GlobalOpt.cpp revision 26f8c27c34b44f7d87de74d1de2128c1a02855bf
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 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 GV->getType()->getAddressSpace()); 491 Globals.insert(GV, NGV); 492 NewGlobals.push_back(NGV); 493 494 // Calculate the known alignment of the field. If the original aggregate 495 // had 256 byte alignment for example, something might depend on that: 496 // propagate info to each field. 497 uint64_t FieldOffset = Layout.getElementOffset(i); 498 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 499 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 500 NGV->setAlignment(NewAlign); 501 } 502 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 503 unsigned NumElements = 0; 504 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 505 NumElements = ATy->getNumElements(); 506 else 507 NumElements = cast<VectorType>(STy)->getNumElements(); 508 509 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 510 return 0; // It's not worth it. 511 NewGlobals.reserve(NumElements); 512 513 uint64_t EltSize = TD.getABITypeSize(STy->getElementType()); 514 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 515 for (unsigned i = 0, e = NumElements; i != e; ++i) { 516 Constant *In = getAggregateConstantElement(Init, 517 ConstantInt::get(Type::Int32Ty, i)); 518 assert(In && "Couldn't get element of initializer?"); 519 520 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 521 GlobalVariable::InternalLinkage, 522 In, GV->getName()+"."+utostr(i), 523 (Module *)NULL, 524 GV->isThreadLocal(), 525 GV->getType()->getAddressSpace()); 526 Globals.insert(GV, NGV); 527 NewGlobals.push_back(NGV); 528 529 // Calculate the known alignment of the field. If the original aggregate 530 // had 256 byte alignment for example, something might depend on that: 531 // propagate info to each field. 532 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 533 if (NewAlign > EltAlign) 534 NGV->setAlignment(NewAlign); 535 } 536 } 537 538 if (NewGlobals.empty()) 539 return 0; 540 541 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 542 543 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 544 545 // Loop over all of the uses of the global, replacing the constantexpr geps, 546 // with smaller constantexpr geps or direct references. 547 while (!GV->use_empty()) { 548 User *GEP = GV->use_back(); 549 assert(((isa<ConstantExpr>(GEP) && 550 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 551 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 552 553 // Ignore the 1th operand, which has to be zero or else the program is quite 554 // broken (undefined). Get the 2nd operand, which is the structure or array 555 // index. 556 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 557 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 558 559 Value *NewPtr = NewGlobals[Val]; 560 561 // Form a shorter GEP if needed. 562 if (GEP->getNumOperands() > 3) { 563 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 564 SmallVector<Constant*, 8> Idxs; 565 Idxs.push_back(NullInt); 566 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 567 Idxs.push_back(CE->getOperand(i)); 568 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 569 &Idxs[0], Idxs.size()); 570 } else { 571 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 572 SmallVector<Value*, 8> Idxs; 573 Idxs.push_back(NullInt); 574 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 575 Idxs.push_back(GEPI->getOperand(i)); 576 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 577 GEPI->getName()+"."+utostr(Val), GEPI); 578 } 579 } 580 GEP->replaceAllUsesWith(NewPtr); 581 582 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 583 GEPI->eraseFromParent(); 584 else 585 cast<ConstantExpr>(GEP)->destroyConstant(); 586 } 587 588 // Delete the old global, now that it is dead. 589 Globals.erase(GV); 590 ++NumSRA; 591 592 // Loop over the new globals array deleting any globals that are obviously 593 // dead. This can arise due to scalarization of a structure or an array that 594 // has elements that are dead. 595 unsigned FirstGlobal = 0; 596 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 597 if (NewGlobals[i]->use_empty()) { 598 Globals.erase(NewGlobals[i]); 599 if (FirstGlobal == i) ++FirstGlobal; 600 } 601 602 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 603} 604 605/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 606/// value will trap if the value is dynamically null. PHIs keeps track of any 607/// phi nodes we've seen to avoid reprocessing them. 608static bool AllUsesOfValueWillTrapIfNull(Value *V, 609 SmallPtrSet<PHINode*, 8> &PHIs) { 610 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 611 if (isa<LoadInst>(*UI)) { 612 // Will trap. 613 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 614 if (SI->getOperand(0) == V) { 615 //cerr << "NONTRAPPING USE: " << **UI; 616 return false; // Storing the value. 617 } 618 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 619 if (CI->getOperand(0) != V) { 620 //cerr << "NONTRAPPING USE: " << **UI; 621 return false; // Not calling the ptr 622 } 623 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 624 if (II->getOperand(0) != V) { 625 //cerr << "NONTRAPPING USE: " << **UI; 626 return false; // Not calling the ptr 627 } 628 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 629 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 630 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 631 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 632 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 633 // If we've already seen this phi node, ignore it, it has already been 634 // checked. 635 if (PHIs.insert(PN)) 636 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 637 } else if (isa<ICmpInst>(*UI) && 638 isa<ConstantPointerNull>(UI->getOperand(1))) { 639 // Ignore setcc X, null 640 } else { 641 //cerr << "NONTRAPPING USE: " << **UI; 642 return false; 643 } 644 return true; 645} 646 647/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 648/// from GV will trap if the loaded value is null. Note that this also permits 649/// comparisons of the loaded value against null, as a special case. 650static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 651 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 652 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 653 SmallPtrSet<PHINode*, 8> PHIs; 654 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 655 return false; 656 } else if (isa<StoreInst>(*UI)) { 657 // Ignore stores to the global. 658 } else { 659 // We don't know or understand this user, bail out. 660 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 661 return false; 662 } 663 664 return true; 665} 666 667static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 668 bool Changed = false; 669 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 670 Instruction *I = cast<Instruction>(*UI++); 671 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 672 LI->setOperand(0, NewV); 673 Changed = true; 674 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 675 if (SI->getOperand(1) == V) { 676 SI->setOperand(1, NewV); 677 Changed = true; 678 } 679 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 680 if (I->getOperand(0) == V) { 681 // Calling through the pointer! Turn into a direct call, but be careful 682 // that the pointer is not also being passed as an argument. 