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 &);
154 Scatterer scatter(Instruction *, Value *);
155 void gather(Instruction *, const ValueVector &);
156 bool canTransferMetadata(unsigned Kind);
157 void transferMetadata(Instruction *, const ValueVector &);
158 bool getVectorLayout(Type *, unsigned, VectorLayout &);
161 template<typename T> bool splitBinary(Instruction &, const T &);
163 ScatterMap Scattered;
165 unsigned ParallelLoopAccessMDKind;
166 const DataLayout *DL;
169 char Scalarizer::ID = 0;
170 } // end anonymous namespace
172 // This is disabled by default because having separate loads and stores makes
173 // it more likely that the -combiner-alias-analysis limits will be reached.
174 static cl::opt<bool> ScalarizeLoadStore
175 ("scalarize-load-store", cl::Hidden, cl::init(false),
176 cl::desc("Allow the scalarizer pass to scalarize loads and store"));
178 INITIALIZE_PASS(Scalarizer, "scalarizer", "Scalarize vector operations",
181 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
182 ValueVector *cachePtr)
183 : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
184 Type *Ty = V->getType();
185 PtrTy = dyn_cast<PointerType>(Ty);
187 Ty = PtrTy->getElementType();
188 Size = Ty->getVectorNumElements();
190 Tmp.resize(Size, nullptr);
191 else if (CachePtr->empty())
192 CachePtr->resize(Size, nullptr);
194 assert(Size == CachePtr->size() && "Inconsistent vector sizes");
197 // Return component I, creating a new Value for it if necessary.
198 Value *Scatterer::operator[](unsigned I) {
199 ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
200 // Try to reuse a previous value.
203 IRBuilder<> Builder(BB, BBI);
207 PointerType::get(PtrTy->getElementType()->getVectorElementType(),
208 PtrTy->getAddressSpace());
209 CV[0] = Builder.CreateBitCast(V, Ty, V->getName() + ".i0");
212 CV[I] = Builder.CreateConstGEP1_32(CV[0], I,
213 V->getName() + ".i" + Twine(I));
215 // Search through a chain of InsertElementInsts looking for element I.
216 // Record other elements in the cache. The new V is still suitable
217 // for all uncached indices.
219 InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
222 ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
225 unsigned J = Idx->getZExtValue();
226 CV[J] = Insert->getOperand(1);
227 V = Insert->getOperand(0);
231 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
232 V->getName() + ".i" + Twine(I));
237 bool Scalarizer::doInitialization(Module &M) {
238 ParallelLoopAccessMDKind =
239 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
243 bool Scalarizer::runOnFunction(Function &F) {
244 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
245 DL = DLP ? &DLP->getDataLayout() : nullptr;
246 for (Function::iterator BBI = F.begin(), BBE = F.end(); BBI != BBE; ++BBI) {
247 BasicBlock *BB = BBI;
248 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
250 bool Done = visit(I);
252 if (Done && I->getType()->isVoidTy())
253 I->eraseFromParent();
259 // Return a scattered form of V that can be accessed by Point. V must be a
260 // vector or a pointer to a vector.
261 Scatterer Scalarizer::scatter(Instruction *Point, Value *V) {
262 if (Argument *VArg = dyn_cast<Argument>(V)) {
263 // Put the scattered form of arguments in the entry block,
264 // so that it can be used everywhere.
265 Function *F = VArg->getParent();
266 BasicBlock *BB = &F->getEntryBlock();
267 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
269 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
270 // Put the scattered form of an instruction directly after the
272 BasicBlock *BB = VOp->getParent();
273 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
276 // In the fallback case, just put the scattered before Point and
277 // keep the result local to Point.
278 return Scatterer(Point->getParent(), Point, V);
281 // Replace Op with the gathered form of the components in CV. Defer the
282 // deletion of Op and creation of the gathered form to the end of the pass,
283 // so that we can avoid creating the gathered form if all uses of Op are
284 // replaced with uses of CV.
285 void Scalarizer::gather(Instruction *Op, const ValueVector &CV) {
286 // Since we're not deleting Op yet, stub out its operands, so that it
287 // doesn't make anything live unnecessarily.
288 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
289 Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
291 transferMetadata(Op, CV);
293 // If we already have a scattered form of Op (created from ExtractElements
294 // of Op itself), replace them with the new form.
295 ValueVector &SV = Scattered[Op];
297 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
298 Instruction *Old = cast<Instruction>(SV[I]);
299 CV[I]->takeName(Old);
300 Old->replaceAllUsesWith(CV[I]);
301 Old->eraseFromParent();
305 Gathered.push_back(GatherList::value_type(Op, &SV));
308 // Return true if it is safe to transfer the given metadata tag from
309 // vector to scalar instructions.
310 bool Scalarizer::canTransferMetadata(unsigned Tag) {
311 return (Tag == LLVMContext::MD_tbaa
312 || Tag == LLVMContext::MD_fpmath
313 || Tag == LLVMContext::MD_tbaa_struct
314 || Tag == LLVMContext::MD_invariant_load
315 || Tag == LLVMContext::MD_alias_scope
316 || Tag == LLVMContext::MD_noalias
317 || Tag == ParallelLoopAccessMDKind);
320 // Transfer metadata from Op to the instructions in CV if it is known
321 // to be safe to do so.
