lib/Transforms/Scalar/Scalarizer.cpp

   1 //===- Scalarizer.cpp - Scalarize vector operations -----------------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This pass converts vector operations into scalar operations, in order
  10 // to expose optimization opportunities on the individual scalar operations.
  11 // It is mainly intended for targets that do not have vector units, but it
  12 // may also be useful for revectorizing code to different vector widths.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #include "llvm/ADT/PostOrderIterator.h"
  17 #include "llvm/ADT/SmallVector.h"
  18 #include "llvm/ADT/Twine.h"
  19 #include "llvm/Analysis/VectorUtils.h"
  20 #include "llvm/IR/Argument.h"
  21 #include "llvm/IR/BasicBlock.h"
  22 #include "llvm/IR/Constants.h"
  23 #include "llvm/IR/DataLayout.h"
  24 #include "llvm/IR/DerivedTypes.h"
  25 #include "llvm/IR/Function.h"
  26 #include "llvm/IR/IRBuilder.h"
  27 #include "llvm/IR/InstVisitor.h"
  28 #include "llvm/IR/InstrTypes.h"
  29 #include "llvm/IR/Instruction.h"
  30 #include "llvm/IR/Instructions.h"
  31 #include "llvm/IR/Intrinsics.h"
  32 #include "llvm/IR/LLVMContext.h"
  33 #include "llvm/IR/Module.h"
  34 #include "llvm/IR/Type.h"
  35 #include "llvm/IR/Value.h"
  36 #include "llvm/Pass.h"
  37 #include "llvm/Support/Casting.h"
  38 #include "llvm/Support/MathExtras.h"
  39 #include "llvm/Support/Options.h"
  40 #include "llvm/Transforms/Scalar.h"
  41 #include "llvm/Transforms/Scalar/Scalarizer.h"
  42 #include <cassert>
  43 #include <cstdint>
  44 #include <iterator>
  45 #include <map>
  46 #include <utility>
  47
  48 using namespace llvm;
  49
  50 #define DEBUG_TYPE "scalarizer"
  51
  52 // This is disabled by default because having separate loads and stores
  53 // makes it more likely that the -combiner-alias-analysis limits will be
  54 // reached.
  55 static cl::opt<bool>
  56     ScalarizeLoadStore("scalarize-load-store", cl::init(false), cl::Hidden,
  57                        cl::desc("Allow the scalarizer pass to scalarize loads and store"));
  58
  59 namespace {
  60
  61 // Used to store the scattered form of a vector.
  62 using ValueVector = SmallVector<Value *, 8>;
  63
  64 // Used to map a vector Value to its scattered form.  We use std::map
  65 // because we want iterators to persist across insertion and because the
  66 // values are relatively large.
  67 using ScatterMap = std::map<Value *, ValueVector>;
  68
  69 // Lists Instructions that have been replaced with scalar implementations,
  70 // along with a pointer to their scattered forms.
  71 using GatherList = SmallVector<std::pair<Instruction *, ValueVector *>, 16>;
  72
  73 // Provides a very limited vector-like interface for lazily accessing one
  74 // component of a scattered vector or vector pointer.
  75 class Scatterer {
  76 public:
  77   Scatterer() = default;
  78
  79   // Scatter V into Size components.  If new instructions are needed,
  80   // insert them before BBI in BB.  If Cache is nonnull, use it to cache
  81   // the results.
  82   Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
  83             ValueVector *cachePtr = nullptr);
  84
  85   // Return component I, creating a new Value for it if necessary.
  86   Value *operator[](unsigned I);
  87
  88   // Return the number of components.
  89   unsigned size() const { return Size; }
  90
  91 private:
  92   BasicBlock *BB;
  93   BasicBlock::iterator BBI;
  94   Value *V;
  95   ValueVector *CachePtr;
  96   PointerType *PtrTy;
  97   ValueVector Tmp;
  98   unsigned Size;
  99 };
 100
 101 // FCmpSpliiter(FCI)(Builder, X, Y, Name) uses Builder to create an FCmp
 102 // called Name that compares X and Y in the same way as FCI.
 103 struct FCmpSplitter {
 104   FCmpSplitter(FCmpInst &fci) : FCI(fci) {}
 105
 106   Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
 107                     const Twine &Name) const {
 108     return Builder.CreateFCmp(FCI.getPredicate(), Op0, Op1, Name);
 109   }
 110
 111   FCmpInst &FCI;
 112 };
 113
 114 // ICmpSpliiter(ICI)(Builder, X, Y, Name) uses Builder to create an ICmp
 115 // called Name that compares X and Y in the same way as ICI.
