GlobalOpt.cpp revision 20eef0f652c5358059d6d0f7f8cf585d971fb28b
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 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 872 ++UI) 873 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 874 // We permit two users of the load: setcc comparing against the null 875 // pointer, and a getelementptr of a specific form. 876 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E; 877 ++UI) { 878 // Comparison against null is ok. 879 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) { 880 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 881 return false; 882 continue; 883 } 884 885 // getelementptr is also ok, but only a simple form. 886 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI); 887 if (!GEPI) return false; 888 889 // Must index into the array and into the struct. 890 if (GEPI->getNumOperands() < 3) 891 return false; 892 893 // Otherwise the GEP is ok. 894 continue; 895 } 896 } 897 return true; 898} 899 900/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd 901/// value, lazily creating it on demand. 902static Value *GetHeapSROALoad(LoadInst *Load, unsigned FieldNo, 903 const std::vector<GlobalVariable*> &FieldGlobals, 904 std::vector<Value *> &InsertedLoadsForPtr) { 905 if (InsertedLoadsForPtr.size() <= FieldNo) 906 InsertedLoadsForPtr.resize(FieldNo+1); 907 if (InsertedLoadsForPtr[FieldNo] == 0) 908 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo], 909 Load->getName()+".f" + 910 utostr(FieldNo), Load); 911 return InsertedLoadsForPtr[FieldNo]; 912} 913 914/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 915/// the load, rewrite the derived value to use the HeapSRoA'd load. 916static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser, 917 const std::vector<GlobalVariable*> &FieldGlobals, 918 std::vector<Value *> &InsertedLoadsForPtr) { 919 // If this is a comparison against null, handle it. 920 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 921 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 922 // If we have a setcc of the loaded pointer, we can use a setcc of any 923 // field. 924 Value *NPtr; 925 if (InsertedLoadsForPtr.empty()) { 926 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr); 927 } else { 928 NPtr = InsertedLoadsForPtr.back(); 929 } 930 931 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 932 Constant::getNullValue(NPtr->getType()), 933 SCI->getName(), SCI); 934 SCI->replaceAllUsesWith(New); 935 SCI->eraseFromParent(); 936 return; 937 } 938 939 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...' 940 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 941 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 942 && "Unexpected GEPI!"); 943 944 // Load the pointer for this field. 945 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 946 Value *NewPtr = GetHeapSROALoad(Load, FieldNo, 947 FieldGlobals, InsertedLoadsForPtr); 948 949 // Create the new GEP idx vector. 950 SmallVector<Value*, 8> GEPIdx; 951 GEPIdx.push_back(GEPI->getOperand(1)); 952 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 953 954 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(), 955 GEPI->getName(), GEPI); 956 GEPI->replaceAllUsesWith(NGEPI); 957 GEPI->eraseFromParent(); 958 return; 959 } 960 961 // Handle PHI nodes. All PHI nodes must be merging in the same values, so 962 // just treat them like a copy. 963 PHINode *PN = cast<PHINode>(LoadUser); 964 while (!PN->use_empty()) 965 RewriteHeapSROALoadUser(Load, PN->use_back(), 966 FieldGlobals, InsertedLoadsForPtr); 967 PN->eraseFromParent(); 968} 969 970/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 971/// is a value loaded from the global. Eliminate all uses of Ptr, making them 972/// use FieldGlobals instead. All uses of loaded values satisfy 973/// GlobalLoadUsesSimpleEnoughForHeapSRA. 974static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 975 const std::vector<GlobalVariable*> &FieldGlobals) { 976 std::vector<Value *> InsertedLoadsForPtr; 977 //InsertedLoadsForPtr.resize(FieldGlobals.size()); 978 while (!Load->use_empty()) 979 RewriteHeapSROALoadUser(Load, Load->use_back(), 980 FieldGlobals, InsertedLoadsForPtr); 981} 982 983/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 984/// it up into multiple allocations of arrays of the fields. 985static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 986 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 987 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 988 989 // There is guaranteed to be at least one use of the malloc (storing 990 // it into GV). If there are other uses, change them to be uses of 991 // the global to simplify later code. This also deletes the store 992 // into GV. 993 ReplaceUsesOfMallocWithGlobal(MI, GV); 994 995 // Okay, at this point, there are no users of the malloc. Insert N 996 // new mallocs at the same place as MI, and N globals. 