1 //===--- Scalarizer.cpp - Scalarize vector operations ---------------------===//
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 pass converts vector operations into scalar operations, in order
11 // to expose optimization opportunities on the individual scalar operations.
12 // It is mainly intended for targets that do not have vector units, but it
13 // may also be useful for revectorizing code to different vector widths.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/Analysis/VectorUtils.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/InstVisitor.h"
22 #include "llvm/Pass.h"
23 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
27 #define DEBUG_TYPE "scalarizer"
30 // Used to store the scattered form of a vector.
31 typedef SmallVector<Value *, 8> ValueVector;
33 // Used to map a vector Value to its scattered form. We use std::map
34 // because we want iterators to persist across insertion and because the
35 // values are relatively large.
36 typedef std::map<Value *, ValueVector> ScatterMap;
38 // Lists Instructions that have been replaced with scalar implementations,
39 // along with a pointer to their scattered forms.
40 typedef SmallVector<std::pair<Instruction *, ValueVector *>, 16> GatherList;
42 // Provides a very limited vector-like interface for lazily accessing one
43 // component of a scattered vector or vector pointer.
48 // Scatter V into Size components. If new instructions are needed,
49 // insert them before BBI in BB. If Cache is nonnull, use it to cache
51 Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
52 ValueVector *cachePtr = nullptr);
54 // Return component I, creating a new Value for it if necessary.
55 Value *operator[](unsigned I);
57 // Return the number of components.
58 unsigned size() const { return Size; }
62 BasicBlock::iterator BBI;
64 ValueVector *CachePtr;
70 // FCmpSpliiter(FCI)(Builder, X, Y, Name) uses Builder to create an FCmp
71 // called Name that compares X and Y in the same way as FCI.
73 FCmpSplitter(FCmpInst &fci) : FCI(fci) {}
74 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
75 const Twine &Name) const {
76 return Builder.CreateFCmp(FCI.getPredicate(), Op0, Op1, Name);
81 // ICmpSpliiter(ICI)(Builder, X, Y, Name) uses Builder to create an ICmp
82 // called Name that compares X and Y in the same way as ICI.
84 ICmpSplitter(ICmpInst &ici) : ICI(ici) {}
85 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
86 const Twine &Name) const {
87 return Builder.CreateICmp(ICI.getPredicate(), Op0, Op1, Name);
92 // BinarySpliiter(BO)(Builder, X, Y, Name) uses Builder to create
93 // a binary operator like BO called Name with operands X and Y.
94 struct BinarySplitter {
95 BinarySplitter(BinaryOperator &bo) : BO(bo) {}
96 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
97 const Twine &Name) const {
98 return Builder.CreateBinOp(BO.getOpcode(), Op0, Op1, Name);
103 // Information about a load or store that we're scalarizing.
104 struct VectorLayout {
105 VectorLayout() : VecTy(nullptr), ElemTy(nullptr), VecAlign(0), ElemSize(0) {}
107 // Return the alignment of element I.
108 uint64_t getElemAlign(unsigned I) {
109 return MinAlign(VecAlign, I * ElemSize);
112 // The type of the vector.
115 // The type of each element.
118 // The alignment of the vector.
121 // The size of each element.
125 class Scalarizer : public FunctionPass,
126 public InstVisitor<Scalarizer, bool> {
132 initializeScalarizerPass(*PassRegistry::getPassRegistry());
135 bool doInitialization(Module &M) override;
136 bool runOnFunction(Function &F) override;
138 // InstVisitor methods. They return true if the instruction was scalarized,
139 // false if nothing changed.
