GlobalOpt.cpp revision 089efffd7d1ca0d10522ace38d36e0a67f4fac2d
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((intptr_t)&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 OptimizeGlobalCtorsList(GlobalVariable *&GCL); 69 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 70 }; 71} 72 73char GlobalOpt::ID = 0; 74static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer"); 75 76ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 77 78namespace { 79 80/// GlobalStatus - As we analyze each global, keep track of some information 81/// about it. If we find out that the address of the global is taken, none of 82/// this info will be accurate. 83struct VISIBILITY_HIDDEN GlobalStatus { 84 /// isLoaded - True if the global is ever loaded. If the global isn't ever 85 /// loaded it can be deleted. 86 bool isLoaded; 87 88 /// StoredType - Keep track of what stores to the global look like. 89 /// 90 enum StoredType { 91 /// NotStored - There is no store to this global. It can thus be marked 92 /// constant. 93 NotStored, 94 95 /// isInitializerStored - This global is stored to, but the only thing 96 /// stored is the constant it was initialized with. This is only tracked 97 /// for scalar globals. 98 isInitializerStored, 99 100 /// isStoredOnce - This global is stored to, but only its initializer and 101 /// one other value is ever stored to it. If this global isStoredOnce, we 102 /// track the value stored to it in StoredOnceValue below. This is only 103 /// tracked for scalar globals. 104 isStoredOnce, 105 106 /// isStored - This global is stored to by multiple values or something else 107 /// that we cannot track. 108 isStored 109 } StoredType; 110 111 /// StoredOnceValue - If only one value (besides the initializer constant) is 112 /// ever stored to this global, keep track of what value it is. 113 Value *StoredOnceValue; 114 115 /// AccessingFunction/HasMultipleAccessingFunctions - These start out 116 /// null/false. When the first accessing function is noticed, it is recorded. 117 /// When a second different accessing function is noticed, 118 /// HasMultipleAccessingFunctions is set to true. 119 Function *AccessingFunction; 120 bool HasMultipleAccessingFunctions; 121 122 /// HasNonInstructionUser - Set to true if this global has a user that is not 123 /// an instruction (e.g. a constant expr or GV initializer). 124 bool HasNonInstructionUser; 125 126 /// HasPHIUser - Set to true if this global has a user that is a PHI node. 127 bool HasPHIUser; 128 129 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0), 130 AccessingFunction(0), HasMultipleAccessingFunctions(false), 131 HasNonInstructionUser(false), HasPHIUser(false) {} 132}; 133 134} 135 136/// ConstantIsDead - Return true if the specified constant is (transitively) 137/// dead. The constant may be used by other constants (e.g. constant arrays and 138/// constant exprs) as long as they are dead, but it cannot be used by anything 139/// else. 140static bool ConstantIsDead(Constant *C) { 141 if (isa<GlobalValue>(C)) return false; 142 143 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI) 144 if (Constant *CU = dyn_cast<Constant>(*UI)) { 145 if (!ConstantIsDead(CU)) return false; 146 } else 147 return false; 148 return true; 149} 150 151 152/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 153/// structure. If the global has its address taken, return true to indicate we 154/// can't do anything with it. 155/// 156static bool AnalyzeGlobal(Value *V, GlobalStatus &GS, 157 std::set<PHINode*> &PHIUsers) { 158 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) { 160 GS.HasNonInstructionUser = true; 161 162 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 163 164 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) { 165 if (!GS.HasMultipleAccessingFunctions) { 166 Function *F = I->getParent()->getParent(); 167 if (GS.AccessingFunction == 0) 168 GS.AccessingFunction = F; 169 else if (GS.AccessingFunction != F) 170 GS.HasMultipleAccessingFunctions = true; 171 } 172 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 173 GS.isLoaded = true; 174 if (LI->isVolatile()) return true; // Don't hack on volatile loads. 175 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 176 // Don't allow a store OF the address, only stores TO the address. 177 if (SI->getOperand(0) == V) return true; 178 179 if (SI->isVolatile()) return true; // Don't hack on volatile stores. 180 181 // If this is a direct store to the global (i.e., the global is a scalar 182 // value, not an aggregate), keep more specific information about 183 // stores. 184 if (GS.StoredType != GlobalStatus::isStored) { 185 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){ 186 Value *StoredVal = SI->getOperand(0); 187 if (StoredVal == GV->getInitializer()) { 188 if (GS.StoredType < GlobalStatus::isInitializerStored) 189 GS.StoredType = GlobalStatus::isInitializerStored; 190 } else if (isa<LoadInst>(StoredVal) && 191 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 192 // G = G 193 if (GS.StoredType < GlobalStatus::isInitializerStored) 194 GS.StoredType = GlobalStatus::isInitializerStored; 195 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 196 GS.StoredType = GlobalStatus::isStoredOnce; 197 GS.StoredOnceValue = StoredVal; 198 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 199 GS.StoredOnceValue == StoredVal) { 200 // noop. 201 } else { 202 GS.StoredType = GlobalStatus::isStored; 203 } 204 } else { 205 GS.StoredType = GlobalStatus::isStored; 206 } 207 } 208 } else if (isa<GetElementPtrInst>(I)) { 209 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 210 } else if (isa<SelectInst>(I)) { 211 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 212 } else if (PHINode *PN = dyn_cast<PHINode>(I)) { 213 // PHI nodes we can check just like select or GEP instructions, but we 214 // have to be careful about infinite recursion. 215 if (PHIUsers.insert(PN).second) // Not already visited. 216 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 217 GS.HasPHIUser = true; 218 } else if (isa<CmpInst>(I)) { 219 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) { 220 if (I->getOperand(1) == V) 221 GS.StoredType = GlobalStatus::isStored; 222 if (I->getOperand(2) == V) 223 GS.isLoaded = true; 224 } else if (isa<MemSetInst>(I)) { 225 assert(I->getOperand(1) == V && "Memset only takes one pointer!"); 226 GS.StoredType = GlobalStatus::isStored; 227 } else { 228 return true; // Any other non-load instruction might take address! 229 } 230 } else if (Constant *C = dyn_cast<Constant>(*UI)) { 231 GS.HasNonInstructionUser = true; 232 // We might have a dead and dangling constant hanging off of here. 233 if (!ConstantIsDead(C)) 234 return true; 235 } else { 236 GS.HasNonInstructionUser = true; 237 // Otherwise must be some other user. 238 return true; 239 } 240 241 return false; 242} 243 244static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 245 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 246 if (!CI) return 0; 247 unsigned IdxV = CI->getZExtValue(); 248 249 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 250 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 251 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 252 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 253 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 254 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 255 } else if (isa<ConstantAggregateZero>(Agg)) { 256 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 257 if (IdxV < STy->getNumElements()) 258 return Constant::getNullValue(STy->getElementType(IdxV)); 259 } else if (const SequentialType *STy = 260 dyn_cast<SequentialType>(Agg->getType())) { 261 return Constant::getNullValue(STy->getElementType()); 262 } 263 } else if (isa<UndefValue>(Agg)) { 264 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 265 if (IdxV < STy->getNumElements()) 266 return UndefValue::get(STy->getElementType(IdxV)); 267 } else if (const SequentialType *STy = 268 dyn_cast<SequentialType>(Agg->getType())) { 269 return UndefValue::get(STy->getElementType()); 270 } 271 } 272 return 0; 273} 274 275 276/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 277/// users of the global, cleaning up the obvious ones. This is largely just a 278/// quick scan over the use list to clean up the easy and obvious cruft. This 279/// returns true if it made a change. 280static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 281 bool Changed = false; 282 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 283 User *U = *UI++; 284 285 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 286 if (Init) { 287 // Replace the load with the initializer. 288 LI->replaceAllUsesWith(Init); 289 LI->eraseFromParent(); 290 Changed = true; 291 } 292 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 293 // Store must be unreachable or storing Init into the global. 294 SI->eraseFromParent(); 295 Changed = true; 296 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 297 if (CE->getOpcode() == Instruction::GetElementPtr) { 298 Constant *SubInit = 0; 299 if (Init) 300 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 301 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 302 } else if (CE->getOpcode() == Instruction::BitCast && 303 isa<PointerType>(CE->getType())) { 304 // Pointer cast, delete any stores and memsets to the global. 305 Changed |= CleanupConstantGlobalUsers(CE, 0); 306 } 307 308 if (CE->use_empty()) { 309 CE->destroyConstant(); 310 Changed = true; 311 } 312 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 313 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 314 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 315 // and will invalidate our notion of what Init is. 316 Constant *SubInit = 0; 317 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 318 ConstantExpr *CE = 319 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 320 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 321 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 322 } 323 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 324 325 if (GEP->use_empty()) { 326 GEP->eraseFromParent(); 327 Changed = true; 328 } 329 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 330 if (MI->getRawDest() == V) { 331 MI->eraseFromParent(); 332 Changed = true; 333 } 334 335 } else if (Constant *C = dyn_cast<Constant>(U)) { 336 // If we have a chain of dead constantexprs or other things dangling from 337 // us, and if they are all dead, nuke them without remorse. 