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/ADT/STLExtras.h"
18 #include "llvm/IR/IRBuilder.h"
19 #include "llvm/IR/InstVisitor.h"
20 #include "llvm/Pass.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Transforms/Scalar.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 &);
153 static void registerOptions() {
154 // This is disabled by default because having separate loads and stores
155 // makes it more likely that the -combiner-alias-analysis limits will be
157 OptionRegistry::registerOption<bool, Scalarizer,
158 &Scalarizer::ScalarizeLoadStore>(
159 "scalarize-load-store",
160 "Allow the scalarizer pass to scalarize loads and store", false);
164 Scatterer scatter(Instruction *, Value *);
165 void gather(Instruction *, const ValueVector &);
166 bool canTransferMetadata(unsigned Kind);
167 void transferMetadata(Instruction *, const ValueVector &);
168 bool getVectorLayout(Type *, unsigned, VectorLayout &, const DataLayout &);
171 template<typename T> bool splitBinary(Instruction &, const T &);
173 ScatterMap Scattered;
175 unsigned ParallelLoopAccessMDKind;
176 bool ScalarizeLoadStore;
179 char Scalarizer::ID = 0;
180 } // end anonymous namespace
182 INITIALIZE_PASS_WITH_OPTIONS(Scalarizer, "scalarizer",
183 "Scalarize vector operations", false, false)
185 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
186 ValueVector *cachePtr)
187 : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
188 Type *Ty = V->getType();
189 PtrTy = dyn_cast<PointerType>(Ty);
191 Ty = PtrTy->getElementType();
192 Size = Ty->getVectorNumElements();
194 Tmp.resize(Size, nullptr);
195 else if (CachePtr->empty())
196 CachePtr->resize(Size, nullptr);
198 assert(Size == CachePtr->size() && "Inconsistent vector sizes");
201 // Return component I, creating a new Value for it if necessary.
202 Value *Scatterer::operator[](unsigned I) {
203 ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
204 // Try to reuse a previous value.
207 IRBuilder<> Builder(BB, BBI);
211 PointerType::get(PtrTy->getElementType()->getVectorElementType(),
212 PtrTy->getAddressSpace());
213 CV[0] = Builder.CreateBitCast(V, Ty, V->getName() + ".i0");
216 CV[I] = Builder.CreateConstGEP1_32(nullptr, CV[0], I,
217 V->getName() + ".i" + Twine(I));
219 // Search through a chain of InsertElementInsts looking for element I.
220 // Record other elements in the cache. The new V is still suitable
221 // for all uncached indices.
223 InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
226 ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
229 unsigned J = Idx->getZExtValue();
230 V = Insert->getOperand(0);
232 CV[J] = Insert->getOperand(1);
235 // Only cache the first entry we find for each index we're not actively
236 // searching for. This prevents us from going too far up the chain and
237 // caching incorrect entries.
238 CV[J] = Insert->getOperand(1);
241 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
242 V->getName() + ".i" + Twine(I));
247 bool Scalarizer::doInitialization(Module &M) {
248 ParallelLoopAccessMDKind =
249 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
251 M.getContext().getOption<bool, Scalarizer, &Scalarizer::ScalarizeLoadStore>();
255 bool Scalarizer::runOnFunction(Function &F) {
256 assert(Gathered.empty() && Scattered.empty());
257 for (BasicBlock &BB : F) {
258 for (BasicBlock::iterator II = BB.begin(), IE = BB.end(); II != IE;) {
259 Instruction *I = &*II;
260 bool Done = visit(I);
262 if (Done && I->getType()->isVoidTy())
263 I->eraseFromParent();
269 // Return a scattered form of V that can be accessed by Point. V must be a
270 // vector or a pointer to a vector.
271 Scatterer Scalarizer::scatter(Instruction *Point, Value *V) {
272 if (Argument *VArg = dyn_cast<Argument>(V)) {
273 // Put the scattered form of arguments in the entry block,
274 // so that it can be used everywhere.