140 bool visitInstruction(Instruction &) { return false; }
141 bool visitSelectInst(SelectInst &SI);
142 bool visitICmpInst(ICmpInst &);
143 bool visitFCmpInst(FCmpInst &);
144 bool visitBinaryOperator(BinaryOperator &);
145 bool visitGetElementPtrInst(GetElementPtrInst &);
146 bool visitCastInst(CastInst &);
147 bool visitBitCastInst(BitCastInst &);
148 bool visitShuffleVectorInst(ShuffleVectorInst &);
149 bool visitPHINode(PHINode &);
150 bool visitLoadInst(LoadInst &);
151 bool visitStoreInst(StoreInst &);
152 bool visitCallInst(CallInst &I);
154 static void registerOptions() {
155 // This is disabled by default because having separate loads and stores
156 // makes it more likely that the -combiner-alias-analysis limits will be
158 OptionRegistry::registerOption<bool, Scalarizer,
159 &Scalarizer::ScalarizeLoadStore>(
160 "scalarize-load-store",
161 "Allow the scalarizer pass to scalarize loads and store", false);
165 Scatterer scatter(Instruction *, Value *);
166 void gather(Instruction *, const ValueVector &);
167 bool canTransferMetadata(unsigned Kind);
168 void transferMetadata(Instruction *, const ValueVector &);
169 bool getVectorLayout(Type *, unsigned, VectorLayout &, const DataLayout &);
172 template<typename T> bool splitBinary(Instruction &, const T &);
174 bool splitCall(CallInst &CI);
176 ScatterMap Scattered;
178 unsigned ParallelLoopAccessMDKind;
179 bool ScalarizeLoadStore;
182 char Scalarizer::ID = 0;
183 } // end anonymous namespace
185 INITIALIZE_PASS_WITH_OPTIONS(Scalarizer, "scalarizer",
186 "Scalarize vector operations", false, false)
188 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
189 ValueVector *cachePtr)
190 : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
191 Type *Ty = V->getType();
192 PtrTy = dyn_cast<PointerType>(Ty);
194 Ty = PtrTy->getElementType();
195 Size = Ty->getVectorNumElements();
197 Tmp.resize(Size, nullptr);
198 else if (CachePtr->empty())
199 CachePtr->resize(Size, nullptr);
201 assert(Size == CachePtr->size() && "Inconsistent vector sizes");
204 // Return component I, creating a new Value for it if necessary.
205 Value *Scatterer::operator[](unsigned I) {
206 ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
207 // Try to reuse a previous value.
210 IRBuilder<> Builder(BB, BBI);
214 PointerType::get(PtrTy->getElementType()->getVectorElementType(),
215 PtrTy->getAddressSpace());
216 CV[0] = Builder.CreateBitCast(V, Ty, V->getName() + ".i0");
219 CV[I] = Builder.CreateConstGEP1_32(nullptr, CV[0], I,
220 V->getName() + ".i" + Twine(I));
222 // Search through a chain of InsertElementInsts looking for element I.
223 // Record other elements in the cache. The new V is still suitable
224 // for all uncached indices.
226 InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
229 ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
232 unsigned J = Idx->getZExtValue();
233 V = Insert->getOperand(0);
235 CV[J] = Insert->getOperand(1);
238 // Only cache the first entry we find for each index we're not actively
239 // searching for. This prevents us from going too far up the chain and
240 // caching incorrect entries.
241 CV[J] = Insert->getOperand(1);
244 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
245 V->getName() + ".i" + Twine(I));
250 bool Scalarizer::doInitialization(Module &M) {
251 ParallelLoopAccessMDKind =
252 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
254 M.getContext().getOption<bool, Scalarizer, &Scalarizer::ScalarizeLoadStore>();
258 bool Scalarizer::runOnFunction(Function &F) {
261 assert(Gathered.empty() && Scattered.empty());
262 for (BasicBlock &BB : F) {
263 for (BasicBlock::iterator II = BB.begin(), IE = BB.end(); II != IE;) {
264 Instruction *I = &*II;
265 bool Done = visit(I);
267 if (Done && I->getType()->isVoidTy())
268 I->eraseFromParent();
274 // Return a scattered form of V that can be accessed by Point. V must be a
275 // vector or a pointer to a vector.
276 Scatterer Scalarizer::scatter(Instruction *Point, Value *V) {
277 if (Argument *VArg = dyn_cast<Argument>(V)) {
278 // Put the scattered form of arguments in the entry block,
279 // so that it can be used everywhere.
280 Function *F = VArg->getParent();
281 BasicBlock *BB = &F->getEntryBlock();
282 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
284 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
285 // Put the scattered form of an instruction directly after the
287 BasicBlock *BB = VOp->getParent();
288 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
291 // In the fallback case, just put the scattered before Point and
292 // keep the result local to Point.
293 return Scatterer(Point->getParent(), Point->getIterator(), V);
296 // Replace Op with the gathered form of the components in CV. Defer the
297 // deletion of Op and creation of the gathered form to the end of the pass,
298 // so that we can avoid creating the gathered form if all uses of Op are
299 // replaced with uses of CV.
300 void Scalarizer::gather(Instruction *Op, const ValueVector &CV) {
301 // Since we're not deleting Op yet, stub out its operands, so that it
302 // doesn't make anything live unnecessarily.
303 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
304 Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
306 transferMetadata(Op, CV);
308 // If we already have a scattered form of Op (created from ExtractElements
309 // of Op itself), replace them with the new form.
