GlobalOpt.cpp revision 309f20fc45a599ca45ad1262525d9f4c332b1911
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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#define DEBUG_TYPE "globalopt" 17#include "llvm/Transforms/IPO.h" 18#include "llvm/CallingConv.h" 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Instructions.h" 22#include "llvm/IntrinsicInst.h" 23#include "llvm/Module.h" 24#include "llvm/Pass.h" 25#include "llvm/Analysis/ConstantFolding.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Support/Compiler.h" 28#include "llvm/Support/Debug.h" 29#include "llvm/ADT/SmallPtrSet.h" 30#include "llvm/ADT/SmallVector.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/StringExtras.h" 33#include <algorithm> 34#include <set> 35using namespace llvm; 36 37STATISTIC(NumMarked , "Number of globals marked constant"); 38STATISTIC(NumSRA , "Number of aggregate globals broken into scalars"); 39STATISTIC(NumHeapSRA , "Number of heap objects SRA'd"); 40STATISTIC(NumSubstitute,"Number of globals with initializers stored into them"); 41STATISTIC(NumDeleted , "Number of globals deleted"); 42STATISTIC(NumFnDeleted , "Number of functions deleted"); 43STATISTIC(NumGlobUses , "Number of global uses devirtualized"); 44STATISTIC(NumLocalized , "Number of globals localized"); 45STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans"); 46STATISTIC(NumFastCallFns , "Number of functions converted to fastcc"); 47STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated"); 48 49namespace { 50 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass { 51 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 52 AU.addRequired<TargetData>(); 53 } 54 static char ID; // Pass identification, replacement for typeid 55 GlobalOpt() : ModulePass((intptr_t)&ID) {} 56 57 bool runOnModule(Module &M); 58 59 private: 60 GlobalVariable *FindGlobalCtors(Module &M); 61 bool OptimizeFunctions(Module &M); 62 bool OptimizeGlobalVars(Module &M); 63 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL); 64 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 65 }; 66 67 char GlobalOpt::ID = 0; 68 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer"); 69} 70 71ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 72 73/// GlobalStatus - As we analyze each global, keep track of some information 74/// about it. If we find out that the address of the global is taken, none of 75/// this info will be accurate. 76struct VISIBILITY_HIDDEN GlobalStatus { 77 /// isLoaded - True if the global is ever loaded. If the global isn't ever 78 /// loaded it can be deleted. 79 bool isLoaded; 80 81 /// StoredType - Keep track of what stores to the global look like. 82 /// 83 enum StoredType { 84 /// NotStored - There is no store to this global. It can thus be marked 85 /// constant. 86 NotStored, 87 88 /// isInitializerStored - This global is stored to, but the only thing 89 /// stored is the constant it was initialized with. This is only tracked 90 /// for scalar globals. 91 isInitializerStored, 92 93 /// isStoredOnce - This global is stored to, but only its initializer and 94 /// one other value is ever stored to it. If this global isStoredOnce, we 95 /// track the value stored to it in StoredOnceValue below. This is only 96 /// tracked for scalar globals. 97 isStoredOnce, 98 99 /// isStored - This global is stored to by multiple values or something else 100 /// that we cannot track. 101 isStored 102 } StoredType; 103 104 /// StoredOnceValue - If only one value (besides the initializer constant) is 105 /// ever stored to this global, keep track of what value it is. 106 Value *StoredOnceValue; 107 108 /// AccessingFunction/HasMultipleAccessingFunctions - These start out 109 /// null/false. When the first accessing function is noticed, it is recorded. 110 /// When a second different accessing function is noticed, 111 /// HasMultipleAccessingFunctions is set to true. 112 Function *AccessingFunction; 113 bool HasMultipleAccessingFunctions; 114 115 /// HasNonInstructionUser - Set to true if this global has a user that is not 116 /// an instruction (e.g. a constant expr or GV initializer). 117 bool HasNonInstructionUser; 118 119 /// HasPHIUser - Set to true if this global has a user that is a PHI node. 120 bool HasPHIUser; 121 122 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of 123 /// the global exist. Such users include GEP instruction with variable 124 /// indexes, and non-gep/load/store users like constant expr casts. 125 bool isNotSuitableForSRA; 126 127 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0), 128 AccessingFunction(0), HasMultipleAccessingFunctions(false), 129 HasNonInstructionUser(false), HasPHIUser(false), 130 isNotSuitableForSRA(false) {} 131}; 132 133 134 135/// ConstantIsDead - Return true if the specified constant is (transitively) 136/// dead. The constant may be used by other constants (e.g. constant arrays and 137/// constant exprs) as long as they are dead, but it cannot be used by anything 138/// else. 139static bool ConstantIsDead(Constant *C) { 140 if (isa<GlobalValue>(C)) return false; 141 142 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI) 143 if (Constant *CU = dyn_cast<Constant>(*UI)) { 144 if (!ConstantIsDead(CU)) return false; 145 } else 146 return false; 147 return true; 148} 149 150 151/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 152/// structure. If the global has its address taken, return true to indicate we 153/// can't do anything with it. 154/// 155static bool AnalyzeGlobal(Value *V, GlobalStatus &GS, 156 std::set<PHINode*> &PHIUsers) { 157 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) { 159 GS.HasNonInstructionUser = true; 160 161 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 162 if (CE->getOpcode() != Instruction::GetElementPtr) 163 GS.isNotSuitableForSRA = true; 164 else if (!GS.isNotSuitableForSRA) { 165 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 166 // don't like < 3 operand CE's, and we don't like non-constant integer 167 // indices. 168 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue()) 169 GS.isNotSuitableForSRA = true; 170 else { 171 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 172 if (!isa<ConstantInt>(CE->getOperand(i))) { 173 GS.isNotSuitableForSRA = true; 174 break; 175 } 176 } 177 } 178 179 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) { 180 if (!GS.HasMultipleAccessingFunctions) { 181 Function *F = I->getParent()->getParent(); 182 if (GS.AccessingFunction == 0) 183 GS.AccessingFunction = F; 184 else if (GS.AccessingFunction != F) 185 GS.HasMultipleAccessingFunctions = true; 186 } 187 if (isa<LoadInst>(I)) { 188 GS.isLoaded = true; 189 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 190 // Don't allow a store OF the address, only stores TO the address. 191 if (SI->getOperand(0) == V) return true; 192 193 // If this is a direct store to the global (i.e., the global is a scalar 194 // value, not an aggregate), keep more specific information about 195 // stores. 196 if (GS.StoredType != GlobalStatus::isStored) 197 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){ 198 Value *StoredVal = SI->getOperand(0); 199 if (StoredVal == GV->getInitializer()) { 200 if (GS.StoredType < GlobalStatus::isInitializerStored) 201 GS.StoredType = GlobalStatus::isInitializerStored; 202 } else if (isa<LoadInst>(StoredVal) && 203 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 204 // G = G 205 if (GS.StoredType < GlobalStatus::isInitializerStored) 206 GS.StoredType = GlobalStatus::isInitializerStored; 207 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 208 GS.StoredType = GlobalStatus::isStoredOnce; 209 GS.StoredOnceValue = StoredVal; 210 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 211 GS.StoredOnceValue == StoredVal) { 212 // noop. 213 } else { 214 GS.StoredType = GlobalStatus::isStored; 215 } 216 } else { 217 GS.StoredType = GlobalStatus::isStored; 218 } 219 } else if (isa<GetElementPtrInst>(I)) { 220 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 221 222 // If the first two indices are constants, this can be SRA'd. 223 if (isa<GlobalVariable>(I->getOperand(0))) { 224 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) || 225 !cast<Constant>(I->getOperand(1))->isNullValue() || 226 !isa<ConstantInt>(I->getOperand(2))) 227 GS.isNotSuitableForSRA = true; 228 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){ 229 if (CE->getOpcode() != Instruction::GetElementPtr || 230 CE->getNumOperands() < 3 || I->getNumOperands() < 2 || 231 !isa<Constant>(I->getOperand(0)) || 232 !cast<Constant>(I->getOperand(0))->isNullValue()) 233 GS.isNotSuitableForSRA = true; 234 } else { 235 GS.isNotSuitableForSRA = true; 236 } 237 } else if (isa<SelectInst>(I)) { 238 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 239 GS.isNotSuitableForSRA = true; 240 } else if (PHINode *PN = dyn_cast<PHINode>(I)) { 241 // PHI nodes we can check just like select or GEP instructions, but we 242 // have to be careful about infinite recursion. 243 if (PHIUsers.insert(PN).second) // Not already visited. 244 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 245 GS.isNotSuitableForSRA = true; 246 GS.HasPHIUser = true; 247 } else if (isa<CmpInst>(I)) { 248 GS.isNotSuitableForSRA = true; 249 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) { 250 if (I->getOperand(1) == V) 251 GS.StoredType = GlobalStatus::isStored; 252 if (I->getOperand(2) == V) 253 GS.isLoaded = true; 254 GS.isNotSuitableForSRA = true; 255 } else if (isa<MemSetInst>(I)) { 256 assert(I->getOperand(1) == V && "Memset only takes one pointer!"); 257 GS.StoredType = GlobalStatus::isStored; 258 GS.isNotSuitableForSRA = true; 259 } else { 260 return true; // Any other non-load instruction might take address! 