix the PHI nodes
930 // that we vectorized. The PHI nodes are currently empty because we did
931 // not want to introduce cycles. Notice that the remaining PHI nodes
932 // that we need to fix are reduction variables.
934 // Create the 'reduced' values for each of the induction vars.
935 // The reduced values are the vector values that we scalarize and combine
936 // after the loop is finished.
937 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
939 PHINode *RdxPhi = *it;
940 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
941 assert(RdxPhi && "Unable to recover vectorized PHI");
943 // Find the reduction variable descriptor.
944 assert(Legal->getReductionVars()->count(RdxPhi) &&
945 "Unable to find the reduction variable");
946 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
947 (*Legal->getReductionVars())[RdxPhi];
949 // We need to generate a reduction vector from the incoming scalar.
950 // To do so, we need to generate the 'identity' vector and overide
951 // one of the elements with the incoming scalar reduction. We need
952 // to do it in the vector-loop preheader.
953 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
955 // This is the vector-clone of the value that leaves the loop.
956 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
957 Type *VecTy = VectorExit->getType();
959 // Find the reduction identity variable. The value of the enum is the
960 // identity. Zero for addition. One for Multiplication.
961 unsigned IdentitySclr = RdxDesc.Kind;
962 Constant *Identity = getUniformVector(IdentitySclr,
963 VecTy->getScalarType());
965 // This vector is the Identity vector where the first element is the
966 // incoming scalar reduction.
967 Value *VectorStart = Builder.CreateInsertElement(Identity,
968 RdxDesc.StartValue, Zero);
971 // Fix the vector-loop phi.
972 // We created the induction variable so we know that the
973 // preheader is the first entry.
974 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
976 // Reductions do not have to start at zero. They can start with
977 // any loop invariant values.
978 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
979 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
980 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
981 VecRdxPhi->addIncoming(Val, LoopVectorBody);
983 // Before each round, move the insertion point right between
984 // the PHIs and the values we are going to write.
985 // This allows us to write both PHINodes and the extractelement
987 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
989 // This PHINode contains the vectorized reduction variable, or
990 // the initial value vector, if we bypass the vector loop.
991 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
992 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
993 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
995 // Extract the first scalar.
997 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
998 // Extract and reduce the remaining vector elements.
999 for (unsigned i=1; i < VF; ++i) {
1001 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1002 switch (RdxDesc.Kind) {
1003 case LoopVectorizationLegality::IntegerAdd:
1004 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1006 case LoopVectorizationLegality::IntegerMult:
1007 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1009 case LoopVectorizationLegality::IntegerOr:
1010 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1012 case LoopVectorizationLegality::IntegerAnd:
1013 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1015 case LoopVectorizationLegality::IntegerXor:
1016 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1019 llvm_unreachable("Unknown reduction operation");
1023 // Now, we need to fix the users of the reduction variable
1024 // inside and outside of the scalar remainder loop.
1025 // We know that the loop is in LCSSA form. We need to update the
1026 // PHI nodes in the exit blocks.
1027 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1028 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1029 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1030 if (!LCSSAPhi) continue;
1032 // All PHINodes need to have a single entry edge, or two if
1033 // we already fixed them.
1034 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1036 // We found our reduction value exit-PHI. Update it with the
1037 // incoming bypass edge.
1038 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1039 // Add an edge coming from the bypass.
1040 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1043 }// end of the LCSSA phi scan.
1045 // Fix the scalar loop reduction variable with the incoming reduction sum
1046 // from the vector body and from the backedge value.
1047 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1048 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1049 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1050 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1051 }// end of for each redux variable.
1054 void SingleBlockLoopVectorizer::cleanup() {
1055 // The original basic block.
1056 SE->forgetLoop(OrigLoop);
1059 bool LoopVectorizationLegality::canVectorize() {
1060 if (!TheLoop->getLoopPreheader()) {
1061 assert(false && "No preheader!!");
1062 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1066 // We can only vectorize single basic block loops.