310 ValueVector &SV = Scattered[Op];
312 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
317 Instruction *Old = cast<Instruction>(V);
318 CV[I]->takeName(Old);
319 Old->replaceAllUsesWith(CV[I]);
320 Old->eraseFromParent();
324 Gathered.push_back(GatherList::value_type(Op, &SV));
327 // Return true if it is safe to transfer the given metadata tag from
328 // vector to scalar instructions.
329 bool Scalarizer::canTransferMetadata(unsigned Tag) {
330 return (Tag == LLVMContext::MD_tbaa
331 || Tag == LLVMContext::MD_fpmath
332 || Tag == LLVMContext::MD_tbaa_struct
333 || Tag == LLVMContext::MD_invariant_load
334 || Tag == LLVMContext::MD_alias_scope
335 || Tag == LLVMContext::MD_noalias
336 || Tag == ParallelLoopAccessMDKind);
339 // Transfer metadata from Op to the instructions in CV if it is known
340 // to be safe to do so.
341 void Scalarizer::transferMetadata(Instruction *Op, const ValueVector &CV) {
342 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
343 Op->getAllMetadataOtherThanDebugLoc(MDs);
344 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
345 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
346 for (const auto &MD : MDs)
347 if (canTransferMetadata(MD.first))
348 New->setMetadata(MD.first, MD.second);
349 if (Op->getDebugLoc() && !New->getDebugLoc())
350 New->setDebugLoc(Op->getDebugLoc());
355 // Try to fill in Layout from Ty, returning true on success. Alignment is
356 // the alignment of the vector, or 0 if the ABI default should be used.
357 bool Scalarizer::getVectorLayout(Type *Ty, unsigned Alignment,
358 VectorLayout &Layout, const DataLayout &DL) {
359 // Make sure we're dealing with a vector.
360 Layout.VecTy = dyn_cast<VectorType>(Ty);
364 // Check that we're dealing with full-byte elements.
365 Layout.ElemTy = Layout.VecTy->getElementType();
366 if (DL.getTypeSizeInBits(Layout.ElemTy) !=
367 DL.getTypeStoreSizeInBits(Layout.ElemTy))
371 Layout.VecAlign = Alignment;
373 Layout.VecAlign = DL.getABITypeAlignment(Layout.VecTy);
374 Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
378 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
379 // to create an instruction like I with operands X and Y and name Name.
380 template<typename Splitter>
381 bool Scalarizer::splitBinary(Instruction &I, const Splitter &Split) {
382 VectorType *VT = dyn_cast<VectorType>(I.getType());
386 unsigned NumElems = VT->getNumElements();
387 IRBuilder<> Builder(&I);
388 Scatterer Op0 = scatter(&I, I.getOperand(0));
389 Scatterer Op1 = scatter(&I, I.getOperand(1));
390 assert(Op0.size() == NumElems && "Mismatched binary operation");
391 assert(Op1.size() == NumElems && "Mismatched binary operation");
393 Res.resize(NumElems);
394 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
395 Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
396 I.getName() + ".i" + Twine(Elem));
401 static bool isTriviallyScalariable(Intrinsic::ID ID) {
402 return isTriviallyVectorizable(ID);
405 // All of the current scalarizable intrinsics only have one mangled type.
406 static Function *getScalarIntrinsicDeclaration(Module *M,
409 return Intrinsic::getDeclaration(M, ID, { Ty->getScalarType() });
412 /// If a call to a vector typed intrinsic function, split into a scalar call per
413 /// element if possible for the intrinsic.
414 bool Scalarizer::splitCall(CallInst &CI) {
415 VectorType *VT = dyn_cast<VectorType>(CI.getType());
419 Function *F = CI.getCalledFunction();
423 Intrinsic::ID ID = F->getIntrinsicID();
424 if (ID == Intrinsic::not_intrinsic || !isTriviallyScalariable(ID))
427 unsigned NumElems = VT->getNumElements();
428 unsigned NumArgs = CI.getNumArgOperands();
430 ValueVector ScalarOperands(NumArgs);
431 SmallVector<Scatterer, 8> Scattered(NumArgs);
433 Scattered.resize(NumArgs);
435 // Assumes that any vector type has the same number of elements as the return
436 // vector type, which is true for all current intrinsics.
437 for (unsigned I = 0; I != NumArgs; ++I) {
438 Value *OpI = CI.getOperand(I);
439 if (OpI->getType()->isVectorTy()) {
440 Scattered[I] = scatter(&CI, OpI);
441 assert(Scattered[I].size() == NumElems && "mismatched call operands");
443 ScalarOperands[I] = OpI;
447 ValueVector Res(NumElems);
448 ValueVector ScalarCallOps(NumArgs);
450 Function *NewIntrin = getScalarIntrinsicDeclaration(F->getParent(), ID, VT);
451 IRBuilder<> Builder(&CI);
453 // Perform actual scalarization, taking care to preserve any scalar operands.