683 I->setOperand(0, NewV); 684 Changed = true; 685 bool PassedAsArg = false; 686 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 687 if (I->getOperand(i) == V) { 688 PassedAsArg = true; 689 I->setOperand(i, NewV); 690 } 691 692 if (PassedAsArg) { 693 // Being passed as an argument also. Be careful to not invalidate UI! 694 UI = V->use_begin(); 695 } 696 } 697 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 698 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 699 ConstantExpr::getCast(CI->getOpcode(), 700 NewV, CI->getType())); 701 if (CI->use_empty()) { 702 Changed = true; 703 CI->eraseFromParent(); 704 } 705 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 706 // Should handle GEP here. 707 SmallVector<Constant*, 8> Idxs; 708 Idxs.reserve(GEPI->getNumOperands()-1); 709 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 710 i != e; ++i) 711 if (Constant *C = dyn_cast<Constant>(*i)) 712 Idxs.push_back(C); 713 else 714 break; 715 if (Idxs.size() == GEPI->getNumOperands()-1) 716 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 717 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 718 Idxs.size())); 719 if (GEPI->use_empty()) { 720 Changed = true; 721 GEPI->eraseFromParent(); 722 } 723 } 724 } 725 726 return Changed; 727} 728 729 730/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 731/// value stored into it. If there are uses of the loaded value that would trap 732/// if the loaded value is dynamically null, then we know that they cannot be 733/// reachable with a null optimize away the load. 734static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 735 std::vector<LoadInst*> Loads; 736 bool Changed = false; 737 738 // Replace all uses of loads with uses of uses of the stored value. 739 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); 740 GUI != E; ++GUI) 741 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) { 742 Loads.push_back(LI); 743 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 744 } else { 745 // If we get here we could have stores, selects, or phi nodes whose values 746 // are loaded. 747 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) || 748 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) && 749 "Only expect load and stores!"); 750 } 751 752 if (Changed) { 753 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 754 ++NumGlobUses; 755 } 756 757 // Delete all of the loads we can, keeping track of whether we nuked them all! 758 bool AllLoadsGone = true; 759 while (!Loads.empty()) { 760 LoadInst *L = Loads.back(); 761 if (L->use_empty()) { 762 L->eraseFromParent(); 763 Changed = true; 764 } else { 765 AllLoadsGone = false; 766 } 767 Loads.pop_back(); 768 } 769 770 // If we nuked all of the loads, then none of the stores are needed either, 771 // nor is the global. 772 if (AllLoadsGone) { 773 DOUT << " *** GLOBAL NOW DEAD!\n"; 774 CleanupConstantGlobalUsers(GV, 0); 775 if (GV->use_empty()) { 776 GV->eraseFromParent(); 777 ++NumDeleted; 778 } 779 Changed = true; 780 } 781 return Changed; 782} 783 784/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 785/// instructions that are foldable. 786static void ConstantPropUsersOf(Value *V) { 787 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 788 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 789 if (Constant *NewC = ConstantFoldInstruction(I)) { 790 I->replaceAllUsesWith(NewC); 791 792 // Advance UI to the next non-I use to avoid invalidating it! 793 // Instructions could multiply use V. 794 while (UI != E && *UI == I) 795 ++UI; 796 I->eraseFromParent(); 797 } 798} 799 800/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 801/// variable, and transforms the program as if it always contained the result of 802/// the specified malloc. Because it is always the result of the specified 803/// malloc, there is no reason to actually DO the malloc. Instead, turn the 804/// malloc into a global, and any loads of GV as uses of the new global. 805static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 806 MallocInst *MI) { 807 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 808 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 809 810 if (NElements->getZExtValue() != 1) { 811 // If we have an array allocation, transform it to a single element 812 // allocation to make the code below simpler. 813 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 814 NElements->getZExtValue()); 815 MallocInst *NewMI = 816 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 817 MI->getAlignment(), MI->getName(), MI); 818 Value* Indices[2]; 819 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 820 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2, 821 NewMI->getName()+".el0", MI); 822 MI->replaceAllUsesWith(NewGEP); 823 MI->eraseFromParent(); 824 MI = NewMI; 825 } 826 827 // Create the new global variable. The contents of the malloc'd memory is 828 // undefined, so initialize with an undef value. 829 Constant *Init = UndefValue::get(MI->getAllocatedType()); 830 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 831 GlobalValue::InternalLinkage, Init, 832 GV->getName()+".body", 833 (Module *)NULL, 834 GV->isThreadLocal()); 835 // FIXME: This new global should have the alignment returned by malloc. Code 836 // could depend on malloc returning large alignment (on the mac, 16 bytes) but 837 // this would only guarantee some lower alignment. 838 GV->getParent()->getGlobalList().insert(GV, NewGV); 839 840 // Anything that used the malloc now uses the global directly. 841 MI->replaceAllUsesWith(NewGV); 842 843 Constant *RepValue = NewGV; 844 if (NewGV->getType() != GV->getType()->getElementType()) 845 RepValue = ConstantExpr::getBitCast(RepValue, 846 GV->getType()->getElementType()); 847 848 // If there is a comparison against null, we will insert a global bool to 849 // keep track of whether the global was initialized yet or not. 850 GlobalVariable *InitBool = 851 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 852 ConstantInt::getFalse(), GV->getName()+".init", 853 (Module *)NULL, GV->isThreadLocal()); 854 bool InitBoolUsed = false; 855 856 // Loop over all uses of GV, processing them in turn. 857 std::vector<StoreInst*> Stores; 858 while (!GV->use_empty()) 859 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 860 while (!LI->use_empty()) { 861 Use &LoadUse = LI->use_begin().getUse(); 862 if (!isa<ICmpInst>(LoadUse.getUser())) 863 LoadUse = RepValue; 864 else { 865 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 866 // Replace the cmp X, 0 with a use of the bool value. 867 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 868 InitBoolUsed = true; 869 switch (CI->getPredicate()) { 870 default: assert(0 && "Unknown ICmp Predicate!"); 871 case ICmpInst::ICMP_ULT: 872 case ICmpInst::ICMP_SLT: 873 LV = ConstantInt::getFalse(); // X < null -> always false 874 break; 875 case ICmpInst::ICMP_ULE: 876 case ICmpInst::ICMP_SLE: 877 case ICmpInst::ICMP_EQ: 878 LV = BinaryOperator::CreateNot(LV, "notinit", CI); 879 break; 880 case ICmpInst::ICMP_NE: 881 case ICmpInst::ICMP_UGE: 882 case ICmpInst::ICMP_SGE: 883 case ICmpInst::ICMP_UGT: 884 case ICmpInst::ICMP_SGT: 885 break; // no change. 886 } 887 CI->replaceAllUsesWith(LV); 888 CI->eraseFromParent(); 889 } 890 } 891 LI->eraseFromParent(); 892 } else { 893 StoreInst *SI = cast<StoreInst>(GV->use_back()); 894 // The global is initialized when the store to it occurs. 895 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 896 SI->eraseFromParent(); 897 } 898 899 // If the initialization boolean was used, insert it, otherwise delete it. 900 if (!InitBoolUsed) { 901 while (!InitBool->use_empty()) // Delete initializations 902 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 903 delete InitBool; 904 } else 905 GV->getParent()->getGlobalList().insert(GV, InitBool); 906 907 908 // Now the GV is dead, nuke it and the malloc. 909 GV->eraseFromParent(); 910 MI->eraseFromParent(); 911 912 // To further other optimizations, loop over all users of NewGV and try to 913 // constant prop them. This will promote GEP instructions with constant 914 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 915 ConstantPropUsersOf(NewGV); 916 if (RepValue != NewGV) 917 ConstantPropUsersOf(RepValue); 918 919 return NewGV; 920} 921 922/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 923/// to make sure that there are no complex uses of V. We permit simple things 924/// like dereferencing the pointer, but not storing through the address, unless 925/// it is to the specified global. 926static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 927 GlobalVariable *GV, 928 SmallPtrSet<PHINode*, 8> &PHIs) { 929 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 930 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) { 931 // Fine, ignore. 