lib/Transforms/Scalar/DeadStoreElimination.cpp

   1 //===- DeadStoreElimination.cpp - Fast Dead Store Elimination -------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements a trivial dead store elimination that only considers
  11 // basic-block local redundant stores.
  12 //
  13 // FIXME: This should eventually be extended to be a post-dominator tree
  14 // traversal.  Doing so would be pretty trivial.
  15 //
  16 //===----------------------------------------------------------------------===//
  17
  18 #define DEBUG_TYPE "dse"
  19 #include "llvm/Transforms/Scalar.h"
  20 #include "llvm/Constants.h"
  21 #include "llvm/Function.h"
  22 #include "llvm/Instructions.h"
  23 #include "llvm/IntrinsicInst.h"
  24 #include "llvm/Pass.h"
  25 #include "llvm/ADT/SetVector.h"
  26 #include "llvm/ADT/SmallPtrSet.h"
  27 #include "llvm/ADT/Statistic.h"
  28 #include "llvm/Analysis/AliasAnalysis.h"
  29 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
  30 #include "llvm/Target/TargetData.h"
  31 #include "llvm/Transforms/Utils/Local.h"
  32 #include "llvm/Support/Compiler.h"
  33 using namespace llvm;
  34
  35 STATISTIC(NumFastStores, "Number of stores deleted");
  36 STATISTIC(NumFastOther , "Number of other instrs removed");
  37
  38 namespace {
  39   struct VISIBILITY_HIDDEN DSE : public FunctionPass {
  40     static char ID; // Pass identification, replacement for typeid
  41     DSE() : FunctionPass((intptr_t)&ID) {}
  42
  43     virtual bool runOnFunction(Function &F) {
  44       bool Changed = false;
  45       for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
  46         Changed |= runOnBasicBlock(*I);
  47       return Changed;
  48     }
  49
  50     bool runOnBasicBlock(BasicBlock &BB);
  51     bool handleFreeWithNonTrivialDependency(FreeInst* F,
  52                                             Instruction* dependency,
  53                                         SetVector<Instruction*>& possiblyDead);
  54     bool handleEndBlock(BasicBlock& BB, SetVector<Instruction*>& possiblyDead);
  55     bool RemoveUndeadPointers(Value* pointer, uint64_t killPointerSize,
  56                               BasicBlock::iterator& BBI,
  57                               SmallPtrSet<Value*, 64>& deadPointers,
  58                               SetVector<Instruction*>& possiblyDead);
  59     void DeleteDeadInstructionChains(Instruction *I,
  60                                      SetVector<Instruction*> &DeadInsts);
  61
  62     /// Find the base pointer that a pointer came from
  63     /// Because this is used to find pointers that originate
  64     /// from allocas, it is safe to ignore GEP indices, since
  65     /// either the store will be in the alloca, and thus dead,
  66     /// or beyond the end of the alloca, and thus undefined.
  67     void TranslatePointerBitCasts(Value*& v, bool zeroGepsOnly = false) {
  68       assert(isa<PointerType>(v->getType()) &&
  69              "Translating a non-pointer type?");
  70       while (true) {
  71         if (BitCastInst* C = dyn_cast<BitCastInst>(v))
  72           v = C->getOperand(0);
  73         else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(v))
  74           if (!zeroGepsOnly || G->hasAllZeroIndices()) {
  75             v = G->getOperand(0);
  76           } else {
  77             break;
  78           }
  79         else
  80           break;
  81       }
  82     }
  83
  84     // getAnalysisUsage - We require post dominance frontiers (aka Control
  85     // Dependence Graph)
  86     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  87       AU.setPreservesCFG();
  88       AU.addRequired<TargetData>();
  89       AU.addRequired<AliasAnalysis>();
  90       AU.addRequired<MemoryDependenceAnalysis>();
  91       AU.addPreserved<AliasAnalysis>();
  92       AU.addPreserved<MemoryDependenceAnalysis>();
  93     }
  94   };
  95   char DSE::ID = 0;
  96   RegisterPass<DSE> X("dse", "Dead Store Elimination");
  97 }
  98
  99 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
 100
 101 bool DSE::runOnBasicBlock(BasicBlock &BB) {
 102   MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
 103   TargetData &TD = getAnalysis<TargetData>();
 104
 105   // Record the last-seen store to this pointer
 106   DenseMap<Value*, StoreInst*> lastStore;
 107   // Record instructions possibly made dead by deleting a store
 108   SetVector<Instruction*> possiblyDead;
 109
 110   bool MadeChange = false;
 111
 112   // Do a top-down walk on the BB
 113   for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end();
 114        BBI != BBE; ++BBI) {
 115     // If we find a store or a free...