997 std::vector<GlobalVariable*> FieldGlobals; 998 std::vector<MallocInst*> FieldMallocs; 999 1000 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1001 const Type *FieldTy = STy->getElementType(FieldNo); 1002 const Type *PFieldTy = PointerType::get(FieldTy); 1003 1004 GlobalVariable *NGV = 1005 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1006 Constant::getNullValue(PFieldTy), 1007 GV->getName() + ".f" + utostr(FieldNo), GV, 1008 GV->isThreadLocal()); 1009 FieldGlobals.push_back(NGV); 1010 1011 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1012 MI->getName() + ".f" + utostr(FieldNo),MI); 1013 FieldMallocs.push_back(NMI); 1014 new StoreInst(NMI, NGV, MI); 1015 } 1016 1017 // The tricky aspect of this transformation is handling the case when malloc 1018 // fails. In the original code, malloc failing would set the result pointer 1019 // of malloc to null. In this case, some mallocs could succeed and others 1020 // could fail. As such, we emit code that looks like this: 1021 // F0 = malloc(field0) 1022 // F1 = malloc(field1) 1023 // F2 = malloc(field2) 1024 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1025 // if (F0) { free(F0); F0 = 0; } 1026 // if (F1) { free(F1); F1 = 0; } 1027 // if (F2) { free(F2); F2 = 0; } 1028 // } 1029 Value *RunningOr = 0; 1030 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1031 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1032 Constant::getNullValue(FieldMallocs[i]->getType()), 1033 "isnull", MI); 1034 if (!RunningOr) 1035 RunningOr = Cond; // First seteq 1036 else 1037 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI); 1038 } 1039 1040 // Split the basic block at the old malloc. 1041 BasicBlock *OrigBB = MI->getParent(); 1042 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1043 1044 // Create the block to check the first condition. Put all these blocks at the 1045 // end of the function as they are unlikely to be executed. 1046 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null", 1047 OrigBB->getParent()); 1048 1049 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1050 // branch on RunningOr. 1051 OrigBB->getTerminator()->eraseFromParent(); 1052 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB); 1053 1054 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1055 // pointer, because some may be null while others are not. 1056 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1057 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1058 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1059 Constant::getNullValue(GVVal->getType()), 1060 "tmp", NullPtrBlock); 1061 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent()); 1062 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent()); 1063 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1064 1065 // Fill in FreeBlock. 1066 new FreeInst(GVVal, FreeBlock); 1067 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1068 FreeBlock); 1069 new BranchInst(NextBlock, FreeBlock); 1070 1071 NullPtrBlock = NextBlock; 1072 } 1073 1074 new BranchInst(ContBB, NullPtrBlock); 1075 1076 1077 // MI is no longer needed, remove it. 1078 MI->eraseFromParent(); 1079 1080 1081 // Okay, the malloc site is completely handled. All of the uses of GV are now 1082 // loads, and all uses of those loads are simple. Rewrite them to use loads 1083 // of the per-field globals instead. 1084 while (!GV->use_empty()) { 1085 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 1086 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals); 1087 LI->eraseFromParent(); 1088 } else { 1089 // Must be a store of null. 1090 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1091 assert(isa<Constant>(SI->getOperand(0)) && 1092 cast<Constant>(SI->getOperand(0))->isNullValue() && 1093 "Unexpected heap-sra user!"); 1094 1095 // Insert a store of null into each global. 1096 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1097 Constant *Null = 1098 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType()); 1099 new StoreInst(Null, FieldGlobals[i], SI); 1100 } 1101 // Erase the original store. 1102 SI->eraseFromParent(); 1103 } 1104 } 1105 1106 // The old global is now dead, remove it. 1107 GV->eraseFromParent(); 1108 1109 ++NumHeapSRA; 1110 return FieldGlobals[0]; 1111} 1112 1113 1114// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1115// that only one value (besides its initializer) is ever stored to the global. 1116static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1117 Module::global_iterator &GVI, 1118 TargetData &TD) { 1119 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal)) 1120 StoredOnceVal = CI->getOperand(0); 1121 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){ 1122 // "getelementptr Ptr, 0, 0, 0" is really just a cast. 1123 bool IsJustACast = true; 1124 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 1125 if (!isa<Constant>(GEPI->getOperand(i)) || 1126 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) { 1127 IsJustACast = false; 1128 break; 1129 } 1130 if (IsJustACast) 1131 StoredOnceVal = GEPI->getOperand(0); 1132 } 1133 1134 // If we are dealing with a pointer global that is initialized to null and 1135 // only has one (non-null) value stored into it, then we can optimize any 1136 // users of the loaded value (often calls and loads) that would trap if the 1137 // value was null. 1138 if (isa<PointerType>(GV->getInitializer()->getType()) && 1139 GV->getInitializer()->isNullValue()) { 1140 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1141 if (GV->getInitializer()->getType() != SOVC->getType()) 1142 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1143 1144 // Optimize away any trapping uses of the loaded value. 1145 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1146 return true; 1147 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1148 // If this is a malloc of an abstract type, don't touch it. 1149 if (!MI->getAllocatedType()->isSized()) 1150 return false; 1151 1152 // We can't optimize this global unless all uses of it are *known* to be 1153 // of the malloc value, not of the null initializer value (consider a use 1154 // that compares the global's value against zero to see if the malloc has 1155 // been reached). To do this, we check to see if all uses of the global 1156 // would trap if the global were null: this proves that they must all 1157 // happen after the malloc. 