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/PostOrderIterator.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/VectorUtils.h"
21 #include "llvm/IR/Argument.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstVisitor.h"
29 #include "llvm/IR/InstrTypes.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Casting.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/Options.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Scalar/Scalarizer.h"
51 #define DEBUG_TYPE "scalarizer"
53 // This is disabled by default because having separate loads and stores
54 // makes it more likely that the -combiner-alias-analysis limits will be
57 ScalarizeLoadStore("scalarize-load-store", cl::init(false), cl::Hidden,
58 cl::desc("Allow the scalarizer pass to scalarize loads and store"));
62 // Used to store the scattered form of a vector.
63 using ValueVector = SmallVector<Value *, 8>;
65 // Used to map a vector Value to its scattered form. We use std::map
66 // because we want iterators to persist across insertion and because the
67 // values are relatively large.
68 using ScatterMap = std::map<Value *, ValueVector>;
70 // Lists Instructions that have been replaced with scalar implementations,
71 // along with a pointer to their scattered forms.
72 using GatherList = SmallVector<std::pair<Instruction *, ValueVector *>, 16>;
74 // Provides a very limited vector-like interface for lazily accessing one
75 // component of a scattered vector or vector pointer.
78 Scatterer() = default;
80 // Scatter V into Size components. If new instructions are needed,
81 // insert them before BBI in BB. If Cache is nonnull, use it to cache
83 Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
84 ValueVector *cachePtr = nullptr);
86 // Return component I, creating a new Value for it if necessary.
87 Value *operator[](unsigned I);
89 // Return the number of components.
90 unsigned size() const { return Size; }
94 BasicBlock::iterator BBI;
96 ValueVector *CachePtr;
102 // FCmpSpliiter(FCI)(Builder, X, Y, Name) uses Builder to create an FCmp
103 // called Name that compares X and Y in the same way as FCI.
104 struct FCmpSplitter {
105 FCmpSplitter(FCmpInst &fci) : FCI(fci) {}
107 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
108 const Twine &Name) const {
109 return Builder.CreateFCmp(FCI.getPredicate(), Op0, Op1, Name);
115 // ICmpSpliiter(ICI)(Builder, X, Y, Name) uses Builder to create an ICmp
116 // called Name that compares X and Y in the same way as ICI.
117 struct ICmpSplitter {
118 ICmpSplitter(ICmpInst &ici) : ICI(ici) {}
120 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
121 const Twine &Name) const {
122 return Builder.CreateICmp(ICI.getPredicate(), Op0, Op1, Name);
128 // BinarySpliiter(BO)(Builder, X, Y, Name) uses Builder to create
129 // a binary operator like BO called Name with operands X and Y.
130 struct BinarySplitter {
131 BinarySplitter(BinaryOperator &bo) : BO(bo) {}
133 Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
134 const Twine &Name) const {
135 return Builder.CreateBinOp(BO.getOpcode(), Op0, Op1, Name);
141 // Information about a load or store that we're scalarizing.
142 struct VectorLayout {
143 VectorLayout() = default;
145 // Return the alignment of element I.
146 uint64_t getElemAlign(unsigned I) {
147 return MinAlign(VecAlign, I * ElemSize);
150 // The type of the vector.
151 VectorType *VecTy = nullptr;
153 // The type of each element.
154 Type *ElemTy = nullptr;
156 // The alignment of the vector.
157 uint64_t VecAlign = 0;
159 // The size of each element.
160 uint64_t ElemSize = 0;
163 class ScalarizerVisitor : public InstVisitor<ScalarizerVisitor, bool> {
165 ScalarizerVisitor(unsigned ParallelLoopAccessMDKind)
166 : ParallelLoopAccessMDKind(ParallelLoopAccessMDKind) {
169 bool visit(Function &F);
171 // InstVisitor methods. They return true if the instruction was scalarized,
172 // false if nothing changed.