932 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 933 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 934 return false; // Storing the pointer itself... bad. 935 // Otherwise, storing through it, or storing into GV... fine. 936 } else if (isa<GetElementPtrInst>(*UI)) { 937 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI), 938 GV, PHIs)) 939 return false; 940 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 941 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 942 // cycles. 943 if (PHIs.insert(PN)) 944 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 945 return false; 946 } else { 947 return false; 948 } 949 return true; 950} 951 952/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 953/// somewhere. Transform all uses of the allocation into loads from the 954/// global and uses of the resultant pointer. Further, delete the store into 955/// GV. This assumes that these value pass the 956/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 957static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 958 GlobalVariable *GV) { 959 while (!Alloc->use_empty()) { 960 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 961 Instruction *InsertPt = U; 962 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 963 // If this is the store of the allocation into the global, remove it. 964 if (SI->getOperand(1) == GV) { 965 SI->eraseFromParent(); 966 continue; 967 } 968 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 969 // Insert the load in the corresponding predecessor, not right before the 970 // PHI. 971 unsigned PredNo = Alloc->use_begin().getOperandNo()/2; 972 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator(); 973 } 974 975 // Insert a load from the global, and use it instead of the malloc. 976 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 977 U->replaceUsesOfWith(Alloc, NL); 978 } 979} 980 981/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 982/// GV are simple enough to perform HeapSRA, return true. 983static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 984 MallocInst *MI) { 985 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 986 ++UI) 987 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 988 // We permit two users of the load: setcc comparing against the null 989 // pointer, and a getelementptr of a specific form. 990 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); 991 UI != E; ++UI) { 992 // Comparison against null is ok. 993 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) { 994 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 995 return false; 996 continue; 997 } 998 999 // getelementptr is also ok, but only a simple form. 1000 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 1001 // Must index into the array and into the struct. 1002 if (GEPI->getNumOperands() < 3) 1003 return false; 1004 1005 // Otherwise the GEP is ok. 1006 continue; 1007 } 1008 1009 if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 1010 // We have a phi of a load from the global. We can only handle this 1011 // if the other PHI'd values are actually the same. In this case, 1012 // the rewriter will just drop the phi entirely. 1013 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1014 Value *IV = PN->getIncomingValue(i); 1015 if (IV == LI) continue; // Trivial the same. 1016 1017 // If the phi'd value is from the malloc that initializes the value, 1018 // we can xform it. 1019 if (IV == MI) continue; 1020 1021 // Otherwise, we don't know what it is. 1022 return false; 1023 } 1024 return true; 1025 } 1026 1027 // Otherwise we don't know what this is, not ok. 1028 return false; 1029 } 1030 } 1031 return true; 1032} 1033 1034/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd 1035/// value, lazily creating it on demand. 1036static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo, 1037 const std::vector<GlobalVariable*> &FieldGlobals, 1038 std::vector<Value *> &InsertedLoadsForPtr) { 1039 if (InsertedLoadsForPtr.size() <= FieldNo) 1040 InsertedLoadsForPtr.resize(FieldNo+1); 1041 if (InsertedLoadsForPtr[FieldNo] == 0) 1042 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo], 1043 Load->getName()+".f" + 1044 utostr(FieldNo), Load); 1045 return InsertedLoadsForPtr[FieldNo]; 1046} 1047 1048/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1049/// the load, rewrite the derived value to use the HeapSRoA'd load. 1050static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser, 1051 const std::vector<GlobalVariable*> &FieldGlobals, 1052 std::vector<Value *> &InsertedLoadsForPtr) { 1053 // If this is a comparison against null, handle it. 1054 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1055 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1056 // If we have a setcc of the loaded pointer, we can use a setcc of any 1057 // field. 1058 Value *NPtr; 1059 if (InsertedLoadsForPtr.empty()) { 1060 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr); 1061 } else { 1062 NPtr = InsertedLoadsForPtr.back(); 1063 } 1064 1065 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1066 Constant::getNullValue(NPtr->getType()), 1067 SCI->getName(), SCI); 1068 SCI->replaceAllUsesWith(New); 1069 SCI->eraseFromParent(); 1070 return; 1071 } 1072 1073 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...' 1074 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1075 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1076 && "Unexpected GEPI!"); 1077 1078 // Load the pointer for this field. 1079 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1080 Value *NewPtr = GetHeapSROALoad(Load, FieldNo, 1081 FieldGlobals, InsertedLoadsForPtr); 1082 1083 // Create the new GEP idx vector. 1084 SmallVector<Value*, 8> GEPIdx; 1085 GEPIdx.push_back(GEPI->getOperand(1)); 1086 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1087 1088 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1089 GEPIdx.begin(), GEPIdx.end(), 1090 GEPI->getName(), GEPI); 1091 GEPI->replaceAllUsesWith(NGEPI); 1092 GEPI->eraseFromParent(); 1093 return; 1094 } 1095 1096 // Handle PHI nodes. PHI nodes must be merging in the same values, plus 1097 // potentially the original malloc. Insert phi nodes for each field, then 1098 // process uses of the PHI. 1099 PHINode *PN = cast<PHINode>(LoadUser); 1100 std::vector<Value *> PHIsForField; 1101 PHIsForField.resize(FieldGlobals.size()); 1102 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1103 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr); 1104 1105 PHINode *FieldPN = PHINode::Create(LoadV->getType(), 1106 PN->getName()+"."+utostr(i), PN); 1107 // Fill in the predecessor values. 1108 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) { 1109 // Each predecessor either uses the load or the original malloc. 1110 Value *InVal = PN->getIncomingValue(pred); 1111 BasicBlock *BB = PN->getIncomingBlock(pred); 1112 Value *NewVal; 1113 if (isa<MallocInst>(InVal)) { 1114 // Insert a reload from the global in the predecessor. 1115 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals, 1116 PHIsForField); 1117 } else { 1118 NewVal = InsertedLoadsForPtr[i]; 1119 } 1120 FieldPN->addIncoming(NewVal, BB); 1121 } 1122 PHIsForField[i] = FieldPN; 1123 } 1124 1125 // Since PHIsForField specifies a phi for every input value, the lazy inserter 1126 // will never insert a load. 1127 while (!PN->use_empty()) 1128 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField); 1129 PN->eraseFromParent(); 1130} 1131 1132/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1133/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1134/// use FieldGlobals instead. All uses of loaded values satisfy 1135/// GlobalLoadUsesSimpleEnoughForHeapSRA. 1136static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1137 const std::vector<GlobalVariable*> &FieldGlobals) { 1138 std::vector<Value *> InsertedLoadsForPtr; 1139 //InsertedLoadsForPtr.resize(FieldGlobals.size()); 1140 while (!Load->use_empty()) 1141 RewriteHeapSROALoadUser(Load, Load->use_back(), 1142 FieldGlobals, InsertedLoadsForPtr); 1143} 1144 1145/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1146/// it up into multiple allocations of arrays of the fields. 