454 for (unsigned Elem = 0; Elem < NumElems; ++Elem) {
455 ScalarCallOps.clear();
457 for (unsigned J = 0; J != NumArgs; ++J) {
458 if (hasVectorInstrinsicScalarOpd(ID, J))
459 ScalarCallOps.push_back(ScalarOperands[J]);
461 ScalarCallOps.push_back(Scattered[J][Elem]);
464 Res[Elem] = Builder.CreateCall(NewIntrin, ScalarCallOps,
465 CI.getName() + ".i" + Twine(Elem));
472 bool Scalarizer::visitSelectInst(SelectInst &SI) {
473 VectorType *VT = dyn_cast<VectorType>(SI.getType());
477 unsigned NumElems = VT->getNumElements();
478 IRBuilder<> Builder(&SI);
479 Scatterer Op1 = scatter(&SI, SI.getOperand(1));
480 Scatterer Op2 = scatter(&SI, SI.getOperand(2));
481 assert(Op1.size() == NumElems && "Mismatched select");
482 assert(Op2.size() == NumElems && "Mismatched select");
484 Res.resize(NumElems);
486 if (SI.getOperand(0)->getType()->isVectorTy()) {
487 Scatterer Op0 = scatter(&SI, SI.getOperand(0));
488 assert(Op0.size() == NumElems && "Mismatched select");
489 for (unsigned I = 0; I < NumElems; ++I)
490 Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
491 SI.getName() + ".i" + Twine(I));
493 Value *Op0 = SI.getOperand(0);
494 for (unsigned I = 0; I < NumElems; ++I)
495 Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
496 SI.getName() + ".i" + Twine(I));
502 bool Scalarizer::visitICmpInst(ICmpInst &ICI) {
503 return splitBinary(ICI, ICmpSplitter(ICI));
506 bool Scalarizer::visitFCmpInst(FCmpInst &FCI) {
507 return splitBinary(FCI, FCmpSplitter(FCI));
510 bool Scalarizer::visitBinaryOperator(BinaryOperator &BO) {
511 return splitBinary(BO, BinarySplitter(BO));
514 bool Scalarizer::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
515 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
519 IRBuilder<> Builder(&GEPI);
520 unsigned NumElems = VT->getNumElements();
521 unsigned NumIndices = GEPI.getNumIndices();
523 // The base pointer might be scalar even if it's a vector GEP. In those cases,
524 // splat the pointer into a vector value, and scatter that vector.
525 Value *Op0 = GEPI.getOperand(0);
526 if (!Op0->getType()->isVectorTy())
527 Op0 = Builder.CreateVectorSplat(NumElems, Op0);
528 Scatterer Base = scatter(&GEPI, Op0);
530 SmallVector<Scatterer, 8> Ops;
531 Ops.resize(NumIndices);
532 for (unsigned I = 0; I < NumIndices; ++I) {
533 Value *Op = GEPI.getOperand(I + 1);
535 // The indices might be scalars even if it's a vector GEP. In those cases,
536 // splat the scalar into a vector value, and scatter that vector.
537 if (!Op->getType()->isVectorTy())
538 Op = Builder.CreateVectorSplat(NumElems, Op);
540 Ops[I] = scatter(&GEPI, Op);
544 Res.resize(NumElems);
545 for (unsigned I = 0; I < NumElems; ++I) {
546 SmallVector<Value *, 8> Indices;
547 Indices.resize(NumIndices);
548 for (unsigned J = 0; J < NumIndices; ++J)
549 Indices[J] = Ops[J][I];
550 Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
551 GEPI.getName() + ".i" + Twine(I));
552 if (GEPI.isInBounds())
553 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
554 NewGEPI->setIsInBounds();
560 bool Scalarizer::visitCastInst(CastInst &CI) {
561 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
565 unsigned NumElems = VT->getNumElements();
566 IRBuilder<> Builder(&CI);
567 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
568 assert(Op0.size() == NumElems && "Mismatched cast");
570 Res.resize(NumElems);
571 for (unsigned I = 0; I < NumElems; ++I)
572 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
573 CI.getName() + ".i" + Twine(I));
578 bool Scalarizer::visitBitCastInst(BitCastInst &BCI) {
579 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
580 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
581 if (!DstVT || !SrcVT)
584 unsigned DstNumElems = DstVT->getNumElements();
585 unsigned SrcNumElems = SrcVT->getNumElements();
586 IRBuilder<> Builder(&BCI);
587 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
589 Res.resize(DstNumElems);
591 if (DstNumElems == SrcNumElems) {
592 for (unsigned I = 0; I < DstNumElems; ++I)
593 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
594 BCI.getName() + ".i" + Twine(I));
595 } else if (DstNumElems > SrcNumElems) {
596 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
597 // individual elements to the destination.