338 if (ConstantIsDead(C)) { 339 C->destroyConstant(); 340 // This could have invalidated UI, start over from scratch. 341 CleanupConstantGlobalUsers(V, Init); 342 return true; 343 } 344 } 345 } 346 return Changed; 347} 348 349/// isSafeSROAElementUse - Return true if the specified instruction is a safe 350/// user of a derived expression from a global that we want to SROA. 351static bool isSafeSROAElementUse(Value *V) { 352 // We might have a dead and dangling constant hanging off of here. 353 if (Constant *C = dyn_cast<Constant>(V)) 354 return ConstantIsDead(C); 355 356 Instruction *I = dyn_cast<Instruction>(V); 357 if (!I) return false; 358 359 // Loads are ok. 360 if (isa<LoadInst>(I)) return true; 361 362 // Stores *to* the pointer are ok. 363 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 364 return SI->getOperand(0) != V; 365 366 // Otherwise, it must be a GEP. 367 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 368 if (GEPI == 0) return false; 369 370 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 371 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 372 return false; 373 374 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end(); 375 I != E; ++I) 376 if (!isSafeSROAElementUse(*I)) 377 return false; 378 return true; 379} 380 381 382/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 383/// Look at it and its uses and decide whether it is safe to SROA this global. 384/// 385static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 386 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 387 if (!isa<GetElementPtrInst>(U) && 388 (!isa<ConstantExpr>(U) || 389 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 390 return false; 391 392 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 393 // don't like < 3 operand CE's, and we don't like non-constant integer 394 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 395 // value of C. 396 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 397 !cast<Constant>(U->getOperand(1))->isNullValue() || 398 !isa<ConstantInt>(U->getOperand(2))) 399 return false; 400 401 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 402 ++GEPI; // Skip over the pointer index. 403 404 // If this is a use of an array allocation, do a bit more checking for sanity. 405 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 406 uint64_t NumElements = AT->getNumElements(); 407 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 408 409 // Check to make sure that index falls within the array. If not, 410 // something funny is going on, so we won't do the optimization. 411 // 412 if (Idx->getZExtValue() >= NumElements) 413 return false; 414 415 // We cannot scalar repl this level of the array unless any array 416 // sub-indices are in-range constants. In particular, consider: 417 // A[0][i]. We cannot know that the user isn't doing invalid things like 418 // allowing i to index an out-of-range subscript that accesses A[1]. 419 // 420 // Scalar replacing *just* the outer index of the array is probably not 421 // going to be a win anyway, so just give up. 422 for (++GEPI; // Skip array index. 423 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI)); 424 ++GEPI) { 425 uint64_t NumElements; 426 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 427 NumElements = SubArrayTy->getNumElements(); 428 else 429 NumElements = cast<VectorType>(*GEPI)->getNumElements(); 430 431 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 432 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 433 return false; 434 } 435 } 436 437 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) 438 if (!isSafeSROAElementUse(*I)) 439 return false; 440 return true; 441} 442 443/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 444/// is safe for us to perform this transformation. 445/// 446static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 447 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 448 UI != E; ++UI) { 449 if (!IsUserOfGlobalSafeForSRA(*UI, GV)) 450 return false; 451 } 452 return true; 453} 454 455 456/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 457/// variable. This opens the door for other optimizations by exposing the 458/// behavior of the program in a more fine-grained way. We have determined that 459/// this transformation is safe already. We return the first global variable we 460/// insert so that the caller can reprocess it. 461static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { 462 // Make sure this global only has simple uses that we can SRA. 463 if (!GlobalUsersSafeToSRA(GV)) 464 return 0; 465 466 assert(GV->hasInternalLinkage() && !GV->isConstant()); 467 Constant *Init = GV->getInitializer(); 468 const Type *Ty = Init->getType(); 469 470 std::vector<GlobalVariable*> NewGlobals; 471 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 472 473 // Get the alignment of the global, either explicit or target-specific. 474 unsigned StartAlignment = GV->getAlignment(); 475 if (StartAlignment == 0) 476 StartAlignment = TD.getABITypeAlignment(GV->getType()); 477 478 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 479 NewGlobals.reserve(STy->getNumElements()); 480 const StructLayout &Layout = *TD.getStructLayout(STy); 481 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 482 Constant *In = getAggregateConstantElement(Init, 483 ConstantInt::get(Type::Int32Ty, i)); 484 assert(In && "Couldn't get element of initializer?"); 485 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 486 GlobalVariable::InternalLinkage, 487 In, GV->getName()+"."+utostr(i), 488 (Module *)NULL, 489 GV->isThreadLocal()); 490 Globals.insert(GV, NGV); 491 NewGlobals.push_back(NGV); 492 493 // Calculate the known alignment of the field. If the original aggregate 494 // had 256 byte alignment for example, something might depend on that: 495 // propagate info to each field. 496 uint64_t FieldOffset = Layout.getElementOffset(i); 497 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 498 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 499 NGV->setAlignment(NewAlign); 500 } 501 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 502 unsigned NumElements = 0; 503 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 504 NumElements = ATy->getNumElements(); 505 else 506 NumElements = cast<VectorType>(STy)->getNumElements(); 507 508 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 509 return 0; // It's not worth it. 510 NewGlobals.reserve(NumElements); 511 512 uint64_t EltSize = TD.getABITypeSize(STy->getElementType()); 513 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 514 for (unsigned i = 0, e = NumElements; i != e; ++i) { 515 Constant *In = getAggregateConstantElement(Init, 516 ConstantInt::get(Type::Int32Ty, i)); 517 assert(In && "Couldn't get element of initializer?"); 518 519 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 520 GlobalVariable::InternalLinkage, 521 In, GV->getName()+"."+utostr(i), 522 (Module *)NULL, 523 GV->isThreadLocal()); 524 Globals.insert(GV, NGV); 525 NewGlobals.push_back(NGV); 526 527 // Calculate the known alignment of the field. If the original aggregate 528 // had 256 byte alignment for example, something might depend on that: 529 // propagate info to each field. 530 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 531 if (NewAlign > EltAlign) 532 NGV->setAlignment(NewAlign); 533 } 534 } 535 536 if (NewGlobals.empty()) 537 return 0; 538 539 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 540 541 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 542 543 // Loop over all of the uses of the global, replacing the constantexpr geps, 544 // with smaller constantexpr geps or direct references. 545 while (!GV->use_empty()) { 546 User *GEP = GV->use_back(); 547 assert(((isa<ConstantExpr>(GEP) && 548 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 549 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 550 551 // Ignore the 1th operand, which has to be zero or else the program is quite 552 // broken (undefined). Get the 2nd operand, which is the structure or array 553 // index. 554 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 555 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 556 557 Value *NewPtr = NewGlobals[Val]; 558 559 // Form a shorter GEP if needed. 560 if (GEP->getNumOperands() > 3) { 561 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 562 SmallVector<Constant*, 8> Idxs; 563 Idxs.push_back(NullInt); 564 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 565 Idxs.push_back(CE->getOperand(i)); 566 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 567 &Idxs[0], Idxs.size()); 568 } else { 569 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 570 SmallVector<Value*, 8> Idxs; 571 Idxs.push_back(NullInt); 572 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 573 Idxs.push_back(GEPI->getOperand(i)); 574 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 575 GEPI->getName()+"."+utostr(Val), GEPI); 576 } 577 } 578 GEP->replaceAllUsesWith(NewPtr); 579 580 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 581 GEPI->eraseFromParent(); 582 else 583 cast<ConstantExpr>(GEP)->destroyConstant(); 584 } 585 586 // Delete the old global, now that it is dead. 587 Globals.erase(GV); 588 ++NumSRA; 589 590 // Loop over the new globals array deleting any globals that are obviously 591 // dead. This can arise due to scalarization of a structure or an array that 592 // has elements that are dead. 593 unsigned FirstGlobal = 0; 594 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 595 if (NewGlobals[i]->use_empty()) { 596 Globals.erase(NewGlobals[i]); 597 if (FirstGlobal == i) ++FirstGlobal; 598 } 599 600 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 601} 602 603/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 604/// value will trap if the value is dynamically null. PHIs keeps track of any 605/// phi nodes we've seen to avoid reprocessing them. 606static bool AllUsesOfValueWillTrapIfNull(Value *V, 607 SmallPtrSet<PHINode*, 8> &PHIs) { 608 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 609 if (isa<LoadInst>(*UI)) { 610 // Will trap. 611 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 612 if (SI->getOperand(0) == V) { 613 //cerr << "NONTRAPPING USE: " << **UI; 614 return false; // Storing the value. 