1067 unsigned NumBlocks = TheLoop->getNumBlocks();
1068 if (NumBlocks != 1) {
1069 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1073 // We need to have a loop header.
1074 BasicBlock *BB = TheLoop->getHeader();
1075 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1077 // Go over each instruction and look at memory deps.
1078 if (!canVectorizeBlock(*BB)) {
1079 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1083 // ScalarEvolution needs to be able to find the exit count.
1084 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1085 if (ExitCount == SE->getCouldNotCompute()) {
1086 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1090 DEBUG(dbgs() << "LV: We can vectorize this loop!\n");
1092 // Okay! We can vectorize. At this point we don't have any other mem analysis
1093 // which may limit our maximum vectorization factor, so just return true with
1098 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1099 // Scan the instructions in the block and look for hazards.
1100 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1101 Instruction *I = it;
1103 PHINode *Phi = dyn_cast<PHINode>(I);
1105 // This should not happen because the loop should be normalized.
1106 if (Phi->getNumIncomingValues() != 2) {
1107 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1110 // We only look at integer phi nodes.
1111 if (!Phi->getType()->isIntegerTy()) {
1112 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
1116 if (isInductionVariable(Phi)) {
1118 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1121 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1125 if (AddReductionVar(Phi, IntegerAdd)) {
1126 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1129 if (AddReductionVar(Phi, IntegerMult)) {
1130 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1133 if (AddReductionVar(Phi, IntegerOr)) {
1134 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1137 if (AddReductionVar(Phi, IntegerAnd)) {
1138 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1141 if (AddReductionVar(Phi, IntegerXor)) {
1142 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1146 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1148 }// end of PHI handling
1150 // We still don't handle functions.
1151 CallInst *CI = dyn_cast<CallInst>(I);
1153 DEBUG(dbgs() << "LV: Found a call site:"<<
1154 CI->getCalledFunction()->getName() << "\n");
1158 // We do not re-vectorize vectors.
1159 if (!VectorType::isValidElementType(I->getType()) &&
1160 !I->getType()->isVoidTy()) {
1161 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1165 // Reduction instructions are allowed to have exit users.
1166 // All other instructions must not have external users.
1167 if (!AllowedExit.count(I))
1168 //Check that all of the users of the loop are inside the BB.
1169 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1171 Instruction *U = cast<Instruction>(*it);
1172 // This user may be a reduction exit value.
1173 BasicBlock *Parent = U->getParent();
1174 if (Parent != &BB) {
1175 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1182 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1186 // Don't vectorize if the memory dependencies do not allow vectorization.
1187 if (!canVectorizeMemory(BB))
1190 // We now know that the loop is vectorizable!
1191 // Collect variables that will remain uniform after vectorization.
1192 std::vector<Value*> Worklist;
1194 // Start with the conditional branch and walk up the block.
1195 Worklist.push_back(BB.getTerminator()->getOperand(0));
1197 while (Worklist.size()) {
1198 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1199 Worklist.pop_back();
1200 // Look at instructions inside this block.
1202 if (I->getParent() != &BB) continue;
1204 // Stop when reaching PHI nodes.
1205 if (isa<PHINode>(I)) {
1206 assert(I == Induction && "Found a uniform PHI that is not the induction");
1210 // This is a known uniform.
1213 // Insert all operands.
1214 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1215 Worklist.push_back(I->getOperand(i));
1222 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1223 typedef SmallVector<Value*, 16> ValueVector;
1224 typedef SmallPtrSet<Value*, 16> ValueSet;
1225 // Holds the Load and Store *instructions*.
1229 // Scan the BB and collect legal loads and stores.
1230 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1231 Instruction *I = it;
1233 // If this is a load, save it. If this instruction can read from memory
1234 // but is not a load, then we quit. Notice that we don't handle function
1235 // calls that read or write.