 116 struct ICmpSplitter {
 117   ICmpSplitter(ICmpInst &ici) : ICI(ici) {}
 118
 119   Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
 120                     const Twine &Name) const {
 121     return Builder.CreateICmp(ICI.getPredicate(), Op0, Op1, Name);
 122   }
 123
 124   ICmpInst &ICI;
 125 };
 126
 127 // UnarySpliiter(UO)(Builder, X, Name) uses Builder to create
 128 // a unary operator like UO called Name with operand X.
 129 struct UnarySplitter {
 130   UnarySplitter(UnaryOperator &uo) : UO(uo) {}
 131
 132   Value *operator()(IRBuilder<> &Builder, Value *Op, const Twine &Name) const {
 133     return Builder.CreateUnOp(UO.getOpcode(), Op, Name);
 134   }
 135
 136   UnaryOperator &UO;
 137 };
 138
 139 // BinarySpliiter(BO)(Builder, X, Y, Name) uses Builder to create
 140 // a binary operator like BO called Name with operands X and Y.
 141 struct BinarySplitter {
 142   BinarySplitter(BinaryOperator &bo) : BO(bo) {}
 143
 144   Value *operator()(IRBuilder<> &Builder, Value *Op0, Value *Op1,
 145                     const Twine &Name) const {
 146     return Builder.CreateBinOp(BO.getOpcode(), Op0, Op1, Name);
 147   }
 148
 149   BinaryOperator &BO;
 150 };
 151
 152 // Information about a load or store that we're scalarizing.
 153 struct VectorLayout {
 154   VectorLayout() = default;
 155
 156   // Return the alignment of element I.
 157   uint64_t getElemAlign(unsigned I) {
 158     return MinAlign(VecAlign, I * ElemSize);
 159   }
 160
 161   // The type of the vector.
 162   VectorType *VecTy = nullptr;
 163
 164   // The type of each element.
 165   Type *ElemTy = nullptr;
 166
 167   // The alignment of the vector.
 168   uint64_t VecAlign = 0;
 169
 170   // The size of each element.
 171   uint64_t ElemSize = 0;
 172 };
 173
 174 class ScalarizerVisitor : public InstVisitor<ScalarizerVisitor, bool> {
 175 public:
 176   ScalarizerVisitor(unsigned ParallelLoopAccessMDKind)
 177     : ParallelLoopAccessMDKind(ParallelLoopAccessMDKind) {
 178   }
 179
 180   bool visit(Function &F);
 181
 182   // InstVisitor methods.  They return true if the instruction was scalarized,
 183   // false if nothing changed.
 184   bool visitInstruction(Instruction &I) { return false; }
 185   bool visitSelectInst(SelectInst &SI);
 186   bool visitICmpInst(ICmpInst &ICI);
 187   bool visitFCmpInst(FCmpInst &FCI);
 188   bool visitUnaryOperator(UnaryOperator &UO);
 189   bool visitBinaryOperator(BinaryOperator &BO);
 190   bool visitGetElementPtrInst(GetElementPtrInst &GEPI);
 191   bool visitCastInst(CastInst &CI);
 192   bool visitBitCastInst(BitCastInst &BCI);
 193   bool visitShuffleVectorInst(ShuffleVectorInst &SVI);
 194   bool visitPHINode(PHINode &PHI);
 195   bool visitLoadInst(LoadInst &LI);
 196   bool visitStoreInst(StoreInst &SI);
 197   bool visitCallInst(CallInst &ICI);
 198
 199 private:
 200   Scatterer scatter(Instruction *Point, Value *V);
 201   void gather(Instruction *Op, const ValueVector &CV);
 202   bool canTransferMetadata(unsigned Kind);
 203   void transferMetadataAndIRFlags(Instruction *Op, const ValueVector &CV);
 204   bool getVectorLayout(Type *Ty, unsigned Alignment, VectorLayout &Layout,
 205                        const DataLayout &DL);
 206   bool finish();
 207
 208   template<typename T> bool splitUnary(Instruction &, const T &);
 209   template<typename T> bool splitBinary(Instruction &, const T &);
 210
 211   bool splitCall(CallInst &CI);
 212
 213   ScatterMap Scattered;
 214   GatherList Gathered;
 215
 216   unsigned ParallelLoopAccessMDKind;
 217 };
 218
 219 class ScalarizerLegacyPass : public FunctionPass {
 220 public:
 221   static char ID;
 222
 223   ScalarizerLegacyPass() : FunctionPass(ID) {
 224     initializeScalarizerLegacyPassPass(*PassRegistry::getPassRegistry());
 225   }
 226
 227   bool runOnFunction(Function &F) override;
 228 };
 229
 230 } // end anonymous namespace
 231
 232 char ScalarizerLegacyPass::ID = 0;
 233 INITIALIZE_PASS_BEGIN(ScalarizerLegacyPass, "scalarizer",
 234                       "Scalarize vector operations", false, false)
 235 INITIALIZE_PASS_END(ScalarizerLegacyPass, "scalarizer",
 236                     "Scalarize vector operations", false, false)
 237
 238 Scatterer::Scatterer(BasicBlock *bb, BasicBlock::iterator bbi, Value *v,
 239                      ValueVector *cachePtr)
 240   : BB(bb), BBI(bbi), V(v), CachePtr(cachePtr) {
 241   Type *Ty = V->getType();
 242   PtrTy = dyn_cast<PointerType>(Ty);
 243   if (PtrTy)
 244     Ty = PtrTy->getElementType();
 245   Size = Ty->getVectorNumElements();
 246   if (!CachePtr)
 247     Tmp.resize(Size, nullptr);
 248   else if (CachePtr->empty())
 249     CachePtr->resize(Size, nullptr);
 250   else
 251     assert(Size == CachePtr->size() && "Inconsistent vector sizes");
 252 }
 253
 254 // Return component I, creating a new Value for it if necessary.
