1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===// 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 pass transforms simple global variables that never have their address 11 // taken. If obviously true, it marks read/write globals as constant, deletes 12 // variables only stored to, etc. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Transforms/IPO.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/SmallSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/ConstantFolding.h" 24 #include "llvm/Analysis/MemoryBuiltins.h" 25 #include "llvm/Analysis/TargetLibraryInfo.h" 26 #include "llvm/IR/CallSite.h" 27 #include "llvm/IR/CallingConv.h" 28 #include "llvm/IR/Constants.h" 29 #include "llvm/IR/DataLayout.h" 30 #include "llvm/IR/DerivedTypes.h" 31 #include "llvm/IR/GetElementPtrTypeIterator.h" 32 #include "llvm/IR/Instructions.h" 33 #include "llvm/IR/IntrinsicInst.h" 34 #include "llvm/IR/Module.h" 35 #include "llvm/IR/Operator.h" 36 #include "llvm/IR/ValueHandle.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/Debug.h" 39 #include "llvm/Support/ErrorHandling.h" 40 #include "llvm/Support/MathExtras.h" 41 #include "llvm/Support/raw_ostream.h" 42 #include "llvm/Transforms/Utils/CtorUtils.h" 43 #include "llvm/Transforms/Utils/GlobalStatus.h" 44 #include "llvm/Transforms/Utils/ModuleUtils.h" 45 #include <algorithm> 46 #include <deque> 47 using namespace llvm; 48 49 #define DEBUG_TYPE "globalopt" 50 51 STATISTIC(NumMarked , "Number of globals marked constant"); 52 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr"); 53 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars"); 54 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd"); 55 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them"); 56 STATISTIC(NumDeleted , "Number of globals deleted"); 57 STATISTIC(NumFnDeleted , "Number of functions deleted"); 58 STATISTIC(NumGlobUses , "Number of global uses devirtualized"); 59 STATISTIC(NumLocalized , "Number of globals localized"); 60 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans"); 61 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc"); 62 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated"); 63 STATISTIC(NumNestRemoved , "Number of nest attributes removed"); 64 STATISTIC(NumAliasesResolved, "Number of global aliases resolved"); 65 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated"); 66 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed"); 67 68 namespace { 69 struct GlobalOpt : public ModulePass { 70 void getAnalysisUsage(AnalysisUsage &AU) const override { 71 AU.addRequired<TargetLibraryInfoWrapperPass>(); 72 } 73 static char ID; // Pass identification, replacement for typeid 74 GlobalOpt() : ModulePass(ID) { 75 initializeGlobalOptPass(*PassRegistry::getPassRegistry()); 76 } 77 78 bool runOnModule(Module &M) override; 79 80 private: 81 bool OptimizeFunctions(Module &M); 82 bool OptimizeGlobalVars(Module &M); 83 bool OptimizeGlobalAliases(Module &M); 84 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 85 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI, 86 const GlobalStatus &GS); 87 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn); 88 89 TargetLibraryInfo *TLI; 90 SmallSet<const Comdat *, 8> NotDiscardableComdats; 91 }; 92 } 93 94 char GlobalOpt::ID = 0; 95 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt", 96 "Global Variable Optimizer", false, false) 97 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 98 INITIALIZE_PASS_END(GlobalOpt, "globalopt", 99 "Global Variable Optimizer", false, false) 100 101 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 102 103 /// isLeakCheckerRoot - Is this global variable possibly used by a leak checker 104 /// as a root? If so, we might not really want to eliminate the stores to it. 105 static bool isLeakCheckerRoot(GlobalVariable *GV) { 106 // A global variable is a root if it is a pointer, or could plausibly contain 107 // a pointer. There are two challenges; one is that we could have a struct 108 // the has an inner member which is a pointer. We recurse through the type to 109 // detect these (up to a point). The other is that we may actually be a union 110 // of a pointer and another type, and so our LLVM type is an integer which 111 // gets converted into a pointer, or our type is an [i8 x #] with a pointer 112 // potentially contained here. 113 114 if (GV->hasPrivateLinkage()) 115 return false; 116 117 SmallVector<Type *, 4> Types; 118 Types.push_back(cast<PointerType>(GV->getType())->getElementType()); 119 120 unsigned Limit = 20; 121 do { 122 Type *Ty = Types.pop_back_val(); 123 switch (Ty->getTypeID()) { 124 default: break; 125 case Type::PointerTyID: return true; 126 case Type::ArrayTyID: 127 case Type::VectorTyID: { 128 SequentialType *STy = cast<SequentialType>(Ty); 129 Types.push_back(STy->getElementType()); 130 break; 131 } 132 case Type::StructTyID: { 133 StructType *STy = cast<StructType>(Ty); 134 if (STy->isOpaque()) return true; 135 for (StructType::element_iterator I = STy->element_begin(), 136 E = STy->element_end(); I != E; ++I) { 137 Type *InnerTy = *I; 138 if (isa<PointerType>(InnerTy)) return true; 139 if (isa<CompositeType>(InnerTy)) 140 Types.push_back(InnerTy); 141 } 142 break; 143 } 144 } 145 if (--Limit == 0) return true; 146 } while (!Types.empty()); 147 return false; 148 } 149 150 /// Given a value that is stored to a global but never read, determine whether 151 /// it's safe to remove the store and the chain of computation that feeds the 152 /// store. 153 static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) { 154 do { 155 if (isa<Constant>(V)) 156 return true; 157 if (!V->hasOneUse()) 158 return false; 159 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) || 160 isa<GlobalValue>(V)) 161 return false; 162 if (isAllocationFn(V, TLI)) 163 return true; 164 165 Instruction *I = cast<Instruction>(V); 166 if (I->mayHaveSideEffects()) 167 return false; 168 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 169 if (!GEP->hasAllConstantIndices()) 170 return false; 171 } else if (I->getNumOperands() != 1) { 172 return false; 173 } 174 175 V = I->getOperand(0); 176 } while (1); 177 } 178 179 /// CleanupPointerRootUsers - This GV is a pointer root. Loop over all users 180 /// of the global and clean up any that obviously don't assign the global a 181 /// value that isn't dynamically allocated. 182 /// 183 static bool CleanupPointerRootUsers(GlobalVariable *GV, 184 const TargetLibraryInfo *TLI) { 185 // A brief explanation of leak checkers. The goal is to find bugs where 186 // pointers are forgotten, causing an accumulating growth in memory 187 // usage over time. The common strategy for leak checkers is to whitelist the 188 // memory pointed to by globals at exit. This is popular because it also 189 // solves another problem where the main thread of a C++ program may shut down 190 // before other threads that are still expecting to use those globals. To 191 // handle that case, we expect the program may create a singleton and never 192 // destroy it. 193 194 bool Changed = false; 195 196 // If Dead[n].first is the only use of a malloc result, we can delete its 197 // chain of computation and the store to the global in Dead[n].second. 198 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead; 199 200 // Constants can't be pointers to dynamically allocated memory. 201 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end(); 202 UI != E;) { 203 User *U = *UI++; 204 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 205 Value *V = SI->getValueOperand(); 206 if (isa<Constant>(V)) { 207 Changed = true; 208 SI->eraseFromParent(); 209 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 210 if (I->hasOneUse()) 211 Dead.push_back(std::make_pair(I, SI)); 212 } 213 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) { 214 if (isa<Constant>(MSI->getValue())) { 215 Changed = true; 216 MSI->eraseFromParent(); 217 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) { 218 if (I->hasOneUse()) 219 Dead.push_back(std::make_pair(I, MSI)); 220 } 221 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) { 222 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource()); 223 if (MemSrc && MemSrc->isConstant()) { 224 Changed = true; 225 MTI->eraseFromParent(); 226 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) { 227 if (I->hasOneUse()) 228 Dead.push_back(std::make_pair(I, MTI)); 229 } 230 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 231 if (CE->use_empty()) { 232 CE->destroyConstant(); 233 Changed = true; 234 } 235 } else if (Constant *C = dyn_cast<Constant>(U)) { 236 if (isSafeToDestroyConstant(C)) { 237 C->destroyConstant(); 238 // This could have invalidated UI, start over from scratch. 239 Dead.clear(); 240 CleanupPointerRootUsers(GV, TLI); 241 return true; 242 } 243 } 244 } 245 246 for (int i = 0, e = Dead.size(); i != e; ++i) { 247 if (IsSafeComputationToRemove(Dead[i].first, TLI)) { 248 Dead[i].second->eraseFromParent(); 249 Instruction *I = Dead[i].first; 250 do { 251 if (isAllocationFn(I, TLI)) 252 break; 253 Instruction *J = dyn_cast<Instruction>(I->getOperand(0)); 254 if (!J) 255 break; 256 I->eraseFromParent(); 257 I = J; 258 } while (1); 259 I->eraseFromParent(); 260 } 261 } 262 263 return Changed; 264 } 265 266 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 267 /// users of the global, cleaning up the obvious ones. This is largely just a 268 /// quick scan over the use list to clean up the easy and obvious cruft. This 269 /// returns true if it made a change. 270 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init, 271 const DataLayout &DL, 272 TargetLibraryInfo *TLI) { 273 bool Changed = false; 274 // Note that we need to use a weak value handle for the worklist items. When 275 // we delete a constant array, we may also be holding pointer to one of its 276 // elements (or an element of one of its elements if we're dealing with an 277 // array of arrays) in the worklist. 278 SmallVector<WeakVH, 8> WorkList(V->user_begin(), V->user_end()); 279 while (!WorkList.empty()) { 280 Value *UV = WorkList.pop_back_val(); 281 if (!UV) 282 continue; 283 284 User *U = cast<User>(UV); 285 286 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 287 if (Init) { 288 // Replace the load with the initializer. 289 LI->replaceAllUsesWith(Init); 290 LI->eraseFromParent(); 291 Changed = true; 292 } 293 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 294 // Store must be unreachable or storing Init into the global. 295 SI->eraseFromParent(); 296 Changed = true; 297 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 298 if (CE->getOpcode() == Instruction::GetElementPtr) { 299 Constant *SubInit = nullptr; 300 if (Init) 301 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 302 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, TLI); 303 } else if ((CE->getOpcode() == Instruction::BitCast && 304 CE->getType()->isPointerTy()) || 305 CE->getOpcode() == Instruction::AddrSpaceCast) { 306 // Pointer cast, delete any stores and memsets to the global. 307 Changed |= CleanupConstantGlobalUsers(CE, nullptr, DL, TLI); 308 } 309 310 if (CE->use_empty()) { 311 CE->destroyConstant(); 312 Changed = true; 313 } 314 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 315 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 316 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 317 // and will invalidate our notion of what Init is. 318 Constant *SubInit = nullptr; 319 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 320 ConstantExpr *CE = dyn_cast_or_null<ConstantExpr>( 321 ConstantFoldInstruction(GEP, DL, TLI)); 322 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 323 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 324 325 // If the initializer is an all-null value and we have an inbounds GEP, 326 // we already know what the result of any load from that GEP is. 327 // TODO: Handle splats. 328 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds()) 329 SubInit = Constant::getNullValue(GEP->getType()->getElementType()); 330 } 331 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, TLI); 332 333 if (GEP->use_empty()) { 334 GEP->eraseFromParent(); 335 Changed = true; 336 } 337 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 338 if (MI->getRawDest() == V) { 339 MI->eraseFromParent(); 340 Changed = true; 341 } 342 343 } else if (Constant *C = dyn_cast<Constant>(U)) { 344 // If we have a chain of dead constantexprs or other things dangling from 345 // us, and if they are all dead, nuke them without remorse. 346 if (isSafeToDestroyConstant(C)) { 347 C->destroyConstant(); 348 CleanupConstantGlobalUsers(V, Init, DL, TLI); 349 return true; 350 } 351 } 352 } 353 return Changed; 354 } 355 356 /// isSafeSROAElementUse - Return true if the specified instruction is a safe 357 /// user of a derived expression from a global that we want to SROA. 358 static bool isSafeSROAElementUse(Value *V) { 359 // We might have a dead and dangling constant hanging off of here. 360 if (Constant *C = dyn_cast<Constant>(V)) 361 return isSafeToDestroyConstant(C); 362 363 Instruction *I = dyn_cast<Instruction>(V); 364 if (!I) return false; 365 366 // Loads are ok. 367 if (isa<LoadInst>(I)) return true; 368 369 // Stores *to* the pointer are ok. 370 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 371 return SI->getOperand(0) != V; 372 373 // Otherwise, it must be a GEP. 374 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 375 if (!GEPI) return false; 376 377 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 378 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 379 return false; 380 381 for (User *U : GEPI->users()) 382 if (!isSafeSROAElementUse(U)) 383 return false; 384 return true; 385 } 386 387 388 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 389 /// Look at it and its uses and decide whether it is safe to SROA this global. 390 /// 391 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 392 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 393 if (!isa<GetElementPtrInst>(U) && 394 (!isa<ConstantExpr>(U) || 395 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 396 return false; 397 398 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 399 // don't like < 3 operand CE's, and we don't like non-constant integer 400 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 401 // value of C. 402 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 403 !cast<Constant>(U->getOperand(1))->isNullValue() || 404 !isa<ConstantInt>(U->getOperand(2))) 405 return false; 406 407 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 408 ++GEPI; // Skip over the pointer index. 