261 } 262 } else if (Constant *C = dyn_cast<Constant>(*UI)) { 263 GS.HasNonInstructionUser = true; 264 // We might have a dead and dangling constant hanging off of here. 265 if (!ConstantIsDead(C)) 266 return true; 267 } else { 268 GS.HasNonInstructionUser = true; 269 // Otherwise must be some other user. 270 return true; 271 } 272 273 return false; 274} 275 276static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 277 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 278 if (!CI) return 0; 279 unsigned IdxV = CI->getZExtValue(); 280 281 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 282 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 283 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 284 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 285 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 286 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 287 } else if (isa<ConstantAggregateZero>(Agg)) { 288 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 289 if (IdxV < STy->getNumElements()) 290 return Constant::getNullValue(STy->getElementType(IdxV)); 291 } else if (const SequentialType *STy = 292 dyn_cast<SequentialType>(Agg->getType())) { 293 return Constant::getNullValue(STy->getElementType()); 294 } 295 } else if (isa<UndefValue>(Agg)) { 296 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 297 if (IdxV < STy->getNumElements()) 298 return UndefValue::get(STy->getElementType(IdxV)); 299 } else if (const SequentialType *STy = 300 dyn_cast<SequentialType>(Agg->getType())) { 301 return UndefValue::get(STy->getElementType()); 302 } 303 } 304 return 0; 305} 306 307 308/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 309/// users of the global, cleaning up the obvious ones. This is largely just a 310/// quick scan over the use list to clean up the easy and obvious cruft. This 311/// returns true if it made a change. 312static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 313 bool Changed = false; 314 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 315 User *U = *UI++; 316 317 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 318 if (Init) { 319 // Replace the load with the initializer. 320 LI->replaceAllUsesWith(Init); 321 LI->eraseFromParent(); 322 Changed = true; 323 } 324 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 325 // Store must be unreachable or storing Init into the global. 326 SI->eraseFromParent(); 327 Changed = true; 328 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 329 if (CE->getOpcode() == Instruction::GetElementPtr) { 330 Constant *SubInit = 0; 331 if (Init) 332 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 333 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 334 } else if (CE->getOpcode() == Instruction::BitCast && 335 isa<PointerType>(CE->getType())) { 336 // Pointer cast, delete any stores and memsets to the global. 337 Changed |= CleanupConstantGlobalUsers(CE, 0); 338 } 339 340 if (CE->use_empty()) { 341 CE->destroyConstant(); 342 Changed = true; 343 } 344 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 345 Constant *SubInit = 0; 346 ConstantExpr *CE = 347 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 348 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 349 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 350 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 351 352 if (GEP->use_empty()) { 353 GEP->eraseFromParent(); 354 Changed = true; 355 } 356 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 357 if (MI->getRawDest() == V) { 358 MI->eraseFromParent(); 359 Changed = true; 360 } 361 362 } else if (Constant *C = dyn_cast<Constant>(U)) { 363 // If we have a chain of dead constantexprs or other things dangling from 364 // us, and if they are all dead, nuke them without remorse. 365 if (ConstantIsDead(C)) { 366 C->destroyConstant(); 367 // This could have invalidated UI, start over from scratch. 368 CleanupConstantGlobalUsers(V, Init); 369 return true; 370 } 371 } 372 } 373 return Changed; 374} 375 376/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 377/// variable. This opens the door for other optimizations by exposing the 378/// behavior of the program in a more fine-grained way. We have determined that 379/// this transformation is safe already. We return the first global variable we 380/// insert so that the caller can reprocess it. 381static GlobalVariable *SRAGlobal(GlobalVariable *GV) { 382 assert(GV->hasInternalLinkage() && !GV->isConstant()); 383 Constant *Init = GV->getInitializer(); 384 const Type *Ty = Init->getType(); 385 386 std::vector<GlobalVariable*> NewGlobals; 387 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 388 389 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 390 NewGlobals.reserve(STy->getNumElements()); 391 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 392 Constant *In = getAggregateConstantElement(Init, 393 ConstantInt::get(Type::Int32Ty, i)); 394 assert(In && "Couldn't get element of initializer?"); 395 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 396 GlobalVariable::InternalLinkage, 397 In, GV->getName()+"."+utostr(i), 398 (Module *)NULL, 399 GV->isThreadLocal()); 400 Globals.insert(GV, NGV); 401 NewGlobals.push_back(NGV); 402 } 403 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 404 unsigned NumElements = 0; 405 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 406 NumElements = ATy->getNumElements(); 407 else if (const VectorType *PTy = dyn_cast<VectorType>(STy)) 408 NumElements = PTy->getNumElements(); 409 else 410 assert(0 && "Unknown aggregate sequential type!"); 411 412 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 413 return 0; // It's not worth it. 414 NewGlobals.reserve(NumElements); 415 for (unsigned i = 0, e = NumElements; i != e; ++i) { 416 Constant *In = getAggregateConstantElement(Init, 417 ConstantInt::get(Type::Int32Ty, i)); 418 assert(In && "Couldn't get element of initializer?"); 419 420 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 421 GlobalVariable::InternalLinkage, 422 In, GV->getName()+"."+utostr(i), 423 (Module *)NULL, 424 GV->isThreadLocal()); 425 Globals.insert(GV, NGV); 426 NewGlobals.push_back(NGV); 427 } 428 } 429 430 if (NewGlobals.empty()) 431 return 0; 432 433 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 434 435 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 436 437 // Loop over all of the uses of the global, replacing the constantexpr geps, 438 // with smaller constantexpr geps or direct references. 439 while (!GV->use_empty()) { 440 User *GEP = GV->use_back(); 441 assert(((isa<ConstantExpr>(GEP) && 442 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 443 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 444 445 // Ignore the 1th operand, which has to be zero or else the program is quite 446 // broken (undefined). Get the 2nd operand, which is the structure or array 447 // index. 448 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 449 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 450 451 Value *NewPtr = NewGlobals[Val]; 452 453 // Form a shorter GEP if needed. 454 if (GEP->getNumOperands() > 3) 455 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 456 SmallVector<Constant*, 8> Idxs; 457 Idxs.push_back(NullInt); 458 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 459 Idxs.push_back(CE->getOperand(i)); 460 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 461 &Idxs[0], Idxs.size()); 462 } else { 463 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 464 SmallVector<Value*, 8> Idxs; 465 Idxs.push_back(NullInt); 466 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 467 Idxs.push_back(GEPI->getOperand(i)); 468 NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(), 469 GEPI->getName()+"."+utostr(Val), GEPI); 470 } 471 GEP->replaceAllUsesWith(NewPtr); 472 473 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 474 GEPI->eraseFromParent(); 475 else 476 cast<ConstantExpr>(GEP)->destroyConstant(); 477 } 478 479 // Delete the old global, now that it is dead. 480 Globals.erase(GV); 481 ++NumSRA; 482 483 // Loop over the new globals array deleting any globals that are obviously 484 // dead. This can arise due to scalarization of a structure or an array that 485 // has elements that are dead. 486 unsigned FirstGlobal = 0; 487 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 488 if (NewGlobals[i]->use_empty()) { 489 Globals.erase(NewGlobals[i]); 490 if (FirstGlobal == i) ++FirstGlobal; 491 } 492 493 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 494} 495 496/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 497/// value will trap if the value is dynamically null. PHIs keeps track of any 498/// phi nodes we've seen to avoid reprocessing them. 499static bool AllUsesOfValueWillTrapIfNull(Value *V, 500 SmallPtrSet<PHINode*, 8> &PHIs) { 501 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 502 if (isa<LoadInst>(*UI)) { 503 // Will trap. 504 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 505 if (SI->getOperand(0) == V) { 506 //cerr << "NONTRAPPING USE: " << **UI; 507 return false; // Storing the value. 508 } 509 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 510 if (CI->getOperand(0) != V) { 511 //cerr << "NONTRAPPING USE: " << **UI; 512 return false; // Not calling the ptr 513 } 514 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 515 if (II->getOperand(0) != V) { 516 //cerr << "NONTRAPPING USE: " << **UI; 517 return false; // Not calling the ptr 518 } 519 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 520 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 521 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 522 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 523 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 524 // If we've already seen this phi node, ignore it, it has already been 525 // checked. 