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/DenseMap.h" 33#include "llvm/ADT/SmallPtrSet.h" 34#include "llvm/ADT/SmallVector.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 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"); 53STATISTIC(NumAliasesResolved, "Number of global aliases resolved"); 54STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated"); 55 56namespace { 57 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass { 58 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 59 AU.addRequired<TargetData>(); 60 } 61 static char ID; // Pass identification, replacement for typeid 62 GlobalOpt() : ModulePass(&ID) {} 63 64 bool runOnModule(Module &M); 65 66 private: 67 GlobalVariable *FindGlobalCtors(Module &M); 68 bool OptimizeFunctions(Module &M); 69 bool OptimizeGlobalVars(Module &M); 70 bool OptimizeGlobalAliases(Module &M); 71 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL); 72 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 73 }; 74} 75 76char GlobalOpt::ID = 0; 77static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer"); 78 79ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 80 81namespace { 82 83/// GlobalStatus - As we analyze each global, keep track of some information 84/// about it. If we find out that the address of the global is taken, none of 85/// this info will be accurate. 86struct VISIBILITY_HIDDEN GlobalStatus { 87 /// isLoaded - True if the global is ever loaded. If the global isn't ever 88 /// loaded it can be deleted. 89 bool isLoaded; 90 91 /// StoredType - Keep track of what stores to the global look like. 92 /// 93 enum StoredType { 94 /// NotStored - There is no store to this global. It can thus be marked 95 /// constant. 96 NotStored, 97 98 /// isInitializerStored - This global is stored to, but the only thing 99 /// stored is the constant it was initialized with. This is only tracked 100 /// for scalar globals. 101 isInitializerStored, 102 103 /// isStoredOnce - This global is stored to, but only its initializer and 104 /// one other value is ever stored to it. If this global isStoredOnce, we 105 /// track the value stored to it in StoredOnceValue below. This is only 106 /// tracked for scalar globals. 107 isStoredOnce, 108 109 /// isStored - This global is stored to by multiple values or something else 110 /// that we cannot track. 111 isStored 112 } StoredType; 113 114 /// StoredOnceValue - If only one value (besides the initializer constant) is 115 /// ever stored to this global, keep track of what value it is. 116 Value *StoredOnceValue; 117 118 /// AccessingFunction/HasMultipleAccessingFunctions - These start out 119 /// null/false. When the first accessing function is noticed, it is recorded. 120 /// When a second different accessing function is noticed, 121 /// HasMultipleAccessingFunctions is set to true. 122 Function *AccessingFunction; 123 bool HasMultipleAccessingFunctions; 124 125 /// HasNonInstructionUser - Set to true if this global has a user that is not 126 /// an instruction (e.g. a constant expr or GV initializer). 127 bool HasNonInstructionUser; 128 129 /// HasPHIUser - Set to true if this global has a user that is a PHI node. 130 bool HasPHIUser; 131 132 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0), 133 AccessingFunction(0), HasMultipleAccessingFunctions(false), 134 HasNonInstructionUser(false), HasPHIUser(false) {} 135}; 136 137} 138 139/// ConstantIsDead - Return true if the specified constant is (transitively) 140/// dead. The constant may be used by other constants (e.g. constant arrays and 141/// constant exprs) as long as they are dead, but it cannot be used by anything 142/// else. 143static bool ConstantIsDead(Constant *C) { 144 if (isa<GlobalValue>(C)) return false; 145 146 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI) 147 if (Constant *CU = dyn_cast<Constant>(*UI)) { 148 if (!ConstantIsDead(CU)) return false; 149 } else 150 return false; 151 return true; 152} 153 154 155/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 156/// structure. If the global has its address taken, return true to indicate we 157/// can't do anything with it. 158/// 159static bool AnalyzeGlobal(Value *V, GlobalStatus &GS, 160 SmallPtrSet<PHINode*, 16> &PHIUsers) { 161 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 162 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) { 163 GS.HasNonInstructionUser = true; 164 165 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 166 167 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) { 168 if (!GS.HasMultipleAccessingFunctions) { 169 Function *F = I->getParent()->getParent(); 170 if (GS.AccessingFunction == 0) 171 GS.AccessingFunction = F; 172 else if (GS.AccessingFunction != F) 173 GS.HasMultipleAccessingFunctions = true; 174 } 175 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 176 GS.isLoaded = true; 177 if (LI->isVolatile()) return true; // Don't hack on volatile loads. 178 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 179 // Don't allow a store OF the address, only stores TO the address. 180 if (SI->getOperand(0) == V) return true; 181 182 if (SI->isVolatile()) return true; // Don't hack on volatile stores. 183 184 // If this is a direct store to the global (i.e., the global is a scalar 185 // value, not an aggregate), keep more specific information about 186 // stores. 187 if (GS.StoredType != GlobalStatus::isStored) { 188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){ 189 Value *StoredVal = SI->getOperand(0); 190 if (StoredVal == GV->getInitializer()) { 191 if (GS.StoredType < GlobalStatus::isInitializerStored) 192 GS.StoredType = GlobalStatus::isInitializerStored; 193 } else if (isa<LoadInst>(StoredVal) && 194 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 195 // G = G 196 if (GS.StoredType < GlobalStatus::isInitializerStored) 197 GS.StoredType = GlobalStatus::isInitializerStored; 198 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 199 GS.StoredType = GlobalStatus::isStoredOnce; 200 GS.StoredOnceValue = StoredVal; 201 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 202 GS.StoredOnceValue == StoredVal) { 203 // noop. 204 } else { 205 GS.StoredType = GlobalStatus::isStored; 206 } 207 } else { 208 GS.StoredType = GlobalStatus::isStored; 209 } 210 } 211 } else if (isa<GetElementPtrInst>(I)) { 212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 213 } else if (isa<SelectInst>(I)) { 214 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 215 } else if (PHINode *PN = dyn_cast<PHINode>(I)) { 216 // PHI nodes we can check just like select or GEP instructions, but we 217 // have to be careful about infinite recursion. 218 if (PHIUsers.insert(PN)) // Not already visited. 219 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 220 GS.HasPHIUser = true; 221 } else if (isa<CmpInst>(I)) { 222 } else if (isa<MemTransferInst>(I)) { 223 if (I->getOperand(1) == V) 224 GS.StoredType = GlobalStatus::isStored; 225 if (I->getOperand(2) == V) 226 GS.isLoaded = true; 227 } else if (isa<MemSetInst>(I)) { 228 assert(I->getOperand(1) == V && "Memset only takes one pointer!"); 229 GS.StoredType = GlobalStatus::isStored; 230 } else { 231 return true; // Any other non-load instruction might take address! 232 } 233 } else if (Constant *C = dyn_cast<Constant>(*UI)) { 234 GS.HasNonInstructionUser = true; 235 // We might have a dead and dangling constant hanging off of here. 236 if (!ConstantIsDead(C)) 237 return true; 238 } else { 239 GS.HasNonInstructionUser = true; 240 // Otherwise must be some other user. 241 return true; 242 } 243 244 return false; 245} 246 247static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 248 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 249 if (!CI) return 0; 250 unsigned IdxV = CI->getZExtValue(); 251 252 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 253 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 254 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 255 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 256 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 257 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 258 } else if (isa<ConstantAggregateZero>(Agg)) { 259 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 260 if (IdxV < STy->getNumElements()) 261 return Constant::getNullValue(STy->getElementType(IdxV)); 262 } else if (const SequentialType *STy = 263 dyn_cast<SequentialType>(Agg->getType())) { 264 return Constant::getNullValue(STy->getElementType()); 265 } 266 } else if (isa<UndefValue>(Agg)) { 267 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 268 if (IdxV < STy->getNumElements()) 269 return UndefValue::get(STy->getElementType(IdxV)); 270 } else if (const SequentialType *STy = 271 dyn_cast<SequentialType>(Agg->getType())) { 272 return UndefValue::get(STy->getElementType()); 273 } 274 } 275 return 0; 276} 277 278 279/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 280/// users of the global, cleaning up the obvious ones. This is largely just a 281/// quick scan over the use list to clean up the easy and obvious cruft. This 282/// returns true if it made a change. 283static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 284 bool Changed = false; 285 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 286 User *U = *UI++; 287 288 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 289 if (Init) { 290 // Replace the load with the initializer. 291 LI->replaceAllUsesWith(Init); 292 LI->eraseFromParent(); 293 Changed = true; 294 } 295 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 296 // Store must be unreachable or storing Init into the global. 297 SI->eraseFromParent(); 298 Changed = true; 299 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 300 if (CE->getOpcode() == Instruction::GetElementPtr) { 301 Constant *SubInit = 0; 302 if (Init) 303 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 304 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 305 } else if (CE->getOpcode() == Instruction::BitCast && 306 isa<PointerType>(CE->getType())) { 307 // Pointer cast, delete any stores and memsets to the global. 308 Changed |= CleanupConstantGlobalUsers(CE, 0); 309 } 310 311 if (CE->use_empty()) { 312 CE->destroyConstant(); 313 Changed = true; 314 } 315 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 316 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 317 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 318 // and will invalidate our notion of what Init is. 319 Constant *SubInit = 0; 320 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 321 ConstantExpr *CE = 322 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 323 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 324 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 325 } 326 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 327 328 if (GEP->use_empty()) { 329 GEP->eraseFromParent(); 330 Changed = true; 331 } 332 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 333 if (MI->getRawDest() == V) { 334 MI->eraseFromParent(); 335 Changed = true; 336 } 337 338 } else if (Constant *C = dyn_cast<Constant>(U)) { 339 // If we have a chain of dead constantexprs or other things dangling from 340 // us, and if they are all dead, nuke them without remorse. 341 if (ConstantIsDead(C)) { 342 C->destroyConstant(); 343 // This could have invalidated UI, start over from scratch. 344 CleanupConstantGlobalUsers(V, Init); 345 return true; 346 } 347 } 348 } 349 return Changed; 350} 351 352/// isSafeSROAElementUse - Return true if the specified instruction is a safe 353/// user of a derived expression from a global that we want to SROA. 354static bool isSafeSROAElementUse(Value *V) { 355 // We might have a dead and dangling constant hanging off of here. 356 if (Constant *C = dyn_cast<Constant>(V)) 357 return ConstantIsDead(C); 358 359 Instruction *I = dyn_cast<Instruction>(V); 360 if (!I) return false; 361 362 // Loads are ok. 363 if (isa<LoadInst>(I)) return true; 364 365 // Stores *to* the pointer are ok. 366 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 367 return SI->getOperand(0) != V; 368 369 // Otherwise, it must be a GEP. 370 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 371 if (GEPI == 0) return false; 372 373 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 374 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 375 return false; 376 377 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end(); 378 I != E; ++I) 379 if (!isSafeSROAElementUse(*I)) 380 return false; 381 return true; 382} 383 384 385/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 386/// Look at it and its uses and decide whether it is safe to SROA this global. 387/// 388static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 389 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 390 if (!isa<GetElementPtrInst>(U) && 391 (!isa<ConstantExpr>(U) || 392 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 393 return false; 394 395 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 396 // don't like < 3 operand CE's, and we don't like non-constant integer 397 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 398 // value of C. 399 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 400 !cast<Constant>(U->getOperand(1))->isNullValue() || 401 !isa<ConstantInt>(U->getOperand(2))) 402 return false; 403 404 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 405 ++GEPI; // Skip over the pointer index. 406 407 // If this is a use of an array allocation, do a bit more checking for sanity. 408 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 409 uint64_t NumElements = AT->getNumElements(); 410 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 411 412 // Check to make sure that index falls within the array. If not, 413 // something funny is going on, so we won't do the optimization. 414 // 415 if (Idx->getZExtValue() >= NumElements) 416 return false; 417 418 // We cannot scalar repl this level of the array unless any array 419 // sub-indices are in-range constants. In particular, consider: 420 // A[0][i]. We cannot know that the user isn't doing invalid things like 421 // allowing i to index an out-of-range subscript that accesses A[1]. 422 // 423 // Scalar replacing *just* the outer index of the array is probably not 424 // going to be a win anyway, so just give up. 425 for (++GEPI; // Skip array index. 426 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI)); 427 ++GEPI) { 428 uint64_t NumElements; 429 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 430 NumElements = SubArrayTy->getNumElements(); 431 else 432 NumElements = cast<VectorType>(*GEPI)->getNumElements(); 433 434 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 435 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 436 return false; 437 } 438 } 439 440 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) 441 if (!isSafeSROAElementUse(*I)) 442 return false; 443 return true; 444} 445 446/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 447/// is safe for us to perform this transformation. 448/// 449static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 450 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 451 UI != E; ++UI) { 452 if (!IsUserOfGlobalSafeForSRA(*UI, GV)) 453 return false; 454 } 455 return true; 456} 457 458 459/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 460/// variable. This opens the door for other optimizations by exposing the 461/// behavior of the program in a more fine-grained way. We have determined that 462/// this transformation is safe already. We return the first global variable we 463/// insert so that the caller can reprocess it. 464static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { 465 // Make sure this global only has simple uses that we can SRA. 466 if (!GlobalUsersSafeToSRA(GV)) 467 return 0; 468 469 assert(GV->hasLocalLinkage() && !GV->isConstant()); 470 Constant *Init = GV->getInitializer(); 471 const Type *Ty = Init->getType(); 472 473 std::vector<GlobalVariable*> NewGlobals; 474 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 475 476 // Get the alignment of the global, either explicit or target-specific. 477 unsigned StartAlignment = GV->getAlignment(); 478 if (StartAlignment == 0) 479 StartAlignment = TD.getABITypeAlignment(GV->getType()); 480 481 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 482 NewGlobals.reserve(STy->getNumElements()); 483 const StructLayout &Layout = *TD.getStructLayout(STy); 484 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 485 Constant *In = getAggregateConstantElement(Init, 486 ConstantInt::get(Type::Int32Ty, i)); 487 assert(In && "Couldn't get element of initializer?"); 488 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 489 GlobalVariable::InternalLinkage, 490 In, GV->getName()+"."+utostr(i), 491 (Module *)NULL, 492 GV->isThreadLocal(), 493 GV->getType()->getAddressSpace()); 494 Globals.insert(GV, NGV); 495 NewGlobals.push_back(NGV); 496 497 // Calculate the known alignment of the field. If the original aggregate 498 // had 256 byte alignment for example, something might depend on that: 499 // propagate info to each field. 