GlobalOpt.cpp revision 2e729117046dc836fe8e19ebe797caa0be7e2a99
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass transforms simple global variables that never have their address 11// taken. If obviously true, it marks read/write globals as constant, deletes 12// variables only stored to, etc. 13// 14//===----------------------------------------------------------------------===// 15 16#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/CallSite.h" 28#include "llvm/Support/Compiler.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/Support/GetElementPtrTypeIterator.h" 31#include "llvm/Support/MathExtras.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/SmallVector.h" 34#include "llvm/ADT/Statistic.h" 35#include "llvm/ADT/StringExtras.h" 36#include <algorithm> 37#include <map> 38#include <set> 39using namespace llvm; 40 41STATISTIC(NumMarked , "Number of globals marked constant"); 42STATISTIC(NumSRA , "Number of aggregate globals broken into scalars"); 43STATISTIC(NumHeapSRA , "Number of heap objects SRA'd"); 44STATISTIC(NumSubstitute,"Number of globals with initializers stored into them"); 45STATISTIC(NumDeleted , "Number of globals deleted"); 46STATISTIC(NumFnDeleted , "Number of functions deleted"); 47STATISTIC(NumGlobUses , "Number of global uses devirtualized"); 48STATISTIC(NumLocalized , "Number of globals localized"); 49STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans"); 50STATISTIC(NumFastCallFns , "Number of functions converted to fastcc"); 51STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated"); 52STATISTIC(NumNestRemoved , "Number of nest attributes removed"); 53 54namespace { 55 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass { 56 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 57 AU.addRequired<TargetData>(); 58 } 59 static char ID; // Pass identification, replacement for typeid 60 GlobalOpt() : ModulePass(&ID) {} 61 62 bool runOnModule(Module &M); 63 64 private: 65 GlobalVariable *FindGlobalCtors(Module &M); 66 bool OptimizeFunctions(Module &M); 67 bool OptimizeGlobalVars(Module &M); 68 bool ResolveAliases(Module &M); 69 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL); 70 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 71 }; 72} 73 74char GlobalOpt::ID = 0; 75static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer"); 76 77ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 78 79namespace { 80 81/// GlobalStatus - As we analyze each global, keep track of some information 82/// about it. If we find out that the address of the global is taken, none of 83/// this info will be accurate. 84struct VISIBILITY_HIDDEN GlobalStatus { 85 /// isLoaded - True if the global is ever loaded. If the global isn't ever 86 /// loaded it can be deleted. 87 bool isLoaded; 88 89 /// StoredType - Keep track of what stores to the global look like. 90 /// 91 enum StoredType { 92 /// NotStored - There is no store to this global. It can thus be marked 93 /// constant. 94 NotStored, 95 96 /// isInitializerStored - This global is stored to, but the only thing 97 /// stored is the constant it was initialized with. This is only tracked 98 /// for scalar globals. 99 isInitializerStored, 100 101 /// isStoredOnce - This global is stored to, but only its initializer and 102 /// one other value is ever stored to it. If this global isStoredOnce, we 103 /// track the value stored to it in StoredOnceValue below. This is only 104 /// tracked for scalar globals. 105 isStoredOnce, 106 107 /// isStored - This global is stored to by multiple values or something else 108 /// that we cannot track. 109 isStored 110 } StoredType; 111 112 /// StoredOnceValue - If only one value (besides the initializer constant) is 113 /// ever stored to this global, keep track of what value it is. 114 Value *StoredOnceValue; 115 116 /// AccessingFunction/HasMultipleAccessingFunctions - These start out 117 /// null/false. When the first accessing function is noticed, it is recorded. 118 /// When a second different accessing function is noticed, 119 /// HasMultipleAccessingFunctions is set to true. 120 Function *AccessingFunction; 121 bool HasMultipleAccessingFunctions; 122 123 /// HasNonInstructionUser - Set to true if this global has a user that is not 124 /// an instruction (e.g. a constant expr or GV initializer). 125 bool HasNonInstructionUser; 126 127 /// HasPHIUser - Set to true if this global has a user that is a PHI node. 128 bool HasPHIUser; 129 130 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0), 131 AccessingFunction(0), HasMultipleAccessingFunctions(false), 132 HasNonInstructionUser(false), HasPHIUser(false) {} 133}; 134 135} 136 137/// ConstantIsDead - Return true if the specified constant is (transitively) 138/// dead. The constant may be used by other constants (e.g. constant arrays and 139/// constant exprs) as long as they are dead, but it cannot be used by anything 140/// else. 141static bool ConstantIsDead(Constant *C) { 142 if (isa<GlobalValue>(C)) return false; 143 144 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI) 145 if (Constant *CU = dyn_cast<Constant>(*UI)) { 146 if (!ConstantIsDead(CU)) return false; 147 } else 148 return false; 149 return true; 150} 151 152 153/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 154/// structure. If the global has its address taken, return true to indicate we 155/// can't do anything with it. 156/// 157static bool AnalyzeGlobal(Value *V, GlobalStatus &GS, 158 std::set<PHINode*> &PHIUsers) { 159 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 160 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) { 161 GS.HasNonInstructionUser = true; 162 163 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 164 165 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) { 166 if (!GS.HasMultipleAccessingFunctions) { 167 Function *F = I->getParent()->getParent(); 168 if (GS.AccessingFunction == 0) 169 GS.AccessingFunction = F; 170 else if (GS.AccessingFunction != F) 171 GS.HasMultipleAccessingFunctions = true; 172 } 173 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 174 GS.isLoaded = true; 175 if (LI->isVolatile()) return true; // Don't hack on volatile loads. 176 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 177 // Don't allow a store OF the address, only stores TO the address. 178 if (SI->getOperand(0) == V) return true; 179 180 if (SI->isVolatile()) return true; // Don't hack on volatile stores. 181 182 // If this is a direct store to the global (i.e., the global is a scalar 183 // value, not an aggregate), keep more specific information about 184 // stores. 185 if (GS.StoredType != GlobalStatus::isStored) { 186 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){ 187 Value *StoredVal = SI->getOperand(0); 188 if (StoredVal == GV->getInitializer()) { 189 if (GS.StoredType < GlobalStatus::isInitializerStored) 190 GS.StoredType = GlobalStatus::isInitializerStored; 191 } else if (isa<LoadInst>(StoredVal) && 192 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 193 // G = G 194 if (GS.StoredType < GlobalStatus::isInitializerStored) 195 GS.StoredType = GlobalStatus::isInitializerStored; 196 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 197 GS.StoredType = GlobalStatus::isStoredOnce; 198 GS.StoredOnceValue = StoredVal; 199 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 200 GS.StoredOnceValue == StoredVal) { 201 // noop. 202 } else { 203 GS.StoredType = GlobalStatus::isStored; 204 } 205 } else { 206 GS.StoredType = GlobalStatus::isStored; 207 } 208 } 209 } else if (isa<GetElementPtrInst>(I)) { 210 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 211 } else if (isa<SelectInst>(I)) { 212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 213 } else if (PHINode *PN = dyn_cast<PHINode>(I)) { 214 // PHI nodes we can check just like select or GEP instructions, but we 215 // have to be careful about infinite recursion. 216 if (PHIUsers.insert(PN).second) // Not already visited. 217 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 218 GS.HasPHIUser = true; 219 } else if (isa<CmpInst>(I)) { 220 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) { 221 if (I->getOperand(1) == V) 222 GS.StoredType = GlobalStatus::isStored; 223 if (I->getOperand(2) == V) 224 GS.isLoaded = true; 225 } else if (isa<MemSetInst>(I)) { 226 assert(I->getOperand(1) == V && "Memset only takes one pointer!"); 227 GS.StoredType = GlobalStatus::isStored; 228 } else { 229 return true; // Any other non-load instruction might take address! 230 } 231 } else if (Constant *C = dyn_cast<Constant>(*UI)) { 232 GS.HasNonInstructionUser = true; 233 // We might have a dead and dangling constant hanging off of here. 234 if (!ConstantIsDead(C)) 235 return true; 236 } else { 237 GS.HasNonInstructionUser = true; 238 // Otherwise must be some other user. 239 return true; 240 } 241 242 return false; 243} 244 245static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 246 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 247 if (!CI) return 0; 248 unsigned IdxV = CI->getZExtValue(); 249 250 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 251 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 252 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 253 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 254 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 255 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 256 } else if (isa<ConstantAggregateZero>(Agg)) { 257 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 258 if (IdxV < STy->getNumElements()) 259 return Constant::getNullValue(STy->getElementType(IdxV)); 260 } else if (const SequentialType *STy = 261 dyn_cast<SequentialType>(Agg->getType())) { 262 return Constant::getNullValue(STy->getElementType()); 263 } 264 } else if (isa<UndefValue>(Agg)) { 265 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 266 if (IdxV < STy->getNumElements()) 267 return UndefValue::get(STy->getElementType(IdxV)); 268 } else if (const SequentialType *STy = 269 dyn_cast<SequentialType>(Agg->getType())) { 270 return UndefValue::get(STy->getElementType()); 271 } 272 } 273 return 0; 274} 275 276 277/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 278/// users of the global, cleaning up the obvious ones. This is largely just a 279/// quick scan over the use list to clean up the easy and obvious cruft. This 280/// returns true if it made a change. 281static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 282 bool Changed = false; 283 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 284 User *U = *UI++; 285 286 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 287 if (Init) { 288 // Replace the load with the initializer. 