409 410 // If this is a use of an array allocation, do a bit more checking for sanity. 411 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 412 uint64_t NumElements = AT->getNumElements(); 413 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 414 415 // Check to make sure that index falls within the array. If not, 416 // something funny is going on, so we won't do the optimization. 417 // 418 if (Idx->getZExtValue() >= NumElements) 419 return false; 420 421 // We cannot scalar repl this level of the array unless any array 422 // sub-indices are in-range constants. In particular, consider: 423 // A[0][i]. We cannot know that the user isn't doing invalid things like 424 // allowing i to index an out-of-range subscript that accesses A[1]. 425 // 426 // Scalar replacing *just* the outer index of the array is probably not 427 // going to be a win anyway, so just give up. 428 for (++GEPI; // Skip array index. 429 GEPI != E; 430 ++GEPI) { 431 uint64_t NumElements; 432 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 433 NumElements = SubArrayTy->getNumElements(); 434 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI)) 435 NumElements = SubVectorTy->getNumElements(); 436 else { 437 assert((*GEPI)->isStructTy() && 438 "Indexed GEP type is not array, vector, or struct!"); 439 continue; 440 } 441 442 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 443 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 444 return false; 445 } 446 } 447 448 for (User *UU : U->users()) 449 if (!isSafeSROAElementUse(UU)) 450 return false; 451 452 return true; 453 } 454 455 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 456 /// is safe for us to perform this transformation. 457 /// 458 static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 459 for (User *U : GV->users()) 460 if (!IsUserOfGlobalSafeForSRA(U, GV)) 461 return false; 462 463 return true; 464 } 465 466 467 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global 468 /// variable. This opens the door for other optimizations by exposing the 469 /// behavior of the program in a more fine-grained way. We have determined that 470 /// this transformation is safe already. We return the first global variable we 471 /// insert so that the caller can reprocess it. 472 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) { 473 // Make sure this global only has simple uses that we can SRA. 474 if (!GlobalUsersSafeToSRA(GV)) 475 return nullptr; 476 477 assert(GV->hasLocalLinkage() && !GV->isConstant()); 478 Constant *Init = GV->getInitializer(); 479 Type *Ty = Init->getType(); 480 481 std::vector<GlobalVariable*> NewGlobals; 482 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 483 484 // Get the alignment of the global, either explicit or target-specific. 485 unsigned StartAlignment = GV->getAlignment(); 486 if (StartAlignment == 0) 487 StartAlignment = DL.getABITypeAlignment(GV->getType()); 488 489 if (StructType *STy = dyn_cast<StructType>(Ty)) { 490 NewGlobals.reserve(STy->getNumElements()); 491 const StructLayout &Layout = *DL.getStructLayout(STy); 492 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 493 Constant *In = Init->getAggregateElement(i); 494 assert(In && "Couldn't get element of initializer?"); 495 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 496 GlobalVariable::InternalLinkage, 497 In, GV->getName()+"."+Twine(i), 498 GV->getThreadLocalMode(), 499 GV->getType()->getAddressSpace()); 500 Globals.insert(GV, NGV); 501 NewGlobals.push_back(NGV); 502 503 // Calculate the known alignment of the field. If the original aggregate 504 // had 256 byte alignment for example, something might depend on that: 505 // propagate info to each field. 506 uint64_t FieldOffset = Layout.getElementOffset(i); 507 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 508 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i))) 509 NGV->setAlignment(NewAlign); 510 } 511 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 512 unsigned NumElements = 0; 513 if (ArrayType *ATy = dyn_cast<ArrayType>(STy)) 514 NumElements = ATy->getNumElements(); 515 else 516 NumElements = cast<VectorType>(STy)->getNumElements(); 517 518 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 519 return nullptr; // It's not worth it. 520 NewGlobals.reserve(NumElements); 521 522 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType()); 523 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType()); 524 for (unsigned i = 0, e = NumElements; i != e; ++i) { 525 Constant *In = Init->getAggregateElement(i); 526 assert(In && "Couldn't get element of initializer?"); 527 528 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 529 GlobalVariable::InternalLinkage, 530 In, GV->getName()+"."+Twine(i), 531 GV->getThreadLocalMode(), 532 GV->getType()->getAddressSpace()); 533 Globals.insert(GV, NGV); 534 NewGlobals.push_back(NGV); 535 536 // Calculate the known alignment of the field. If the original aggregate 537 // had 256 byte alignment for example, something might depend on that: 538 // propagate info to each field. 539 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 540 if (NewAlign > EltAlign) 541 NGV->setAlignment(NewAlign); 542 } 543 } 544 545 if (NewGlobals.empty()) 546 return nullptr; 547 548 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV); 549 550 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext())); 551 552 // Loop over all of the uses of the global, replacing the constantexpr geps, 553 // with smaller constantexpr geps or direct references. 554 while (!GV->use_empty()) { 555 User *GEP = GV->user_back(); 556 assert(((isa<ConstantExpr>(GEP) && 557 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 558 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 559 560 // Ignore the 1th operand, which has to be zero or else the program is quite 561 // broken (undefined). Get the 2nd operand, which is the structure or array 562 // index. 563 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 564 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 565 566 Value *NewPtr = NewGlobals[Val]; 567 Type *NewTy = NewGlobals[Val]->getType(); 568 569 // Form a shorter GEP if needed. 570 if (GEP->getNumOperands() > 3) { 571 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 572 SmallVector<Constant*, 8> Idxs; 573 Idxs.push_back(NullInt); 574 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 575 Idxs.push_back(CE->getOperand(i)); 576 NewPtr = 577 ConstantExpr::getGetElementPtr(NewTy, cast<Constant>(NewPtr), Idxs); 578 NewTy = GetElementPtrInst::getIndexedType(NewTy, Idxs); 579 } else { 580 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 581 SmallVector<Value*, 8> Idxs; 582 Idxs.push_back(NullInt); 583 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 584 Idxs.push_back(GEPI->getOperand(i)); 585 NewPtr = GetElementPtrInst::Create( 586 NewPtr->getType()->getPointerElementType(), NewPtr, Idxs, 587 GEPI->getName() + "." + Twine(Val), GEPI); 588 } 589 } 590 GEP->replaceAllUsesWith(NewPtr); 591 592 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 593 GEPI->eraseFromParent(); 594 else 595 cast<ConstantExpr>(GEP)->destroyConstant(); 596 } 597 598 // Delete the old global, now that it is dead. 599 Globals.erase(GV); 600 ++NumSRA; 601 602 // Loop over the new globals array deleting any globals that are obviously 603 // dead. This can arise due to scalarization of a structure or an array that 604 // has elements that are dead. 605 unsigned FirstGlobal = 0; 606 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 607 if (NewGlobals[i]->use_empty()) { 608 Globals.erase(NewGlobals[i]); 609 if (FirstGlobal == i) ++FirstGlobal; 610 } 611 612 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr; 613 } 614 615 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 616 /// value will trap if the value is dynamically null. PHIs keeps track of any 617 /// phi nodes we've seen to avoid reprocessing them. 618 static bool AllUsesOfValueWillTrapIfNull(const Value *V, 619 SmallPtrSetImpl<const PHINode*> &PHIs) { 620 for (const User *U : V->users()) 621 if (isa<LoadInst>(U)) { 622 // Will trap. 623 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) { 624 if (SI->getOperand(0) == V) { 625 //cerr << "NONTRAPPING USE: " << *U; 626 return false; // Storing the value. 627 } 628 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) { 629 if (CI->getCalledValue() != V) { 630 //cerr << "NONTRAPPING USE: " << *U; 631 return false; // Not calling the ptr 632 } 633 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) { 634 if (II->getCalledValue() != V) { 635 //cerr << "NONTRAPPING USE: " << *U; 636 return false; // Not calling the ptr 637 } 638 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) { 639 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 640 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 641 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 642 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) { 643 // If we've already seen this phi node, ignore it, it has already been 644 // checked. 645 if (PHIs.insert(PN).second && !AllUsesOfValueWillTrapIfNull(PN, PHIs)) 646 return false; 647 } else if (isa<ICmpInst>(U) && 648 isa<ConstantPointerNull>(U->getOperand(1))) { 649 // Ignore icmp X, null 650 } else { 651 //cerr << "NONTRAPPING USE: " << *U; 652 return false; 653 } 654 655 return true; 656 } 657 658 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 659 /// from GV will trap if the loaded value is null. Note that this also permits 660 /// comparisons of the loaded value against null, as a special case. 661 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) { 662 for (const User *U : GV->users()) 663 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 664 SmallPtrSet<const PHINode*, 8> PHIs; 665 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 666 return false; 667 } else if (isa<StoreInst>(U)) { 668 // Ignore stores to the global. 669 } else { 670 // We don't know or understand this user, bail out. 671 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U; 672 return false; 673 } 674 return true; 675 } 676 677 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 678 bool Changed = false; 679 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) { 680 Instruction *I = cast<Instruction>(*UI++); 681 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 682 LI->setOperand(0, NewV); 683 Changed = true; 684 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 685 if (SI->getOperand(1) == V) { 686 SI->setOperand(1, NewV); 687 Changed = true; 688 } 689 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 690 CallSite CS(I); 691 if (CS.getCalledValue() == V) { 692 // Calling through the pointer! Turn into a direct call, but be careful 693 // that the pointer is not also being passed as an argument. 694 CS.setCalledFunction(NewV); 695 Changed = true; 696 bool PassedAsArg = false; 697 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 698 if (CS.getArgument(i) == V) { 699 PassedAsArg = true; 700 CS.setArgument(i, NewV); 701 } 702 703 if (PassedAsArg) { 704 // Being passed as an argument also. Be careful to not invalidate UI! 705 UI = V->user_begin(); 706 } 707 } 708 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 709 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 710 ConstantExpr::getCast(CI->getOpcode(), 711 NewV, CI->getType())); 712 if (CI->use_empty()) { 713 Changed = true; 714 CI->eraseFromParent(); 715 } 716 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 717 // Should handle GEP here. 718 SmallVector<Constant*, 8> Idxs; 719 Idxs.reserve(GEPI->getNumOperands()-1); 720 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 721 i != e; ++i) 722 if (Constant *C = dyn_cast<Constant>(*i)) 723 Idxs.push_back(C); 724 else 725 break; 726 if (Idxs.size() == GEPI->getNumOperands()-1) 727 Changed |= OptimizeAwayTrappingUsesOfValue( 728 GEPI, ConstantExpr::getGetElementPtr(nullptr, NewV, Idxs)); 729 if (GEPI->use_empty()) { 730 Changed = true; 731 GEPI->eraseFromParent(); 732 } 733 } 734 } 735 736 return Changed; 737 } 738 739 740 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 741 /// value stored into it. If there are uses of the loaded value that would trap 742 /// if the loaded value is dynamically null, then we know that they cannot be 743 /// reachable with a null optimize away the load. 744 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, 745 const DataLayout &DL, 746 TargetLibraryInfo *TLI) { 747 bool Changed = false; 748 749 // Keep track of whether we are able to remove all the uses of the global 750 // other than the store that defines it. 751 bool AllNonStoreUsesGone = true; 752 753 // Replace all uses of loads with uses of uses of the stored value. 754 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){ 755 User *GlobalUser = *GUI++; 756 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) { 757 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 758 // If we were able to delete all uses of the loads 759 if (LI->use_empty()) { 760 LI->eraseFromParent(); 761 Changed = true; 762 } else { 763 AllNonStoreUsesGone = false; 764 } 765 } else if (isa<StoreInst>(GlobalUser)) { 766 // Ignore the store that stores "LV" to the global. 767 assert(GlobalUser->getOperand(1) == GV && 768 "Must be storing *to* the global"); 769 } else { 770 AllNonStoreUsesGone = false; 771 772 // If we get here we could have other crazy uses that are transitively 773 // loaded. 774 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || 775 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) || 776 isa<BitCastInst>(GlobalUser) || 777 isa<GetElementPtrInst>(GlobalUser)) && 778 "Only expect load and stores!"); 779 } 780 } 781 782 if (Changed) { 783 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV); 784 ++NumGlobUses; 785 } 786 787 // If we nuked all of the loads, then none of the stores are needed either, 788 // nor is the global. 