500 uint64_t FieldOffset = Layout.getElementOffset(i); 501 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 502 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 503 NGV->setAlignment(NewAlign); 504 } 505 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 506 unsigned NumElements = 0; 507 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 508 NumElements = ATy->getNumElements(); 509 else 510 NumElements = cast<VectorType>(STy)->getNumElements(); 511 512 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 513 return 0; // It's not worth it. 514 NewGlobals.reserve(NumElements); 515 516 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType()); 517 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 518 for (unsigned i = 0, e = NumElements; i != e; ++i) { 519 Constant *In = getAggregateConstantElement(Init, 520 ConstantInt::get(Type::Int32Ty, i)); 521 assert(In && "Couldn't get element of initializer?"); 522 523 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 524 GlobalVariable::InternalLinkage, 525 In, GV->getName()+"."+utostr(i), 526 (Module *)NULL, 527 GV->isThreadLocal(), 528 GV->getType()->getAddressSpace()); 529 Globals.insert(GV, NGV); 530 NewGlobals.push_back(NGV); 531 532 // Calculate the known alignment of the field. If the original aggregate 533 // had 256 byte alignment for example, something might depend on that: 534 // propagate info to each field. 535 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 536 if (NewAlign > EltAlign) 537 NGV->setAlignment(NewAlign); 538 } 539 } 540 541 if (NewGlobals.empty()) 542 return 0; 543 544 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 545 546 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 547 548 // Loop over all of the uses of the global, replacing the constantexpr geps, 549 // with smaller constantexpr geps or direct references. 550 while (!GV->use_empty()) { 551 User *GEP = GV->use_back(); 552 assert(((isa<ConstantExpr>(GEP) && 553 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 554 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 555 556 // Ignore the 1th operand, which has to be zero or else the program is quite 557 // broken (undefined). Get the 2nd operand, which is the structure or array 558 // index. 559 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 560 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 561 562 Value *NewPtr = NewGlobals[Val]; 563 564 // Form a shorter GEP if needed. 565 if (GEP->getNumOperands() > 3) { 566 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 567 SmallVector<Constant*, 8> Idxs; 568 Idxs.push_back(NullInt); 569 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 570 Idxs.push_back(CE->getOperand(i)); 571 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 572 &Idxs[0], Idxs.size()); 573 } else { 574 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 575 SmallVector<Value*, 8> Idxs; 576 Idxs.push_back(NullInt); 577 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 578 Idxs.push_back(GEPI->getOperand(i)); 579 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 580 GEPI->getName()+"."+utostr(Val), GEPI); 581 } 582 } 583 GEP->replaceAllUsesWith(NewPtr); 584 585 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 586 GEPI->eraseFromParent(); 587 else 588 cast<ConstantExpr>(GEP)->destroyConstant(); 589 } 590 591 // Delete the old global, now that it is dead. 592 Globals.erase(GV); 593 ++NumSRA; 594 595 // Loop over the new globals array deleting any globals that are obviously 596 // dead. This can arise due to scalarization of a structure or an array that 597 // has elements that are dead. 598 unsigned FirstGlobal = 0; 599 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 600 if (NewGlobals[i]->use_empty()) { 601 Globals.erase(NewGlobals[i]); 602 if (FirstGlobal == i) ++FirstGlobal; 603 } 604 605 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 606} 607 608/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 609/// value will trap if the value is dynamically null. PHIs keeps track of any 610/// phi nodes we've seen to avoid reprocessing them. 611static bool AllUsesOfValueWillTrapIfNull(Value *V, 612 SmallPtrSet<PHINode*, 8> &PHIs) { 613 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 614 if (isa<LoadInst>(*UI)) { 615 // Will trap. 616 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 617 if (SI->getOperand(0) == V) { 618 //cerr << "NONTRAPPING USE: " << **UI; 619 return false; // Storing the value. 620 } 621 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 622 if (CI->getOperand(0) != V) { 623 //cerr << "NONTRAPPING USE: " << **UI; 624 return false; // Not calling the ptr 625 } 626 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 627 if (II->getOperand(0) != V) { 628 //cerr << "NONTRAPPING USE: " << **UI; 629 return false; // Not calling the ptr 630 } 631 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 632 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 633 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 634 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 635 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 636 // If we've already seen this phi node, ignore it, it has already been 637 // checked. 638 if (PHIs.insert(PN)) 639 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 640 } else if (isa<ICmpInst>(*UI) && 641 isa<ConstantPointerNull>(UI->getOperand(1))) { 642 // Ignore setcc X, null 643 } else { 644 //cerr << "NONTRAPPING USE: " << **UI; 645 return false; 646 } 647 return true; 648} 649 650/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 651/// from GV will trap if the loaded value is null. Note that this also permits 652/// comparisons of the loaded value against null, as a special case. 653static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 654 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 655 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 656 SmallPtrSet<PHINode*, 8> PHIs; 657 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 658 return false; 659 } else if (isa<StoreInst>(*UI)) { 660 // Ignore stores to the global. 661 } else { 662 // We don't know or understand this user, bail out. 663 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 664 return false; 665 } 666 667 return true; 668} 669 670static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 671 bool Changed = false; 672 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 673 Instruction *I = cast<Instruction>(*UI++); 674 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 675 LI->setOperand(0, NewV); 676 Changed = true; 677 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 678 if (SI->getOperand(1) == V) { 679 SI->setOperand(1, NewV); 680 Changed = true; 681 } 682 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 683 if (I->getOperand(0) == V) { 684 // Calling through the pointer! Turn into a direct call, but be careful 685 // that the pointer is not also being passed as an argument. 686 I->setOperand(0, NewV); 687 Changed = true; 688 bool PassedAsArg = false; 689 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 690 if (I->getOperand(i) == V) { 691 PassedAsArg = true; 692 I->setOperand(i, NewV); 693 } 694 695 if (PassedAsArg) { 696 // Being passed as an argument also. Be careful to not invalidate UI! 697 UI = V->use_begin(); 698 } 699 } 700 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 701 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 702 ConstantExpr::getCast(CI->getOpcode(), 703 NewV, CI->getType())); 704 if (CI->use_empty()) { 705 Changed = true; 706 CI->eraseFromParent(); 707 } 708 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 709 // Should handle GEP here. 710 SmallVector<Constant*, 8> Idxs; 711 Idxs.reserve(GEPI->getNumOperands()-1); 712 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 713 i != e; ++i) 714 if (Constant *C = dyn_cast<Constant>(*i)) 715 Idxs.push_back(C); 716 else 717 break; 718 if (Idxs.size() == GEPI->getNumOperands()-1) 719 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 720 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 721 Idxs.size())); 722 if (GEPI->use_empty()) { 723 Changed = true; 724 GEPI->eraseFromParent(); 725 } 726 } 727 } 728 729 return Changed; 730} 731 732 733/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 734/// value stored into it. If there are uses of the loaded value that would trap 735/// if the loaded value is dynamically null, then we know that they cannot be 736/// reachable with a null optimize away the load. 737static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 738 bool Changed = false; 739 740 // Keep track of whether we are able to remove all the uses of the global 741 // other than the store that defines it. 742 bool AllNonStoreUsesGone = true; 743 744 // Replace all uses of loads with uses of uses of the stored value. 745 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){ 746 User *GlobalUser = *GUI++; 747 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) { 748 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 749 // If we were able to delete all uses of the loads 750 if (LI->use_empty()) { 751 LI->eraseFromParent(); 752 Changed = true; 753 } else { 754 AllNonStoreUsesGone = false; 755 } 756 } else if (isa<StoreInst>(GlobalUser)) { 757 // Ignore the store that stores "LV" to the global. 758 assert(GlobalUser->getOperand(1) == GV && 759 "Must be storing *to* the global"); 760 } else { 761 AllNonStoreUsesGone = false; 762 763 // If we get here we could have other crazy uses that are transitively 764 // loaded. 765 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || 766 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!"); 767 } 768 } 769 770 if (Changed) { 771 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 772 ++NumGlobUses; 773 } 774 775 // If we nuked all of the loads, then none of the stores are needed either, 776 // nor is the global. 777 if (AllNonStoreUsesGone) { 778 DOUT << " *** GLOBAL NOW DEAD!\n"; 779 CleanupConstantGlobalUsers(GV, 0); 780 if (GV->use_empty()) { 781 GV->eraseFromParent(); 782 ++NumDeleted; 783 } 784 Changed = true; 785 } 786 return Changed; 787} 788 789/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 790/// instructions that are foldable. 791static void ConstantPropUsersOf(Value *V) { 792 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 793 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 794 if (Constant *NewC = ConstantFoldInstruction(I)) { 795 I->replaceAllUsesWith(NewC); 796 797 // Advance UI to the next non-I use to avoid invalidating it! 798 // Instructions could multiply use V. 799 while (UI != E && *UI == I) 800 ++UI; 801 I->eraseFromParent(); 802 } 803} 804 805/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 806/// variable, and transforms the program as if it always contained the result of 807/// the specified malloc. Because it is always the result of the specified 808/// malloc, there is no reason to actually DO the malloc. Instead, turn the 809/// malloc into a global, and any loads of GV as uses of the new global. 810static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 811 MallocInst *MI) { 812 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 813 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 814 815 if (NElements->getZExtValue() != 1) { 816 // If we have an array allocation, transform it to a single element 817 // allocation to make the code below simpler. 818 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 819 NElements->getZExtValue()); 820 MallocInst *NewMI = 821 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 822 MI->getAlignment(), MI->getName(), MI); 823 Value* Indices[2]; 824 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 825 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2, 826 NewMI->getName()+".el0", MI); 827 MI->replaceAllUsesWith(NewGEP); 828 MI->eraseFromParent(); 829 MI = NewMI; 830 } 831 832 // Create the new global variable. The contents of the malloc'd memory is 833 // undefined, so initialize with an undef value. 834 Constant *Init = UndefValue::get(MI->getAllocatedType()); 835 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 836 GlobalValue::InternalLinkage, Init, 837 GV->getName()+".body", 838 (Module *)NULL, 839 GV->isThreadLocal()); 840 // FIXME: This new global should have the alignment returned by malloc. Code 841 // could depend on malloc returning large alignment (on the mac, 16 bytes) but 842 // this would only guarantee some lower alignment. 843 GV->getParent()->getGlobalList().insert(GV, NewGV); 844 845 // Anything that used the malloc now uses the global directly. 846 MI->replaceAllUsesWith(NewGV); 847 848 Constant *RepValue = NewGV; 849 if (NewGV->getType() != GV->getType()->getElementType()) 850 RepValue = ConstantExpr::getBitCast(RepValue, 851 GV->getType()->getElementType()); 852 853 // If there is a comparison against null, we will insert a global bool to 854 // keep track of whether the global was initialized yet or not. 855 GlobalVariable *InitBool = 856 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 857 ConstantInt::getFalse(), GV->getName()+".init", 858 (Module *)NULL, GV->isThreadLocal()); 859 bool InitBoolUsed = false; 860 861 // Loop over all uses of GV, processing them in turn. 862 std::vector<StoreInst*> Stores; 863 while (!GV->use_empty()) 864 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 865 while (!LI->use_empty()) { 866 Use &LoadUse = LI->use_begin().getUse(); 867 if (!isa<ICmpInst>(LoadUse.getUser())) 868 LoadUse = RepValue; 869 else { 870 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 871 // Replace the cmp X, 0 with a use of the bool value. 872 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 873 InitBoolUsed = true; 874 switch (CI->getPredicate()) { 875 default: assert(0 && "Unknown ICmp Predicate!"); 876 case ICmpInst::ICMP_ULT: 877 case ICmpInst::ICMP_SLT: 878 LV = ConstantInt::getFalse(); // X < null -> always false 879 break; 880 case ICmpInst::ICMP_ULE: 881 case ICmpInst::ICMP_SLE: 882 case ICmpInst::ICMP_EQ: 883 LV = BinaryOperator::CreateNot(LV, "notinit", CI); 884 break; 885 case ICmpInst::ICMP_NE: 886 case ICmpInst::ICMP_UGE: 887 case ICmpInst::ICMP_SGE: 888 case ICmpInst::ICMP_UGT: 889 case ICmpInst::ICMP_SGT: 890 break; // no change. 891 } 892 CI->replaceAllUsesWith(LV); 893 CI->eraseFromParent(); 894 } 895 } 896 LI->eraseFromParent(); 897 } else { 898 StoreInst *SI = cast<StoreInst>(GV->use_back()); 899 // The global is initialized when the store to it occurs. 900 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 901 SI->eraseFromParent(); 902 } 903 904 // If the initialization boolean was used, insert it, otherwise delete it. 905 if (!InitBoolUsed) { 906 while (!InitBool->use_empty()) // Delete initializations 907 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 908 delete InitBool; 909 } else 910 GV->getParent()->getGlobalList().insert(GV, InitBool); 911 912 913 // Now the GV is dead, nuke it and the malloc. 914 GV->eraseFromParent(); 915 MI->eraseFromParent(); 916 917 // To further other optimizations, loop over all users of NewGV and try to 918 // constant prop them. This will promote GEP instructions with constant 919 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 920 ConstantPropUsersOf(NewGV); 921 if (RepValue != NewGV) 922 ConstantPropUsersOf(RepValue); 923 924 return NewGV; 925} 926 927/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 928/// to make sure that there are no complex uses of V. We permit simple things 929/// like dereferencing the pointer, but not storing through the address, unless 930/// it is to the specified global. 931static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 932 GlobalVariable *GV, 933 SmallPtrSet<PHINode*, 8> &PHIs) { 934 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