289 LI->replaceAllUsesWith(Init); 290 LI->eraseFromParent(); 291 Changed = true; 292 } 293 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 294 // Store must be unreachable or storing Init into the global. 295 SI->eraseFromParent(); 296 Changed = true; 297 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 298 if (CE->getOpcode() == Instruction::GetElementPtr) { 299 Constant *SubInit = 0; 300 if (Init) 301 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 302 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 303 } else if (CE->getOpcode() == Instruction::BitCast && 304 isa<PointerType>(CE->getType())) { 305 // Pointer cast, delete any stores and memsets to the global. 306 Changed |= CleanupConstantGlobalUsers(CE, 0); 307 } 308 309 if (CE->use_empty()) { 310 CE->destroyConstant(); 311 Changed = true; 312 } 313 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 314 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 315 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 316 // and will invalidate our notion of what Init is. 317 Constant *SubInit = 0; 318 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 319 ConstantExpr *CE = 320 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 321 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 322 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 323 } 324 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 325 326 if (GEP->use_empty()) { 327 GEP->eraseFromParent(); 328 Changed = true; 329 } 330 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 331 if (MI->getRawDest() == V) { 332 MI->eraseFromParent(); 333 Changed = true; 334 } 335 336 } else if (Constant *C = dyn_cast<Constant>(U)) { 337 // If we have a chain of dead constantexprs or other things dangling from 338 // us, and if they are all dead, nuke them without remorse. 339 if (ConstantIsDead(C)) { 340 C->destroyConstant(); 341 // This could have invalidated UI, start over from scratch. 342 CleanupConstantGlobalUsers(V, Init); 343 return true; 344 } 345 } 346 } 347 return Changed; 348} 349 350/// isSafeSROAElementUse - Return true if the specified instruction is a safe 351/// user of a derived expression from a global that we want to SROA. 352static bool isSafeSROAElementUse(Value *V) { 353 // We might have a dead and dangling constant hanging off of here. 354 if (Constant *C = dyn_cast<Constant>(V)) 355 return ConstantIsDead(C); 356 357 Instruction *I = dyn_cast<Instruction>(V); 358 if (!I) return false; 359 360 // Loads are ok. 361 if (isa<LoadInst>(I)) return true; 362 363 // Stores *to* the pointer are ok. 364 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 365 return SI->getOperand(0) != V; 366 367 // Otherwise, it must be a GEP. 368 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 369 if (GEPI == 0) return false; 370 371 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 372 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 373 return false; 374 375 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end(); 376 I != E; ++I) 377 if (!isSafeSROAElementUse(*I)) 378 return false; 379 return true; 380} 381 382 383/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 384/// Look at it and its uses and decide whether it is safe to SROA this global. 385/// 386static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 387 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 388 if (!isa<GetElementPtrInst>(U) && 389 (!isa<ConstantExpr>(U) || 390 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 391 return false; 392 393 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 394 // don't like < 3 operand CE's, and we don't like non-constant integer 395 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 396 // value of C. 397 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 398 !cast<Constant>(U->getOperand(1))->isNullValue() || 399 !isa<ConstantInt>(U->getOperand(2))) 400 return false; 401 402 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 403 ++GEPI; // Skip over the pointer index. 404 405 // If this is a use of an array allocation, do a bit more checking for sanity. 406 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 407 uint64_t NumElements = AT->getNumElements(); 408 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 409 410 // Check to make sure that index falls within the array. If not, 411 // something funny is going on, so we won't do the optimization. 412 // 413 if (Idx->getZExtValue() >= NumElements) 414 return false; 415 416 // We cannot scalar repl this level of the array unless any array 417 // sub-indices are in-range constants. In particular, consider: 418 // A[0][i]. We cannot know that the user isn't doing invalid things like 419 // allowing i to index an out-of-range subscript that accesses A[1]. 420 // 421 // Scalar replacing *just* the outer index of the array is probably not 422 // going to be a win anyway, so just give up. 423 for (++GEPI; // Skip array index. 424 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI)); 425 ++GEPI) { 426 uint64_t NumElements; 427 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 428 NumElements = SubArrayTy->getNumElements(); 429 else 430 NumElements = cast<VectorType>(*GEPI)->getNumElements(); 431 432 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 433 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 434 return false; 435 } 436 } 437 438 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) 439 if (!isSafeSROAElementUse(*I)) 440 return false; 441 return true; 442} 443 444/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 445/// is safe for us to perform this transformation. 446/// 447static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 448 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 449 UI != E; ++UI) { 450 if (!IsUserOfGlobalSafeForSRA(*UI, GV)) 451 return false; 452 } 453 return true; 454} 455 456 457/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 458/// variable. This opens the door for other optimizations by exposing the 459/// behavior of the program in a more fine-grained way. We have determined that 460/// this transformation is safe already. We return the first global variable we 461/// insert so that the caller can reprocess it. 462static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { 463 // Make sure this global only has simple uses that we can SRA. 464 if (!GlobalUsersSafeToSRA(GV)) 465 return 0; 466 467 assert(GV->hasInternalLinkage() && !GV->isConstant()); 468 Constant *Init = GV->getInitializer(); 469 const Type *Ty = Init->getType(); 470 471 std::vector<GlobalVariable*> NewGlobals; 472 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 473 474 // Get the alignment of the global, either explicit or target-specific. 475 unsigned StartAlignment = GV->getAlignment(); 476 if (StartAlignment == 0) 477 StartAlignment = TD.getABITypeAlignment(GV->getType()); 478 479 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 480 NewGlobals.reserve(STy->getNumElements()); 481 const StructLayout &Layout = *TD.getStructLayout(STy); 482 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 483 Constant *In = getAggregateConstantElement(Init, 484 ConstantInt::get(Type::Int32Ty, i)); 485 assert(In && "Couldn't get element of initializer?"); 486 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 487 GlobalVariable::InternalLinkage, 488 In, GV->getName()+"."+utostr(i), 489 (Module *)NULL, 490 GV->isThreadLocal(), 491 GV->getType()->getAddressSpace()); 492 Globals.insert(GV, NGV); 493 NewGlobals.push_back(NGV); 494 495 // Calculate the known alignment of the field. If the original aggregate 496 // had 256 byte alignment for example, something might depend on that: 497 // propagate info to each field. 498 uint64_t FieldOffset = Layout.getElementOffset(i); 499 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 500 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 501 NGV->setAlignment(NewAlign); 502 } 503 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 504 unsigned NumElements = 0; 505 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 506 NumElements = ATy->getNumElements(); 507 else 508 NumElements = cast<VectorType>(STy)->getNumElements(); 509 510 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 511 return 0; // It's not worth it. 512 NewGlobals.reserve(NumElements); 513 514 uint64_t EltSize = TD.getABITypeSize(STy->getElementType()); 515 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 516 for (unsigned i = 0, e = NumElements; i != e; ++i) { 517 Constant *In = getAggregateConstantElement(Init, 518 ConstantInt::get(Type::Int32Ty, i)); 519 assert(In && "Couldn't get element of initializer?"); 520 521 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 522 GlobalVariable::InternalLinkage, 523 In, GV->getName()+"."+utostr(i), 524 (Module *)NULL, 525 GV->isThreadLocal(), 526 GV->getType()->getAddressSpace()); 527 Globals.insert(GV, NGV); 528 NewGlobals.push_back(NGV); 529 530 // Calculate the known alignment of the field. If the original aggregate 531 // had 256 byte alignment for example, something might depend on that: 532 // propagate info to each field. 533 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 534 if (NewAlign > EltAlign) 535 NGV->setAlignment(NewAlign); 536 } 537 } 538 539 if (NewGlobals.empty()) 540 return 0; 541 542 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 543 544 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 545 546 // Loop over all of the uses of the global, replacing the constantexpr geps, 547 // with smaller constantexpr geps or direct references. 548 while (!GV->use_empty()) { 549 User *GEP = GV->use_back(); 550 assert(((isa<ConstantExpr>(GEP) && 551 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 552 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 553 554 // Ignore the 1th operand, which has to be zero or else the program is quite 555 // broken (undefined). Get the 2nd operand, which is the structure or array 556 // index. 557 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 558 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 559 560 Value *NewPtr = NewGlobals[Val]; 561 562 // Form a shorter GEP if needed. 