789 if (AllNonStoreUsesGone) { 790 if (isLeakCheckerRoot(GV)) { 791 Changed |= CleanupPointerRootUsers(GV, TLI); 792 } else { 793 Changed = true; 794 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI); 795 } 796 if (GV->use_empty()) { 797 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n"); 798 Changed = true; 799 GV->eraseFromParent(); 800 ++NumDeleted; 801 } 802 } 803 return Changed; 804 } 805 806 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 807 /// instructions that are foldable. 808 static void ConstantPropUsersOf(Value *V, const DataLayout &DL, 809 TargetLibraryInfo *TLI) { 810 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; ) 811 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 812 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) { 813 I->replaceAllUsesWith(NewC); 814 815 // Advance UI to the next non-I use to avoid invalidating it! 816 // Instructions could multiply use V. 817 while (UI != E && *UI == I) 818 ++UI; 819 I->eraseFromParent(); 820 } 821 } 822 823 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global 824 /// variable, and transforms the program as if it always contained the result of 825 /// the specified malloc. Because it is always the result of the specified 826 /// malloc, there is no reason to actually DO the malloc. Instead, turn the 827 /// malloc into a global, and any loads of GV as uses of the new global. 828 static GlobalVariable * 829 OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy, 830 ConstantInt *NElements, const DataLayout &DL, 831 TargetLibraryInfo *TLI) { 832 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n'); 833 834 Type *GlobalType; 835 if (NElements->getZExtValue() == 1) 836 GlobalType = AllocTy; 837 else 838 // If we have an array allocation, the global variable is of an array. 839 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue()); 840 841 // Create the new global variable. The contents of the malloc'd memory is 842 // undefined, so initialize with an undef value. 843 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(), 844 GlobalType, false, 845 GlobalValue::InternalLinkage, 846 UndefValue::get(GlobalType), 847 GV->getName()+".body", 848 GV, 849 GV->getThreadLocalMode()); 850 851 // If there are bitcast users of the malloc (which is typical, usually we have 852 // a malloc + bitcast) then replace them with uses of the new global. Update 853 // other users to use the global as well. 854 BitCastInst *TheBC = nullptr; 855 while (!CI->use_empty()) { 856 Instruction *User = cast<Instruction>(CI->user_back()); 857 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 858 if (BCI->getType() == NewGV->getType()) { 859 BCI->replaceAllUsesWith(NewGV); 860 BCI->eraseFromParent(); 861 } else { 862 BCI->setOperand(0, NewGV); 863 } 864 } else { 865 if (!TheBC) 866 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI); 867 User->replaceUsesOfWith(CI, TheBC); 868 } 869 } 870 871 Constant *RepValue = NewGV; 872 if (NewGV->getType() != GV->getType()->getElementType()) 873 RepValue = ConstantExpr::getBitCast(RepValue, 874 GV->getType()->getElementType()); 875 876 // If there is a comparison against null, we will insert a global bool to 877 // keep track of whether the global was initialized yet or not. 878 GlobalVariable *InitBool = 879 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false, 880 GlobalValue::InternalLinkage, 881 ConstantInt::getFalse(GV->getContext()), 882 GV->getName()+".init", GV->getThreadLocalMode()); 883 bool InitBoolUsed = false; 884 885 // Loop over all uses of GV, processing them in turn. 886 while (!GV->use_empty()) { 887 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) { 888 // The global is initialized when the store to it occurs. 889 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0, 890 SI->getOrdering(), SI->getSynchScope(), SI); 891 SI->eraseFromParent(); 892 continue; 893 } 894 895 LoadInst *LI = cast<LoadInst>(GV->user_back()); 896 while (!LI->use_empty()) { 897 Use &LoadUse = *LI->use_begin(); 898 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser()); 899 if (!ICI) { 900 LoadUse = RepValue; 901 continue; 902 } 903 904 // Replace the cmp X, 0 with a use of the bool value. 905 // Sink the load to where the compare was, if atomic rules allow us to. 906 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0, 907 LI->getOrdering(), LI->getSynchScope(), 908 LI->isUnordered() ? (Instruction*)ICI : LI); 909 InitBoolUsed = true; 910 switch (ICI->getPredicate()) { 911 default: llvm_unreachable("Unknown ICmp Predicate!"); 912 case ICmpInst::ICMP_ULT: 913 case ICmpInst::ICMP_SLT: // X < null -> always false 914 LV = ConstantInt::getFalse(GV->getContext()); 915 break; 916 case ICmpInst::ICMP_ULE: 917 case ICmpInst::ICMP_SLE: 918 case ICmpInst::ICMP_EQ: 919 LV = BinaryOperator::CreateNot(LV, "notinit", ICI); 920 break; 921 case ICmpInst::ICMP_NE: 922 case ICmpInst::ICMP_UGE: 923 case ICmpInst::ICMP_SGE: 924 case ICmpInst::ICMP_UGT: 925 case ICmpInst::ICMP_SGT: 926 break; // no change. 927 } 928 ICI->replaceAllUsesWith(LV); 929 ICI->eraseFromParent(); 930 } 931 LI->eraseFromParent(); 932 } 933 934 // If the initialization boolean was used, insert it, otherwise delete it. 935 if (!InitBoolUsed) { 936 while (!InitBool->use_empty()) // Delete initializations 937 cast<StoreInst>(InitBool->user_back())->eraseFromParent(); 938 delete InitBool; 939 } else 940 GV->getParent()->getGlobalList().insert(GV, InitBool); 941 942 // Now the GV is dead, nuke it and the malloc.. 943 GV->eraseFromParent(); 944 CI->eraseFromParent(); 945 946 // To further other optimizations, loop over all users of NewGV and try to 947 // constant prop them. This will promote GEP instructions with constant 948 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 949 ConstantPropUsersOf(NewGV, DL, TLI); 950 if (RepValue != NewGV) 951 ConstantPropUsersOf(RepValue, DL, TLI); 952 953 return NewGV; 954 } 955 956 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 957 /// to make sure that there are no complex uses of V. We permit simple things 958 /// like dereferencing the pointer, but not storing through the address, unless 959 /// it is to the specified global. 960 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V, 961 const GlobalVariable *GV, 962 SmallPtrSetImpl<const PHINode*> &PHIs) { 963 for (const User *U : V->users()) { 964 const Instruction *Inst = cast<Instruction>(U); 965 966 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { 967 continue; // Fine, ignore. 968 } 969 970 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 971 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 972 return false; // Storing the pointer itself... bad. 973 continue; // Otherwise, storing through it, or storing into GV... fine. 974 } 975 976 // Must index into the array and into the struct. 977 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) { 978 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs)) 979 return false; 980 continue; 981 } 982 983 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) { 984 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 985 // cycles. 986 if (PHIs.insert(PN).second) 987 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 988 return false; 989 continue; 990 } 991 992 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { 993 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) 994 return false; 995 continue; 996 } 997 998 return false; 999 } 1000 return true; 1001 } 1002 1003 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 1004 /// somewhere. Transform all uses of the allocation into loads from the 1005 /// global and uses of the resultant pointer. Further, delete the store into 1006 /// GV. This assumes that these value pass the 1007 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 1008 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 1009 GlobalVariable *GV) { 1010 while (!Alloc->use_empty()) { 1011 Instruction *U = cast<Instruction>(*Alloc->user_begin()); 1012 Instruction *InsertPt = U; 1013 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1014 // If this is the store of the allocation into the global, remove it. 1015 if (SI->getOperand(1) == GV) { 1016 SI->eraseFromParent(); 1017 continue; 1018 } 1019 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 1020 // Insert the load in the corresponding predecessor, not right before the 1021 // PHI. 1022 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator(); 1023 } else if (isa<BitCastInst>(U)) { 1024 // Must be bitcast between the malloc and store to initialize the global. 1025 ReplaceUsesOfMallocWithGlobal(U, GV); 1026 U->eraseFromParent(); 1027 continue; 1028 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 1029 // If this is a "GEP bitcast" and the user is a store to the global, then 1030 // just process it as a bitcast. 1031 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse()) 1032 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back())) 1033 if (SI->getOperand(1) == GV) { 1034 // Must be bitcast GEP between the malloc and store to initialize 1035 // the global. 1036 ReplaceUsesOfMallocWithGlobal(GEPI, GV); 1037 GEPI->eraseFromParent(); 1038 continue; 1039 } 1040 } 1041 1042 // Insert a load from the global, and use it instead of the malloc. 1043 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 1044 U->replaceUsesOfWith(Alloc, NL); 1045 } 1046 } 1047 1048 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi 1049 /// of a load) are simple enough to perform heap SRA on. This permits GEP's 1050 /// that index through the array and struct field, icmps of null, and PHIs. 1051 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V, 1052 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs, 1053 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) { 1054 // We permit two users of the load: setcc comparing against the null 1055 // pointer, and a getelementptr of a specific form. 1056 for (const User *U : V->users()) { 1057 const Instruction *UI = cast<Instruction>(U); 1058 1059 // Comparison against null is ok. 1060 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) { 1061 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 1062 return false; 1063 continue; 1064 } 1065 1066 // getelementptr is also ok, but only a simple form. 1067 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) { 1068 // Must index into the array and into the struct. 1069 if (GEPI->getNumOperands() < 3) 1070 return false; 1071 1072 // Otherwise the GEP is ok. 1073 continue; 1074 } 1075 1076 if (const PHINode *PN = dyn_cast<PHINode>(UI)) { 1077 if (!LoadUsingPHIsPerLoad.insert(PN).second) 1078 // This means some phi nodes are dependent on each other. 1079 // Avoid infinite looping! 1080 return false; 1081 if (!LoadUsingPHIs.insert(PN).second) 1082 // If we have already analyzed this PHI, then it is safe. 1083 continue; 1084 1085 // Make sure all uses of the PHI are simple enough to transform. 1086 if (!LoadUsesSimpleEnoughForHeapSRA(PN, 1087 LoadUsingPHIs, LoadUsingPHIsPerLoad)) 1088 return false; 1089 1090 continue; 1091 } 1092 1093 // Otherwise we don't know what this is, not ok. 1094 return false; 1095 } 1096 1097 return true; 1098 } 1099 1100 1101 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1102 /// GV are simple enough to perform HeapSRA, return true. 1103 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV, 1104 Instruction *StoredVal) { 1105 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs; 1106 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad; 1107 for (const User *U : GV->users()) 1108 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 1109 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs, 1110 LoadUsingPHIsPerLoad)) 1111 return false; 1112 LoadUsingPHIsPerLoad.clear(); 1113 } 1114 1115 // If we reach here, we know that all uses of the loads and transitive uses 1116 // (through PHI nodes) are simple enough to transform. However, we don't know 1117 // that all inputs the to the PHI nodes are in the same equivalence sets. 1118 // Check to verify that all operands of the PHIs are either PHIS that can be 1119 // transformed, loads from GV, or MI itself. 1120 for (const PHINode *PN : LoadUsingPHIs) { 1121 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) { 1122 Value *InVal = PN->getIncomingValue(op); 1123 1124 // PHI of the stored value itself is ok. 1125 if (InVal == StoredVal) continue; 1126 1127 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) { 1128 // One of the PHIs in our set is (optimistically) ok. 1129 if (LoadUsingPHIs.count(InPN)) 1130 continue; 1131 return false; 1132 } 1133 1134 // Load from GV is ok. 1135 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal)) 1136 if (LI->getOperand(0) == GV) 1137 continue; 1138 1139 // UNDEF? NULL? 1140 1141 // Anything else is rejected. 1142 return false; 1143 } 1144 } 1145 1146 return true; 1147 } 1148 1149 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, 1150 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1151 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1152 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V]; 1153 1154 if (FieldNo >= FieldVals.size()) 1155 FieldVals.resize(FieldNo+1); 1156 1157 // If we already have this value, just reuse the previously scalarized 1158 // version. 1159 if (Value *FieldVal = FieldVals[FieldNo]) 1160 return FieldVal; 1161 1162 // Depending on what instruction this is, we have several cases. 1163 Value *Result; 1164 if (LoadInst *LI = dyn_cast<LoadInst>(V)) { 1165 // This is a scalarized version of the load from the global. Just create 1166 // a new Load of the scalarized global. 1167 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo, 1168 InsertedScalarizedValues, 1169 PHIsToRewrite), 1170 LI->getName()+".f"+Twine(FieldNo), LI); 1171 } else { 1172 PHINode *PN = cast<PHINode>(V); 1173 // PN's type is pointer to struct. Make a new PHI of pointer to struct 1174 // field. 