526 if (PHIs.insert(PN)) 527 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 528 } else if (isa<ICmpInst>(*UI) && 529 isa<ConstantPointerNull>(UI->getOperand(1))) { 530 // Ignore setcc X, null 531 } else { 532 //cerr << "NONTRAPPING USE: " << **UI; 533 return false; 534 } 535 return true; 536} 537 538/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 539/// from GV will trap if the loaded value is null. Note that this also permits 540/// comparisons of the loaded value against null, as a special case. 541static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 542 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 543 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 544 SmallPtrSet<PHINode*, 8> PHIs; 545 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 546 return false; 547 } else if (isa<StoreInst>(*UI)) { 548 // Ignore stores to the global. 549 } else { 550 // We don't know or understand this user, bail out. 551 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 552 return false; 553 } 554 555 return true; 556} 557 558static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 559 bool Changed = false; 560 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 561 Instruction *I = cast<Instruction>(*UI++); 562 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 563 LI->setOperand(0, NewV); 564 Changed = true; 565 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 566 if (SI->getOperand(1) == V) { 567 SI->setOperand(1, NewV); 568 Changed = true; 569 } 570 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 571 if (I->getOperand(0) == V) { 572 // Calling through the pointer! Turn into a direct call, but be careful 573 // that the pointer is not also being passed as an argument. 574 I->setOperand(0, NewV); 575 Changed = true; 576 bool PassedAsArg = false; 577 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 578 if (I->getOperand(i) == V) { 579 PassedAsArg = true; 580 I->setOperand(i, NewV); 581 } 582 583 if (PassedAsArg) { 584 // Being passed as an argument also. Be careful to not invalidate UI! 585 UI = V->use_begin(); 586 } 587 } 588 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 589 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 590 ConstantExpr::getCast(CI->getOpcode(), 591 NewV, CI->getType())); 592 if (CI->use_empty()) { 593 Changed = true; 594 CI->eraseFromParent(); 595 } 596 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 597 // Should handle GEP here. 598 SmallVector<Constant*, 8> Idxs; 599 Idxs.reserve(GEPI->getNumOperands()-1); 600 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 601 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i))) 602 Idxs.push_back(C); 603 else 604 break; 605 if (Idxs.size() == GEPI->getNumOperands()-1) 606 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 607 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 608 Idxs.size())); 609 if (GEPI->use_empty()) { 610 Changed = true; 611 GEPI->eraseFromParent(); 612 } 613 } 614 } 615 616 return Changed; 617} 618 619 620/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 621/// value stored into it. If there are uses of the loaded value that would trap 622/// if the loaded value is dynamically null, then we know that they cannot be 623/// reachable with a null optimize away the load. 624static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 625 std::vector<LoadInst*> Loads; 626 bool Changed = false; 627 628 // Replace all uses of loads with uses of uses of the stored value. 629 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); 630 GUI != E; ++GUI) 631 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) { 632 Loads.push_back(LI); 633 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 634 } else { 635 // If we get here we could have stores, selects, or phi nodes whose values 636 // are loaded. 637 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) || 638 isa<SelectInst>(*GUI)) && 639 "Only expect load and stores!"); 640 } 641 642 if (Changed) { 643 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 644 ++NumGlobUses; 645 } 646 647 // Delete all of the loads we can, keeping track of whether we nuked them all! 648 bool AllLoadsGone = true; 649 while (!Loads.empty()) { 650 LoadInst *L = Loads.back(); 651 if (L->use_empty()) { 652 L->eraseFromParent(); 653 Changed = true; 654 } else { 655 AllLoadsGone = false; 656 } 657 Loads.pop_back(); 658 } 659 660 // If we nuked all of the loads, then none of the stores are needed either, 661 // nor is the global. 662 if (AllLoadsGone) { 663 DOUT << " *** GLOBAL NOW DEAD!\n"; 664 CleanupConstantGlobalUsers(GV, 0); 665 if (GV->use_empty()) { 666 GV->eraseFromParent(); 667 ++NumDeleted; 668 } 669 Changed = true; 670 } 671 return Changed; 672} 673 674/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 675/// instructions that are foldable. 676static void ConstantPropUsersOf(Value *V) { 677 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 678 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 679 if (Constant *NewC = ConstantFoldInstruction(I)) { 680 I->replaceAllUsesWith(NewC); 681 682 // Advance UI to the next non-I use to avoid invalidating it! 683 // Instructions could multiply use V. 684 while (UI != E && *UI == I) 685 ++UI; 686 I->eraseFromParent(); 687 } 688} 689 690/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 691/// variable, and transforms the program as if it always contained the result of 692/// the specified malloc. Because it is always the result of the specified 693/// malloc, there is no reason to actually DO the malloc. Instead, turn the 694/// malloc into a global, and any loads of GV as uses of the new global. 695static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 696 MallocInst *MI) { 697 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 698 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 699 700 if (NElements->getZExtValue() != 1) { 701 // If we have an array allocation, transform it to a single element 702 // allocation to make the code below simpler. 703 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 704 NElements->getZExtValue()); 705 MallocInst *NewMI = 706 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 707 MI->getAlignment(), MI->getName(), MI); 708 Value* Indices[2]; 709 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 710 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2, 711 NewMI->getName()+".el0", MI); 712 MI->replaceAllUsesWith(NewGEP); 713 MI->eraseFromParent(); 714 MI = NewMI; 715 } 716 717 // Create the new global variable. The contents of the malloc'd memory is 718 // undefined, so initialize with an undef value. 719 Constant *Init = UndefValue::get(MI->getAllocatedType()); 720 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 721 GlobalValue::InternalLinkage, Init, 722 GV->getName()+".body", 723 (Module *)NULL, 724 GV->isThreadLocal()); 725 GV->getParent()->getGlobalList().insert(GV, NewGV); 726 727 // Anything that used the malloc now uses the global directly. 728 MI->replaceAllUsesWith(NewGV); 729 730 Constant *RepValue = NewGV; 731 if (NewGV->getType() != GV->getType()->getElementType()) 732 RepValue = ConstantExpr::getBitCast(RepValue, 733 GV->getType()->getElementType()); 734 735 // If there is a comparison against null, we will insert a global bool to 736 // keep track of whether the global was initialized yet or not. 737 GlobalVariable *InitBool = 738 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 739 ConstantInt::getFalse(), GV->getName()+".init", 740 (Module *)NULL, GV->isThreadLocal()); 741 bool InitBoolUsed = false; 742 743 // Loop over all uses of GV, processing them in turn. 744 std::vector<StoreInst*> Stores; 745 while (!GV->use_empty()) 746 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 747 while (!LI->use_empty()) { 748 Use &LoadUse = LI->use_begin().getUse(); 749 if (!isa<ICmpInst>(LoadUse.getUser())) 750 LoadUse = RepValue; 751 else { 752 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 753 // Replace the cmp X, 0 with a use of the bool value. 754 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 755 InitBoolUsed = true; 756 switch (CI->getPredicate()) { 757 default: assert(0 && "Unknown ICmp Predicate!"); 758 case ICmpInst::ICMP_ULT: 759 case ICmpInst::ICMP_SLT: 760 LV = ConstantInt::getFalse(); // X < null -> always false 761 break; 762 case ICmpInst::ICMP_ULE: 763 case ICmpInst::ICMP_SLE: 764 case ICmpInst::ICMP_EQ: 765 LV = BinaryOperator::createNot(LV, "notinit", CI); 766 break; 767 case ICmpInst::ICMP_NE: 768 case ICmpInst::ICMP_UGE: 769 case ICmpInst::ICMP_SGE: 770 case ICmpInst::ICMP_UGT: 771 case ICmpInst::ICMP_SGT: 772 break; // no change. 773 } 774 CI->replaceAllUsesWith(LV); 775 CI->eraseFromParent(); 776 } 777 } 778 LI->eraseFromParent(); 779 } else { 780 StoreInst *SI = cast<StoreInst>(GV->use_back()); 781 // The global is initialized when the store to it occurs. 782 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 783 SI->eraseFromParent(); 784 } 785 786 // If the initialization boolean was used, insert it, otherwise delete it. 787 if (!InitBoolUsed) { 788 while (!InitBool->use_empty()) // Delete initializations 789 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 790 delete InitBool; 791 } else 792 GV->getParent()->getGlobalList().insert(GV, InitBool); 793 794 795 // Now the GV is dead, nuke it and the malloc. 