563 if (GEP->getNumOperands() > 3) { 564 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 565 SmallVector<Constant*, 8> Idxs; 566 Idxs.push_back(NullInt); 567 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 568 Idxs.push_back(CE->getOperand(i)); 569 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 570 &Idxs[0], Idxs.size()); 571 } else { 572 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 573 SmallVector<Value*, 8> Idxs; 574 Idxs.push_back(NullInt); 575 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 576 Idxs.push_back(GEPI->getOperand(i)); 577 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 578 GEPI->getName()+"."+utostr(Val), GEPI); 579 } 580 } 581 GEP->replaceAllUsesWith(NewPtr); 582 583 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 584 GEPI->eraseFromParent(); 585 else 586 cast<ConstantExpr>(GEP)->destroyConstant(); 587 } 588 589 // Delete the old global, now that it is dead. 590 Globals.erase(GV); 591 ++NumSRA; 592 593 // Loop over the new globals array deleting any globals that are obviously 594 // dead. This can arise due to scalarization of a structure or an array that 595 // has elements that are dead. 596 unsigned FirstGlobal = 0; 597 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 598 if (NewGlobals[i]->use_empty()) { 599 Globals.erase(NewGlobals[i]); 600 if (FirstGlobal == i) ++FirstGlobal; 601 } 602 603 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 604} 605 606/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 607/// value will trap if the value is dynamically null. PHIs keeps track of any 608/// phi nodes we've seen to avoid reprocessing them. 609static bool AllUsesOfValueWillTrapIfNull(Value *V, 610 SmallPtrSet<PHINode*, 8> &PHIs) { 611 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 612 if (isa<LoadInst>(*UI)) { 613 // Will trap. 614 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 615 if (SI->getOperand(0) == V) { 616 //cerr << "NONTRAPPING USE: " << **UI; 617 return false; // Storing the value. 618 } 619 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 620 if (CI->getOperand(0) != V) { 621 //cerr << "NONTRAPPING USE: " << **UI; 622 return false; // Not calling the ptr 623 } 624 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 625 if (II->getOperand(0) != V) { 626 //cerr << "NONTRAPPING USE: " << **UI; 627 return false; // Not calling the ptr 628 } 629 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 630 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 631 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 632 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 633 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 634 // If we've already seen this phi node, ignore it, it has already been 635 // checked. 636 if (PHIs.insert(PN)) 637 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 638 } else if (isa<ICmpInst>(*UI) && 639 isa<ConstantPointerNull>(UI->getOperand(1))) { 640 // Ignore setcc X, null 641 } else { 642 //cerr << "NONTRAPPING USE: " << **UI; 643 return false; 644 } 645 return true; 646} 647 648/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 649/// from GV will trap if the loaded value is null. Note that this also permits 650/// comparisons of the loaded value against null, as a special case. 651static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 652 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 653 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 654 SmallPtrSet<PHINode*, 8> PHIs; 655 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 656 return false; 657 } else if (isa<StoreInst>(*UI)) { 658 // Ignore stores to the global. 659 } else { 660 // We don't know or understand this user, bail out. 661 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 662 return false; 663 } 664 665 return true; 666} 667 668static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 669 bool Changed = false; 670 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 671 Instruction *I = cast<Instruction>(*UI++); 672 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 673 LI->setOperand(0, NewV); 674 Changed = true; 675 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 676 if (SI->getOperand(1) == V) { 677 SI->setOperand(1, NewV); 678 Changed = true; 679 } 680 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 681 if (I->getOperand(0) == V) { 682 // Calling through the pointer! Turn into a direct call, but be careful 683 // that the pointer is not also being passed as an argument. 684 I->setOperand(0, NewV); 685 Changed = true; 686 bool PassedAsArg = false; 687 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 688 if (I->getOperand(i) == V) { 689 PassedAsArg = true; 690 I->setOperand(i, NewV); 691 } 692 693 if (PassedAsArg) { 694 // Being passed as an argument also. Be careful to not invalidate UI! 695 UI = V->use_begin(); 696 } 697 } 698 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 699 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 700 ConstantExpr::getCast(CI->getOpcode(), 701 NewV, CI->getType())); 702 if (CI->use_empty()) { 703 Changed = true; 704 CI->eraseFromParent(); 705 } 706 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 707 // Should handle GEP here. 708 SmallVector<Constant*, 8> Idxs; 709 Idxs.reserve(GEPI->getNumOperands()-1); 710 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 711 i != e; ++i) 712 if (Constant *C = dyn_cast<Constant>(*i)) 713 Idxs.push_back(C); 714 else 715 break; 716 if (Idxs.size() == GEPI->getNumOperands()-1) 717 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 718 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 719 Idxs.size())); 720 if (GEPI->use_empty()) { 721 Changed = true; 722 GEPI->eraseFromParent(); 723 } 724 } 725 } 726 727 return Changed; 728} 729 730 731/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 732/// value stored into it. If there are uses of the loaded value that would trap 733/// if the loaded value is dynamically null, then we know that they cannot be 734/// reachable with a null optimize away the load. 735static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 736 std::vector<LoadInst*> Loads; 737 bool Changed = false; 738 739 // Replace all uses of loads with uses of uses of the stored value. 740 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); 741 GUI != E; ++GUI) 742 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) { 743 Loads.push_back(LI); 744 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 745 } else { 746 // If we get here we could have stores, selects, or phi nodes whose values 747 // are loaded. 748 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) || 749 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) && 750 "Only expect load and stores!"); 751 } 752 753 if (Changed) { 754 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 755 ++NumGlobUses; 756 } 757 758 // Delete all of the loads we can, keeping track of whether we nuked them all! 759 bool AllLoadsGone = true; 760 while (!Loads.empty()) { 761 LoadInst *L = Loads.back(); 762 if (L->use_empty()) { 763 L->eraseFromParent(); 764 Changed = true; 765 } else { 766 AllLoadsGone = false; 767 } 768 Loads.pop_back(); 769 } 770 771 // If we nuked all of the loads, then none of the stores are needed either, 772 // nor is the global. 773 if (AllLoadsGone) { 774 DOUT << " *** GLOBAL NOW DEAD!\n"; 775 CleanupConstantGlobalUsers(GV, 0); 776 if (GV->use_empty()) { 777 GV->eraseFromParent(); 778 ++NumDeleted; 779 } 780 Changed = true; 781 } 782 return Changed; 783} 784 785/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 786/// instructions that are foldable. 787static void ConstantPropUsersOf(Value *V) { 788 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 789 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 790 if (Constant *NewC = ConstantFoldInstruction(I)) { 791 I->replaceAllUsesWith(NewC); 792 793 // Advance UI to the next non-I use to avoid invalidating it! 794 // Instructions could multiply use V. 795 while (UI != E && *UI == I) 796 ++UI; 797 I->eraseFromParent(); 798 } 799} 800 801/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 802/// variable, and transforms the program as if it always contained the result of 803/// the specified malloc. Because it is always the result of the specified 804/// malloc, there is no reason to actually DO the malloc. Instead, turn the 805/// malloc into a global, and any loads of GV as uses of the new global. 806static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 807 MallocInst *MI) { 808 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 809 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 810 811 if (NElements->getZExtValue() != 1) { 812 // If we have an array allocation, transform it to a single element 813 // allocation to make the code below simpler. 814 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 815 NElements->getZExtValue()); 816 MallocInst *NewMI = 817 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 818 MI->getAlignment(), MI->getName(), MI); 819 Value* Indices[2]; 820 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 821 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2, 822 NewMI->getName()+".el0", MI); 823 MI->replaceAllUsesWith(NewGEP); 824 MI->eraseFromParent(); 825 MI = NewMI; 826 } 827 828 // Create the new global variable. The contents of the malloc'd memory is 829 // undefined, so initialize with an undef value. 830 Constant *Init = UndefValue::get(MI->getAllocatedType()); 831 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 832 GlobalValue::InternalLinkage, Init, 833 GV->getName()+".body", 834 (Module *)NULL, 835 GV->isThreadLocal()); 836 // FIXME: This new global should have the alignment returned by malloc. Code 837 // could depend on malloc returning large alignment (on the mac, 16 bytes) but 838 // this would only guarantee some lower alignment. 839 GV->getParent()->getGlobalList().insert(GV, NewGV); 840 841 // Anything that used the malloc now uses the global directly. 842 MI->replaceAllUsesWith(NewGV); 843 844 Constant *RepValue = NewGV; 845 if (NewGV->getType() != GV->getType()->getElementType()) 846 RepValue = ConstantExpr::getBitCast(RepValue, 847 GV->getType()->getElementType()); 848 849 // If there is a comparison against null, we will insert a global bool to 850 // keep track of whether the global was initialized yet or not. 851 GlobalVariable *InitBool = 852 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 853 ConstantInt::getFalse(), GV->getName()+".init", 854 (Module *)NULL, GV->isThreadLocal()); 855 bool InitBoolUsed = false; 856 857 // Loop over all uses of GV, processing them in turn. 858 std::vector<StoreInst*> Stores; 859 while (!GV->use_empty()) 860 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 861 while (!LI->use_empty()) { 862 Use &LoadUse = LI->use_begin().getUse(); 863 if (!isa<ICmpInst>(LoadUse.getUser())) 864 LoadUse = RepValue; 865 else { 866 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 867 // Replace the cmp X, 0 with a use of the bool value. 