615 } 616 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 617 if (CI->getOperand(0) != V) { 618 //cerr << "NONTRAPPING USE: " << **UI; 619 return false; // Not calling the ptr 620 } 621 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 622 if (II->getOperand(0) != V) { 623 //cerr << "NONTRAPPING USE: " << **UI; 624 return false; // Not calling the ptr 625 } 626 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 627 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 628 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 629 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 630 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 631 // If we've already seen this phi node, ignore it, it has already been 632 // checked. 633 if (PHIs.insert(PN)) 634 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 635 } else if (isa<ICmpInst>(*UI) && 636 isa<ConstantPointerNull>(UI->getOperand(1))) { 637 // Ignore setcc X, null 638 } else { 639 //cerr << "NONTRAPPING USE: " << **UI; 640 return false; 641 } 642 return true; 643} 644 645/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 646/// from GV will trap if the loaded value is null. Note that this also permits 647/// comparisons of the loaded value against null, as a special case. 648static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 649 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 650 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 651 SmallPtrSet<PHINode*, 8> PHIs; 652 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 653 return false; 654 } else if (isa<StoreInst>(*UI)) { 655 // Ignore stores to the global. 656 } else { 657 // We don't know or understand this user, bail out. 658 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 659 return false; 660 } 661 662 return true; 663} 664 665static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 666 bool Changed = false; 667 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 668 Instruction *I = cast<Instruction>(*UI++); 669 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 670 LI->setOperand(0, NewV); 671 Changed = true; 672 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 673 if (SI->getOperand(1) == V) { 674 SI->setOperand(1, NewV); 675 Changed = true; 676 } 677 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 678 if (I->getOperand(0) == V) { 679 // Calling through the pointer! Turn into a direct call, but be careful 680 // that the pointer is not also being passed as an argument. 681 I->setOperand(0, NewV); 682 Changed = true; 683 bool PassedAsArg = false; 684 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 685 if (I->getOperand(i) == V) { 686 PassedAsArg = true; 687 I->setOperand(i, NewV); 688 } 689 690 if (PassedAsArg) { 691 // Being passed as an argument also. Be careful to not invalidate UI! 692 UI = V->use_begin(); 693 } 694 } 695 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 696 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 697 ConstantExpr::getCast(CI->getOpcode(), 698 NewV, CI->getType())); 699 if (CI->use_empty()) { 700 Changed = true; 701 CI->eraseFromParent(); 702 } 703 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 704 // Should handle GEP here. 705 SmallVector<Constant*, 8> Idxs; 706 Idxs.reserve(GEPI->getNumOperands()-1); 707 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 708 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i))) 709 Idxs.push_back(C); 710 else 711 break; 712 if (Idxs.size() == GEPI->getNumOperands()-1) 713 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 714 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 715 Idxs.size())); 716 if (GEPI->use_empty()) { 717 Changed = true; 718 GEPI->eraseFromParent(); 719 } 720 } 721 } 722 723 return Changed; 724} 725 726 727/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 728/// value stored into it. If there are uses of the loaded value that would trap 729/// if the loaded value is dynamically null, then we know that they cannot be 730/// reachable with a null optimize away the load. 731static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 732 std::vector<LoadInst*> Loads; 733 bool Changed = false; 734 735 // Replace all uses of loads with uses of uses of the stored value. 736 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); 737 GUI != E; ++GUI) 738 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) { 739 Loads.push_back(LI); 740 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 741 } else { 742 // If we get here we could have stores, selects, or phi nodes whose values 743 // are loaded. 744 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) || 745 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) && 746 "Only expect load and stores!"); 747 } 748 749 if (Changed) { 750 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 751 ++NumGlobUses; 752 } 753 754 // Delete all of the loads we can, keeping track of whether we nuked them all! 755 bool AllLoadsGone = true; 756 while (!Loads.empty()) { 757 LoadInst *L = Loads.back(); 758 if (L->use_empty()) { 759 L->eraseFromParent(); 760 Changed = true; 761 } else { 762 AllLoadsGone = false; 763 } 764 Loads.pop_back(); 765 } 766 767 // If we nuked all of the loads, then none of the stores are needed either, 768 // nor is the global. 769 if (AllLoadsGone) { 770 DOUT << " *** GLOBAL NOW DEAD!\n"; 771 CleanupConstantGlobalUsers(GV, 0); 772 if (GV->use_empty()) { 773 GV->eraseFromParent(); 774 ++NumDeleted; 775 } 776 Changed = true; 777 } 778 return Changed; 779} 780 781/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 782/// instructions that are foldable. 783static void ConstantPropUsersOf(Value *V) { 784 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 785 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 786 if (Constant *NewC = ConstantFoldInstruction(I)) { 787 I->replaceAllUsesWith(NewC); 788 789 // Advance UI to the next non-I use to avoid invalidating it! 790 // Instructions could multiply use V. 791 while (UI != E && *UI == I) 792 ++UI; 793 I->eraseFromParent(); 794 } 795} 796 797/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 798/// variable, and transforms the program as if it always contained the result of 799/// the specified malloc. Because it is always the result of the specified 800/// malloc, there is no reason to actually DO the malloc. Instead, turn the 801/// malloc into a global, and any loads of GV as uses of the new global. 802static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 803 MallocInst *MI) { 804 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 805 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 806 807 if (NElements->getZExtValue() != 1) { 808 // If we have an array allocation, transform it to a single element 809 // allocation to make the code below simpler. 810 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 811 NElements->getZExtValue()); 812 MallocInst *NewMI = 813 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 814 MI->getAlignment(), MI->getName(), MI); 815 Value* Indices[2]; 816 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 817 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2, 818 NewMI->getName()+".el0", MI); 819 MI->replaceAllUsesWith(NewGEP); 820 MI->eraseFromParent(); 821 MI = NewMI; 822 } 823 824 // Create the new global variable. The contents of the malloc'd memory is 825 // undefined, so initialize with an undef value. 826 Constant *Init = UndefValue::get(MI->getAllocatedType()); 827 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 828 GlobalValue::InternalLinkage, Init, 829 GV->getName()+".body", 830 (Module *)NULL, 831 GV->isThreadLocal()); 832 // FIXME: This new global should have the alignment returned by malloc. Code 833 // could depend on malloc returning large alignment (on the mac, 16 bytes) but 834 // this would only guarantee some lower alignment. 835 GV->getParent()->getGlobalList().insert(GV, NewGV); 836 837 // Anything that used the malloc now uses the global directly. 838 MI->replaceAllUsesWith(NewGV); 839 840 Constant *RepValue = NewGV; 841 if (NewGV->getType() != GV->getType()->getElementType()) 842 RepValue = ConstantExpr::getBitCast(RepValue, 843 GV->getType()->getElementType()); 844 845 // If there is a comparison against null, we will insert a global bool to 846 // keep track of whether the global was initialized yet or not. 847 GlobalVariable *InitBool = 848 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 849 ConstantInt::getFalse(), GV->getName()+".init", 850 (Module *)NULL, GV->isThreadLocal()); 851 bool InitBoolUsed = false; 852 853 // Loop over all uses of GV, processing them in turn. 854 std::vector<StoreInst*> Stores; 855 while (!GV->use_empty()) 856 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 857 while (!LI->use_empty()) { 858 Use &LoadUse = LI->use_begin().getUse(); 859 if (!isa<ICmpInst>(LoadUse.getUser())) 860 LoadUse = RepValue; 861 else { 862 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 863 // Replace the cmp X, 0 with a use of the bool value. 864 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 865 InitBoolUsed = true; 866 switch (CI->getPredicate()) { 867 default: assert(0 && "Unknown ICmp Predicate!"); 868 case ICmpInst::ICMP_ULT: 869 case ICmpInst::ICMP_SLT: 870 LV = ConstantInt::getFalse(); // X < null -> always false 871 break; 872 case ICmpInst::ICMP_ULE: 873 case ICmpInst::ICMP_SLE: 874 case ICmpInst::ICMP_EQ: 875 LV = BinaryOperator::createNot(LV, "notinit", CI); 876 break; 877 case ICmpInst::ICMP_NE: 878 case ICmpInst::ICMP_UGE: 879 case ICmpInst::ICMP_SGE: 880 case ICmpInst::ICMP_UGT: 881 case ICmpInst::ICMP_SGT: 882 break; // no change. 