322 void Scalarizer::transferMetadata(Instruction *Op, const ValueVector &CV) {
323 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
324 Op->getAllMetadataOtherThanDebugLoc(MDs);
325 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
326 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
327 for (SmallVectorImpl<std::pair<unsigned, MDNode *> >::iterator
328 MI = MDs.begin(), ME = MDs.end(); MI != ME; ++MI)
329 if (canTransferMetadata(MI->first))
330 New->setMetadata(MI->first, MI->second);
331 New->setDebugLoc(Op->getDebugLoc());
336 // Try to fill in Layout from Ty, returning true on success. Alignment is
337 // the alignment of the vector, or 0 if the ABI default should be used.
338 bool Scalarizer::getVectorLayout(Type *Ty, unsigned Alignment,
339 VectorLayout &Layout) {
343 // Make sure we're dealing with a vector.
344 Layout.VecTy = dyn_cast<VectorType>(Ty);
348 // Check that we're dealing with full-byte elements.
349 Layout.ElemTy = Layout.VecTy->getElementType();
350 if (DL->getTypeSizeInBits(Layout.ElemTy) !=
351 DL->getTypeStoreSizeInBits(Layout.ElemTy))
355 Layout.VecAlign = Alignment;
357 Layout.VecAlign = DL->getABITypeAlignment(Layout.VecTy);
358 Layout.ElemSize = DL->getTypeStoreSize(Layout.ElemTy);
362 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
363 // to create an instruction like I with operands X and Y and name Name.
364 template<typename Splitter>
365 bool Scalarizer::splitBinary(Instruction &I, const Splitter &Split) {
366 VectorType *VT = dyn_cast<VectorType>(I.getType());
370 unsigned NumElems = VT->getNumElements();
371 IRBuilder<> Builder(I.getParent(), &I);
372 Scatterer Op0 = scatter(&I, I.getOperand(0));
373 Scatterer Op1 = scatter(&I, I.getOperand(1));
374 assert(Op0.size() == NumElems && "Mismatched binary operation");
375 assert(Op1.size() == NumElems && "Mismatched binary operation");
377 Res.resize(NumElems);
378 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
379 Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
380 I.getName() + ".i" + Twine(Elem));
385 bool Scalarizer::visitSelectInst(SelectInst &SI) {
386 VectorType *VT = dyn_cast<VectorType>(SI.getType());
390 unsigned NumElems = VT->getNumElements();
391 IRBuilder<> Builder(SI.getParent(), &SI);
392 Scatterer Op1 = scatter(&SI, SI.getOperand(1));
393 Scatterer Op2 = scatter(&SI, SI.getOperand(2));
394 assert(Op1.size() == NumElems && "Mismatched select");
395 assert(Op2.size() == NumElems && "Mismatched select");
397 Res.resize(NumElems);
399 if (SI.getOperand(0)->getType()->isVectorTy()) {
400 Scatterer Op0 = scatter(&SI, SI.getOperand(0));
401 assert(Op0.size() == NumElems && "Mismatched select");
402 for (unsigned I = 0; I < NumElems; ++I)
403 Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
404 SI.getName() + ".i" + Twine(I));
406 Value *Op0 = SI.getOperand(0);
407 for (unsigned I = 0; I < NumElems; ++I)
408 Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
409 SI.getName() + ".i" + Twine(I));
415 bool Scalarizer::visitICmpInst(ICmpInst &ICI) {
416 return splitBinary(ICI, ICmpSplitter(ICI));
419 bool Scalarizer::visitFCmpInst(FCmpInst &FCI) {
420 return splitBinary(FCI, FCmpSplitter(FCI));
423 bool Scalarizer::visitBinaryOperator(BinaryOperator &BO) {
424 return splitBinary(BO, BinarySplitter(BO));
427 bool Scalarizer::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
428 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
432 IRBuilder<> Builder(GEPI.getParent(), &GEPI);
433 unsigned NumElems = VT->getNumElements();
434 unsigned NumIndices = GEPI.getNumIndices();
436 Scatterer Base = scatter(&GEPI, GEPI.getOperand(0));
438 SmallVector<Scatterer, 8> Ops;
439 Ops.resize(NumIndices);
440 for (unsigned I = 0; I < NumIndices; ++I)
441 Ops[I] = scatter(&GEPI, GEPI.getOperand(I + 1));
444 Res.resize(NumElems);
445 for (unsigned I = 0; I < NumElems; ++I) {
446 SmallVector<Value *, 8> Indices;
447 Indices.resize(NumIndices);
448 for (unsigned J = 0; J < NumIndices; ++J)
449 Indices[J] = Ops[J][I];
450 Res[I] = Builder.CreateGEP(Base[I], Indices,
451 GEPI.getName() + ".i" + Twine(I));
452 if (GEPI.isInBounds())
453 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
454 NewGEPI->setIsInBounds();
460 bool Scalarizer::visitCastInst(CastInst &CI) {
461 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
465 unsigned NumElems = VT->getNumElements();
466 IRBuilder<> Builder(CI.getParent(), &CI);
467 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
468 assert(Op0.size() == NumElems && "Mismatched cast");
470 Res.resize(NumElems);
471 for (unsigned I = 0; I < NumElems; ++I)
472 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
473 CI.getName() + ".i" + Twine(I));
478 bool Scalarizer::visitBitCastInst(BitCastInst &BCI) {
479 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
480 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
481 if (!DstVT || !SrcVT)
484 unsigned DstNumElems = DstVT->getNumElements();
485 unsigned SrcNumElems = SrcVT->getNumElements();
486 IRBuilder<> Builder(BCI.getParent(), &BCI);
487 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
489 Res.resize(DstNumElems);
491 if (DstNumElems == SrcNumElems) {
492 for (unsigned I = 0; I < DstNumElems; ++I)
493 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
494 BCI.getName() + ".i" + Twine(I));
495 } else if (DstNumElems > SrcNumElems) {
496 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
497 // individual elements to the destination.