 116     if (!isa<StoreInst>(BBI) && !isa<FreeInst>(BBI))
 117       continue;
 118
 119     Value* pointer = 0;
 120     if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
 121       if (!S->isVolatile())
 122         pointer = S->getPointerOperand();
 123       else
 124         continue;
 125     } else
 126       pointer = cast<FreeInst>(BBI)->getPointerOperand();
 127
 128     TranslatePointerBitCasts(pointer, true);
 129     StoreInst*& last = lastStore[pointer];
 130     bool deletedStore = false;
 131
 132     // ... to a pointer that has been stored to before...
 133     if (last) {
 134       Instruction* dep = MD.getDependency(BBI);
 135
 136       // ... and no other memory dependencies are between them....
 137       while (dep != MemoryDependenceAnalysis::None &&
 138              dep != MemoryDependenceAnalysis::NonLocal &&
 139              isa<StoreInst>(dep)) {
 140         if (dep != last ||
 141              TD.getTypeStoreSize(last->getOperand(0)->getType()) >
 142              TD.getTypeStoreSize(BBI->getOperand(0)->getType())) {
 143           dep = MD.getDependency(BBI, dep);
 144           continue;
 145         }
 146
 147         // Remove it!
 148         MD.removeInstruction(last);
 149
 150         // DCE instructions only used to calculate that store
 151         if (Instruction* D = dyn_cast<Instruction>(last->getOperand(0)))
 152           possiblyDead.insert(D);
 153         if (Instruction* D = dyn_cast<Instruction>(last->getOperand(1)))
 154           possiblyDead.insert(D);
 155
 156         last->eraseFromParent();
 157         NumFastStores++;
 158         deletedStore = true;
 159         MadeChange = true;
 160
 161         break;
 162       }
 163     }
 164
 165     // Handle frees whose dependencies are non-trivial.
 166     if (FreeInst* F = dyn_cast<FreeInst>(BBI)) {
 167       if (!deletedStore)
 168         MadeChange |= handleFreeWithNonTrivialDependency(F,
 169                                                          MD.getDependency(F),
 170                                                          possiblyDead);
 171       // No known stores after the free
 172       last = 0;
 173     } else {
 174       // Update our most-recent-store map.
 175       last = cast<StoreInst>(BBI);
 176     }
 177   }
 178
 179   // If this block ends in a return, unwind, unreachable, and eventually
 180   // tailcall, then all allocas are dead at its end.
 181   if (BB.getTerminator()->getNumSuccessors() == 0)
 182     MadeChange |= handleEndBlock(BB, possiblyDead);
 183
 184   // Do a trivial DCE
 185   while (!possiblyDead.empty()) {
 186     Instruction *I = possiblyDead.back();
 187     possiblyDead.pop_back();
 188     DeleteDeadInstructionChains(I, possiblyDead);
 189   }
 190
 191   return MadeChange;
 192 }
 193
 194 /// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
 195 /// dependency is a store to a field of that structure
 196 bool DSE::handleFreeWithNonTrivialDependency(FreeInst* F, Instruction* dep,
 197                                        SetVector<Instruction*>& possiblyDead) {
 198   TargetData &TD = getAnalysis<TargetData>();
 199   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
 200   MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
 201
 202   if (dep == MemoryDependenceAnalysis::None ||
 203       dep == MemoryDependenceAnalysis::NonLocal)
 204     return false;
 205
 206   StoreInst* dependency = dyn_cast<StoreInst>(dep);
 207   if (!dependency)
 208     return false;
 209   else if (dependency->isVolatile())
 210     return false;
 211
 212   Value* depPointer = dependency->getPointerOperand();
 213   const Type* depType = dependency->getOperand(0)->getType();
 214   unsigned depPointerSize = TD.getTypeStoreSize(depType);
 215
 216   // Check for aliasing
 217   AliasAnalysis::AliasResult A = AA.alias(F->getPointerOperand(), ~0U,
 218                                           depPointer, depPointerSize);
 219
 220   if (A == AliasAnalysis::MustAlias) {
 221     // Remove it!