883 } 884 CI->replaceAllUsesWith(LV); 885 CI->eraseFromParent(); 886 } 887 } 888 LI->eraseFromParent(); 889 } else { 890 StoreInst *SI = cast<StoreInst>(GV->use_back()); 891 // The global is initialized when the store to it occurs. 892 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 893 SI->eraseFromParent(); 894 } 895 896 // If the initialization boolean was used, insert it, otherwise delete it. 897 if (!InitBoolUsed) { 898 while (!InitBool->use_empty()) // Delete initializations 899 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 900 delete InitBool; 901 } else 902 GV->getParent()->getGlobalList().insert(GV, InitBool); 903 904 905 // Now the GV is dead, nuke it and the malloc. 906 GV->eraseFromParent(); 907 MI->eraseFromParent(); 908 909 // To further other optimizations, loop over all users of NewGV and try to 910 // constant prop them. This will promote GEP instructions with constant 911 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 912 ConstantPropUsersOf(NewGV); 913 if (RepValue != NewGV) 914 ConstantPropUsersOf(RepValue); 915 916 return NewGV; 917} 918 919/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 920/// to make sure that there are no complex uses of V. We permit simple things 921/// like dereferencing the pointer, but not storing through the address, unless 922/// it is to the specified global. 923static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 924 GlobalVariable *GV, 925 SmallPtrSet<PHINode*, 8> &PHIs) { 926 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 927 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) { 928 // Fine, ignore. 929 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 930 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 931 return false; // Storing the pointer itself... bad. 932 // Otherwise, storing through it, or storing into GV... fine. 933 } else if (isa<GetElementPtrInst>(*UI)) { 934 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI), 935 GV, PHIs)) 936 return false; 937 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 938 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 939 // cycles. 940 if (PHIs.insert(PN)) 941 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 942 return false; 943 } else { 944 return false; 945 } 946 return true; 947} 948 949/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 950/// somewhere. Transform all uses of the allocation into loads from the 951/// global and uses of the resultant pointer. Further, delete the store into 952/// GV. This assumes that these value pass the 953/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 954static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 955 GlobalVariable *GV) { 956 while (!Alloc->use_empty()) { 957 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 958 Instruction *InsertPt = U; 959 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 960 // If this is the store of the allocation into the global, remove it. 961 if (SI->getOperand(1) == GV) { 962 SI->eraseFromParent(); 963 continue; 964 } 965 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 966 // Insert the load in the corresponding predecessor, not right before the 967 // PHI. 968 unsigned PredNo = Alloc->use_begin().getOperandNo()/2; 969 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator(); 970 } 971 972 // Insert a load from the global, and use it instead of the malloc. 973 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 974 U->replaceUsesOfWith(Alloc, NL); 975 } 976} 977 978/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 979/// GV are simple enough to perform HeapSRA, return true. 980static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 981 MallocInst *MI) { 982 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 983 ++UI) 984 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 985 // We permit two users of the load: setcc comparing against the null 986 // pointer, and a getelementptr of a specific form. 987 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E; 988 ++UI) { 989 // Comparison against null is ok. 990 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) { 991 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 992 return false; 993 continue; 994 } 995 996 // getelementptr is also ok, but only a simple form. 997 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 998 // Must index into the array and into the struct. 999 if (GEPI->getNumOperands() < 3) 1000 return false; 1001 1002 // Otherwise the GEP is ok. 1003 continue; 1004 } 1005 1006 if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 1007 // We have a phi of a load from the global. We can only handle this 1008 // if the other PHI'd values are actually the same. In this case, 1009 // the rewriter will just drop the phi entirely. 1010 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1011 Value *IV = PN->getIncomingValue(i); 1012 if (IV == LI) continue; // Trivial the same. 1013 1014 // If the phi'd value is from the malloc that initializes the value, 1015 // we can xform it. 1016 if (IV == MI) continue; 1017 1018 // Otherwise, we don't know what it is. 1019 return false; 1020 } 1021 return true; 1022 } 1023 1024 // Otherwise we don't know what this is, not ok. 1025 return false; 1026 } 1027 } 1028 return true; 1029} 1030 1031/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd 1032/// value, lazily creating it on demand. 1033static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo, 1034 const std::vector<GlobalVariable*> &FieldGlobals, 1035 std::vector<Value *> &InsertedLoadsForPtr) { 1036 if (InsertedLoadsForPtr.size() <= FieldNo) 1037 InsertedLoadsForPtr.resize(FieldNo+1); 1038 if (InsertedLoadsForPtr[FieldNo] == 0) 1039 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo], 1040 Load->getName()+".f" + 1041 utostr(FieldNo), Load); 1042 return InsertedLoadsForPtr[FieldNo]; 1043} 1044 1045/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1046/// the load, rewrite the derived value to use the HeapSRoA'd load. 1047static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser, 1048 const std::vector<GlobalVariable*> &FieldGlobals, 1049 std::vector<Value *> &InsertedLoadsForPtr) { 1050 // If this is a comparison against null, handle it. 1051 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1052 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1053 // If we have a setcc of the loaded pointer, we can use a setcc of any 1054 // field. 1055 Value *NPtr; 1056 if (InsertedLoadsForPtr.empty()) { 1057 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr); 1058 } else { 1059 NPtr = InsertedLoadsForPtr.back(); 1060 } 1061 1062 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1063 Constant::getNullValue(NPtr->getType()), 1064 SCI->getName(), SCI); 1065 SCI->replaceAllUsesWith(New); 1066 SCI->eraseFromParent(); 1067 return; 1068 } 1069 1070 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...' 1071 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1072 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1073 && "Unexpected GEPI!"); 1074 1075 // Load the pointer for this field. 1076 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1077 Value *NewPtr = GetHeapSROALoad(Load, FieldNo, 1078 FieldGlobals, InsertedLoadsForPtr); 1079 1080 // Create the new GEP idx vector. 1081 SmallVector<Value*, 8> GEPIdx; 1082 GEPIdx.push_back(GEPI->getOperand(1)); 1083 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1084 1085 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx.begin(), GEPIdx.end(), 1086 GEPI->getName(), GEPI); 1087 GEPI->replaceAllUsesWith(NGEPI); 1088 GEPI->eraseFromParent(); 1089 return; 1090 } 1091 1092 // Handle PHI nodes. PHI nodes must be merging in the same values, plus 1093 // potentially the original malloc. Insert phi nodes for each field, then 1094 // process uses of the PHI. 1095 PHINode *PN = cast<PHINode>(LoadUser); 1096 std::vector<Value *> PHIsForField; 1097 PHIsForField.resize(FieldGlobals.size()); 1098 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1099 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr); 1100 1101 PHINode *FieldPN = PHINode::Create(LoadV->getType(), 1102 PN->getName()+"."+utostr(i), PN); 1103 // Fill in the predecessor values. 1104 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) { 1105 // Each predecessor either uses the load or the original malloc. 1106 Value *InVal = PN->getIncomingValue(pred); 1107 BasicBlock *BB = PN->getIncomingBlock(pred); 1108 Value *NewVal; 1109 if (isa<MallocInst>(InVal)) { 1110 // Insert a reload from the global in the predecessor. 1111 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals, 1112 PHIsForField); 1113 } else { 1114 NewVal = InsertedLoadsForPtr[i]; 1115 } 1116 FieldPN->addIncoming(NewVal, BB); 1117 } 1118 PHIsForField[i] = FieldPN; 1119 } 1120 1121 // Since PHIsForField specifies a phi for every input value, the lazy inserter 1122 // will never insert a load. 1123 while (!PN->use_empty()) 1124 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField); 1125 PN->eraseFromParent(); 1126} 1127 1128/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1129/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1130/// use FieldGlobals instead. All uses of loaded values satisfy 1131/// GlobalLoadUsesSimpleEnoughForHeapSRA. 1132static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1133 const std::vector<GlobalVariable*> &FieldGlobals) { 1134 std::vector<Value *> InsertedLoadsForPtr; 1135 //InsertedLoadsForPtr.resize(FieldGlobals.size()); 1136 while (!Load->use_empty()) 1137 RewriteHeapSROALoadUser(Load, Load->use_back(), 1138 FieldGlobals, InsertedLoadsForPtr); 1139} 1140 1141/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1142/// it up into multiple allocations of arrays of the fields. 1143static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1144 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1145 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1146 1147 // There is guaranteed to be at least one use of the malloc (storing 1148 // it into GV). If there are other uses, change them to be uses of 1149 // the global to simplify later code. This also deletes the store 1150 // into GV. 1151 ReplaceUsesOfMallocWithGlobal(MI, GV); 1152 1153 // Okay, at this point, there are no users of the malloc. Insert N 1154 // new mallocs at the same place as MI, and N globals. 