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 //===----------------------------------------------------------------------===//
9 #include "LoopVectorize.h"
10 #include "llvm/ADT/StringExtras.h"
11 #include "llvm/Analysis/AliasAnalysis.h"
12 #include "llvm/Analysis/AliasSetTracker.h"
13 #include "llvm/Analysis/Dominators.h"
14 #include "llvm/Analysis/LoopInfo.h"
15 #include "llvm/Analysis/LoopIterator.h"
16 #include "llvm/Analysis/LoopPass.h"
17 #include "llvm/Analysis/ScalarEvolutionExpander.h"
18 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/Analysis/Verifier.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DataLayout.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/Module.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/TargetTransformInfo.h"
34 #include "llvm/Transforms/Scalar.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Vectorize.h"
38 #include "llvm/Type.h"
39 #include "llvm/Value.h"
41 static cl::opt<unsigned>
42 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
43 cl::desc("Sets the SIMD width. Zero is autoselect."));
46 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
47 cl::desc("Enable if-conversion during vectorization."));
51 /// The LoopVectorize Pass.
52 struct LoopVectorize : public LoopPass {
53 /// Pass identification, replacement for typeid
56 explicit LoopVectorize() : LoopPass(ID) {
57 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
63 TargetTransformInfo *TTI;
66 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
67 // We only vectorize innermost loops.
71 SE = &getAnalysis<ScalarEvolution>();
72 DL = getAnalysisIfAvailable<DataLayout>();
73 LI = &getAnalysis<LoopInfo>();
74 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
75 DT = &getAnalysis<DominatorTree>();
77 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
78 L->getHeader()->getParent()->getName() << "\"\n");
80 // Check if it is legal to vectorize the loop.
81 LoopVectorizationLegality LVL(L, SE, DL, DT);
82 if (!LVL.canVectorize()) {
83 DEBUG(dbgs() << "LV: Not vectorizing.\n");
87 // Select the preffered vectorization factor.
88 const VectorTargetTransformInfo *VTTI = 0;
90 VTTI = TTI->getVectorTargetTransformInfo();
91 // Use the cost model.
92 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
94 // Check the function attribues to find out if this function should be
95 // optimized for size.
96 Function *F = L->getHeader()->getParent();
97 Attribute::AttrKind SzAttr= Attribute::OptimizeForSize;
98 bool OptForSize = F->getFnAttributes().hasAttribute(SzAttr);
100 unsigned VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
103 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
107 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
108 F->getParent()->getModuleIdentifier()<<"\n");
110 // If we decided that it is *legal* to vectorizer the loop then do it.
111 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF);
114 DEBUG(verifyFunction(*L->getHeader()->getParent()));
118 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
119 LoopPass::getAnalysisUsage(AU);
120 AU.addRequiredID(LoopSimplifyID);
121 AU.addRequiredID(LCSSAID);
122 AU.addRequired<LoopInfo>();
123 AU.addRequired<ScalarEvolution>();
124 AU.addRequired<DominatorTree>();
125 AU.addPreserved<LoopInfo>();
126 AU.addPreserved<DominatorTree>();
133 //===----------------------------------------------------------------------===//
134 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
135 // LoopVectorizationCostModel.
136 //===----------------------------------------------------------------------===//
139 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
140 Loop *Lp, Value *Ptr) {
141 const SCEV *Sc = SE->getSCEV(Ptr);
142 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
143 assert(AR && "Invalid addrec expression");
144 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
145 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
146 Pointers.push_back(Ptr);
147 Starts.push_back(AR->getStart());
148 Ends.push_back(ScEnd);
151 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
153 LLVMContext &C = V->getContext();
154 Type *VTy = VectorType::get(V->getType(), VF);
155 Type *I32 = IntegerType::getInt32Ty(C);
157 // Save the current insertion location.
158 Instruction *Loc = Builder.GetInsertPoint();
160 // We need to place the broadcast of invariant variables outside the loop.
161 Instruction *Instr = dyn_cast<Instruction>(V);
162 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
163 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
165 // Place the code for broadcasting invariant variables in the new preheader.
167 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
169 Constant *Zero = ConstantInt::get(I32, 0);
170 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
171 Value *UndefVal = UndefValue::get(VTy);
172 // Insert the value into a new vector.
173 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
174 // Broadcast the scalar into all locations in the vector.
175 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
178 // Restore the builder insertion point.
180 Builder.SetInsertPoint(Loc);
185 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) {
186 assert(Val->getType()->isVectorTy() && "Must be a vector");
187 assert(Val->getType()->getScalarType()->isIntegerTy() &&
188 "Elem must be an integer");
190 Type *ITy = Val->getType()->getScalarType();
191 VectorType *Ty = cast<VectorType>(Val->getType());
192 int VLen = Ty->getNumElements();
193 SmallVector<Constant*, 8> Indices;
195 // Create a vector of consecutive numbers from zero to VF.
196 for (int i = 0; i < VLen; ++i)
197 Indices.push_back(ConstantInt::get(ITy, Negate ? (-i): i ));
199 // Add the consecutive indices to the vector value.
200 Constant *Cv = ConstantVector::get(Indices);
201 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
202 return Builder.CreateAdd(Val, Cv, "induction");
205 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
206 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
208 // If this value is a pointer induction variable we know it is consecutive.
209 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
210 if (Phi && Inductions.count(Phi)) {
211 InductionInfo II = Inductions[Phi];
212 if (PtrInduction == II.IK)
216 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
220 unsigned NumOperands = Gep->getNumOperands();
221 Value *LastIndex = Gep->getOperand(NumOperands - 1);
223 // Check that all of the gep indices are uniform except for the last.
224 for (unsigned i = 0; i < NumOperands - 1; ++i)
225 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
228 // We can emit wide load/stores only if the last index is the induction
230 const SCEV *Last = SE->getSCEV(LastIndex);
231 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
232 const SCEV *Step = AR->getStepRecurrence(*SE);
234 // The memory is consecutive because the last index is consecutive
235 // and all other indices are loop invariant.
243 bool LoopVectorizationLegality::isUniform(Value *V) {
244 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
247 Value *InnerLoopVectorizer::getVectorValue(Value *V) {
248 assert(V != Induction && "The new induction variable should not be used.");
249 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
250 // If we saved a vectorized copy of V, use it.
251 Value *&MapEntry = WidenMap[V];
255 // Broadcast V and save the value for future uses.
256 Value *B = getBroadcastInstrs(V);
262 InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
263 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
266 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
267 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
268 // Holds vector parameters or scalars, in case of uniform vals.
269 SmallVector<Value*, 8> Params;
271 // Find all of the vectorized parameters.
272 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
273 Value *SrcOp = Instr->getOperand(op);
275 // If we are accessing the old induction variable, use the new one.
276 if (SrcOp == OldInduction) {
277 Params.push_back(getVectorValue(SrcOp));
281 // Try using previously calculated values.
282 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
284 // If the src is an instruction that appeared earlier in the basic block
285 // then it should already be vectorized.
286 if (SrcInst && OrigLoop->contains(SrcInst)) {
287 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
288 // The parameter is a vector value from earlier.
289 Params.push_back(WidenMap[SrcInst]);
291 // The parameter is a scalar from outside the loop. Maybe even a constant.