868 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 869 InitBoolUsed = true; 870 switch (CI->getPredicate()) { 871 default: assert(0 && "Unknown ICmp Predicate!"); 872 case ICmpInst::ICMP_ULT: 873 case ICmpInst::ICMP_SLT: 874 LV = ConstantInt::getFalse(); // X < null -> always false 875 break; 876 case ICmpInst::ICMP_ULE: 877 case ICmpInst::ICMP_SLE: 878 case ICmpInst::ICMP_EQ: 879 LV = BinaryOperator::CreateNot(LV, "notinit", CI); 880 break; 881 case ICmpInst::ICMP_NE: 882 case ICmpInst::ICMP_UGE: 883 case ICmpInst::ICMP_SGE: 884 case ICmpInst::ICMP_UGT: 885 case ICmpInst::ICMP_SGT: 886 break; // no change. 887 } 888 CI->replaceAllUsesWith(LV); 889 CI->eraseFromParent(); 890 } 891 } 892 LI->eraseFromParent(); 893 } else { 894 StoreInst *SI = cast<StoreInst>(GV->use_back()); 895 // The global is initialized when the store to it occurs. 896 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 897 SI->eraseFromParent(); 898 } 899 900 // If the initialization boolean was used, insert it, otherwise delete it. 901 if (!InitBoolUsed) { 902 while (!InitBool->use_empty()) // Delete initializations 903 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 904 delete InitBool; 905 } else 906 GV->getParent()->getGlobalList().insert(GV, InitBool); 907 908 909 // Now the GV is dead, nuke it and the malloc. 910 GV->eraseFromParent(); 911 MI->eraseFromParent(); 912 913 // To further other optimizations, loop over all users of NewGV and try to 914 // constant prop them. This will promote GEP instructions with constant 915 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 916 ConstantPropUsersOf(NewGV); 917 if (RepValue != NewGV) 918 ConstantPropUsersOf(RepValue); 919 920 return NewGV; 921} 922 923/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 924/// to make sure that there are no complex uses of V. We permit simple things 925/// like dereferencing the pointer, but not storing through the address, unless 926/// it is to the specified global. 927static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 928 GlobalVariable *GV, 929 SmallPtrSet<PHINode*, 8> &PHIs) { 930 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 931 Instruction *Inst = dyn_cast<Instruction>(*UI); 932 if (Inst == 0) return false; 933 934 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { 935 continue; // Fine, ignore. 936 } 937 938 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 939 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 940 return false; // Storing the pointer itself... bad. 941 continue; // Otherwise, storing through it, or storing into GV... fine. 942 } 943 944 if (isa<GetElementPtrInst>(Inst)) { 945 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs)) 946 return false; 947 continue; 948 } 949 950 if (PHINode *PN = dyn_cast<PHINode>(Inst)) { 951 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 952 // cycles. 953 if (PHIs.insert(PN)) 954 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 955 return false; 956 continue; 957 } 958 959 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { 960 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) 961 return false; 962 continue; 963 } 964 965 return false; 966 } 967 return true; 968} 969 970/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 971/// somewhere. Transform all uses of the allocation into loads from the 972/// global and uses of the resultant pointer. Further, delete the store into 973/// GV. This assumes that these value pass the 974/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 975static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 976 GlobalVariable *GV) { 977 while (!Alloc->use_empty()) { 978 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 979 Instruction *InsertPt = U; 980 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 981 // If this is the store of the allocation into the global, remove it. 982 if (SI->getOperand(1) == GV) { 983 SI->eraseFromParent(); 984 continue; 985 } 986 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 987 // Insert the load in the corresponding predecessor, not right before the 988 // PHI. 989 unsigned PredNo = Alloc->use_begin().getOperandNo()/2; 990 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator(); 991 } 992 993 // Insert a load from the global, and use it instead of the malloc. 994 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 995 U->replaceUsesOfWith(Alloc, NL); 996 } 997} 998 999/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1000/// GV are simple enough to perform HeapSRA, return true. 1001static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 1002 MallocInst *MI) { 1003 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 1004 ++UI) 1005 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 1006 // We permit two users of the load: setcc comparing against the null 1007 // pointer, and a getelementptr of a specific form. 1008 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); 1009 UI != E; ++UI) { 1010 // Comparison against null is ok. 1011 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) { 1012 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 1013 return false; 1014 continue; 1015 } 1016 1017 // getelementptr is also ok, but only a simple form. 1018 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 1019 // Must index into the array and into the struct. 1020 if (GEPI->getNumOperands() < 3) 1021 return false; 1022 1023 // Otherwise the GEP is ok. 1024 continue; 1025 } 1026 1027 if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 1028 // We have a phi of a load from the global. We can only handle this 1029 // if the other PHI'd values are actually the same. In this case, 1030 // the rewriter will just drop the phi entirely. 1031 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1032 Value *IV = PN->getIncomingValue(i); 1033 if (IV == LI) continue; // Trivial the same. 1034 1035 // If the phi'd value is from the malloc that initializes the value, 1036 // we can xform it. 1037 if (IV == MI) continue; 1038 1039 // Otherwise, we don't know what it is. 1040 return false; 1041 } 1042 return true; 1043 } 1044 1045 // Otherwise we don't know what this is, not ok. 1046 return false; 1047 } 1048 } 1049 return true; 1050} 1051 1052/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd 1053/// value, lazily creating it on demand. 1054static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo, 1055 const std::vector<GlobalVariable*> &FieldGlobals, 1056 std::vector<Value *> &InsertedLoadsForPtr) { 1057 if (InsertedLoadsForPtr.size() <= FieldNo) 1058 InsertedLoadsForPtr.resize(FieldNo+1); 1059 if (InsertedLoadsForPtr[FieldNo] == 0) 1060 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo], 1061 Load->getName()+".f" + 1062 utostr(FieldNo), Load); 1063 return InsertedLoadsForPtr[FieldNo]; 1064} 1065 1066/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1067/// the load, rewrite the derived value to use the HeapSRoA'd load. 1068static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser, 1069 const std::vector<GlobalVariable*> &FieldGlobals, 1070 std::vector<Value *> &InsertedLoadsForPtr) { 1071 // If this is a comparison against null, handle it. 1072 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1073 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1074 // If we have a setcc of the loaded pointer, we can use a setcc of any 1075 // field. 1076 Value *NPtr; 1077 if (InsertedLoadsForPtr.empty()) { 1078 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr); 1079 } else { 1080 NPtr = InsertedLoadsForPtr.back(); 1081 } 1082 1083 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1084 Constant::getNullValue(NPtr->getType()), 1085 SCI->getName(), SCI); 1086 SCI->replaceAllUsesWith(New); 1087 SCI->eraseFromParent(); 1088 return; 1089 } 1090 1091 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...' 1092 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1093 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1094 && "Unexpected GEPI!"); 1095 1096 // Load the pointer for this field. 1097 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1098 Value *NewPtr = GetHeapSROALoad(Load, FieldNo, 1099 FieldGlobals, InsertedLoadsForPtr); 1100 1101 // Create the new GEP idx vector. 1102 SmallVector<Value*, 8> GEPIdx; 1103 GEPIdx.push_back(GEPI->getOperand(1)); 1104 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1105 1106 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1107 GEPIdx.begin(), GEPIdx.end(), 1108 GEPI->getName(), GEPI); 1109 GEPI->replaceAllUsesWith(NGEPI); 1110 GEPI->eraseFromParent(); 1111 return; 1112 } 1113 1114 // Handle PHI nodes. PHI nodes must be merging in the same values, plus 1115 // potentially the original malloc. Insert phi nodes for each field, then 1116 // process uses of the PHI. 1117 PHINode *PN = cast<PHINode>(LoadUser); 1118 std::vector<Value *> PHIsForField; 1119 PHIsForField.resize(FieldGlobals.size()); 1120 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1121 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr); 1122 1123 PHINode *FieldPN = PHINode::Create(LoadV->getType(), 1124 PN->getName()+"."+utostr(i), PN); 1125 // Fill in the predecessor values. 1126 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) { 1127 // Each predecessor either uses the load or the original malloc. 1128 Value *InVal = PN->getIncomingValue(pred); 1129 BasicBlock *BB = PN->getIncomingBlock(pred); 1130 Value *NewVal; 1131 if (isa<MallocInst>(InVal)) { 1132 // Insert a reload from the global in the predecessor. 1133 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals, 1134 PHIsForField); 1135 } else { 1136 NewVal = InsertedLoadsForPtr[i]; 1137 } 1138 FieldPN->addIncoming(NewVal, BB); 1139 } 1140 PHIsForField[i] = FieldPN; 1141 } 1142 1143 // Since PHIsForField specifies a phi for every input value, the lazy inserter 1144 // will never insert a load. 1145 while (!PN->use_empty()) 1146 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField); 1147 PN->eraseFromParent(); 1148} 1149 1150/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1151/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1152/// use FieldGlobals instead. All uses of loaded values satisfy 1153/// GlobalLoadUsesSimpleEnoughForHeapSRA. 