1236 if (I->mayReadFromMemory()) {
1237 LoadInst *Ld = dyn_cast<LoadInst>(I);
1238 if (!Ld) return false;
1239 if (!Ld->isSimple()) {
1240 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1243 Loads.push_back(Ld);
1247 // Save store instructions. Abort if other instructions write to memory.
1248 if (I->mayWriteToMemory()) {
1249 StoreInst *St = dyn_cast<StoreInst>(I);
1250 if (!St) return false;
1251 if (!St->isSimple()) {
1252 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1255 Stores.push_back(St);
1259 // Now we have two lists that hold the loads and the stores.
1260 // Next, we find the pointers that they use.
1262 // Check if we see any stores. If there are no stores, then we don't
1263 // care if the pointers are *restrict*.
1264 if (!Stores.size()) {
1265 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1269 // Holds the read and read-write *pointers* that we find.
1271 ValueVector ReadWrites;
1273 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1274 // multiple times on the same object. If the ptr is accessed twice, once
1275 // for read and once for write, it will only appear once (on the write
1276 // list). This is okay, since we are going to check for conflicts between
1277 // writes and between reads and writes, but not between reads and reads.
1280 ValueVector::iterator I, IE;
1281 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1282 StoreInst *ST = dyn_cast<StoreInst>(*I);
1283 assert(ST && "Bad StoreInst");
1284 Value* Ptr = ST->getPointerOperand();
1285 // If we did *not* see this pointer before, insert it to
1286 // the read-write list. At this phase it is only a 'write' list.
1287 if (Seen.insert(Ptr))
1288 ReadWrites.push_back(Ptr);
1291 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1292 LoadInst *LD = dyn_cast<LoadInst>(*I);
1293 assert(LD && "Bad LoadInst");
1294 Value* Ptr = LD->getPointerOperand();
1295 // If we did *not* see this pointer before, insert it to the
1296 // read list. If we *did* see it before, then it is already in
1297 // the read-write list. This allows us to vectorize expressions
1298 // such as A[i] += x; Because the address of A[i] is a read-write
1299 // pointer. This only works if the index of A[i] is consecutive.
1300 // If the address of i is unknown (for example A[B[i]]) then we may
1301 // read a few words, modify, and write a few words, and some of the
1302 // words may be written to the same address.
1303 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1304 Reads.push_back(Ptr);
1307 // Now that the pointers are in two lists (Reads and ReadWrites), we
1308 // can check that there are no conflicts between each of the writes and
1309 // between the writes to the reads.
1310 ValueSet WriteObjects;
1311 ValueVector TempObjects;
1313 // Check that the read-writes do not conflict with other read-write
1315 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1316 GetUnderlyingObjects(*I, TempObjects, DL);
1317 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1319 if (!isIdentifiedObject(*it)) {
1320 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1323 if (!WriteObjects.insert(*it)) {
1324 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1329 TempObjects.clear();
1332 /// Check that the reads don't conflict with the read-writes.
1333 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1334 GetUnderlyingObjects(*I, TempObjects, DL);
1335 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1337 if (!isIdentifiedObject(*it)) {
1338 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1341 if (WriteObjects.count(*it)) {
1342 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1347 TempObjects.clear();
1354 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1355 ReductionKind Kind) {
1356 if (Phi->getNumIncomingValues() != 2)
1359 // Find the possible incoming reduction variable.
1360 BasicBlock *BB = Phi->getParent();
1361 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1362 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1363 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1365 // ExitInstruction is the single value which is used outside the loop.
1366 // We only allow for a single reduction value to be used outside the loop.
1367 // This includes users of the reduction, variables (which form a cycle
1368 // which ends in the phi node).
1369 Instruction *ExitInstruction = 0;
1371 // Iter is our iterator. We start with the PHI node and scan for all of the
1372 // users of this instruction. All users must be instructions which can be
1373 // used as reduction variables (such as ADD). We may have a single
1374 // out-of-block user. They cycle must end with the original PHI.