796 GV->eraseFromParent(); 797 MI->eraseFromParent(); 798 799 // To further other optimizations, loop over all users of NewGV and try to 800 // constant prop them. This will promote GEP instructions with constant 801 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 802 ConstantPropUsersOf(NewGV); 803 if (RepValue != NewGV) 804 ConstantPropUsersOf(RepValue); 805 806 return NewGV; 807} 808 809/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 810/// to make sure that there are no complex uses of V. We permit simple things 811/// like dereferencing the pointer, but not storing through the address, unless 812/// it is to the specified global. 813static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 814 GlobalVariable *GV, 815 SmallPtrSet<PHINode*, 8> &PHIs) { 816 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI) 817 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) { 818 // Fine, ignore. 819 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 820 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 821 return false; // Storing the pointer itself... bad. 822 // Otherwise, storing through it, or storing into GV... fine. 823 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI) || 824 isa<BitCastInst>(*UI)) { 825 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI), 826 GV, PHIs)) 827 return false; 828 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 829 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 830 // cycles. 831 if (PHIs.insert(PN)) 832 return ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs); 833 } else { 834 return false; 835 } 836 return true; 837} 838 839/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 840/// somewhere. Transform all uses of the allocation into loads from the 841/// global and uses of the resultant pointer. Further, delete the store into 842/// GV. This assumes that these value pass the 843/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 844static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 845 GlobalVariable *GV) { 846 while (!Alloc->use_empty()) { 847 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 848 Instruction *InsertPt = U; 849 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 850 // If this is the store of the allocation into the global, remove it. 851 if (SI->getOperand(1) == GV) { 852 SI->eraseFromParent(); 853 continue; 854 } 855 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 856 // Insert the load in the corresponding predecessor, not right before the 857 // PHI. 858 unsigned PredNo = Alloc->use_begin().getOperandNo()/2; 859 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator(); 860 } 861 862 // Insert a load from the global, and use it instead of the malloc. 863 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 864 U->replaceUsesOfWith(Alloc, NL); 865 } 866} 867 868/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 869/// GV are simple enough to perform HeapSRA, return true. 870static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 871 MallocInst *MI) { 872 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 873 ++UI) 874 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 875 // We permit two users of the load: setcc comparing against the null 876 // pointer, and a getelementptr of a specific form. 877 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E; 878 ++UI) { 879 // Comparison against null is ok. 880 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) { 881 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 882 return false; 883 continue; 884 } 885 886 // getelementptr is also ok, but only a simple form. 887 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 888 // Must index into the array and into the struct. 889 if (GEPI->getNumOperands() < 3) 890 return false; 891 892 // Otherwise the GEP is ok. 893 continue; 894 } 895 896 if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 897 // We have a phi of a load from the global. We can only handle this 898 // if the other PHI'd values are actually the same. In this case, 899 // the rewriter will just drop the phi entirely. 900 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 901 Value *IV = PN->getIncomingValue(i); 902 if (IV == LI) continue; // Trivial the same. 903 904 // If the phi'd value is from the malloc that initializes the value, 905 // we can xform it. 906 if (IV == MI) continue; 907 908 // Otherwise, we don't know what it is. 909 return false; 910 } 911 return true; 912 } 913 914 // Otherwise we don't know what this is, not ok. 915 return false; 916 } 917 } 918 return true; 919} 920 921/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd 922/// value, lazily creating it on demand. 923static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo, 924 const std::vector<GlobalVariable*> &FieldGlobals, 925 std::vector<Value *> &InsertedLoadsForPtr) { 926 if (InsertedLoadsForPtr.size() <= FieldNo) 927 InsertedLoadsForPtr.resize(FieldNo+1); 928 if (InsertedLoadsForPtr[FieldNo] == 0) 929 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo], 930 Load->getName()+".f" + 931 utostr(FieldNo), Load); 932 return InsertedLoadsForPtr[FieldNo]; 933} 934 935/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 936/// the load, rewrite the derived value to use the HeapSRoA'd load. 937static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser, 938 const std::vector<GlobalVariable*> &FieldGlobals, 939 std::vector<Value *> &InsertedLoadsForPtr) { 940 // If this is a comparison against null, handle it. 941 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 942 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 943 // If we have a setcc of the loaded pointer, we can use a setcc of any 944 // field. 945 Value *NPtr; 946 if (InsertedLoadsForPtr.empty()) { 947 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr); 948 } else { 949 NPtr = InsertedLoadsForPtr.back(); 950 } 951 952 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 953 Constant::getNullValue(NPtr->getType()), 954 SCI->getName(), SCI); 955 SCI->replaceAllUsesWith(New); 956 SCI->eraseFromParent(); 957 return; 958 } 959 960 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...' 961 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 962 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 963 && "Unexpected GEPI!"); 964 965 // Load the pointer for this field. 966 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 967 Value *NewPtr = GetHeapSROALoad(Load, FieldNo, 968 FieldGlobals, InsertedLoadsForPtr); 969 970 // Create the new GEP idx vector. 971 SmallVector<Value*, 8> GEPIdx; 972 GEPIdx.push_back(GEPI->getOperand(1)); 973 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 974 975 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(), 976 GEPI->getName(), GEPI); 977 GEPI->replaceAllUsesWith(NGEPI); 978 GEPI->eraseFromParent(); 979 return; 980 } 981 982 // Handle PHI nodes. PHI nodes must be merging in the same values, plus 983 // potentially the original malloc. Insert phi nodes for each field, then 984 // process uses of the PHI. 985 PHINode *PN = cast<PHINode>(LoadUser); 986 std::vector<Value *> PHIsForField; 987 PHIsForField.resize(FieldGlobals.size()); 988 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 989 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr); 990 991 PHINode *FieldPN = new PHINode(LoadV->getType(), 992 PN->getName()+"."+utostr(i), PN); 993 // Fill in the predecessor values. 994 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) { 995 // Each predecessor either uses the load or the original malloc. 996 Value *InVal = PN->getIncomingValue(pred); 997 BasicBlock *BB = PN->getIncomingBlock(pred); 998 Value *NewVal; 999 if (isa<MallocInst>(InVal)) { 1000 // Insert a reload from the global in the predecessor. 1001 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals, 1002 PHIsForField); 1003 } else { 1004 NewVal = InsertedLoadsForPtr[i]; 1005 } 1006 FieldPN->addIncoming(NewVal, BB); 1007 } 1008 PHIsForField[i] = FieldPN; 1009 } 1010 1011 // Since PHIsForField specifies a phi for every input value, the lazy inserter 1012 // will never insert a load. 1013 while (!PN->use_empty()) 1014 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField); 1015 PN->eraseFromParent(); 1016} 1017 1018/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1019/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1020/// use FieldGlobals instead. All uses of loaded values satisfy 1021/// GlobalLoadUsesSimpleEnoughForHeapSRA. 1022static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1023 const std::vector<GlobalVariable*> &FieldGlobals) { 1024 std::vector<Value *> InsertedLoadsForPtr; 1025 //InsertedLoadsForPtr.resize(FieldGlobals.size()); 1026 while (!Load->use_empty()) 1027 RewriteHeapSROALoadUser(Load, Load->use_back(), 1028 FieldGlobals, InsertedLoadsForPtr); 1029} 1030 1031/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1032/// it up into multiple allocations of arrays of the fields. 1033static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1034 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1035 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1036 1037 // There is guaranteed to be at least one use of the malloc (storing 1038 // it into GV). If there are other uses, change them to be uses of 1039 // the global to simplify later code. This also deletes the store 1040 // into GV. 1041 ReplaceUsesOfMallocWithGlobal(MI, GV); 1042 1043 // Okay, at this point, there are no users of the malloc. Insert N 1044 // new mallocs at the same place as MI, and N globals. 