 222     MD.removeInstruction(dependency);
 223
 224     // DCE instructions only used to calculate that store
 225     if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(0)))
 226       possiblyDead.insert(D);
 227     if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(1)))
 228       possiblyDead.insert(D);
 229
 230     dependency->eraseFromParent();
 231     NumFastStores++;
 232     return true;
 233   }
 234
 235   return false;
 236 }
 237
 238 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
 239 /// function end block.  Ex:
 240 /// %A = alloca i32
 241 /// ...
 242 /// store i32 1, i32* %A
 243 /// ret void
 244 bool DSE::handleEndBlock(BasicBlock& BB,
 245                          SetVector<Instruction*>& possiblyDead) {
 246   TargetData &TD = getAnalysis<TargetData>();
 247   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
 248   MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
 249
 250   bool MadeChange = false;
 251
 252   // Pointers alloca'd in this function are dead in the end block
 253   SmallPtrSet<Value*, 64> deadPointers;
 254
 255   // Find all of the alloca'd pointers in the entry block
 256   BasicBlock *Entry = BB.getParent()->begin();
 257   for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
 258     if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
 259       deadPointers.insert(AI);
 260   for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
 261        AE = BB.getParent()->arg_end(); AI != AE; ++AI)
 262     if (AI->hasByValAttr())
 263       deadPointers.insert(AI);
 264
 265   // Scan the basic block backwards
 266   for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
 267     --BBI;
 268
 269     // If we find a store whose pointer is dead...
 270     if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
 271       if (!S->isVolatile()) {
 272         Value* pointerOperand = S->getPointerOperand();
 273         // See through pointer-to-pointer bitcasts
 274         TranslatePointerBitCasts(pointerOperand);
 275
 276         // Alloca'd pointers or byval arguments (which are functionally like
 277         // alloca's) are valid candidates for removal.
 278         if (deadPointers.count(pointerOperand)) {
 279           // Remove it!
 280           MD.removeInstruction(S);
 281
 282           // DCE instructions only used to calculate that store
 283           if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
 284             possiblyDead.insert(D);
 285           if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
 286             possiblyDead.insert(D);
 287
 288           BBI++;
 289           S->eraseFromParent();
 290           NumFastStores++;
 291           MadeChange = true;
 292         }
 293       }
 294
 295       continue;
 296
 297     // We can also remove memcpy's to local variables at the end of a function
 298     } else if (MemCpyInst* M = dyn_cast<MemCpyInst>(BBI)) {
 299       Value* dest = M->getDest();
 300       TranslatePointerBitCasts(dest);
 301
 302       if (deadPointers.count(dest)) {
 303         MD.removeInstruction(M);
 304
 305         // DCE instructions only used to calculate that memcpy
 306         if (Instruction* D = dyn_cast<Instruction>(M->getRawSource()))
 307           possiblyDead.insert(D);
 308         if (Instruction* D = dyn_cast<Instruction>(M->getLength()))
 309           possiblyDead.insert(D);
 310         if (Instruction* D = dyn_cast<Instruction>(M->getRawDest()))
 311           possiblyDead.insert(D);
 312
 313         BBI++;
 314         M->eraseFromParent();
 315         NumFastOther++;
 316         MadeChange = true;
 317
 318         continue;
 319       }
 320
 321       // Because a memcpy is also a load, we can't skip it if we didn't remove it
 322     }
 323
 324     Value* killPointer = 0;
 325     uint64_t killPointerSize = ~0UL;
 326
 327     // If we encounter a use of the pointer, it is no longer considered dead
 328     if (LoadInst* L = dyn_cast<LoadInst>(BBI)) {
 329       // However, if this load is unused, we can go ahead and remove it, and
 330       // not have to worry about it making our pointer undead!