| 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/DenseMap.h" 33#include "llvm/ADT/SmallPtrSet.h" 34#include "llvm/ADT/SmallVector.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 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"); 53STATISTIC(NumAliasesResolved, "Number of global aliases resolved"); 54STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated"); 55 56namespace { 57 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass { 58 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 59 AU.addRequired<TargetData>(); 60 } 61 static char ID; // Pass identification, replacement for typeid 62 GlobalOpt() : ModulePass(&ID) {} 63 64 bool runOnModule(Module &M); 65 66 private: 67 GlobalVariable *FindGlobalCtors(Module &M); 68 bool OptimizeFunctions(Module &M); 69 bool OptimizeGlobalVars(Module &M); 70 bool OptimizeGlobalAliases(Module &M); 71 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL); 72 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI); 73 }; 74} 75 76char GlobalOpt::ID = 0; 77static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer"); 78 79ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } 80 81namespace { 82 83/// GlobalStatus - As we analyze each global, keep track of some information 84/// about it. If we find out that the address of the global is taken, none of 85/// this info will be accurate. 86struct VISIBILITY_HIDDEN GlobalStatus { 87 /// isLoaded - True if the global is ever loaded. If the global isn't ever 88 /// loaded it can be deleted. 89 bool isLoaded; 90 91 /// StoredType - Keep track of what stores to the global look like. 92 /// 93 enum StoredType { 94 /// NotStored - There is no store to this global. It can thus be marked 95 /// constant. 96 NotStored, 97 98 /// isInitializerStored - This global is stored to, but the only thing 99 /// stored is the constant it was initialized with. This is only tracked 100 /// for scalar globals. 101 isInitializerStored, 102 103 /// isStoredOnce - This global is stored to, but only its initializer and 104 /// one other value is ever stored to it. If this global isStoredOnce, we 105 /// track the value stored to it in StoredOnceValue below. This is only 106 /// tracked for scalar globals. 107 isStoredOnce, 108 109 /// isStored - This global is stored to by multiple values or something else 110 /// that we cannot track. 111 isStored 112 } StoredType; 113 114 /// StoredOnceValue - If only one value (besides the initializer constant) is 115 /// ever stored to this global, keep track of what value it is. 116 Value *StoredOnceValue; 117 118 /// AccessingFunction/HasMultipleAccessingFunctions - These start out 119 /// null/false. When the first accessing function is noticed, it is recorded. 120 /// When a second different accessing function is noticed, 121 /// HasMultipleAccessingFunctions is set to true. 122 Function *AccessingFunction; 123 bool HasMultipleAccessingFunctions; 124 125 /// HasNonInstructionUser - Set to true if this global has a user that is not 126 /// an instruction (e.g. a constant expr or GV initializer). 127 bool HasNonInstructionUser; 128 129 /// HasPHIUser - Set to true if this global has a user that is a PHI node. 130 bool HasPHIUser; 131 132 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0), 133 AccessingFunction(0), HasMultipleAccessingFunctions(false), 134 HasNonInstructionUser(false), HasPHIUser(false) {} 135}; 136 137} 138 139/// ConstantIsDead - Return true if the specified constant is (transitively) 140/// dead. The constant may be used by other constants (e.g. constant arrays and 141/// constant exprs) as long as they are dead, but it cannot be used by anything 142/// else. 143static bool ConstantIsDead(Constant *C) { 144 if (isa<GlobalValue>(C)) return false; 145 146 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI) 147 if (Constant *CU = dyn_cast<Constant>(*UI)) { 148 if (!ConstantIsDead(CU)) return false; 149 } else 150 return false; 151 return true; 152} 153 154 155/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 156/// structure. If the global has its address taken, return true to indicate we 157/// can't do anything with it. 158/// 159static bool AnalyzeGlobal(Value *V, GlobalStatus &GS, 160 SmallPtrSet<PHINode*, 16> &PHIUsers) { 161 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 162 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) { 163 GS.HasNonInstructionUser = true; 164 165 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 166 167 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) { 168 if (!GS.HasMultipleAccessingFunctions) { 169 Function *F = I->getParent()->getParent(); 170 if (GS.AccessingFunction == 0) 171 GS.AccessingFunction = F; 172 else if (GS.AccessingFunction != F) 173 GS.HasMultipleAccessingFunctions = true; 174 } 175 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 176 GS.isLoaded = true; 177 if (LI->isVolatile()) return true; // Don't hack on volatile loads. 178 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 179 // Don't allow a store OF the address, only stores TO the address. 180 if (SI->getOperand(0) == V) return true; 181 182 if (SI->isVolatile()) return true; // Don't hack on volatile stores. 183 184 // If this is a direct store to the global (i.e., the global is a scalar 185 // value, not an aggregate), keep more specific information about 186 // stores. 187 if (GS.StoredType != GlobalStatus::isStored) { 188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){ 189 Value *StoredVal = SI->getOperand(0); 190 if (StoredVal == GV->getInitializer()) { 191 if (GS.StoredType < GlobalStatus::isInitializerStored) 192 GS.StoredType = GlobalStatus::isInitializerStored; 193 } else if (isa<LoadInst>(StoredVal) && 194 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 195 // G = G 196 if (GS.StoredType < GlobalStatus::isInitializerStored) 197 GS.StoredType = GlobalStatus::isInitializerStored; 198 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 199 GS.StoredType = GlobalStatus::isStoredOnce; 200 GS.StoredOnceValue = StoredVal; 201 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 202 GS.StoredOnceValue == StoredVal) { 203 // noop. 204 } else { 205 GS.StoredType = GlobalStatus::isStored; 206 } 207 } else { 208 GS.StoredType = GlobalStatus::isStored; 209 } 210 } 211 } else if (isa<GetElementPtrInst>(I)) { 212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 213 } else if (isa<SelectInst>(I)) { 214 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 215 } else if (PHINode *PN = dyn_cast<PHINode>(I)) { 216 // PHI nodes we can check just like select or GEP instructions, but we 217 // have to be careful about infinite recursion. 218 if (PHIUsers.insert(PN)) // Not already visited. 219 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 220 GS.HasPHIUser = true; 221 } else if (isa<CmpInst>(I)) { 222 } else if (isa<MemTransferInst>(I)) { 223 if (I->getOperand(1) == V) 224 GS.StoredType = GlobalStatus::isStored; 225 if (I->getOperand(2) == V) 226 GS.isLoaded = true; 227 } else if (isa<MemSetInst>(I)) { 228 assert(I->getOperand(1) == V && "Memset only takes one pointer!"); 229 GS.StoredType = GlobalStatus::isStored; 230 } else { 231 return true; // Any other non-load instruction might take address! 232 } 233 } else if (Constant *C = dyn_cast<Constant>(*UI)) { 234 GS.HasNonInstructionUser = true; 235 // We might have a dead and dangling constant hanging off of here. 236 if (!ConstantIsDead(C)) 237 return true; 238 } else { 239 GS.HasNonInstructionUser = true; 240 // Otherwise must be some other user. 241 return true; 242 } 243 244 return false; 245} 246 247static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 248 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 249 if (!CI) return 0; 250 unsigned IdxV = CI->getZExtValue(); 251 252 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 253 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 254 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 255 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 256 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 257 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 258 } else if (isa<ConstantAggregateZero>(Agg)) { 259 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 260 if (IdxV < STy->getNumElements()) 261 return Constant::getNullValue(STy->getElementType(IdxV)); 262 } else if (const SequentialType *STy = 263 dyn_cast<SequentialType>(Agg->getType())) { 264 return Constant::getNullValue(STy->getElementType()); 265 } 266 } else if (isa<UndefValue>(Agg)) { 267 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 268 if (IdxV < STy->getNumElements()) 269 return UndefValue::get(STy->getElementType(IdxV)); 270 } else if (const SequentialType *STy = 271 dyn_cast<SequentialType>(Agg->getType())) { 272 return UndefValue::get(STy->getElementType()); 273 } 274 } 275 return 0; 276} 277 278 279/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 280/// users of the global, cleaning up the obvious ones. This is largely just a 281/// quick scan over the use list to clean up the easy and obvious cruft. This 282/// returns true if it made a change. 283static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 284 bool Changed = false; 285 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 286 User *U = *UI++; 287 288 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 289 if (Init) { 290 // Replace the load with the initializer. 291 LI->replaceAllUsesWith(Init); 292 LI->eraseFromParent(); 293 Changed = true; 294 } 295 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 296 // Store must be unreachable or storing Init into the global. 297 SI->eraseFromParent(); 298 Changed = true; 299 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 300 if (CE->getOpcode() == Instruction::GetElementPtr) { 301 Constant *SubInit = 0; 302 if (Init) 303 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 304 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 305 } else if (CE->getOpcode() == Instruction::BitCast && 306 isa<PointerType>(CE->getType())) { 307 // Pointer cast, delete any stores and memsets to the global. 308 Changed |= CleanupConstantGlobalUsers(CE, 0); 309 } 310 311 if (CE->use_empty()) { 312 CE->destroyConstant(); 313 Changed = true; 314 } 315 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 316 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 317 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 318 // and will invalidate our notion of what Init is. 319 Constant *SubInit = 0; 320 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 321 ConstantExpr *CE = 322 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 323 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 324 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 325 } 326 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 327 328 if (GEP->use_empty()) { 329 GEP->eraseFromParent(); 330 Changed = true; 331 } 332 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 333 if (MI->getRawDest() == V) { 334 MI->eraseFromParent(); 335 Changed = true; 336 } 337 338 } else if (Constant *C = dyn_cast<Constant>(U)) { 339 // If we have a chain of dead constantexprs or other things dangling from 340 // us, and if they are all dead, nuke them without remorse. 341 if (ConstantIsDead(C)) { 342 C->destroyConstant(); 343 // This could have invalidated UI, start over from scratch. 344 CleanupConstantGlobalUsers(V, Init); 345 return true; 346 } 347 } 348 } 349 return Changed; 350} 351 352/// isSafeSROAElementUse - Return true if the specified instruction is a safe 353/// user of a derived expression from a global that we want to SROA. 354static bool isSafeSROAElementUse(Value *V) { 355 // We might have a dead and dangling constant hanging off of here. 356 if (Constant *C = dyn_cast<Constant>(V)) 357 return ConstantIsDead(C); 358 359 Instruction *I = dyn_cast<Instruction>(V); 360 if (!I) return false; 361 362 // Loads are ok. 363 if (isa<LoadInst>(I)) return true; 364 365 // Stores *to* the pointer are ok. 366 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 367 return SI->getOperand(0) != V; 368 369 // Otherwise, it must be a GEP. 370 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 371 if (GEPI == 0) return false; 372 373 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 374 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 375 return false; 376 377 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end(); 378 I != E; ++I) 379 if (!isSafeSROAElementUse(*I)) 380 return false; 381 return true; 382} 383 384 385/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 386/// Look at it and its uses and decide whether it is safe to SROA this global. 387/// 388static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 389 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 390 if (!isa<GetElementPtrInst>(U) && 391 (!isa<ConstantExpr>(U) || 392 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 393 return false; 394 395 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 396 // don't like < 3 operand CE's, and we don't like non-constant integer 397 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 398 // value of C. 399 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 400 !cast<Constant>(U->getOperand(1))->isNullValue() || 401 !isa<ConstantInt>(U->getOperand(2))) 402 return false; 403 404 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 405 ++GEPI; // Skip over the pointer index. 406 407 // If this is a use of an array allocation, do a bit more checking for sanity. 408 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 409 uint64_t NumElements = AT->getNumElements(); 410 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 411 412 // Check to make sure that index falls within the array. If not, 413 // something funny is going on, so we won't do the optimization. 414 // 415 if (Idx->getZExtValue() >= NumElements) 416 return false; 417 418 // We cannot scalar repl this level of the array unless any array 419 // sub-indices are in-range constants. In particular, consider: 420 // A[0][i]. We cannot know that the user isn't doing invalid things like 421 // allowing i to index an out-of-range subscript that accesses A[1]. 422 // 423 // Scalar replacing *just* the outer index of the array is probably not 424 // going to be a win anyway, so just give up. 425 for (++GEPI; // Skip array index. 426 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI)); 427 ++GEPI) { 428 uint64_t NumElements; 429 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 430 NumElements = SubArrayTy->getNumElements(); 431 else 432 NumElements = cast<VectorType>(*GEPI)->getNumElements(); 433 434 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 435 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 436 return false; 437 } 438 } 439 440 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) 441 if (!isSafeSROAElementUse(*I)) 442 return false; 443 return true; 444} 445 446/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 447/// is safe for us to perform this transformation. 448/// 449static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 450 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 451 UI != E; ++UI) { 452 if (!IsUserOfGlobalSafeForSRA(*UI, GV)) 453 return false; 454 } 455 return true; 456} 457 458 459/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 460/// variable. This opens the door for other optimizations by exposing the 461/// behavior of the program in a more fine-grained way. We have determined that 462/// this transformation is safe already. We return the first global variable we 463/// insert so that the caller can reprocess it. 464static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { 465 // Make sure this global only has simple uses that we can SRA. 466 if (!GlobalUsersSafeToSRA(GV)) 467 return 0; 468 469 assert(GV->hasLocalLinkage() && !GV->isConstant()); 470 Constant *Init = GV->getInitializer(); 471 const Type *Ty = Init->getType(); 472 473 std::vector<GlobalVariable*> NewGlobals; 474 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 475 476 // Get the alignment of the global, either explicit or target-specific. 477 unsigned StartAlignment = GV->getAlignment(); 478 if (StartAlignment == 0) 479 StartAlignment = TD.getABITypeAlignment(GV->getType()); 480 481 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 482 NewGlobals.reserve(STy->getNumElements()); 483 const StructLayout &Layout = *TD.getStructLayout(STy); 484 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 485 Constant *In = getAggregateConstantElement(Init, 486 ConstantInt::get(Type::Int32Ty, i)); 487 assert(In && "Couldn't get element of initializer?"); 488 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 489 GlobalVariable::InternalLinkage, 490 In, GV->getName()+"."+utostr(i), 491 (Module *)NULL, 492 GV->isThreadLocal(), 493 GV->getType()->getAddressSpace()); 494 Globals.insert(GV, NGV); 495 NewGlobals.push_back(NGV); 496 497 // Calculate the known alignment of the field. If the original aggregate 498 // had 256 byte alignment for example, something might depend on that: 499 // propagate info to each field. 500 uint64_t FieldOffset = Layout.getElementOffset(i); 501 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 502 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 503 NGV->setAlignment(NewAlign); 504 } 505 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 506 unsigned NumElements = 0; 507 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 508 NumElements = ATy->getNumElements(); 509 else 510 NumElements = cast<VectorType>(STy)->getNumElements(); 511 512 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 513 return 0; // It's not worth it. 514 NewGlobals.reserve(NumElements); 515 516 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType()); 517 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 518 for (unsigned i = 0, e = NumElements; i != e; ++i) { 519 Constant *In = getAggregateConstantElement(Init, 520 ConstantInt::get(Type::Int32Ty, i)); 521 assert(In && "Couldn't get element of initializer?"); 522 523 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 524 GlobalVariable::InternalLinkage, 525 In, GV->getName()+"."+utostr(i), 526 (Module *)NULL, 527 GV->isThreadLocal(), 528 GV->getType()->getAddressSpace()); 529 Globals.insert(GV, NGV); 530 NewGlobals.push_back(NGV); 531 532 // Calculate the known alignment of the field. If the original aggregate 533 // had 256 byte alignment for example, something might depend on that: 534 // propagate info to each field. 535 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 536 if (NewAlign > EltAlign) 537 NGV->setAlignment(NewAlign); 538 } 539 } 540 541 if (NewGlobals.empty()) 542 return 0; 543 544 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV; 545 546 Constant *NullInt = Constant::getNullValue(Type::Int32Ty); 547 548 // Loop over all of the uses of the global, replacing the constantexpr geps, 549 // with smaller constantexpr geps or direct references. 550 while (!GV->use_empty()) { 551 User *GEP = GV->use_back(); 552 assert(((isa<ConstantExpr>(GEP) && 553 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 554 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 555 556 // Ignore the 1th operand, which has to be zero or else the program is quite 557 // broken (undefined). Get the 2nd operand, which is the structure or array 558 // index. 559 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 560 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 561 562 Value *NewPtr = NewGlobals[Val]; 563 564 // Form a shorter GEP if needed. 