1147static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1148 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1149 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1150 1151 // There is guaranteed to be at least one use of the malloc (storing 1152 // it into GV). If there are other uses, change them to be uses of 1153 // the global to simplify later code. This also deletes the store 1154 // into GV. 1155 ReplaceUsesOfMallocWithGlobal(MI, GV); 1156 1157 // Okay, at this point, there are no users of the malloc. Insert N 1158 // new mallocs at the same place as MI, and N globals. 1159 std::vector<GlobalVariable*> FieldGlobals; 1160 std::vector<MallocInst*> FieldMallocs; 1161 1162 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1163 const Type *FieldTy = STy->getElementType(FieldNo); 1164 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1165 1166 GlobalVariable *NGV = 1167 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1168 Constant::getNullValue(PFieldTy), 1169 GV->getName() + ".f" + utostr(FieldNo), GV, 1170 GV->isThreadLocal()); 1171 FieldGlobals.push_back(NGV); 1172 1173 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1174 MI->getName() + ".f" + utostr(FieldNo),MI); 1175 FieldMallocs.push_back(NMI); 1176 new StoreInst(NMI, NGV, MI); 1177 } 1178 1179 // The tricky aspect of this transformation is handling the case when malloc 1180 // fails. In the original code, malloc failing would set the result pointer 1181 // of malloc to null. In this case, some mallocs could succeed and others 1182 // could fail. As such, we emit code that looks like this: 1183 // F0 = malloc(field0) 1184 // F1 = malloc(field1) 1185 // F2 = malloc(field2) 1186 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1187 // if (F0) { free(F0); F0 = 0; } 1188 // if (F1) { free(F1); F1 = 0; } 1189 // if (F2) { free(F2); F2 = 0; } 1190 // } 1191 Value *RunningOr = 0; 1192 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1193 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1194 Constant::getNullValue(FieldMallocs[i]->getType()), 1195 "isnull", MI); 1196 if (!RunningOr) 1197 RunningOr = Cond; // First seteq 1198 else 1199 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI); 1200 } 1201 1202 // Split the basic block at the old malloc. 1203 BasicBlock *OrigBB = MI->getParent(); 1204 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1205 1206 // Create the block to check the first condition. Put all these blocks at the 1207 // end of the function as they are unlikely to be executed. 1208 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1209 OrigBB->getParent()); 1210 1211 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1212 // branch on RunningOr. 1213 OrigBB->getTerminator()->eraseFromParent(); 1214 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1215 1216 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1217 // pointer, because some may be null while others are not. 1218 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1219 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1220 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1221 Constant::getNullValue(GVVal->getType()), 1222 "tmp", NullPtrBlock); 1223 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1224 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1225 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1226 1227 // Fill in FreeBlock. 1228 new FreeInst(GVVal, FreeBlock); 1229 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1230 FreeBlock); 1231 BranchInst::Create(NextBlock, FreeBlock); 1232 1233 NullPtrBlock = NextBlock; 1234 } 1235 1236 BranchInst::Create(ContBB, NullPtrBlock); 1237 1238 // MI is no longer needed, remove it. 1239 MI->eraseFromParent(); 1240 1241 1242 // Okay, the malloc site is completely handled. All of the uses of GV are now 1243 // loads, and all uses of those loads are simple. Rewrite them to use loads 1244 // of the per-field globals instead. 1245 while (!GV->use_empty()) { 1246 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 1247 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals); 1248 LI->eraseFromParent(); 1249 } else { 1250 // Must be a store of null. 1251 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1252 assert(isa<Constant>(SI->getOperand(0)) && 1253 cast<Constant>(SI->getOperand(0))->isNullValue() && 1254 "Unexpected heap-sra user!"); 1255 1256 // Insert a store of null into each global. 1257 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1258 Constant *Null = 1259 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType()); 1260 new StoreInst(Null, FieldGlobals[i], SI); 1261 } 1262 // Erase the original store. 1263 SI->eraseFromParent(); 1264 } 1265 } 1266 1267 // The old global is now dead, remove it. 1268 GV->eraseFromParent(); 1269 1270 ++NumHeapSRA; 1271 return FieldGlobals[0]; 1272} 1273 1274 1275// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1276// that only one value (besides its initializer) is ever stored to the global. 1277static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1278 Module::global_iterator &GVI, 1279 TargetData &TD) { 1280 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal)) 1281 StoredOnceVal = CI->getOperand(0); 1282 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){ 1283 // "getelementptr Ptr, 0, 0, 0" is really just a cast. 1284 bool IsJustACast = true; 1285 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 1286 i != e; ++i) 1287 if (!isa<Constant>(*i) || 1288 !cast<Constant>(*i)->isNullValue()) { 1289 IsJustACast = false; 1290 break; 1291 } 1292 if (IsJustACast) 1293 StoredOnceVal = GEPI->getOperand(0); 1294 } 1295 1296 // If we are dealing with a pointer global that is initialized to null and 1297 // only has one (non-null) value stored into it, then we can optimize any 1298 // users of the loaded value (often calls and loads) that would trap if the 1299 // value was null. 1300 if (isa<PointerType>(GV->getInitializer()->getType()) && 1301 GV->getInitializer()->isNullValue()) { 1302 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1303 if (GV->getInitializer()->getType() != SOVC->getType()) 1304 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1305 1306 // Optimize away any trapping uses of the loaded value. 1307 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1308 return true; 1309 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1310 // If this is a malloc of an abstract type, don't touch it. 1311 if (!MI->getAllocatedType()->isSized()) 1312 return false; 1313 1314 // We can't optimize this global unless all uses of it are *known* to be 1315 // of the malloc value, not of the null initializer value (consider a use 1316 // that compares the global's value against zero to see if the malloc has 1317 // been reached). To do this, we check to see if all uses of the global 1318 // would trap if the global were null: this proves that they must all 1319 // happen after the malloc. 1320 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1321 return false; 1322 1323 // We can't optimize this if the malloc itself is used in a complex way, 1324 // for example, being stored into multiple globals. This allows the 1325 // malloc to be stored into the specified global, loaded setcc'd, and 1326 // GEP'd. These are all things we could transform to using the global 1327 // for. 1328 { 1329 SmallPtrSet<PHINode*, 8> PHIs; 1330 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1331 return false; 1332 } 1333 1334 1335 // If we have a global that is only initialized with a fixed size malloc, 1336 // transform the program to use global memory instead of malloc'd memory. 1337 // This eliminates dynamic allocation, avoids an indirection accessing the 1338 // data, and exposes the resultant global to further GlobalOpt. 1339 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1340 // Restrict this transformation to only working on small allocations 1341 // (2048 bytes currently), as we don't want to introduce a 16M global or 1342 // something. 1343 if (NElements->getZExtValue()* 1344 TD.getABITypeSize(MI->getAllocatedType()) < 2048) { 1345 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1346 return true; 1347 } 1348 } 1349 1350 // If the allocation is an array of structures, consider transforming this 1351 // into multiple malloc'd arrays, one for each field. This is basically 1352 // SRoA for malloc'd memory. 