1045 std::vector<GlobalVariable*> FieldGlobals; 1046 std::vector<MallocInst*> FieldMallocs; 1047 1048 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1049 const Type *FieldTy = STy->getElementType(FieldNo); 1050 const Type *PFieldTy = PointerType::get(FieldTy); 1051 1052 GlobalVariable *NGV = 1053 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1054 Constant::getNullValue(PFieldTy), 1055 GV->getName() + ".f" + utostr(FieldNo), GV, 1056 GV->isThreadLocal()); 1057 FieldGlobals.push_back(NGV); 1058 1059 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1060 MI->getName() + ".f" + utostr(FieldNo),MI); 1061 FieldMallocs.push_back(NMI); 1062 new StoreInst(NMI, NGV, MI); 1063 } 1064 1065 // The tricky aspect of this transformation is handling the case when malloc 1066 // fails. In the original code, malloc failing would set the result pointer 1067 // of malloc to null. In this case, some mallocs could succeed and others 1068 // could fail. As such, we emit code that looks like this: 1069 // F0 = malloc(field0) 1070 // F1 = malloc(field1) 1071 // F2 = malloc(field2) 1072 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1073 // if (F0) { free(F0); F0 = 0; } 1074 // if (F1) { free(F1); F1 = 0; } 1075 // if (F2) { free(F2); F2 = 0; } 1076 // } 1077 Value *RunningOr = 0; 1078 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1079 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1080 Constant::getNullValue(FieldMallocs[i]->getType()), 1081 "isnull", MI); 1082 if (!RunningOr) 1083 RunningOr = Cond; // First seteq 1084 else 1085 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI); 1086 } 1087 1088 // Split the basic block at the old malloc. 1089 BasicBlock *OrigBB = MI->getParent(); 1090 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1091 1092 // Create the block to check the first condition. Put all these blocks at the 1093 // end of the function as they are unlikely to be executed. 1094 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null", 1095 OrigBB->getParent()); 1096 1097 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1098 // branch on RunningOr. 1099 OrigBB->getTerminator()->eraseFromParent(); 1100 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB); 1101 1102 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1103 // pointer, because some may be null while others are not. 1104 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1105 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1106 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1107 Constant::getNullValue(GVVal->getType()), 1108 "tmp", NullPtrBlock); 1109 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent()); 1110 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent()); 1111 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1112 1113 // Fill in FreeBlock. 1114 new FreeInst(GVVal, FreeBlock); 1115 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1116 FreeBlock); 1117 new BranchInst(NextBlock, FreeBlock); 1118 1119 NullPtrBlock = NextBlock; 1120 } 1121 1122 new BranchInst(ContBB, NullPtrBlock); 1123 1124 1125 // MI is no longer needed, remove it. 1126 MI->eraseFromParent(); 1127 1128 1129 // Okay, the malloc site is completely handled. All of the uses of GV are now 1130 // loads, and all uses of those loads are simple. Rewrite them to use loads 1131 // of the per-field globals instead. 1132 while (!GV->use_empty()) { 1133 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 1134 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals); 1135 LI->eraseFromParent(); 1136 } else { 1137 // Must be a store of null. 1138 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1139 assert(isa<Constant>(SI->getOperand(0)) && 1140 cast<Constant>(SI->getOperand(0))->isNullValue() && 1141 "Unexpected heap-sra user!"); 1142 1143 // Insert a store of null into each global. 1144 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1145 Constant *Null = 1146 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType()); 1147 new StoreInst(Null, FieldGlobals[i], SI); 1148 } 1149 // Erase the original store. 1150 SI->eraseFromParent(); 1151 } 1152 } 1153 1154 // The old global is now dead, remove it. 1155 GV->eraseFromParent(); 1156 1157 ++NumHeapSRA; 1158 return FieldGlobals[0]; 1159} 1160 1161 1162// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1163// that only one value (besides its initializer) is ever stored to the global. 1164static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1165 Module::global_iterator &GVI, 1166 TargetData &TD) { 1167 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal)) 1168 StoredOnceVal = CI->getOperand(0); 1169 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){ 1170 // "getelementptr Ptr, 0, 0, 0" is really just a cast. 1171 bool IsJustACast = true; 1172 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 1173 if (!isa<Constant>(GEPI->getOperand(i)) || 1174 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) { 1175 IsJustACast = false; 1176 break; 1177 } 1178 if (IsJustACast) 1179 StoredOnceVal = GEPI->getOperand(0); 1180 } 1181 1182 // If we are dealing with a pointer global that is initialized to null and 1183 // only has one (non-null) value stored into it, then we can optimize any 1184 // users of the loaded value (often calls and loads) that would trap if the 1185 // value was null. 1186 if (isa<PointerType>(GV->getInitializer()->getType()) && 1187 GV->getInitializer()->isNullValue()) { 1188 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1189 if (GV->getInitializer()->getType() != SOVC->getType()) 1190 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1191 1192 // Optimize away any trapping uses of the loaded value. 1193 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1194 return true; 1195 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1196 // If this is a malloc of an abstract type, don't touch it. 1197 if (!MI->getAllocatedType()->isSized()) 1198 return false; 1199 1200 // We can't optimize this global unless all uses of it are *known* to be 1201 // of the malloc value, not of the null initializer value (consider a use 1202 // that compares the global's value against zero to see if the malloc has 1203 // been reached). To do this, we check to see if all uses of the global 1204 // would trap if the global were null: this proves that they must all 1205 // happen after the malloc. 1206 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1207 return false; 1208 1209 // We can't optimize this if the malloc itself is used in a complex way, 1210 // for example, being stored into multiple globals. This allows the 1211 // malloc to be stored into the specified global, loaded setcc'd, and 1212 // GEP'd. These are all things we could transform to using the global 1213 // for. 1214 { 1215 SmallPtrSet<PHINode*, 8> PHIs; 1216 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1217 return false; 1218 } 1219 1220 1221 // If we have a global that is only initialized with a fixed size malloc, 1222 // transform the program to use global memory instead of malloc'd memory. 1223 // This eliminates dynamic allocation, avoids an indirection accessing the 1224 // data, and exposes the resultant global to further GlobalOpt. 1225 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1226 // Restrict this transformation to only working on small allocations 1227 // (2048 bytes currently), as we don't want to introduce a 16M global or 1228 // something. 1229 if (NElements->getZExtValue()* 1230 TD.getTypeSize(MI->getAllocatedType()) < 2048) { 1231 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1232 return true; 1233 } 1234 } 1235 1236 // If the allocation is an array of structures, consider transforming this 1237 // into multiple malloc'd arrays, one for each field. This is basically 1238 // SRoA for malloc'd memory. 1239 if (const StructType *AllocTy = 1240 dyn_cast<StructType>(MI->getAllocatedType())) { 1241 // This the structure has an unreasonable number of fields, leave it 1242 // alone. 1243 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 && 1244 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1245 GVI = PerformHeapAllocSRoA(GV, MI); 1246 return true; 1247 } 1248 } 1249 } 1250 } 1251 1252 return false; 1253} 1254 1255/// ShrinkGlobalToBoolean - At this point, we have learned that the only two 1256/// values ever stored into GV are its initializer and OtherVal. 1257static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1258 // Create the new global, initializing it to false. 1259 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1260 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1261 GV->getName()+".b", 1262 (Module *)NULL, 1263 GV->isThreadLocal()); 1264 GV->getParent()->getGlobalList().insert(GV, NewGV); 1265 1266 Constant *InitVal = GV->getInitializer(); 1267 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1268 1269 // If initialized to zero and storing one into the global, we can use a cast 1270 // instead of a select to synthesize the desired value. 1271 bool IsOneZero = false; 1272 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1273 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1274 1275 while (!GV->use_empty()) { 1276 Instruction *UI = cast<Instruction>(GV->use_back()); 1277 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1278 // Change the store into a boolean store. 1279 bool StoringOther = SI->getOperand(0) == OtherVal; 1280 // Only do this if we weren't storing a loaded value. 1281 Value *StoreVal; 1282 if (StoringOther || SI->getOperand(0) == InitVal) 1283 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1284 else { 1285 // Otherwise, we are storing a previously loaded copy. To do this, 1286 // change the copy from copying the original value to just copying the 1287 // bool. 1288 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1289 1290 // If we're already replaced the input, StoredVal will be a cast or 1291 // select instruction. If not, it will be a load of the original 1292 // global. 1293 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1294 assert(LI->getOperand(0) == GV && "Not a copy!"); 1295 // Insert a new load, to preserve the saved value. 