565 if (GEP->getNumOperands() > 3) { 566 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 567 SmallVector<Constant*, 8> Idxs; 568 Idxs.push_back(NullInt); 569 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 570 Idxs.push_back(CE->getOperand(i)); 571 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 572 &Idxs[0], Idxs.size()); 573 } else { 574 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 575 SmallVector<Value*, 8> Idxs; 576 Idxs.push_back(NullInt); 577 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 578 Idxs.push_back(GEPI->getOperand(i)); 579 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 580 GEPI->getName()+"."+utostr(Val), GEPI); 581 } 582 } 583 GEP->replaceAllUsesWith(NewPtr); 584 585 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 586 GEPI->eraseFromParent(); 587 else 588 cast<ConstantExpr>(GEP)->destroyConstant(); 589 } 590 591 // Delete the old global, now that it is dead. 592 Globals.erase(GV); 593 ++NumSRA; 594 595 // Loop over the new globals array deleting any globals that are obviously 596 // dead. This can arise due to scalarization of a structure or an array that 597 // has elements that are dead. 598 unsigned FirstGlobal = 0; 599 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 600 if (NewGlobals[i]->use_empty()) { 601 Globals.erase(NewGlobals[i]); 602 if (FirstGlobal == i) ++FirstGlobal; 603 } 604 605 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 606} 607 608/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 609/// value will trap if the value is dynamically null. PHIs keeps track of any 610/// phi nodes we've seen to avoid reprocessing them. 611static bool AllUsesOfValueWillTrapIfNull(Value *V, 612 SmallPtrSet<PHINode*, 8> &PHIs) { 613 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 614 if (isa<LoadInst>(*UI)) { 615 // Will trap. 616 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 617 if (SI->getOperand(0) == V) { 618 //cerr << "NONTRAPPING USE: " << **UI; 619 return false; // Storing the value. 620 } 621 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 622 if (CI->getOperand(0) != V) { 623 //cerr << "NONTRAPPING USE: " << **UI; 624 return false; // Not calling the ptr 625 } 626 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 627 if (II->getOperand(0) != V) { 628 //cerr << "NONTRAPPING USE: " << **UI; 629 return false; // Not calling the ptr 630 } 631 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) { 632 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 633 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) { 634 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 635 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { 636 // If we've already seen this phi node, ignore it, it has already been 637 // checked. 638 if (PHIs.insert(PN)) 639 return AllUsesOfValueWillTrapIfNull(PN, PHIs); 640 } else if (isa<ICmpInst>(*UI) && 641 isa<ConstantPointerNull>(UI->getOperand(1))) { 642 // Ignore setcc X, null 643 } else { 644 //cerr << "NONTRAPPING USE: " << **UI; 645 return false; 646 } 647 return true; 648} 649 650/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 651/// from GV will trap if the loaded value is null. Note that this also permits 652/// comparisons of the loaded value against null, as a special case. 653static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) { 654 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI) 655 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 656 SmallPtrSet<PHINode*, 8> PHIs; 657 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 658 return false; 659 } else if (isa<StoreInst>(*UI)) { 660 // Ignore stores to the global. 661 } else { 662 // We don't know or understand this user, bail out. 663 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI; 664 return false; 665 } 666 667 return true; 668} 669 670static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 671 bool Changed = false; 672 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 673 Instruction *I = cast<Instruction>(*UI++); 674 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 675 LI->setOperand(0, NewV); 676 Changed = true; 677 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 678 if (SI->getOperand(1) == V) { 679 SI->setOperand(1, NewV); 680 Changed = true; 681 } 682 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 683 if (I->getOperand(0) == V) { 684 // Calling through the pointer! Turn into a direct call, but be careful 685 // that the pointer is not also being passed as an argument. 686 I->setOperand(0, NewV); 687 Changed = true; 688 bool PassedAsArg = false; 689 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i) 690 if (I->getOperand(i) == V) { 691 PassedAsArg = true; 692 I->setOperand(i, NewV); 693 } 694 695 if (PassedAsArg) { 696 // Being passed as an argument also. Be careful to not invalidate UI! 697 UI = V->use_begin(); 698 } 699 } 700 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 701 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 702 ConstantExpr::getCast(CI->getOpcode(), 703 NewV, CI->getType())); 704 if (CI->use_empty()) { 705 Changed = true; 706 CI->eraseFromParent(); 707 } 708 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 709 // Should handle GEP here. 710 SmallVector<Constant*, 8> Idxs; 711 Idxs.reserve(GEPI->getNumOperands()-1); 712 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 713 i != e; ++i) 714 if (Constant *C = dyn_cast<Constant>(*i)) 715 Idxs.push_back(C); 716 else 717 break; 718 if (Idxs.size() == GEPI->getNumOperands()-1) 719 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 720 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 721 Idxs.size())); 722 if (GEPI->use_empty()) { 723 Changed = true; 724 GEPI->eraseFromParent(); 725 } 726 } 727 } 728 729 return Changed; 730} 731 732 733/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 734/// value stored into it. If there are uses of the loaded value that would trap 735/// if the loaded value is dynamically null, then we know that they cannot be 736/// reachable with a null optimize away the load. 737static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 738 bool Changed = false; 739 740 // Keep track of whether we are able to remove all the uses of the global 741 // other than the store that defines it. 742 bool AllNonStoreUsesGone = true; 743 744 // Replace all uses of loads with uses of uses of the stored value. 745 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){ 746 User *GlobalUser = *GUI++; 747 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) { 748 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 749 // If we were able to delete all uses of the loads 750 if (LI->use_empty()) { 751 LI->eraseFromParent(); 752 Changed = true; 753 } else { 754 AllNonStoreUsesGone = false; 755 } 756 } else if (isa<StoreInst>(GlobalUser)) { 757 // Ignore the store that stores "LV" to the global. 758 assert(GlobalUser->getOperand(1) == GV && 759 "Must be storing *to* the global"); 760 } else { 761 AllNonStoreUsesGone = false; 762 763 // If we get here we could have other crazy uses that are transitively 764 // loaded. 765 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || 766 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!"); 767 } 768 } 769 770 if (Changed) { 771 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV; 772 ++NumGlobUses; 773 } 774 775 // If we nuked all of the loads, then none of the stores are needed either, 776 // nor is the global. 777 if (AllNonStoreUsesGone) { 778 DOUT << " *** GLOBAL NOW DEAD!\n"; 779 CleanupConstantGlobalUsers(GV, 0); 780 if (GV->use_empty()) { 781 GV->eraseFromParent(); 782 ++NumDeleted; 783 } 784 Changed = true; 785 } 786 return Changed; 787} 788 789/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 790/// instructions that are foldable. 791static void ConstantPropUsersOf(Value *V) { 792 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 793 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 794 if (Constant *NewC = ConstantFoldInstruction(I)) { 795 I->replaceAllUsesWith(NewC); 796 797 // Advance UI to the next non-I use to avoid invalidating it! 798 // Instructions could multiply use V. 799 while (UI != E && *UI == I) 800 ++UI; 801 I->eraseFromParent(); 802 } 803} 804 805/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 806/// variable, and transforms the program as if it always contained the result of 807/// the specified malloc. Because it is always the result of the specified 808/// malloc, there is no reason to actually DO the malloc. Instead, turn the 809/// malloc into a global, and any loads of GV as uses of the new global. 810static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 811 MallocInst *MI) { 812 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI; 813 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize()); 814 815 if (NElements->getZExtValue() != 1) { 816 // If we have an array allocation, transform it to a single element 817 // allocation to make the code below simpler. 818 Type *NewTy = ArrayType::get(MI->getAllocatedType(), 819 NElements->getZExtValue()); 820 MallocInst *NewMI = 821 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty), 822 MI->getAlignment(), MI->getName(), MI); 823 Value* Indices[2]; 824 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty); 825 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2, 826 NewMI->getName()+".el0", MI); 827 MI->replaceAllUsesWith(NewGEP); 828 MI->eraseFromParent(); 829 MI = NewMI; 830 } 831 832 // Create the new global variable. The contents of the malloc'd memory is 833 // undefined, so initialize with an undef value. 834 Constant *Init = UndefValue::get(MI->getAllocatedType()); 835 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false, 836 GlobalValue::InternalLinkage, Init, 837 GV->getName()+".body", 838 (Module *)NULL, 839 GV->isThreadLocal()); 840 // FIXME: This new global should have the alignment returned by malloc. Code 841 // could depend on malloc returning large alignment (on the mac, 16 bytes) but 842 // this would only guarantee some lower alignment. 843 GV->getParent()->getGlobalList().insert(GV, NewGV); 844 845 // Anything that used the malloc now uses the global directly. 846 MI->replaceAllUsesWith(NewGV); 847 848 Constant *RepValue = NewGV; 849 if (NewGV->getType() != GV->getType()->getElementType()) 850 RepValue = ConstantExpr::getBitCast(RepValue, 851 GV->getType()->getElementType()); 852 853 // If there is a comparison against null, we will insert a global bool to 854 // keep track of whether the global was initialized yet or not. 855 GlobalVariable *InitBool = 856 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage, 857 ConstantInt::getFalse(), GV->getName()+".init", 858 (Module *)NULL, GV->isThreadLocal()); 859 bool InitBoolUsed = false; 860 861 // Loop over all uses of GV, processing them in turn. 862 std::vector<StoreInst*> Stores; 863 while (!GV->use_empty()) 864 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) { 865 while (!LI->use_empty()) { 866 Use &LoadUse = LI->use_begin().getUse(); 867 if (!isa<ICmpInst>(LoadUse.getUser())) 868 LoadUse = RepValue; 869 else { 870 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser()); 871 // Replace the cmp X, 0 with a use of the bool value. 872 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI); 873 InitBoolUsed = true; 874 switch (CI->getPredicate()) { 875 default: assert(0 && "Unknown ICmp Predicate!"); 876 case ICmpInst::ICMP_ULT: 877 case ICmpInst::ICMP_SLT: 878 LV = ConstantInt::getFalse(); // X < null -> always false 879 break; 880 case ICmpInst::ICMP_ULE: 881 case ICmpInst::ICMP_SLE: 882 case ICmpInst::ICMP_EQ: 883 LV = BinaryOperator::CreateNot(LV, "notinit", CI); 884 break; 885 case ICmpInst::ICMP_NE: 886 case ICmpInst::ICMP_UGE: 887 case ICmpInst::ICMP_SGE: 888 case ICmpInst::ICMP_UGT: 889 case ICmpInst::ICMP_SGT: 890 break; // no change. 891 } 892 CI->replaceAllUsesWith(LV); 893 CI->eraseFromParent(); 894 } 895 } 896 LI->eraseFromParent(); 897 } else { 898 StoreInst *SI = cast<StoreInst>(GV->use_back()); 899 // The global is initialized when the store to it occurs. 900 new StoreInst(ConstantInt::getTrue(), InitBool, SI); 901 SI->eraseFromParent(); 902 } 903 904 // If the initialization boolean was used, insert it, otherwise delete it. 905 if (!InitBoolUsed) { 906 while (!InitBool->use_empty()) // Delete initializations 907 cast<Instruction>(InitBool->use_back())->eraseFromParent(); 908 delete InitBool; 909 } else 910 GV->getParent()->getGlobalList().insert(GV, InitBool); 911 912 913 // Now the GV is dead, nuke it and the malloc. 914 GV->eraseFromParent(); 915 MI->eraseFromParent(); 916 917 // To further other optimizations, loop over all users of NewGV and try to 918 // constant prop them. This will promote GEP instructions with constant 919 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 920 ConstantPropUsersOf(NewGV); 921 if (RepValue != NewGV) 922 ConstantPropUsersOf(RepValue); 923 924 return NewGV; 925} 926 927/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 928/// to make sure that there are no complex uses of V. We permit simple things 929/// like dereferencing the pointer, but not storing through the address, unless 930/// it is to the specified global. 931static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V, 932 GlobalVariable *GV, 933 SmallPtrSet<PHINode*, 8> &PHIs) { 934 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
|
937 938 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { 939 continue; // Fine, ignore. 940 } 941 942 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 943 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 944 return false; // Storing the pointer itself... bad. 945 continue; // Otherwise, storing through it, or storing into GV... fine. 946 } 947 948 if (isa<GetElementPtrInst>(Inst)) { 949 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs)) 950 return false; 951 continue; 952 } 953 954 if (PHINode *PN = dyn_cast<PHINode>(Inst)) { 955 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 956 // cycles. 957 if (PHIs.insert(PN)) 958 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 959 return false; 960 continue; 961 } 962 963 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { 964 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) 965 return false; 966 continue; 967 } 968 969 return false; 970 } 971 return true; 972} 973 974/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 975/// somewhere. Transform all uses of the allocation into loads from the 976/// global and uses of the resultant pointer. Further, delete the store into 977/// GV. This assumes that these value pass the 978/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 979static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 980 GlobalVariable *GV) { 981 while (!Alloc->use_empty()) { 982 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 983 Instruction *InsertPt = U; 984 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 985 // If this is the store of the allocation into the global, remove it. 986 if (SI->getOperand(1) == GV) { 987 SI->eraseFromParent(); 988 continue; 989 } 990 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 991 // Insert the load in the corresponding predecessor, not right before the 992 // PHI. 993 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator(); 994 } else if (isa<BitCastInst>(U)) { 995 // Must be bitcast between the malloc and store to initialize the global. 996 ReplaceUsesOfMallocWithGlobal(U, GV); 997 U->eraseFromParent(); 998 continue; 999 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 1000 // If this is a "GEP bitcast" and the user is a store to the global, then 1001 // just process it as a bitcast. 1002 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse()) 1003 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back())) 1004 if (SI->getOperand(1) == GV) { 1005 // Must be bitcast GEP between the malloc and store to initialize 1006 // the global. 1007 ReplaceUsesOfMallocWithGlobal(GEPI, GV); 1008 GEPI->eraseFromParent(); 1009 continue; 1010 } 1011 } 1012 1013 // Insert a load from the global, and use it instead of the malloc. 1014 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 1015 U->replaceUsesOfWith(Alloc, NL); 1016 } 1017} 1018 1019/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi 1020/// of a load) are simple enough to perform heap SRA on. This permits GEP's 1021/// that index through the array and struct field, icmps of null, and PHIs. 1022static bool LoadUsesSimpleEnoughForHeapSRA(Value *V, 1023 SmallPtrSet<PHINode*, 32> &LoadUsingPHIs, 1024 SmallPtrSet<PHINode*, 32> &LoadUsingPHIsPerLoad) { 1025 // We permit two users of the load: setcc comparing against the null 1026 // pointer, and a getelementptr of a specific form. 1027 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 1028 Instruction *User = cast<Instruction>(*UI); 1029 1030 // Comparison against null is ok. 1031 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) { 1032 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 1033 return false; 1034 continue; 1035 } 1036 1037 // getelementptr is also ok, but only a simple form. 1038 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1039 // Must index into the array and into the struct. 1040 if (GEPI->getNumOperands() < 3) 1041 return false; 1042 1043 // Otherwise the GEP is ok. 1044 continue; 1045 } 1046 1047 if (PHINode *PN = dyn_cast<PHINode>(User)) { 1048 if (!LoadUsingPHIsPerLoad.insert(PN)) 1049 // This means some phi nodes are dependent on each other. 1050 // Avoid infinite looping! 