1155 std::vector<GlobalVariable*> FieldGlobals; 1156 std::vector<MallocInst*> FieldMallocs; 1157 1158 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1159 const Type *FieldTy = STy->getElementType(FieldNo); 1160 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1161 1162 GlobalVariable *NGV = 1163 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1164 Constant::getNullValue(PFieldTy), 1165 GV->getName() + ".f" + utostr(FieldNo), GV, 1166 GV->isThreadLocal()); 1167 FieldGlobals.push_back(NGV); 1168 1169 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1170 MI->getName() + ".f" + utostr(FieldNo),MI); 1171 FieldMallocs.push_back(NMI); 1172 new StoreInst(NMI, NGV, MI); 1173 } 1174 1175 // The tricky aspect of this transformation is handling the case when malloc 1176 // fails. In the original code, malloc failing would set the result pointer 1177 // of malloc to null. In this case, some mallocs could succeed and others 1178 // could fail. As such, we emit code that looks like this: 1179 // F0 = malloc(field0) 1180 // F1 = malloc(field1) 1181 // F2 = malloc(field2) 1182 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1183 // if (F0) { free(F0); F0 = 0; } 1184 // if (F1) { free(F1); F1 = 0; } 1185 // if (F2) { free(F2); F2 = 0; } 1186 // } 1187 Value *RunningOr = 0; 1188 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1189 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1190 Constant::getNullValue(FieldMallocs[i]->getType()), 1191 "isnull", MI); 1192 if (!RunningOr) 1193 RunningOr = Cond; // First seteq 1194 else 1195 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI); 1196 } 1197 1198 // Split the basic block at the old malloc. 1199 BasicBlock *OrigBB = MI->getParent(); 1200 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1201 1202 // Create the block to check the first condition. Put all these blocks at the 1203 // end of the function as they are unlikely to be executed. 1204 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1205 OrigBB->getParent()); 1206 1207 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1208 // branch on RunningOr. 1209 OrigBB->getTerminator()->eraseFromParent(); 1210 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1211 1212 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1213 // pointer, because some may be null while others are not. 1214 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1215 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1216 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1217 Constant::getNullValue(GVVal->getType()), 1218 "tmp", NullPtrBlock); 1219 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1220 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1221 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1222 1223 // Fill in FreeBlock. 1224 new FreeInst(GVVal, FreeBlock); 1225 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1226 FreeBlock); 1227 BranchInst::Create(NextBlock, FreeBlock); 1228 1229 NullPtrBlock = NextBlock; 1230 } 1231 1232 BranchInst::Create(ContBB, NullPtrBlock); 1233 1234 // MI is no longer needed, remove it. 1235 MI->eraseFromParent(); 1236 1237 1238 // Okay, the malloc site is completely handled. All of the uses of GV are now 1239 // loads, and all uses of those loads are simple. Rewrite them to use loads 1240 // of the per-field globals instead. 1241 while (!GV->use_empty()) { 1242 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 1243 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals); 1244 LI->eraseFromParent(); 1245 } else { 1246 // Must be a store of null. 1247 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1248 assert(isa<Constant>(SI->getOperand(0)) && 1249 cast<Constant>(SI->getOperand(0))->isNullValue() && 1250 "Unexpected heap-sra user!"); 1251 1252 // Insert a store of null into each global. 1253 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1254 Constant *Null = 1255 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType()); 1256 new StoreInst(Null, FieldGlobals[i], SI); 1257 } 1258 // Erase the original store. 1259 SI->eraseFromParent(); 1260 } 1261 } 1262 1263 // The old global is now dead, remove it. 1264 GV->eraseFromParent(); 1265 1266 ++NumHeapSRA; 1267 return FieldGlobals[0]; 1268} 1269 1270 1271// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1272// that only one value (besides its initializer) is ever stored to the global. 1273static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1274 Module::global_iterator &GVI, 1275 TargetData &TD) { 1276 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal)) 1277 StoredOnceVal = CI->getOperand(0); 1278 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){ 1279 // "getelementptr Ptr, 0, 0, 0" is really just a cast. 1280 bool IsJustACast = true; 1281 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 1282 if (!isa<Constant>(GEPI->getOperand(i)) || 1283 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) { 1284 IsJustACast = false; 1285 break; 1286 } 1287 if (IsJustACast) 1288 StoredOnceVal = GEPI->getOperand(0); 1289 } 1290 1291 // If we are dealing with a pointer global that is initialized to null and 1292 // only has one (non-null) value stored into it, then we can optimize any 1293 // users of the loaded value (often calls and loads) that would trap if the 1294 // value was null. 1295 if (isa<PointerType>(GV->getInitializer()->getType()) && 1296 GV->getInitializer()->isNullValue()) { 1297 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1298 if (GV->getInitializer()->getType() != SOVC->getType()) 1299 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1300 1301 // Optimize away any trapping uses of the loaded value. 1302 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1303 return true; 1304 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1305 // If this is a malloc of an abstract type, don't touch it. 1306 if (!MI->getAllocatedType()->isSized()) 1307 return false; 1308 1309 // We can't optimize this global unless all uses of it are *known* to be 1310 // of the malloc value, not of the null initializer value (consider a use 1311 // that compares the global's value against zero to see if the malloc has 1312 // been reached). To do this, we check to see if all uses of the global 1313 // would trap if the global were null: this proves that they must all 1314 // happen after the malloc. 1315 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1316 return false; 1317 1318 // We can't optimize this if the malloc itself is used in a complex way, 1319 // for example, being stored into multiple globals. This allows the 1320 // malloc to be stored into the specified global, loaded setcc'd, and 1321 // GEP'd. These are all things we could transform to using the global 1322 // for. 1323 { 1324 SmallPtrSet<PHINode*, 8> PHIs; 1325 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1326 return false; 1327 } 1328 1329 1330 // If we have a global that is only initialized with a fixed size malloc, 1331 // transform the program to use global memory instead of malloc'd memory. 1332 // This eliminates dynamic allocation, avoids an indirection accessing the 1333 // data, and exposes the resultant global to further GlobalOpt. 1334 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1335 // Restrict this transformation to only working on small allocations 1336 // (2048 bytes currently), as we don't want to introduce a 16M global or 1337 // something. 1338 if (NElements->getZExtValue()* 1339 TD.getABITypeSize(MI->getAllocatedType()) < 2048) { 1340 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1341 return true; 1342 } 1343 } 1344 1345 // If the allocation is an array of structures, consider transforming this 1346 // into multiple malloc'd arrays, one for each field. This is basically 1347 // SRoA for malloc'd memory. 1348 if (const StructType *AllocTy = 1349 dyn_cast<StructType>(MI->getAllocatedType())) { 1350 // This the structure has an unreasonable number of fields, leave it 1351 // alone. 1352 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 && 1353 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1354 GVI = PerformHeapAllocSRoA(GV, MI); 1355 return true; 1356 } 1357 } 1358 } 1359 } 1360 1361 return false; 1362} 1363 1364/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1365/// two values ever stored into GV are its initializer and OtherVal. See if we 1366/// can shrink the global into a boolean and select between the two values 1367/// whenever it is used. This exposes the values to other scalar optimizations. 1368static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1369 const Type *GVElType = GV->getType()->getElementType(); 1370 1371 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1372 // an FP value or vector, don't do this optimization because a select between 1373 // them is very expensive and unlikely to lead to later simplification. 1374 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1375 isa<VectorType>(GVElType)) 1376 return false; 1377 1378 // Walk the use list of the global seeing if all the uses are load or store. 1379 // If there is anything else, bail out. 1380 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1381 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1382 return false; 1383 1384 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1385 1386 // Create the new global, initializing it to false. 1387 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1388 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1389 GV->getName()+".b", 1390 (Module *)NULL, 1391 GV->isThreadLocal()); 1392 GV->getParent()->getGlobalList().insert(GV, NewGV); 1393 1394 Constant *InitVal = GV->getInitializer(); 1395 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1396 1397 // If initialized to zero and storing one into the global, we can use a cast 1398 // instead of a select to synthesize the desired value. 1399 bool IsOneZero = false; 1400 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1401 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1402 1403 while (!GV->use_empty()) { 1404 Instruction *UI = cast<Instruction>(GV->use_back()); 1405 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1406 // Change the store into a boolean store. 