173 bool visitInstruction(Instruction &I) { return false; }
174 bool visitSelectInst(SelectInst &SI);
175 bool visitICmpInst(ICmpInst &ICI);
176 bool visitFCmpInst(FCmpInst &FCI);
177 bool visitBinaryOperator(BinaryOperator &BO);
178 bool visitGetElementPtrInst(GetElementPtrInst &GEPI);
179 bool visitCastInst(CastInst &CI);
180 bool visitBitCastInst(BitCastInst &BCI);
181 bool visitShuffleVectorInst(ShuffleVectorInst &SVI);
182 bool visitPHINode(PHINode &PHI);
183 bool visitLoadInst(LoadInst &LI);
184 bool visitStoreInst(StoreInst &SI);
185 bool visitCallInst(CallInst &ICI);
188 Scatterer scatter(Instruction *Point, Value *V);
189 void gather(Instruction *Op, const ValueVector &CV);
190 bool canTransferMetadata(unsigned Kind);
191 void transferMetadata(Instruction *Op, const ValueVector &CV);
192 bool getVectorLayout(Type *Ty, unsigned Alignment, VectorLayout &Layout,
193 const DataLayout &DL);
196 template<typename T> bool splitBinary(Instruction &, const T &);
198 bool splitCall(CallInst &CI);
200 ScatterMap Scattered;
203 unsigned ParallelLoopAccessMDKind;
206 class ScalarizerLegacyPass : public FunctionPass {
210 ScalarizerLegacyPass() : FunctionPass(ID) {
211 initializeScalarizerLegacyPassPass(*PassRegistry::getPassRegistry());
214 bool runOnFunction(Function &F) override;
217 } // end anonymous namespace
219 char ScalarizerLegacyPass::ID = 0;
220 INITIALIZE_PASS_BEGIN(ScalarizerLegacyPass, "scalarizer",
221 "Scalarize vector operations", false, false)
222 INITIALIZE_PASS_END(ScalarizerLegacyPass, "scalarizer",
223 "Scalarize vector operations", false, false)
225 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
226 ValueVector *cachePtr)
227 : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
228 Type *Ty = V->getType();
229 PtrTy = dyn_cast<PointerType>(Ty);
231 Ty = PtrTy->getElementType();
232 Size = Ty->getVectorNumElements();
234 Tmp.resize(Size, nullptr);
235 else if (CachePtr->empty())
236 CachePtr->resize(Size, nullptr);
238 assert(Size == CachePtr->size() && "Inconsistent vector sizes");
241 // Return component I, creating a new Value for it if necessary.
242 Value *Scatterer::operator[](unsigned I) {
243 ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
244 // Try to reuse a previous value.
247 IRBuilder<> Builder(BB, BBI);
251 PointerType::get(PtrTy->getElementType()->getVectorElementType(),
252 PtrTy->getAddressSpace());
253 CV[0] = Builder.CreateBitCast(V, Ty, V->getName() + ".i0");
256 CV[I] = Builder.CreateConstGEP1_32(nullptr, CV[0], I,
257 V->getName() + ".i" + Twine(I));
259 // Search through a chain of InsertElementInsts looking for element I.
260 // Record other elements in the cache. The new V is still suitable
261 // for all uncached indices.
263 InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
266 ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
269 unsigned J = Idx->getZExtValue();
270 V = Insert->getOperand(0);
272 CV[J] = Insert->getOperand(1);
275 // Only cache the first entry we find for each index we're not actively
276 // searching for. This prevents us from going too far up the chain and
277 // caching incorrect entries.
278 CV[J] = Insert->getOperand(1);
281 CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
282 V->getName() + ".i" + Twine(I));
287 bool ScalarizerLegacyPass::runOnFunction(Function &F) {
291 Module &M = *F.getParent();
292 unsigned ParallelLoopAccessMDKind =
293 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
294 ScalarizerVisitor Impl(ParallelLoopAccessMDKind);
295 return Impl.visit(F);
298 FunctionPass *llvm::createScalarizerPass() {
299 return new ScalarizerLegacyPass();
302 bool ScalarizerVisitor::visit(Function &F) {
303 assert(Gathered.empty() && Scattered.empty());
305 // To ensure we replace gathered components correctly we need to do an ordered
306 // traversal of the basic blocks in the function.
307 ReversePostOrderTraversal<BasicBlock *> RPOT(&F.getEntryBlock());
308 for (BasicBlock *BB : RPOT) {
309 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
310 Instruction *I = &*II;
311 bool Done = InstVisitor::visit(I);
313 if (Done && I->getType()->isVoidTy())
314 I->eraseFromParent();
320 // Return a scattered form of V that can be accessed by Point. V must be a
321 // vector or a pointer to a vector.