1051 return false; 1052 if (!LoadUsingPHIs.insert(PN)) 1053 // If we have already analyzed this PHI, then it is safe. 1054 continue; 1055 1056 // Make sure all uses of the PHI are simple enough to transform. 1057 if (!LoadUsesSimpleEnoughForHeapSRA(PN, 1058 LoadUsingPHIs, LoadUsingPHIsPerLoad)) 1059 return false; 1060 1061 continue; 1062 } 1063 1064 // Otherwise we don't know what this is, not ok. 1065 return false; 1066 } 1067 1068 return true; 1069} 1070 1071 1072/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1073/// GV are simple enough to perform HeapSRA, return true. 1074static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 1075 MallocInst *MI) { 1076 SmallPtrSet<PHINode*, 32> LoadUsingPHIs; 1077 SmallPtrSet<PHINode*, 32> LoadUsingPHIsPerLoad; 1078 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 1079 ++UI) 1080 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 1081 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs, 1082 LoadUsingPHIsPerLoad)) 1083 return false; 1084 LoadUsingPHIsPerLoad.clear(); 1085 } 1086 1087 // If we reach here, we know that all uses of the loads and transitive uses 1088 // (through PHI nodes) are simple enough to transform. However, we don't know 1089 // that all inputs the to the PHI nodes are in the same equivalence sets. 1090 // Check to verify that all operands of the PHIs are either PHIS that can be 1091 // transformed, loads from GV, or MI itself. 1092 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(), 1093 E = LoadUsingPHIs.end(); I != E; ++I) { 1094 PHINode *PN = *I; 1095 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) { 1096 Value *InVal = PN->getIncomingValue(op); 1097 1098 // PHI of the stored value itself is ok. 1099 if (InVal == MI) continue; 1100 1101 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) { 1102 // One of the PHIs in our set is (optimistically) ok. 1103 if (LoadUsingPHIs.count(InPN)) 1104 continue; 1105 return false; 1106 } 1107 1108 // Load from GV is ok. 1109 if (LoadInst *LI = dyn_cast<LoadInst>(InVal)) 1110 if (LI->getOperand(0) == GV) 1111 continue; 1112 1113 // UNDEF? NULL? 1114 1115 // Anything else is rejected. 1116 return false; 1117 } 1118 } 1119 1120 return true; 1121} 1122 1123static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, 1124 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1125 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1126 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V]; 1127 1128 if (FieldNo >= FieldVals.size()) 1129 FieldVals.resize(FieldNo+1); 1130 1131 // If we already have this value, just reuse the previously scalarized 1132 // version. 1133 if (Value *FieldVal = FieldVals[FieldNo]) 1134 return FieldVal; 1135 1136 // Depending on what instruction this is, we have several cases. 1137 Value *Result; 1138 if (LoadInst *LI = dyn_cast<LoadInst>(V)) { 1139 // This is a scalarized version of the load from the global. Just create 1140 // a new Load of the scalarized global. 1141 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo, 1142 InsertedScalarizedValues, 1143 PHIsToRewrite), 1144 LI->getName()+".f" + utostr(FieldNo), LI); 1145 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 1146 // PN's type is pointer to struct. Make a new PHI of pointer to struct 1147 // field. 1148 const StructType *ST = 1149 cast<StructType>(cast<PointerType>(PN->getType())->getElementType()); 1150 1151 Result =PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)), 1152 PN->getName()+".f"+utostr(FieldNo), PN); 1153 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); 1154 } else { 1155 assert(0 && "Unknown usable value"); 1156 Result = 0; 1157 } 1158 1159 return FieldVals[FieldNo] = Result; 1160} 1161 1162/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1163/// the load, rewrite the derived value to use the HeapSRoA'd load. 1164static void RewriteHeapSROALoadUser(Instruction *LoadUser, 1165 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1166 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1167 // If this is a comparison against null, handle it. 1168 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1169 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1170 // If we have a setcc of the loaded pointer, we can use a setcc of any 1171 // field. 1172 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0, 1173 InsertedScalarizedValues, PHIsToRewrite); 1174 1175 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1176 Constant::getNullValue(NPtr->getType()), 1177 SCI->getName(), SCI); 1178 SCI->replaceAllUsesWith(New); 1179 SCI->eraseFromParent(); 1180 return; 1181 } 1182 1183 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...' 1184 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1185 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1186 && "Unexpected GEPI!"); 1187 1188 // Load the pointer for this field. 1189 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1190 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo, 1191 InsertedScalarizedValues, PHIsToRewrite); 1192 1193 // Create the new GEP idx vector. 1194 SmallVector<Value*, 8> GEPIdx; 1195 GEPIdx.push_back(GEPI->getOperand(1)); 1196 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1197 1198 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1199 GEPIdx.begin(), GEPIdx.end(), 1200 GEPI->getName(), GEPI); 1201 GEPI->replaceAllUsesWith(NGEPI); 1202 GEPI->eraseFromParent(); 1203 return; 1204 } 1205 1206 // Recursively transform the users of PHI nodes. This will lazily create the 1207 // PHIs that are needed for individual elements. Keep track of what PHIs we 1208 // see in InsertedScalarizedValues so that we don't get infinite loops (very 1209 // antisocial). If the PHI is already in InsertedScalarizedValues, it has 1210 // already been seen first by another load, so its uses have already been 1211 // processed. 1212 PHINode *PN = cast<PHINode>(LoadUser); 1213 bool Inserted; 1214 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos; 1215 tie(InsertPos, Inserted) = 1216 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>())); 1217 if (!Inserted) return; 1218 1219 // If this is the first time we've seen this PHI, recursively process all 1220 // users. 1221 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) { 1222 Instruction *User = cast<Instruction>(*UI++); 1223 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1224 } 1225} 1226 1227/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1228/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1229/// use FieldGlobals instead. All uses of loaded values satisfy 1230/// AllGlobalLoadUsesSimpleEnoughForHeapSRA. 1231static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1232 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1233 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1234 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end(); 1235 UI != E; ) { 1236 Instruction *User = cast<Instruction>(*UI++); 1237 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1238 } 1239 1240 if (Load->use_empty()) { 1241 Load->eraseFromParent(); 1242 InsertedScalarizedValues.erase(Load); 1243 } 1244} 1245 1246/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1247/// it up into multiple allocations of arrays of the fields. 1248static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1249 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1250 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1251 1252 // There is guaranteed to be at least one use of the malloc (storing 1253 // it into GV). If there are other uses, change them to be uses of 1254 // the global to simplify later code. This also deletes the store 1255 // into GV. 1256 ReplaceUsesOfMallocWithGlobal(MI, GV); 1257 1258 // Okay, at this point, there are no users of the malloc. Insert N 1259 // new mallocs at the same place as MI, and N globals. 1260 std::vector<Value*> FieldGlobals; 1261 std::vector<MallocInst*> FieldMallocs; 1262 1263 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1264 const Type *FieldTy = STy->getElementType(FieldNo); 1265 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1266 1267 GlobalVariable *NGV = 1268 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1269 Constant::getNullValue(PFieldTy), 1270 GV->getName() + ".f" + utostr(FieldNo), GV, 1271 GV->isThreadLocal()); 1272 FieldGlobals.push_back(NGV); 1273 1274 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1275 MI->getName() + ".f" + utostr(FieldNo),MI); 1276 FieldMallocs.push_back(NMI); 1277 new StoreInst(NMI, NGV, MI); 1278 } 1279 1280 // The tricky aspect of this transformation is handling the case when malloc 1281 // fails. In the original code, malloc failing would set the result pointer 1282 // of malloc to null. In this case, some mallocs could succeed and others 1283 // could fail. As such, we emit code that looks like this: 1284 // F0 = malloc(field0) 1285 // F1 = malloc(field1) 1286 // F2 = malloc(field2) 1287 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1288 // if (F0) { free(F0); F0 = 0; } 1289 // if (F1) { free(F1); F1 = 0; } 1290 // if (F2) { free(F2); F2 = 0; } 1291 // } 1292 Value *RunningOr = 0; 1293 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1294 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1295 Constant::getNullValue(FieldMallocs[i]->getType()), 1296 "isnull", MI); 1297 if (!RunningOr) 1298 RunningOr = Cond; // First seteq 1299 else 1300 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI); 1301 } 1302 1303 // Split the basic block at the old malloc. 1304 BasicBlock *OrigBB = MI->getParent(); 1305 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1306 1307 // Create the block to check the first condition. Put all these blocks at the 1308 // end of the function as they are unlikely to be executed. 1309 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1310 OrigBB->getParent()); 1311 1312 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1313 // branch on RunningOr. 1314 OrigBB->getTerminator()->eraseFromParent(); 1315 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1316 1317 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1318 // pointer, because some may be null while others are not. 1319 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1320 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1321 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1322 Constant::getNullValue(GVVal->getType()), 1323 "tmp", NullPtrBlock); 1324 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1325 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1326 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1327 1328 // Fill in FreeBlock. 1329 new FreeInst(GVVal, FreeBlock); 1330 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1331 FreeBlock); 1332 BranchInst::Create(NextBlock, FreeBlock); 1333 1334 NullPtrBlock = NextBlock; 1335 } 1336 1337 BranchInst::Create(ContBB, NullPtrBlock); 1338 1339 // MI is no longer needed, remove it. 1340 MI->eraseFromParent(); 1341 1342 /// InsertedScalarizedLoads - As we process loads, if we can't immediately 1343 /// update all uses of the load, keep track of what scalarized loads are 1344 /// inserted for a given load. 1345 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues; 1346 InsertedScalarizedValues[GV] = FieldGlobals; 1347 1348 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite; 1349 1350 // Okay, the malloc site is completely handled. All of the uses of GV are now 1351 // loads, and all uses of those loads are simple. Rewrite them to use loads 1352 // of the per-field globals instead. 1353 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) { 1354 Instruction *User = cast<Instruction>(*UI++); 1355 1356 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1357 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite); 1358 continue; 1359 } 1360 1361 // Must be a store of null. 1362 StoreInst *SI = cast<StoreInst>(User); 1363 assert(isa<ConstantPointerNull>(SI->getOperand(0)) && 1364 "Unexpected heap-sra user!"); 1365 1366 // Insert a store of null into each global. 1367 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1368 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType()); 1369 Constant *Null = Constant::getNullValue(PT->getElementType()); 1370 new StoreInst(Null, FieldGlobals[i], SI); 1371 } 1372 // Erase the original store. 1373 SI->eraseFromParent(); 1374 } 1375 1376 // While we have PHIs that are interesting to rewrite, do it. 1377 while (!PHIsToRewrite.empty()) { 1378 PHINode *PN = PHIsToRewrite.back().first; 1379 unsigned FieldNo = PHIsToRewrite.back().second; 1380 PHIsToRewrite.pop_back(); 1381 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]); 1382 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi"); 1383 1384 // Add all the incoming values. This can materialize more phis. 1385 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1386 Value *InVal = PN->getIncomingValue(i); 1387 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues, 1388 PHIsToRewrite); 1389 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i)); 1390 } 1391 } 1392 1393 // Drop all inter-phi links and any loads that made it this far. 1394 for (DenseMap<Value*, std::vector<Value*> >::iterator 1395 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1396 I != E; ++I) { 1397 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1398 PN->dropAllReferences(); 1399 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1400 LI->dropAllReferences(); 1401 } 1402 1403 // Delete all the phis and loads now that inter-references are dead. 1404 for (DenseMap<Value*, std::vector<Value*> >::iterator 1405 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1406 I != E; ++I) { 1407 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1408 PN->eraseFromParent(); 1409 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1410 LI->eraseFromParent(); 1411 } 1412 1413 // The old global is now dead, remove it. 1414 GV->eraseFromParent(); 1415 1416 ++NumHeapSRA; 1417 return cast<GlobalVariable>(FieldGlobals[0]); 1418} 1419 1420/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1421/// pointer global variable with a single value stored it that is a malloc or 1422/// cast of malloc. 1423static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, 1424 MallocInst *MI, 1425 Module::global_iterator &GVI, 1426 TargetData &TD) { 1427 // If this is a malloc of an abstract type, don't touch it. 1428 if (!MI->getAllocatedType()->isSized()) 1429 return false; 1430 1431 // We can't optimize this global unless all uses of it are *known* to be 1432 // of the malloc value, not of the null initializer value (consider a use 1433 // that compares the global's value against zero to see if the malloc has 1434 // been reached). To do this, we check to see if all uses of the global 1435 // would trap if the global were null: this proves that they must all 1436 // happen after the malloc. 1437 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1438 return false; 1439 1440 // We can't optimize this if the malloc itself is used in a complex way, 1441 // for example, being stored into multiple globals. This allows the 1442 // malloc to be stored into the specified global, loaded setcc'd, and 1443 // GEP'd. These are all things we could transform to using the global 1444 // for. 1445 { 1446 SmallPtrSet<PHINode*, 8> PHIs; 1447 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1448 return false; 1449 } 1450 1451 1452 // If we have a global that is only initialized with a fixed size malloc, 1453 // transform the program to use global memory instead of malloc'd memory. 1454 // This eliminates dynamic allocation, avoids an indirection accessing the 1455 // data, and exposes the resultant global to further GlobalOpt. 1456 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1457 // Restrict this transformation to only working on small allocations 1458 // (2048 bytes currently), as we don't want to introduce a 16M global or 1459 // something. 1460 if (NElements->getZExtValue()* 1461 TD.getTypeAllocSize(MI->getAllocatedType()) < 2048) { 1462 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1463 return true; 1464 } 1465 } 1466 1467 // If the allocation is an array of structures, consider transforming this 1468 // into multiple malloc'd arrays, one for each field. This is basically 1469 // SRoA for malloc'd memory. 1470 const Type *AllocTy = MI->getAllocatedType(); 1471 1472 // If this is an allocation of a fixed size array of structs, analyze as a 1473 // variable size array. malloc [100 x struct],1 -> malloc struct, 100 1474 if (!MI->isArrayAllocation()) 1475 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy)) 1476 AllocTy = AT->getElementType(); 1477 1478 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) { 1479 // This the structure has an unreasonable number of fields, leave it 1480 // alone. 1481 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 && 1482 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1483 1484 // If this is a fixed size array, transform the Malloc to be an alloc of 1485 // structs. malloc [100 x struct],1 -> malloc struct, 100 1486 if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) { 1487 MallocInst *NewMI = 1488 new MallocInst(AllocSTy, 1489 ConstantInt::get(Type::Int32Ty, AT->getNumElements()), 1490 "", MI); 1491 NewMI->takeName(MI); 1492 Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI); 1493 MI->replaceAllUsesWith(Cast); 1494 MI->eraseFromParent(); 1495 MI = NewMI; 1496 } 1497 1498 GVI = PerformHeapAllocSRoA(GV, MI); 1499 return true; 1500 } 1501 } 1502 1503 return false; 1504} 1505 1506// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1507// that only one value (besides its initializer) is ever stored to the global. 