275 Function *F = VArg->getParent();
276 BasicBlock *BB = &F->getEntryBlock();
277 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
279 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
280 // Put the scattered form of an instruction directly after the
282 BasicBlock *BB = VOp->getParent();
283 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
286 // In the fallback case, just put the scattered before Point and
287 // keep the result local to Point.
288 return Scatterer(Point->getParent(), Point->getIterator(), V);
291 // Replace Op with the gathered form of the components in CV. Defer the
292 // deletion of Op and creation of the gathered form to the end of the pass,
293 // so that we can avoid creating the gathered form if all uses of Op are
294 // replaced with uses of CV.
295 void Scalarizer::gather(Instruction *Op, const ValueVector &CV) {
296 // Since we're not deleting Op yet, stub out its operands, so that it
297 // doesn't make anything live unnecessarily.
298 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
299 Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
301 transferMetadata(Op, CV);
303 // If we already have a scattered form of Op (created from ExtractElements
304 // of Op itself), replace them with the new form.
305 ValueVector &SV = Scattered[Op];
307 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
308 Instruction *Old = cast<Instruction>(SV[I]);
309 CV[I]->takeName(Old);
310 Old->replaceAllUsesWith(CV[I]);
311 Old->eraseFromParent();
315 Gathered.push_back(GatherList::value_type(Op, &SV));
318 // Return true if it is safe to transfer the given metadata tag from
319 // vector to scalar instructions.
320 bool Scalarizer::canTransferMetadata(unsigned Tag) {
321 return (Tag == LLVMContext::MD_tbaa
322 || Tag == LLVMContext::MD_fpmath
323 || Tag == LLVMContext::MD_tbaa_struct
324 || Tag == LLVMContext::MD_invariant_load
325 || Tag == LLVMContext::MD_alias_scope
326 || Tag == LLVMContext::MD_noalias
327 || Tag == ParallelLoopAccessMDKind);
330 // Transfer metadata from Op to the instructions in CV if it is known
331 // to be safe to do so.
332 void Scalarizer::transferMetadata(Instruction *Op, const ValueVector &CV) {
333 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
334 Op->getAllMetadataOtherThanDebugLoc(MDs);
335 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
336 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
337 for (SmallVectorImpl<std::pair<unsigned, MDNode *>>::iterator
341 if (canTransferMetadata(MI->first))
342 New->setMetadata(MI->first, MI->second);
343 New->setDebugLoc(Op->getDebugLoc());
348 // Try to fill in Layout from Ty, returning true on success. Alignment is
349 // the alignment of the vector, or 0 if the ABI default should be used.
350 bool Scalarizer::getVectorLayout(Type *Ty, unsigned Alignment,
351 VectorLayout &Layout, const DataLayout &DL) {
352 // Make sure we're dealing with a vector.
353 Layout.VecTy = dyn_cast<VectorType>(Ty);
357 // Check that we're dealing with full-byte elements.
358 Layout.ElemTy = Layout.VecTy->getElementType();
359 if (DL.getTypeSizeInBits(Layout.ElemTy) !=
360 DL.getTypeStoreSizeInBits(Layout.ElemTy))
364 Layout.VecAlign = Alignment;
366 Layout.VecAlign = DL.getABITypeAlignment(Layout.VecTy);
367 Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
371 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
372 // to create an instruction like I with operands X and Y and name Name.