1353 if (const StructType *AllocTy = 1354 dyn_cast<StructType>(MI->getAllocatedType())) { 1355 // This the structure has an unreasonable number of fields, leave it 1356 // alone. 1357 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 && 1358 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1359 GVI = PerformHeapAllocSRoA(GV, MI); 1360 return true; 1361 } 1362 } 1363 } 1364 } 1365 1366 return false; 1367} 1368 1369/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1370/// two values ever stored into GV are its initializer and OtherVal. See if we 1371/// can shrink the global into a boolean and select between the two values 1372/// whenever it is used. This exposes the values to other scalar optimizations. 1373static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1374 const Type *GVElType = GV->getType()->getElementType(); 1375 1376 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1377 // an FP value or vector, don't do this optimization because a select between 1378 // them is very expensive and unlikely to lead to later simplification. 1379 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1380 isa<VectorType>(GVElType)) 1381 return false; 1382 1383 // Walk the use list of the global seeing if all the uses are load or store. 1384 // If there is anything else, bail out. 1385 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1386 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1387 return false; 1388 1389 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1390 1391 // Create the new global, initializing it to false. 1392 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1393 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1394 GV->getName()+".b", 1395 (Module *)NULL, 1396 GV->isThreadLocal()); 1397 GV->getParent()->getGlobalList().insert(GV, NewGV); 1398 1399 Constant *InitVal = GV->getInitializer(); 1400 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1401 1402 // If initialized to zero and storing one into the global, we can use a cast 1403 // instead of a select to synthesize the desired value. 1404 bool IsOneZero = false; 1405 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1406 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1407 1408 while (!GV->use_empty()) { 1409 Instruction *UI = cast<Instruction>(GV->use_back()); 1410 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1411 // Change the store into a boolean store. 1412 bool StoringOther = SI->getOperand(0) == OtherVal; 1413 // Only do this if we weren't storing a loaded value. 1414 Value *StoreVal; 1415 if (StoringOther || SI->getOperand(0) == InitVal) 1416 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1417 else { 1418 // Otherwise, we are storing a previously loaded copy. To do this, 1419 // change the copy from copying the original value to just copying the 1420 // bool. 1421 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1422 1423 // If we're already replaced the input, StoredVal will be a cast or 1424 // select instruction. If not, it will be a load of the original 1425 // global. 1426 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1427 assert(LI->getOperand(0) == GV && "Not a copy!"); 1428 // Insert a new load, to preserve the saved value. 1429 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1430 } else { 1431 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1432 "This is not a form that we understand!"); 1433 StoreVal = StoredVal->getOperand(0); 1434 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1435 } 1436 } 1437 new StoreInst(StoreVal, NewGV, SI); 1438 } else { 1439 // Change the load into a load of bool then a select. 1440 LoadInst *LI = cast<LoadInst>(UI); 1441 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1442 Value *NSI; 1443 if (IsOneZero) 1444 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1445 else 1446 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1447 NSI->takeName(LI); 1448 LI->replaceAllUsesWith(NSI); 1449 } 1450 UI->eraseFromParent(); 1451 } 1452 1453 GV->eraseFromParent(); 1454 return true; 1455} 1456 1457 1458/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1459/// it if possible. If we make a change, return true. 1460bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1461 Module::global_iterator &GVI) { 1462 std::set<PHINode*> PHIUsers; 1463 GlobalStatus GS; 1464 GV->removeDeadConstantUsers(); 1465 1466 if (GV->use_empty()) { 1467 DOUT << "GLOBAL DEAD: " << *GV; 1468 GV->eraseFromParent(); 1469 ++NumDeleted; 1470 return true; 1471 } 1472 1473 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1474#if 0 1475 cerr << "Global: " << *GV; 1476 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1477 cerr << " StoredType = "; 1478 switch (GS.StoredType) { 1479 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1480 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1481 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1482 case GlobalStatus::isStored: cerr << "stored\n"; break; 1483 } 1484 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1485 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1486 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1487 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1488 << "\n"; 1489 cerr << " HasMultipleAccessingFunctions = " 1490 << GS.HasMultipleAccessingFunctions << "\n"; 1491 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1492 cerr << "\n"; 1493#endif 1494 1495 // If this is a first class global and has only one accessing function 1496 // and this function is main (which we know is not recursive we can make 1497 // this global a local variable) we replace the global with a local alloca 1498 // in this function. 1499 // 1500 // NOTE: It doesn't make sense to promote non single-value types since we 1501 // are just replacing static memory to stack memory. 1502 if (!GS.HasMultipleAccessingFunctions && 1503 GS.AccessingFunction && !GS.HasNonInstructionUser && 1504 GV->getType()->getElementType()->isSingleValueType() && 1505 GS.AccessingFunction->getName() == "main" && 1506 GS.AccessingFunction->hasExternalLinkage()) { 1507 DOUT << "LOCALIZING GLOBAL: " << *GV; 1508 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1509 const Type* ElemTy = GV->getType()->getElementType(); 1510 // FIXME: Pass Global's alignment when globals have alignment 1511 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1512 if (!isa<UndefValue>(GV->getInitializer())) 1513 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1514 1515 GV->replaceAllUsesWith(Alloca); 1516 GV->eraseFromParent(); 1517 ++NumLocalized; 1518 return true; 1519 } 1520 1521 // If the global is never loaded (but may be stored to), it is dead. 1522 // Delete it now. 1523 if (!GS.isLoaded) { 1524 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1525 1526 // Delete any stores we can find to the global. We may not be able to 1527 // make it completely dead though. 1528 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1529 1530 // If the global is dead now, delete it. 1531 if (GV->use_empty()) { 1532 GV->eraseFromParent(); 1533 ++NumDeleted; 1534 Changed = true; 1535 } 1536 return Changed; 1537 1538 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1539 DOUT << "MARKING CONSTANT: " << *GV; 1540 GV->setConstant(true); 1541 1542 // Clean up any obviously simplifiable users now. 1543 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1544 1545 // If the global is dead now, just nuke it. 1546 if (GV->use_empty()) { 1547 DOUT << " *** Marking constant allowed us to simplify " 1548 << "all users and delete global!\n"; 1549 GV->eraseFromParent(); 1550 ++NumDeleted; 1551 } 1552 1553 ++NumMarked; 1554 return true; 1555 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1556 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1557 getAnalysis<TargetData>())) { 1558 GVI = FirstNewGV; // Don't skip the newly produced globals! 1559 return true; 1560 } 1561 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1562 // If the initial value for the global was an undef value, and if only 1563 // one other value was stored into it, we can just change the 1564 // initializer to be an undef value, then delete all stores to the 1565 // global. This allows us to mark it constant. 