1175 1176 PointerType *PTy = cast<PointerType>(PN->getType()); 1177 StructType *ST = cast<StructType>(PTy->getElementType()); 1178 1179 unsigned AS = PTy->getAddressSpace(); 1180 PHINode *NewPN = 1181 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS), 1182 PN->getNumIncomingValues(), 1183 PN->getName()+".f"+Twine(FieldNo), PN); 1184 Result = NewPN; 1185 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); 1186 } 1187 1188 return FieldVals[FieldNo] = Result; 1189 } 1190 1191 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1192 /// the load, rewrite the derived value to use the HeapSRoA'd load. 1193 static void RewriteHeapSROALoadUser(Instruction *LoadUser, 1194 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1195 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1196 // If this is a comparison against null, handle it. 1197 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1198 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1199 // If we have a setcc of the loaded pointer, we can use a setcc of any 1200 // field. 1201 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0, 1202 InsertedScalarizedValues, PHIsToRewrite); 1203 1204 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr, 1205 Constant::getNullValue(NPtr->getType()), 1206 SCI->getName()); 1207 SCI->replaceAllUsesWith(New); 1208 SCI->eraseFromParent(); 1209 return; 1210 } 1211 1212 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...' 1213 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1214 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1215 && "Unexpected GEPI!"); 1216 1217 // Load the pointer for this field. 1218 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1219 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo, 1220 InsertedScalarizedValues, PHIsToRewrite); 1221 1222 // Create the new GEP idx vector. 1223 SmallVector<Value*, 8> GEPIdx; 1224 GEPIdx.push_back(GEPI->getOperand(1)); 1225 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1226 1227 Value *NGEPI = GetElementPtrInst::Create(GEPI->getResultElementType(), NewPtr, GEPIdx, 1228 GEPI->getName(), GEPI); 1229 GEPI->replaceAllUsesWith(NGEPI); 1230 GEPI->eraseFromParent(); 1231 return; 1232 } 1233 1234 // Recursively transform the users of PHI nodes. This will lazily create the 1235 // PHIs that are needed for individual elements. Keep track of what PHIs we 1236 // see in InsertedScalarizedValues so that we don't get infinite loops (very 1237 // antisocial). If the PHI is already in InsertedScalarizedValues, it has 1238 // already been seen first by another load, so its uses have already been 1239 // processed. 1240 PHINode *PN = cast<PHINode>(LoadUser); 1241 if (!InsertedScalarizedValues.insert(std::make_pair(PN, 1242 std::vector<Value*>())).second) 1243 return; 1244 1245 // If this is the first time we've seen this PHI, recursively process all 1246 // users. 1247 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) { 1248 Instruction *User = cast<Instruction>(*UI++); 1249 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1250 } 1251 } 1252 1253 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1254 /// is a value loaded from the global. Eliminate all uses of Ptr, making them 1255 /// use FieldGlobals instead. All uses of loaded values satisfy 1256 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA. 1257 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1258 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1259 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1260 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) { 1261 Instruction *User = cast<Instruction>(*UI++); 1262 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1263 } 1264 1265 if (Load->use_empty()) { 1266 Load->eraseFromParent(); 1267 InsertedScalarizedValues.erase(Load); 1268 } 1269 } 1270 1271 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break 1272 /// it up into multiple allocations of arrays of the fields. 1273 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, 1274 Value *NElems, const DataLayout &DL, 1275 const TargetLibraryInfo *TLI) { 1276 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n'); 1277 Type *MAT = getMallocAllocatedType(CI, TLI); 1278 StructType *STy = cast<StructType>(MAT); 1279 1280 // There is guaranteed to be at least one use of the malloc (storing 1281 // it into GV). If there are other uses, change them to be uses of 1282 // the global to simplify later code. This also deletes the store 1283 // into GV. 1284 ReplaceUsesOfMallocWithGlobal(CI, GV); 1285 1286 // Okay, at this point, there are no users of the malloc. Insert N 1287 // new mallocs at the same place as CI, and N globals. 1288 std::vector<Value*> FieldGlobals; 1289 std::vector<Value*> FieldMallocs; 1290 1291 unsigned AS = GV->getType()->getPointerAddressSpace(); 1292 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1293 Type *FieldTy = STy->getElementType(FieldNo); 1294 PointerType *PFieldTy = PointerType::get(FieldTy, AS); 1295 1296 GlobalVariable *NGV = 1297 new GlobalVariable(*GV->getParent(), 1298 PFieldTy, false, GlobalValue::InternalLinkage, 1299 Constant::getNullValue(PFieldTy), 1300 GV->getName() + ".f" + Twine(FieldNo), GV, 1301 GV->getThreadLocalMode()); 1302 FieldGlobals.push_back(NGV); 1303 1304 unsigned TypeSize = DL.getTypeAllocSize(FieldTy); 1305 if (StructType *ST = dyn_cast<StructType>(FieldTy)) 1306 TypeSize = DL.getStructLayout(ST)->getSizeInBytes(); 1307 Type *IntPtrTy = DL.getIntPtrType(CI->getType()); 1308 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy, 1309 ConstantInt::get(IntPtrTy, TypeSize), 1310 NElems, nullptr, 1311 CI->getName() + ".f" + Twine(FieldNo)); 1312 FieldMallocs.push_back(NMI); 1313 new StoreInst(NMI, NGV, CI); 1314 } 1315 1316 // The tricky aspect of this transformation is handling the case when malloc 1317 // fails. In the original code, malloc failing would set the result pointer 1318 // of malloc to null. In this case, some mallocs could succeed and others 1319 // could fail. As such, we emit code that looks like this: 1320 // F0 = malloc(field0) 1321 // F1 = malloc(field1) 1322 // F2 = malloc(field2) 1323 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1324 // if (F0) { free(F0); F0 = 0; } 1325 // if (F1) { free(F1); F1 = 0; } 1326 // if (F2) { free(F2); F2 = 0; } 1327 // } 1328 // The malloc can also fail if its argument is too large. 1329 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0); 1330 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0), 1331 ConstantZero, "isneg"); 1332 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1333 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i], 1334 Constant::getNullValue(FieldMallocs[i]->getType()), 1335 "isnull"); 1336 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI); 1337 } 1338 1339 // Split the basic block at the old malloc. 1340 BasicBlock *OrigBB = CI->getParent(); 1341 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont"); 1342 1343 // Create the block to check the first condition. Put all these blocks at the 1344 // end of the function as they are unlikely to be executed. 1345 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(), 1346 "malloc_ret_null", 1347 OrigBB->getParent()); 1348 1349 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1350 // branch on RunningOr. 1351 OrigBB->getTerminator()->eraseFromParent(); 1352 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1353 1354 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1355 // pointer, because some may be null while others are not. 1356 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1357 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1358 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal, 1359 Constant::getNullValue(GVVal->getType())); 1360 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it", 1361 OrigBB->getParent()); 1362 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next", 1363 OrigBB->getParent()); 1364 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock, 1365 Cmp, NullPtrBlock); 1366 1367 // Fill in FreeBlock. 1368 CallInst::CreateFree(GVVal, BI); 1369 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1370 FreeBlock); 1371 BranchInst::Create(NextBlock, FreeBlock); 1372 1373 NullPtrBlock = NextBlock; 1374 } 1375 1376 BranchInst::Create(ContBB, NullPtrBlock); 1377 1378 // CI is no longer needed, remove it. 1379 CI->eraseFromParent(); 1380 1381 /// InsertedScalarizedLoads - As we process loads, if we can't immediately 1382 /// update all uses of the load, keep track of what scalarized loads are 1383 /// inserted for a given load. 1384 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues; 1385 InsertedScalarizedValues[GV] = FieldGlobals; 1386 1387 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite; 1388 1389 // Okay, the malloc site is completely handled. All of the uses of GV are now 1390 // loads, and all uses of those loads are simple. Rewrite them to use loads 1391 // of the per-field globals instead. 1392 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) { 1393 Instruction *User = cast<Instruction>(*UI++); 1394 1395 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1396 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite); 1397 continue; 1398 } 1399 1400 // Must be a store of null. 1401 StoreInst *SI = cast<StoreInst>(User); 1402 assert(isa<ConstantPointerNull>(SI->getOperand(0)) && 1403 "Unexpected heap-sra user!"); 1404 1405 // Insert a store of null into each global. 1406 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1407 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType()); 1408 Constant *Null = Constant::getNullValue(PT->getElementType()); 1409 new StoreInst(Null, FieldGlobals[i], SI); 1410 } 1411 // Erase the original store. 1412 SI->eraseFromParent(); 1413 } 1414 1415 // While we have PHIs that are interesting to rewrite, do it. 1416 while (!PHIsToRewrite.empty()) { 1417 PHINode *PN = PHIsToRewrite.back().first; 1418 unsigned FieldNo = PHIsToRewrite.back().second; 1419 PHIsToRewrite.pop_back(); 1420 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]); 1421 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi"); 1422 1423 // Add all the incoming values. This can materialize more phis. 1424 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1425 Value *InVal = PN->getIncomingValue(i); 1426 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues, 1427 PHIsToRewrite); 1428 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i)); 1429 } 1430 } 1431 1432 // Drop all inter-phi links and any loads that made it this far. 1433 for (DenseMap<Value*, std::vector<Value*> >::iterator 1434 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1435 I != E; ++I) { 1436 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1437 PN->dropAllReferences(); 1438 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1439 LI->dropAllReferences(); 1440 } 1441 1442 // Delete all the phis and loads now that inter-references are dead. 1443 for (DenseMap<Value*, std::vector<Value*> >::iterator 1444 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1445 I != E; ++I) { 1446 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1447 PN->eraseFromParent(); 1448 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1449 LI->eraseFromParent(); 1450 } 1451 1452 // The old global is now dead, remove it. 1453 GV->eraseFromParent(); 1454 1455 ++NumHeapSRA; 1456 return cast<GlobalVariable>(FieldGlobals[0]); 1457 } 1458 1459 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1460 /// pointer global variable with a single value stored it that is a malloc or 1461 /// cast of malloc. 1462 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI, 1463 Type *AllocTy, 1464 AtomicOrdering Ordering, 1465 Module::global_iterator &GVI, 1466 const DataLayout &DL, 1467 TargetLibraryInfo *TLI) { 1468 // If this is a malloc of an abstract type, don't touch it. 1469 if (!AllocTy->isSized()) 1470 return false; 1471 1472 // We can't optimize this global unless all uses of it are *known* to be 1473 // of the malloc value, not of the null initializer value (consider a use 1474 // that compares the global's value against zero to see if the malloc has 1475 // been reached). To do this, we check to see if all uses of the global 1476 // would trap if the global were null: this proves that they must all 1477 // happen after the malloc. 1478 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1479 return false; 1480 1481 // We can't optimize this if the malloc itself is used in a complex way, 1482 // for example, being stored into multiple globals. This allows the 1483 // malloc to be stored into the specified global, loaded icmp'd, and 1484 // GEP'd. These are all things we could transform to using the global 1485 // for. 1486 SmallPtrSet<const PHINode*, 8> PHIs; 1487 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs)) 1488 return false; 1489 1490 // If we have a global that is only initialized with a fixed size malloc, 1491 // transform the program to use global memory instead of malloc'd memory. 1492 // This eliminates dynamic allocation, avoids an indirection accessing the 1493 // data, and exposes the resultant global to further GlobalOpt. 1494 // We cannot optimize the malloc if we cannot determine malloc array size. 1495 Value *NElems = getMallocArraySize(CI, DL, TLI, true); 1496 if (!NElems) 1497 return false; 1498 1499 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 1500 // Restrict this transformation to only working on small allocations 1501 // (2048 bytes currently), as we don't want to introduce a 16M global or 1502 // something. 1503 if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) { 1504 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI); 1505 return true; 1506 } 1507 1508 // If the allocation is an array of structures, consider transforming this 1509 // into multiple malloc'd arrays, one for each field. This is basically 1510 // SRoA for malloc'd memory. 1511 1512 if (Ordering != NotAtomic) 1513 return false; 1514 1515 // If this is an allocation of a fixed size array of structs, analyze as a 1516 // variable size array. malloc [100 x struct],1 -> malloc struct, 100 1517 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1)) 1518 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy)) 1519 AllocTy = AT->getElementType(); 1520 1521 StructType *AllocSTy = dyn_cast<StructType>(AllocTy); 1522 if (!AllocSTy) 1523 return false; 1524 1525 // This the structure has an unreasonable number of fields, leave it 1526 // alone. 