373 template<typename Splitter>
374 bool Scalarizer::splitBinary(Instruction &I, const Splitter &Split) {
375 VectorType *VT = dyn_cast<VectorType>(I.getType());
379 unsigned NumElems = VT->getNumElements();
380 IRBuilder<> Builder(&I);
381 Scatterer Op0 = scatter(&I, I.getOperand(0));
382 Scatterer Op1 = scatter(&I, I.getOperand(1));
383 assert(Op0.size() == NumElems && "Mismatched binary operation");
384 assert(Op1.size() == NumElems && "Mismatched binary operation");
386 Res.resize(NumElems);
387 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
388 Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
389 I.getName() + ".i" + Twine(Elem));
394 bool Scalarizer::visitSelectInst(SelectInst &SI) {
395 VectorType *VT = dyn_cast<VectorType>(SI.getType());
399 unsigned NumElems = VT->getNumElements();
400 IRBuilder<> Builder(&SI);
401 Scatterer Op1 = scatter(&SI, SI.getOperand(1));
402 Scatterer Op2 = scatter(&SI, SI.getOperand(2));
403 assert(Op1.size() == NumElems && "Mismatched select");
404 assert(Op2.size() == NumElems && "Mismatched select");
406 Res.resize(NumElems);
408 if (SI.getOperand(0)->getType()->isVectorTy()) {
409 Scatterer Op0 = scatter(&SI, SI.getOperand(0));
410 assert(Op0.size() == NumElems && "Mismatched select");
411 for (unsigned I = 0; I < NumElems; ++I)
412 Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
413 SI.getName() + ".i" + Twine(I));
415 Value *Op0 = SI.getOperand(0);
416 for (unsigned I = 0; I < NumElems; ++I)
417 Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
418 SI.getName() + ".i" + Twine(I));
424 bool Scalarizer::visitICmpInst(ICmpInst &ICI) {
425 return splitBinary(ICI, ICmpSplitter(ICI));
428 bool Scalarizer::visitFCmpInst(FCmpInst &FCI) {
429 return splitBinary(FCI, FCmpSplitter(FCI));
432 bool Scalarizer::visitBinaryOperator(BinaryOperator &BO) {
433 return splitBinary(BO, BinarySplitter(BO));
436 bool Scalarizer::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
437 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
441 IRBuilder<> Builder(&GEPI);
442 unsigned NumElems = VT->getNumElements();
443 unsigned NumIndices = GEPI.getNumIndices();
445 Scatterer Base = scatter(&GEPI, GEPI.getOperand(0));
447 SmallVector<Scatterer, 8> Ops;
448 Ops.resize(NumIndices);
449 for (unsigned I = 0; I < NumIndices; ++I)
450 Ops[I] = scatter(&GEPI, GEPI.getOperand(I + 1));
453 Res.resize(NumElems);
454 for (unsigned I = 0; I < NumElems; ++I) {
455 SmallVector<Value *, 8> Indices;
456 Indices.resize(NumIndices);
457 for (unsigned J = 0; J < NumIndices; ++J)
458 Indices[J] = Ops[J][I];
459 Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
460 GEPI.getName() + ".i" + Twine(I));
461 if (GEPI.isInBounds())
462 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
463 NewGEPI->setIsInBounds();
469 bool Scalarizer::visitCastInst(CastInst &CI) {
470 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
474 unsigned NumElems = VT->getNumElements();
475 IRBuilder<> Builder(&CI);
476 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
477 assert(Op0.size() == NumElems && "Mismatched cast");
479 Res.resize(NumElems);
480 for (unsigned I = 0; I < NumElems; ++I)
481 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
482 CI.getName() + ".i" + Twine(I));
487 bool Scalarizer::visitBitCastInst(BitCastInst &BCI) {
488 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
489 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
490 if (!DstVT || !SrcVT)
493 unsigned DstNumElems = DstVT->getNumElements();
494 unsigned SrcNumElems = SrcVT->getNumElements();
495 IRBuilder<> Builder(&BCI);
496 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
498 Res.resize(DstNumElems);
500 if (DstNumElems == SrcNumElems) {
501 for (unsigned I = 0; I < DstNumElems; ++I)
502 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
503 BCI.getName() + ".i" + Twine(I));
504 } else if (DstNumElems > SrcNumElems) {
505 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
506 // individual elements to the destination.
507 unsigned FanOut = DstNumElems / SrcNumElems;
508 Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
510 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
511 Value *V = Op0[Op0I];
513 // Look through any existing bitcasts before converting to <N x t2>.
514 // In the best case, the resulting conversion might be a no-op.