1375 // Also, we can't have multiple block-local users.
1376 Instruction *Iter = Phi;
1378 // Any reduction instr must be of one of the allowed kinds.
1379 if (!isReductionInstr(Iter, Kind))
1382 // Did we found a user inside this block ?
1383 bool FoundInBlockUser = false;
1384 // Did we reach the initial PHI node ?
1385 bool FoundStartPHI = false;
1387 // If the instruction has no users then this is a broken
1388 // chain and can't be a reduction variable.
1389 if (Iter->use_empty())
1392 // For each of the *users* of iter.
1393 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1395 Instruction *U = cast<Instruction>(*it);
1396 // We already know that the PHI is a user.
1398 FoundStartPHI = true;
1401 // Check if we found the exit user.
1402 BasicBlock *Parent = U->getParent();
1404 // We must have a single exit instruction.
1405 if (ExitInstruction != 0)
1407 ExitInstruction = Iter;
1409 // We can't have multiple inside users.
1410 if (FoundInBlockUser)
1412 FoundInBlockUser = true;
1416 // We found a reduction var if we have reached the original
1417 // phi node and we only have a single instruction with out-of-loop
1419 if (FoundStartPHI && ExitInstruction) {
1420 // This instruction is allowed to have out-of-loop users.
1421 AllowedExit.insert(ExitInstruction);
1423 // Save the description of this reduction variable.
1424 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1425 Reductions[Phi] = RD;
1432 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1433 ReductionKind Kind) {
1434 switch (I->getOpcode()) {
1437 case Instruction::PHI:
1440 case Instruction::Add:
1441 case Instruction::Sub:
1442 return Kind == IntegerAdd;
1443 case Instruction::Mul:
1444 case Instruction::UDiv:
1445 case Instruction::SDiv:
1446 return Kind == IntegerMult;
1447 case Instruction::And:
1448 return Kind == IntegerAnd;
1449 case Instruction::Or:
1450 return Kind == IntegerOr;
1451 case Instruction::Xor:
1452 return Kind == IntegerXor;
1456 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1457 // Check that the PHI is consecutive and starts at zero.
1458 const SCEV *PhiScev = SE->getSCEV(Phi);
1459 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1461 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1464 const SCEV *Step = AR->getStepRecurrence(*SE);
1465 const SCEV *Start = AR->getStart();
1467 if (!Step->isOne() || !Start->isZero()) {
1468 DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
1475 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1477 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1481 float Cost = expectedCost(1);
1483 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1484 for (unsigned i=2; i <= VF; i*=2) {
1485 // Notice that the vector loop needs to be executed less times, so
1486 // we need to divide the cost of the vector loops by the width of
1487 // the vector elements.
1488 float VectorCost = expectedCost(i) / (float)i;
1489 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1490 (int)VectorCost << ".\n");
1491 if (VectorCost < Cost) {
1497 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1501 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1502 // We can only estimate the cost of single basic block loops.
1503 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1505 BasicBlock *BB = TheLoop->getHeader();
1508 // For each instruction in the old loop.
1509 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1510 Instruction *Inst = it;
1511 unsigned C = getInstructionCost(Inst, VF);
1513 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1514 " For instruction: "<< *Inst << "\n");
1521 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1522 assert(VTTI && "Invalid vector target transformation info");
1524 // If we know that this instruction will remain uniform, check the cost of
1525 // the scalar version.
1526 if (Legal->isUniformAfterVectorization(I))
1529 Type *RetTy = I->getType();
1530 Type *VectorTy = ToVectorTy(RetTy, VF);
1533 // TODO: We need to estimate the cost of intrinsic calls.
1534 switch (I->getOpcode()) {
1535 case Instruction::GetElementPtr:
1536 // We mark this instruction as zero-cost because scalar GEPs are usually
1537 // lowered to the intruction addressing mode. At the moment we don't
1538 // generate vector geps.
