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