1527 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 && 1528 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) { 1529 1530 // If this is a fixed size array, transform the Malloc to be an alloc of 1531 // structs. malloc [100 x struct],1 -> malloc struct, 100 1532 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) { 1533 Type *IntPtrTy = DL.getIntPtrType(CI->getType()); 1534 unsigned TypeSize = DL.getStructLayout(AllocSTy)->getSizeInBytes(); 1535 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize); 1536 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements()); 1537 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy, 1538 AllocSize, NumElements, 1539 nullptr, CI->getName()); 1540 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI); 1541 CI->replaceAllUsesWith(Cast); 1542 CI->eraseFromParent(); 1543 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc)) 1544 CI = cast<CallInst>(BCI->getOperand(0)); 1545 else 1546 CI = cast<CallInst>(Malloc); 1547 } 1548 1549 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true), 1550 DL, TLI); 1551 return true; 1552 } 1553 1554 return false; 1555 } 1556 1557 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1558 // that only one value (besides its initializer) is ever stored to the global. 1559 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1560 AtomicOrdering Ordering, 1561 Module::global_iterator &GVI, 1562 const DataLayout &DL, 1563 TargetLibraryInfo *TLI) { 1564 // Ignore no-op GEPs and bitcasts. 1565 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1566 1567 // If we are dealing with a pointer global that is initialized to null and 1568 // only has one (non-null) value stored into it, then we can optimize any 1569 // users of the loaded value (often calls and loads) that would trap if the 1570 // value was null. 1571 if (GV->getInitializer()->getType()->isPointerTy() && 1572 GV->getInitializer()->isNullValue()) { 1573 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1574 if (GV->getInitializer()->getType() != SOVC->getType()) 1575 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1576 1577 // Optimize away any trapping uses of the loaded value. 1578 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI)) 1579 return true; 1580 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) { 1581 Type *MallocType = getMallocAllocatedType(CI, TLI); 1582 if (MallocType && 1583 TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI, 1584 DL, TLI)) 1585 return true; 1586 } 1587 } 1588 1589 return false; 1590 } 1591 1592 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1593 /// two values ever stored into GV are its initializer and OtherVal. See if we 1594 /// can shrink the global into a boolean and select between the two values 1595 /// whenever it is used. This exposes the values to other scalar optimizations. 1596 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1597 Type *GVElType = GV->getType()->getElementType(); 1598 1599 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1600 // an FP value, pointer or vector, don't do this optimization because a select 1601 // between them is very expensive and unlikely to lead to later 1602 // simplification. In these cases, we typically end up with "cond ? v1 : v2" 1603 // where v1 and v2 both require constant pool loads, a big loss. 1604 if (GVElType == Type::getInt1Ty(GV->getContext()) || 1605 GVElType->isFloatingPointTy() || 1606 GVElType->isPointerTy() || GVElType->isVectorTy()) 1607 return false; 1608 1609 // Walk the use list of the global seeing if all the uses are load or store. 1610 // If there is anything else, bail out. 1611 for (User *U : GV->users()) 1612 if (!isa<LoadInst>(U) && !isa<StoreInst>(U)) 1613 return false; 1614 1615 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV); 1616 1617 // Create the new global, initializing it to false. 1618 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()), 1619 false, 1620 GlobalValue::InternalLinkage, 1621 ConstantInt::getFalse(GV->getContext()), 1622 GV->getName()+".b", 1623 GV->getThreadLocalMode(), 1624 GV->getType()->getAddressSpace()); 1625 GV->getParent()->getGlobalList().insert(GV, NewGV); 1626 1627 Constant *InitVal = GV->getInitializer(); 1628 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) && 1629 "No reason to shrink to bool!"); 1630 1631 // If initialized to zero and storing one into the global, we can use a cast 1632 // instead of a select to synthesize the desired value. 1633 bool IsOneZero = false; 1634 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1635 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1636 1637 while (!GV->use_empty()) { 1638 Instruction *UI = cast<Instruction>(GV->user_back()); 1639 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1640 // Change the store into a boolean store. 1641 bool StoringOther = SI->getOperand(0) == OtherVal; 1642 // Only do this if we weren't storing a loaded value. 1643 Value *StoreVal; 1644 if (StoringOther || SI->getOperand(0) == InitVal) { 1645 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()), 1646 StoringOther); 1647 } else { 1648 // Otherwise, we are storing a previously loaded copy. To do this, 1649 // change the copy from copying the original value to just copying the 1650 // bool. 1651 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1652 1653 // If we've already replaced the input, StoredVal will be a cast or 1654 // select instruction. If not, it will be a load of the original 1655 // global. 1656 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1657 assert(LI->getOperand(0) == GV && "Not a copy!"); 1658 // Insert a new load, to preserve the saved value. 1659 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0, 1660 LI->getOrdering(), LI->getSynchScope(), LI); 1661 } else { 1662 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1663 "This is not a form that we understand!"); 1664 StoreVal = StoredVal->getOperand(0); 1665 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1666 } 1667 } 1668 new StoreInst(StoreVal, NewGV, false, 0, 1669 SI->getOrdering(), SI->getSynchScope(), SI); 1670 } else { 1671 // Change the load into a load of bool then a select. 1672 LoadInst *LI = cast<LoadInst>(UI); 1673 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0, 1674 LI->getOrdering(), LI->getSynchScope(), LI); 1675 Value *NSI; 1676 if (IsOneZero) 1677 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1678 else 1679 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1680 NSI->takeName(LI); 1681 LI->replaceAllUsesWith(NSI); 1682 } 1683 UI->eraseFromParent(); 1684 } 1685 1686 // Retain the name of the old global variable. People who are debugging their 1687 // programs may expect these variables to be named the same. 1688 NewGV->takeName(GV); 1689 GV->eraseFromParent(); 1690 return true; 1691 } 1692 1693 1694 /// ProcessGlobal - Analyze the specified global variable and optimize it if 1695 /// possible. If we make a change, return true. 1696 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV, 1697 Module::global_iterator &GVI) { 1698 // Do more involved optimizations if the global is internal. 1699 GV->removeDeadConstantUsers(); 1700 1701 if (GV->use_empty()) { 1702 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV); 1703 GV->eraseFromParent(); 1704 ++NumDeleted; 1705 return true; 1706 } 1707 1708 if (!GV->hasLocalLinkage()) 1709 return false; 1710 1711 GlobalStatus GS; 1712 1713 if (GlobalStatus::analyzeGlobal(GV, GS)) 1714 return false; 1715 1716 if (!GS.IsCompared && !GV->hasUnnamedAddr()) { 1717 GV->setUnnamedAddr(true); 1718 NumUnnamed++; 1719 } 1720 1721 if (GV->isConstant() || !GV->hasInitializer()) 1722 return false; 1723 1724 return ProcessInternalGlobal(GV, GVI, GS); 1725 } 1726 1727 /// ProcessInternalGlobal - Analyze the specified global variable and optimize 1728 /// it if possible. If we make a change, return true. 1729 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1730 Module::global_iterator &GVI, 1731 const GlobalStatus &GS) { 1732 auto &DL = GV->getParent()->getDataLayout(); 1733 // If this is a first class global and has only one accessing function 1734 // and this function is main (which we know is not recursive), we replace 1735 // the global with a local alloca in this function. 1736 // 1737 // NOTE: It doesn't make sense to promote non-single-value types since we 1738 // are just replacing static memory to stack memory. 1739 // 1740 // If the global is in different address space, don't bring it to stack. 1741 if (!GS.HasMultipleAccessingFunctions && 1742 GS.AccessingFunction && !GS.HasNonInstructionUser && 1743 GV->getType()->getElementType()->isSingleValueType() && 1744 GS.AccessingFunction->getName() == "main" && 1745 GS.AccessingFunction->hasExternalLinkage() && 1746 GV->getType()->getAddressSpace() == 0) { 1747 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV); 1748 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction 1749 ->getEntryBlock().begin()); 1750 Type *ElemTy = GV->getType()->getElementType(); 1751 // FIXME: Pass Global's alignment when globals have alignment 1752 AllocaInst *Alloca = new AllocaInst(ElemTy, nullptr, 1753 GV->getName(), &FirstI); 1754 if (!isa<UndefValue>(GV->getInitializer())) 1755 new StoreInst(GV->getInitializer(), Alloca, &FirstI); 1756 1757 GV->replaceAllUsesWith(Alloca); 1758 GV->eraseFromParent(); 1759 ++NumLocalized; 1760 return true; 1761 } 1762 1763 // If the global is never loaded (but may be stored to), it is dead. 1764 // Delete it now. 1765 if (!GS.IsLoaded) { 1766 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV); 1767 1768 bool Changed; 1769 if (isLeakCheckerRoot(GV)) { 1770 // Delete any constant stores to the global. 1771 Changed = CleanupPointerRootUsers(GV, TLI); 1772 } else { 1773 // Delete any stores we can find to the global. We may not be able to 1774 // make it completely dead though. 1775 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI); 1776 } 1777 1778 // If the global is dead now, delete it. 1779 if (GV->use_empty()) { 1780 GV->eraseFromParent(); 1781 ++NumDeleted; 1782 Changed = true; 1783 } 1784 return Changed; 1785 1786 } else if (GS.StoredType <= GlobalStatus::InitializerStored) { 1787 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n"); 1788 GV->setConstant(true); 1789 1790 // Clean up any obviously simplifiable users now. 1791 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI); 1792 1793 // If the global is dead now, just nuke it. 1794 if (GV->use_empty()) { 1795 DEBUG(dbgs() << " *** Marking constant allowed us to simplify " 1796 << "all users and delete global!\n"); 1797 GV->eraseFromParent(); 1798 ++NumDeleted; 1799 } 1800 1801 ++NumMarked; 1802 return true; 1803 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1804 const DataLayout &DL = GV->getParent()->getDataLayout(); 1805 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, DL)) { 1806 GVI = FirstNewGV; // Don't skip the newly produced globals! 1807 return true; 1808 } 1809 } else if (GS.StoredType == GlobalStatus::StoredOnce) { 1810 // If the initial value for the global was an undef value, and if only 1811 // one other value was stored into it, we can just change the 1812 // initializer to be the stored value, then delete all stores to the 1813 // global. This allows us to mark it constant. 1814 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1815 if (isa<UndefValue>(GV->getInitializer())) { 1816 // Change the initial value here. 1817 GV->setInitializer(SOVConstant); 1818 1819 // Clean up any obviously simplifiable users now. 1820 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI); 1821 1822 if (GV->use_empty()) { 1823 DEBUG(dbgs() << " *** Substituting initializer allowed us to " 1824 << "simplify all users and delete global!\n"); 1825 GV->eraseFromParent(); 1826 ++NumDeleted; 1827 } else { 1828 GVI = GV; 1829 } 1830 ++NumSubstitute; 1831 return true; 1832 } 1833 1834 // Try to optimize globals based on the knowledge that only one value 1835 // (besides its initializer) is ever stored to the global. 1836 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI, 1837 DL, TLI)) 1838 return true; 1839 1840 // Otherwise, if the global was not a boolean, we can shrink it to be a 1841 // boolean. 1842 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) { 1843 if (GS.Ordering == NotAtomic) { 1844 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1845 ++NumShrunkToBool; 1846 return true; 1847 } 1848 } 1849 } 1850 } 1851 1852 return false; 1853 } 1854 1855 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1856 /// function, changing them to FastCC. 1857 static void ChangeCalleesToFastCall(Function *F) { 1858 for (User *U : F->users()) { 1859 if (isa<BlockAddress>(U)) 1860 continue; 1861 CallSite CS(cast<Instruction>(U)); 1862 CS.setCallingConv(CallingConv::Fast); 1863 } 1864 } 1865 1866 static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) { 1867 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1868 unsigned Index = Attrs.getSlotIndex(i); 1869 if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest)) 1870 continue; 1871 1872 // There can be only one. 1873 return Attrs.removeAttribute(C, Index, Attribute::Nest); 1874 } 1875 1876 return Attrs; 1877 } 1878 1879 static void RemoveNestAttribute(Function *F) { 1880 F->setAttributes(StripNest(F->getContext(), F->getAttributes())); 1881 for (User *U : F->users()) { 1882 if (isa<BlockAddress>(U)) 1883 continue; 1884 CallSite CS(cast<Instruction>(U)); 1885 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes())); 1886 } 1887 } 1888 1889 /// Return true if this is a calling convention that we'd like to change. The 1890 /// idea here is that we don't want to mess with the convention if the user 1891 /// explicitly requested something with performance implications like coldcc, 1892 /// GHC, or anyregcc. 1893 static bool isProfitableToMakeFastCC(Function *F) { 1894 CallingConv::ID CC = F->getCallingConv(); 1895 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc? 1896 return CC == CallingConv::C || CC == CallingConv::X86_ThisCall; 1897 } 1898 1899 bool GlobalOpt::OptimizeFunctions(Module &M) { 1900 bool Changed = false; 1901 // Optimize functions. 1902 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1903 Function *F = FI++; 1904 // Functions without names cannot be referenced outside this module. 