 255 Value *Scatterer::operator[](unsigned I) {
 256   ValueVector &CV = (CachePtr ? *CachePtr : Tmp);
 257   // Try to reuse a previous value.
 258   if (CV[I])
 259     return CV[I];
 260   IRBuilder<> Builder(BB, BBI);
 261   if (PtrTy) {
 262     Type *ElTy = PtrTy->getElementType()->getVectorElementType();
 263     if (!CV[0]) {
 264       Type *NewPtrTy = PointerType::get(ElTy, PtrTy->getAddressSpace());
 265       CV[0] = Builder.CreateBitCast(V, NewPtrTy, V->getName() + ".i0");
 266     }
 267     if (I != 0)
 268       CV[I] = Builder.CreateConstGEP1_32(ElTy, CV[0], I,
 269                                          V->getName() + ".i" + Twine(I));
 270   } else {
 271     // Search through a chain of InsertElementInsts looking for element I.
 272     // Record other elements in the cache.  The new V is still suitable
 273     // for all uncached indices.
 274     while (true) {
 275       InsertElementInst *Insert = dyn_cast<InsertElementInst>(V);
 276       if (!Insert)
 277         break;
 278       ConstantInt *Idx = dyn_cast<ConstantInt>(Insert->getOperand(2));
 279       if (!Idx)
 280         break;
 281       unsigned J = Idx->getZExtValue();
 282       V = Insert->getOperand(0);
 283       if (I == J) {
 284         CV[J] = Insert->getOperand(1);
 285         return CV[J];
 286       } else if (!CV[J]) {
 287         // Only cache the first entry we find for each index we're not actively
 288         // searching for. This prevents us from going too far up the chain and
 289         // caching incorrect entries.
 290         CV[J] = Insert->getOperand(1);
 291       }
 292     }
 293     CV[I] = Builder.CreateExtractElement(V, Builder.getInt32(I),
 294                                          V->getName() + ".i" + Twine(I));
 295   }
 296   return CV[I];
 297 }
 298
 299 bool ScalarizerLegacyPass::runOnFunction(Function &F) {
 300   if (skipFunction(F))
 301     return false;
 302
 303   Module &M = *F.getParent();
 304   unsigned ParallelLoopAccessMDKind =
 305       M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
 306   ScalarizerVisitor Impl(ParallelLoopAccessMDKind);
 307   return Impl.visit(F);
 308 }
 309
 310 FunctionPass *llvm::createScalarizerPass() {
 311   return new ScalarizerLegacyPass();
 312 }
 313
 314 bool ScalarizerVisitor::visit(Function &F) {
 315   assert(Gathered.empty() && Scattered.empty());
 316
 317   // To ensure we replace gathered components correctly we need to do an ordered
 318   // traversal of the basic blocks in the function.
 319   ReversePostOrderTraversal<BasicBlock *> RPOT(&F.getEntryBlock());
 320   for (BasicBlock *BB : RPOT) {
 321     for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
 322       Instruction *I = &*II;
 323       bool Done = InstVisitor::visit(I);
 324       ++II;
 325       if (Done && I->getType()->isVoidTy())
 326         I->eraseFromParent();
 327     }
 328   }
 329   return finish();
 330 }
 331
 332 // Return a scattered form of V that can be accessed by Point.  V must be a
 333 // vector or a pointer to a vector.
 334 Scatterer ScalarizerVisitor::scatter(Instruction *Point, Value *V) {
 335   if (Argument *VArg = dyn_cast<Argument>(V)) {
 336     // Put the scattered form of arguments in the entry block,
 337     // so that it can be used everywhere.
 338     Function *F = VArg->getParent();
 339     BasicBlock *BB = &F->getEntryBlock();
 340     return Scatterer(BB, BB->begin(), V, &Scattered[V]);
 341   }
 342   if (Instruction *VOp = dyn_cast<Instruction>(V)) {
 343     // Put the scattered form of an instruction directly after the
 344     // instruction.