1540 case Instruction::Br: {
1541 return VTTI->getCFInstrCost(I->getOpcode());
1543 case Instruction::PHI:
1545 case Instruction::Add:
1546 case Instruction::FAdd:
1547 case Instruction::Sub:
1548 case Instruction::FSub:
1549 case Instruction::Mul:
1550 case Instruction::FMul:
1551 case Instruction::UDiv:
1552 case Instruction::SDiv:
1553 case Instruction::FDiv:
1554 case Instruction::URem:
1555 case Instruction::SRem:
1556 case Instruction::FRem:
1557 case Instruction::Shl:
1558 case Instruction::LShr:
1559 case Instruction::AShr:
1560 case Instruction::And:
1561 case Instruction::Or:
1562 case Instruction::Xor: {
1563 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1565 case Instruction::Select: {
1566 SelectInst *SI = cast<SelectInst>(I);
1567 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1568 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1569 Type *CondTy = SI->getCondition()->getType();
1571 CondTy = VectorType::get(CondTy, VF);
1573 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
1575 case Instruction::ICmp:
1576 case Instruction::FCmp: {
1577 Type *ValTy = I->getOperand(0)->getType();
1578 VectorTy = ToVectorTy(ValTy, VF);
1579 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
1581 case Instruction::Store: {
1582 StoreInst *SI = cast<StoreInst>(I);
1583 Type *ValTy = SI->getValueOperand()->getType();
1584 VectorTy = ToVectorTy(ValTy, VF);
1587 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
1588 SI->getAlignment(), SI->getPointerAddressSpace());
1590 // Scalarized stores.
1591 if (!Legal->isConsecutiveGep(SI->getPointerOperand())) {
1593 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1595 // The cost of extracting from the value vector.
1596 Cost += VF * (ExtCost);
1597 // The cost of the scalar stores.
1598 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1599 ValTy->getScalarType(),
1601 SI->getPointerAddressSpace());
1606 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
1607 SI->getPointerAddressSpace());
1609 case Instruction::Load: {
1610 LoadInst *LI = cast<LoadInst>(I);
1613 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
1615 LI->getPointerAddressSpace());
1617 // Scalarized loads.
1618 if (!Legal->isConsecutiveGep(LI->getPointerOperand())) {
1620 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
1621 // The cost of inserting the loaded value into the result vector.
1622 Cost += VF * (InCost);
1623 // The cost of the scalar stores.
1624 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1625 RetTy->getScalarType(),
1627 LI->getPointerAddressSpace());
1632 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
1633 LI->getPointerAddressSpace());
1635 case Instruction::ZExt:
1636 case Instruction::SExt:
1637 case Instruction::FPToUI:
1638 case Instruction::FPToSI:
1639 case Instruction::FPExt:
1640 case Instruction::PtrToInt:
1641 case Instruction::IntToPtr:
1642 case Instruction::SIToFP:
1643 case Instruction::UIToFP:
1644 case Instruction::Trunc:
1645 case Instruction::FPTrunc:
1646 case Instruction::BitCast: {
1647 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
1648 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
1651 // We are scalarizing the instruction. Return the cost of the scalar
1652 // instruction, plus the cost of insert and extract into vector
1653 // elements, times the vector width.
1656 bool IsVoid = RetTy->isVoidTy();
1658 unsigned InsCost = (IsVoid ? 0 :
1659 VTTI->getInstrCost(Instruction::InsertElement,
1662 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1665 // The cost of inserting the results plus extracting each one of the
1667 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
1669 // The cost of executing VF copies of the scalar instruction.
1670 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
1676 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
1677 if (Scalar->isVoidTy() || VF == 1)
1679 return VectorType::get(Scalar, VF);
1684 char LoopVectorize::ID = 0;
1685 static const char lv_name[] = "Loop Vectorization";
1686 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1687 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1688 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1689 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1690 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1693 Pass *createLoopVectorizePass() {
1694 return new LoopVectorize();