1508static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1509 Module::global_iterator &GVI, 1510 TargetData &TD) { 1511 // Ignore no-op GEPs and bitcasts. 1512 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1513 1514 // If we are dealing with a pointer global that is initialized to null and 1515 // only has one (non-null) value stored into it, then we can optimize any 1516 // users of the loaded value (often calls and loads) that would trap if the 1517 // value was null. 1518 if (isa<PointerType>(GV->getInitializer()->getType()) && 1519 GV->getInitializer()->isNullValue()) { 1520 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1521 if (GV->getInitializer()->getType() != SOVC->getType()) 1522 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1523 1524 // Optimize away any trapping uses of the loaded value. 1525 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1526 return true; 1527 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1528 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD)) 1529 return true; 1530 } 1531 } 1532 1533 return false; 1534} 1535 1536/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1537/// two values ever stored into GV are its initializer and OtherVal. See if we 1538/// can shrink the global into a boolean and select between the two values 1539/// whenever it is used. This exposes the values to other scalar optimizations. 1540static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1541 const Type *GVElType = GV->getType()->getElementType(); 1542 1543 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1544 // an FP value, pointer or vector, don't do this optimization because a select 1545 // between them is very expensive and unlikely to lead to later 1546 // simplification. In these cases, we typically end up with "cond ? v1 : v2" 1547 // where v1 and v2 both require constant pool loads, a big loss. 1548 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1549 isa<PointerType>(GVElType) || isa<VectorType>(GVElType)) 1550 return false; 1551 1552 // Walk the use list of the global seeing if all the uses are load or store. 1553 // If there is anything else, bail out. 1554 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1555 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1556 return false; 1557 1558 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1559 1560 // Create the new global, initializing it to false. 1561 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1562 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1563 GV->getName()+".b", 1564 (Module *)NULL, 1565 GV->isThreadLocal()); 1566 GV->getParent()->getGlobalList().insert(GV, NewGV); 1567 1568 Constant *InitVal = GV->getInitializer(); 1569 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1570 1571 // If initialized to zero and storing one into the global, we can use a cast 1572 // instead of a select to synthesize the desired value. 1573 bool IsOneZero = false; 1574 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1575 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1576 1577 while (!GV->use_empty()) { 1578 Instruction *UI = cast<Instruction>(GV->use_back()); 1579 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1580 // Change the store into a boolean store. 1581 bool StoringOther = SI->getOperand(0) == OtherVal; 1582 // Only do this if we weren't storing a loaded value. 1583 Value *StoreVal; 1584 if (StoringOther || SI->getOperand(0) == InitVal) 1585 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1586 else { 1587 // Otherwise, we are storing a previously loaded copy. To do this, 1588 // change the copy from copying the original value to just copying the 1589 // bool. 1590 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1591 1592 // If we're already replaced the input, StoredVal will be a cast or 1593 // select instruction. If not, it will be a load of the original 1594 // global. 1595 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1596 assert(LI->getOperand(0) == GV && "Not a copy!"); 1597 // Insert a new load, to preserve the saved value. 1598 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1599 } else { 1600 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1601 "This is not a form that we understand!"); 1602 StoreVal = StoredVal->getOperand(0); 1603 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1604 } 1605 } 1606 new StoreInst(StoreVal, NewGV, SI); 1607 } else { 1608 // Change the load into a load of bool then a select. 1609 LoadInst *LI = cast<LoadInst>(UI); 1610 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1611 Value *NSI; 1612 if (IsOneZero) 1613 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1614 else 1615 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1616 NSI->takeName(LI); 1617 LI->replaceAllUsesWith(NSI); 1618 } 1619 UI->eraseFromParent(); 1620 } 1621 1622 GV->eraseFromParent(); 1623 return true; 1624} 1625 1626 1627/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1628/// it if possible. If we make a change, return true. 1629bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1630 Module::global_iterator &GVI) { 1631 SmallPtrSet<PHINode*, 16> PHIUsers; 1632 GlobalStatus GS; 1633 GV->removeDeadConstantUsers(); 1634 1635 if (GV->use_empty()) { 1636 DOUT << "GLOBAL DEAD: " << *GV; 1637 GV->eraseFromParent(); 1638 ++NumDeleted; 1639 return true; 1640 } 1641 1642 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1643#if 0 1644 cerr << "Global: " << *GV; 1645 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1646 cerr << " StoredType = "; 1647 switch (GS.StoredType) { 1648 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1649 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1650 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1651 case GlobalStatus::isStored: cerr << "stored\n"; break; 1652 } 1653 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1654 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1655 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1656 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1657 << "\n"; 1658 cerr << " HasMultipleAccessingFunctions = " 1659 << GS.HasMultipleAccessingFunctions << "\n"; 1660 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1661 cerr << "\n"; 1662#endif 1663 1664 // If this is a first class global and has only one accessing function 1665 // and this function is main (which we know is not recursive we can make 1666 // this global a local variable) we replace the global with a local alloca 1667 // in this function. 1668 // 1669 // NOTE: It doesn't make sense to promote non single-value types since we 1670 // are just replacing static memory to stack memory. 1671 if (!GS.HasMultipleAccessingFunctions && 1672 GS.AccessingFunction && !GS.HasNonInstructionUser && 1673 GV->getType()->getElementType()->isSingleValueType() && 1674 GS.AccessingFunction->getName() == "main" && 1675 GS.AccessingFunction->hasExternalLinkage()) { 1676 DOUT << "LOCALIZING GLOBAL: " << *GV; 1677 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1678 const Type* ElemTy = GV->getType()->getElementType(); 1679 // FIXME: Pass Global's alignment when globals have alignment 1680 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1681 if (!isa<UndefValue>(GV->getInitializer())) 1682 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1683 1684 GV->replaceAllUsesWith(Alloca); 1685 GV->eraseFromParent(); 1686 ++NumLocalized; 1687 return true; 1688 } 1689 1690 // If the global is never loaded (but may be stored to), it is dead. 1691 // Delete it now. 1692 if (!GS.isLoaded) { 1693 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1694 1695 // Delete any stores we can find to the global. We may not be able to 1696 // make it completely dead though. 1697 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1698 1699 // If the global is dead now, delete it. 1700 if (GV->use_empty()) { 1701 GV->eraseFromParent(); 1702 ++NumDeleted; 1703 Changed = true; 1704 } 1705 return Changed; 1706 1707 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1708 DOUT << "MARKING CONSTANT: " << *GV; 1709 GV->setConstant(true); 1710 1711 // Clean up any obviously simplifiable users now. 1712 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1713 1714 // If the global is dead now, just nuke it. 1715 if (GV->use_empty()) { 1716 DOUT << " *** Marking constant allowed us to simplify " 1717 << "all users and delete global!\n"; 1718 GV->eraseFromParent(); 1719 ++NumDeleted; 1720 } 1721 1722 ++NumMarked; 1723 return true; 1724 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1725 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1726 getAnalysis<TargetData>())) { 1727 GVI = FirstNewGV; // Don't skip the newly produced globals! 1728 return true; 1729 } 1730 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1731 // If the initial value for the global was an undef value, and if only 1732 // one other value was stored into it, we can just change the 1733 // initializer to be the stored value, then delete all stores to the 1734 // global. This allows us to mark it constant. 1735 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1736 if (isa<UndefValue>(GV->getInitializer())) { 1737 // Change the initial value here. 1738 GV->setInitializer(SOVConstant); 1739 1740 // Clean up any obviously simplifiable users now. 1741 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1742 1743 if (GV->use_empty()) { 1744 DOUT << " *** Substituting initializer allowed us to " 1745 << "simplify all users and delete global!\n"; 1746 GV->eraseFromParent(); 1747 ++NumDeleted; 1748 } else { 1749 GVI = GV; 1750 } 1751 ++NumSubstitute; 1752 return true; 1753 } 1754 1755 // Try to optimize globals based on the knowledge that only one value 1756 // (besides its initializer) is ever stored to the global. 1757 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1758 getAnalysis<TargetData>())) 1759 return true; 1760 1761 // Otherwise, if the global was not a boolean, we can shrink it to be a 1762 // boolean. 1763 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1764 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1765 ++NumShrunkToBool; 1766 return true; 1767 } 1768 } 1769 } 1770 return false; 1771} 1772 1773/// OnlyCalledDirectly - Return true if the specified function is only called 1774/// directly. In other words, its address is never taken. 1775static bool OnlyCalledDirectly(Function *F) { 1776 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1777 Instruction *User = dyn_cast<Instruction>(*UI); 1778 if (!User) return false; 1779 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1780 1781 // See if the function address is passed as an argument. 1782 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end(); 1783 i != e; ++i) 1784 if (*i == F) return false; 1785 } 1786 return true; 1787} 1788 1789/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1790/// function, changing them to FastCC. 1791static void ChangeCalleesToFastCall(Function *F) { 1792 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1793 CallSite User(cast<Instruction>(*UI)); 1794 User.setCallingConv(CallingConv::Fast); 1795 } 1796} 1797 1798static AttrListPtr StripNest(const AttrListPtr &Attrs) { 1799 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1800 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) 1801 continue; 1802 1803 // There can be only one. 1804 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); 1805 } 1806 1807 return Attrs; 1808} 1809 1810static void RemoveNestAttribute(Function *F) { 1811 F->setAttributes(StripNest(F->getAttributes())); 1812 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1813 CallSite User(cast<Instruction>(*UI)); 1814 User.setAttributes(StripNest(User.getAttributes())); 1815 } 1816} 1817 1818bool GlobalOpt::OptimizeFunctions(Module &M) { 1819 bool Changed = false; 1820 // Optimize functions. 1821 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1822 Function *F = FI++; 1823 // Functions without names cannot be referenced outside this module. 1824 if (!F->hasName() && !F->isDeclaration()) 1825 F->setLinkage(GlobalValue::InternalLinkage); 1826 F->removeDeadConstantUsers(); 1827 if (F->use_empty() && (F->hasLocalLinkage() || 1828 F->hasLinkOnceLinkage())) { 1829 M.getFunctionList().erase(F); 1830 Changed = true; 1831 ++NumFnDeleted; 1832 } else if (F->hasLocalLinkage()) { 1833 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1834 OnlyCalledDirectly(F)) { 1835 // If this function has C calling conventions, is not a varargs 1836 // function, and is only called directly, promote it to use the Fast 1837 // calling convention. 1838 F->setCallingConv(CallingConv::Fast); 1839 ChangeCalleesToFastCall(F); 1840 ++NumFastCallFns; 1841 Changed = true; 1842 } 1843 1844 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1845 OnlyCalledDirectly(F)) { 1846 // The function is not used by a trampoline intrinsic, so it is safe 1847 // to remove the 'nest' attribute. 1848 RemoveNestAttribute(F); 1849 ++NumNestRemoved; 1850 Changed = true; 1851 } 1852 } 1853 } 1854 return Changed; 1855} 1856 1857bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1858 bool Changed = false; 1859 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1860 GVI != E; ) { 1861 GlobalVariable *GV = GVI++; 1862 // Global variables without names cannot be referenced outside this module. 1863 if (!GV->hasName() && !GV->isDeclaration()) 1864 GV->setLinkage(GlobalValue::InternalLinkage); 1865 if (!GV->isConstant() && GV->hasLocalLinkage() && 1866 GV->hasInitializer()) 1867 Changed |= ProcessInternalGlobal(GV, GVI); 1868 } 1869 return Changed; 1870} 1871 1872/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1873/// initializers have an init priority of 65535. 1874GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1875 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1876 I != E; ++I) 1877 if (I->getName() == "llvm.global_ctors") { 1878 // Found it, verify it's an array of { int, void()* }. 1879 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1880 if (!ATy) return 0; 1881 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1882 if (!STy || STy->getNumElements() != 2 || 1883 STy->getElementType(0) != Type::Int32Ty) return 0; 1884 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1885 if (!PFTy) return 0; 1886 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1887 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1888 FTy->getNumParams() != 0) 1889 return 0; 1890 1891 // Verify that the initializer is simple enough for us to handle. 1892 if (!I->hasInitializer()) return 0; 1893 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1894 if (!CA) return 0; 1895 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1896 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1897 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1898 continue; 1899 1900 // Must have a function or null ptr. 1901 if (!isa<Function>(CS->getOperand(1))) 1902 return 0; 1903 1904 // Init priority must be standard. 1905 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1906 if (!CI || CI->getZExtValue() != 65535) 1907 return 0; 1908 } else { 1909 return 0; 1910 } 1911 1912 return I; 1913 } 1914 return 0; 1915} 1916 1917/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1918/// return a list of the functions and null terminator as a vector. 1919static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1920 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1921 std::vector<Function*> Result; 1922 Result.reserve(CA->getNumOperands()); 1923 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1924 ConstantStruct *CS = cast<ConstantStruct>(*i); 1925 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1926 } 1927 return Result; 1928} 1929 1930/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1931/// specified array, returning the new global to use. 1932static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1933 const std::vector<Function*> &Ctors) { 1934 // If we made a change, reassemble the initializer list. 1935 std::vector<Constant*> CSVals; 1936 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1937 CSVals.push_back(0); 1938 1939 // Create the new init list. 1940 std::vector<Constant*> CAList; 1941 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1942 if (Ctors[i]) { 1943 CSVals[1] = Ctors[i]; 1944 } else { 1945 const Type *FTy = FunctionType::get(Type::VoidTy, 1946 std::vector<const Type*>(), false); 1947 const PointerType *PFTy = PointerType::getUnqual(FTy); 1948 CSVals[1] = Constant::getNullValue(PFTy); 1949 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1950 } 1951 CAList.push_back(ConstantStruct::get(CSVals)); 1952 } 1953 1954 // Create the array initializer. 