 331       if (L->use_empty()) {
 332         MD.removeInstruction(L);
 333
 334         // DCE instructions only used to calculate that load
 335         if (Instruction* D = dyn_cast<Instruction>(L->getPointerOperand()))
 336           possiblyDead.insert(D);
 337
 338         BBI++;
 339         L->eraseFromParent();
 340         NumFastOther++;
 341         MadeChange = true;
 342         possiblyDead.remove(L);
 343
 344         continue;
 345       }
 346
 347       killPointer = L->getPointerOperand();
 348     } else if (VAArgInst* V = dyn_cast<VAArgInst>(BBI)) {
 349       killPointer = V->getOperand(0);
 350     } else if (isa<MemCpyInst>(BBI) &&
 351                isa<ConstantInt>(cast<MemCpyInst>(BBI)->getLength())) {
 352       killPointer = cast<MemCpyInst>(BBI)->getSource();
 353       killPointerSize = cast<ConstantInt>(
 354                             cast<MemCpyInst>(BBI)->getLength())->getZExtValue();
 355     } else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
 356       deadPointers.erase(A);
 357
 358       // Dead alloca's can be DCE'd when we reach them
 359       if (A->use_empty()) {
 360         MD.removeInstruction(A);
 361
 362         // DCE instructions only used to calculate that load
 363         if (Instruction* D = dyn_cast<Instruction>(A->getArraySize()))
 364           possiblyDead.insert(D);
 365
 366         BBI++;
 367         A->eraseFromParent();
 368         NumFastOther++;
 369         MadeChange = true;
 370         possiblyDead.remove(A);
 371       }
 372
 373       continue;
 374     } else if (CallSite::get(BBI).getInstruction() != 0) {
 375       // If this call does not access memory, it can't
 376       // be undeadifying any of our pointers.
 377       CallSite CS = CallSite::get(BBI);
 378       if (AA.doesNotAccessMemory(CS))
 379         continue;
 380
 381       unsigned modRef = 0;
 382       unsigned other = 0;
 383
 384       // Remove any pointers made undead by the call from the dead set
 385       std::vector<Value*> dead;
 386       for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
 387            E = deadPointers.end(); I != E; ++I) {
 388         // HACK: if we detect that our AA is imprecise, it's not
 389         // worth it to scan the rest of the deadPointers set.  Just
 390         // assume that the AA will return ModRef for everything, and
 391         // go ahead and bail.
 392         if (modRef >= 16 && other == 0) {
 393           deadPointers.clear();
 394           return MadeChange;
 395         }
 396
 397         // Get size information for the alloca
 398         unsigned pointerSize = ~0U;
 399         if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
 400           if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
 401             pointerSize = C->getZExtValue() * \
 402                           TD.getABITypeSize(A->getAllocatedType());
 403         } else {
 404           const PointerType* PT = cast<PointerType>(
 405                                                  cast<Argument>(*I)->getType());
 406           pointerSize = TD.getABITypeSize(PT->getElementType());
 407         }
 408
 409         // See if the call site touches it
 410         AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
 411
 412         if (A == AliasAnalysis::ModRef)
 413           modRef++;
 414         else
 415           other++;
 416
 417         if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
 418           dead.push_back(*I);
 419       }
 420
 421       for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
 422            I != E; ++I)
 423         deadPointers.erase(*I);
 424
 425       continue;
 426     } else {
 427       // For any non-memory-affecting non-terminators, DCE them as we reach them
 428       Instruction *CI = BBI;
 429       if (!CI->isTerminator() && CI->use_empty() && !isa<FreeInst>(CI)) {
 430
 431         // DCE instructions only used to calculate that load
 432         for (Instruction::op_iterator OI = CI->op_begin(), OE = CI->op_end();
 433              OI != OE; ++OI)
 434           if (Instruction* D = dyn_cast<Instruction>(OI))
 435             possiblyDead.insert(D);
 436
 437         BBI++;
 438         CI->eraseFromParent();
 439         NumFastOther++;
 440         MadeChange = true;
 441         possiblyDead.remove(CI);
 442
 443         continue;
 444       }
 445     }
 446
 447     if (!killPointer)
 448       continue;
 449
 450     TranslatePointerBitCasts(killPointer);
 451
 452     // Deal with undead pointers
 453     MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
 454                                        deadPointers, possiblyDead);
 455   }
 456
 457   return MadeChange;
 458 }
 459
 460 /// RemoveUndeadPointers - check for uses of a pointer that make it
 461 /// undead when scanning for dead stores to alloca's.