1158 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1159 return false; 1160 1161 // We can't optimize this if the malloc itself is used in a complex way, 1162 // for example, being stored into multiple globals. This allows the 1163 // malloc to be stored into the specified global, loaded setcc'd, and 1164 // GEP'd. These are all things we could transform to using the global 1165 // for. 1166 { 1167 SmallPtrSet<PHINode*, 8> PHIs; 1168 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1169 return false; 1170 } 1171 1172 1173 // If we have a global that is only initialized with a fixed size malloc, 1174 // transform the program to use global memory instead of malloc'd memory. 1175 // This eliminates dynamic allocation, avoids an indirection accessing the 1176 // data, and exposes the resultant global to further GlobalOpt. 1177 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1178 // Restrict this transformation to only working on small allocations 1179 // (2048 bytes currently), as we don't want to introduce a 16M global or 1180 // something. 1181 if (NElements->getZExtValue()* 1182 TD.getTypeSize(MI->getAllocatedType()) < 2048) { 1183 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1184 return true; 1185 } 1186 } 1187 1188 // If the allocation is an array of structures, consider transforming this 1189 // into multiple malloc'd arrays, one for each field. This is basically 1190 // SRoA for malloc'd memory. 1191 if (const StructType *AllocTy = 1192 dyn_cast<StructType>(MI->getAllocatedType())) { 1193 // This the structure has an unreasonable number of fields, leave it 1194 // alone. 1195 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 && 1196 GlobalLoadUsesSimpleEnoughForHeapSRA(GV)) { 1197 GVI = PerformHeapAllocSRoA(GV, MI); 1198 return true; 1199 } 1200 } 1201 } 1202 } 1203 1204 return false; 1205} 1206 1207/// ShrinkGlobalToBoolean - At this point, we have learned that the only two 1208/// values ever stored into GV are its initializer and OtherVal. 1209static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1210 // Create the new global, initializing it to false. 1211 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1212 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1213 GV->getName()+".b", 1214 (Module *)NULL, 1215 GV->isThreadLocal()); 1216 GV->getParent()->getGlobalList().insert(GV, NewGV); 1217 1218 Constant *InitVal = GV->getInitializer(); 1219 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1220 1221 // If initialized to zero and storing one into the global, we can use a cast 1222 // instead of a select to synthesize the desired value. 1223 bool IsOneZero = false; 1224 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1225 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1226 1227 while (!GV->use_empty()) { 1228 Instruction *UI = cast<Instruction>(GV->use_back()); 1229 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1230 // Change the store into a boolean store. 1231 bool StoringOther = SI->getOperand(0) == OtherVal; 1232 // Only do this if we weren't storing a loaded value. 1233 Value *StoreVal; 1234 if (StoringOther || SI->getOperand(0) == InitVal) 1235 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1236 else { 1237 // Otherwise, we are storing a previously loaded copy. To do this, 1238 // change the copy from copying the original value to just copying the 1239 // bool. 1240 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1241 1242 // If we're already replaced the input, StoredVal will be a cast or 1243 // select instruction. If not, it will be a load of the original 1244 // global. 1245 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1246 assert(LI->getOperand(0) == GV && "Not a copy!"); 1247 // Insert a new load, to preserve the saved value. 1248 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1249 } else { 1250 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1251 "This is not a form that we understand!"); 1252 StoreVal = StoredVal->getOperand(0); 1253 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1254 } 1255 } 1256 new StoreInst(StoreVal, NewGV, SI); 1257 } else if (!UI->use_empty()) { 1258 // Change the load into a load of bool then a select. 1259 LoadInst *LI = cast<LoadInst>(UI); 1260 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1261 Value *NSI; 1262 if (IsOneZero) 1263 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1264 else 1265 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI); 1266 NSI->takeName(LI); 1267 LI->replaceAllUsesWith(NSI); 1268 } 1269 UI->eraseFromParent(); 1270 } 1271 1272 GV->eraseFromParent(); 1273} 1274 1275 1276/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1277/// it if possible. If we make a change, return true. 1278bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1279 Module::global_iterator &GVI) { 1280 std::set<PHINode*> PHIUsers; 1281 GlobalStatus GS; 1282 GV->removeDeadConstantUsers(); 1283 1284 if (GV->use_empty()) { 1285 DOUT << "GLOBAL DEAD: " << *GV; 1286 GV->eraseFromParent(); 1287 ++NumDeleted; 1288 return true; 1289 } 1290 1291 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1292#if 0 1293 cerr << "Global: " << *GV; 1294 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1295 cerr << " StoredType = "; 1296 switch (GS.StoredType) { 1297 