1407 bool StoringOther = SI->getOperand(0) == OtherVal; 1408 // Only do this if we weren't storing a loaded value. 1409 Value *StoreVal; 1410 if (StoringOther || SI->getOperand(0) == InitVal) 1411 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1412 else { 1413 // Otherwise, we are storing a previously loaded copy. To do this, 1414 // change the copy from copying the original value to just copying the 1415 // bool. 1416 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1417 1418 // If we're already replaced the input, StoredVal will be a cast or 1419 // select instruction. If not, it will be a load of the original 1420 // global. 1421 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1422 assert(LI->getOperand(0) == GV && "Not a copy!"); 1423 // Insert a new load, to preserve the saved value. 1424 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1425 } else { 1426 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1427 "This is not a form that we understand!"); 1428 StoreVal = StoredVal->getOperand(0); 1429 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1430 } 1431 } 1432 new StoreInst(StoreVal, NewGV, SI); 1433 } else { 1434 // Change the load into a load of bool then a select. 1435 LoadInst *LI = cast<LoadInst>(UI); 1436 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1437 Value *NSI; 1438 if (IsOneZero) 1439 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1440 else 1441 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1442 NSI->takeName(LI); 1443 LI->replaceAllUsesWith(NSI); 1444 } 1445 UI->eraseFromParent(); 1446 } 1447 1448 GV->eraseFromParent(); 1449 return true; 1450} 1451 1452 1453/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1454/// it if possible. If we make a change, return true. 1455bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1456 Module::global_iterator &GVI) { 1457 std::set<PHINode*> PHIUsers; 1458 GlobalStatus GS; 1459 GV->removeDeadConstantUsers(); 1460 1461 if (GV->use_empty()) { 1462 DOUT << "GLOBAL DEAD: " << *GV; 1463 GV->eraseFromParent(); 1464 ++NumDeleted; 1465 return true; 1466 } 1467 1468 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1469#if 0 1470 cerr << "Global: " << *GV; 1471 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1472 cerr << " StoredType = "; 1473 switch (GS.StoredType) { 1474 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1475 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1476 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1477 case GlobalStatus::isStored: cerr << "stored\n"; break; 1478 } 1479 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1480 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1481 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1482 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1483 << "\n"; 1484 cerr << " HasMultipleAccessingFunctions = " 1485 << GS.HasMultipleAccessingFunctions << "\n"; 1486 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1487 cerr << "\n"; 1488#endif 1489 1490 // If this is a first class global and has only one accessing function 1491 // and this function is main (which we know is not recursive we can make 1492 // this global a local variable) we replace the global with a local alloca 1493 // in this function. 1494 // 1495 // NOTE: It doesn't make sense to promote non first class types since we 1496 // are just replacing static memory to stack memory. 1497 if (!GS.HasMultipleAccessingFunctions && 1498 GS.AccessingFunction && !GS.HasNonInstructionUser && 1499 GV->getType()->getElementType()->isFirstClassType() && 1500 GS.AccessingFunction->getName() == "main" && 1501 GS.AccessingFunction->hasExternalLinkage()) { 1502 DOUT << "LOCALIZING GLOBAL: " << *GV; 1503 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1504 const Type* ElemTy = GV->getType()->getElementType(); 1505 // FIXME: Pass Global's alignment when globals have alignment 1506 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1507 if (!isa<UndefValue>(GV->getInitializer())) 1508 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1509 1510 GV->replaceAllUsesWith(Alloca); 1511 GV->eraseFromParent(); 1512 ++NumLocalized; 1513 return true; 1514 } 1515 1516 // If the global is never loaded (but may be stored to), it is dead. 1517 // Delete it now. 1518 if (!GS.isLoaded) { 1519 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1520 1521 // Delete any stores we can find to the global. We may not be able to 1522 // make it completely dead though. 1523 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1524 1525 // If the global is dead now, delete it. 1526 if (GV->use_empty()) { 1527 GV->eraseFromParent(); 1528 ++NumDeleted; 1529 Changed = true; 1530 } 1531 return Changed; 1532 1533 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1534 DOUT << "MARKING CONSTANT: " << *GV; 1535 GV->setConstant(true); 1536 1537 // Clean up any obviously simplifiable users now. 1538 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1539 1540 // If the global is dead now, just nuke it. 1541 if (GV->use_empty()) { 1542 DOUT << " *** Marking constant allowed us to simplify " 1543 << "all users and delete global!\n"; 1544 GV->eraseFromParent(); 1545 ++NumDeleted; 1546 } 1547 1548 ++NumMarked; 1549 return true; 1550 } else if (!GV->getInitializer()->getType()->isFirstClassType()) { 1551 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1552 getAnalysis<TargetData>())) { 1553 GVI = FirstNewGV; // Don't skip the newly produced globals! 1554 return true; 1555 } 1556 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1557 // If the initial value for the global was an undef value, and if only 1558 // one other value was stored into it, we can just change the 1559 // initializer to be an undef value, then delete all stores to the 1560 // global. This allows us to mark it constant. 1561 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1562 if (isa<UndefValue>(GV->getInitializer())) { 1563 // Change the initial value here. 1564 GV->setInitializer(SOVConstant); 1565 1566 // Clean up any obviously simplifiable users now. 1567 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1568 1569 if (GV->use_empty()) { 1570 DOUT << " *** Substituting initializer allowed us to " 1571 << "simplify all users and delete global!\n"; 1572 GV->eraseFromParent(); 1573 ++NumDeleted; 1574 } else { 1575 GVI = GV; 1576 } 1577 ++NumSubstitute; 1578 return true; 1579 } 1580 1581 // Try to optimize globals based on the knowledge that only one value 1582 // (besides its initializer) is ever stored to the global. 1583 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1584 getAnalysis<TargetData>())) 1585 return true; 1586 1587 // Otherwise, if the global was not a boolean, we can shrink it to be a 1588 // boolean. 1589 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1590 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1591 ++NumShrunkToBool; 1592 return true; 1593 } 1594 } 1595 } 1596 return false; 1597} 1598 1599/// OnlyCalledDirectly - Return true if the specified function is only called 1600/// directly. In other words, its address is never taken. 1601static bool OnlyCalledDirectly(Function *F) { 1602 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1603 Instruction *User = dyn_cast<Instruction>(*UI); 1604 if (!User) return false; 1605 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1606 1607 // See if the function address is passed as an argument. 1608 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i) 1609 if (User->getOperand(i) == F) return false; 1610 } 1611 return true; 1612} 1613 1614/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1615/// function, changing them to FastCC. 1616static void ChangeCalleesToFastCall(Function *F) { 1617 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1618 CallSite User(cast<Instruction>(*UI)); 1619 User.setCallingConv(CallingConv::Fast); 1620 } 1621} 1622 1623static PAListPtr StripNest(const PAListPtr &Attrs) { 1624 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1625 if ((Attrs.getSlot(i).Attrs & ParamAttr::Nest) == 0) 1626 continue; 1627 1628 // There can be only one. 1629 return Attrs.removeAttr(Attrs.getSlot(i).Index, ParamAttr::Nest); 1630 } 1631 1632 return Attrs; 1633} 1634 1635static void RemoveNestAttribute(Function *F) { 1636 F->setParamAttrs(StripNest(F->getParamAttrs())); 1637 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1638 CallSite User(cast<Instruction>(*UI)); 1639 User.setParamAttrs(StripNest(User.getParamAttrs())); 1640 } 1641} 1642 1643bool GlobalOpt::OptimizeFunctions(Module &M) { 1644 bool Changed = false; 1645 // Optimize functions. 1646 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1647 Function *F = FI++; 1648 F->removeDeadConstantUsers(); 1649 if (F->use_empty() && (F->hasInternalLinkage() || 1650 F->hasLinkOnceLinkage())) { 1651 M.getFunctionList().erase(F); 1652 Changed = true; 1653 ++NumFnDeleted; 1654 } else if (F->hasInternalLinkage()) { 1655 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1656 OnlyCalledDirectly(F)) { 1657 // If this function has C calling conventions, is not a varargs 1658 // function, and is only called directly, promote it to use the Fast 1659 // calling convention. 1660 F->setCallingConv(CallingConv::Fast); 1661 ChangeCalleesToFastCall(F); 1662 ++NumFastCallFns; 1663 Changed = true; 1664 } 1665 1666 if (F->getParamAttrs().hasAttrSomewhere(ParamAttr::Nest) && 1667 OnlyCalledDirectly(F)) { 1668 // The function is not used by a trampoline intrinsic, so it is safe 1669 // to remove the 'nest' attribute. 1670 RemoveNestAttribute(F); 1671 ++NumNestRemoved; 1672 Changed = true; 1673 } 1674 } 1675 } 1676 return Changed; 1677} 1678 1679bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1680 bool Changed = false; 1681 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1682 GVI != E; ) { 1683 GlobalVariable *GV = GVI++; 1684 if (!GV->isConstant() && GV->hasInternalLinkage() && 1685 GV->hasInitializer()) 1686 Changed |= ProcessInternalGlobal(GV, GVI); 1687 } 1688 return Changed; 1689} 1690 1691/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1692/// initializers have an init priority of 65535. 