598 unsigned FanOut = DstNumElems / SrcNumElems;
599 Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
601 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
602 Value *V = Op0[Op0I];
604 // Look through any existing bitcasts before converting to <N x t2>.
605 // In the best case, the resulting conversion might be a no-op.
606 while ((VI = dyn_cast<Instruction>(V)) &&
607 VI->getOpcode() == Instruction::BitCast)
608 V = VI->getOperand(0);
609 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
610 Scatterer Mid = scatter(&BCI, V);
611 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
612 Res[ResI++] = Mid[MidI];
615 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
616 unsigned FanIn = SrcNumElems / DstNumElems;
617 Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
619 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
620 Value *V = UndefValue::get(MidTy);
621 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
622 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
623 BCI.getName() + ".i" + Twine(ResI)
624 + ".upto" + Twine(MidI));
625 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
626 BCI.getName() + ".i" + Twine(ResI));
633 bool Scalarizer::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
634 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
638 unsigned NumElems = VT->getNumElements();
639 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
640 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
642 Res.resize(NumElems);
644 for (unsigned I = 0; I < NumElems; ++I) {
645 int Selector = SVI.getMaskValue(I);
647 Res[I] = UndefValue::get(VT->getElementType());
648 else if (unsigned(Selector) < Op0.size())
649 Res[I] = Op0[Selector];
651 Res[I] = Op1[Selector - Op0.size()];
657 bool Scalarizer::visitPHINode(PHINode &PHI) {
658 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
662 unsigned NumElems = VT->getNumElements();
663 IRBuilder<> Builder(&PHI);
665 Res.resize(NumElems);
667 unsigned NumOps = PHI.getNumOperands();
668 for (unsigned I = 0; I < NumElems; ++I)
669 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
670 PHI.getName() + ".i" + Twine(I));
672 for (unsigned I = 0; I < NumOps; ++I) {
673 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
674 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
675 for (unsigned J = 0; J < NumElems; ++J)
676 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
682 bool Scalarizer::visitLoadInst(LoadInst &LI) {
683 if (!ScalarizeLoadStore)
689 if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout,
690 LI.getModule()->getDataLayout()))
693 unsigned NumElems = Layout.VecTy->getNumElements();
694 IRBuilder<> Builder(&LI);
695 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
697 Res.resize(NumElems);
699 for (unsigned I = 0; I < NumElems; ++I)
700 Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
701 LI.getName() + ".i" + Twine(I));
706 bool Scalarizer::visitStoreInst(StoreInst &SI) {
707 if (!ScalarizeLoadStore)
713 Value *FullValue = SI.getValueOperand();
714 if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout,
715 SI.getModule()->getDataLayout()))
718 unsigned NumElems = Layout.VecTy->getNumElements();
719 IRBuilder<> Builder(&SI);
720 Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
721 Scatterer Val = scatter(&SI, FullValue);
724 Stores.resize(NumElems);
725 for (unsigned I = 0; I < NumElems; ++I) {
726 unsigned Align = Layout.getElemAlign(I);
727 Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
729 transferMetadata(&SI, Stores);
733 bool Scalarizer::visitCallInst(CallInst &CI) {
734 return splitCall(CI);
737 // Delete the instructions that we scalarized. If a full vector result
738 // is still needed, recreate it using InsertElements.
739 bool Scalarizer::finish() {
740 // The presence of data in Gathered or Scattered indicates changes
741 // made to the Function.
742 if (Gathered.empty() && Scattered.empty())
744 for (const auto &GMI : Gathered) {
745 Instruction *Op = GMI.first;
746 ValueVector &CV = *GMI.second;
747 if (!Op->use_empty()) {
748 // The value is still needed, so recreate it using a series of
750 Type *Ty = Op->getType();
751 Value *Res = UndefValue::get(Ty);
752 BasicBlock *BB = Op->getParent();
753 unsigned Count = Ty->getVectorNumElements();
754 IRBuilder<> Builder(Op);
755 if (isa<PHINode>(Op))
756 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
757 for (unsigned I = 0; I < Count; ++I)
758 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
759 Op->getName() + ".upto" + Twine(I));
761 Op->replaceAllUsesWith(Res);
763 Op->eraseFromParent();
770 FunctionPass *llvm::createScalarizerPass() {
771 return new Scalarizer();