1296 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1297 } else { 1298 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1299 "This is not a form that we understand!"); 1300 StoreVal = StoredVal->getOperand(0); 1301 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1302 } 1303 } 1304 new StoreInst(StoreVal, NewGV, SI); 1305 } else if (!UI->use_empty()) { 1306 // Change the load into a load of bool then a select. 1307 LoadInst *LI = cast<LoadInst>(UI); 1308 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1309 Value *NSI; 1310 if (IsOneZero) 1311 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1312 else 1313 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI); 1314 NSI->takeName(LI); 1315 LI->replaceAllUsesWith(NSI); 1316 } 1317 UI->eraseFromParent(); 1318 } 1319 1320 GV->eraseFromParent(); 1321} 1322 1323 1324/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1325/// it if possible. If we make a change, return true. 1326bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1327 Module::global_iterator &GVI) { 1328 std::set<PHINode*> PHIUsers; 1329 GlobalStatus GS; 1330 GV->removeDeadConstantUsers(); 1331 1332 if (GV->use_empty()) { 1333 DOUT << "GLOBAL DEAD: " << *GV; 1334 GV->eraseFromParent(); 1335 ++NumDeleted; 1336 return true; 1337 } 1338 1339 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1340#if 0 1341 cerr << "Global: " << *GV; 1342 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1343 cerr << " StoredType = "; 1344 switch (GS.StoredType) { 1345 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1346 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1347 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1348 case GlobalStatus::isStored: cerr << "stored\n"; break; 1349 } 1350 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1351 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1352 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1353 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1354 << "\n"; 1355 cerr << " HasMultipleAccessingFunctions = " 1356 << GS.HasMultipleAccessingFunctions << "\n"; 1357 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1358 cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n"; 1359 cerr << "\n"; 1360#endif 1361 1362 // If this is a first class global and has only one accessing function 1363 // and this function is main (which we know is not recursive we can make 1364 // this global a local variable) we replace the global with a local alloca 1365 // in this function. 1366 // 1367 // NOTE: It doesn't make sense to promote non first class types since we 1368 // are just replacing static memory to stack memory. 1369 if (!GS.HasMultipleAccessingFunctions && 1370 GS.AccessingFunction && !GS.HasNonInstructionUser && 1371 GV->getType()->getElementType()->isFirstClassType() && 1372 GS.AccessingFunction->getName() == "main" && 1373 GS.AccessingFunction->hasExternalLinkage()) { 1374 DOUT << "LOCALIZING GLOBAL: " << *GV; 1375 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1376 const Type* ElemTy = GV->getType()->getElementType(); 1377 // FIXME: Pass Global's alignment when globals have alignment 1378 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1379 if (!isa<UndefValue>(GV->getInitializer())) 1380 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1381 1382 GV->replaceAllUsesWith(Alloca); 1383 GV->eraseFromParent(); 1384 ++NumLocalized; 1385 return true; 1386 } 1387 1388 // If the global is never loaded (but may be stored to), it is dead. 1389 // Delete it now. 1390 if (!GS.isLoaded) { 1391 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1392 1393 // Delete any stores we can find to the global. We may not be able to 1394 // make it completely dead though. 1395 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1396 1397 // If the global is dead now, delete it. 1398 if (GV->use_empty()) { 1399 GV->eraseFromParent(); 1400 ++NumDeleted; 1401 Changed = true; 1402 } 1403 return Changed; 1404 1405 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1406 DOUT << "MARKING CONSTANT: " << *GV; 1407 GV->setConstant(true); 1408 1409 // Clean up any obviously simplifiable users now. 1410 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1411 1412 // If the global is dead now, just nuke it. 1413 if (GV->use_empty()) { 1414 DOUT << " *** Marking constant allowed us to simplify " 1415 << "all users and delete global!\n"; 1416 GV->eraseFromParent(); 1417 ++NumDeleted; 1418 } 1419 1420 ++NumMarked; 1421 return true; 1422 } else if (!GS.isNotSuitableForSRA && 1423 !GV->getInitializer()->getType()->isFirstClassType()) { 1424 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) { 1425 GVI = FirstNewGV; // Don't skip the newly produced globals! 1426 return true; 1427 } 1428 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1429 // If the initial value for the global was an undef value, and if only 1430 // one other value was stored into it, we can just change the 1431 // initializer to be an undef value, then delete all stores to the 1432 // global. This allows us to mark it constant. 1433 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1434 if (isa<UndefValue>(GV->getInitializer())) { 1435 // Change the initial value here. 1436 GV->setInitializer(SOVConstant); 1437 1438 // Clean up any obviously simplifiable users now. 1439 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1440 1441 if (GV->use_empty()) { 1442 DOUT << " *** Substituting initializer allowed us to " 1443 << "simplify all users and delete global!\n"; 1444 GV->eraseFromParent(); 1445 ++NumDeleted; 1446 } else { 1447 GVI = GV; 1448 } 1449 ++NumSubstitute; 1450 return true; 1451 } 1452 1453 // Try to optimize globals based on the knowledge that only one value 1454 // (besides its initializer) is ever stored to the global. 1455 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1456 getAnalysis<TargetData>())) 1457 return true; 1458 1459 // Otherwise, if the global was not a boolean, we can shrink it to be a 1460 // boolean. 1461 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1462 if (GV->getType()->getElementType() != Type::Int1Ty && 1463 !GV->getType()->getElementType()->isFloatingPoint() && 1464 !isa<VectorType>(GV->getType()->getElementType()) && 1465 !GS.HasPHIUser && !GS.isNotSuitableForSRA) { 1466 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1467 ShrinkGlobalToBoolean(GV, SOVConstant); 1468 ++NumShrunkToBool; 1469 return true; 1470 } 1471 } 1472 } 1473 return false; 1474} 1475 1476/// OnlyCalledDirectly - Return true if the specified function is only called 1477/// directly. In other words, its address is never taken. 1478static bool OnlyCalledDirectly(Function *F) { 1479 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1480 Instruction *User = dyn_cast<Instruction>(*UI); 1481 if (!User) return false; 1482 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1483 1484 // See if the function address is passed as an argument. 1485 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i) 1486 if (User->getOperand(i) == F) return false; 1487 } 1488 return true; 1489} 1490 1491/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1492/// function, changing them to FastCC. 1493static void ChangeCalleesToFastCall(Function *F) { 1494 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1495 Instruction *User = cast<Instruction>(*UI); 1496 if (CallInst *CI = dyn_cast<CallInst>(User)) 1497 CI->setCallingConv(CallingConv::Fast); 1498 else 1499 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast); 1500 } 1501} 1502 1503bool GlobalOpt::OptimizeFunctions(Module &M) { 1504 bool Changed = false; 1505 // Optimize functions. 1506 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1507 Function *F = FI++; 1508 F->removeDeadConstantUsers(); 1509 if (F->use_empty() && (F->hasInternalLinkage() || 1510 F->hasLinkOnceLinkage())) { 1511 M.getFunctionList().erase(F); 1512 Changed = true; 1513 ++NumFnDeleted; 1514 } else if (F->hasInternalLinkage() && 1515 F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1516 OnlyCalledDirectly(F)) { 1517 // If this function has C calling conventions, is not a varargs 1518 // function, and is only called directly, promote it to use the Fast 1519 // calling convention. 1520 F->setCallingConv(CallingConv::Fast); 1521 ChangeCalleesToFastCall(F); 1522 ++NumFastCallFns; 1523 Changed = true; 1524 } 1525 } 1526 return Changed; 1527} 1528 1529bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1530 bool Changed = false; 1531 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1532 GVI != E; ) { 1533 GlobalVariable *GV = GVI++; 1534 if (!GV->isConstant() && GV->hasInternalLinkage() && 1535 GV->hasInitializer()) 1536 Changed |= ProcessInternalGlobal(GV, GVI); 1537 } 1538 return Changed; 1539} 1540 1541/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1542/// initializers have an init priority of 65535. 1543GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1544 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1545 I != E; ++I) 1546 if (I->getName() == "llvm.global_ctors") { 1547 // Found it, verify it's an array of { int, void()* }. 1548 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1549 if (!ATy) return 0; 1550 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1551 if (!STy || STy->getNumElements() != 2 || 1552 STy->getElementType(0) != Type::Int32Ty) return 0; 1553 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1554 if (!PFTy) return 0; 1555 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1556 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1557 FTy->getNumParams() != 0) 1558 return 0; 1559 1560 // Verify that the initializer is simple enough for us to handle. 