1154static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1155 const std::vector<GlobalVariable*> &FieldGlobals) { 1156 std::vector<Value *> InsertedLoadsForPtr; 1157 //InsertedLoadsForPtr.resize(FieldGlobals.size()); 1158 while (!Load->use_empty()) 1159 RewriteHeapSROALoadUser(Load, Load->use_back(), 1160 FieldGlobals, InsertedLoadsForPtr); 1161} 1162 1163/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1164/// it up into multiple allocations of arrays of the fields. 1165static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1166 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1167 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1168 1169 // There is guaranteed to be at least one use of the malloc (storing 1170 // it into GV). If there are other uses, change them to be uses of 1171 // the global to simplify later code. This also deletes the store 1172 // into GV. 1173 ReplaceUsesOfMallocWithGlobal(MI, GV); 1174 1175 // Okay, at this point, there are no users of the malloc. Insert N 1176 // new mallocs at the same place as MI, and N globals. 1177 std::vector<GlobalVariable*> FieldGlobals; 1178 std::vector<MallocInst*> FieldMallocs; 1179 1180 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1181 const Type *FieldTy = STy->getElementType(FieldNo); 1182 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1183 1184 GlobalVariable *NGV = 1185 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1186 Constant::getNullValue(PFieldTy), 1187 GV->getName() + ".f" + utostr(FieldNo), GV, 1188 GV->isThreadLocal()); 1189 FieldGlobals.push_back(NGV); 1190 1191 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1192 MI->getName() + ".f" + utostr(FieldNo),MI); 1193 FieldMallocs.push_back(NMI); 1194 new StoreInst(NMI, NGV, MI); 1195 } 1196 1197 // The tricky aspect of this transformation is handling the case when malloc 1198 // fails. In the original code, malloc failing would set the result pointer 1199 // of malloc to null. In this case, some mallocs could succeed and others 1200 // could fail. As such, we emit code that looks like this: 1201 // F0 = malloc(field0) 1202 // F1 = malloc(field1) 1203 // F2 = malloc(field2) 1204 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1205 // if (F0) { free(F0); F0 = 0; } 1206 // if (F1) { free(F1); F1 = 0; } 1207 // if (F2) { free(F2); F2 = 0; } 1208 // } 1209 Value *RunningOr = 0; 1210 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1211 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1212 Constant::getNullValue(FieldMallocs[i]->getType()), 1213 "isnull", MI); 1214 if (!RunningOr) 1215 RunningOr = Cond; // First seteq 1216 else 1217 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI); 1218 } 1219 1220 // Split the basic block at the old malloc. 1221 BasicBlock *OrigBB = MI->getParent(); 1222 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1223 1224 // Create the block to check the first condition. Put all these blocks at the 1225 // end of the function as they are unlikely to be executed. 1226 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1227 OrigBB->getParent()); 1228 1229 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1230 // branch on RunningOr. 1231 OrigBB->getTerminator()->eraseFromParent(); 1232 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1233 1234 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1235 // pointer, because some may be null while others are not. 1236 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1237 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1238 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1239 Constant::getNullValue(GVVal->getType()), 1240 "tmp", NullPtrBlock); 1241 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1242 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1243 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1244 1245 // Fill in FreeBlock. 1246 new FreeInst(GVVal, FreeBlock); 1247 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1248 FreeBlock); 1249 BranchInst::Create(NextBlock, FreeBlock); 1250 1251 NullPtrBlock = NextBlock; 1252 } 1253 1254 BranchInst::Create(ContBB, NullPtrBlock); 1255 1256 // MI is no longer needed, remove it. 1257 MI->eraseFromParent(); 1258 1259 1260 // Okay, the malloc site is completely handled. All of the uses of GV are now 1261 // loads, and all uses of those loads are simple. Rewrite them to use loads 1262 // of the per-field globals instead. 1263 while (!GV->use_empty()) { 1264 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 1265 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals); 1266 LI->eraseFromParent(); 1267 } else { 1268 // Must be a store of null. 1269 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1270 assert(isa<Constant>(SI->getOperand(0)) && 1271 cast<Constant>(SI->getOperand(0))->isNullValue() && 1272 "Unexpected heap-sra user!"); 1273 1274 // Insert a store of null into each global. 1275 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1276 Constant *Null = 1277 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType()); 1278 new StoreInst(Null, FieldGlobals[i], SI); 1279 } 1280 // Erase the original store. 1281 SI->eraseFromParent(); 1282 } 1283 } 1284 1285 // The old global is now dead, remove it. 1286 GV->eraseFromParent(); 1287 1288 ++NumHeapSRA; 1289 return FieldGlobals[0]; 1290} 1291 1292/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1293/// pointer global variable with a single value stored it that is a malloc or 1294/// cast of malloc. 1295static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, 1296 MallocInst *MI, 1297 Module::global_iterator &GVI, 1298 TargetData &TD) { 1299 // If this is a malloc of an abstract type, don't touch it. 1300 if (!MI->getAllocatedType()->isSized()) 1301 return false; 1302 1303 // We can't optimize this global unless all uses of it are *known* to be 1304 // of the malloc value, not of the null initializer value (consider a use 1305 // that compares the global's value against zero to see if the malloc has 1306 // been reached). To do this, we check to see if all uses of the global 1307 // would trap if the global were null: this proves that they must all 1308 // happen after the malloc. 1309 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1310 return false; 1311 1312 // We can't optimize this if the malloc itself is used in a complex way, 1313 // for example, being stored into multiple globals. This allows the 1314 // malloc to be stored into the specified global, loaded setcc'd, and 1315 // GEP'd. These are all things we could transform to using the global 1316 // for. 1317 { 1318 SmallPtrSet<PHINode*, 8> PHIs; 1319 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1320 return false; 1321 } 1322 1323 1324 // If we have a global that is only initialized with a fixed size malloc, 1325 // transform the program to use global memory instead of malloc'd memory. 1326 // This eliminates dynamic allocation, avoids an indirection accessing the 1327 // data, and exposes the resultant global to further GlobalOpt. 1328 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1329 // Restrict this transformation to only working on small allocations 1330 // (2048 bytes currently), as we don't want to introduce a 16M global or 1331 // something. 1332 if (NElements->getZExtValue()* 1333 TD.getABITypeSize(MI->getAllocatedType()) < 2048) { 1334 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1335 return true; 1336 } 1337 } 1338 1339 // If the allocation is an array of structures, consider transforming this 1340 // into multiple malloc'd arrays, one for each field. This is basically 1341 // SRoA for malloc'd memory. 1342 if (const StructType *AllocTy = 1343 dyn_cast<StructType>(MI->getAllocatedType())) { 1344 // This the structure has an unreasonable number of fields, leave it 1345 // alone. 1346 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 && 1347 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1348 GVI = PerformHeapAllocSRoA(GV, MI); 1349 return true; 1350 } 1351 } 1352 1353 return false; 1354} 1355 1356// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1357// that only one value (besides its initializer) is ever stored to the global. 1358static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1359 Module::global_iterator &GVI, 1360 TargetData &TD) { 1361 // Ignore no-op GEPs and bitcasts. 1362 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1363 1364 // If we are dealing with a pointer global that is initialized to null and 1365 // only has one (non-null) value stored into it, then we can optimize any 1366 // users of the loaded value (often calls and loads) that would trap if the 1367 // value was null. 1368 if (isa<PointerType>(GV->getInitializer()->getType()) && 1369 GV->getInitializer()->isNullValue()) { 1370 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1371 if (GV->getInitializer()->getType() != SOVC->getType()) 1372 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1373 1374 // Optimize away any trapping uses of the loaded value. 1375 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1376 return true; 1377 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1378 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD)) 1379 return true; 1380 } 1381 } 1382 1383 return false; 1384} 1385 1386/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1387/// two values ever stored into GV are its initializer and OtherVal. See if we 1388/// can shrink the global into a boolean and select between the two values 1389/// whenever it is used. This exposes the values to other scalar optimizations. 1390static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1391 const Type *GVElType = GV->getType()->getElementType(); 1392 1393 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1394 // an FP value or vector, don't do this optimization because a select between 1395 // them is very expensive and unlikely to lead to later simplification. 1396 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1397 isa<VectorType>(GVElType)) 1398 return false; 1399 1400 // Walk the use list of the global seeing if all the uses are load or store. 1401 // If there is anything else, bail out. 1402 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1403 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1404 return false; 1405 1406 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1407 1408 // Create the new global, initializing it to false. 