322 Scatterer ScalarizerVisitor::scatter(Instruction *Point, Value *V) {
323 if (Argument *VArg = dyn_cast<Argument>(V)) {
324 // Put the scattered form of arguments in the entry block,
325 // so that it can be used everywhere.
326 Function *F = VArg->getParent();
327 BasicBlock *BB = &F->getEntryBlock();
328 return Scatterer(BB, BB->begin(), V, &Scattered[V]);
330 if (Instruction *VOp = dyn_cast<Instruction>(V)) {
331 // Put the scattered form of an instruction directly after the
333 BasicBlock *BB = VOp->getParent();
334 return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
337 // In the fallback case, just put the scattered before Point and
338 // keep the result local to Point.
339 return Scatterer(Point->getParent(), Point->getIterator(), V);
342 // Replace Op with the gathered form of the components in CV. Defer the
343 // deletion of Op and creation of the gathered form to the end of the pass,
344 // so that we can avoid creating the gathered form if all uses of Op are
345 // replaced with uses of CV.
346 void ScalarizerVisitor::gather(Instruction *Op, const ValueVector &CV) {
347 // Since we're not deleting Op yet, stub out its operands, so that it
348 // doesn't make anything live unnecessarily.
349 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
350 Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
352 transferMetadata(Op, CV);
354 // If we already have a scattered form of Op (created from ExtractElements
355 // of Op itself), replace them with the new form.
356 ValueVector &SV = Scattered[Op];
358 for (unsigned I = 0, E = SV.size(); I != E; ++I) {
363 Instruction *Old = cast<Instruction>(V);
364 CV[I]->takeName(Old);
365 Old->replaceAllUsesWith(CV[I]);
366 Old->eraseFromParent();
370 Gathered.push_back(GatherList::value_type(Op, &SV));
373 // Return true if it is safe to transfer the given metadata tag from
374 // vector to scalar instructions.
375 bool ScalarizerVisitor::canTransferMetadata(unsigned Tag) {
376 return (Tag == LLVMContext::MD_tbaa
377 || Tag == LLVMContext::MD_fpmath
378 || Tag == LLVMContext::MD_tbaa_struct
379 || Tag == LLVMContext::MD_invariant_load
380 || Tag == LLVMContext::MD_alias_scope
381 || Tag == LLVMContext::MD_noalias
382 || Tag == ParallelLoopAccessMDKind
383 || Tag == LLVMContext::MD_access_group);
386 // Transfer metadata from Op to the instructions in CV if it is known
387 // to be safe to do so.
388 void ScalarizerVisitor::transferMetadata(Instruction *Op, const ValueVector &CV) {
389 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
390 Op->getAllMetadataOtherThanDebugLoc(MDs);
391 for (unsigned I = 0, E = CV.size(); I != E; ++I) {
392 if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
393 for (const auto &MD : MDs)
394 if (canTransferMetadata(MD.first))
395 New->setMetadata(MD.first, MD.second);
396 if (Op->getDebugLoc() && !New->getDebugLoc())
397 New->setDebugLoc(Op->getDebugLoc());
402 // Try to fill in Layout from Ty, returning true on success. Alignment is
403 // the alignment of the vector, or 0 if the ABI default should be used.
404 bool ScalarizerVisitor::getVectorLayout(Type *Ty, unsigned Alignment,
405 VectorLayout &Layout, const DataLayout &DL) {
406 // Make sure we're dealing with a vector.
407 Layout.VecTy = dyn_cast<VectorType>(Ty);
411 // Check that we're dealing with full-byte elements.
412 Layout.ElemTy = Layout.VecTy->getElementType();
413 if (DL.getTypeSizeInBits(Layout.ElemTy) !=
414 DL.getTypeStoreSizeInBits(Layout.ElemTy))
418 Layout.VecAlign = Alignment;
420 Layout.VecAlign = DL.getABITypeAlignment(Layout.VecTy);
421 Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
425 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
426 // to create an instruction like I with operands X and Y and name Name.