1566 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1567 if (isa<UndefValue>(GV->getInitializer())) { 1568 // Change the initial value here. 1569 GV->setInitializer(SOVConstant); 1570 1571 // Clean up any obviously simplifiable users now. 1572 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1573 1574 if (GV->use_empty()) { 1575 DOUT << " *** Substituting initializer allowed us to " 1576 << "simplify all users and delete global!\n"; 1577 GV->eraseFromParent(); 1578 ++NumDeleted; 1579 } else { 1580 GVI = GV; 1581 } 1582 ++NumSubstitute; 1583 return true; 1584 } 1585 1586 // Try to optimize globals based on the knowledge that only one value 1587 // (besides its initializer) is ever stored to the global. 1588 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1589 getAnalysis<TargetData>())) 1590 return true; 1591 1592 // Otherwise, if the global was not a boolean, we can shrink it to be a 1593 // boolean. 1594 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1595 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1596 ++NumShrunkToBool; 1597 return true; 1598 } 1599 } 1600 } 1601 return false; 1602} 1603 1604/// OnlyCalledDirectly - Return true if the specified function is only called 1605/// directly. In other words, its address is never taken. 1606static bool OnlyCalledDirectly(Function *F) { 1607 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1608 Instruction *User = dyn_cast<Instruction>(*UI); 1609 if (!User) return false; 1610 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1611 1612 // See if the function address is passed as an argument. 1613 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end(); 1614 i != e; ++i) 1615 if (*i == F) return false; 1616 } 1617 return true; 1618} 1619 1620/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1621/// function, changing them to FastCC. 1622static void ChangeCalleesToFastCall(Function *F) { 1623 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1624 CallSite User(cast<Instruction>(*UI)); 1625 User.setCallingConv(CallingConv::Fast); 1626 } 1627} 1628 1629static PAListPtr StripNest(const PAListPtr &Attrs) { 1630 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1631 if ((Attrs.getSlot(i).Attrs & ParamAttr::Nest) == 0) 1632 continue; 1633 1634 // There can be only one. 1635 return Attrs.removeAttr(Attrs.getSlot(i).Index, ParamAttr::Nest); 1636 } 1637 1638 return Attrs; 1639} 1640 1641static void RemoveNestAttribute(Function *F) { 1642 F->setParamAttrs(StripNest(F->getParamAttrs())); 1643 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1644 CallSite User(cast<Instruction>(*UI)); 1645 User.setParamAttrs(StripNest(User.getParamAttrs())); 1646 } 1647} 1648 1649bool GlobalOpt::OptimizeFunctions(Module &M) { 1650 bool Changed = false; 1651 // Optimize functions. 1652 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1653 Function *F = FI++; 1654 F->removeDeadConstantUsers(); 1655 if (F->use_empty() && (F->hasInternalLinkage() || 1656 F->hasLinkOnceLinkage())) { 1657 M.getFunctionList().erase(F); 1658 Changed = true; 1659 ++NumFnDeleted; 1660 } else if (F->hasInternalLinkage()) { 1661 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1662 OnlyCalledDirectly(F)) { 1663 // If this function has C calling conventions, is not a varargs 1664 // function, and is only called directly, promote it to use the Fast 1665 // calling convention. 1666 F->setCallingConv(CallingConv::Fast); 1667 ChangeCalleesToFastCall(F); 1668 ++NumFastCallFns; 1669 Changed = true; 1670 } 1671 1672 if (F->getParamAttrs().hasAttrSomewhere(ParamAttr::Nest) && 1673 OnlyCalledDirectly(F)) { 1674 // The function is not used by a trampoline intrinsic, so it is safe 1675 // to remove the 'nest' attribute. 1676 RemoveNestAttribute(F); 1677 ++NumNestRemoved; 1678 Changed = true; 1679 } 1680 } 1681 } 1682 return Changed; 1683} 1684 1685bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1686 bool Changed = false; 1687 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1688 GVI != E; ) { 1689 GlobalVariable *GV = GVI++; 1690 if (!GV->isConstant() && GV->hasInternalLinkage() && 1691 GV->hasInitializer()) 1692 Changed |= ProcessInternalGlobal(GV, GVI); 1693 } 1694 return Changed; 1695} 1696 1697/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1698/// initializers have an init priority of 65535. 1699GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1700 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1701 I != E; ++I) 1702 if (I->getName() == "llvm.global_ctors") { 1703 // Found it, verify it's an array of { int, void()* }. 1704 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1705 if (!ATy) return 0; 1706 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1707 if (!STy || STy->getNumElements() != 2 || 1708 STy->getElementType(0) != Type::Int32Ty) return 0; 1709 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1710 if (!PFTy) return 0; 1711 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1712 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1713 FTy->getNumParams() != 0) 1714 return 0; 1715 1716 // Verify that the initializer is simple enough for us to handle. 1717 if (!I->hasInitializer()) return 0; 1718 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1719 if (!CA) return 0; 1720 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1721 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1722 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1723 continue; 1724 1725 // Must have a function or null ptr. 1726 if (!isa<Function>(CS->getOperand(1))) 1727 return 0; 1728 1729 // Init priority must be standard. 1730 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1731 if (!CI || CI->getZExtValue() != 65535) 1732 return 0; 1733 } else { 1734 return 0; 1735 } 1736 1737 return I; 1738 } 1739 return 0; 1740} 1741 1742/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1743/// return a list of the functions and null terminator as a vector. 1744static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1745 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1746 std::vector<Function*> Result; 1747 Result.reserve(CA->getNumOperands()); 1748 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1749 ConstantStruct *CS = cast<ConstantStruct>(*i); 1750 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1751 } 1752 return Result; 1753} 1754 1755/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1756/// specified array, returning the new global to use. 1757static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1758 const std::vector<Function*> &Ctors) { 1759 // If we made a change, reassemble the initializer list. 1760 std::vector<Constant*> CSVals; 1761 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1762 CSVals.push_back(0); 1763 1764 // Create the new init list. 1765 std::vector<Constant*> CAList; 1766 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1767 if (Ctors[i]) { 1768 CSVals[1] = Ctors[i]; 1769 } else { 1770 const Type *FTy = FunctionType::get(Type::VoidTy, 1771 std::vector<const Type*>(), false); 1772 const PointerType *PFTy = PointerType::getUnqual(FTy); 1773 CSVals[1] = Constant::getNullValue(PFTy); 1774 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1775 } 1776 CAList.push_back(ConstantStruct::get(CSVals)); 1777 } 1778 1779 // Create the array initializer. 1780 const Type *StructTy = 1781 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1782 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1783 CAList); 1784 1785 // If we didn't change the number of elements, don't create a new GV. 1786 if (CA->getType() == GCL->getInitializer()->getType()) { 1787 GCL->setInitializer(CA); 1788 return GCL; 1789 } 1790 1791 // Create the new global and insert it next to the existing list. 1792 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1793 GCL->getLinkage(), CA, "", 1794 (Module *)NULL, 1795 GCL->isThreadLocal()); 1796 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1797 NGV->takeName(GCL); 1798 1799 // Nuke the old list, replacing any uses with the new one. 1800 if (!GCL->use_empty()) { 1801 Constant *V = NGV; 1802 if (V->getType() != GCL->getType()) 1803 V = ConstantExpr::getBitCast(V, GCL->getType()); 1804 GCL->replaceAllUsesWith(V); 1805 } 1806 GCL->eraseFromParent(); 1807 1808 if (Ctors.size()) 1809 return NGV; 1810 else 1811 return 0; 1812} 1813 1814 1815static Constant *getVal(std::map<Value*, Constant*> &ComputedValues, 1816 Value *V) { 1817 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1818 Constant *R = ComputedValues[V]; 1819 assert(R && "Reference to an uncomputed value!"); 1820 return R; 1821} 1822 1823/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1824/// enough for us to understand. In particular, if it is a cast of something, 1825/// we punt. We basically just support direct accesses to globals and GEP's of 1826/// globals. This should be kept up to date with CommitValueTo. 