1693GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1694 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1695 I != E; ++I) 1696 if (I->getName() == "llvm.global_ctors") { 1697 // Found it, verify it's an array of { int, void()* }. 1698 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1699 if (!ATy) return 0; 1700 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1701 if (!STy || STy->getNumElements() != 2 || 1702 STy->getElementType(0) != Type::Int32Ty) return 0; 1703 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1704 if (!PFTy) return 0; 1705 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1706 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1707 FTy->getNumParams() != 0) 1708 return 0; 1709 1710 // Verify that the initializer is simple enough for us to handle. 1711 if (!I->hasInitializer()) return 0; 1712 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1713 if (!CA) return 0; 1714 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1715 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) { 1716 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1717 continue; 1718 1719 // Must have a function or null ptr. 1720 if (!isa<Function>(CS->getOperand(1))) 1721 return 0; 1722 1723 // Init priority must be standard. 1724 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1725 if (!CI || CI->getZExtValue() != 65535) 1726 return 0; 1727 } else { 1728 return 0; 1729 } 1730 1731 return I; 1732 } 1733 return 0; 1734} 1735 1736/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1737/// return a list of the functions and null terminator as a vector. 1738static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1739 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1740 std::vector<Function*> Result; 1741 Result.reserve(CA->getNumOperands()); 1742 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) { 1743 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i)); 1744 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1745 } 1746 return Result; 1747} 1748 1749/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1750/// specified array, returning the new global to use. 1751static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1752 const std::vector<Function*> &Ctors) { 1753 // If we made a change, reassemble the initializer list. 1754 std::vector<Constant*> CSVals; 1755 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1756 CSVals.push_back(0); 1757 1758 // Create the new init list. 1759 std::vector<Constant*> CAList; 1760 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1761 if (Ctors[i]) { 1762 CSVals[1] = Ctors[i]; 1763 } else { 1764 const Type *FTy = FunctionType::get(Type::VoidTy, 1765 std::vector<const Type*>(), false); 1766 const PointerType *PFTy = PointerType::getUnqual(FTy); 1767 CSVals[1] = Constant::getNullValue(PFTy); 1768 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1769 } 1770 CAList.push_back(ConstantStruct::get(CSVals)); 1771 } 1772 1773 // Create the array initializer. 1774 const Type *StructTy = 1775 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1776 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1777 CAList); 1778 1779 // If we didn't change the number of elements, don't create a new GV. 1780 if (CA->getType() == GCL->getInitializer()->getType()) { 1781 GCL->setInitializer(CA); 1782 return GCL; 1783 } 1784 1785 // Create the new global and insert it next to the existing list. 1786 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1787 GCL->getLinkage(), CA, "", 1788 (Module *)NULL, 1789 GCL->isThreadLocal()); 1790 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1791 NGV->takeName(GCL); 1792 1793 // Nuke the old list, replacing any uses with the new one. 1794 if (!GCL->use_empty()) { 1795 Constant *V = NGV; 1796 if (V->getType() != GCL->getType()) 1797 V = ConstantExpr::getBitCast(V, GCL->getType()); 1798 GCL->replaceAllUsesWith(V); 1799 } 1800 GCL->eraseFromParent(); 1801 1802 if (Ctors.size()) 1803 return NGV; 1804 else 1805 return 0; 1806} 1807 1808 1809static Constant *getVal(std::map<Value*, Constant*> &ComputedValues, 1810 Value *V) { 1811 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1812 Constant *R = ComputedValues[V]; 1813 assert(R && "Reference to an uncomputed value!"); 1814 return R; 1815} 1816 1817/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1818/// enough for us to understand. In particular, if it is a cast of something, 1819/// we punt. We basically just support direct accesses to globals and GEP's of 1820/// globals. This should be kept up to date with CommitValueTo. 1821static bool isSimpleEnoughPointerToCommit(Constant *C) { 1822 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1823 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1824 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1825 return !GV->isDeclaration(); // reject external globals. 1826 } 1827 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 1828 // Handle a constantexpr gep. 1829 if (CE->getOpcode() == Instruction::GetElementPtr && 1830 isa<GlobalVariable>(CE->getOperand(0))) { 1831 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1832 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1833 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1834 return GV->hasInitializer() && 1835 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1836 } 1837 return false; 1838} 1839 1840/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 1841/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 1842/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 1843static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 1844 ConstantExpr *Addr, unsigned OpNo) { 1845 // Base case of the recursion. 1846 if (OpNo == Addr->getNumOperands()) { 1847 assert(Val->getType() == Init->getType() && "Type mismatch!"); 1848 return Val; 1849 } 1850 1851 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1852 std::vector<Constant*> Elts; 1853 1854 // Break up the constant into its elements. 1855 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1856 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) 1857 Elts.push_back(CS->getOperand(i)); 1858 } else if (isa<ConstantAggregateZero>(Init)) { 1859 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1860 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 1861 } else if (isa<UndefValue>(Init)) { 1862 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1863 Elts.push_back(UndefValue::get(STy->getElementType(i))); 1864 } else { 1865 assert(0 && "This code is out of sync with " 1866 " ConstantFoldLoadThroughGEPConstantExpr"); 1867 } 1868 1869 // Replace the element that we are supposed to. 1870 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 1871 unsigned Idx = CU->getZExtValue(); 1872 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 1873 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 1874 1875 // Return the modified struct. 1876 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 1877 } else { 1878 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 1879 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 1880 1881 // Break up the array into elements. 1882 std::vector<Constant*> Elts; 1883 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1884 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) 1885 Elts.push_back(CA->getOperand(i)); 1886 } else if (isa<ConstantAggregateZero>(Init)) { 1887 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 1888 Elts.assign(ATy->getNumElements(), Elt); 1889 } else if (isa<UndefValue>(Init)) { 1890 Constant *Elt = UndefValue::get(ATy->getElementType()); 1891 Elts.assign(ATy->getNumElements(), Elt); 1892 } else { 1893 assert(0 && "This code is out of sync with " 1894 " ConstantFoldLoadThroughGEPConstantExpr"); 1895 } 1896 1897 assert(CI->getZExtValue() < ATy->getNumElements()); 1898 Elts[CI->getZExtValue()] = 1899 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 1900 return ConstantArray::get(ATy, Elts); 1901 } 1902} 1903 1904/// CommitValueTo - We have decided that Addr (which satisfies the predicate 1905/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 1906static void CommitValueTo(Constant *Val, Constant *Addr) { 1907 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 1908 assert(GV->hasInitializer()); 1909 GV->setInitializer(Val); 1910 return; 1911 } 1912 1913 ConstantExpr *CE = cast<ConstantExpr>(Addr); 1914 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1915 1916 Constant *Init = GV->getInitializer(); 1917 Init = EvaluateStoreInto(Init, Val, CE, 2); 1918 GV->setInitializer(Init); 1919} 1920 1921/// ComputeLoadResult - Return the value that would be computed by a load from 1922/// P after the stores reflected by 'memory' have been performed. If we can't 1923/// decide, return null. 1924static Constant *ComputeLoadResult(Constant *P, 1925 const std::map<Constant*, Constant*> &Memory) { 1926 // If this memory location has been recently stored, use the stored value: it 1927 // is the most up-to-date. 1928 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P); 1929 if (I != Memory.end()) return I->second; 1930 1931 // Access it. 1932 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 1933 if (GV->hasInitializer()) 1934 return GV->getInitializer(); 1935 return 0; 1936 } 1937 1938 // Handle a constantexpr getelementptr. 1939 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 1940 if (CE->getOpcode() == Instruction::GetElementPtr && 1941 isa<GlobalVariable>(CE->getOperand(0))) { 1942 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1943 if (GV->hasInitializer()) 1944 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1945 } 1946 1947 return 0; // don't know how to evaluate. 1948} 1949 1950/// EvaluateFunction - Evaluate a call to function F, returning true if 1951/// successful, false if we can't evaluate it. ActualArgs contains the formal 1952/// arguments for the function. 