498 unsigned FanOut = DstNumElems / SrcNumElems;
499 Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
501 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
502 Value *V = Op0[Op0I];
504 // Look through any existing bitcasts before converting to <N x t2>.
505 // In the best case, the resulting conversion might be a no-op.
506 while ((VI = dyn_cast<Instruction>(V)) &&
507 VI->getOpcode() == Instruction::BitCast)
508 V = VI->getOperand(0);
509 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
510 Scatterer Mid = scatter(&BCI, V);
511 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
512 Res[ResI++] = Mid[MidI];
515 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
516 unsigned FanIn = SrcNumElems / DstNumElems;
517 Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
519 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
520 Value *V = UndefValue::get(MidTy);
521 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
522 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
523 BCI.getName() + ".i" + Twine(ResI)
524 + ".upto" + Twine(MidI));
525 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
526 BCI.getName() + ".i" + Twine(ResI));
533 bool Scalarizer::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
534 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
538 unsigned NumElems = VT->getNumElements();
539 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
540 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
542 Res.resize(NumElems);
544 for (unsigned I = 0; I < NumElems; ++I) {
545 int Selector = SVI.getMaskValue(I);
547 Res[I] = UndefValue::get(VT->getElementType());
548 else if (unsigned(Selector) < Op0.size())
549 Res[I] = Op0[Selector];
551 Res[I] = Op1[Selector - Op0.size()];
557 bool Scalarizer::visitPHINode(PHINode &PHI) {
558 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
562 unsigned NumElems = VT->getNumElements();
563 IRBuilder<> Builder(PHI.getParent(), &PHI);
565 Res.resize(NumElems);
567 unsigned NumOps = PHI.getNumOperands();
568 for (unsigned I = 0; I < NumElems; ++I)
569 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
570 PHI.getName() + ".i" + Twine(I));
572 for (unsigned I = 0; I < NumOps; ++I) {
573 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
574 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
575 for (unsigned J = 0; J < NumElems; ++J)
576 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
582 bool Scalarizer::visitLoadInst(LoadInst &LI) {
583 if (!ScalarizeLoadStore)
589 if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout))
592 unsigned NumElems = Layout.VecTy->getNumElements();
593 IRBuilder<> Builder(LI.getParent(), &LI);
594 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
596 Res.resize(NumElems);
598 for (unsigned I = 0; I < NumElems; ++I)
599 Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
600 LI.getName() + ".i" + Twine(I));
605 bool Scalarizer::visitStoreInst(StoreInst &SI) {
606 if (!ScalarizeLoadStore)
612 Value *FullValue = SI.getValueOperand();
613 if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout))
616 unsigned NumElems = Layout.VecTy->getNumElements();
617 IRBuilder<> Builder(SI.getParent(), &SI);
618 Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
619 Scatterer Val = scatter(&SI, FullValue);
622 Stores.resize(NumElems);
623 for (unsigned I = 0; I < NumElems; ++I) {
624 unsigned Align = Layout.getElemAlign(I);
625 Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
627 transferMetadata(&SI, Stores);
631 // Delete the instructions that we scalarized. If a full vector result
632 // is still needed, recreate it using InsertElements.
633 bool Scalarizer::finish() {
634 if (Gathered.empty())
636 for (GatherList::iterator GMI = Gathered.begin(), GME = Gathered.end();
638 Instruction *Op = GMI->first;
639 ValueVector &CV = *GMI->second;
640 if (!Op->use_empty()) {
641 // The value is still needed, so recreate it using a series of
643 Type *Ty = Op->getType();
644 Value *Res = UndefValue::get(Ty);
645 BasicBlock *BB = Op->getParent();
646 unsigned Count = Ty->getVectorNumElements();
647 IRBuilder<> Builder(BB, Op);
648 if (isa<PHINode>(Op))
649 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
650 for (unsigned I = 0; I < Count; ++I)
651 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
652 Op->getName() + ".upto" + Twine(I));
654 Op->replaceAllUsesWith(Res);
656 Op->eraseFromParent();
663 FunctionPass *llvm::createScalarizerPass() {
664 return new Scalarizer();