292 Params.push_back(SrcOp);
296 assert(Params.size() == Instr->getNumOperands() &&
297 "Invalid number of operands");
299 // Does this instruction return a value ?
300 bool IsVoidRetTy = Instr->getType()->isVoidTy();
301 Value *VecResults = 0;
303 // If we have a return value, create an empty vector. We place the scalarized
304 // instructions in this vector.
306 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
308 // For each scalar that we create:
309 for (unsigned i = 0; i < VF; ++i) {
310 Instruction *Cloned = Instr->clone();
312 Cloned->setName(Instr->getName() + ".cloned");
313 // Replace the operands of the cloned instrucions with extracted scalars.
314 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
315 Value *Op = Params[op];
316 // Param is a vector. Need to extract the right lane.
317 if (Op->getType()->isVectorTy())
318 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
319 Cloned->setOperand(op, Op);
322 // Place the cloned scalar in the new loop.
323 Builder.Insert(Cloned);
325 // If the original scalar returns a value we need to place it in a vector
326 // so that future users will be able to use it.
328 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
329 Builder.getInt32(i));
333 WidenMap[Instr] = VecResults;
337 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
339 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
340 Legal->getRuntimePointerCheck();
342 if (!PtrRtCheck->Need)
345 Value *MemoryRuntimeCheck = 0;
346 unsigned NumPointers = PtrRtCheck->Pointers.size();
347 SmallVector<Value* , 2> Starts;
348 SmallVector<Value* , 2> Ends;
350 SCEVExpander Exp(*SE, "induction");
352 // Use this type for pointer arithmetic.
353 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
355 for (unsigned i = 0; i < NumPointers; ++i) {
356 Value *Ptr = PtrRtCheck->Pointers[i];
357 const SCEV *Sc = SE->getSCEV(Ptr);
359 if (SE->isLoopInvariant(Sc, OrigLoop)) {
360 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
362 Starts.push_back(Ptr);
365 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
367 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
368 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
369 Starts.push_back(Start);
374 for (unsigned i = 0; i < NumPointers; ++i) {
375 for (unsigned j = i+1; j < NumPointers; ++j) {
376 Instruction::CastOps Op = Instruction::BitCast;
377 Value *Start0 = CastInst::Create(Op, Starts[i], PtrArithTy, "bc", Loc);
378 Value *Start1 = CastInst::Create(Op, Starts[j], PtrArithTy, "bc", Loc);
379 Value *End0 = CastInst::Create(Op, Ends[i], PtrArithTy, "bc", Loc);
380 Value *End1 = CastInst::Create(Op, Ends[j], PtrArithTy, "bc", Loc);
382 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
383 Start0, End1, "bound0", Loc);
384 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
385 Start1, End0, "bound1", Loc);
386 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
387 "found.conflict", Loc);
388 if (MemoryRuntimeCheck)
389 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
392 "conflict.rdx", Loc);
394 MemoryRuntimeCheck = IsConflict;
399 return MemoryRuntimeCheck;
403 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
405 In this function we generate a new loop. The new loop will contain
406 the vectorized instructions while the old loop will continue to run the
409 [ ] <-- vector loop bypass.
412 | [ ] <-- vector pre header.
416 | [ ]_| <-- vector loop.
419 >[ ] <--- middle-block.
422 | [ ] <--- new preheader.
426 | [ ]_| <-- old scalar loop to handle remainder.
433 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
434 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
435 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
436 assert(ExitBlock && "Must have an exit block");
438 // Some loops have a single integer induction variable, while other loops
439 // don't. One example is c++ iterators that often have multiple pointer
440 // induction variables. In the code below we also support a case where we
441 // don't have a single induction variable.
442 OldInduction = Legal->getInduction();
443 Type *IdxTy = OldInduction ? OldInduction->getType() :
444 DL->getIntPtrType(SE->getContext());
446 // Find the loop boundaries.
447 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
448 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
450 // Get the total trip count from the count by adding 1.
451 ExitCount = SE->getAddExpr(ExitCount,
452 SE->getConstant(ExitCount->getType(), 1));
454 // Expand the trip count and place the new instructions in the preheader.
455 // Notice that the pre-header does not change, only the loop body.
456 SCEVExpander Exp(*SE, "induction");
458 // Count holds the overall loop count (N).
459 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
460 BypassBlock->getTerminator());
462 // The loop index does not have to start at Zero. Find the original start
463 // value from the induction PHI node. If we don't have an induction variable
464 // then we know that it starts at zero.
465 Value *StartIdx = OldInduction ?
466 OldInduction->getIncomingValueForBlock(BypassBlock):
467 ConstantInt::get(IdxTy, 0);
469 assert(BypassBlock && "Invalid loop structure");
471 // Generate the code that checks in runtime if arrays overlap.
472 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal,
473 BypassBlock->getTerminator());
475 // Split the single block loop into the two loop structure described above.
476 BasicBlock *VectorPH =
477 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
478 BasicBlock *VecBody =
479 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
480 BasicBlock *MiddleBlock =
481 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
482 BasicBlock *ScalarPH =
483 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
485 // This is the location in which we add all of the logic for bypassing
486 // the new vector loop.
487 Instruction *Loc = BypassBlock->getTerminator();
489 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
491 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
493 // Generate the induction variable.
494 Induction = Builder.CreatePHI(IdxTy, 2, "index");
495 Constant *Step = ConstantInt::get(IdxTy, VF);
497 // We may need to extend the index in case there is a type mismatch.
498 // We know that the count starts at zero and does not overflow.
499 if (Count->getType() != IdxTy) {
500 // The exit count can be of pointer type. Convert it to the correct
502 if (ExitCount->getType()->isPointerTy())
503 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
505 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
508 // Add the start index to the loop count to get the new end index.
509 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
511 // Now we need to generate the expression for N - (N % VF), which is
512 // the part that the vectorized body will execute.
513 Constant *CIVF = ConstantInt::get(IdxTy, VF);
514 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
515 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
516 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
517 "end.idx.rnd.down", Loc);
519 // Now, compare the new count to zero. If it is zero skip the vector loop and
520 // jump to the scalar loop.
521 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
526 // If we are using memory runtime checks, include them in.
527 if (MemoryRuntimeCheck)
528 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
531 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
532 // Remove the old terminator.
533 Loc->eraseFromParent();
535 // We are going to resume the execution of the scalar loop.
536 // Go over all of the induction variables that we found and fix the
537 // PHIs that are left in the scalar version of the loop.
538 // The starting values of PHI nodes depend on the counter of the last
539 // iteration in the vectorized loop.
540 // If we come from a bypass edge then we need to start from the original
543 // This variable saves the new starting index for the scalar loop.
544 PHINode *ResumeIndex = 0;
545 LoopVectorizationLegality::InductionList::iterator I, E;
546 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
547 for (I = List->begin(), E = List->end(); I != E; ++I) {
548 PHINode *OrigPhi = I->first;
549 LoopVectorizationLegality::InductionInfo II = I->second;
550 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
551 MiddleBlock->getTerminator());
554 case LoopVectorizationLegality::NoInduction:
555 llvm_unreachable("Unknown induction");
556 case LoopVectorizationLegality::IntInduction: {
557 // Handle the integer induction counter:
558 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
559 assert(OrigPhi == OldInduction && "Unknown integer PHI");
560 // We know what the end value is.
561 EndValue = IdxEndRoundDown;
562 // We also know which PHI node holds it.