427 template<typename Splitter>
428 bool ScalarizerVisitor::splitBinary(Instruction &I, const Splitter &Split) {
429 VectorType *VT = dyn_cast<VectorType>(I.getType());
433 unsigned NumElems = VT->getNumElements();
434 IRBuilder<> Builder(&I);
435 Scatterer Op0 = scatter(&I, I.getOperand(0));
436 Scatterer Op1 = scatter(&I, I.getOperand(1));
437 assert(Op0.size() == NumElems && "Mismatched binary operation");
438 assert(Op1.size() == NumElems && "Mismatched binary operation");
440 Res.resize(NumElems);
441 for (unsigned Elem = 0; Elem < NumElems; ++Elem)
442 Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
443 I.getName() + ".i" + Twine(Elem));
448 static bool isTriviallyScalariable(Intrinsic::ID ID) {
449 return isTriviallyVectorizable(ID);
452 // All of the current scalarizable intrinsics only have one mangled type.
453 static Function *getScalarIntrinsicDeclaration(Module *M,
456 return Intrinsic::getDeclaration(M, ID, { Ty->getScalarType() });
459 /// If a call to a vector typed intrinsic function, split into a scalar call per
460 /// element if possible for the intrinsic.
461 bool ScalarizerVisitor::splitCall(CallInst &CI) {
462 VectorType *VT = dyn_cast<VectorType>(CI.getType());
466 Function *F = CI.getCalledFunction();
470 Intrinsic::ID ID = F->getIntrinsicID();
471 if (ID == Intrinsic::not_intrinsic || !isTriviallyScalariable(ID))
474 unsigned NumElems = VT->getNumElements();
475 unsigned NumArgs = CI.getNumArgOperands();
477 ValueVector ScalarOperands(NumArgs);
478 SmallVector<Scatterer, 8> Scattered(NumArgs);
480 Scattered.resize(NumArgs);
482 // Assumes that any vector type has the same number of elements as the return
483 // vector type, which is true for all current intrinsics.
484 for (unsigned I = 0; I != NumArgs; ++I) {
485 Value *OpI = CI.getOperand(I);
486 if (OpI->getType()->isVectorTy()) {
487 Scattered[I] = scatter(&CI, OpI);
488 assert(Scattered[I].size() == NumElems && "mismatched call operands");
490 ScalarOperands[I] = OpI;
494 ValueVector Res(NumElems);
495 ValueVector ScalarCallOps(NumArgs);
497 Function *NewIntrin = getScalarIntrinsicDeclaration(F->getParent(), ID, VT);
498 IRBuilder<> Builder(&CI);
500 // Perform actual scalarization, taking care to preserve any scalar operands.
501 for (unsigned Elem = 0; Elem < NumElems; ++Elem) {
502 ScalarCallOps.clear();
504 for (unsigned J = 0; J != NumArgs; ++J) {
505 if (hasVectorInstrinsicScalarOpd(ID, J))
506 ScalarCallOps.push_back(ScalarOperands[J]);
508 ScalarCallOps.push_back(Scattered[J][Elem]);
511 Res[Elem] = Builder.CreateCall(NewIntrin, ScalarCallOps,
512 CI.getName() + ".i" + Twine(Elem));
519 bool ScalarizerVisitor::visitSelectInst(SelectInst &SI) {
520 VectorType *VT = dyn_cast<VectorType>(SI.getType());
524 unsigned NumElems = VT->getNumElements();
525 IRBuilder<> Builder(&SI);
526 Scatterer Op1 = scatter(&SI, SI.getOperand(1));
527 Scatterer Op2 = scatter(&SI, SI.getOperand(2));
528 assert(Op1.size() == NumElems && "Mismatched select");
529 assert(Op2.size() == NumElems && "Mismatched select");
531 Res.resize(NumElems);
533 if (SI.getOperand(0)->getType()->isVectorTy()) {
534 Scatterer Op0 = scatter(&SI, SI.getOperand(0));
535 assert(Op0.size() == NumElems && "Mismatched select");
536 for (unsigned I = 0; I < NumElems; ++I)
537 Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
538 SI.getName() + ".i" + Twine(I));
540 Value *Op0 = SI.getOperand(0);
541 for (unsigned I = 0; I < NumElems; ++I)
542 Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
543 SI.getName() + ".i" + Twine(I));
549 bool ScalarizerVisitor::visitICmpInst(ICmpInst &ICI) {
550 return splitBinary(ICI, ICmpSplitter(ICI));
553 bool ScalarizerVisitor::visitFCmpInst(FCmpInst &FCI) {
554 return splitBinary(FCI, FCmpSplitter(FCI));
557 bool ScalarizerVisitor::visitBinaryOperator(BinaryOperator &BO) {
558 return splitBinary(BO, BinarySplitter(BO));
561 bool ScalarizerVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
562 VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
566 IRBuilder<> Builder(&GEPI);
567 unsigned NumElems = VT->getNumElements();
568 unsigned NumIndices = GEPI.getNumIndices();
570 // The base pointer might be scalar even if it's a vector GEP. In those cases,
571 // splat the pointer into a vector value, and scatter that vector.