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1298 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1299 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1300 case GlobalStatus::isStored: cerr << "stored\n"; break; 1301 } 1302 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1303 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1304 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1305 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1306 << "\n"; 1307 cerr << " HasMultipleAccessingFunctions = " 1308 << GS.HasMultipleAccessingFunctions << "\n"; 1309 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1310 cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n"; 1311 cerr << "\n"; 1312#endif 1313 1314 // If this is a first class global and has only one accessing function 1315 // and this function is main (which we know is not recursive we can make 1316 // this global a local variable) we replace the global with a local alloca 1317 // in this function. 1318 // 1319 // NOTE: It doesn't make sense to promote non first class types since we 1320 // are just replacing static memory to stack memory. 1321 if (!GS.HasMultipleAccessingFunctions && 1322 GS.AccessingFunction && !GS.HasNonInstructionUser && 1323 GV->getType()->getElementType()->isFirstClassType() && 1324 GS.AccessingFunction->getName() == "main" && 1325 GS.AccessingFunction->hasExternalLinkage()) { 1326 DOUT << "LOCALIZING GLOBAL: " << *GV; 1327 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1328 const Type* ElemTy = GV->getType()->getElementType(); 1329 // FIXME: Pass Global's alignment when globals have alignment 1330 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1331 if (!isa<UndefValue>(GV->getInitializer())) 1332 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1333 1334 GV->replaceAllUsesWith(Alloca); 1335 GV->eraseFromParent(); 1336 ++NumLocalized; 1337 return true; 1338 } 1339 1340 // If the global is never loaded (but may be stored to), it is dead. 1341 // Delete it now. 1342 if (!GS.isLoaded) { 1343 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1344 1345 // Delete any stores we can find to the global. We may not be able to 1346 // make it completely dead though. 1347 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1348 1349 // If the global is dead now, delete it. 1350 if (GV->use_empty()) { 1351 GV->eraseFromParent(); 1352 ++NumDeleted; 1353 Changed = true; 1354 } 1355 return Changed; 1356 1357 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1358 DOUT << "MARKING CONSTANT: " << *GV; 1359 GV->setConstant(true); 1360 1361 // Clean up any obviously simplifiable users now. 1362 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1363 1364 // If the global is dead now, just nuke it. 1365 if (GV->use_empty()) { 1366 DOUT << " *** Marking constant allowed us to simplify " 1367 << "all users and delete global!\n"; 1368 GV->eraseFromParent(); 1369 ++NumDeleted; 1370 } 1371 1372 ++NumMarked; 1373 return true; 1374 } else if (!GS.isNotSuitableForSRA && 1375 !GV->getInitializer()->getType()->isFirstClassType()) { 1376 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) { 1377 GVI = FirstNewGV; // Don't skip the newly produced globals! 1378 return true; 1379 } 1380 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1381 // If the initial value for the global was an undef value, and if only 1382 // one other value was stored into it, we can just change the 1383 // initializer to be an undef value, then delete all stores to the 1384 // global. This allows us to mark it constant. 1385 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1386 if (isa<UndefValue>(GV->getInitializer())) { 1387 // Change the initial value here. 1388 GV->setInitializer(SOVConstant); 1389 1390 // Clean up any obviously simplifiable users now. 1391 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1392 1393 if (GV->use_empty()) { 1394 DOUT << " *** Substituting initializer allowed us to " 1395 << "simplify all users and delete global!\n"; 1396 GV->eraseFromParent(); 1397 ++NumDeleted; 1398 } else { 1399 GVI = GV; 1400 } 1401 ++NumSubstitute; 1402 return true; 1403 } 1404 1405 // Try to optimize globals based on the knowledge that only one value 1406 // (besides its initializer) is ever stored to the global. 1407 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1408 getAnalysis<TargetData>())) 1409 return true; 1410 1411 // Otherwise, if the global was not a boolean, we can shrink it to be a 1412 // boolean. 1413 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1414 if (GV->getType()->getElementType() != Type::Int1Ty && 1415 !GV->getType()->getElementType()->isFloatingPoint() && 1416 !isa<VectorType>(GV->getType()->getElementType()) && 1417 !GS.HasPHIUser && !GS.isNotSuitableForSRA) { 1418 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1419 ShrinkGlobalToBoolean(GV, SOVConstant); 1420 ++NumShrunkToBool; 1421 return true; 1422 } 1423 } 1424 } 1425 return false; 1426} 1427 1428/// OnlyCalledDirectly - Return true if the specified function is only called 1429/// directly. In other words, its address is never taken. 1430static bool OnlyCalledDirectly(Function *F) { 1431 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1432 Instruction *User = dyn_cast<Instruction>(*UI); 1433 if (!User) return false; 1434 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1435 1436 // See if the function address is passed as an argument. 