563 ResumeIndex = ResumeVal;
566 case LoopVectorizationLegality::ReverseIntInduction: {
567 // Convert the CountRoundDown variable to the PHI size.
568 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
569 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
570 Value *CRD = CountRoundDown;
571 if (CRDSize > IISize)
572 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
573 II.StartValue->getType(),
574 "tr.crd", BypassBlock->getTerminator());
575 else if (CRDSize < IISize)
576 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
577 II.StartValue->getType(),
578 "sext.crd", BypassBlock->getTerminator());
579 // Handle reverse integer induction counter:
580 EndValue = BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
581 BypassBlock->getTerminator());
584 case LoopVectorizationLegality::PtrInduction: {
585 // For pointer induction variables, calculate the offset using
587 EndValue = GetElementPtrInst::Create(II.StartValue, CountRoundDown,
589 BypassBlock->getTerminator());
594 // The new PHI merges the original incoming value, in case of a bypass,
595 // or the value at the end of the vectorized loop.
596 ResumeVal->addIncoming(II.StartValue, BypassBlock);
597 ResumeVal->addIncoming(EndValue, VecBody);
599 // Fix the scalar body counter (PHI node).
600 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
601 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
604 // If we are generating a new induction variable then we also need to
605 // generate the code that calculates the exit value. This value is not
606 // simply the end of the counter because we may skip the vectorized body
607 // in case of a runtime check.
609 assert(!ResumeIndex && "Unexpected resume value found");
610 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
611 MiddleBlock->getTerminator());
612 ResumeIndex->addIncoming(StartIdx, BypassBlock);
613 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
616 // Make sure that we found the index where scalar loop needs to continue.
617 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
618 "Invalid resume Index");
620 // Add a check in the middle block to see if we have completed
621 // all of the iterations in the first vector loop.
622 // If (N - N%VF) == N, then we *don't* need to run the remainder.
623 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
624 ResumeIndex, "cmp.n",
625 MiddleBlock->getTerminator());
627 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
628 // Remove the old terminator.
629 MiddleBlock->getTerminator()->eraseFromParent();
631 // Create i+1 and fill the PHINode.
632 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
633 Induction->addIncoming(StartIdx, VectorPH);
634 Induction->addIncoming(NextIdx, VecBody);
635 // Create the compare.
636 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
637 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
639 // Now we have two terminators. Remove the old one from the block.
640 VecBody->getTerminator()->eraseFromParent();
642 // Get ready to start creating new instructions into the vectorized body.
643 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
645 // Create and register the new vector loop.
646 Loop* Lp = new Loop();
647 Loop *ParentLoop = OrigLoop->getParentLoop();
649 // Insert the new loop into the loop nest and register the new basic blocks.
651 ParentLoop->addChildLoop(Lp);
652 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
653 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
654 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
656 LI->addTopLevelLoop(Lp);
659 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
662 LoopVectorPreHeader = VectorPH;
663 LoopScalarPreHeader = ScalarPH;
664 LoopMiddleBlock = MiddleBlock;
665 LoopExitBlock = ExitBlock;
666 LoopVectorBody = VecBody;
667 LoopScalarBody = OldBasicBlock;
668 LoopBypassBlock = BypassBlock;
671 /// This function returns the identity element (or neutral element) for
674 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
676 case LoopVectorizationLegality::IntegerXor:
677 case LoopVectorizationLegality::IntegerAdd:
678 case LoopVectorizationLegality::IntegerOr:
679 // Adding, Xoring, Oring zero to a number does not change it.
681 case LoopVectorizationLegality::IntegerMult:
682 // Multiplying a number by 1 does not change it.
684 case LoopVectorizationLegality::IntegerAnd:
685 // AND-ing a number with an all-1 value does not change it.
688 llvm_unreachable("Unknown reduction kind");
693 isTriviallyVectorizableIntrinsic(Instruction *Inst) {
694 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
697 switch (II->getIntrinsicID()) {
698 case Intrinsic::sqrt:
702 case Intrinsic::exp2:
704 case Intrinsic::log10:
705 case Intrinsic::log2:
706 case Intrinsic::fabs:
707 case Intrinsic::floor:
708 case Intrinsic::ceil:
709 case Intrinsic::trunc:
710 case Intrinsic::rint:
711 case Intrinsic::nearbyint:
714 case Intrinsic::fmuladd:
723 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
724 //===------------------------------------------------===//
726 // Notice: any optimization or new instruction that go
727 // into the code below should be also be implemented in
730 //===------------------------------------------------===//
731 BasicBlock &BB = *OrigLoop->getHeader();
733 ConstantInt::get(IntegerType::getInt32Ty(BB.getContext()), 0);
735 // In order to support reduction variables we need to be able to vectorize
736 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
737 // stages. First, we create a new vector PHI node with no incoming edges.
738 // We use this value when we vectorize all of the instructions that use the
739 // PHI. Next, after all of the instructions in the block are complete we
740 // add the new incoming edges to the PHI. At this point all of the
741 // instructions in the basic block are vectorized, so we can use them to
742 // construct the PHI.
743 PhiVector RdxPHIsToFix;
745 // Scan the loop in a topological order to ensure that defs are vectorized
747 LoopBlocksDFS DFS(OrigLoop);
750 // Vectorize all of the blocks in the original loop.
751 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
752 be = DFS.endRPO(); bb != be; ++bb)
753 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
755 // At this point every instruction in the original loop is widened to
756 // a vector form. We are almost done. Now, we need to fix the PHI nodes
757 // that we vectorized. The PHI nodes are currently empty because we did
758 // not want to introduce cycles. Notice that the remaining PHI nodes
759 // that we need to fix are reduction variables.
761 // Create the 'reduced' values for each of the induction vars.
762 // The reduced values are the vector values that we scalarize and combine
763 // after the loop is finished.
764 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
766 PHINode *RdxPhi = *it;
767 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
768 assert(RdxPhi && "Unable to recover vectorized PHI");
770 // Find the reduction variable descriptor.
771 assert(Legal->getReductionVars()->count(RdxPhi) &&
772 "Unable to find the reduction variable");
773 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
774 (*Legal->getReductionVars())[RdxPhi];
776 // We need to generate a reduction vector from the incoming scalar.
777 // To do so, we need to generate the 'identity' vector and overide
778 // one of the elements with the incoming scalar reduction. We need
779 // to do it in the vector-loop preheader.
780 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
782 // This is the vector-clone of the value that leaves the loop.
783 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
784 Type *VecTy = VectorExit->getType();
786 // Find the reduction identity variable. Zero for addition, or, xor,
787 // one for multiplication, -1 for And.
788 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
789 VecTy->getScalarType());
791 // This vector is the Identity vector where the first element is the
792 // incoming scalar reduction.
793 Value *VectorStart = Builder.CreateInsertElement(Identity,
794 RdxDesc.StartValue, Zero);
796 // Fix the vector-loop phi.
797 // We created the induction variable so we know that the
798 // preheader is the first entry.
799 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
801 // Reductions do not have to start at zero. They can start with
802 // any loop invariant values.
803 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
805 getVectorValue(RdxPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()));
806 VecRdxPhi->addIncoming(Val, LoopVectorBody);
808 // Before each round, move the insertion point right between
809 // the PHIs and the values we are going to write.