1905 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage()) 1906 F->setLinkage(GlobalValue::InternalLinkage); 1907 1908 const Comdat *C = F->getComdat(); 1909 bool inComdat = C && NotDiscardableComdats.count(C); 1910 F->removeDeadConstantUsers(); 1911 if ((!inComdat || F->hasLocalLinkage()) && F->isDefTriviallyDead()) { 1912 F->eraseFromParent(); 1913 Changed = true; 1914 ++NumFnDeleted; 1915 } else if (F->hasLocalLinkage()) { 1916 if (isProfitableToMakeFastCC(F) && !F->isVarArg() && 1917 !F->hasAddressTaken()) { 1918 // If this function has a calling convention worth changing, is not a 1919 // varargs function, and is only called directly, promote it to use the 1920 // Fast calling convention. 1921 F->setCallingConv(CallingConv::Fast); 1922 ChangeCalleesToFastCall(F); 1923 ++NumFastCallFns; 1924 Changed = true; 1925 } 1926 1927 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1928 !F->hasAddressTaken()) { 1929 // The function is not used by a trampoline intrinsic, so it is safe 1930 // to remove the 'nest' attribute. 1931 RemoveNestAttribute(F); 1932 ++NumNestRemoved; 1933 Changed = true; 1934 } 1935 } 1936 } 1937 return Changed; 1938 } 1939 1940 bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1941 bool Changed = false; 1942 1943 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1944 GVI != E; ) { 1945 GlobalVariable *GV = GVI++; 1946 // Global variables without names cannot be referenced outside this module. 1947 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage()) 1948 GV->setLinkage(GlobalValue::InternalLinkage); 1949 // Simplify the initializer. 1950 if (GV->hasInitializer()) 1951 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) { 1952 auto &DL = M.getDataLayout(); 1953 Constant *New = ConstantFoldConstantExpression(CE, DL, TLI); 1954 if (New && New != CE) 1955 GV->setInitializer(New); 1956 } 1957 1958 if (GV->isDiscardableIfUnused()) { 1959 if (const Comdat *C = GV->getComdat()) 1960 if (NotDiscardableComdats.count(C) && !GV->hasLocalLinkage()) 1961 continue; 1962 Changed |= ProcessGlobal(GV, GVI); 1963 } 1964 } 1965 return Changed; 1966 } 1967 1968 static inline bool 1969 isSimpleEnoughValueToCommit(Constant *C, 1970 SmallPtrSetImpl<Constant *> &SimpleConstants, 1971 const DataLayout &DL); 1972 1973 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be 1974 /// handled by the code generator. We don't want to generate something like: 1975 /// void *X = &X/42; 1976 /// because the code generator doesn't have a relocation that can handle that. 1977 /// 1978 /// This function should be called if C was not found (but just got inserted) 1979 /// in SimpleConstants to avoid having to rescan the same constants all the 1980 /// time. 1981 static bool 1982 isSimpleEnoughValueToCommitHelper(Constant *C, 1983 SmallPtrSetImpl<Constant *> &SimpleConstants, 1984 const DataLayout &DL) { 1985 // Simple global addresses are supported, do not allow dllimport or 1986 // thread-local globals. 1987 if (auto *GV = dyn_cast<GlobalValue>(C)) 1988 return !GV->hasDLLImportStorageClass() && !GV->isThreadLocal(); 1989 1990 // Simple integer, undef, constant aggregate zero, etc are all supported. 1991 if (C->getNumOperands() == 0 || isa<BlockAddress>(C)) 1992 return true; 1993 1994 // Aggregate values are safe if all their elements are. 1995 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) || 1996 isa<ConstantVector>(C)) { 1997 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { 1998 Constant *Op = cast<Constant>(C->getOperand(i)); 1999 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, DL)) 2000 return false; 2001 } 2002 return true; 2003 } 2004 2005 // We don't know exactly what relocations are allowed in constant expressions, 2006 // so we allow &global+constantoffset, which is safe and uniformly supported 2007 // across targets. 2008 ConstantExpr *CE = cast<ConstantExpr>(C); 2009 switch (CE->getOpcode()) { 2010 case Instruction::BitCast: 2011 // Bitcast is fine if the casted value is fine. 2012 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 2013 2014 case Instruction::IntToPtr: 2015 case Instruction::PtrToInt: 2016 // int <=> ptr is fine if the int type is the same size as the 2017 // pointer type. 2018 if (DL.getTypeSizeInBits(CE->getType()) != 2019 DL.getTypeSizeInBits(CE->getOperand(0)->getType())) 2020 return false; 2021 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 2022 2023 // GEP is fine if it is simple + constant offset. 2024 case Instruction::GetElementPtr: 2025 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 2026 if (!isa<ConstantInt>(CE->getOperand(i))) 2027 return false; 2028 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 2029 2030 case Instruction::Add: 2031 // We allow simple+cst. 2032 if (!isa<ConstantInt>(CE->getOperand(1))) 2033 return false; 2034 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 2035 } 2036 return false; 2037 } 2038 2039 static inline bool 2040 isSimpleEnoughValueToCommit(Constant *C, 2041 SmallPtrSetImpl<Constant *> &SimpleConstants, 2042 const DataLayout &DL) { 2043 // If we already checked this constant, we win. 2044 if (!SimpleConstants.insert(C).second) 2045 return true; 2046 // Check the constant. 2047 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL); 2048 } 2049 2050 2051 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple 2052 /// enough for us to understand. In particular, if it is a cast to anything 2053 /// other than from one pointer type to another pointer type, we punt. 2054 /// We basically just support direct accesses to globals and GEP's of 2055 /// globals. This should be kept up to date with CommitValueTo. 2056 static bool isSimpleEnoughPointerToCommit(Constant *C) { 2057 // Conservatively, avoid aggregate types. This is because we don't 2058 // want to worry about them partially overlapping other stores. 2059 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType()) 2060 return false; 2061 2062 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 2063 // Do not allow weak/*_odr/linkonce linkage or external globals. 2064 return GV->hasUniqueInitializer(); 2065 2066 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2067 // Handle a constantexpr gep. 2068 if (CE->getOpcode() == Instruction::GetElementPtr && 2069 isa<GlobalVariable>(CE->getOperand(0)) && 2070 cast<GEPOperator>(CE)->isInBounds()) { 2071 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2072 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or 2073 // external globals. 2074 if (!GV->hasUniqueInitializer()) 2075 return false; 2076 2077 // The first index must be zero. 2078 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin())); 2079 if (!CI || !CI->isZero()) return false; 2080 2081 // The remaining indices must be compile-time known integers within the 2082 // notional bounds of the corresponding static array types. 2083 if (!CE->isGEPWithNoNotionalOverIndexing()) 2084 return false; 2085 2086 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2087 2088 // A constantexpr bitcast from a pointer to another pointer is a no-op, 2089 // and we know how to evaluate it by moving the bitcast from the pointer 2090 // operand to the value operand. 2091 } else if (CE->getOpcode() == Instruction::BitCast && 2092 isa<GlobalVariable>(CE->getOperand(0))) { 2093 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or 2094 // external globals. 2095 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer(); 2096 } 2097 } 2098 2099 return false; 2100 } 2101 2102 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 2103 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 2104 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 2105 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 2106 ConstantExpr *Addr, unsigned OpNo) { 2107 // Base case of the recursion. 2108 if (OpNo == Addr->getNumOperands()) { 2109 assert(Val->getType() == Init->getType() && "Type mismatch!"); 2110 return Val; 2111 } 2112 2113 SmallVector<Constant*, 32> Elts; 2114 if (StructType *STy = dyn_cast<StructType>(Init->getType())) { 2115 // Break up the constant into its elements. 2116 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2117 Elts.push_back(Init->getAggregateElement(i)); 2118 2119 // Replace the element that we are supposed to. 2120 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 2121 unsigned Idx = CU->getZExtValue(); 2122 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 2123 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 2124 2125 // Return the modified struct. 2126 return ConstantStruct::get(STy, Elts); 2127 } 2128 2129 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 2130 SequentialType *InitTy = cast<SequentialType>(Init->getType()); 2131 2132 uint64_t NumElts; 2133 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy)) 2134 NumElts = ATy->getNumElements(); 2135 else 2136 NumElts = InitTy->getVectorNumElements(); 2137 2138 // Break up the array into elements. 2139 for (uint64_t i = 0, e = NumElts; i != e; ++i) 2140 Elts.push_back(Init->getAggregateElement(i)); 2141 2142 assert(CI->getZExtValue() < NumElts); 2143 Elts[CI->getZExtValue()] = 2144 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 2145 2146 if (Init->getType()->isArrayTy()) 2147 return ConstantArray::get(cast<ArrayType>(InitTy), Elts); 2148 return ConstantVector::get(Elts); 2149 } 2150 2151 /// CommitValueTo - We have decided that Addr (which satisfies the predicate 2152 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 2153 static void CommitValueTo(Constant *Val, Constant *Addr) { 2154 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 2155 assert(GV->hasInitializer()); 2156 GV->setInitializer(Val); 2157 return; 2158 } 2159 2160 ConstantExpr *CE = cast<ConstantExpr>(Addr); 2161 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2162 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2)); 2163 } 2164 2165 namespace { 2166 2167 /// Evaluator - This class evaluates LLVM IR, producing the Constant 2168 /// representing each SSA instruction. Changes to global variables are stored 2169 /// in a mapping that can be iterated over after the evaluation is complete. 2170 /// Once an evaluation call fails, the evaluation object should not be reused. 2171 class Evaluator { 2172 public: 2173 Evaluator(const DataLayout &DL, const TargetLibraryInfo *TLI) 2174 : DL(DL), TLI(TLI) { 2175 ValueStack.emplace_back(); 2176 } 2177 2178 ~Evaluator() { 2179 for (auto &Tmp : AllocaTmps) 2180 // If there are still users of the alloca, the program is doing something 2181 // silly, e.g. storing the address of the alloca somewhere and using it 2182 // later. Since this is undefined, we'll just make it be null. 2183 if (!Tmp->use_empty()) 2184 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2185 } 2186 2187 /// EvaluateFunction - Evaluate a call to function F, returning true if 2188 /// successful, false if we can't evaluate it. ActualArgs contains the formal 2189 /// arguments for the function. 2190 bool EvaluateFunction(Function *F, Constant *&RetVal, 2191 const SmallVectorImpl<Constant*> &ActualArgs); 2192 2193 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if 2194 /// successful, false if we can't evaluate it. NewBB returns the next BB that 2195 /// control flows into, or null upon return. 2196 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB); 2197 2198 Constant *getVal(Value *V) { 2199 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 2200 Constant *R = ValueStack.back().lookup(V); 2201 assert(R && "Reference to an uncomputed value!"); 2202 return R; 2203 } 2204 2205 void setVal(Value *V, Constant *C) { 2206 ValueStack.back()[V] = C; 2207 } 2208 2209 const DenseMap<Constant*, Constant*> &getMutatedMemory() const { 2210 return MutatedMemory; 2211 } 2212 2213 const SmallPtrSetImpl<GlobalVariable*> &getInvariants() const { 2214 return Invariants; 2215 } 2216 2217 private: 2218 Constant *ComputeLoadResult(Constant *P); 2219 2220 /// ValueStack - As we compute SSA register values, we store their contents 2221 /// here. The back of the deque contains the current function and the stack 2222 /// contains the values in the calling frames. 2223 std::deque<DenseMap<Value*, Constant*>> ValueStack; 2224 2225 /// CallStack - This is used to detect recursion. In pathological situations 2226 /// we could hit exponential behavior, but at least there is nothing 2227 /// unbounded. 2228 SmallVector<Function*, 4> CallStack; 2229 2230 /// MutatedMemory - For each store we execute, we update this map. Loads 2231 /// check this to get the most up-to-date value. If evaluation is successful, 2232 /// this state is committed to the process. 2233 DenseMap<Constant*, Constant*> MutatedMemory; 2234 2235 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2236 /// to represent its body. This vector is needed so we can delete the 2237 /// temporary globals when we are done. 2238 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps; 2239 2240 /// Invariants - These global variables have been marked invariant by the 2241 /// static constructor. 2242 SmallPtrSet<GlobalVariable*, 8> Invariants; 2243 2244 /// SimpleConstants - These are constants we have checked and know to be 2245 /// simple enough to live in a static initializer of a global. 2246 SmallPtrSet<Constant*, 8> SimpleConstants; 2247 2248 const DataLayout &DL; 2249 const TargetLibraryInfo *TLI; 2250 }; 2251 2252 } // anonymous namespace 2253 2254 /// ComputeLoadResult - Return the value that would be computed by a load from 2255 /// P after the stores reflected by 'memory' have been performed. If we can't 2256 /// decide, return null. 2257 Constant *Evaluator::ComputeLoadResult(Constant *P) { 2258 // If this memory location has been recently stored, use the stored value: it 2259 // is the most up-to-date. 2260 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P); 2261 if (I != MutatedMemory.end()) return I->second; 2262 2263 // Access it. 2264 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 2265 if (GV->hasDefinitiveInitializer()) 2266 return GV->getInitializer(); 2267 return nullptr; 2268 } 2269 2270 // Handle a constantexpr getelementptr. 2271 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 2272 if (CE->getOpcode() == Instruction::GetElementPtr && 2273 isa<GlobalVariable>(CE->getOperand(0))) { 2274 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2275 if (GV->hasDefinitiveInitializer()) 2276 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2277 } 2278 2279 return nullptr; // don't know how to evaluate. 