1437 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i) 1438 if (User->getOperand(i) == F) return false; 1439 } 1440 return true; 1441} 1442 1443/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1444/// function, changing them to FastCC. 1445static void ChangeCalleesToFastCall(Function *F) { 1446 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1447 Instruction *User = cast<Instruction>(*UI); 1448 if (CallInst *CI = dyn_cast<CallInst>(User)) 1449 CI->setCallingConv(CallingConv::Fast); 1450 else 1451 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast); 1452 } 1453} 1454 1455bool GlobalOpt::OptimizeFunctions(Module &M) { 1456 bool Changed = false; 1457 // Optimize functions. 1458 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1459 Function *F = FI++; 1460 F->removeDeadConstantUsers(); 1461 if (F->use_empty() && (F->hasInternalLinkage() || 1462 F->hasLinkOnceLinkage())) { 1463 M.getFunctionList().erase(F); 1464 Changed = true; 1465 ++NumFnDeleted; 1466 } else if (F->hasInternalLinkage() && 1467 F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1468 OnlyCalledDirectly(F)) { 1469 // If this function has C calling conventions, is not a varargs 1470 // function, and is only called directly, promote it to use the Fast 1471 // calling convention. 1472 F->setCallingConv(CallingConv::Fast); 1473 ChangeCalleesToFastCall(F); 1474 ++NumFastCallFns; 1475 Changed = true; 1476 } 1477 } 1478 return Changed; 1479} 1480 1481bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1482 bool Changed = false; 1483 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1484 GVI != E; ) { 1485 GlobalVariable *GV = GVI++; 1486 if (!GV->isConstant() && GV->hasInternalLinkage() && 1487 GV->hasInitializer()) 1488 Changed |= ProcessInternalGlobal(GV, GVI); 1489 } 1490 return Changed; 1491} 1492 1493/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1494/// initializers have an init priority of 65535. 1495GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1496 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1497 I != E; ++I) 1498 if (I->getName() == "llvm.global_ctors") { 1499 // Found it, verify it's an array of { int, void()* }. 1500 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1501 if (!ATy) return 0; 1502 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1503 if (!STy || STy->getNumElements() != 2 || 1504 STy->getElementType(0) != Type::Int32Ty) return 0; 1505 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1506 if (!PFTy) return 0; 1507 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1508 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1509 FTy->getNumParams() != 0) 1510 return 0; 1511 1512 // Verify that the initializer is simple enough for us to handle. 1513 if (!I->hasInitializer()) return 0; 1514 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1515 if (!CA) return 0; 1516 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1517 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) { 1518 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1519 continue; 1520 1521 // Must have a function or null ptr. 1522 if (!isa<Function>(CS->getOperand(1))) 1523 return 0; 1524 1525 // Init priority must be standard. 1526 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1527 if (!CI || CI->getZExtValue() != 65535) 1528 return 0; 1529 } else { 1530 return 0; 1531 } 1532 1533 return I; 1534 } 1535 return 0; 1536} 1537 1538/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1539/// return a list of the functions and null terminator as a vector. 1540static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1541 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1542 std::vector<Function*> Result; 1543 Result.reserve(CA->getNumOperands()); 1544 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) { 1545 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i)); 1546 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1547 } 1548 return Result; 1549} 1550 1551/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1552/// specified array, returning the new global to use. 1553static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1554 const std::vector<Function*> &Ctors) { 1555 // If we made a change, reassemble the initializer list. 1556 std::vector<Constant*> CSVals; 1557 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1558 CSVals.push_back(0); 1559 1560 // Create the new init list. 1561 std::vector<Constant*> CAList; 1562 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1563 if (Ctors[i]) { 1564 CSVals[1] = Ctors[i]; 1565 } else { 1566 const Type *FTy = FunctionType::get(Type::VoidTy, 1567 std::vector<const Type*>(), false); 1568 const PointerType *PFTy = PointerType::get(FTy); 1569 CSVals[1] = Constant::getNullValue(PFTy); 1570 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1571 } 1572 CAList.push_back(ConstantStruct::get(CSVals)); 1573 } 1574 1575 // Create the array initializer. 