810 // This allows us to write both PHINodes and the extractelement
812 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
814 // This PHINode contains the vectorized reduction variable, or
815 // the initial value vector, if we bypass the vector loop.
816 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
817 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
818 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
820 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
821 // and vector ops, reducing the set of values being computed by half each
823 assert(isPowerOf2_32(VF) &&
824 "Reduction emission only supported for pow2 vectors!");
825 Value *TmpVec = NewPhi;
826 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
827 for (unsigned i = VF; i != 1; i >>= 1) {
828 // Move the upper half of the vector to the lower half.
829 for (unsigned j = 0; j != i/2; ++j)
830 ShuffleMask[j] = Builder.getInt32(i/2 + j);
832 // Fill the rest of the mask with undef.
833 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
834 UndefValue::get(Builder.getInt32Ty()));
837 Builder.CreateShuffleVector(TmpVec,
838 UndefValue::get(TmpVec->getType()),
839 ConstantVector::get(ShuffleMask),
842 // Emit the operation on the shuffled value.
843 switch (RdxDesc.Kind) {
844 case LoopVectorizationLegality::IntegerAdd:
845 TmpVec = Builder.CreateAdd(TmpVec, Shuf, "add.rdx");
847 case LoopVectorizationLegality::IntegerMult:
848 TmpVec = Builder.CreateMul(TmpVec, Shuf, "mul.rdx");
850 case LoopVectorizationLegality::IntegerOr:
851 TmpVec = Builder.CreateOr(TmpVec, Shuf, "or.rdx");
853 case LoopVectorizationLegality::IntegerAnd:
854 TmpVec = Builder.CreateAnd(TmpVec, Shuf, "and.rdx");
856 case LoopVectorizationLegality::IntegerXor:
857 TmpVec = Builder.CreateXor(TmpVec, Shuf, "xor.rdx");
860 llvm_unreachable("Unknown reduction operation");
864 // The result is in the first element of the vector.
865 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
867 // Now, we need to fix the users of the reduction variable
868 // inside and outside of the scalar remainder loop.
869 // We know that the loop is in LCSSA form. We need to update the
870 // PHI nodes in the exit blocks.
871 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
872 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
873 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
874 if (!LCSSAPhi) continue;
876 // All PHINodes need to have a single entry edge, or two if
877 // we already fixed them.
878 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
880 // We found our reduction value exit-PHI. Update it with the
881 // incoming bypass edge.
882 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
883 // Add an edge coming from the bypass.
884 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
887 }// end of the LCSSA phi scan.
889 // Fix the scalar loop reduction variable with the incoming reduction sum
890 // from the vector body and from the backedge value.
891 int IncomingEdgeBlockIdx =
892 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
893 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
894 // Pick the other block.
895 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
896 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
897 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
898 }// end of for each redux variable.
901 Value *InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
902 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
905 Value *SrcMask = createBlockInMask(Src);
907 // The terminator has to be a branch inst!
908 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
909 assert(BI && "Unexpected terminator found");
911 Value *EdgeMask = SrcMask;
912 if (BI->isConditional()) {
913 EdgeMask = getVectorValue(BI->getCondition());
914 if (BI->getSuccessor(0) != Dst)
915 EdgeMask = Builder.CreateNot(EdgeMask);
918 return Builder.CreateAnd(EdgeMask, SrcMask);
921 Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
922 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
924 // Loop incoming mask is all-one.
925 if (OrigLoop->getHeader() == BB) {
926 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
927 return getVectorValue(C);
930 // This is the block mask. We OR all incoming edges, and with zero.
931 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
932 Value *BlockMask = getVectorValue(Zero);
935 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it)
936 BlockMask = Builder.CreateOr(BlockMask, createEdgeMask(*it, BB));
942 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
943 BasicBlock *BB, PhiVector *PV) {
945 ConstantInt::get(IntegerType::getInt32Ty(BB->getContext()), 0);
947 // For each instruction in the old loop.
948 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
949 switch (it->getOpcode()) {
950 case Instruction::Br:
951 // Nothing to do for PHIs and BR, since we already took care of the
952 // loop control flow instructions.
954 case Instruction::PHI:{
955 PHINode* P = cast<PHINode>(it);
956 // Handle reduction variables:
957 if (Legal->getReductionVars()->count(P)) {
958 // This is phase one of vectorizing PHIs.
959 Type *VecTy = VectorType::get(it->getType(), VF);
961 PHINode::Create(VecTy, 2, "vec.phi",
962 LoopVectorBody->getFirstInsertionPt());
967 // Check for PHI nodes that are lowered to vector selects.
968 if (P->getParent() != OrigLoop->getHeader()) {
969 // We know that all PHIs in non header blocks are converted into
970 // selects, so we don't have to worry about the insertion order and we
971 // can just use the builder.
973 // At this point we generate the predication tree. There may be
974 // duplications since this is a simple recursive scan, but future
975 // optimizations will clean it up.
976 Value *Cond = createEdgeMask(P->getIncomingBlock(0), P->getParent());
978 Builder.CreateSelect(Cond,
979 getVectorValue(P->getIncomingValue(0)),
980 getVectorValue(P->getIncomingValue(1)),
985 // This PHINode must be an induction variable.
986 // Make sure that we know about it.
987 assert(Legal->getInductionVars()->count(P) &&
988 "Not an induction variable");
990 LoopVectorizationLegality::InductionInfo II =
991 Legal->getInductionVars()->lookup(P);
994 case LoopVectorizationLegality::NoInduction:
995 llvm_unreachable("Unknown induction");
996 case LoopVectorizationLegality::IntInduction: {
997 assert(P == OldInduction && "Unexpected PHI");
998 Value *Broadcasted = getBroadcastInstrs(Induction);
999 // After broadcasting the induction variable we need to make the
1000 // vector consecutive by adding 0, 1, 2 ...
1001 Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted);
1002 WidenMap[OldInduction] = ConsecutiveInduction;
1005 case LoopVectorizationLegality::ReverseIntInduction:
1006 case LoopVectorizationLegality::PtrInduction:
1007 // Handle reverse integer and pointer inductions.
1008 Value *StartIdx = 0;
1009 // If we have a single integer induction variable then use it.
1010 // Otherwise, start counting at zero.
1012 LoopVectorizationLegality::InductionInfo OldII =
1013 Legal->getInductionVars()->lookup(OldInduction);
1014 StartIdx = OldII.StartValue;
1016 StartIdx = ConstantInt::get(Induction->getType(), 0);
1018 // This is the normalized GEP that starts counting at zero.
1019 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1022 // Handle the reverse integer induction variable case.
1023 if (LoopVectorizationLegality::ReverseIntInduction == II.IK) {
1024 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
1025 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
1027 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
1030 // This is a new value so do not hoist it out.
1031 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
1032 // After broadcasting the induction variable we need to make the
1033 // vector consecutive by adding ... -3, -2, -1, 0.
1034 Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted,
1036 WidenMap[it] = ConsecutiveInduction;
1040 // Handle the pointer induction variable case.
1041 assert(P->getType()->isPointerTy() && "Unexpected type.");
1043 // This is the vector of results. Notice that we don't generate
1044 // vector geps because scalar geps result in better code.