 345     BasicBlock *BB = VOp->getParent();
 346     return Scatterer(BB, std::next(BasicBlock::iterator(VOp)),
 347                      V, &Scattered[V]);
 348   }
 349   // In the fallback case, just put the scattered before Point and
 350   // keep the result local to Point.
 351   return Scatterer(Point->getParent(), Point->getIterator(), V);
 352 }
 353
 354 // Replace Op with the gathered form of the components in CV.  Defer the
 355 // deletion of Op and creation of the gathered form to the end of the pass,
 356 // so that we can avoid creating the gathered form if all uses of Op are
 357 // replaced with uses of CV.
 358 void ScalarizerVisitor::gather(Instruction *Op, const ValueVector &CV) {
 359   // Since we're not deleting Op yet, stub out its operands, so that it
 360   // doesn't make anything live unnecessarily.
 361   for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I)
 362     Op->setOperand(I, UndefValue::get(Op->getOperand(I)->getType()));
 363
 364   transferMetadataAndIRFlags(Op, CV);
 365
 366   // If we already have a scattered form of Op (created from ExtractElements
 367   // of Op itself), replace them with the new form.
 368   ValueVector &SV = Scattered[Op];
 369   if (!SV.empty()) {
 370     for (unsigned I = 0, E = SV.size(); I != E; ++I) {
 371       Value *V = SV[I];
 372       if (V == nullptr)
 373         continue;
 374
 375       Instruction *Old = cast<Instruction>(V);
 376       CV[I]->takeName(Old);
 377       Old->replaceAllUsesWith(CV[I]);
 378       Old->eraseFromParent();
 379     }
 380   }
 381   SV = CV;
 382   Gathered.push_back(GatherList::value_type(Op, &SV));
 383 }
 384
 385 // Return true if it is safe to transfer the given metadata tag from
 386 // vector to scalar instructions.
 387 bool ScalarizerVisitor::canTransferMetadata(unsigned Tag) {
 388   return (Tag == LLVMContext::MD_tbaa
 389           || Tag == LLVMContext::MD_fpmath
 390           || Tag == LLVMContext::MD_tbaa_struct
 391           || Tag == LLVMContext::MD_invariant_load
 392           || Tag == LLVMContext::MD_alias_scope
 393           || Tag == LLVMContext::MD_noalias
 394           || Tag == ParallelLoopAccessMDKind
 395           || Tag == LLVMContext::MD_access_group);
 396 }
 397
 398 // Transfer metadata from Op to the instructions in CV if it is known
 399 // to be safe to do so.
 400 void ScalarizerVisitor::transferMetadataAndIRFlags(Instruction *Op,
 401                                                    const ValueVector &CV) {
 402   SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
 403   Op->getAllMetadataOtherThanDebugLoc(MDs);
 404   for (unsigned I = 0, E = CV.size(); I != E; ++I) {
 405     if (Instruction *New = dyn_cast<Instruction>(CV[I])) {
 406       for (const auto &MD : MDs)
 407         if (canTransferMetadata(MD.first))
 408           New->setMetadata(MD.first, MD.second);
 409       New->copyIRFlags(Op);
 410       if (Op->getDebugLoc() && !New->getDebugLoc())
 411         New->setDebugLoc(Op->getDebugLoc());
 412     }
 413   }
 414 }
 415
 416 // Try to fill in Layout from Ty, returning true on success.  Alignment is
 417 // the alignment of the vector, or 0 if the ABI default should be used.
 418 bool ScalarizerVisitor::getVectorLayout(Type *Ty, unsigned Alignment,
 419                                  VectorLayout &Layout, const DataLayout &DL) {
 420   // Make sure we're dealing with a vector.
 421   Layout.VecTy = dyn_cast<VectorType>(Ty);
 422   if (!Layout.VecTy)
 423     return false;
 424
 425   // Check that we're dealing with full-byte elements.
 426   Layout.ElemTy = Layout.VecTy->getElementType();
 427   if (!DL.typeSizeEqualsStoreSize(Layout.ElemTy))
 428     return false;
 429
 430   if (Alignment)
 431     Layout.VecAlign = Alignment;
 432   else
 433     Layout.VecAlign = DL.getABITypeAlignment(Layout.VecTy);
 434   Layout.ElemSize = DL.getTypeStoreSize(Layout.ElemTy);
 435   return true;
 436 }
 437
 438 // Scalarize one-operand instruction I, using Split(Builder, X, Name)
 439 // to create an instruction like I with operand X and name Name.