1409 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1410 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1411 GV->getName()+".b", 1412 (Module *)NULL, 1413 GV->isThreadLocal()); 1414 GV->getParent()->getGlobalList().insert(GV, NewGV); 1415 1416 Constant *InitVal = GV->getInitializer(); 1417 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1418 1419 // If initialized to zero and storing one into the global, we can use a cast 1420 // instead of a select to synthesize the desired value. 1421 bool IsOneZero = false; 1422 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1423 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1424 1425 while (!GV->use_empty()) { 1426 Instruction *UI = cast<Instruction>(GV->use_back()); 1427 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1428 // Change the store into a boolean store. 1429 bool StoringOther = SI->getOperand(0) == OtherVal; 1430 // Only do this if we weren't storing a loaded value. 1431 Value *StoreVal; 1432 if (StoringOther || SI->getOperand(0) == InitVal) 1433 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1434 else { 1435 // Otherwise, we are storing a previously loaded copy. To do this, 1436 // change the copy from copying the original value to just copying the 1437 // bool. 1438 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1439 1440 // If we're already replaced the input, StoredVal will be a cast or 1441 // select instruction. If not, it will be a load of the original 1442 // global. 1443 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1444 assert(LI->getOperand(0) == GV && "Not a copy!"); 1445 // Insert a new load, to preserve the saved value. 1446 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1447 } else { 1448 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1449 "This is not a form that we understand!"); 1450 StoreVal = StoredVal->getOperand(0); 1451 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1452 } 1453 } 1454 new StoreInst(StoreVal, NewGV, SI); 1455 } else { 1456 // Change the load into a load of bool then a select. 1457 LoadInst *LI = cast<LoadInst>(UI); 1458 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1459 Value *NSI; 1460 if (IsOneZero) 1461 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1462 else 1463 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1464 NSI->takeName(LI); 1465 LI->replaceAllUsesWith(NSI); 1466 } 1467 UI->eraseFromParent(); 1468 } 1469 1470 GV->eraseFromParent(); 1471 return true; 1472} 1473 1474 1475/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1476/// it if possible. If we make a change, return true. 1477bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1478 Module::global_iterator &GVI) { 1479 std::set<PHINode*> PHIUsers; 1480 GlobalStatus GS; 1481 GV->removeDeadConstantUsers(); 1482 1483 if (GV->use_empty()) { 1484 DOUT << "GLOBAL DEAD: " << *GV; 1485 GV->eraseFromParent(); 1486 ++NumDeleted; 1487 return true; 1488 } 1489 1490 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1491#if 0 1492 cerr << "Global: " << *GV; 1493 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1494 cerr << " StoredType = "; 1495 switch (GS.StoredType) { 1496 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1497 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1498 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1499 case GlobalStatus::isStored: cerr << "stored\n"; break; 1500 } 1501 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1502 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1503 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1504 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1505 << "\n"; 1506 cerr << " HasMultipleAccessingFunctions = " 1507 << GS.HasMultipleAccessingFunctions << "\n"; 1508 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1509 cerr << "\n"; 1510#endif 1511 1512 // If this is a first class global and has only one accessing function 1513 // and this function is main (which we know is not recursive we can make 1514 // this global a local variable) we replace the global with a local alloca 1515 // in this function. 1516 // 1517 // NOTE: It doesn't make sense to promote non single-value types since we 1518 // are just replacing static memory to stack memory. 1519 if (!GS.HasMultipleAccessingFunctions && 1520 GS.AccessingFunction && !GS.HasNonInstructionUser && 1521 GV->getType()->getElementType()->isSingleValueType() && 1522 GS.AccessingFunction->getName() == "main" && 1523 GS.AccessingFunction->hasExternalLinkage()) { 1524 DOUT << "LOCALIZING GLOBAL: " << *GV; 1525 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1526 const Type* ElemTy = GV->getType()->getElementType(); 1527 // FIXME: Pass Global's alignment when globals have alignment 1528 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1529 if (!isa<UndefValue>(GV->getInitializer())) 1530 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1531 1532 GV->replaceAllUsesWith(Alloca); 1533 GV->eraseFromParent(); 1534 ++NumLocalized; 1535 return true; 1536 } 1537 1538 // If the global is never loaded (but may be stored to), it is dead. 1539 // Delete it now. 1540 if (!GS.isLoaded) { 1541 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1542 1543 // Delete any stores we can find to the global. We may not be able to 1544 // make it completely dead though. 1545 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1546 1547 // If the global is dead now, delete it. 1548 if (GV->use_empty()) { 1549 GV->eraseFromParent(); 1550 ++NumDeleted; 1551 Changed = true; 1552 } 1553 return Changed; 1554 1555 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1556 DOUT << "MARKING CONSTANT: " << *GV; 1557 GV->setConstant(true); 1558 1559 // Clean up any obviously simplifiable users now. 1560 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1561 1562 // If the global is dead now, just nuke it. 1563 if (GV->use_empty()) { 1564 DOUT << " *** Marking constant allowed us to simplify " 1565 << "all users and delete global!\n"; 1566 GV->eraseFromParent(); 1567 ++NumDeleted; 1568 } 1569 1570 ++NumMarked; 1571 return true; 1572 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1573 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1574 getAnalysis<TargetData>())) { 1575 GVI = FirstNewGV; // Don't skip the newly produced globals! 1576 return true; 1577 } 1578 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1579 // If the initial value for the global was an undef value, and if only 1580 // one other value was stored into it, we can just change the 1581 // initializer to be an undef value, then delete all stores to the 1582 // global. This allows us to mark it constant. 1583 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1584 if (isa<UndefValue>(GV->getInitializer())) { 1585 // Change the initial value here. 1586 GV->setInitializer(SOVConstant); 1587 1588 // Clean up any obviously simplifiable users now. 1589 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1590 1591 if (GV->use_empty()) { 1592 DOUT << " *** Substituting initializer allowed us to " 1593 << "simplify all users and delete global!\n"; 1594 GV->eraseFromParent(); 1595 ++NumDeleted; 1596 } else { 1597 GVI = GV; 1598 } 1599 ++NumSubstitute; 1600 return true; 1601 } 1602 1603 // Try to optimize globals based on the knowledge that only one value 1604 // (besides its initializer) is ever stored to the global. 1605 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1606 getAnalysis<TargetData>())) 1607 return true; 1608 1609 // Otherwise, if the global was not a boolean, we can shrink it to be a 1610 // boolean. 1611 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1612 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1613 ++NumShrunkToBool; 1614 return true; 1615 } 1616 } 1617 } 1618 return false; 1619} 1620 1621/// OnlyCalledDirectly - Return true if the specified function is only called 1622/// directly. In other words, its address is never taken. 1623static bool OnlyCalledDirectly(Function *F) { 1624 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1625 Instruction *User = dyn_cast<Instruction>(*UI); 1626 if (!User) return false; 1627 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1628 1629 // See if the function address is passed as an argument. 1630 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end(); 1631 i != e; ++i) 1632 if (*i == F) return false; 1633 } 1634 return true; 1635} 1636 1637/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1638/// function, changing them to FastCC. 1639static void ChangeCalleesToFastCall(Function *F) { 1640 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1641 CallSite User(cast<Instruction>(*UI)); 1642 User.setCallingConv(CallingConv::Fast); 1643 } 1644} 1645 1646static AttrListPtr StripNest(const AttrListPtr &Attrs) { 1647 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1648 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) 1649 continue; 1650 1651 // There can be only one. 1652 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); 1653 } 1654 1655 return Attrs; 1656} 1657 1658static void RemoveNestAttribute(Function *F) { 1659 F->setAttributes(StripNest(F->getAttributes())); 1660 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1661 CallSite User(cast<Instruction>(*UI)); 1662 User.setAttributes(StripNest(User.getAttributes())); 1663 } 1664} 1665 1666bool GlobalOpt::OptimizeFunctions(Module &M) { 1667 bool Changed = false; 1668 // Optimize functions. 1669 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1670 Function *F = FI++; 1671 F->removeDeadConstantUsers(); 1672 if (F->use_empty() && (F->hasInternalLinkage() || 1673 F->hasLinkOnceLinkage())) { 1674 M.getFunctionList().erase(F); 1675 Changed = true; 1676 ++NumFnDeleted; 1677 } else if (F->hasInternalLinkage()) { 1678 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1679 OnlyCalledDirectly(F)) { 1680 // If this function has C calling conventions, is not a varargs 1681 // function, and is only called directly, promote it to use the Fast 1682 // calling convention. 1683 F->setCallingConv(CallingConv::Fast); 1684 ChangeCalleesToFastCall(F); 1685 ++NumFastCallFns; 1686 Changed = true; 1687 } 1688 1689 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1690 OnlyCalledDirectly(F)) { 1691 // The function is not used by a trampoline intrinsic, so it is safe 1692 // to remove the 'nest' attribute. 