572 Value *Op0 = GEPI.getOperand(0);
573 if (!Op0->getType()->isVectorTy())
574 Op0 = Builder.CreateVectorSplat(NumElems, Op0);
575 Scatterer Base = scatter(&GEPI, Op0);
577 SmallVector<Scatterer, 8> Ops;
578 Ops.resize(NumIndices);
579 for (unsigned I = 0; I < NumIndices; ++I) {
580 Value *Op = GEPI.getOperand(I + 1);
582 // The indices might be scalars even if it's a vector GEP. In those cases,
583 // splat the scalar into a vector value, and scatter that vector.
584 if (!Op->getType()->isVectorTy())
585 Op = Builder.CreateVectorSplat(NumElems, Op);
587 Ops[I] = scatter(&GEPI, Op);
591 Res.resize(NumElems);
592 for (unsigned I = 0; I < NumElems; ++I) {
593 SmallVector<Value *, 8> Indices;
594 Indices.resize(NumIndices);
595 for (unsigned J = 0; J < NumIndices; ++J)
596 Indices[J] = Ops[J][I];
597 Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
598 GEPI.getName() + ".i" + Twine(I));
599 if (GEPI.isInBounds())
600 if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
601 NewGEPI->setIsInBounds();
607 bool ScalarizerVisitor::visitCastInst(CastInst &CI) {
608 VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
612 unsigned NumElems = VT->getNumElements();
613 IRBuilder<> Builder(&CI);
614 Scatterer Op0 = scatter(&CI, CI.getOperand(0));
615 assert(Op0.size() == NumElems && "Mismatched cast");
617 Res.resize(NumElems);
618 for (unsigned I = 0; I < NumElems; ++I)
619 Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
620 CI.getName() + ".i" + Twine(I));
625 bool ScalarizerVisitor::visitBitCastInst(BitCastInst &BCI) {
626 VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
627 VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
628 if (!DstVT || !SrcVT)
631 unsigned DstNumElems = DstVT->getNumElements();
632 unsigned SrcNumElems = SrcVT->getNumElements();
633 IRBuilder<> Builder(&BCI);
634 Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
636 Res.resize(DstNumElems);
638 if (DstNumElems == SrcNumElems) {
639 for (unsigned I = 0; I < DstNumElems; ++I)
640 Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
641 BCI.getName() + ".i" + Twine(I));
642 } else if (DstNumElems > SrcNumElems) {
643 // <M x t1> -> <N*M x t2>. Convert each t1 to <N x t2> and copy the
644 // individual elements to the destination.
645 unsigned FanOut = DstNumElems / SrcNumElems;
646 Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
648 for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
649 Value *V = Op0[Op0I];
651 // Look through any existing bitcasts before converting to <N x t2>.
652 // In the best case, the resulting conversion might be a no-op.
653 while ((VI = dyn_cast<Instruction>(V)) &&
654 VI->getOpcode() == Instruction::BitCast)
655 V = VI->getOperand(0);
656 V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
657 Scatterer Mid = scatter(&BCI, V);
658 for (unsigned MidI = 0; MidI < FanOut; ++MidI)
659 Res[ResI++] = Mid[MidI];
662 // <N*M x t1> -> <M x t2>. Convert each group of <N x t1> into a t2.