1955 const Type *StructTy = 1956 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1957 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1958 CAList); 1959 1960 // If we didn't change the number of elements, don't create a new GV. 1961 if (CA->getType() == GCL->getInitializer()->getType()) { 1962 GCL->setInitializer(CA); 1963 return GCL; 1964 } 1965 1966 // Create the new global and insert it next to the existing list. 1967 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1968 GCL->getLinkage(), CA, "", 1969 (Module *)NULL, 1970 GCL->isThreadLocal()); 1971 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1972 NGV->takeName(GCL); 1973 1974 // Nuke the old list, replacing any uses with the new one. 1975 if (!GCL->use_empty()) { 1976 Constant *V = NGV; 1977 if (V->getType() != GCL->getType()) 1978 V = ConstantExpr::getBitCast(V, GCL->getType()); 1979 GCL->replaceAllUsesWith(V); 1980 } 1981 GCL->eraseFromParent(); 1982 1983 if (Ctors.size()) 1984 return NGV; 1985 else 1986 return 0; 1987} 1988 1989 1990static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, 1991 Value *V) { 1992 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1993 Constant *R = ComputedValues[V]; 1994 assert(R && "Reference to an uncomputed value!"); 1995 return R; 1996} 1997 1998/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1999/// enough for us to understand. In particular, if it is a cast of something, 2000/// we punt. We basically just support direct accesses to globals and GEP's of 2001/// globals. This should be kept up to date with CommitValueTo. 2002static bool isSimpleEnoughPointerToCommit(Constant *C) { 2003 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 2004 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 2005 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 2006 return !GV->isDeclaration(); // reject external globals. 2007 } 2008 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2009 // Handle a constantexpr gep. 2010 if (CE->getOpcode() == Instruction::GetElementPtr && 2011 isa<GlobalVariable>(CE->getOperand(0))) { 2012 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2013 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 2014 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 2015 return GV->hasInitializer() && 2016 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2017 } 2018 return false; 2019} 2020 2021/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 2022/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 2023/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 2024static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 2025 ConstantExpr *Addr, unsigned OpNo) { 2026 // Base case of the recursion. 2027 if (OpNo == Addr->getNumOperands()) { 2028 assert(Val->getType() == Init->getType() && "Type mismatch!"); 2029 return Val; 2030 } 2031 2032 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 2033 std::vector<Constant*> Elts; 2034 2035 // Break up the constant into its elements. 2036 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 2037 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 2038 Elts.push_back(cast<Constant>(*i)); 2039 } else if (isa<ConstantAggregateZero>(Init)) { 2040 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2041 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 2042 } else if (isa<UndefValue>(Init)) { 2043 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2044 Elts.push_back(UndefValue::get(STy->getElementType(i))); 2045 } else { 2046 assert(0 && "This code is out of sync with " 2047 " ConstantFoldLoadThroughGEPConstantExpr"); 2048 } 2049 2050 // Replace the element that we are supposed to. 2051 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 2052 unsigned Idx = CU->getZExtValue(); 2053 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 2054 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 2055 2056 // Return the modified struct. 2057 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 2058 } else { 2059 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 2060 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 2061 2062 // Break up the array into elements. 2063 std::vector<Constant*> Elts; 2064 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 2065 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 2066 Elts.push_back(cast<Constant>(*i)); 2067 } else if (isa<ConstantAggregateZero>(Init)) { 2068 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 2069 Elts.assign(ATy->getNumElements(), Elt); 2070 } else if (isa<UndefValue>(Init)) { 2071 Constant *Elt = UndefValue::get(ATy->getElementType()); 2072 Elts.assign(ATy->getNumElements(), Elt); 2073 } else { 2074 assert(0 && "This code is out of sync with " 2075 " ConstantFoldLoadThroughGEPConstantExpr"); 2076 } 2077 2078 assert(CI->getZExtValue() < ATy->getNumElements()); 2079 Elts[CI->getZExtValue()] = 2080 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 2081 return ConstantArray::get(ATy, Elts); 2082 } 2083} 2084 2085/// CommitValueTo - We have decided that Addr (which satisfies the predicate 2086/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 2087static void CommitValueTo(Constant *Val, Constant *Addr) { 2088 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 2089 assert(GV->hasInitializer()); 2090 GV->setInitializer(Val); 2091 return; 2092 } 2093 2094 ConstantExpr *CE = cast<ConstantExpr>(Addr); 2095 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2096 2097 Constant *Init = GV->getInitializer(); 2098 Init = EvaluateStoreInto(Init, Val, CE, 2); 2099 GV->setInitializer(Init); 2100} 2101 2102/// ComputeLoadResult - Return the value that would be computed by a load from 2103/// P after the stores reflected by 'memory' have been performed. If we can't 2104/// decide, return null. 2105static Constant *ComputeLoadResult(Constant *P, 2106 const DenseMap<Constant*, Constant*> &Memory) { 2107 // If this memory location has been recently stored, use the stored value: it 2108 // is the most up-to-date. 2109 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P); 2110 if (I != Memory.end()) return I->second; 2111 2112 // Access it. 2113 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 2114 if (GV->hasInitializer()) 2115 return GV->getInitializer(); 2116 return 0; 2117 } 2118 2119 // Handle a constantexpr getelementptr. 2120 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 2121 if (CE->getOpcode() == Instruction::GetElementPtr && 2122 isa<GlobalVariable>(CE->getOperand(0))) { 2123 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2124 if (GV->hasInitializer()) 2125 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2126 } 2127 2128 return 0; // don't know how to evaluate. 2129} 2130 2131/// EvaluateFunction - Evaluate a call to function F, returning true if 2132/// successful, false if we can't evaluate it. ActualArgs contains the formal 2133/// arguments for the function. 2134static bool EvaluateFunction(Function *F, Constant *&RetVal, 2135 const std::vector<Constant*> &ActualArgs, 2136 std::vector<Function*> &CallStack, 2137 DenseMap<Constant*, Constant*> &MutatedMemory, 2138 std::vector<GlobalVariable*> &AllocaTmps) { 2139 // Check to see if this function is already executing (recursion). If so, 2140 // bail out. TODO: we might want to accept limited recursion. 2141 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 2142 return false; 2143 2144 CallStack.push_back(F); 2145 2146 /// Values - As we compute SSA register values, we store their contents here. 2147 DenseMap<Value*, Constant*> Values; 2148 2149 // Initialize arguments to the incoming values specified. 2150 unsigned ArgNo = 0; 2151 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 2152 ++AI, ++ArgNo) 2153 Values[AI] = ActualArgs[ArgNo]; 2154 2155 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 2156 /// we can only evaluate any one basic block at most once. This set keeps 2157 /// track of what we have executed so we can detect recursive cases etc. 2158 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 2159 2160 // CurInst - The current instruction we're evaluating. 2161 BasicBlock::iterator CurInst = F->begin()->begin(); 2162 2163 // This is the main evaluation loop. 2164 while (1) { 2165 Constant *InstResult = 0; 2166 2167 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2168 if (SI->isVolatile()) return false; // no volatile accesses. 2169 Constant *Ptr = getVal(Values, SI->getOperand(1)); 2170 if (!isSimpleEnoughPointerToCommit(Ptr)) 2171 // If this is too complex for us to commit, reject it. 2172 return false; 2173 Constant *Val = getVal(Values, SI->getOperand(0)); 2174 MutatedMemory[Ptr] = Val; 2175 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2176 InstResult = ConstantExpr::get(BO->getOpcode(), 2177 getVal(Values, BO->getOperand(0)), 2178 getVal(Values, BO->getOperand(1))); 2179 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2180 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2181 getVal(Values, CI->getOperand(0)), 2182 getVal(Values, CI->getOperand(1))); 2183 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2184 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2185 getVal(Values, CI->getOperand(0)), 2186 CI->getType()); 2187 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2188 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2189 getVal(Values, SI->getOperand(1)), 2190 getVal(Values, SI->getOperand(2))); 2191 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2192 Constant *P = getVal(Values, GEP->getOperand(0)); 2193 SmallVector<Constant*, 8> GEPOps; 2194 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2195 i != e; ++i) 2196 GEPOps.push_back(getVal(Values, *i)); 2197 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2198 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2199 if (LI->isVolatile()) return false; // no volatile accesses. 2200 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2201 MutatedMemory); 2202 if (InstResult == 0) return false; // Could not evaluate load. 2203 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2204 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2205 const Type *Ty = AI->getType()->getElementType(); 2206 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2207 GlobalValue::InternalLinkage, 2208 UndefValue::get(Ty), 2209 AI->getName())); 2210 InstResult = AllocaTmps.back(); 2211 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2212 2213 // Debug info can safely be ignored here. 2214 if (isa<DbgInfoIntrinsic>(CI)) { 2215 ++CurInst; 2216 continue; 2217 } 2218 2219 // Cannot handle inline asm. 2220 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2221 2222 // Resolve function pointers. 2223 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2224 if (!Callee) return false; // Cannot resolve. 2225 2226 std::vector<Constant*> Formals; 2227 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end(); 2228 i != e; ++i) 2229 Formals.push_back(getVal(Values, *i)); 2230 2231 if (Callee->isDeclaration()) { 2232 // If this is a function we can constant fold, do it. 2233 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2234 Formals.size())) { 2235 InstResult = C; 2236 } else { 2237 return false; 2238 } 2239 } else { 2240 if (Callee->getFunctionType()->isVarArg()) 2241 return false; 2242 2243 Constant *RetVal; 2244 // Execute the call, if successful, use the return value. 2245 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2246 MutatedMemory, AllocaTmps)) 2247 return false; 2248 InstResult = RetVal; 2249 } 2250 } else if (isa<TerminatorInst>(CurInst)) { 2251 BasicBlock *NewBB = 0; 2252 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2253 if (BI->isUnconditional()) { 2254 NewBB = BI->getSuccessor(0); 2255 } else { 2256 ConstantInt *Cond = 2257 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2258 if (!Cond) return false; // Cannot determine. 2259 2260 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2261 } 2262 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2263 ConstantInt *Val = 2264 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2265 if (!Val) return false; // Cannot determine. 2266 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2267 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2268 if (RI->getNumOperands()) 2269 RetVal = getVal(Values, RI->getOperand(0)); 2270 2271 CallStack.pop_back(); // return from fn. 2272 return true; // We succeeded at evaluating this ctor! 2273 } else { 2274 // invoke, unwind, unreachable. 2275 return false; // Cannot handle this terminator. 2276 } 2277 2278 // Okay, we succeeded in evaluating this control flow. See if we have 2279 // executed the new block before. If so, we have a looping function, 2280 // which we cannot evaluate in reasonable time. 2281 if (!ExecutedBlocks.insert(NewBB)) 2282 return false; // looped! 2283 2284 // Okay, we have never been in this block before. Check to see if there 2285 // are any PHI nodes. If so, evaluate them with information about where 2286 // we came from. 2287 BasicBlock *OldBB = CurInst->getParent(); 2288 CurInst = NewBB->begin(); 2289 PHINode *PN; 2290 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2291 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2292 2293 // Do NOT increment CurInst. We know that the terminator had no value. 2294 continue; 2295 } else { 2296 // Did not know how to evaluate this! 2297 return false; 2298 } 2299 2300 if (!CurInst->use_empty()) 2301 Values[CurInst] = InstResult; 2302 2303 // Advance program counter. 2304 ++CurInst; 2305 } 2306} 2307 2308/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2309/// we can. Return true if we can, false otherwise. 2310static bool EvaluateStaticConstructor(Function *F) { 2311 /// MutatedMemory - For each store we execute, we update this map. Loads 2312 /// check this to get the most up-to-date value. If evaluation is successful, 2313 /// this state is committed to the process. 2314 DenseMap<Constant*, Constant*> MutatedMemory; 2315 2316 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2317 /// to represent its body. This vector is needed so we can delete the 2318 /// temporary globals when we are done. 2319 std::vector<GlobalVariable*> AllocaTmps; 2320 2321 /// CallStack - This is used to detect recursion. In pathological situations 2322 /// we could hit exponential behavior, but at least there is nothing 2323 /// unbounded. 2324 std::vector<Function*> CallStack; 2325 2326 // Call the function. 2327 Constant *RetValDummy; 2328 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2329 CallStack, MutatedMemory, AllocaTmps); 2330 if (EvalSuccess) { 2331 // We succeeded at evaluation: commit the result. 2332 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2333 << F->getName() << "' to " << MutatedMemory.size() 2334 << " stores.\n"; 2335 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2336 E = MutatedMemory.end(); I != E; ++I) 2337 CommitValueTo(I->second, I->first); 2338 } 2339 2340 // At this point, we are done interpreting. If we created any 'alloca' 2341 // temporaries, release them now. 2342 while (!AllocaTmps.empty()) { 2343 GlobalVariable *Tmp = AllocaTmps.back(); 2344 AllocaTmps.pop_back(); 2345 2346 // If there are still users of the alloca, the program is doing something 2347 // silly, e.g. storing the address of the alloca somewhere and using it 2348 // later. Since this is undefined, we'll just make it be null. 2349 if (!Tmp->use_empty()) 2350 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2351 delete Tmp; 2352 } 2353 2354 return EvalSuccess; 2355} 2356 2357 2358 2359/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2360/// Return true if anything changed. 2361bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2362 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2363 bool MadeChange = false; 2364 if (Ctors.empty()) return false; 2365 2366 // Loop over global ctors, optimizing them when we can. 2367 for (unsigned i = 0; i != Ctors.size(); ++i) { 2368 Function *F = Ctors[i]; 2369 // Found a null terminator in the middle of the list, prune off the rest of 2370 // the list. 2371 if (F == 0) { 2372 if (i != Ctors.size()-1) { 2373 Ctors.resize(i+1); 2374 MadeChange = true; 2375 } 2376 break; 2377 } 2378 2379 // We cannot simplify external ctor functions. 2380 if (F->empty()) continue; 2381 2382 // If we can evaluate the ctor at compile time, do. 2383 if (EvaluateStaticConstructor(F)) { 2384 Ctors.erase(Ctors.begin()+i); 2385 MadeChange = true; 2386 --i; 2387 ++NumCtorsEvaluated; 2388 continue; 2389 } 2390 } 2391 2392 if (!MadeChange) return false; 2393 2394 GCL = InstallGlobalCtors(GCL, Ctors); 2395 return true; 2396} 2397 2398bool GlobalOpt::OptimizeGlobalAliases(Module &M) { 2399 bool Changed = false; 2400 2401 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 2402 I != E;) { 2403 Module::alias_iterator J = I++; 2404 // Aliases without names cannot be referenced outside this module. 2405 if (!J->hasName() && !J->isDeclaration()) 2406 J->setLinkage(GlobalValue::InternalLinkage); 2407 // If the aliasee may change at link time, nothing can be done - bail out. 2408 if (J->mayBeOverridden()) 2409 continue; 2410 2411 Constant *Aliasee = J->getAliasee(); 2412 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts()); 2413 Target->removeDeadConstantUsers(); 2414 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse(); 2415 2416 // Make all users of the alias use the aliasee instead. 2417 if (!J->use_empty()) { 2418 J->replaceAllUsesWith(Aliasee); 2419 ++NumAliasesResolved; 2420 Changed = true; 2421 } 2422 2423 // If the aliasee has internal linkage, give it the name and linkage 2424 // of the alias, and delete the alias. This turns: 2425 // define internal ... @f(...) 2426 // @a = alias ... @f 2427 // into: 2428 // define ... @a(...) 2429 if (!Target->hasLocalLinkage()) 2430 continue; 2431 2432 // The transform is only useful if the alias does not have internal linkage. 2433 if (J->hasLocalLinkage()) 2434 continue; 2435 2436 // Do not perform the transform if multiple aliases potentially target the 2437 // aliasee. This check also ensures that it is safe to replace the section 2438 // and other attributes of the aliasee with those of the alias. 2439 if (!hasOneUse) 2440 continue; 2441 2442 // Give the aliasee the name, linkage and other attributes of the alias. 2443 Target->takeName(J); 2444 Target->setLinkage(J->getLinkage()); 2445 Target->GlobalValue::copyAttributesFrom(J); 2446 2447 // Delete the alias. 2448 M.getAliasList().erase(J); 2449 ++NumAliasesRemoved; 2450 Changed = true; 2451 } 2452 2453 return Changed; 2454} 2455 2456bool GlobalOpt::runOnModule(Module &M) { 2457 bool Changed = false; 2458 2459 // Try to find the llvm.globalctors list. 2460 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2461 2462 bool LocalChange = true; 2463 while (LocalChange) { 2464 LocalChange = false; 2465 2466 // Delete functions that are trivially dead, ccc -> fastcc 2467 LocalChange |= OptimizeFunctions(M); 2468 2469 // Optimize global_ctors list. 2470 if (GlobalCtors) 2471 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2472 2473 // Optimize non-address-taken globals. 2474 LocalChange |= OptimizeGlobalVars(M); 2475 2476 // Resolve aliases, when possible. 2477 LocalChange |= OptimizeGlobalAliases(M); 2478 Changed |= LocalChange; 2479 } 2480 2481 // TODO: Move all global ctors functions to the end of the module for code 2482 // layout. 2483 2484 return Changed; 2485}