 440 template<typename Splitter>
 441 bool ScalarizerVisitor::splitUnary(Instruction &I, const Splitter &Split) {
 442   VectorType *VT = dyn_cast<VectorType>(I.getType());
 443   if (!VT)
 444     return false;
 445
 446   unsigned NumElems = VT->getNumElements();
 447   IRBuilder<> Builder(&I);
 448   Scatterer Op = scatter(&I, I.getOperand(0));
 449   assert(Op.size() == NumElems && "Mismatched unary operation");
 450   ValueVector Res;
 451   Res.resize(NumElems);
 452   for (unsigned Elem = 0; Elem < NumElems; ++Elem)
 453     Res[Elem] = Split(Builder, Op[Elem], I.getName() + ".i" + Twine(Elem));
 454   gather(&I, Res);
 455   return true;
 456 }
 457
 458 // Scalarize two-operand instruction I, using Split(Builder, X, Y, Name)
 459 // to create an instruction like I with operands X and Y and name Name.
 460 template<typename Splitter>
 461 bool ScalarizerVisitor::splitBinary(Instruction &I, const Splitter &Split) {
 462   VectorType *VT = dyn_cast<VectorType>(I.getType());
 463   if (!VT)
 464     return false;
 465
 466   unsigned NumElems = VT->getNumElements();
 467   IRBuilder<> Builder(&I);
 468   Scatterer Op0 = scatter(&I, I.getOperand(0));
 469   Scatterer Op1 = scatter(&I, I.getOperand(1));
 470   assert(Op0.size() == NumElems && "Mismatched binary operation");
 471   assert(Op1.size() == NumElems && "Mismatched binary operation");
 472   ValueVector Res;
 473   Res.resize(NumElems);
 474   for (unsigned Elem = 0; Elem < NumElems; ++Elem)
 475     Res[Elem] = Split(Builder, Op0[Elem], Op1[Elem],
 476                       I.getName() + ".i" + Twine(Elem));
 477   gather(&I, Res);
 478   return true;
 479 }
 480
 481 static bool isTriviallyScalariable(Intrinsic::ID ID) {
 482   return isTriviallyVectorizable(ID);
 483 }
 484
 485 // All of the current scalarizable intrinsics only have one mangled type.
 486 static Function *getScalarIntrinsicDeclaration(Module *M,
 487                                                Intrinsic::ID ID,
 488                                                VectorType *Ty) {
 489   return Intrinsic::getDeclaration(M, ID, { Ty->getScalarType() });
 490 }
 491
 492 /// If a call to a vector typed intrinsic function, split into a scalar call per
 493 /// element if possible for the intrinsic.
 494 bool ScalarizerVisitor::splitCall(CallInst &CI) {
 495   VectorType *VT = dyn_cast<VectorType>(CI.getType());
 496   if (!VT)
 497     return false;
 498
 499   Function *F = CI.getCalledFunction();
 500   if (!F)
 501     return false;
 502
 503   Intrinsic::ID ID = F->getIntrinsicID();
 504   if (ID == Intrinsic::not_intrinsic || !isTriviallyScalariable(ID))
 505     return false;
 506
 507   unsigned NumElems = VT->getNumElements();
 508   unsigned NumArgs = CI.getNumArgOperands();
 509
 510   ValueVector ScalarOperands(NumArgs);
 511   SmallVector<Scatterer, 8> Scattered(NumArgs);
 512
 513   Scattered.resize(NumArgs);
 514
 515   // Assumes that any vector type has the same number of elements as the return
 516   // vector type, which is true for all current intrinsics.
 517   for (unsigned I = 0; I != NumArgs; ++I) {
 518     Value *OpI = CI.getOperand(I);
 519     if (OpI->getType()->isVectorTy()) {
 520       Scattered[I] = scatter(&CI, OpI);
 521       assert(Scattered[I].size() == NumElems && "mismatched call operands");
 522     } else {
 523       ScalarOperands[I] = OpI;
 524     }
 525   }
 526
 527   ValueVector Res(NumElems);
 528   ValueVector ScalarCallOps(NumArgs);
 529
 530   Function *NewIntrin = getScalarIntrinsicDeclaration(F->getParent(), ID, VT);
 531   IRBuilder<> Builder(&CI);
 532
 533   // Perform actual scalarization, taking care to preserve any scalar operands.