1045 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1046 for (unsigned int i = 0; i < VF; ++i) {
1047 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1048 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx,
1050 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
1052 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1053 Builder.getInt32(i),
1057 WidenMap[it] = VecVal;
1063 case Instruction::Add:
1064 case Instruction::FAdd:
1065 case Instruction::Sub:
1066 case Instruction::FSub:
1067 case Instruction::Mul:
1068 case Instruction::FMul:
1069 case Instruction::UDiv:
1070 case Instruction::SDiv:
1071 case Instruction::FDiv:
1072 case Instruction::URem:
1073 case Instruction::SRem:
1074 case Instruction::FRem:
1075 case Instruction::Shl:
1076 case Instruction::LShr:
1077 case Instruction::AShr:
1078 case Instruction::And:
1079 case Instruction::Or:
1080 case Instruction::Xor: {
1081 // Just widen binops.
1082 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
1083 Value *A = getVectorValue(it->getOperand(0));
1084 Value *B = getVectorValue(it->getOperand(1));
1086 // Use this vector value for all users of the original instruction.
1087 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1090 // Update the NSW, NUW and Exact flags.
1091 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1092 if (isa<OverflowingBinaryOperator>(BinOp)) {
1093 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1094 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1096 if (isa<PossiblyExactOperator>(VecOp))
1097 VecOp->setIsExact(BinOp->isExact());
1100 case Instruction::Select: {
1102 // If the selector is loop invariant we can create a select
1103 // instruction with a scalar condition. Otherwise, use vector-select.
1104 Value *Cond = it->getOperand(0);
1105 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1107 // The condition can be loop invariant but still defined inside the
1108 // loop. This means that we can't just use the original 'cond' value.
1109 // We have to take the 'vectorized' value and pick the first lane.
1110 // Instcombine will make this a no-op.
1111 Cond = getVectorValue(Cond);
1113 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1115 Value *Op0 = getVectorValue(it->getOperand(1));
1116 Value *Op1 = getVectorValue(it->getOperand(2));
1117 WidenMap[it] = Builder.CreateSelect(Cond, Op0, Op1);
1121 case Instruction::ICmp:
1122 case Instruction::FCmp: {
1123 // Widen compares. Generate vector compares.
1124 bool FCmp = (it->getOpcode() == Instruction::FCmp);
1125 CmpInst *Cmp = dyn_cast<CmpInst>(it);
1126 Value *A = getVectorValue(it->getOperand(0));
1127 Value *B = getVectorValue(it->getOperand(1));
1129 WidenMap[it] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1131 WidenMap[it] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1135 case Instruction::Store: {
1136 // Attempt to issue a wide store.
1137 StoreInst *SI = dyn_cast<StoreInst>(it);
1138 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1139 Value *Ptr = SI->getPointerOperand();
1140 unsigned Alignment = SI->getAlignment();
1142 assert(!Legal->isUniform(Ptr) &&
1143 "We do not allow storing to uniform addresses");
1145 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1147 // This store does not use GEPs.
1148 if (!Legal->isConsecutivePtr(Ptr)) {
1149 scalarizeInstruction(it);
1154 // The last index does not have to be the induction. It can be
1155 // consecutive and be a function of the index. For example A[I+1];
1156 unsigned NumOperands = Gep->getNumOperands();
1157 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1158 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1160 // Create the new GEP with the new induction variable.
1161 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1162 Gep2->setOperand(NumOperands - 1, LastIndex);
1163 Ptr = Builder.Insert(Gep2);
1165 // Use the induction element ptr.
1166 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1167 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1169 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1170 Value *Val = getVectorValue(SI->getValueOperand());
1171 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1174 case Instruction::Load: {
1175 // Attempt to issue a wide load.
1176 LoadInst *LI = dyn_cast<LoadInst>(it);
1177 Type *RetTy = VectorType::get(LI->getType(), VF);
1178 Value *Ptr = LI->getPointerOperand();
1179 unsigned Alignment = LI->getAlignment();
1180 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1182 // If the pointer is loop invariant or if it is non consecutive,
1183 // scalarize the load.
1184 bool Con = Legal->isConsecutivePtr(Ptr);
1185 if (Legal->isUniform(Ptr) || !Con) {
1186 scalarizeInstruction(it);
1191 // The last index does not have to be the induction. It can be
1192 // consecutive and be a function of the index. For example A[I+1];
1193 unsigned NumOperands = Gep->getNumOperands();
1194 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1195 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1197 // Create the new GEP with the new induction variable.
1198 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1199 Gep2->setOperand(NumOperands - 1, LastIndex);
1200 Ptr = Builder.Insert(Gep2);
1202 // Use the induction element ptr.
1203 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1204 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1207 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1208 LI = Builder.CreateLoad(Ptr);
1209 LI->setAlignment(Alignment);
1210 // Use this vector value for all users of the load.
1214 case Instruction::ZExt:
1215 case Instruction::SExt:
1216 case Instruction::FPToUI:
1217 case Instruction::FPToSI:
1218 case Instruction::FPExt:
1219 case Instruction::PtrToInt:
1220 case Instruction::IntToPtr:
1221 case Instruction::SIToFP:
1222 case Instruction::UIToFP:
1223 case Instruction::Trunc:
1224 case Instruction::FPTrunc:
1225 case Instruction::BitCast: {
1226 CastInst *CI = dyn_cast<CastInst>(it);
1227 /// Optimize the special case where the source is the induction
1228 /// variable. Notice that we can only optimize the 'trunc' case
1229 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
1230 /// c. other casts depend on pointer size.
1231 if (CI->getOperand(0) == OldInduction &&
1232 it->getOpcode() == Instruction::Trunc) {
1233 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
1235 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
1236 WidenMap[it] = getConsecutiveVector(Broadcasted);
1239 /// Vectorize casts.
1240 Value *A = getVectorValue(it->getOperand(0));
1241 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1242 WidenMap[it] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1246 case Instruction::Call: {
1247 assert(isTriviallyVectorizableIntrinsic(it));
1248 Module *M = BB->getParent()->getParent();
1249 IntrinsicInst *II = cast<IntrinsicInst>(it);
1250 Intrinsic::ID ID = II->getIntrinsicID();
1251 SmallVector<Value*, 4> Args;
1252 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
1253 Args.push_back(getVectorValue(II->getArgOperand(i)));
1254 Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
1255 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
1256 WidenMap[it] = Builder.CreateCall(F, Args);
1261 // All other instructions are unsupported. Scalarize them.
1262 scalarizeInstruction(it);
1265 }// end of for_each instr.
1268 void InnerLoopVectorizer::updateAnalysis() {
1269 // Forget the original basic block.
1270 SE->forgetLoop(OrigLoop);
1272 // Update the dominator tree information.
1273 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1274 "Entry does not dominate exit.");
1276 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1277 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1278 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1279 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1280 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1281 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1283 DEBUG(DT->verifyAnalysis());
1286 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
1287 if (!EnableIfConversion)
1290 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
1291 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
1293 // Collect the blocks that need predication.
1294 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
1295 BasicBlock *BB = LoopBlocks[i];
1297 // We don't support switch statements inside loops.
1298 if (!isa<BranchInst>(BB->getTerminator()))
1301 // We must have at most two predecessors because we need to convert
1302 // all PHIs to selects.
1303 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
1307 // We must be able to predicate all blocks that need to be predicated.
1308 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
1312 // We can if-convert this loop.
1316 bool LoopVectorizationLegality::canVectorize() {
1317 assert(TheLoop->getLoopPreheader() && "No preheader!!");
1319 // We can only vectorize innermost loops.