1693 RemoveNestAttribute(F); 1694 ++NumNestRemoved; 1695 Changed = true; 1696 } 1697 } 1698 } 1699 return Changed; 1700} 1701 1702bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1703 bool Changed = false; 1704 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1705 GVI != E; ) { 1706 GlobalVariable *GV = GVI++; 1707 if (!GV->isConstant() && GV->hasInternalLinkage() && 1708 GV->hasInitializer()) 1709 Changed |= ProcessInternalGlobal(GV, GVI); 1710 } 1711 return Changed; 1712} 1713 1714/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1715/// initializers have an init priority of 65535. 1716GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1717 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1718 I != E; ++I) 1719 if (I->getName() == "llvm.global_ctors") { 1720 // Found it, verify it's an array of { int, void()* }. 1721 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1722 if (!ATy) return 0; 1723 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1724 if (!STy || STy->getNumElements() != 2 || 1725 STy->getElementType(0) != Type::Int32Ty) return 0; 1726 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1727 if (!PFTy) return 0; 1728 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1729 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1730 FTy->getNumParams() != 0) 1731 return 0; 1732 1733 // Verify that the initializer is simple enough for us to handle. 1734 if (!I->hasInitializer()) return 0; 1735 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1736 if (!CA) return 0; 1737 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1738 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1739 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1740 continue; 1741 1742 // Must have a function or null ptr. 1743 if (!isa<Function>(CS->getOperand(1))) 1744 return 0; 1745 1746 // Init priority must be standard. 1747 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1748 if (!CI || CI->getZExtValue() != 65535) 1749 return 0; 1750 } else { 1751 return 0; 1752 } 1753 1754 return I; 1755 } 1756 return 0; 1757} 1758 1759/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1760/// return a list of the functions and null terminator as a vector. 1761static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1762 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1763 std::vector<Function*> Result; 1764 Result.reserve(CA->getNumOperands()); 1765 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1766 ConstantStruct *CS = cast<ConstantStruct>(*i); 1767 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1768 } 1769 return Result; 1770} 1771 1772/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1773/// specified array, returning the new global to use. 1774static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1775 const std::vector<Function*> &Ctors) { 1776 // If we made a change, reassemble the initializer list. 1777 std::vector<Constant*> CSVals; 1778 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1779 CSVals.push_back(0); 1780 1781 // Create the new init list. 1782 std::vector<Constant*> CAList; 1783 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1784 if (Ctors[i]) { 1785 CSVals[1] = Ctors[i]; 1786 } else { 1787 const Type *FTy = FunctionType::get(Type::VoidTy, 1788 std::vector<const Type*>(), false); 1789 const PointerType *PFTy = PointerType::getUnqual(FTy); 1790 CSVals[1] = Constant::getNullValue(PFTy); 1791 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1792 } 1793 CAList.push_back(ConstantStruct::get(CSVals)); 1794 } 1795 1796 // Create the array initializer. 1797 const Type *StructTy = 1798 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1799 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1800 CAList); 1801 1802 // If we didn't change the number of elements, don't create a new GV. 1803 if (CA->getType() == GCL->getInitializer()->getType()) { 1804 GCL->setInitializer(CA); 1805 return GCL; 1806 } 1807 1808 // Create the new global and insert it next to the existing list. 1809 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1810 GCL->getLinkage(), CA, "", 1811 (Module *)NULL, 1812 GCL->isThreadLocal()); 1813 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1814 NGV->takeName(GCL); 1815 1816 // Nuke the old list, replacing any uses with the new one. 1817 if (!GCL->use_empty()) { 1818 Constant *V = NGV; 1819 if (V->getType() != GCL->getType()) 1820 V = ConstantExpr::getBitCast(V, GCL->getType()); 1821 GCL->replaceAllUsesWith(V); 1822 } 1823 GCL->eraseFromParent(); 1824 1825 if (Ctors.size()) 1826 return NGV; 1827 else 1828 return 0; 1829} 1830 1831 1832static Constant *getVal(std::map<Value*, Constant*> &ComputedValues, 1833 Value *V) { 1834 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1835 Constant *R = ComputedValues[V]; 1836 assert(R && "Reference to an uncomputed value!"); 1837 return R; 1838} 1839 1840/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1841/// enough for us to understand. In particular, if it is a cast of something, 1842/// we punt. We basically just support direct accesses to globals and GEP's of 1843/// globals. This should be kept up to date with CommitValueTo. 1844static bool isSimpleEnoughPointerToCommit(Constant *C) { 1845 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1846 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1847 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1848 return !GV->isDeclaration(); // reject external globals. 1849 } 1850 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 1851 // Handle a constantexpr gep. 1852 if (CE->getOpcode() == Instruction::GetElementPtr && 1853 isa<GlobalVariable>(CE->getOperand(0))) { 1854 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1855 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1856 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1857 return GV->hasInitializer() && 1858 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1859 } 1860 return false; 1861} 1862 1863/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 1864/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 1865/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 1866static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 1867 ConstantExpr *Addr, unsigned OpNo) { 1868 // Base case of the recursion. 1869 if (OpNo == Addr->getNumOperands()) { 1870 assert(Val->getType() == Init->getType() && "Type mismatch!"); 1871 return Val; 1872 } 1873 1874 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1875 std::vector<Constant*> Elts; 1876 1877 // Break up the constant into its elements. 1878 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1879 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 1880 Elts.push_back(cast<Constant>(*i)); 1881 } else if (isa<ConstantAggregateZero>(Init)) { 1882 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1883 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 1884 } else if (isa<UndefValue>(Init)) { 1885 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1886 Elts.push_back(UndefValue::get(STy->getElementType(i))); 1887 } else { 1888 assert(0 && "This code is out of sync with " 1889 " ConstantFoldLoadThroughGEPConstantExpr"); 1890 } 1891 1892 // Replace the element that we are supposed to. 1893 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 1894 unsigned Idx = CU->getZExtValue(); 1895 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 1896 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 1897 1898 // Return the modified struct. 1899 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 1900 } else { 1901 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 1902 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 1903 1904 // Break up the array into elements. 1905 std::vector<Constant*> Elts; 1906 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1907 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1908 Elts.push_back(cast<Constant>(*i)); 1909 } else if (isa<ConstantAggregateZero>(Init)) { 1910 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 1911 Elts.assign(ATy->getNumElements(), Elt); 1912 } else if (isa<UndefValue>(Init)) { 1913 Constant *Elt = UndefValue::get(ATy->getElementType()); 1914 Elts.assign(ATy->getNumElements(), Elt); 1915 } else { 1916 assert(0 && "This code is out of sync with " 1917 " ConstantFoldLoadThroughGEPConstantExpr"); 1918 } 1919 1920 assert(CI->getZExtValue() < ATy->getNumElements()); 1921 Elts[CI->getZExtValue()] = 1922 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 1923 return ConstantArray::get(ATy, Elts); 1924 } 1925} 1926 1927/// CommitValueTo - We have decided that Addr (which satisfies the predicate 1928/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 1929static void CommitValueTo(Constant *Val, Constant *Addr) { 1930 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 1931 assert(GV->hasInitializer()); 1932 GV->setInitializer(Val); 1933 return; 1934 } 1935 1936 ConstantExpr *CE = cast<ConstantExpr>(Addr); 1937 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1938 1939 Constant *Init = GV->getInitializer(); 1940 Init = EvaluateStoreInto(Init, Val, CE, 2); 1941 GV->setInitializer(Init); 1942} 1943 1944/// ComputeLoadResult - Return the value that would be computed by a load from 1945/// P after the stores reflected by 'memory' have been performed. If we can't 1946/// decide, return null. 1947static Constant *ComputeLoadResult(Constant *P, 1948 const std::map<Constant*, Constant*> &Memory) { 1949 // If this memory location has been recently stored, use the stored value: it 1950 // is the most up-to-date. 1951 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P); 1952 if (I != Memory.end()) return I->second; 1953 1954 // Access it. 1955 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 1956 if (GV->hasInitializer()) 1957 return GV->getInitializer(); 1958 return 0; 1959 } 1960 1961 // Handle a constantexpr getelementptr. 1962 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 1963 if (CE->getOpcode() == Instruction::GetElementPtr && 1964 isa<GlobalVariable>(CE->getOperand(0))) { 1965 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1966 if (GV->hasInitializer()) 1967 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1968 } 1969 1970 return 0; // don't know how to evaluate. 1971} 1972 1973/// EvaluateFunction - Evaluate a call to function F, returning true if 1974/// successful, false if we can't evaluate it. ActualArgs contains the formal 1975/// arguments for the function. 1976static bool EvaluateFunction(Function *F, Constant *&RetVal, 1977 const std::vector<Constant*> &ActualArgs, 1978 std::vector<Function*> &CallStack, 1979 std::map<Constant*, Constant*> &MutatedMemory, 1980 std::vector<GlobalVariable*> &AllocaTmps) { 1981 // Check to see if this function is already executing (recursion). If so, 1982 // bail out. TODO: we might want to accept limited recursion. 