1576 const Type *StructTy = 1577 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1578 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1579 CAList); 1580 1581 // If we didn't change the number of elements, don't create a new GV. 1582 if (CA->getType() == GCL->getInitializer()->getType()) { 1583 GCL->setInitializer(CA); 1584 return GCL; 1585 } 1586 1587 // Create the new global and insert it next to the existing list. 1588 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1589 GCL->getLinkage(), CA, "", 1590 (Module *)NULL, 1591 GCL->isThreadLocal()); 1592 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1593 NGV->takeName(GCL); 1594 1595 // Nuke the old list, replacing any uses with the new one. 1596 if (!GCL->use_empty()) { 1597 Constant *V = NGV; 1598 if (V->getType() != GCL->getType()) 1599 V = ConstantExpr::getBitCast(V, GCL->getType()); 1600 GCL->replaceAllUsesWith(V); 1601 } 1602 GCL->eraseFromParent(); 1603 1604 if (Ctors.size()) 1605 return NGV; 1606 else 1607 return 0; 1608} 1609 1610 1611static Constant *getVal(std::map<Value*, Constant*> &ComputedValues, 1612 Value *V) { 1613 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1614 Constant *R = ComputedValues[V]; 1615 assert(R && "Reference to an uncomputed value!"); 1616 return R; 1617} 1618 1619/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1620/// enough for us to understand. In particular, if it is a cast of something, 1621/// we punt. We basically just support direct accesses to globals and GEP's of 1622/// globals. This should be kept up to date with CommitValueTo. 1623static bool isSimpleEnoughPointerToCommit(Constant *C) { 1624 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1625 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1626 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1627 return !GV->isDeclaration(); // reject external globals. 1628 } 1629 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 1630 // Handle a constantexpr gep. 1631 if (CE->getOpcode() == Instruction::GetElementPtr && 1632 isa<GlobalVariable>(CE->getOperand(0))) { 1633 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1634 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1635 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1636 return GV->hasInitializer() && 1637 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1638 } 1639 return false; 1640} 1641 1642/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 1643/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 1644/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 1645static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 1646 ConstantExpr *Addr, unsigned OpNo) { 1647 // Base case of the recursion. 1648 if (OpNo == Addr->getNumOperands()) { 1649 assert(Val->getType() == Init->getType() && "Type mismatch!"); 1650 return Val; 1651 } 1652 1653 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1654 std::vector<Constant*> Elts; 1655 1656 // Break up the constant into its elements. 1657 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1658 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) 1659 Elts.push_back(CS->getOperand(i)); 1660 } else if (isa<ConstantAggregateZero>(Init)) { 1661 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1662 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 1663 } else if (isa<UndefValue>(Init)) { 1664 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1665 Elts.push_back(UndefValue::get(STy->getElementType(i))); 1666 } else { 1667 assert(0 && "This code is out of sync with " 1668 " ConstantFoldLoadThroughGEPConstantExpr"); 1669 } 1670 1671 // Replace the element that we are supposed to. 1672 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 1673 unsigned Idx = CU->getZExtValue(); 1674 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 1675 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 1676 1677 // Return the modified struct. 1678 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 1679 } else { 1680 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 1681 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 1682 1683 // Break up the array into elements. 1684 std::vector<Constant*> Elts; 1685 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1686 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1687 Elts.push_back(CA->getOperand(i)); 1688 } else if (isa<ConstantAggregateZero>(Init)) { 1689 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 1690 Elts.assign(ATy->getNumElements(), Elt); 1691 } else if (isa<UndefValue>(Init)) { 1692 Constant *Elt = UndefValue::get(ATy->getElementType()); 1693 Elts.assign(ATy->getNumElements(), Elt); 1694 } else { 1695 assert(0 && "This code is out of sync with " 1696 " ConstantFoldLoadThroughGEPConstantExpr"); 1697 } 1698 1699 assert(CI->getZExtValue() < ATy->getNumElements()); 1700 Elts[CI->getZExtValue()] = 1701 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 1702 return ConstantArray::get(ATy, Elts); 1703 } 1704} 1705 1706/// CommitValueTo - We have decided that Addr (which satisfies the predicate 1707/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 1708static void CommitValueTo(Constant *Val, Constant *Addr) { 1709 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 1710 assert(GV->hasInitializer()); 1711 GV->setInitializer(Val); 1712 return; 1713 } 1714 1715 ConstantExpr *CE = cast<ConstantExpr>(Addr); 1716 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1717 1718 Constant *Init = GV->getInitializer(); 1719 Init = EvaluateStoreInto(Init, Val, CE, 2); 1720 GV->setInitializer(Init); 1721} 1722 1723/// ComputeLoadResult - Return the value that would be computed by a load from 1724/// P after the stores reflected by 'memory' have been performed. If we can't 1725/// decide, return null. 1726static Constant *ComputeLoadResult(Constant *P, 1727 const std::map<Constant*, Constant*> &Memory) { 1728 // If this memory location has been recently stored, use the stored value: it 1729 // is the most up-to-date. 1730 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P); 1731 if (I != Memory.end()) return I->second; 1732 1733 // Access it. 1734 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 1735 if (GV->hasInitializer()) 1736 return GV->getInitializer(); 1737 return 0; 1738 } 1739 1740 // Handle a constantexpr getelementptr. 