1320 if (TheLoop->getSubLoopsVector().size())
1323 // We must have a single backedge.
1324 if (TheLoop->getNumBackEdges() != 1)
1327 // We must have a single exiting block.
1328 if (!TheLoop->getExitingBlock())
1331 unsigned NumBlocks = TheLoop->getNumBlocks();
1333 // Check if we can if-convert non single-bb loops.
1334 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
1335 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
1339 // We need to have a loop header.
1340 BasicBlock *Latch = TheLoop->getLoopLatch();
1341 DEBUG(dbgs() << "LV: Found a loop: " <<
1342 TheLoop->getHeader()->getName() << "\n");
1344 // ScalarEvolution needs to be able to find the exit count.
1345 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
1346 if (ExitCount == SE->getCouldNotCompute()) {
1347 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1351 // Do not loop-vectorize loops with a tiny trip count.
1352 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
1353 if (TC > 0u && TC < TinyTripCountThreshold) {
1354 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1355 "This loop is not worth vectorizing.\n");
1359 // Check if we can vectorize the instructions and CFG in this loop.
1360 if (!canVectorizeInstrs()) {
1361 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
1365 // Go over each instruction and look at memory deps.
1366 if (!canVectorizeMemory()) {
1367 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
1371 // Collect all of the variables that remain uniform after vectorization.
1372 collectLoopUniforms();
1374 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1375 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1378 // Okay! We can vectorize. At this point we don't have any other mem analysis
1379 // which may limit our maximum vectorization factor, so just return true with
1384 bool LoopVectorizationLegality::canVectorizeInstrs() {
1385 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1386 BasicBlock *Header = TheLoop->getHeader();
1388 // For each block in the loop.
1389 for (Loop::block_iterator bb = TheLoop->block_begin(),
1390 be = TheLoop->block_end(); bb != be; ++bb) {
1392 // Scan the instructions in the block and look for hazards.
1393 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1396 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
1397 // This should not happen because the loop should be normalized.
1398 if (Phi->getNumIncomingValues() != 2) {
1399 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1403 // Check that this PHI type is allowed.
1404 if (!Phi->getType()->isIntegerTy() &&
1405 !Phi->getType()->isPointerTy()) {
1406 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1410 // If this PHINode is not in the header block, then we know that we
1411 // can convert it to select during if-conversion. No need to check if
1412 // the PHIs in this block are induction or reduction variables.
1416 // This is the value coming from the preheader.
1417 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1418 // Check if this is an induction variable.
1419 InductionKind IK = isInductionVariable(Phi);
1421 if (NoInduction != IK) {
1422 // Int inductions are special because we only allow one IV.
1423 if (IK == IntInduction) {
1425 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1431 DEBUG(dbgs() << "LV: Found an induction variable.\n");
1432 Inductions[Phi] = InductionInfo(StartValue, IK);
1436 if (AddReductionVar(Phi, IntegerAdd)) {
1437 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1440 if (AddReductionVar(Phi, IntegerMult)) {
1441 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1444 if (AddReductionVar(Phi, IntegerOr)) {
1445 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1448 if (AddReductionVar(Phi, IntegerAnd)) {
1449 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1452 if (AddReductionVar(Phi, IntegerXor)) {
1453 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1457 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1459 }// end of PHI handling
1461 // We still don't handle functions.
1462 CallInst *CI = dyn_cast<CallInst>(it);
1463 if (CI && !isTriviallyVectorizableIntrinsic(it)) {
1464 DEBUG(dbgs() << "LV: Found a call site.\n");
1468 // Check that the instruction return type is vectorizable.
1469 if (!VectorType::isValidElementType(it->getType()) &&
1470 !it->getType()->isVoidTy()) {
1471 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1475 // Check that the stored type is vectorizable.
1476 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
1477 Type *T = ST->getValueOperand()->getType();
1478 if (!VectorType::isValidElementType(T))
1482 // Reduction instructions are allowed to have exit users.
1483 // All other instructions must not have external users.
1484 if (!AllowedExit.count(it))
1485 //Check that all of the users of the loop are inside the BB.
1486 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
1488 Instruction *U = cast<Instruction>(*I);
1489 // This user may be a reduction exit value.
1490 if (!TheLoop->contains(U)) {
1491 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1500 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1501 assert(getInductionVars()->size() && "No induction variables");
1507 void LoopVectorizationLegality::collectLoopUniforms() {
1508 // We now know that the loop is vectorizable!
1509 // Collect variables that will remain uniform after vectorization.
1510 std::vector<Value*> Worklist;
1511 BasicBlock *Latch = TheLoop->getLoopLatch();
1513 // Start with the conditional branch and walk up the block.
1514 Worklist.push_back(Latch->getTerminator()->getOperand(0));
1516 while (Worklist.size()) {
1517 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1518 Worklist.pop_back();
1520 // Look at instructions inside this loop.
1521 // Stop when reaching PHI nodes.
1522 // TODO: we need to follow values all over the loop, not only in this block.
1523 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
1526 // This is a known uniform.
1529 // Insert all operands.
1530 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
1531 Worklist.push_back(I->getOperand(i));
1536 bool LoopVectorizationLegality::canVectorizeMemory() {
1537 typedef SmallVector<Value*, 16> ValueVector;
1538 typedef SmallPtrSet<Value*, 16> ValueSet;
1539 // Holds the Load and Store *instructions*.
1542 PtrRtCheck.Pointers.clear();
1543 PtrRtCheck.Need = false;
1546 for (Loop::block_iterator bb = TheLoop->block_begin(),
1547 be = TheLoop->block_end(); bb != be; ++bb) {
1549 // Scan the BB and collect legal loads and stores.
1550 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1553 // If this is a load, save it. If this instruction can read from memory
1554 // but is not a load, then we quit. Notice that we don't handle function
1555 // calls that read or write.
1556 if (it->mayReadFromMemory()) {
1557 LoadInst *Ld = dyn_cast<LoadInst>(it);
1558 if (!Ld) return false;
1559 if (!Ld->isSimple()) {
1560 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1563 Loads.push_back(Ld);
1567 // Save 'store' instructions. Abort if other instructions write to memory.
1568 if (it->mayWriteToMemory()) {
1569 StoreInst *St = dyn_cast<StoreInst>(it);
1570 if (!St) return false;
1571 if (!St->isSimple()) {
1572 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1575 Stores.push_back(St);
1580 // Now we have two lists that hold the loads and the stores.
1581 // Next, we find the pointers that they use.
1583 // Check if we see any stores. If there are no stores, then we don't
1584 // care if the pointers are *restrict*.
1585 if (!Stores.size()) {
1586 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1590 // Holds the read and read-write *pointers* that we find.
1592 ValueVector ReadWrites;
1594 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1595 // multiple times on the same object. If the ptr is accessed twice, once
1596 // for read and once for write, it will only appear once (on the write
1597 // list). This is okay, since we are going to check for conflicts between
1598 // writes and between reads and writes, but not between reads and reads.
1601 ValueVector::iterator I, IE;
1602 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1603 StoreInst *ST = cast<StoreInst>(*I);
1604 Value* Ptr = ST->getPointerOperand();
1606 if (isUniform(Ptr)) {
1607 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1611 // If we did *not* see this pointer before, insert it to
1612 // the read-write list. At this phase it is only a 'write' list.