 462 bool DSE::RemoveUndeadPointers(Value* killPointer, uint64_t killPointerSize,
 463                                 BasicBlock::iterator& BBI,
 464                                 SmallPtrSet<Value*, 64>& deadPointers,
 465                                 SetVector<Instruction*>& possiblyDead) {
 466   TargetData &TD = getAnalysis<TargetData>();
 467   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
 468   MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
 469
 470   // If the kill pointer can be easily reduced to an alloca,
 471   // don't bother doing extraneous AA queries
 472   if (deadPointers.count(killPointer)) {
 473     deadPointers.erase(killPointer);
 474     return false;
 475   } else if (isa<GlobalValue>(killPointer)) {
 476     // A global can't be in the dead pointer set
 477     return false;
 478   }
 479
 480   bool MadeChange = false;
 481
 482   std::vector<Value*> undead;
 483
 484   for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
 485       E = deadPointers.end(); I != E; ++I) {
 486     // Get size information for the alloca
 487     unsigned pointerSize = ~0U;
 488     if (AllocaInst* A = dyn_cast<AllocaInst>(*I)) {
 489       if (ConstantInt* C = dyn_cast<ConstantInt>(A->getArraySize()))
 490         pointerSize = C->getZExtValue() * \
 491                       TD.getABITypeSize(A->getAllocatedType());
 492     } else {
 493       const PointerType* PT = cast<PointerType>(
 494                                                 cast<Argument>(*I)->getType());
 495       pointerSize = TD.getABITypeSize(PT->getElementType());
 496     }
 497
 498     // See if this pointer could alias it
 499     AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
 500                                             killPointer, killPointerSize);
 501
 502     // If it must-alias and a store, we can delete it
 503     if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
 504       StoreInst* S = cast<StoreInst>(BBI);
 505
 506       // Remove it!
 507       MD.removeInstruction(S);
 508
 509       // DCE instructions only used to calculate that store
 510       if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
 511         possiblyDead.insert(D);
 512       if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
 513         possiblyDead.insert(D);
 514
 515       BBI++;
 516       S->eraseFromParent();
 517       NumFastStores++;
 518       MadeChange = true;
 519
 520       continue;
 521
 522       // Otherwise, it is undead
 523       } else if (A != AliasAnalysis::NoAlias)
 524         undead.push_back(*I);
 525   }
 526
 527   for (std::vector<Value*>::iterator I = undead.begin(), E = undead.end();
 528        I != E; ++I)
 529       deadPointers.erase(*I);
 530
 531   return MadeChange;
 532 }
 533
 534 /// DeleteDeadInstructionChains - takes an instruction and a setvector of
 535 /// dead instructions.  If I is dead, it is erased, and its operands are
 536 /// checked for deadness.  If they are dead, they are added to the dead
 537 /// setvector.
 538 void DSE::DeleteDeadInstructionChains(Instruction *I,
 539                                       SetVector<Instruction*> &DeadInsts) {
 540   // Instruction must be dead.
 541   if (!I->use_empty() || !isInstructionTriviallyDead(I)) return;
 542
 543   // Let the memory dependence know
 544   getAnalysis<MemoryDependenceAnalysis>().removeInstruction(I);
 545
 546   // See if this made any operands dead.  We do it this way in case the
 547   // instruction uses the same operand twice.  We don't want to delete a
 548   // value then reference it.
 549   for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
 550     if (I->getOperand(i)->hasOneUse())
 551       if (Instruction* Op = dyn_cast<Instruction>(I->getOperand(i)))
 552         DeadInsts.insert(Op);      // Attempt to nuke it later.
 553
 554     I->setOperand(i, 0);         // Drop from the operand list.
 555   }
 556
 557   I->eraseFromParent();
 558   ++NumFastOther;
 559 }