 534   for (unsigned Elem = 0; Elem < NumElems; ++Elem) {
 535     ScalarCallOps.clear();
 536
 537     for (unsigned J = 0; J != NumArgs; ++J) {
 538       if (hasVectorInstrinsicScalarOpd(ID, J))
 539         ScalarCallOps.push_back(ScalarOperands[J]);
 540       else
 541         ScalarCallOps.push_back(Scattered[J][Elem]);
 542     }
 543
 544     Res[Elem] = Builder.CreateCall(NewIntrin, ScalarCallOps,
 545                                    CI.getName() + ".i" + Twine(Elem));
 546   }
 547
 548   gather(&CI, Res);
 549   return true;
 550 }
 551
 552 bool ScalarizerVisitor::visitSelectInst(SelectInst &SI) {
 553   VectorType *VT = dyn_cast<VectorType>(SI.getType());
 554   if (!VT)
 555     return false;
 556
 557   unsigned NumElems = VT->getNumElements();
 558   IRBuilder<> Builder(&SI);
 559   Scatterer Op1 = scatter(&SI, SI.getOperand(1));
 560   Scatterer Op2 = scatter(&SI, SI.getOperand(2));
 561   assert(Op1.size() == NumElems && "Mismatched select");
 562   assert(Op2.size() == NumElems && "Mismatched select");
 563   ValueVector Res;
 564   Res.resize(NumElems);
 565
 566   if (SI.getOperand(0)->getType()->isVectorTy()) {
 567     Scatterer Op0 = scatter(&SI, SI.getOperand(0));
 568     assert(Op0.size() == NumElems && "Mismatched select");
 569     for (unsigned I = 0; I < NumElems; ++I)
 570       Res[I] = Builder.CreateSelect(Op0[I], Op1[I], Op2[I],
 571                                     SI.getName() + ".i" + Twine(I));
 572   } else {
 573     Value *Op0 = SI.getOperand(0);
 574     for (unsigned I = 0; I < NumElems; ++I)
 575       Res[I] = Builder.CreateSelect(Op0, Op1[I], Op2[I],
 576                                     SI.getName() + ".i" + Twine(I));
 577   }
 578   gather(&SI, Res);
 579   return true;
 580 }
 581
 582 bool ScalarizerVisitor::visitICmpInst(ICmpInst &ICI) {
 583   return splitBinary(ICI, ICmpSplitter(ICI));
 584 }
 585
 586 bool ScalarizerVisitor::visitFCmpInst(FCmpInst &FCI) {
 587   return splitBinary(FCI, FCmpSplitter(FCI));
 588 }
 589
 590 bool ScalarizerVisitor::visitUnaryOperator(UnaryOperator &UO) {
 591   return splitUnary(UO, UnarySplitter(UO));
 592 }
 593
 594 bool ScalarizerVisitor::visitBinaryOperator(BinaryOperator &BO) {
 595   return splitBinary(BO, BinarySplitter(BO));
 596 }
 597
 598 bool ScalarizerVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
 599   VectorType *VT = dyn_cast<VectorType>(GEPI.getType());
 600   if (!VT)
 601     return false;
 602
 603   IRBuilder<> Builder(&GEPI);
 604   unsigned NumElems = VT->getNumElements();
 605   unsigned NumIndices = GEPI.getNumIndices();
 606
 607   // The base pointer might be scalar even if it's a vector GEP. In those cases,
 608   // splat the pointer into a vector value, and scatter that vector.
 609   Value *Op0 = GEPI.getOperand(0);
 610   if (!Op0->getType()->isVectorTy())
 611     Op0 = Builder.CreateVectorSplat(NumElems, Op0);
 612   Scatterer Base = scatter(&GEPI, Op0);
 613
 614   SmallVector<Scatterer, 8> Ops;
 615   Ops.resize(NumIndices);
 616   for (unsigned I = 0; I < NumIndices; ++I) {
 617     Value *Op = GEPI.getOperand(I + 1);
 618
 619     // The indices might be scalars even if it's a vector GEP. In those cases,
 620     // splat the scalar into a vector value, and scatter that vector.
 621     if (!Op->getType()->isVectorTy())
 622       Op = Builder.CreateVectorSplat(NumElems, Op);
 623
 624     Ops[I] = scatter(&GEPI, Op);
 625   }
 626
 627   ValueVector Res;
 628   Res.resize(NumElems);
 629   for (unsigned I = 0; I < NumElems; ++I) {
 630     SmallVector<Value *, 8> Indices;
 631     Indices.resize(NumIndices);
 632     for (unsigned J = 0; J < NumIndices; ++J)
 633       Indices[J] = Ops[J][I];
 634     Res[I] = Builder.CreateGEP(GEPI.getSourceElementType(), Base[I], Indices,
 635                                GEPI.getName() + ".i" + Twine(I));
 636     if (GEPI.isInBounds())
 637       if (GetElementPtrInst *NewGEPI = dyn_cast<GetElementPtrInst>(Res[I]))
 638         NewGEPI->setIsInBounds();
 639   }
 640   gather(&GEPI, Res);
 641   return true;
 642 }
 643
 644 bool ScalarizerVisitor::visitCastInst(CastInst &CI) {
 645   VectorType *VT = dyn_cast<VectorType>(CI.getDestTy());
 646   if (!VT)
 647     return false;
 648
 649   unsigned NumElems = VT->getNumElements();
 650   IRBuilder<> Builder(&CI);
 651   Scatterer Op0 = scatter(&CI, CI.getOperand(0));
 652   assert(Op0.size() == NumElems && "Mismatched cast");
 653   ValueVector Res;
 654   Res.resize(NumElems);
 655   for (unsigned I = 0; I < NumElems; ++I)
 656     Res[I] = Builder.CreateCast(CI.getOpcode(), Op0[I], VT->getElementType(),
 657                                 CI.getName() + ".i" + Twine(I));
 658   gather(&CI, Res);
 659   return true;
 660 }
 661
 662 bool ScalarizerVisitor::visitBitCastInst(BitCastInst &BCI) {
 663   VectorType *DstVT = dyn_cast<VectorType>(BCI.getDestTy());
 664   VectorType *SrcVT = dyn_cast<VectorType>(BCI.getSrcTy());
 665   if (!DstVT || !SrcVT)
 666     return false;
 667
 668   unsigned DstNumElems = DstVT->getNumElements();
 669   unsigned SrcNumElems = SrcVT->getNumElements();
 670   IRBuilder<> Builder(&BCI);
 671   Scatterer Op0 = scatter(&BCI, BCI.getOperand(0));
 672   ValueVector Res;
 673   Res.resize(DstNumElems);
 674
 675   if (DstNumElems == SrcNumElems) {
 676     for (unsigned I = 0; I < DstNumElems; ++I)
 677       Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
 678                                      BCI.getName() + ".i" + Twine(I));
 679   } else if (DstNumElems > SrcNumElems) {
 680     // <M x t1> -> <N*M x t2>.  Convert each t1 to <N x t2> and copy the
 681     // individual elements to the destination.