1561 if (!I->hasInitializer()) return 0; 1562 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1563 if (!CA) return 0; 1564 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1565 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) { 1566 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1567 continue; 1568 1569 // Must have a function or null ptr. 1570 if (!isa<Function>(CS->getOperand(1))) 1571 return 0; 1572 1573 // Init priority must be standard. 1574 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1575 if (!CI || CI->getZExtValue() != 65535) 1576 return 0; 1577 } else { 1578 return 0; 1579 } 1580 1581 return I; 1582 } 1583 return 0; 1584} 1585 1586/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1587/// return a list of the functions and null terminator as a vector. 1588static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1589 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1590 std::vector<Function*> Result; 1591 Result.reserve(CA->getNumOperands()); 1592 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) { 1593 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i)); 1594 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1595 } 1596 return Result; 1597} 1598 1599/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1600/// specified array, returning the new global to use. 1601static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1602 const std::vector<Function*> &Ctors) { 1603 // If we made a change, reassemble the initializer list. 1604 std::vector<Constant*> CSVals; 1605 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1606 CSVals.push_back(0); 1607 1608 // Create the new init list. 1609 std::vector<Constant*> CAList; 1610 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1611 if (Ctors[i]) { 1612 CSVals[1] = Ctors[i]; 1613 } else { 1614 const Type *FTy = FunctionType::get(Type::VoidTy, 1615 std::vector<const Type*>(), false); 1616 const PointerType *PFTy = PointerType::get(FTy); 1617 CSVals[1] = Constant::getNullValue(PFTy); 1618 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1619 } 1620 CAList.push_back(ConstantStruct::get(CSVals)); 1621 } 1622 1623 // Create the array initializer. 1624 const Type *StructTy = 1625 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1626 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1627 CAList); 1628 1629 // If we didn't change the number of elements, don't create a new GV. 1630 if (CA->getType() == GCL->getInitializer()->getType()) { 1631 GCL->setInitializer(CA); 1632 return GCL; 1633 } 1634 1635 // Create the new global and insert it next to the existing list. 1636 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1637 GCL->getLinkage(), CA, "", 1638 (Module *)NULL, 1639 GCL->isThreadLocal()); 1640 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1641 NGV->takeName(GCL); 1642 1643 // Nuke the old list, replacing any uses with the new one. 1644 if (!GCL->use_empty()) { 1645 Constant *V = NGV; 1646 if (V->getType() != GCL->getType()) 1647 V = ConstantExpr::getBitCast(V, GCL->getType()); 1648 GCL->replaceAllUsesWith(V); 1649 } 1650 GCL->eraseFromParent(); 1651 1652 if (Ctors.size()) 1653 return NGV; 1654 else 1655 return 0; 1656} 1657 1658 1659static Constant *getVal(std::map<Value*, Constant*> &ComputedValues, 1660 Value *V) { 1661 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1662 Constant *R = ComputedValues[V]; 1663 assert(R && "Reference to an uncomputed value!"); 1664 return R; 1665} 1666 1667/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1668/// enough for us to understand. In particular, if it is a cast of something, 1669/// we punt. We basically just support direct accesses to globals and GEP's of 1670/// globals. This should be kept up to date with CommitValueTo. 1671static bool isSimpleEnoughPointerToCommit(Constant *C) { 1672 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1673 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1674 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1675 return !GV->isDeclaration(); // reject external globals. 1676 } 1677 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 1678 // Handle a constantexpr gep. 1679 if (CE->getOpcode() == Instruction::GetElementPtr && 1680 isa<GlobalVariable>(CE->getOperand(0))) { 1681 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1682 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1683 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1684 return GV->hasInitializer() && 1685 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1686 } 1687 return false; 1688} 1689 1690/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 1691/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 1692/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 1693static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 1694 ConstantExpr *Addr, unsigned OpNo) { 1695 // Base case of the recursion. 1696 if (OpNo == Addr->getNumOperands()) { 1697 assert(Val->getType() == Init->getType() && "Type mismatch!"); 1698 return Val; 1699 } 1700 1701 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1702 std::vector<Constant*> Elts; 1703 1704 // Break up the constant into its elements. 1705 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1706 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) 1707 Elts.push_back(CS->getOperand(i)); 1708 } else if (isa<ConstantAggregateZero>(Init)) { 1709 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1710 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 1711 } else if (isa<UndefValue>(Init)) { 1712 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1713 Elts.push_back(UndefValue::get(STy->getElementType(i))); 1714 } else { 1715 assert(0 && "This code is out of sync with " 1716 " ConstantFoldLoadThroughGEPConstantExpr"); 1717 } 1718 1719 // Replace the element that we are supposed to. 1720 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 1721 unsigned Idx = CU->getZExtValue(); 1722 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 1723 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 1724 1725 // Return the modified struct. 1726 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 1727 } else { 1728 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 1729 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 1730 1731 // Break up the array into elements. 1732 std::vector<Constant*> Elts; 1733 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1734 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1735 Elts.push_back(CA->getOperand(i)); 1736 } else if (isa<ConstantAggregateZero>(Init)) { 1737 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 1738 Elts.assign(ATy->getNumElements(), Elt); 1739 } else if (isa<UndefValue>(Init)) { 1740 Constant *Elt = UndefValue::get(ATy->getElementType()); 1741 Elts.assign(ATy->getNumElements(), Elt); 1742 } else { 1743 assert(0 && "This code is out of sync with " 1744 " ConstantFoldLoadThroughGEPConstantExpr"); 1745 } 1746 1747 assert(CI->getZExtValue() < ATy->getNumElements()); 1748 Elts[CI->getZExtValue()] = 1749 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 1750 return ConstantArray::get(ATy, Elts); 1751 } 1752} 1753 1754/// CommitValueTo - We have decided that Addr (which satisfies the predicate 1755/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 1756static void CommitValueTo(Constant *Val, Constant *Addr) { 1757 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 1758 assert(GV->hasInitializer()); 1759 GV->setInitializer(Val); 1760 return; 1761 } 1762 1763 ConstantExpr *CE = cast<ConstantExpr>(Addr); 1764 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1765 1766 Constant *Init = GV->getInitializer(); 1767 Init = EvaluateStoreInto(Init, Val, CE, 2); 1768 GV->setInitializer(Init); 1769} 1770 1771/// ComputeLoadResult - Return the value that would be computed by a load from 1772/// P after the stores reflected by 'memory' have been performed. If we can't 1773/// decide, return null. 1774static Constant *ComputeLoadResult(Constant *P, 1775 const std::map<Constant*, Constant*> &Memory) { 1776 // If this memory location has been recently stored, use the stored value: it 1777 // is the most up-to-date. 1778 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P); 1779 if (I != Memory.end()) return I->second; 1780 1781 // Access it. 1782 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 1783 if (GV->hasInitializer()) 1784 return GV->getInitializer(); 1785 return 0; 1786 } 1787 1788 // Handle a constantexpr getelementptr. 1789 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 1790 if (CE->getOpcode() == Instruction::GetElementPtr && 1791 isa<GlobalVariable>(CE->getOperand(0))) { 1792 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1793 if (GV->hasInitializer()) 1794 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1795 } 1796 1797 return 0; // don't know how to evaluate. 1798} 1799 1800/// EvaluateFunction - Evaluate a call to function F, returning true if 1801/// successful, false if we can't evaluate it. ActualArgs contains the formal 1802/// arguments for the function. 1803static bool EvaluateFunction(Function *F, Constant *&RetVal, 1804 const std::vector<Constant*> &ActualArgs, 1805 std::vector<Function*> &CallStack, 1806 std::map<Constant*, Constant*> &MutatedMemory, 1807 std::vector<GlobalVariable*> &AllocaTmps) { 1808 // Check to see if this function is already executing (recursion). If so, 1809 // bail out. TODO: we might want to accept limited recursion. 1810 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 1811 return false; 1812 1813 CallStack.push_back(F); 1814 1815 /// Values - As we compute SSA register values, we store their contents here. 1816 std::map<Value*, Constant*> Values; 1817 1818 // Initialize arguments to the incoming values specified. 