1741 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 1742 if (CE->getOpcode() == Instruction::GetElementPtr && 1743 isa<GlobalVariable>(CE->getOperand(0))) { 1744 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1745 if (GV->hasInitializer()) 1746 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1747 } 1748 1749 return 0; // don't know how to evaluate. 1750} 1751 1752/// EvaluateFunction - Evaluate a call to function F, returning true if 1753/// successful, false if we can't evaluate it. ActualArgs contains the formal 1754/// arguments for the function. 1755static bool EvaluateFunction(Function *F, Constant *&RetVal, 1756 const std::vector<Constant*> &ActualArgs, 1757 std::vector<Function*> &CallStack, 1758 std::map<Constant*, Constant*> &MutatedMemory, 1759 std::vector<GlobalVariable*> &AllocaTmps) { 1760 // Check to see if this function is already executing (recursion). If so, 1761 // bail out. TODO: we might want to accept limited recursion. 1762 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 1763 return false; 1764 1765 CallStack.push_back(F); 1766 1767 /// Values - As we compute SSA register values, we store their contents here. 1768 std::map<Value*, Constant*> Values; 1769 1770 // Initialize arguments to the incoming values specified. 1771 unsigned ArgNo = 0; 1772 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1773 ++AI, ++ArgNo) 1774 Values[AI] = ActualArgs[ArgNo]; 1775 1776 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 1777 /// we can only evaluate any one basic block at most once. This set keeps 1778 /// track of what we have executed so we can detect recursive cases etc. 1779 std::set<BasicBlock*> ExecutedBlocks; 1780 1781 // CurInst - The current instruction we're evaluating. 1782 BasicBlock::iterator CurInst = F->begin()->begin(); 1783 1784 // This is the main evaluation loop. 1785 while (1) { 1786 Constant *InstResult = 0; 1787 1788 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 1789 if (SI->isVolatile()) return false; // no volatile accesses. 1790 Constant *Ptr = getVal(Values, SI->getOperand(1)); 1791 if (!isSimpleEnoughPointerToCommit(Ptr)) 1792 // If this is too complex for us to commit, reject it. 1793 return false; 1794 Constant *Val = getVal(Values, SI->getOperand(0)); 1795 MutatedMemory[Ptr] = Val; 1796 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 1797 InstResult = ConstantExpr::get(BO->getOpcode(), 1798 getVal(Values, BO->getOperand(0)), 1799 getVal(Values, BO->getOperand(1))); 1800 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 1801 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 1802 getVal(Values, CI->getOperand(0)), 1803 getVal(Values, CI->getOperand(1))); 1804 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 1805 InstResult = ConstantExpr::getCast(CI->getOpcode(), 1806 getVal(Values, CI->getOperand(0)), 1807 CI->getType()); 1808 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 1809 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 1810 getVal(Values, SI->getOperand(1)), 1811 getVal(Values, SI->getOperand(2))); 1812 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 1813 Constant *P = getVal(Values, GEP->getOperand(0)); 1814 SmallVector<Constant*, 8> GEPOps; 1815 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i) 1816 GEPOps.push_back(getVal(Values, GEP->getOperand(i))); 1817 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 1818 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 1819 if (LI->isVolatile()) return false; // no volatile accesses. 1820 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 1821 MutatedMemory); 1822 if (InstResult == 0) return false; // Could not evaluate load. 1823 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 1824 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 1825 const Type *Ty = AI->getType()->getElementType(); 1826 AllocaTmps.push_back(new GlobalVariable(Ty, false, 1827 GlobalValue::InternalLinkage, 1828 UndefValue::get(Ty), 1829 AI->getName())); 1830 InstResult = AllocaTmps.back(); 1831 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 1832 // Cannot handle inline asm. 1833 if (isa<InlineAsm>(CI->getOperand(0))) return false; 1834 1835 // Resolve function pointers. 1836 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 1837 if (!Callee) return false; // Cannot resolve. 1838 1839 std::vector<Constant*> Formals; 1840 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i) 1841 Formals.push_back(getVal(Values, CI->getOperand(i))); 1842 1843 if (Callee->isDeclaration()) { 1844 // If this is a function we can constant fold, do it. 1845 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 1846 Formals.size())) { 1847 InstResult = C; 1848 } else { 1849 return false; 1850 } 1851 } else { 1852 if (Callee->getFunctionType()->isVarArg()) 1853 return false; 1854 1855 Constant *RetVal; 1856 1857 // Execute the call, if successful, use the return value. 1858 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 1859 MutatedMemory, AllocaTmps)) 1860 return false; 1861 InstResult = RetVal; 1862 } 1863 } else if (isa<TerminatorInst>(CurInst)) { 1864 BasicBlock *NewBB = 0; 1865 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 1866 if (BI->isUnconditional()) { 1867 NewBB = BI->getSuccessor(0); 1868 } else { 1869 ConstantInt *Cond = 1870 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 1871 if (!Cond) return false; // Cannot determine. 