2280 } 2281 2282 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if 2283 /// successful, false if we can't evaluate it. NewBB returns the next BB that 2284 /// control flows into, or null upon return. 2285 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst, 2286 BasicBlock *&NextBB) { 2287 // This is the main evaluation loop. 2288 while (1) { 2289 Constant *InstResult = nullptr; 2290 2291 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n"); 2292 2293 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2294 if (!SI->isSimple()) { 2295 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n"); 2296 return false; // no volatile/atomic accesses. 2297 } 2298 Constant *Ptr = getVal(SI->getOperand(1)); 2299 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) { 2300 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr); 2301 Ptr = ConstantFoldConstantExpression(CE, DL, TLI); 2302 DEBUG(dbgs() << "; To: " << *Ptr << "\n"); 2303 } 2304 if (!isSimpleEnoughPointerToCommit(Ptr)) { 2305 // If this is too complex for us to commit, reject it. 2306 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store."); 2307 return false; 2308 } 2309 2310 Constant *Val = getVal(SI->getOperand(0)); 2311 2312 // If this might be too difficult for the backend to handle (e.g. the addr 2313 // of one global variable divided by another) then we can't commit it. 2314 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) { 2315 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val 2316 << "\n"); 2317 return false; 2318 } 2319 2320 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) { 2321 if (CE->getOpcode() == Instruction::BitCast) { 2322 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n"); 2323 // If we're evaluating a store through a bitcast, then we need 2324 // to pull the bitcast off the pointer type and push it onto the 2325 // stored value. 2326 Ptr = CE->getOperand(0); 2327 2328 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType(); 2329 2330 // In order to push the bitcast onto the stored value, a bitcast 2331 // from NewTy to Val's type must be legal. If it's not, we can try 2332 // introspecting NewTy to find a legal conversion. 2333 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) { 2334 // If NewTy is a struct, we can convert the pointer to the struct 2335 // into a pointer to its first member. 2336 // FIXME: This could be extended to support arrays as well. 2337 if (StructType *STy = dyn_cast<StructType>(NewTy)) { 2338 NewTy = STy->getTypeAtIndex(0U); 2339 2340 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32); 2341 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false); 2342 Constant * const IdxList[] = {IdxZero, IdxZero}; 2343 2344 Ptr = ConstantExpr::getGetElementPtr(nullptr, Ptr, IdxList); 2345 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 2346 Ptr = ConstantFoldConstantExpression(CE, DL, TLI); 2347 2348 // If we can't improve the situation by introspecting NewTy, 2349 // we have to give up. 2350 } else { 2351 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not " 2352 "evaluate.\n"); 2353 return false; 2354 } 2355 } 2356 2357 // If we found compatible types, go ahead and push the bitcast 2358 // onto the stored value. 2359 Val = ConstantExpr::getBitCast(Val, NewTy); 2360 2361 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n"); 2362 } 2363 } 2364 2365 MutatedMemory[Ptr] = Val; 2366 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2367 InstResult = ConstantExpr::get(BO->getOpcode(), 2368 getVal(BO->getOperand(0)), 2369 getVal(BO->getOperand(1))); 2370 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult 2371 << "\n"); 2372 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2373 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2374 getVal(CI->getOperand(0)), 2375 getVal(CI->getOperand(1))); 2376 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult 2377 << "\n"); 2378 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2379 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2380 getVal(CI->getOperand(0)), 2381 CI->getType()); 2382 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult 2383 << "\n"); 2384 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2385 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)), 2386 getVal(SI->getOperand(1)), 2387 getVal(SI->getOperand(2))); 2388 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult 2389 << "\n"); 2390 } else if (auto *EVI = dyn_cast<ExtractValueInst>(CurInst)) { 2391 InstResult = ConstantExpr::getExtractValue( 2392 getVal(EVI->getAggregateOperand()), EVI->getIndices()); 2393 DEBUG(dbgs() << "Found an ExtractValueInst! Simplifying: " << *InstResult 2394 << "\n"); 2395 } else if (auto *IVI = dyn_cast<InsertValueInst>(CurInst)) { 2396 InstResult = ConstantExpr::getInsertValue( 2397 getVal(IVI->getAggregateOperand()), 2398 getVal(IVI->getInsertedValueOperand()), IVI->getIndices()); 2399 DEBUG(dbgs() << "Found an InsertValueInst! Simplifying: " << *InstResult 2400 << "\n"); 2401 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2402 Constant *P = getVal(GEP->getOperand(0)); 2403 SmallVector<Constant*, 8> GEPOps; 2404 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2405 i != e; ++i) 2406 GEPOps.push_back(getVal(*i)); 2407 InstResult = 2408 ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), P, GEPOps, 2409 cast<GEPOperator>(GEP)->isInBounds()); 2410 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult 2411 << "\n"); 2412 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2413 2414 if (!LI->isSimple()) { 2415 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n"); 2416 return false; // no volatile/atomic accesses. 2417 } 2418 2419 Constant *Ptr = getVal(LI->getOperand(0)); 2420 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) { 2421 Ptr = ConstantFoldConstantExpression(CE, DL, TLI); 2422 DEBUG(dbgs() << "Found a constant pointer expression, constant " 2423 "folding: " << *Ptr << "\n"); 2424 } 2425 InstResult = ComputeLoadResult(Ptr); 2426 if (!InstResult) { 2427 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load." 2428 "\n"); 2429 return false; // Could not evaluate load. 2430 } 2431 2432 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n"); 2433 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2434 if (AI->isArrayAllocation()) { 2435 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n"); 2436 return false; // Cannot handle array allocs. 2437 } 2438 Type *Ty = AI->getType()->getElementType(); 2439 AllocaTmps.push_back( 2440 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage, 2441 UndefValue::get(Ty), AI->getName())); 2442 InstResult = AllocaTmps.back().get(); 2443 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n"); 2444 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) { 2445 CallSite CS(CurInst); 2446 2447 // Debug info can safely be ignored here. 2448 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) { 2449 DEBUG(dbgs() << "Ignoring debug info.\n"); 2450 ++CurInst; 2451 continue; 2452 } 2453 2454 // Cannot handle inline asm. 2455 if (isa<InlineAsm>(CS.getCalledValue())) { 2456 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n"); 2457 return false; 2458 } 2459 2460 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 2461 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) { 2462 if (MSI->isVolatile()) { 2463 DEBUG(dbgs() << "Can not optimize a volatile memset " << 2464 "intrinsic.\n"); 2465 return false; 2466 } 2467 Constant *Ptr = getVal(MSI->getDest()); 2468 Constant *Val = getVal(MSI->getValue()); 2469 Constant *DestVal = ComputeLoadResult(getVal(Ptr)); 2470 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) { 2471 // This memset is a no-op. 2472 DEBUG(dbgs() << "Ignoring no-op memset.\n"); 2473 ++CurInst; 2474 continue; 2475 } 2476 } 2477 2478 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2479 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2480 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n"); 2481 ++CurInst; 2482 continue; 2483 } 2484 2485 if (II->getIntrinsicID() == Intrinsic::invariant_start) { 2486 // We don't insert an entry into Values, as it doesn't have a 2487 // meaningful return value. 2488 if (!II->use_empty()) { 2489 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n"); 2490 return false; 2491 } 2492 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0)); 2493 Value *PtrArg = getVal(II->getArgOperand(1)); 2494 Value *Ptr = PtrArg->stripPointerCasts(); 2495 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) { 2496 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType(); 2497 if (!Size->isAllOnesValue() && 2498 Size->getValue().getLimitedValue() >= 2499 DL.getTypeStoreSize(ElemTy)) { 2500 Invariants.insert(GV); 2501 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV 2502 << "\n"); 2503 } else { 2504 DEBUG(dbgs() << "Found a global var, but can not treat it as an " 2505 "invariant.\n"); 2506 } 2507 } 2508 // Continue even if we do nothing. 2509 ++CurInst; 2510 continue; 2511 } 2512 2513 DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n"); 2514 return false; 2515 } 2516 2517 // Resolve function pointers. 2518 Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue())); 2519 if (!Callee || Callee->mayBeOverridden()) { 2520 DEBUG(dbgs() << "Can not resolve function pointer.\n"); 2521 return false; // Cannot resolve. 2522 } 2523 2524 SmallVector<Constant*, 8> Formals; 2525 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i) 2526 Formals.push_back(getVal(*i)); 2527 2528 if (Callee->isDeclaration()) { 2529 // If this is a function we can constant fold, do it. 2530 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) { 2531 InstResult = C; 2532 DEBUG(dbgs() << "Constant folded function call. Result: " << 2533 *InstResult << "\n"); 2534 } else { 2535 DEBUG(dbgs() << "Can not constant fold function call.\n"); 2536 return false; 2537 } 2538 } else { 2539 if (Callee->getFunctionType()->isVarArg()) { 2540 DEBUG(dbgs() << "Can not constant fold vararg function call.\n"); 2541 return false; 2542 } 2543 2544 Constant *RetVal = nullptr; 2545 // Execute the call, if successful, use the return value. 2546 ValueStack.emplace_back(); 2547 if (!EvaluateFunction(Callee, RetVal, Formals)) { 2548 DEBUG(dbgs() << "Failed to evaluate function.\n"); 2549 return false; 2550 } 2551 ValueStack.pop_back(); 2552 InstResult = RetVal; 2553 2554 if (InstResult) { 2555 DEBUG(dbgs() << "Successfully evaluated function. Result: " << 2556 InstResult << "\n\n"); 2557 } else { 2558 DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n"); 2559 } 2560 } 2561 } else if (isa<TerminatorInst>(CurInst)) { 2562 DEBUG(dbgs() << "Found a terminator instruction.\n"); 2563 2564 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2565 if (BI->isUnconditional()) { 2566 NextBB = BI->getSuccessor(0); 2567 } else { 2568 ConstantInt *Cond = 2569 dyn_cast<ConstantInt>(getVal(BI->getCondition())); 2570 if (!Cond) return false; // Cannot determine. 2571 2572 NextBB = BI->getSuccessor(!Cond->getZExtValue()); 2573 } 2574 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2575 ConstantInt *Val = 2576 dyn_cast<ConstantInt>(getVal(SI->getCondition())); 2577 if (!Val) return false; // Cannot determine. 2578 NextBB = SI->findCaseValue(Val).getCaseSuccessor(); 2579 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) { 2580 Value *Val = getVal(IBI->getAddress())->stripPointerCasts(); 2581 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val)) 2582 NextBB = BA->getBasicBlock(); 2583 else 2584 return false; // Cannot determine. 2585 } else if (isa<ReturnInst>(CurInst)) { 2586 NextBB = nullptr; 2587 } else { 2588 // invoke, unwind, resume, unreachable. 2589 DEBUG(dbgs() << "Can not handle terminator."); 2590 return false; // Cannot handle this terminator. 2591 } 2592 2593 // We succeeded at evaluating this block! 2594 DEBUG(dbgs() << "Successfully evaluated block.\n"); 2595 return true; 2596 } else { 2597 // Did not know how to evaluate this! 2598 DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction." 2599 "\n"); 2600 return false; 2601 } 2602 2603 if (!CurInst->use_empty()) { 2604 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult)) 2605 InstResult = ConstantFoldConstantExpression(CE, DL, TLI); 2606 2607 setVal(CurInst, InstResult); 2608 } 2609 2610 // If we just processed an invoke, we finished evaluating the block. 2611 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) { 2612 NextBB = II->getNormalDest(); 2613 DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n"); 2614 return true; 2615 } 2616 2617 // Advance program counter. 2618 ++CurInst; 2619 } 2620 } 2621 2622 /// EvaluateFunction - Evaluate a call to function F, returning true if 2623 /// successful, false if we can't evaluate it. ActualArgs contains the formal 2624 /// arguments for the function. 2625 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal, 2626 const SmallVectorImpl<Constant*> &ActualArgs) { 2627 // Check to see if this function is already executing (recursion). If so, 2628 // bail out. TODO: we might want to accept limited recursion. 2629 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 2630 return false; 2631 2632 CallStack.push_back(F); 2633 2634 // Initialize arguments to the incoming values specified. 2635 unsigned ArgNo = 0; 2636 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 2637 ++AI, ++ArgNo) 2638 setVal(AI, ActualArgs[ArgNo]); 2639 2640 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 2641 // we can only evaluate any one basic block at most once. This set keeps 2642 // track of what we have executed so we can detect recursive cases etc. 2643 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 2644 2645 // CurBB - The current basic block we're evaluating. 2646 BasicBlock *CurBB = F->begin(); 2647 2648 BasicBlock::iterator CurInst = CurBB->begin(); 2649 2650 while (1) { 2651 BasicBlock *NextBB = nullptr; // Initialized to avoid compiler warnings. 2652 DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n"); 2653 2654 if (!EvaluateBlock(CurInst, NextBB)) 2655 return false; 2656 2657 if (!NextBB) { 2658 // Successfully running until there's no next block means that we found 2659 // the return. Fill it the return value and pop the call stack. 