1613 if (Seen.insert(Ptr))
1614 ReadWrites.push_back(Ptr);
1617 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1618 LoadInst *LD = cast<LoadInst>(*I);
1619 Value* Ptr = LD->getPointerOperand();
1620 // If we did *not* see this pointer before, insert it to the
1621 // read list. If we *did* see it before, then it is already in
1622 // the read-write list. This allows us to vectorize expressions
1623 // such as A[i] += x; Because the address of A[i] is a read-write
1624 // pointer. This only works if the index of A[i] is consecutive.
1625 // If the address of i is unknown (for example A[B[i]]) then we may
1626 // read a few words, modify, and write a few words, and some of the
1627 // words may be written to the same address.
1628 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1629 Reads.push_back(Ptr);
1632 // If we write (or read-write) to a single destination and there are no
1633 // other reads in this loop then is it safe to vectorize.
1634 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1635 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1639 // Find pointers with computable bounds. We are going to use this information
1640 // to place a runtime bound check.
1641 bool CanDoRT = true;
1642 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1643 if (hasComputableBounds(*I)) {
1644 PtrRtCheck.insert(SE, TheLoop, *I);
1645 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1650 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1651 if (hasComputableBounds(*I)) {
1652 PtrRtCheck.insert(SE, TheLoop, *I);
1653 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1659 // Check that we did not collect too many pointers or found a
1660 // unsizeable pointer.
1661 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1667 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1670 bool NeedRTCheck = false;
1672 // Now that the pointers are in two lists (Reads and ReadWrites), we
1673 // can check that there are no conflicts between each of the writes and
1674 // between the writes to the reads.
1675 ValueSet WriteObjects;
1676 ValueVector TempObjects;
1678 // Check that the read-writes do not conflict with other read-write
1680 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1681 GetUnderlyingObjects(*I, TempObjects, DL);
1682 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1684 if (!isIdentifiedObject(*it)) {
1685 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1688 if (!WriteObjects.insert(*it)) {
1689 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1694 TempObjects.clear();
1697 /// Check that the reads don't conflict with the read-writes.
1698 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1699 GetUnderlyingObjects(*I, TempObjects, DL);
1700 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1702 if (!isIdentifiedObject(*it)) {
1703 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1706 if (WriteObjects.count(*it)) {
1707 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1712 TempObjects.clear();
1715 PtrRtCheck.Need = NeedRTCheck;
1716 if (NeedRTCheck && !CanDoRT) {
1717 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
1718 "the array bounds.\n");
1723 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
1724 " need a runtime memory check.\n");
1728 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1729 ReductionKind Kind) {
1730 if (Phi->getNumIncomingValues() != 2)
1733 // Reduction variables are only found in the loop header block.
1734 if (Phi->getParent() != TheLoop->getHeader())
1737 // Obtain the reduction start value from the value that comes from the loop
1739 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
1741 // ExitInstruction is the single value which is used outside the loop.
1742 // We only allow for a single reduction value to be used outside the loop.
1743 // This includes users of the reduction, variables (which form a cycle
1744 // which ends in the phi node).
1745 Instruction *ExitInstruction = 0;
1747 // Iter is our iterator. We start with the PHI node and scan for all of the
1748 // users of this instruction. All users must be instructions that can be
1749 // used as reduction variables (such as ADD). We may have a single
1750 // out-of-block user. The cycle must end with the original PHI.
1751 Instruction *Iter = Phi;
1753 // If the instruction has no users then this is a broken
1754 // chain and can't be a reduction variable.
1755 if (Iter->use_empty())
1758 // Any reduction instr must be of one of the allowed kinds.
1759 if (!isReductionInstr(Iter, Kind))
1762 // Did we find a user inside this loop already ?
1763 bool FoundInBlockUser = false;
1764 // Did we reach the initial PHI node already ?
1765 bool FoundStartPHI = false;
1767 // For each of the *users* of iter.
1768 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1770 Instruction *U = cast<Instruction>(*it);
1771 // We already know that the PHI is a user.
1773 FoundStartPHI = true;
1777 // Check if we found the exit user.
1778 BasicBlock *Parent = U->getParent();
1779 if (!TheLoop->contains(Parent)) {
1780 // Exit if you find multiple outside users.
1781 if (ExitInstruction != 0)
1783 ExitInstruction = Iter;
1786 // We allow in-loop PHINodes which are not the original reduction PHI
1787 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
1788 // structure) then don't skip this PHI.
1789 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
1790 U->getParent() != TheLoop->getHeader() &&
1791 TheLoop->contains(U) &&
1792 Iter->getNumUses() > 1)
1795 // We can't have multiple inside users.
1796 if (FoundInBlockUser)
1798 FoundInBlockUser = true;
1802 // We found a reduction var if we have reached the original
1803 // phi node and we only have a single instruction with out-of-loop
1805 if (FoundStartPHI && ExitInstruction) {
1806 // This instruction is allowed to have out-of-loop users.
1807 AllowedExit.insert(ExitInstruction);
1809 // Save the description of this reduction variable.
1810 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1811 Reductions[Phi] = RD;
1815 // If we've reached the start PHI but did not find an outside user then
1816 // this is dead code. Abort.
1823 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1824 ReductionKind Kind) {
1825 switch (I->getOpcode()) {
1828 case Instruction::PHI:
1831 case Instruction::Add:
1832 case Instruction::Sub:
1833 return Kind == IntegerAdd;
1834 case Instruction::Mul:
1835 return Kind == IntegerMult;
1836 case Instruction::And:
1837 return Kind == IntegerAnd;
1838 case Instruction::Or:
1839 return Kind == IntegerOr;
1840 case Instruction::Xor:
1841 return Kind == IntegerXor;
1845 LoopVectorizationLegality::InductionKind
1846 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1847 Type *PhiTy = Phi->getType();
1848 // We only handle integer and pointer inductions variables.
1849 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1852 // Check that the PHI is consecutive and starts at zero.
1853 const SCEV *PhiScev = SE->getSCEV(Phi);
1854 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1856 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1859 const SCEV *Step = AR->getStepRecurrence(*SE);
1861 // Integer inductions need to have a stride of one.
1862 if (PhiTy->isIntegerTy()) {
1864 return IntInduction;
1865 if (Step->isAllOnesValue())
1866 return ReverseIntInduction;
1870 // Calculate the pointer stride and check if it is consecutive.
1871 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1875 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1876 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1877 if (C->getValue()->equalsInt(Size))
1878 return PtrInduction;
1883 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
1884 Value *In0 = const_cast<Value*>(V);
1885 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
1889 return Inductions.count(PN);
1892 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
1893 assert(TheLoop->contains(BB) && "Unknown block used");
1895 // Blocks that do not dominate the latch need predication.
1896 BasicBlock* Latch = TheLoop->getLoopLatch();
1897 return !DT->dominates(BB, Latch);
1900 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
1901 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1902 // We don't predicate loads/stores at the moment.
1903 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
1906 // The instructions below can trap.
1907 switch (it->getOpcode()) {
1909 case Instruction::UDiv:
1910 case Instruction::SDiv:
1911 case Instruction::URem:
1912 case Instruction::SRem:
1920 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1921 const SCEV *PhiScev = SE->getSCEV(Ptr);
1922 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1926 return AR->isAffine();
1930 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
1932 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
1933 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
1937 // Find the trip count.