1827static bool isSimpleEnoughPointerToCommit(Constant *C) { 1828 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1829 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1830 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1831 return !GV->isDeclaration(); // reject external globals. 1832 } 1833 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 1834 // Handle a constantexpr gep. 1835 if (CE->getOpcode() == Instruction::GetElementPtr && 1836 isa<GlobalVariable>(CE->getOperand(0))) { 1837 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1838 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage()) 1839 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1840 return GV->hasInitializer() && 1841 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1842 } 1843 return false; 1844} 1845 1846/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 1847/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 1848/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 1849static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 1850 ConstantExpr *Addr, unsigned OpNo) { 1851 // Base case of the recursion. 1852 if (OpNo == Addr->getNumOperands()) { 1853 assert(Val->getType() == Init->getType() && "Type mismatch!"); 1854 return Val; 1855 } 1856 1857 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1858 std::vector<Constant*> Elts; 1859 1860 // Break up the constant into its elements. 1861 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1862 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 1863 Elts.push_back(cast<Constant>(*i)); 1864 } else if (isa<ConstantAggregateZero>(Init)) { 1865 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1866 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 1867 } else if (isa<UndefValue>(Init)) { 1868 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1869 Elts.push_back(UndefValue::get(STy->getElementType(i))); 1870 } else { 1871 assert(0 && "This code is out of sync with " 1872 " ConstantFoldLoadThroughGEPConstantExpr"); 1873 } 1874 1875 // Replace the element that we are supposed to. 1876 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 1877 unsigned Idx = CU->getZExtValue(); 1878 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 1879 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 1880 1881 // Return the modified struct. 1882 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 1883 } else { 1884 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 1885 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 1886 1887 // Break up the array into elements. 1888 std::vector<Constant*> Elts; 1889 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1890 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1891 Elts.push_back(cast<Constant>(*i)); 1892 } else if (isa<ConstantAggregateZero>(Init)) { 1893 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 1894 Elts.assign(ATy->getNumElements(), Elt); 1895 } else if (isa<UndefValue>(Init)) { 1896 Constant *Elt = UndefValue::get(ATy->getElementType()); 1897 Elts.assign(ATy->getNumElements(), Elt); 1898 } else { 1899 assert(0 && "This code is out of sync with " 1900 " ConstantFoldLoadThroughGEPConstantExpr"); 1901 } 1902 1903 assert(CI->getZExtValue() < ATy->getNumElements()); 1904 Elts[CI->getZExtValue()] = 1905 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 1906 return ConstantArray::get(ATy, Elts); 1907 } 1908} 1909 1910/// CommitValueTo - We have decided that Addr (which satisfies the predicate 1911/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 1912static void CommitValueTo(Constant *Val, Constant *Addr) { 1913 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 1914 assert(GV->hasInitializer()); 1915 GV->setInitializer(Val); 1916 return; 1917 } 1918 1919 ConstantExpr *CE = cast<ConstantExpr>(Addr); 1920 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1921 1922 Constant *Init = GV->getInitializer(); 1923 Init = EvaluateStoreInto(Init, Val, CE, 2); 1924 GV->setInitializer(Init); 1925} 1926 1927/// ComputeLoadResult - Return the value that would be computed by a load from 1928/// P after the stores reflected by 'memory' have been performed. If we can't 1929/// decide, return null. 1930static Constant *ComputeLoadResult(Constant *P, 1931 const std::map<Constant*, Constant*> &Memory) { 1932 // If this memory location has been recently stored, use the stored value: it 1933 // is the most up-to-date. 1934 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P); 1935 if (I != Memory.end()) return I->second; 1936 1937 // Access it. 1938 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 1939 if (GV->hasInitializer()) 1940 return GV->getInitializer(); 1941 return 0; 1942 } 1943 1944 // Handle a constantexpr getelementptr. 1945 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 1946 if (CE->getOpcode() == Instruction::GetElementPtr && 1947 isa<GlobalVariable>(CE->getOperand(0))) { 1948 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 1949 if (GV->hasInitializer()) 1950 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 1951 } 1952 1953 return 0; // don't know how to evaluate. 1954} 1955 1956/// EvaluateFunction - Evaluate a call to function F, returning true if 1957/// successful, false if we can't evaluate it. ActualArgs contains the formal 1958/// arguments for the function. 1959static bool EvaluateFunction(Function *F, Constant *&RetVal, 1960 const std::vector<Constant*> &ActualArgs, 1961 std::vector<Function*> &CallStack, 1962 std::map<Constant*, Constant*> &MutatedMemory, 1963 std::vector<GlobalVariable*> &AllocaTmps) { 1964 // Check to see if this function is already executing (recursion). If so, 1965 // bail out. TODO: we might want to accept limited recursion. 1966 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 1967 return false; 1968 1969 CallStack.push_back(F); 1970 1971 /// Values - As we compute SSA register values, we store their contents here. 1972 std::map<Value*, Constant*> Values; 1973 1974 // Initialize arguments to the incoming values specified. 1975 unsigned ArgNo = 0; 1976 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1977 ++AI, ++ArgNo) 1978 Values[AI] = ActualArgs[ArgNo]; 1979 1980 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 1981 /// we can only evaluate any one basic block at most once. This set keeps 1982 /// track of what we have executed so we can detect recursive cases etc. 1983 std::set<BasicBlock*> ExecutedBlocks; 1984 1985 // CurInst - The current instruction we're evaluating. 1986 BasicBlock::iterator CurInst = F->begin()->begin(); 1987 1988 // This is the main evaluation loop. 1989 while (1) { 1990 Constant *InstResult = 0; 1991 1992 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 1993 if (SI->isVolatile()) return false; // no volatile accesses. 