2660 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator()); 2661 if (RI->getNumOperands()) 2662 RetVal = getVal(RI->getOperand(0)); 2663 CallStack.pop_back(); 2664 return true; 2665 } 2666 2667 // Okay, we succeeded in evaluating this control flow. See if we have 2668 // executed the new block before. If so, we have a looping function, 2669 // which we cannot evaluate in reasonable time. 2670 if (!ExecutedBlocks.insert(NextBB).second) 2671 return false; // looped! 2672 2673 // Okay, we have never been in this block before. Check to see if there 2674 // are any PHI nodes. If so, evaluate them with information about where 2675 // we came from. 2676 PHINode *PN = nullptr; 2677 for (CurInst = NextBB->begin(); 2678 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2679 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB))); 2680 2681 // Advance to the next block. 2682 CurBB = NextBB; 2683 } 2684 } 2685 2686 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2687 /// we can. Return true if we can, false otherwise. 2688 static bool EvaluateStaticConstructor(Function *F, const DataLayout &DL, 2689 const TargetLibraryInfo *TLI) { 2690 // Call the function. 2691 Evaluator Eval(DL, TLI); 2692 Constant *RetValDummy; 2693 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy, 2694 SmallVector<Constant*, 0>()); 2695 2696 if (EvalSuccess) { 2697 ++NumCtorsEvaluated; 2698 2699 // We succeeded at evaluation: commit the result. 2700 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2701 << F->getName() << "' to " << Eval.getMutatedMemory().size() 2702 << " stores.\n"); 2703 for (DenseMap<Constant*, Constant*>::const_iterator I = 2704 Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end(); 2705 I != E; ++I) 2706 CommitValueTo(I->second, I->first); 2707 for (GlobalVariable *GV : Eval.getInvariants()) 2708 GV->setConstant(true); 2709 } 2710 2711 return EvalSuccess; 2712 } 2713 2714 static int compareNames(Constant *const *A, Constant *const *B) { 2715 return (*A)->getName().compare((*B)->getName()); 2716 } 2717 2718 static void setUsedInitializer(GlobalVariable &V, 2719 const SmallPtrSet<GlobalValue *, 8> &Init) { 2720 if (Init.empty()) { 2721 V.eraseFromParent(); 2722 return; 2723 } 2724 2725 // Type of pointer to the array of pointers. 2726 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0); 2727 2728 SmallVector<llvm::Constant *, 8> UsedArray; 2729 for (GlobalValue *GV : Init) { 2730 Constant *Cast 2731 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy); 2732 UsedArray.push_back(Cast); 2733 } 2734 // Sort to get deterministic order. 2735 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames); 2736 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size()); 2737 2738 Module *M = V.getParent(); 2739 V.removeFromParent(); 2740 GlobalVariable *NV = 2741 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage, 2742 llvm::ConstantArray::get(ATy, UsedArray), ""); 2743 NV->takeName(&V); 2744 NV->setSection("llvm.metadata"); 2745 delete &V; 2746 } 2747 2748 namespace { 2749 /// \brief An easy to access representation of llvm.used and llvm.compiler.used. 2750 class LLVMUsed { 2751 SmallPtrSet<GlobalValue *, 8> Used; 2752 SmallPtrSet<GlobalValue *, 8> CompilerUsed; 2753 GlobalVariable *UsedV; 2754 GlobalVariable *CompilerUsedV; 2755 2756 public: 2757 LLVMUsed(Module &M) { 2758 UsedV = collectUsedGlobalVariables(M, Used, false); 2759 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true); 2760 } 2761 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator; 2762 typedef iterator_range<iterator> used_iterator_range; 2763 iterator usedBegin() { return Used.begin(); } 2764 iterator usedEnd() { return Used.end(); } 2765 used_iterator_range used() { 2766 return used_iterator_range(usedBegin(), usedEnd()); 2767 } 2768 iterator compilerUsedBegin() { return CompilerUsed.begin(); } 2769 iterator compilerUsedEnd() { return CompilerUsed.end(); } 2770 used_iterator_range compilerUsed() { 2771 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd()); 2772 } 2773 bool usedCount(GlobalValue *GV) const { return Used.count(GV); } 2774 bool compilerUsedCount(GlobalValue *GV) const { 2775 return CompilerUsed.count(GV); 2776 } 2777 bool usedErase(GlobalValue *GV) { return Used.erase(GV); } 2778 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); } 2779 bool usedInsert(GlobalValue *GV) { return Used.insert(GV).second; } 2780 bool compilerUsedInsert(GlobalValue *GV) { 2781 return CompilerUsed.insert(GV).second; 2782 } 2783 2784 void syncVariablesAndSets() { 2785 if (UsedV) 2786 setUsedInitializer(*UsedV, Used); 2787 if (CompilerUsedV) 2788 setUsedInitializer(*CompilerUsedV, CompilerUsed); 2789 } 2790 }; 2791 } 2792 2793 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) { 2794 if (GA.use_empty()) // No use at all. 2795 return false; 2796 2797 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) && 2798 "We should have removed the duplicated " 2799 "element from llvm.compiler.used"); 2800 if (!GA.hasOneUse()) 2801 // Strictly more than one use. So at least one is not in llvm.used and 2802 // llvm.compiler.used. 2803 return true; 2804 2805 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used. 2806 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA); 2807 } 2808 2809 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V, 2810 const LLVMUsed &U) { 2811 unsigned N = 2; 2812 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) && 2813 "We should have removed the duplicated " 2814 "element from llvm.compiler.used"); 2815 if (U.usedCount(&V) || U.compilerUsedCount(&V)) 2816 ++N; 2817 return V.hasNUsesOrMore(N); 2818 } 2819 2820 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) { 2821 if (!GA.hasLocalLinkage()) 2822 return true; 2823 2824 return U.usedCount(&GA) || U.compilerUsedCount(&GA); 2825 } 2826 2827 static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U, 2828 bool &RenameTarget) { 2829 RenameTarget = false; 2830 bool Ret = false; 2831 if (hasUseOtherThanLLVMUsed(GA, U)) 2832 Ret = true; 2833 2834 // If the alias is externally visible, we may still be able to simplify it. 2835 if (!mayHaveOtherReferences(GA, U)) 2836 return Ret; 2837 2838 // If the aliasee has internal linkage, give it the name and linkage 2839 // of the alias, and delete the alias. This turns: 2840 // define internal ... @f(...) 2841 // @a = alias ... @f 2842 // into: 2843 // define ... @a(...) 2844 Constant *Aliasee = GA.getAliasee(); 2845 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts()); 2846 if (!Target->hasLocalLinkage()) 2847 return Ret; 2848 2849 // Do not perform the transform if multiple aliases potentially target the 2850 // aliasee. This check also ensures that it is safe to replace the section 2851 // and other attributes of the aliasee with those of the alias. 2852 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U)) 2853 return Ret; 2854 2855 RenameTarget = true; 2856 return true; 2857 } 2858 2859 bool GlobalOpt::OptimizeGlobalAliases(Module &M) { 2860 bool Changed = false; 2861 LLVMUsed Used(M); 2862 2863 for (GlobalValue *GV : Used.used()) 2864 Used.compilerUsedErase(GV); 2865 2866 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 2867 I != E;) { 2868 Module::alias_iterator J = I++; 2869 // Aliases without names cannot be referenced outside this module. 2870 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage()) 2871 J->setLinkage(GlobalValue::InternalLinkage); 2872 // If the aliasee may change at link time, nothing can be done - bail out. 2873 if (J->mayBeOverridden()) 2874 continue; 2875 2876 Constant *Aliasee = J->getAliasee(); 2877 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts()); 2878 // We can't trivially replace the alias with the aliasee if the aliasee is 2879 // non-trivial in some way. 2880 // TODO: Try to handle non-zero GEPs of local aliasees. 2881 if (!Target) 2882 continue; 2883 Target->removeDeadConstantUsers(); 2884 2885 // Make all users of the alias use the aliasee instead. 2886 bool RenameTarget; 2887 if (!hasUsesToReplace(*J, Used, RenameTarget)) 2888 continue; 2889 2890 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType())); 2891 ++NumAliasesResolved; 2892 Changed = true; 2893 2894 if (RenameTarget) { 2895 // Give the aliasee the name, linkage and other attributes of the alias. 2896 Target->takeName(J); 2897 Target->setLinkage(J->getLinkage()); 2898 Target->setVisibility(J->getVisibility()); 2899 Target->setDLLStorageClass(J->getDLLStorageClass()); 2900 2901 if (Used.usedErase(J)) 2902 Used.usedInsert(Target); 2903 2904 if (Used.compilerUsedErase(J)) 2905 Used.compilerUsedInsert(Target); 2906 } else if (mayHaveOtherReferences(*J, Used)) 2907 continue; 2908 2909 // Delete the alias. 2910 M.getAliasList().erase(J); 2911 ++NumAliasesRemoved; 2912 Changed = true; 2913 } 2914 2915 Used.syncVariablesAndSets(); 2916 2917 return Changed; 2918 } 2919 2920 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) { 2921 if (!TLI->has(LibFunc::cxa_atexit)) 2922 return nullptr; 2923 2924 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit)); 2925 2926 if (!Fn) 2927 return nullptr; 2928 2929 FunctionType *FTy = Fn->getFunctionType(); 2930 2931 // Checking that the function has the right return type, the right number of 2932 // parameters and that they all have pointer types should be enough. 2933 if (!FTy->getReturnType()->isIntegerTy() || 2934 FTy->getNumParams() != 3 || 2935 !FTy->getParamType(0)->isPointerTy() || 2936 !FTy->getParamType(1)->isPointerTy() || 2937 !FTy->getParamType(2)->isPointerTy()) 2938 return nullptr; 2939 2940 return Fn; 2941 } 2942 2943 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++ 2944 /// destructor and can therefore be eliminated. 2945 /// Note that we assume that other optimization passes have already simplified 2946 /// the code so we only look for a function with a single basic block, where 2947 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and 2948 /// other side-effect free instructions. 2949 static bool cxxDtorIsEmpty(const Function &Fn, 2950 SmallPtrSet<const Function *, 8> &CalledFunctions) { 2951 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and 2952 // nounwind, but that doesn't seem worth doing. 2953 if (Fn.isDeclaration()) 2954 return false; 2955 2956 if (++Fn.begin() != Fn.end()) 2957 return false; 2958 2959 const BasicBlock &EntryBlock = Fn.getEntryBlock(); 2960 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end(); 2961 I != E; ++I) { 2962 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 2963 // Ignore debug intrinsics. 2964 if (isa<DbgInfoIntrinsic>(CI)) 2965 continue; 2966 2967 const Function *CalledFn = CI->getCalledFunction(); 2968 2969 if (!CalledFn) 2970 return false; 2971 2972 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions); 2973 2974 // Don't treat recursive functions as empty. 2975 if (!NewCalledFunctions.insert(CalledFn).second) 2976 return false; 2977 2978 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions)) 2979 return false; 2980 } else if (isa<ReturnInst>(*I)) 2981 return true; // We're done. 2982 else if (I->mayHaveSideEffects()) 2983 return false; // Destructor with side effects, bail. 2984 } 2985 2986 return false; 2987 } 2988 2989 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) { 2990 /// Itanium C++ ABI p3.3.5: 2991 /// 2992 /// After constructing a global (or local static) object, that will require 2993 /// destruction on exit, a termination function is registered as follows: 2994 /// 2995 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d ); 2996 /// 2997 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the 2998 /// call f(p) when DSO d is unloaded, before all such termination calls 2999 /// registered before this one. It returns zero if registration is 3000 /// successful, nonzero on failure. 3001 3002 // This pass will look for calls to __cxa_atexit where the function is trivial 3003 // and remove them. 3004 bool Changed = false; 3005 3006 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end(); 3007 I != E;) { 3008 // We're only interested in calls. Theoretically, we could handle invoke 3009 // instructions as well, but neither llvm-gcc nor clang generate invokes 3010 // to __cxa_atexit. 3011 CallInst *CI = dyn_cast<CallInst>(*I++); 3012 if (!CI) 3013 continue; 3014 3015 Function *DtorFn = 3016 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts()); 3017 if (!DtorFn) 3018 continue; 3019 3020 SmallPtrSet<const Function *, 8> CalledFunctions; 3021 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions)) 3022 continue; 3023 3024 // Just remove the call. 3025 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 3026 CI->eraseFromParent(); 3027 3028 ++NumCXXDtorsRemoved; 3029 3030 Changed |= true; 3031 } 3032 3033 return Changed; 3034 } 3035 3036 bool GlobalOpt::runOnModule(Module &M) { 3037 bool Changed = false; 3038 3039 auto &DL = M.getDataLayout(); 3040 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 3041 3042 bool LocalChange = true; 3043 while (LocalChange) { 3044 LocalChange = false; 3045 3046 NotDiscardableComdats.clear(); 3047 for (const GlobalVariable &GV : M.globals()) 3048 if (const Comdat *C = GV.getComdat()) 3049 if (!GV.isDiscardableIfUnused() || !GV.use_empty()) 3050 NotDiscardableComdats.insert(C); 3051 for (Function &F : M) 3052 if (const Comdat *C = F.getComdat()) 3053 if (!F.isDefTriviallyDead()) 3054 NotDiscardableComdats.insert(C); 3055 for (GlobalAlias &GA : M.aliases()) 3056 if (const Comdat *C = GA.getComdat()) 3057 if (!GA.isDiscardableIfUnused() || !GA.use_empty()) 3058 NotDiscardableComdats.insert(C); 3059 3060 // Delete functions that are trivially dead, ccc -> fastcc 3061 LocalChange |= OptimizeFunctions(M); 3062 3063 // Optimize global_ctors list. 3064 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) { 3065 return EvaluateStaticConstructor(F, DL, TLI); 3066 }); 3067 3068 // Optimize non-address-taken globals. 3069 LocalChange |= OptimizeGlobalVars(M); 3070 3071 // Resolve aliases, when possible. 3072 LocalChange |= OptimizeGlobalAliases(M); 3073 3074 // Try to remove trivial global destructors if they are not removed 3075 // already. 3076 Function *CXAAtExitFn = FindCXAAtExit(M, TLI); 3077 if (CXAAtExitFn) 3078 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn); 3079 3080 Changed |= LocalChange; 3081 } 3082 3083 // TODO: Move all global ctors functions to the end of the module for code 3084 // layout. 3085 3086 return Changed; 3087 } 3088