1983 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 1984 return false; 1985 1986 CallStack.push_back(F); 1987 1988 /// Values - As we compute SSA register values, we store their contents here. 1989 std::map<Value*, Constant*> Values; 1990 1991 // Initialize arguments to the incoming values specified. 1992 unsigned ArgNo = 0; 1993 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1994 ++AI, ++ArgNo) 1995 Values[AI] = ActualArgs[ArgNo]; 1996 1997 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 1998 /// we can only evaluate any one basic block at most once. This set keeps 1999 /// track of what we have executed so we can detect recursive cases etc. 2000 std::set<BasicBlock*> ExecutedBlocks; 2001 2002 // CurInst - The current instruction we're evaluating. 2003 BasicBlock::iterator CurInst = F->begin()->begin(); 2004 2005 // This is the main evaluation loop. 2006 while (1) { 2007 Constant *InstResult = 0; 2008 2009 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2010 if (SI->isVolatile()) return false; // no volatile accesses. 2011 Constant *Ptr = getVal(Values, SI->getOperand(1)); 2012 if (!isSimpleEnoughPointerToCommit(Ptr)) 2013 // If this is too complex for us to commit, reject it. 2014 return false; 2015 Constant *Val = getVal(Values, SI->getOperand(0)); 2016 MutatedMemory[Ptr] = Val; 2017 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2018 InstResult = ConstantExpr::get(BO->getOpcode(), 2019 getVal(Values, BO->getOperand(0)), 2020 getVal(Values, BO->getOperand(1))); 2021 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2022 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2023 getVal(Values, CI->getOperand(0)), 2024 getVal(Values, CI->getOperand(1))); 2025 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2026 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2027 getVal(Values, CI->getOperand(0)), 2028 CI->getType()); 2029 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2030 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2031 getVal(Values, SI->getOperand(1)), 2032 getVal(Values, SI->getOperand(2))); 2033 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2034 Constant *P = getVal(Values, GEP->getOperand(0)); 2035 SmallVector<Constant*, 8> GEPOps; 2036 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2037 i != e; ++i) 2038 GEPOps.push_back(getVal(Values, *i)); 2039 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2040 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2041 if (LI->isVolatile()) return false; // no volatile accesses. 2042 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2043 MutatedMemory); 2044 if (InstResult == 0) return false; // Could not evaluate load. 2045 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2046 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2047 const Type *Ty = AI->getType()->getElementType(); 2048 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2049 GlobalValue::InternalLinkage, 2050 UndefValue::get(Ty), 2051 AI->getName())); 2052 InstResult = AllocaTmps.back(); 2053 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2054 // Cannot handle inline asm. 2055 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2056 2057 // Resolve function pointers. 2058 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2059 if (!Callee) return false; // Cannot resolve. 2060 2061 std::vector<Constant*> Formals; 2062 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end(); 2063 i != e; ++i) 2064 Formals.push_back(getVal(Values, *i)); 2065 2066 if (Callee->isDeclaration()) { 2067 // If this is a function we can constant fold, do it. 2068 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2069 Formals.size())) { 2070 InstResult = C; 2071 } else { 2072 return false; 2073 } 2074 } else { 2075 if (Callee->getFunctionType()->isVarArg()) 2076 return false; 2077 2078 Constant *RetVal; 2079 2080 // Execute the call, if successful, use the return value. 2081 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2082 MutatedMemory, AllocaTmps)) 2083 return false; 2084 InstResult = RetVal; 2085 } 2086 } else if (isa<TerminatorInst>(CurInst)) { 2087 BasicBlock *NewBB = 0; 2088 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2089 if (BI->isUnconditional()) { 2090 NewBB = BI->getSuccessor(0); 2091 } else { 2092 ConstantInt *Cond = 2093 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2094 if (!Cond) return false; // Cannot determine. 2095 2096 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2097 } 2098 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2099 ConstantInt *Val = 2100 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2101 if (!Val) return false; // Cannot determine. 2102 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2103 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2104 if (RI->getNumOperands()) 2105 RetVal = getVal(Values, RI->getOperand(0)); 2106 2107 CallStack.pop_back(); // return from fn. 2108 return true; // We succeeded at evaluating this ctor! 2109 } else { 2110 // invoke, unwind, unreachable. 2111 return false; // Cannot handle this terminator. 2112 } 2113 2114 // Okay, we succeeded in evaluating this control flow. See if we have 2115 // executed the new block before. If so, we have a looping function, 2116 // which we cannot evaluate in reasonable time. 2117 if (!ExecutedBlocks.insert(NewBB).second) 2118 return false; // looped! 2119 2120 // Okay, we have never been in this block before. Check to see if there 2121 // are any PHI nodes. If so, evaluate them with information about where 2122 // we came from. 2123 BasicBlock *OldBB = CurInst->getParent(); 2124 CurInst = NewBB->begin(); 2125 PHINode *PN; 2126 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2127 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2128 2129 // Do NOT increment CurInst. We know that the terminator had no value. 2130 continue; 2131 } else { 2132 // Did not know how to evaluate this! 2133 return false; 2134 } 2135 2136 if (!CurInst->use_empty()) 2137 Values[CurInst] = InstResult; 2138 2139 // Advance program counter. 2140 ++CurInst; 2141 } 2142} 2143 2144/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2145/// we can. Return true if we can, false otherwise. 2146static bool EvaluateStaticConstructor(Function *F) { 2147 /// MutatedMemory - For each store we execute, we update this map. Loads 2148 /// check this to get the most up-to-date value. If evaluation is successful, 2149 /// this state is committed to the process. 2150 std::map<Constant*, Constant*> MutatedMemory; 2151 2152 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2153 /// to represent its body. This vector is needed so we can delete the 2154 /// temporary globals when we are done. 2155 std::vector<GlobalVariable*> AllocaTmps; 2156 2157 /// CallStack - This is used to detect recursion. In pathological situations 2158 /// we could hit exponential behavior, but at least there is nothing 2159 /// unbounded. 2160 std::vector<Function*> CallStack; 2161 2162 // Call the function. 2163 Constant *RetValDummy; 2164 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2165 CallStack, MutatedMemory, AllocaTmps); 2166 if (EvalSuccess) { 2167 // We succeeded at evaluation: commit the result. 2168 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2169 << F->getName() << "' to " << MutatedMemory.size() 2170 << " stores.\n"; 2171 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2172 E = MutatedMemory.end(); I != E; ++I) 2173 CommitValueTo(I->second, I->first); 2174 } 2175 2176 // At this point, we are done interpreting. If we created any 'alloca' 2177 // temporaries, release them now. 2178 while (!AllocaTmps.empty()) { 2179 GlobalVariable *Tmp = AllocaTmps.back(); 2180 AllocaTmps.pop_back(); 2181 2182 // If there are still users of the alloca, the program is doing something 2183 // silly, e.g. storing the address of the alloca somewhere and using it 2184 // later. Since this is undefined, we'll just make it be null. 2185 if (!Tmp->use_empty()) 2186 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2187 delete Tmp; 2188 } 2189 2190 return EvalSuccess; 2191} 2192 2193 2194 2195/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2196/// Return true if anything changed. 2197bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2198 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2199 bool MadeChange = false; 2200 if (Ctors.empty()) return false; 2201 2202 // Loop over global ctors, optimizing them when we can. 2203 for (unsigned i = 0; i != Ctors.size(); ++i) { 2204 Function *F = Ctors[i]; 2205 // Found a null terminator in the middle of the list, prune off the rest of 2206 // the list. 2207 if (F == 0) { 2208 if (i != Ctors.size()-1) { 2209 Ctors.resize(i+1); 2210 MadeChange = true; 2211 } 2212 break; 2213 } 2214 2215 // We cannot simplify external ctor functions. 2216 if (F->empty()) continue; 2217 2218 // If we can evaluate the ctor at compile time, do. 2219 if (EvaluateStaticConstructor(F)) { 2220 Ctors.erase(Ctors.begin()+i); 2221 MadeChange = true; 2222 --i; 2223 ++NumCtorsEvaluated; 2224 continue; 2225 } 2226 } 2227 2228 if (!MadeChange) return false; 2229 2230 GCL = InstallGlobalCtors(GCL, Ctors); 2231 return true; 2232} 2233 2234bool GlobalOpt::ResolveAliases(Module &M) { 2235 bool Changed = false; 2236 2237 for (Module::alias_iterator I = M.alias_begin(), 2238 E = M.alias_end(); I != E; ++I) { 2239 if (I->use_empty()) 2240 continue; 2241 2242 if (const GlobalValue *GV = I->resolveAliasedGlobal()) 2243 if (GV != I) { 2244 I->replaceAllUsesWith(const_cast<GlobalValue*>(GV)); 2245 Changed = true; 2246 } 2247 } 2248 2249 return Changed; 2250} 2251 2252bool GlobalOpt::runOnModule(Module &M) { 2253 bool Changed = false; 2254 2255 // Try to find the llvm.globalctors list. 2256 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2257 2258 bool LocalChange = true; 2259 while (LocalChange) { 2260 LocalChange = false; 2261 2262 // Delete functions that are trivially dead, ccc -> fastcc 2263 LocalChange |= OptimizeFunctions(M); 2264 2265 // Optimize global_ctors list. 2266 if (GlobalCtors) 2267 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2268 2269 // Optimize non-address-taken globals. 2270 LocalChange |= OptimizeGlobalVars(M); 2271 2272 // Resolve aliases, when possible. 2273 LocalChange |= ResolveAliases(M); 2274 Changed |= LocalChange; 2275 } 2276 2277 // TODO: Move all global ctors functions to the end of the module for code 2278 // layout. 2279 2280 return Changed; 2281} 2282