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 CV[J] = Insert->getOperand(1);
231 V = Insert->getOperand(0);
235 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
236 V->getName() + ".i" + Twine(I));
241 bool Scalarizer::doInitialization(Module &M) {
242 ParallelLoopAccessMDKind =
243 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
245 M.getContext().getOption<bool, Scalarizer, &Scalarizer::ScalarizeLoadStore>();
249 bool Scalarizer::runOnFunction(Function &F) {
250 for (Function::iterator BBI = F.begin(), BBE = F.end(); BBI != BBE; ++BBI) {
251 BasicBlock *BB = BBI;
252 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
254 bool Done = visit(I);
256 if (Done && I->getType()->isVoidTy())
257 I->eraseFromParent();
263 // Return a scattered form of V that can be accessed by Point. V must be a
264 // vector or a pointer to a vector.
265 Scatterer Scalarizer::scatter(Instruction *Point, Value *V) {
266 if (Argument *VArg = dyn_cast<Argument>(V)) {
267 // Put the scattered form of arguments in the entry block,
268 // so that it can be used everywhere.
269 Function *F = VArg->getParent();
270 BasicBlock *BB = &F->getEntryBlock();
271 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
273 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
274 // Put the scattered form of an instruction directly after the
276 BasicBlock *BB = VOp->getParent();
277 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
280 // In the fallback case, just put the scattered before Point and
281 // keep the result local to Point.
282 return Scatterer(Point->getParent(), Point, V);
285 // Replace Op with the gathered form of the components in CV. Defer the
286 // deletion of Op and creation of the gathered form to the end of the pass,
287 // so that we can avoid creating the gathered form if all uses of Op are
288 // replaced with uses of CV.
289 void Scalarizer::gather(Instruction *Op, const ValueVector &CV) {
290 // Since we're not deleting Op yet, stub out its operands, so that it
291 // doesn't make anything live unnecessarily.
292 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
293 Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
295 transferMetadata(Op, CV);
297 // If we already have a scattered form of Op (created from ExtractElements
298 // of Op itself), replace them with the new form.
299 ValueVector &SV = Scattered[Op];
301 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
302 Instruction *Old = cast<Instruction>(SV[I]);
303 CV[I]->takeName(Old);
304 Old->replaceAllUsesWith(CV[I]);
305 Old->eraseFromParent();
309 Gathered.push_back(GatherList::value_type(Op, &SV));
312 // Return true if it is safe to transfer the given metadata tag from
313 // vector to scalar instructions.
314 bool Scalarizer::canTransferMetadata(unsigned Tag) {
315 return (Tag == LLVMContext::MD_tbaa
316 || Tag == LLVMContext::MD_fpmath
317 || Tag == LLVMContext::MD_tbaa_struct
318 || Tag == LLVMContext::MD_invariant_load
319 || Tag == LLVMContext::MD_alias_scope
320 || Tag == LLVMContext::MD_noalias
321 || Tag == ParallelLoopAccessMDKind);
324 // Transfer metadata from Op to the instructions in CV if it is known
325 // to be safe to do so.
326 void Scalarizer::transferMetadata(Instruction *Op, const ValueVector &CV) {
327 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
328 Op->getAllMetadataOtherThanDebugLoc(MDs);
329 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
330 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
331 for (SmallVectorImpl<std::pair<unsigned, MDNode *>>::iterator
335 if (canTransferMetadata(MI->first))
336 New->setMetadata(MI->first, MI->second);
337 New->setDebugLoc(Op->getDebugLoc());
342 // Try to fill in Layout from Ty, returning true on success. Alignment is
343 // the alignment of the vector, or 0 if the ABI default should be used.
344 bool Scalarizer::getVectorLayout(Type *Ty, unsigned Alignment,
345 VectorLayout &Layout, const DataLayout &DL) {
346 // Make sure we're dealing with a vector.
347 Layout.VecTy = dyn_cast<VectorType>(Ty);
351 // Check that we're dealing with full-byte elements.
352 Layout.ElemTy = Layout.VecTy->getElementType();
353 if (DL.getTypeSizeInBits(Layout.ElemTy) !=
354 DL.getTypeStoreSizeInBits(Layout.ElemTy))
358 Layout.VecAlign = Alignment;
360 Layout.VecAlign = DL.getABITypeAlignment(Layout.VecTy);
361 Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
365 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
366 // to create an instruction like I with operands X and Y and name Name.