 682     unsigned FanOut = DstNumElems / SrcNumElems;
 683     Type *MidTy = VectorType::get(DstVT->getElementType(), FanOut);
 684     unsigned ResI = 0;
 685     for (unsigned Op0I = 0; Op0I < SrcNumElems; ++Op0I) {
 686       Value *V = Op0[Op0I];
 687       Instruction *VI;
 688       // Look through any existing bitcasts before converting to <N x t2>.
 689       // In the best case, the resulting conversion might be a no-op.
 690       while ((VI = dyn_cast<Instruction>(V)) &&
 691              VI->getOpcode() == Instruction::BitCast)
 692         V = VI->getOperand(0);
 693       V = Builder.CreateBitCast(V, MidTy, V->getName() + ".cast");
 694       Scatterer Mid = scatter(&BCI, V);
 695       for (unsigned MidI = 0; MidI < FanOut; ++MidI)
 696         Res[ResI++] = Mid[MidI];
 697     }
 698   } else {
 699     // <N*M x t1> -> <M x t2>.  Convert each group of <N x t1> into a t2.
 700     unsigned FanIn = SrcNumElems / DstNumElems;
 701     Type *MidTy = VectorType::get(SrcVT->getElementType(), FanIn);
 702     unsigned Op0I = 0;
 703     for (unsigned ResI = 0; ResI < DstNumElems; ++ResI) {
 704       Value *V = UndefValue::get(MidTy);
 705       for (unsigned MidI = 0; MidI < FanIn; ++MidI)
 706         V = Builder.CreateInsertElement(V, Op0[Op0I++], Builder.getInt32(MidI),
 707                                         BCI.getName() + ".i" + Twine(ResI)
 708                                         + ".upto" + Twine(MidI));
 709       Res[ResI] = Builder.CreateBitCast(V, DstVT->getElementType(),
 710                                         BCI.getName() + ".i" + Twine(ResI));
 711     }
 712   }
 713   gather(&BCI, Res);
 714   return true;
 715 }
 716
 717 bool ScalarizerVisitor::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
 718   VectorType *VT = dyn_cast<VectorType>(SVI.getType());
 719   if (!VT)
 720     return false;
 721
 722   unsigned NumElems = VT->getNumElements();
 723   Scatterer Op0 = scatter(&SVI, SVI.getOperand(0));
 724   Scatterer Op1 = scatter(&SVI, SVI.getOperand(1));
 725   ValueVector Res;
 726   Res.resize(NumElems);
 727
 728   for (unsigned I = 0; I < NumElems; ++I) {
 729     int Selector = SVI.getMaskValue(I);
 730     if (Selector < 0)
 731       Res[I] = UndefValue::get(VT->getElementType());
 732     else if (unsigned(Selector) < Op0.size())
 733       Res[I] = Op0[Selector];
 734     else
 735       Res[I] = Op1[Selector - Op0.size()];
 736   }
 737   gather(&SVI, Res);
 738   return true;
 739 }
 740
 741 bool ScalarizerVisitor::visitPHINode(PHINode &PHI) {
 742   VectorType *VT = dyn_cast<VectorType>(PHI.getType());
 743   if (!VT)
 744     return false;
 745
 746   unsigned NumElems = VT->getNumElements();
 747   IRBuilder<> Builder(&PHI);
 748   ValueVector Res;
 749   Res.resize(NumElems);
 750
 751   unsigned NumOps = PHI.getNumOperands();
 752   for (unsigned I = 0; I < NumElems; ++I)
 753     Res[I] = Builder.CreatePHI(VT->getElementType(), NumOps,
 754                                PHI.getName() + ".i" + Twine(I));
 755
 756   for (unsigned I = 0; I < NumOps; ++I) {
 757     Scatterer Op = scatter(&PHI, PHI.getIncomingValue(I));
 758     BasicBlock *IncomingBlock = PHI.getIncomingBlock(I);
 759     for (unsigned J = 0; J < NumElems; ++J)
 760       cast<PHINode>(Res[J])->addIncoming(Op[J], IncomingBlock);
 761   }
 762   gather(&PHI, Res);
 763   return true;
 764 }
 765
 766 bool ScalarizerVisitor::visitLoadInst(LoadInst &LI) {
 767   if (!ScalarizeLoadStore)
 768     return false;
 769   if (!LI.isSimple())