1819 unsigned ArgNo = 0; 1820 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1821 ++AI, ++ArgNo) 1822 Values[AI] = ActualArgs[ArgNo]; 1823 1824 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 1825 /// we can only evaluate any one basic block at most once. This set keeps 1826 /// track of what we have executed so we can detect recursive cases etc. 1827 std::set<BasicBlock*> ExecutedBlocks; 1828 1829 // CurInst - The current instruction we're evaluating. 1830 BasicBlock::iterator CurInst = F->begin()->begin(); 1831 1832 // This is the main evaluation loop. 1833 while (1) { 1834 Constant *InstResult = 0; 1835 1836 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 1837 if (SI->isVolatile()) return false; // no volatile accesses. 1838 Constant *Ptr = getVal(Values, SI->getOperand(1)); 1839 if (!isSimpleEnoughPointerToCommit(Ptr)) 1840 // If this is too complex for us to commit, reject it. 1841 return false; 1842 Constant *Val = getVal(Values, SI->getOperand(0)); 1843 MutatedMemory[Ptr] = Val; 1844 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 1845 InstResult = ConstantExpr::get(BO->getOpcode(), 1846 getVal(Values, BO->getOperand(0)), 1847 getVal(Values, BO->getOperand(1))); 1848 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 1849 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 1850 getVal(Values, CI->getOperand(0)), 1851 getVal(Values, CI->getOperand(1))); 1852 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 1853 InstResult = ConstantExpr::getCast(CI->getOpcode(), 1854 getVal(Values, CI->getOperand(0)), 1855 CI->getType()); 1856 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 1857 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 1858 getVal(Values, SI->getOperand(1)), 1859 getVal(Values, SI->getOperand(2))); 1860 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 1861 Constant *P = getVal(Values, GEP->getOperand(0)); 1862 SmallVector<Constant*, 8> GEPOps; 1863 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i) 1864 GEPOps.push_back(getVal(Values, GEP->getOperand(i))); 1865 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 1866 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 1867 if (LI->isVolatile()) return false; // no volatile accesses. 1868 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 1869 MutatedMemory); 1870 if (InstResult == 0) return false; // Could not evaluate load. 1871 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 1872 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 1873 const Type *Ty = AI->getType()->getElementType(); 1874 AllocaTmps.push_back(new GlobalVariable(Ty, false, 1875 GlobalValue::InternalLinkage, 1876 UndefValue::get(Ty), 1877 AI->getName())); 1878 InstResult = AllocaTmps.back(); 1879 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 1880 // Cannot handle inline asm. 1881 if (isa<InlineAsm>(CI->getOperand(0))) return false; 1882 1883 // Resolve function pointers. 1884 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 1885 if (!Callee) return false; // Cannot resolve. 1886 1887 std::vector<Constant*> Formals; 1888 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i) 1889 Formals.push_back(getVal(Values, CI->getOperand(i))); 1890 1891 if (Callee->isDeclaration()) { 1892 // If this is a function we can constant fold, do it. 1893 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 1894 Formals.size())) { 1895 InstResult = C; 1896 } else { 1897 return false; 1898 } 1899 } else { 1900 if (Callee->getFunctionType()->isVarArg()) 1901 return false; 1902 1903 Constant *RetVal; 1904 1905 // Execute the call, if successful, use the return value. 1906 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 1907 MutatedMemory, AllocaTmps)) 1908 return false; 1909 InstResult = RetVal; 1910 } 1911 } else if (isa<TerminatorInst>(CurInst)) { 1912 BasicBlock *NewBB = 0; 1913 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 1914 if (BI->isUnconditional()) { 1915 NewBB = BI->getSuccessor(0); 1916 } else { 1917 ConstantInt *Cond = 1918 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 1919 if (!Cond) return false; // Cannot determine. 1920 1921 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 1922 } 1923 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 1924 ConstantInt *Val = 1925 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 1926 if (!Val) return false; // Cannot determine. 1927 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 1928 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 1929 if (RI->getNumOperands()) 1930 RetVal = getVal(Values, RI->getOperand(0)); 1931 1932 CallStack.pop_back(); // return from fn. 1933 return true; // We succeeded at evaluating this ctor! 1934 } else { 1935 // invoke, unwind, unreachable. 1936 return false; // Cannot handle this terminator. 1937 } 1938 1939 // Okay, we succeeded in evaluating this control flow. See if we have 1940 // executed the new block before. If so, we have a looping function, 1941 // which we cannot evaluate in reasonable time. 1942 if (!ExecutedBlocks.insert(NewBB).second) 1943 return false; // looped! 1944 1945 // Okay, we have never been in this block before. Check to see if there 1946 // are any PHI nodes. If so, evaluate them with information about where 1947 // we came from. 1948 BasicBlock *OldBB = CurInst->getParent(); 1949 CurInst = NewBB->begin(); 1950 PHINode *PN; 1951 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 1952 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 1953 1954 // Do NOT increment CurInst. We know that the terminator had no value. 1955 continue; 1956 } else { 1957 // Did not know how to evaluate this! 1958 return false; 1959 } 1960 1961 if (!CurInst->use_empty()) 1962 Values[CurInst] = InstResult; 1963 1964 // Advance program counter. 1965 ++CurInst; 1966 } 1967} 1968 1969/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 1970/// we can. Return true if we can, false otherwise. 1971static bool EvaluateStaticConstructor(Function *F) { 1972 /// MutatedMemory - For each store we execute, we update this map. Loads 1973 /// check this to get the most up-to-date value. If evaluation is successful, 1974 /// this state is committed to the process. 1975 std::map<Constant*, Constant*> MutatedMemory; 1976 1977 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 1978 /// to represent its body. This vector is needed so we can delete the 1979 /// temporary globals when we are done. 1980 std::vector<GlobalVariable*> AllocaTmps; 1981 1982 /// CallStack - This is used to detect recursion. In pathological situations 1983 /// we could hit exponential behavior, but at least there is nothing 1984 /// unbounded. 1985 std::vector<Function*> CallStack; 1986 1987 // Call the function. 1988 Constant *RetValDummy; 1989 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 1990 CallStack, MutatedMemory, AllocaTmps); 1991 if (EvalSuccess) { 1992 // We succeeded at evaluation: commit the result. 1993 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 1994 << F->getName() << "' to " << MutatedMemory.size() 1995 << " stores.\n"; 1996 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 1997 E = MutatedMemory.end(); I != E; ++I) 1998 CommitValueTo(I->second, I->first); 1999 } 2000 2001 // At this point, we are done interpreting. If we created any 'alloca' 2002 // temporaries, release them now. 2003 while (!AllocaTmps.empty()) { 2004 GlobalVariable *Tmp = AllocaTmps.back(); 2005 AllocaTmps.pop_back(); 2006 2007 // If there are still users of the alloca, the program is doing something 2008 // silly, e.g. storing the address of the alloca somewhere and using it 2009 // later. Since this is undefined, we'll just make it be null. 2010 if (!Tmp->use_empty()) 2011 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2012 delete Tmp; 2013 } 2014 2015 return EvalSuccess; 2016} 2017 2018 2019 2020/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2021/// Return true if anything changed. 2022bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2023 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2024 bool MadeChange = false; 2025 if (Ctors.empty()) return false; 2026 2027 // Loop over global ctors, optimizing them when we can. 2028 for (unsigned i = 0; i != Ctors.size(); ++i) { 2029 Function *F = Ctors[i]; 2030 // Found a null terminator in the middle of the list, prune off the rest of 2031 // the list. 2032 if (F == 0) { 2033 if (i != Ctors.size()-1) { 2034 Ctors.resize(i+1); 2035 MadeChange = true; 2036 } 2037 break; 2038 } 2039 2040 // We cannot simplify external ctor functions. 2041 if (F->empty()) continue; 2042 2043 // If we can evaluate the ctor at compile time, do. 2044 if (EvaluateStaticConstructor(F)) { 2045 Ctors.erase(Ctors.begin()+i); 2046 MadeChange = true; 2047 --i; 2048 ++NumCtorsEvaluated; 2049 continue; 2050 } 2051 } 2052 2053 if (!MadeChange) return false; 2054 2055 GCL = InstallGlobalCtors(GCL, Ctors); 2056 return true; 2057} 2058 2059 2060bool GlobalOpt::runOnModule(Module &M) { 2061 bool Changed = false; 2062 2063 // Try to find the llvm.globalctors list. 2064 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2065 2066 bool LocalChange = true; 2067 while (LocalChange) { 2068 LocalChange = false; 2069 2070 // Delete functions that are trivially dead, ccc -> fastcc 2071 LocalChange |= OptimizeFunctions(M); 2072 2073 // Optimize global_ctors list. 2074 if (GlobalCtors) 2075 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2076 2077 // Optimize non-address-taken globals. 2078 LocalChange |= OptimizeGlobalVars(M); 2079 Changed |= LocalChange; 2080 } 2081 2082 // TODO: Move all global ctors functions to the end of the module for code 2083 // layout. 2084 2085 return Changed; 2086} 2087