367 template<typename Splitter>
368 bool Scalarizer::splitBinary(Instruction &I, const Splitter &Split) {
369 VectorType *VT = dyn_cast<VectorType>(I.getType());
373 unsigned NumElems = VT->getNumElements();
374 IRBuilder<> Builder(I.getParent(), &I);
375 Scatterer Op0 = scatter(&I, I.getOperand(0));
376 Scatterer Op1 = scatter(&I, I.getOperand(1));
377 assert(Op0.size() == NumElems && "Mismatched binary operation");
378 assert(Op1.size() == NumElems && "Mismatched binary operation");
380 Res.resize(NumElems);
381 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
382 Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
383 I.getName() + ".i" + Twine(Elem));
388 bool Scalarizer::visitSelectInst(SelectInst &SI) {
389 VectorType *VT = dyn_cast<VectorType>(SI.getType());
393 unsigned NumElems = VT->getNumElements();
394 IRBuilder<> Builder(SI.getParent(), &SI);
395 Scatterer Op1 = scatter(&SI, SI.getOperand(1));
396 Scatterer Op2 = scatter(&SI, SI.getOperand(2));
397 assert(Op1.size() == NumElems && "Mismatched select");
398 assert(Op2.size() == NumElems && "Mismatched select");
400 Res.resize(NumElems);
402 if (SI.getOperand(0)->getType()->isVectorTy()) {
403 Scatterer Op0 = scatter(&SI, SI.getOperand(0));
404 assert(Op0.size() == NumElems && "Mismatched select");
405 for (unsigned I = 0; I < NumElems; ++I)
406 Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
407 SI.getName() + ".i" + Twine(I));
409 Value *Op0 = SI.getOperand(0);
410 for (unsigned I = 0; I < NumElems; ++I)
411 Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
412 SI.getName() + ".i" + Twine(I));
418 bool Scalarizer::visitICmpInst(ICmpInst &ICI) {
419 return splitBinary(ICI, ICmpSplitter(ICI));
422 bool Scalarizer::visitFCmpInst(FCmpInst &FCI) {
423 return splitBinary(FCI, FCmpSplitter(FCI));
426 bool Scalarizer::visitBinaryOperator(BinaryOperator &BO) {
427 return splitBinary(BO, BinarySplitter(BO));
430 bool Scalarizer::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
431 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
435 IRBuilder<> Builder(GEPI.getParent(), &GEPI);
436 unsigned NumElems = VT->getNumElements();
437 unsigned NumIndices = GEPI.getNumIndices();
439 Scatterer Base = scatter(&GEPI, GEPI.getOperand(0));
441 SmallVector<Scatterer, 8> Ops;
442 Ops.resize(NumIndices);
443 for (unsigned I = 0; I < NumIndices; ++I)
444 Ops[I] = scatter(&GEPI, GEPI.getOperand(I + 1));
447 Res.resize(NumElems);
448 for (unsigned I = 0; I < NumElems; ++I) {
449 SmallVector<Value *, 8> Indices;
450 Indices.resize(NumIndices);
451 for (unsigned J = 0; J < NumIndices; ++J)
452 Indices[J] = Ops[J][I];
453 Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
454 GEPI.getName() + ".i" + Twine(I));
455 if (GEPI.isInBounds())
456 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
457 NewGEPI->setIsInBounds();
463 bool Scalarizer::visitCastInst(CastInst &CI) {
464 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
468 unsigned NumElems = VT->getNumElements();
469 IRBuilder<> Builder(CI.getParent(), &CI);
470 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
471 assert(Op0.size() == NumElems && "Mismatched cast");
473 Res.resize(NumElems);
474 for (unsigned I = 0; I < NumElems; ++I)
475 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
476 CI.getName() + ".i" + Twine(I));
481 bool Scalarizer::visitBitCastInst(BitCastInst &BCI) {
482 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
483 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
484 if (!DstVT || !SrcVT)
487 unsigned DstNumElems = DstVT->getNumElements();
488 unsigned SrcNumElems = SrcVT->getNumElements();
489 IRBuilder<> Builder(BCI.getParent(), &BCI);
490 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
492 Res.resize(DstNumElems);
494 if (DstNumElems == SrcNumElems) {
495 for (unsigned I = 0; I < DstNumElems; ++I)
496 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
497 BCI.getName() + ".i" + Twine(I));
498 } else if (DstNumElems > SrcNumElems) {
499 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
500 // individual elements to the destination.