 770     return false;
 771
 772   VectorLayout Layout;
 773   if (!getVectorLayout(LI.getType(), LI.getAlignment(), Layout,
 774                        LI.getModule()->getDataLayout()))
 775     return false;
 776
 777   unsigned NumElems = Layout.VecTy->getNumElements();
 778   IRBuilder<> Builder(&LI);
 779   Scatterer Ptr = scatter(&LI, LI.getPointerOperand());
 780   ValueVector Res;
 781   Res.resize(NumElems);
 782
 783   for (unsigned I = 0; I < NumElems; ++I)
 784     Res[I] = Builder.CreateAlignedLoad(Layout.VecTy->getElementType(), Ptr[I],
 785                                        Layout.getElemAlign(I),
 786                                        LI.getName() + ".i" + Twine(I));
 787   gather(&LI, Res);
 788   return true;
 789 }
 790
 791 bool ScalarizerVisitor::visitStoreInst(StoreInst &SI) {
 792   if (!ScalarizeLoadStore)
 793     return false;
 794   if (!SI.isSimple())
 795     return false;
 796
 797   VectorLayout Layout;
 798   Value *FullValue = SI.getValueOperand();
 799   if (!getVectorLayout(FullValue->getType(), SI.getAlignment(), Layout,
 800                        SI.getModule()->getDataLayout()))
 801     return false;
 802
 803   unsigned NumElems = Layout.VecTy->getNumElements();
 804   IRBuilder<> Builder(&SI);
 805   Scatterer Ptr = scatter(&SI, SI.getPointerOperand());
 806   Scatterer Val = scatter(&SI, FullValue);
 807
 808   ValueVector Stores;
 809   Stores.resize(NumElems);
 810   for (unsigned I = 0; I < NumElems; ++I) {
 811     unsigned Align = Layout.getElemAlign(I);
 812     Stores[I] = Builder.CreateAlignedStore(Val[I], Ptr[I], Align);
 813   }
 814   transferMetadataAndIRFlags(&SI, Stores);
 815   return true;
 816 }
 817
 818 bool ScalarizerVisitor::visitCallInst(CallInst &CI) {
 819   return splitCall(CI);
 820 }
 821
 822 // Delete the instructions that we scalarized.  If a full vector result
 823 // is still needed, recreate it using InsertElements.
 824 bool ScalarizerVisitor::finish() {
 825   // The presence of data in Gathered or Scattered indicates changes
 826   // made to the Function.
 827   if (Gathered.empty() && Scattered.empty())
 828     return false;
 829   for (const auto &GMI : Gathered) {
 830     Instruction *Op = GMI.first;
 831     ValueVector &CV = *GMI.second;
 832     if (!Op->use_empty()) {
 833       // The value is still needed, so recreate it using a series of
 834       // InsertElements.
 835       Type *Ty = Op->getType();
 836       Value *Res = UndefValue::get(Ty);
 837       BasicBlock *BB = Op->getParent();
 838       unsigned Count = Ty->getVectorNumElements();
 839       IRBuilder<> Builder(Op);
 840       if (isa<PHINode>(Op))
 841         Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
 842       for (unsigned I = 0; I < Count; ++I)
 843         Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
 844                                           Op->getName() + ".upto" + Twine(I));
 845       Res->takeName(Op);
 846       Op->replaceAllUsesWith(Res);
 847     }
 848     Op->eraseFromParent();
 849   }
 850   Gathered.clear();
 851   Scattered.clear();
 852   return true;
 853 }
 854
 855 PreservedAnalyses ScalarizerPass::run(Function &F, FunctionAnalysisManager &AM) {
 856   Module &M = *F.getParent();
 857   unsigned ParallelLoopAccessMDKind =
 858       M.getContext().getMDKindID("llvm.mem.parallel_loop_access");
 859   ScalarizerVisitor Impl(ParallelLoopAccessMDKind);
 860   bool Changed = Impl.visit(F);
 861   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
 862 }