1994 Constant *Ptr = getVal(Values, SI->getOperand(1)); 1995 if (!isSimpleEnoughPointerToCommit(Ptr)) 1996 // If this is too complex for us to commit, reject it. 1997 return false; 1998 Constant *Val = getVal(Values, SI->getOperand(0)); 1999 MutatedMemory[Ptr] = Val; 2000 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2001 InstResult = ConstantExpr::get(BO->getOpcode(), 2002 getVal(Values, BO->getOperand(0)), 2003 getVal(Values, BO->getOperand(1))); 2004 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2005 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2006 getVal(Values, CI->getOperand(0)), 2007 getVal(Values, CI->getOperand(1))); 2008 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2009 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2010 getVal(Values, CI->getOperand(0)), 2011 CI->getType()); 2012 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2013 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2014 getVal(Values, SI->getOperand(1)), 2015 getVal(Values, SI->getOperand(2))); 2016 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2017 Constant *P = getVal(Values, GEP->getOperand(0)); 2018 SmallVector<Constant*, 8> GEPOps; 2019 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2020 i != e; ++i) 2021 GEPOps.push_back(getVal(Values, *i)); 2022 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2023 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2024 if (LI->isVolatile()) return false; // no volatile accesses. 2025 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2026 MutatedMemory); 2027 if (InstResult == 0) return false; // Could not evaluate load. 2028 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2029 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2030 const Type *Ty = AI->getType()->getElementType(); 2031 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2032 GlobalValue::InternalLinkage, 2033 UndefValue::get(Ty), 2034 AI->getName())); 2035 InstResult = AllocaTmps.back(); 2036 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2037 // Cannot handle inline asm. 2038 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2039 2040 // Resolve function pointers. 2041 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2042 if (!Callee) return false; // Cannot resolve. 2043 2044 std::vector<Constant*> Formals; 2045 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end(); 2046 i != e; ++i) 2047 Formals.push_back(getVal(Values, *i)); 2048 2049 if (Callee->isDeclaration()) { 2050 // If this is a function we can constant fold, do it. 2051 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2052 Formals.size())) { 2053 InstResult = C; 2054 } else { 2055 return false; 2056 } 2057 } else { 2058 if (Callee->getFunctionType()->isVarArg()) 2059 return false; 2060 2061 Constant *RetVal; 2062 2063 // Execute the call, if successful, use the return value. 2064 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2065 MutatedMemory, AllocaTmps)) 2066 return false; 2067 InstResult = RetVal; 2068 } 2069 } else if (isa<TerminatorInst>(CurInst)) { 2070 BasicBlock *NewBB = 0; 2071 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2072 if (BI->isUnconditional()) { 2073 NewBB = BI->getSuccessor(0); 2074 } else { 2075 ConstantInt *Cond = 2076 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2077 if (!Cond) return false; // Cannot determine. 2078 2079 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2080 } 2081 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2082 ConstantInt *Val = 2083 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2084 if (!Val) return false; // Cannot determine. 2085 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2086 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2087 if (RI->getNumOperands()) 2088 RetVal = getVal(Values, RI->getOperand(0)); 2089 2090 CallStack.pop_back(); // return from fn. 2091 return true; // We succeeded at evaluating this ctor! 2092 } else { 2093 // invoke, unwind, unreachable. 2094 return false; // Cannot handle this terminator. 2095 } 2096 2097 // Okay, we succeeded in evaluating this control flow. See if we have 2098 // executed the new block before. If so, we have a looping function, 2099 // which we cannot evaluate in reasonable time. 2100 if (!ExecutedBlocks.insert(NewBB).second) 2101 return false; // looped! 2102 2103 // Okay, we have never been in this block before. Check to see if there 2104 // are any PHI nodes. If so, evaluate them with information about where 2105 // we came from. 2106 BasicBlock *OldBB = CurInst->getParent(); 2107 CurInst = NewBB->begin(); 2108 PHINode *PN; 2109 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2110 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2111 2112 // Do NOT increment CurInst. We know that the terminator had no value. 2113 continue; 2114 } else { 2115 // Did not know how to evaluate this! 2116 return false; 2117 } 2118 2119 if (!CurInst->use_empty()) 2120 Values[CurInst] = InstResult; 2121 2122 // Advance program counter. 2123 ++CurInst; 2124 } 2125} 2126 2127/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2128/// we can. Return true if we can, false otherwise. 2129static bool EvaluateStaticConstructor(Function *F) { 2130 /// MutatedMemory - For each store we execute, we update this map. Loads 2131 /// check this to get the most up-to-date value. If evaluation is successful, 2132 /// this state is committed to the process. 2133 std::map<Constant*, Constant*> MutatedMemory; 2134 2135 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2136 /// to represent its body. This vector is needed so we can delete the 2137 /// temporary globals when we are done. 2138 std::vector<GlobalVariable*> AllocaTmps; 2139 2140 /// CallStack - This is used to detect recursion. In pathological situations 2141 /// we could hit exponential behavior, but at least there is nothing 2142 /// unbounded. 2143 std::vector<Function*> CallStack; 2144 2145 // Call the function. 2146 Constant *RetValDummy; 2147 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2148 CallStack, MutatedMemory, AllocaTmps); 2149 if (EvalSuccess) { 2150 // We succeeded at evaluation: commit the result. 2151 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2152 << F->getName() << "' to " << MutatedMemory.size() 2153 << " stores.\n"; 2154 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2155 E = MutatedMemory.end(); I != E; ++I) 2156 CommitValueTo(I->second, I->first); 2157 } 2158 2159 // At this point, we are done interpreting. If we created any 'alloca' 2160 // temporaries, release them now. 2161 while (!AllocaTmps.empty()) { 2162 GlobalVariable *Tmp = AllocaTmps.back(); 2163 AllocaTmps.pop_back(); 2164 2165 // If there are still users of the alloca, the program is doing something 2166 // silly, e.g. storing the address of the alloca somewhere and using it 2167 // later. Since this is undefined, we'll just make it be null. 2168 if (!Tmp->use_empty()) 2169 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2170 delete Tmp; 2171 } 2172 2173 return EvalSuccess; 2174} 2175 2176 2177 2178/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2179/// Return true if anything changed. 2180bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2181 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2182 bool MadeChange = false; 2183 if (Ctors.empty()) return false; 2184 2185 // Loop over global ctors, optimizing them when we can. 2186 for (unsigned i = 0; i != Ctors.size(); ++i) { 2187 Function *F = Ctors[i]; 2188 // Found a null terminator in the middle of the list, prune off the rest of 2189 // the list. 2190 if (F == 0) { 2191 if (i != Ctors.size()-1) { 2192 Ctors.resize(i+1); 2193 MadeChange = true; 2194 } 2195 break; 2196 } 2197 2198 // We cannot simplify external ctor functions. 2199 if (F->empty()) continue; 2200 2201 // If we can evaluate the ctor at compile time, do. 2202 if (EvaluateStaticConstructor(F)) { 2203 Ctors.erase(Ctors.begin()+i); 2204 MadeChange = true; 2205 --i; 2206 ++NumCtorsEvaluated; 2207 continue; 2208 } 2209 } 2210 2211 if (!MadeChange) return false; 2212 2213 GCL = InstallGlobalCtors(GCL, Ctors); 2214 return true; 2215} 2216 2217 2218bool GlobalOpt::runOnModule(Module &M) { 2219 bool Changed = false; 2220 2221 // Try to find the llvm.globalctors list. 2222 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2223 2224 bool LocalChange = true; 2225 while (LocalChange) { 2226 LocalChange = false; 2227 2228 // Delete functions that are trivially dead, ccc -> fastcc 2229 LocalChange |= OptimizeFunctions(M); 2230 2231 // Optimize global_ctors list. 2232 if (GlobalCtors) 2233 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2234 2235 // Optimize non-address-taken globals. 2236 LocalChange |= OptimizeGlobalVars(M); 2237 Changed |= LocalChange; 2238 } 2239 2240 // TODO: Move all global ctors functions to the end of the module for code 2241 // layout. 2242 2243 return Changed; 2244} 2245