501 unsigned FanOut = DstNumElems / SrcNumElems;
502 Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
504 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
505 Value *V = Op0[Op0I];
507 // Look through any existing bitcasts before converting to <N x t2>.
508 // In the best case, the resulting conversion might be a no-op.
509 while ((VI = dyn_cast<Instruction>(V)) &&
510 VI->getOpcode() == Instruction::BitCast)
511 V = VI->getOperand(0);
512 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
513 Scatterer Mid = scatter(&BCI, V);
514 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
515 Res[ResI++] = Mid[MidI];
518 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
519 unsigned FanIn = SrcNumElems / DstNumElems;
520 Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
522 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
523 Value *V = UndefValue::get(MidTy);
524 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
525 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
526 BCI.getName() + ".i" + Twine(ResI)
527 + ".upto" + Twine(MidI));
528 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
529 BCI.getName() + ".i" + Twine(ResI));
536 bool Scalarizer::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
537 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
541 unsigned NumElems = VT->getNumElements();
542 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
543 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
545 Res.resize(NumElems);
547 for (unsigned I = 0; I < NumElems; ++I) {
548 int Selector = SVI.getMaskValue(I);
550 Res[I] = UndefValue::get(VT->getElementType());
551 else if (unsigned(Selector) < Op0.size())
552 Res[I] = Op0[Selector];
554 Res[I] = Op1[Selector - Op0.size()];
560 bool Scalarizer::visitPHINode(PHINode &PHI) {
561 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
565 unsigned NumElems = VT->getNumElements();
566 IRBuilder<> Builder(PHI.getParent(), &PHI);
568 Res.resize(NumElems);
570 unsigned NumOps = PHI.getNumOperands();
571 for (unsigned I = 0; I < NumElems; ++I)
572 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
573 PHI.getName() + ".i" + Twine(I));
575 for (unsigned I = 0; I < NumOps; ++I) {
576 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
577 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
578 for (unsigned J = 0; J < NumElems; ++J)
579 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
585 bool Scalarizer::visitLoadInst(LoadInst &LI) {
586 if (!ScalarizeLoadStore)
592 if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout,
593 LI.getModule()->getDataLayout()))
596 unsigned NumElems = Layout.VecTy->getNumElements();
597 IRBuilder<> Builder(LI.getParent(), &LI);
598 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
600 Res.resize(NumElems);
602 for (unsigned I = 0; I < NumElems; ++I)
603 Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
604 LI.getName() + ".i" + Twine(I));
609 bool Scalarizer::visitStoreInst(StoreInst &SI) {
610 if (!ScalarizeLoadStore)
616 Value *FullValue = SI.getValueOperand();
617 if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout,
618 SI.getModule()->getDataLayout()))
621 unsigned NumElems = Layout.VecTy->getNumElements();
622 IRBuilder<> Builder(SI.getParent(), &SI);
623 Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
624 Scatterer Val = scatter(&SI, FullValue);
627 Stores.resize(NumElems);
628 for (unsigned I = 0; I < NumElems; ++I) {
629 unsigned Align = Layout.getElemAlign(I);
630 Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
632 transferMetadata(&SI, Stores);
636 // Delete the instructions that we scalarized. If a full vector result
637 // is still needed, recreate it using InsertElements.
638 bool Scalarizer::finish() {
639 if (Gathered.empty())
641 for (GatherList::iterator GMI = Gathered.begin(), GME = Gathered.end();
643 Instruction *Op = GMI->first;
644 ValueVector &CV = *GMI->second;
645 if (!Op->use_empty()) {
646 // The value is still needed, so recreate it using a series of
648 Type *Ty = Op->getType();
649 Value *Res = UndefValue::get(Ty);
650 BasicBlock *BB = Op->getParent();
651 unsigned Count = Ty->getVectorNumElements();
652 IRBuilder<> Builder(BB, Op);
653 if (isa<PHINode>(Op))
654 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
655 for (unsigned I = 0; I < Count; ++I)
656 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
657 Op->getName() + ".upto" + Twine(I));
659 Op->replaceAllUsesWith(Res);
661 Op->eraseFromParent();
668 FunctionPass *llvm::createScalarizerPass() {
669 return new Scalarizer();