515 while ((VI = dyn_cast<Instruction>(V)) &&
516 VI->getOpcode() == Instruction::BitCast)
517 V = VI->getOperand(0);
518 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
519 Scatterer Mid = scatter(&BCI, V);
520 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
521 Res[ResI++] = Mid[MidI];
524 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
525 unsigned FanIn = SrcNumElems / DstNumElems;
526 Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
528 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
529 Value *V = UndefValue::get(MidTy);
530 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
531 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
532 BCI.getName() + ".i" + Twine(ResI)
533 + ".upto" + Twine(MidI));
534 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
535 BCI.getName() + ".i" + Twine(ResI));
542 bool Scalarizer::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
543 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
547 unsigned NumElems = VT->getNumElements();
548 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
549 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
551 Res.resize(NumElems);
553 for (unsigned I = 0; I < NumElems; ++I) {
554 int Selector = SVI.getMaskValue(I);
556 Res[I] = UndefValue::get(VT->getElementType());
557 else if (unsigned(Selector) < Op0.size())
558 Res[I] = Op0[Selector];
560 Res[I] = Op1[Selector - Op0.size()];
566 bool Scalarizer::visitPHINode(PHINode &PHI) {
567 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
571 unsigned NumElems = VT->getNumElements();
572 IRBuilder<> Builder(&PHI);
574 Res.resize(NumElems);
576 unsigned NumOps = PHI.getNumOperands();
577 for (unsigned I = 0; I < NumElems; ++I)
578 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
579 PHI.getName() + ".i" + Twine(I));
581 for (unsigned I = 0; I < NumOps; ++I) {
582 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
583 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
584 for (unsigned J = 0; J < NumElems; ++J)
585 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
591 bool Scalarizer::visitLoadInst(LoadInst &LI) {
592 if (!ScalarizeLoadStore)
598 if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout,
599 LI.getModule()->getDataLayout()))
602 unsigned NumElems = Layout.VecTy->getNumElements();
603 IRBuilder<> Builder(&LI);
604 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
606 Res.resize(NumElems);
608 for (unsigned I = 0; I < NumElems; ++I)
609 Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
610 LI.getName() + ".i" + Twine(I));
615 bool Scalarizer::visitStoreInst(StoreInst &SI) {
616 if (!ScalarizeLoadStore)
622 Value *FullValue = SI.getValueOperand();
623 if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout,
624 SI.getModule()->getDataLayout()))
627 unsigned NumElems = Layout.VecTy->getNumElements();
628 IRBuilder<> Builder(&SI);
629 Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
630 Scatterer Val = scatter(&SI, FullValue);
633 Stores.resize(NumElems);
634 for (unsigned I = 0; I < NumElems; ++I) {
635 unsigned Align = Layout.getElemAlign(I);
636 Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
638 transferMetadata(&SI, Stores);
642 // Delete the instructions that we scalarized. If a full vector result
643 // is still needed, recreate it using InsertElements.
644 bool Scalarizer::finish() {
645 // The presence of data in Gathered or Scattered indicates changes
646 // made to the Function.
647 if (Gathered.empty() && Scattered.empty())
649 for (GatherList::iterator GMI = Gathered.begin(), GME = Gathered.end();
651 Instruction *Op = GMI->first;
652 ValueVector &CV = *GMI->second;
653 if (!Op->use_empty()) {
654 // The value is still needed, so recreate it using a series of
656 Type *Ty = Op->getType();
657 Value *Res = UndefValue::get(Ty);
658 BasicBlock *BB = Op->getParent();
659 unsigned Count = Ty->getVectorNumElements();
660 IRBuilder<> Builder(Op);
661 if (isa<PHINode>(Op))
662 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
663 for (unsigned I = 0; I < Count; ++I)
664 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
665 Op->getName() + ".upto" + Twine(I));
667 Op->replaceAllUsesWith(Res);
669 Op->eraseFromParent();
676 FunctionPass *llvm::createScalarizerPass() {
677 return new Scalarizer();