1872 1873 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 1874 } 1875 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 1876 ConstantInt *Val = 1877 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 1878 if (!Val) return false; // Cannot determine. 1879 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 1880 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 1881 if (RI->getNumOperands()) 1882 RetVal = getVal(Values, RI->getOperand(0)); 1883 1884 CallStack.pop_back(); // return from fn. 1885 return true; // We succeeded at evaluating this ctor! 1886 } else { 1887 // invoke, unwind, unreachable. 1888 return false; // Cannot handle this terminator. 1889 } 1890 1891 // Okay, we succeeded in evaluating this control flow. See if we have 1892 // executed the new block before. If so, we have a looping function, 1893 // which we cannot evaluate in reasonable time. 1894 if (!ExecutedBlocks.insert(NewBB).second) 1895 return false; // looped! 1896 1897 // Okay, we have never been in this block before. Check to see if there 1898 // are any PHI nodes. If so, evaluate them with information about where 1899 // we came from. 1900 BasicBlock *OldBB = CurInst->getParent(); 1901 CurInst = NewBB->begin(); 1902 PHINode *PN; 1903 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 1904 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 1905 1906 // Do NOT increment CurInst. We know that the terminator had no value. 1907 continue; 1908 } else { 1909 // Did not know how to evaluate this! 1910 return false; 1911 } 1912 1913 if (!CurInst->use_empty()) 1914 Values[CurInst] = InstResult; 1915 1916 // Advance program counter. 1917 ++CurInst; 1918 } 1919} 1920 1921/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 1922/// we can. Return true if we can, false otherwise. 1923static bool EvaluateStaticConstructor(Function *F) { 1924 /// MutatedMemory - For each store we execute, we update this map. Loads 1925 /// check this to get the most up-to-date value. If evaluation is successful, 1926 /// this state is committed to the process. 1927 std::map<Constant*, Constant*> MutatedMemory; 1928 1929 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 1930 /// to represent its body. This vector is needed so we can delete the 1931 /// temporary globals when we are done. 1932 std::vector<GlobalVariable*> AllocaTmps; 1933 1934 /// CallStack - This is used to detect recursion. In pathological situations 1935 /// we could hit exponential behavior, but at least there is nothing 1936 /// unbounded. 1937 std::vector<Function*> CallStack; 1938 1939 // Call the function. 1940 Constant *RetValDummy; 1941 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 1942 CallStack, MutatedMemory, AllocaTmps); 1943 if (EvalSuccess) { 1944 // We succeeded at evaluation: commit the result. 1945 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 1946 << F->getName() << "' to " << MutatedMemory.size() 1947 << " stores.\n"; 1948 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 1949 E = MutatedMemory.end(); I != E; ++I) 1950 CommitValueTo(I->second, I->first); 1951 } 1952 1953 // At this point, we are done interpreting. If we created any 'alloca' 1954 // temporaries, release them now. 1955 while (!AllocaTmps.empty()) { 1956 GlobalVariable *Tmp = AllocaTmps.back(); 1957 AllocaTmps.pop_back(); 1958 1959 // If there are still users of the alloca, the program is doing something 1960 // silly, e.g. storing the address of the alloca somewhere and using it 1961 // later. Since this is undefined, we'll just make it be null. 1962 if (!Tmp->use_empty()) 1963 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 1964 delete Tmp; 1965 } 1966 1967 return EvalSuccess; 1968} 1969 1970 1971 1972/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 1973/// Return true if anything changed. 1974bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 1975 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 1976 bool MadeChange = false; 1977 if (Ctors.empty()) return false; 1978 1979 // Loop over global ctors, optimizing them when we can. 1980 for (unsigned i = 0; i != Ctors.size(); ++i) { 1981 Function *F = Ctors[i]; 1982 // Found a null terminator in the middle of the list, prune off the rest of 1983 // the list. 1984 if (F == 0) { 1985 if (i != Ctors.size()-1) { 1986 Ctors.resize(i+1); 1987 MadeChange = true; 1988 } 1989 break; 1990 } 1991 1992 // We cannot simplify external ctor functions. 1993 if (F->empty()) continue; 1994 1995 // If we can evaluate the ctor at compile time, do. 1996 if (EvaluateStaticConstructor(F)) { 1997 Ctors.erase(Ctors.begin()+i); 1998 MadeChange = true; 1999 --i; 2000 ++NumCtorsEvaluated; 2001 continue; 2002 } 2003 } 2004 2005 if (!MadeChange) return false; 2006 2007 GCL = InstallGlobalCtors(GCL, Ctors); 2008 return true; 2009} 2010 2011 2012bool GlobalOpt::runOnModule(Module &M) { 2013 bool Changed = false; 2014 2015 // Try to find the llvm.globalctors list. 2016 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2017 2018 bool LocalChange = true; 2019 while (LocalChange) { 2020 LocalChange = false; 2021 2022 // Delete functions that are trivially dead, ccc -> fastcc 2023 LocalChange |= OptimizeFunctions(M); 2024 2025 // Optimize global_ctors list. 2026 if (GlobalCtors) 2027 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2028 2029 // Optimize non-address-taken globals. 2030 LocalChange |= OptimizeGlobalVars(M); 2031 Changed |= LocalChange; 2032 } 2033 2034 // TODO: Move all global ctors functions to the end of the module for code 2035 // layout. 2036 2037 return Changed; 2038} 2039