663 unsigned FanIn = SrcNumElems / DstNumElems;
664 Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
666 for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
667 Value *V = UndefValue::get(MidTy);
668 for (unsigned MidI = 0; MidI < FanIn; ++MidI)
669 V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
670 BCI.getName() + ".i" + Twine(ResI)
671 + ".upto" + Twine(MidI));
672 Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
673 BCI.getName() + ".i" + Twine(ResI));
680 bool ScalarizerVisitor::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
681 VectorType *VT = dyn_cast<VectorType>(SVI.getType());
685 unsigned NumElems = VT->getNumElements();
686 Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
687 Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
689 Res.resize(NumElems);
691 for (unsigned I = 0; I < NumElems; ++I) {
692 int Selector = SVI.getMaskValue(I);
694 Res[I] = UndefValue::get(VT->getElementType());
695 else if (unsigned(Selector) < Op0.size())
696 Res[I] = Op0[Selector];
698 Res[I] = Op1[Selector - Op0.size()];
704 bool ScalarizerVisitor::visitPHINode(PHINode &PHI) {
705 VectorType *VT = dyn_cast<VectorType>(PHI.getType());
709 unsigned NumElems = VT->getNumElements();
710 IRBuilder<> Builder(&PHI);
712 Res.resize(NumElems);
714 unsigned NumOps = PHI.getNumOperands();
715 for (unsigned I = 0; I < NumElems; ++I)
716 Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
717 PHI.getName() + ".i" + Twine(I));
719 for (unsigned I = 0; I < NumOps; ++I) {
720 Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
721 BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
722 for (unsigned J = 0; J < NumElems; ++J)
723 cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
729 bool ScalarizerVisitor::visitLoadInst(LoadInst &LI) {
730 if (!ScalarizeLoadStore)
736 if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout,
737 LI.getModule()->getDataLayout()))
740 unsigned NumElems = Layout.VecTy->getNumElements();
741 IRBuilder<> Builder(&LI);
742 Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
744 Res.resize(NumElems);
746 for (unsigned I = 0; I < NumElems; ++I)
747 Res[I] = Builder.CreateAlignedLoad(Ptr[I], Layout.getElemAlign(I),
748 LI.getName() + ".i" + Twine(I));
753 bool ScalarizerVisitor::visitStoreInst(StoreInst &SI) {
754 if (!ScalarizeLoadStore)
760 Value *FullValue = SI.getValueOperand();
761 if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout,
762 SI.getModule()->getDataLayout()))
765 unsigned NumElems = Layout.VecTy->getNumElements();
766 IRBuilder<> Builder(&SI);
767 Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
768 Scatterer Val = scatter(&SI, FullValue);
771 Stores.resize(NumElems);
772 for (unsigned I = 0; I < NumElems; ++I) {
773 unsigned Align = Layout.getElemAlign(I);
774 Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
776 transferMetadata(&SI, Stores);
780 bool ScalarizerVisitor::visitCallInst(CallInst &CI) {
781 return splitCall(CI);
784 // Delete the instructions that we scalarized. If a full vector result
785 // is still needed, recreate it using InsertElements.
786 bool ScalarizerVisitor::finish() {
787 // The presence of data in Gathered or Scattered indicates changes
788 // made to the Function.
789 if (Gathered.empty() && Scattered.empty())
791 for (const auto &GMI : Gathered) {
792 Instruction *Op = GMI.first;
793 ValueVector &CV = *GMI.second;
794 if (!Op->use_empty()) {
795 // The value is still needed, so recreate it using a series of
797 Type *Ty = Op->getType();
798 Value *Res = UndefValue::get(Ty);
799 BasicBlock *BB = Op->getParent();
800 unsigned Count = Ty->getVectorNumElements();
801 IRBuilder<> Builder(Op);
802 if (isa<PHINode>(Op))
803 Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
804 for (unsigned I = 0; I < Count; ++I)
805 Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
806 Op->getName() + ".upto" + Twine(I));
808 Op->replaceAllUsesWith(Res);
810 Op->eraseFromParent();
817 PreservedAnalyses ScalarizerPass::run(Function &F, FunctionAnalysisManager &AM) {
818 Module &M = *F.getParent();
819 unsigned ParallelLoopAccessMDKind =
820 M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
821 ScalarizerVisitor Impl(ParallelLoopAccessMDKind);
822 bool Changed = Impl.visit(F);
823 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();