1953static bool EvaluateFunction(Function *F, Constant *&RetVal, 1954 const std::vector<Constant*> &ActualArgs, 1955 std::vector<Function*> &CallStack, 1956 std::map<Constant*, Constant*> &MutatedMemory, 1957 std::vector<GlobalVariable*> &AllocaTmps) { 1958 // Check to see if this function is already executing (recursion). If so, 1959 // bail out. TODO: we might want to accept limited recursion. 1960 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 1961 return false; 1962 1963 CallStack.push_back(F); 1964 1965 /// Values - As we compute SSA register values, we store their contents here. 1966 std::map<Value*, Constant*> Values; 1967 1968 // Initialize arguments to the incoming values specified. 1969 unsigned ArgNo = 0; 1970 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1971 ++AI, ++ArgNo) 1972 Values[AI] = ActualArgs[ArgNo]; 1973 1974 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 1975 /// we can only evaluate any one basic block at most once. This set keeps 1976 /// track of what we have executed so we can detect recursive cases etc. 1977 std::set<BasicBlock*> ExecutedBlocks; 1978 1979 // CurInst - The current instruction we're evaluating. 1980 BasicBlock::iterator CurInst = F->begin()->begin(); 1981 1982 // This is the main evaluation loop. 1983 while (1) { 1984 Constant *InstResult = 0; 1985 1986 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 1987 if (SI->isVolatile()) return false; // no volatile accesses. 1988 Constant *Ptr = getVal(Values, SI->getOperand(1)); 1989 if (!isSimpleEnoughPointerToCommit(Ptr)) 1990 // If this is too complex for us to commit, reject it. 1991 return false; 1992 Constant *Val = getVal(Values, SI->getOperand(0)); 1993 MutatedMemory[Ptr] = Val; 1994 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 1995 InstResult = ConstantExpr::get(BO->getOpcode(), 1996 getVal(Values, BO->getOperand(0)), 1997 getVal(Values, BO->getOperand(1))); 1998 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 1999 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2000 getVal(Values, CI->getOperand(0)), 2001 getVal(Values, CI->getOperand(1))); 2002 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2003 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2004 getVal(Values, CI->getOperand(0)), 2005 CI->getType()); 2006 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2007 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2008 getVal(Values, SI->getOperand(1)), 2009 getVal(Values, SI->getOperand(2))); 2010 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2011 Constant *P = getVal(Values, GEP->getOperand(0)); 2012 SmallVector<Constant*, 8> GEPOps; 2013 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i) 2014 GEPOps.push_back(getVal(Values, GEP->getOperand(i))); 2015 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2016 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2017 if (LI->isVolatile()) return false; // no volatile accesses. 2018 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2019 MutatedMemory); 2020 if (InstResult == 0) return false; // Could not evaluate load. 2021 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2022 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2023 const Type *Ty = AI->getType()->getElementType(); 2024 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2025 GlobalValue::InternalLinkage, 2026 UndefValue::get(Ty), 2027 AI->getName())); 2028 InstResult = AllocaTmps.back(); 2029 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2030 // Cannot handle inline asm. 2031 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2032 2033 // Resolve function pointers. 2034 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2035 if (!Callee) return false; // Cannot resolve. 2036 2037 std::vector<Constant*> Formals; 2038 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i) 2039 Formals.push_back(getVal(Values, CI->getOperand(i))); 2040 2041 if (Callee->isDeclaration()) { 2042 // If this is a function we can constant fold, do it. 2043 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2044 Formals.size())) { 2045 InstResult = C; 2046 } else { 2047 return false; 2048 } 2049 } else { 2050 if (Callee->getFunctionType()->isVarArg()) 2051 return false; 2052 2053 Constant *RetVal; 2054 2055 // Execute the call, if successful, use the return value. 2056 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2057 MutatedMemory, AllocaTmps)) 2058 return false; 2059 InstResult = RetVal; 2060 } 2061 } else if (isa<TerminatorInst>(CurInst)) { 2062 BasicBlock *NewBB = 0; 2063 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2064 if (BI->isUnconditional()) { 2065 NewBB = BI->getSuccessor(0); 2066 } else { 2067 ConstantInt *Cond = 2068 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2069 if (!Cond) return false; // Cannot determine. 2070 2071 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2072 } 2073 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2074 ConstantInt *Val = 2075 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2076 if (!Val) return false; // Cannot determine. 2077 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2078 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2079 if (RI->getNumOperands()) 2080 RetVal = getVal(Values, RI->getOperand(0)); 2081 2082 CallStack.pop_back(); // return from fn. 2083 return true; // We succeeded at evaluating this ctor! 2084 } else { 2085 // invoke, unwind, unreachable. 2086 return false; // Cannot handle this terminator. 2087 } 2088 2089 // Okay, we succeeded in evaluating this control flow. See if we have 2090 // executed the new block before. If so, we have a looping function, 2091 // which we cannot evaluate in reasonable time. 2092 if (!ExecutedBlocks.insert(NewBB).second) 2093 return false; // looped! 2094 2095 // Okay, we have never been in this block before. Check to see if there 2096 // are any PHI nodes. If so, evaluate them with information about where 2097 // we came from. 2098 BasicBlock *OldBB = CurInst->getParent(); 2099 CurInst = NewBB->begin(); 2100 PHINode *PN; 2101 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2102 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2103 2104 // Do NOT increment CurInst. We know that the terminator had no value. 2105 continue; 2106 } else { 2107 // Did not know how to evaluate this! 2108 return false; 2109 } 2110 2111 if (!CurInst->use_empty()) 2112 Values[CurInst] = InstResult; 2113 2114 // Advance program counter. 2115 ++CurInst; 2116 } 2117} 2118 2119/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2120/// we can. Return true if we can, false otherwise. 2121static bool EvaluateStaticConstructor(Function *F) { 2122 /// MutatedMemory - For each store we execute, we update this map. Loads 2123 /// check this to get the most up-to-date value. If evaluation is successful, 2124 /// this state is committed to the process. 2125 std::map<Constant*, Constant*> MutatedMemory; 2126 2127 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2128 /// to represent its body. This vector is needed so we can delete the 2129 /// temporary globals when we are done. 2130 std::vector<GlobalVariable*> AllocaTmps; 2131 2132 /// CallStack - This is used to detect recursion. In pathological situations 2133 /// we could hit exponential behavior, but at least there is nothing 2134 /// unbounded. 2135 std::vector<Function*> CallStack; 2136 2137 // Call the function. 2138 Constant *RetValDummy; 2139 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2140 CallStack, MutatedMemory, AllocaTmps); 2141 if (EvalSuccess) { 2142 // We succeeded at evaluation: commit the result. 2143 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2144 << F->getName() << "' to " << MutatedMemory.size() 2145 << " stores.\n"; 2146 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2147 E = MutatedMemory.end(); I != E; ++I) 2148 CommitValueTo(I->second, I->first); 2149 } 2150 2151 // At this point, we are done interpreting. If we created any 'alloca' 2152 // temporaries, release them now. 2153 while (!AllocaTmps.empty()) { 2154 GlobalVariable *Tmp = AllocaTmps.back(); 2155 AllocaTmps.pop_back(); 2156 2157 // If there are still users of the alloca, the program is doing something 2158 // silly, e.g. storing the address of the alloca somewhere and using it 2159 // later. Since this is undefined, we'll just make it be null. 2160 if (!Tmp->use_empty()) 2161 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2162 delete Tmp; 2163 } 2164 2165 return EvalSuccess; 2166} 2167 2168 2169 2170/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2171/// Return true if anything changed. 2172bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2173 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2174 bool MadeChange = false; 2175 if (Ctors.empty()) return false; 2176 2177 // Loop over global ctors, optimizing them when we can. 2178 for (unsigned i = 0; i != Ctors.size(); ++i) { 2179 Function *F = Ctors[i]; 2180 // Found a null terminator in the middle of the list, prune off the rest of 2181 // the list. 2182 if (F == 0) { 2183 if (i != Ctors.size()-1) { 2184 Ctors.resize(i+1); 2185 MadeChange = true; 2186 } 2187 break; 2188 } 2189 2190 // We cannot simplify external ctor functions. 2191 if (F->empty()) continue; 2192 2193 // If we can evaluate the ctor at compile time, do. 2194 if (EvaluateStaticConstructor(F)) { 2195 Ctors.erase(Ctors.begin()+i); 2196 MadeChange = true; 2197 --i; 2198 ++NumCtorsEvaluated; 2199 continue; 2200 } 2201 } 2202 2203 if (!MadeChange) return false; 2204 2205 GCL = InstallGlobalCtors(GCL, Ctors); 2206 return true; 2207} 2208 2209 2210bool GlobalOpt::runOnModule(Module &M) { 2211 bool Changed = false; 2212 2213 // Try to find the llvm.globalctors list. 2214 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2215 2216 bool LocalChange = true; 2217 while (LocalChange) { 2218 LocalChange = false; 2219 2220 // Delete functions that are trivially dead, ccc -> fastcc 2221 LocalChange |= OptimizeFunctions(M); 2222 2223 // Optimize global_ctors list. 2224 if (GlobalCtors) 2225 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2226 2227 // Optimize non-address-taken globals. 2228 LocalChange |= OptimizeGlobalVars(M); 2229 Changed |= LocalChange; 2230 } 2231 2232 // TODO: Move all global ctors functions to the end of the module for code 2233 // layout. 2234 2235 return Changed; 2236} 2237