1938 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
1939 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
1941 unsigned VF = MaxVectorSize;
1943 // If we optimize the program for size, avoid creating the tail loop.
1945 // If we are unable to calculate the trip count then don't try to vectorize.
1947 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
1951 // Find the maximum SIMD width that can fit within the trip count.
1952 VF = TC % MaxVectorSize;
1957 // If the trip count that we found modulo the vectorization factor is not
1958 // zero then we require a tail.
1960 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
1966 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
1967 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
1973 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1977 float Cost = expectedCost(1);
1979 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1980 for (unsigned i=2; i <= VF; i*=2) {
1981 // Notice that the vector loop needs to be executed less times, so
1982 // we need to divide the cost of the vector loops by the width of
1983 // the vector elements.
1984 float VectorCost = expectedCost(i) / (float)i;
1985 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1986 (int)VectorCost << ".\n");
1987 if (VectorCost < Cost) {
1993 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1997 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
2001 for (Loop::block_iterator bb = TheLoop->block_begin(),
2002 be = TheLoop->block_end(); bb != be; ++bb) {
2003 unsigned BlockCost = 0;
2004 BasicBlock *BB = *bb;
2006 // For each instruction in the old loop.
2007 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2008 unsigned C = getInstructionCost(it, VF);
2010 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
2011 VF << " For instruction: "<< *it << "\n");
2014 // We assume that if-converted blocks have a 50% chance of being executed.
2015 // When the code is scalar then some of the blocks are avoided due to CF.
2016 // When the code is vectorized we execute all code paths.
2017 if (Legal->blockNeedsPredication(*bb) && VF == 1)
2027 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
2028 assert(VTTI && "Invalid vector target transformation info");
2030 // If we know that this instruction will remain uniform, check the cost of
2031 // the scalar version.
2032 if (Legal->isUniformAfterVectorization(I))
2035 Type *RetTy = I->getType();
2036 Type *VectorTy = ToVectorTy(RetTy, VF);
2038 // TODO: We need to estimate the cost of intrinsic calls.
2039 switch (I->getOpcode()) {
2040 case Instruction::GetElementPtr:
2041 // We mark this instruction as zero-cost because scalar GEPs are usually
2042 // lowered to the intruction addressing mode. At the moment we don't
2043 // generate vector geps.
2045 case Instruction::Br: {
2046 return VTTI->getCFInstrCost(I->getOpcode());
2048 case Instruction::PHI:
2049 //TODO: IF-converted IFs become selects.
2051 case Instruction::Add:
2052 case Instruction::FAdd:
2053 case Instruction::Sub:
2054 case Instruction::FSub:
2055 case Instruction::Mul:
2056 case Instruction::FMul:
2057 case Instruction::UDiv:
2058 case Instruction::SDiv:
2059 case Instruction::FDiv:
2060 case Instruction::URem:
2061 case Instruction::SRem:
2062 case Instruction::FRem:
2063 case Instruction::Shl:
2064 case Instruction::LShr:
2065 case Instruction::AShr:
2066 case Instruction::And:
2067 case Instruction::Or:
2068 case Instruction::Xor:
2069 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
2070 case Instruction::Select: {
2071 SelectInst *SI = cast<SelectInst>(I);
2072 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
2073 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
2074 Type *CondTy = SI->getCondition()->getType();
2076 CondTy = VectorType::get(CondTy, VF);
2078 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2080 case Instruction::ICmp:
2081 case Instruction::FCmp: {
2082 Type *ValTy = I->getOperand(0)->getType();
2083 VectorTy = ToVectorTy(ValTy, VF);
2084 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
2086 case Instruction::Store: {
2087 StoreInst *SI = cast<StoreInst>(I);
2088 Type *ValTy = SI->getValueOperand()->getType();
2089 VectorTy = ToVectorTy(ValTy, VF);
2092 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy,
2094 SI->getPointerAddressSpace());
2096 // Scalarized stores.
2097 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
2100 // The cost of extracting from the value vector and pointer vector.
2101 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2102 for (unsigned i = 0; i < VF; ++i) {
2103 Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
2105 Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
2109 // The cost of the scalar stores.
2110 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2111 ValTy->getScalarType(),
2113 SI->getPointerAddressSpace());
2118 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2119 SI->getPointerAddressSpace());
2121 case Instruction::Load: {
2122 LoadInst *LI = cast<LoadInst>(I);
2125 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy,
2127 LI->getPointerAddressSpace());
2129 // Scalarized loads.
2130 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2132 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2134 // The cost of extracting from the pointer vector.
2135 for (unsigned i = 0; i < VF; ++i)
2136 Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
2139 // The cost of inserting data to the result vector.
2140 for (unsigned i = 0; i < VF; ++i)
2141 Cost += VTTI->getVectorInstrCost(Instruction::InsertElement,
2144 // The cost of the scalar stores.
2145 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2146 RetTy->getScalarType(),
2148 LI->getPointerAddressSpace());
2153 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2154 LI->getPointerAddressSpace());
2156 case Instruction::ZExt:
2157 case Instruction::SExt:
2158 case Instruction::FPToUI:
2159 case Instruction::FPToSI:
2160 case Instruction::FPExt:
2161 case Instruction::PtrToInt:
2162 case Instruction::IntToPtr:
2163 case Instruction::SIToFP:
2164 case Instruction::UIToFP:
2165 case Instruction::Trunc:
2166 case Instruction::FPTrunc:
2167 case Instruction::BitCast: {
2168 // We optimize the truncation of induction variable.
2169 // The cost of these is the same as the scalar operation.
2170 if (I->getOpcode() == Instruction::Trunc &&
2171 Legal->isInductionVariable(I->getOperand(0)))
2172 return VTTI->getCastInstrCost(I->getOpcode(), I->getType(),
2173 I->getOperand(0)->getType());
2175 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2176 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2178 case Instruction::Call: {
2179 assert(isTriviallyVectorizableIntrinsic(I));
2180 IntrinsicInst *II = cast<IntrinsicInst>(I);
2181 Type *RetTy = ToVectorTy(II->getType(), VF);
2182 SmallVector<Type*, 4> Tys;
2183 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
2184 Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF));
2185 return VTTI->getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys);
2188 // We are scalarizing the instruction. Return the cost of the scalar
2189 // instruction, plus the cost of insert and extract into vector
2190 // elements, times the vector width.
2193 if (!RetTy->isVoidTy() && VF != 1) {
2194 unsigned InsCost = VTTI->getVectorInstrCost(Instruction::InsertElement,
2196 unsigned ExtCost = VTTI->getVectorInstrCost(Instruction::ExtractElement,
2199 // The cost of inserting the results plus extracting each one of the
2201 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2204 // The cost of executing VF copies of the scalar instruction. This opcode
2205 // is unknown. Assume that it is the same as 'mul'.
2206 Cost += VF * VTTI->getArithmeticInstrCost(Instruction::Mul, VectorTy);
2212 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2213 if (Scalar->isVoidTy() || VF == 1)
2215 return VectorType::get(Scalar, VF);
2218 char LoopVectorize::ID = 0;
2219 static const char lv_name[] = "Loop Vectorization";
2220 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2221 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2222 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2223 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2224 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2227 Pass *createLoopVectorizePass() {
2228 return new LoopVectorize();