| 936 937 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { 938 continue; // Fine, ignore. 939 } 940 941 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 942 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 943 return false; // Storing the pointer itself... bad. 944 continue; // Otherwise, storing through it, or storing into GV... fine. 945 } 946 947 if (isa<GetElementPtrInst>(Inst)) { 948 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs)) 949 return false; 950 continue; 951 } 952 953 if (PHINode *PN = dyn_cast<PHINode>(Inst)) { 954 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 955 // cycles. 956 if (PHIs.insert(PN)) 957 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 958 return false; 959 continue; 960 } 961 962 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { 963 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) 964 return false; 965 continue; 966 } 967 968 return false; 969 } 970 return true; 971} 972 973/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 974/// somewhere. Transform all uses of the allocation into loads from the 975/// global and uses of the resultant pointer. Further, delete the store into 976/// GV. This assumes that these value pass the 977/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 978static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 979 GlobalVariable *GV) { 980 while (!Alloc->use_empty()) { 981 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 982 Instruction *InsertPt = U; 983 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 984 // If this is the store of the allocation into the global, remove it. 985 if (SI->getOperand(1) == GV) { 986 SI->eraseFromParent(); 987 continue; 988 } 989 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 990 // Insert the load in the corresponding predecessor, not right before the 991 // PHI. 992 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator(); 993 } else if (isa<BitCastInst>(U)) { 994 // Must be bitcast between the malloc and store to initialize the global. 995 ReplaceUsesOfMallocWithGlobal(U, GV); 996 U->eraseFromParent(); 997 continue; 998 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 999 // If this is a "GEP bitcast" and the user is a store to the global, then 1000 // just process it as a bitcast. 1001 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse()) 1002 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back())) 1003 if (SI->getOperand(1) == GV) { 1004 // Must be bitcast GEP between the malloc and store to initialize 1005 // the global. 1006 ReplaceUsesOfMallocWithGlobal(GEPI, GV); 1007 GEPI->eraseFromParent(); 1008 continue; 1009 } 1010 } 1011 1012 // Insert a load from the global, and use it instead of the malloc. 1013 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 1014 U->replaceUsesOfWith(Alloc, NL); 1015 } 1016} 1017 1018/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi 1019/// of a load) are simple enough to perform heap SRA on. This permits GEP's 1020/// that index through the array and struct field, icmps of null, and PHIs. 1021static bool LoadUsesSimpleEnoughForHeapSRA(Value *V, 1022 SmallPtrSet<PHINode*, 32> &LoadUsingPHIs, 1023 SmallPtrSet<PHINode*, 32> &LoadUsingPHIsPerLoad) { 1024 // We permit two users of the load: setcc comparing against the null 1025 // pointer, and a getelementptr of a specific form. 1026 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 1027 Instruction *User = cast<Instruction>(*UI); 1028 1029 // Comparison against null is ok. 1030 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) { 1031 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 1032 return false; 1033 continue; 1034 } 1035 1036 // getelementptr is also ok, but only a simple form. 1037 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1038 // Must index into the array and into the struct. 1039 if (GEPI->getNumOperands() < 3) 1040 return false; 1041 1042 // Otherwise the GEP is ok. 1043 continue; 1044 } 1045 1046 if (PHINode *PN = dyn_cast<PHINode>(User)) { 1047 if (!LoadUsingPHIsPerLoad.insert(PN)) 1048 // This means some phi nodes are dependent on each other. 1049 // Avoid infinite looping! 1050 return false; 1051 if (!LoadUsingPHIs.insert(PN)) 1052 // If we have already analyzed this PHI, then it is safe. 1053 continue; 1054 1055 // Make sure all uses of the PHI are simple enough to transform. 1056 if (!LoadUsesSimpleEnoughForHeapSRA(PN, 1057 LoadUsingPHIs, LoadUsingPHIsPerLoad)) 1058 return false; 1059 1060 continue; 1061 } 1062 1063 // Otherwise we don't know what this is, not ok. 1064 return false; 1065 } 1066 1067 return true; 1068} 1069 1070 1071/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1072/// GV are simple enough to perform HeapSRA, return true. 1073static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 1074 MallocInst *MI) { 1075 SmallPtrSet<PHINode*, 32> LoadUsingPHIs; 1076 SmallPtrSet<PHINode*, 32> LoadUsingPHIsPerLoad; 1077 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 1078 ++UI) 1079 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 1080 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs, 1081 LoadUsingPHIsPerLoad)) 1082 return false; 1083 LoadUsingPHIsPerLoad.clear(); 1084 } 1085 1086 // If we reach here, we know that all uses of the loads and transitive uses 1087 // (through PHI nodes) are simple enough to transform. However, we don't know 1088 // that all inputs the to the PHI nodes are in the same equivalence sets. 1089 // Check to verify that all operands of the PHIs are either PHIS that can be 1090 // transformed, loads from GV, or MI itself. 1091 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(), 1092 E = LoadUsingPHIs.end(); I != E; ++I) { 1093 PHINode *PN = *I; 1094 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) { 1095 Value *InVal = PN->getIncomingValue(op); 1096 1097 // PHI of the stored value itself is ok. 1098 if (InVal == MI) continue; 1099 1100 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) { 1101 // One of the PHIs in our set is (optimistically) ok. 1102 if (LoadUsingPHIs.count(InPN)) 1103 continue; 1104 return false; 1105 } 1106 1107 // Load from GV is ok. 1108 if (LoadInst *LI = dyn_cast<LoadInst>(InVal)) 1109 if (LI->getOperand(0) == GV) 1110 continue; 1111 1112 // UNDEF? NULL? 1113 1114 // Anything else is rejected. 1115 return false; 1116 } 1117 } 1118 1119 return true; 1120} 1121 1122static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, 1123 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1124 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1125 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V]; 1126 1127 if (FieldNo >= FieldVals.size()) 1128 FieldVals.resize(FieldNo+1); 1129 1130 // If we already have this value, just reuse the previously scalarized 1131 // version. 1132 if (Value *FieldVal = FieldVals[FieldNo]) 1133 return FieldVal; 1134 1135 // Depending on what instruction this is, we have several cases. 1136 Value *Result; 1137 if (LoadInst *LI = dyn_cast<LoadInst>(V)) { 1138 // This is a scalarized version of the load from the global. Just create 1139 // a new Load of the scalarized global. 1140 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo, 1141 InsertedScalarizedValues, 1142 PHIsToRewrite), 1143 LI->getName()+".f" + utostr(FieldNo), LI); 1144 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 1145 // PN's type is pointer to struct. Make a new PHI of pointer to struct 1146 // field. 1147 const StructType *ST = 1148 cast<StructType>(cast<PointerType>(PN->getType())->getElementType()); 1149 1150 Result =PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)), 1151 PN->getName()+".f"+utostr(FieldNo), PN); 1152 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); 1153 } else { 1154 assert(0 && "Unknown usable value"); 1155 Result = 0; 1156 } 1157 1158 return FieldVals[FieldNo] = Result; 1159} 1160 1161/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1162/// the load, rewrite the derived value to use the HeapSRoA'd load. 1163static void RewriteHeapSROALoadUser(Instruction *LoadUser, 1164 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1165 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1166 // If this is a comparison against null, handle it. 1167 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1168 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1169 // If we have a setcc of the loaded pointer, we can use a setcc of any 1170 // field. 1171 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0, 1172 InsertedScalarizedValues, PHIsToRewrite); 1173 1174 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1175 Constant::getNullValue(NPtr->getType()), 1176 SCI->getName(), SCI); 1177 SCI->replaceAllUsesWith(New); 1178 SCI->eraseFromParent(); 1179 return; 1180 } 1181 1182 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...' 1183 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1184 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1185 && "Unexpected GEPI!"); 1186 1187 // Load the pointer for this field. 1188 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1189 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo, 1190 InsertedScalarizedValues, PHIsToRewrite); 1191 1192 // Create the new GEP idx vector. 1193 SmallVector<Value*, 8> GEPIdx; 1194 GEPIdx.push_back(GEPI->getOperand(1)); 1195 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1196 1197 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1198 GEPIdx.begin(), GEPIdx.end(), 1199 GEPI->getName(), GEPI); 1200 GEPI->replaceAllUsesWith(NGEPI); 1201 GEPI->eraseFromParent(); 1202 return; 1203 } 1204 1205 // Recursively transform the users of PHI nodes. This will lazily create the 1206 // PHIs that are needed for individual elements. Keep track of what PHIs we 1207 // see in InsertedScalarizedValues so that we don't get infinite loops (very 1208 // antisocial). If the PHI is already in InsertedScalarizedValues, it has 1209 // already been seen first by another load, so its uses have already been 1210 // processed. 1211 PHINode *PN = cast<PHINode>(LoadUser); 1212 bool Inserted; 1213 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos; 1214 tie(InsertPos, Inserted) = 1215 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>())); 1216 if (!Inserted) return; 1217 1218 // If this is the first time we've seen this PHI, recursively process all 1219 // users. 1220 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) { 1221 Instruction *User = cast<Instruction>(*UI++); 1222 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1223 } 1224} 1225 1226/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1227/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1228/// use FieldGlobals instead. All uses of loaded values satisfy 1229/// AllGlobalLoadUsesSimpleEnoughForHeapSRA. 1230static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1231 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1232 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1233 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end(); 1234 UI != E; ) { 1235 Instruction *User = cast<Instruction>(*UI++); 1236 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1237 } 1238 1239 if (Load->use_empty()) { 1240 Load->eraseFromParent(); 1241 InsertedScalarizedValues.erase(Load); 1242 } 1243} 1244 1245/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1246/// it up into multiple allocations of arrays of the fields. 1247static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1248 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1249 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1250 1251 // There is guaranteed to be at least one use of the malloc (storing 1252 // it into GV). If there are other uses, change them to be uses of 1253 // the global to simplify later code. This also deletes the store 1254 // into GV. 1255 ReplaceUsesOfMallocWithGlobal(MI, GV); 1256 1257 // Okay, at this point, there are no users of the malloc. Insert N 1258 // new mallocs at the same place as MI, and N globals. 1259 std::vector<Value*> FieldGlobals; 1260 std::vector<MallocInst*> FieldMallocs; 1261 1262 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1263 const Type *FieldTy = STy->getElementType(FieldNo); 1264 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1265 1266 GlobalVariable *NGV = 1267 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1268 Constant::getNullValue(PFieldTy), 1269 GV->getName() + ".f" + utostr(FieldNo), GV, 1270 GV->isThreadLocal()); 1271 FieldGlobals.push_back(NGV); 1272 1273 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1274 MI->getName() + ".f" + utostr(FieldNo),MI); 1275 FieldMallocs.push_back(NMI); 1276 new StoreInst(NMI, NGV, MI); 1277 } 1278 1279 // The tricky aspect of this transformation is handling the case when malloc 1280 // fails. In the original code, malloc failing would set the result pointer 1281 // of malloc to null. In this case, some mallocs could succeed and others 1282 // could fail. As such, we emit code that looks like this: 1283 // F0 = malloc(field0) 1284 // F1 = malloc(field1) 1285 // F2 = malloc(field2) 1286 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1287 // if (F0) { free(F0); F0 = 0; } 1288 // if (F1) { free(F1); F1 = 0; } 1289 // if (F2) { free(F2); F2 = 0; } 1290 // } 1291 Value *RunningOr = 0; 1292 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1293 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1294 Constant::getNullValue(FieldMallocs[i]->getType()), 1295 "isnull", MI); 1296 if (!RunningOr) 1297 RunningOr = Cond; // First seteq 1298 else 1299 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI); 1300 } 1301 1302 // Split the basic block at the old malloc. 1303 BasicBlock *OrigBB = MI->getParent(); 1304 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1305 1306 // Create the block to check the first condition. Put all these blocks at the 1307 // end of the function as they are unlikely to be executed. 1308 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1309 OrigBB->getParent()); 1310 1311 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1312 // branch on RunningOr. 1313 OrigBB->getTerminator()->eraseFromParent(); 1314 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1315 1316 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1317 // pointer, because some may be null while others are not. 1318 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1319 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1320 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1321 Constant::getNullValue(GVVal->getType()), 1322 "tmp", NullPtrBlock); 1323 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1324 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1325 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1326 1327 // Fill in FreeBlock. 1328 new FreeInst(GVVal, FreeBlock); 1329 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1330 FreeBlock); 1331 BranchInst::Create(NextBlock, FreeBlock); 1332 1333 NullPtrBlock = NextBlock; 1334 } 1335 1336 BranchInst::Create(ContBB, NullPtrBlock); 1337 1338 // MI is no longer needed, remove it. 1339 MI->eraseFromParent(); 1340 1341 /// InsertedScalarizedLoads - As we process loads, if we can't immediately 1342 /// update all uses of the load, keep track of what scalarized loads are 1343 /// inserted for a given load. 1344 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues; 1345 InsertedScalarizedValues[GV] = FieldGlobals; 1346 1347 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite; 1348 1349 // Okay, the malloc site is completely handled. All of the uses of GV are now 1350 // loads, and all uses of those loads are simple. Rewrite them to use loads 1351 // of the per-field globals instead. 1352 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) { 1353 Instruction *User = cast<Instruction>(*UI++); 1354 1355 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1356 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite); 1357 continue; 1358 } 1359 1360 // Must be a store of null. 1361 StoreInst *SI = cast<StoreInst>(User); 1362 assert(isa<ConstantPointerNull>(SI->getOperand(0)) && 1363 "Unexpected heap-sra user!"); 1364 1365 // Insert a store of null into each global. 1366 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1367 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType()); 1368 Constant *Null = Constant::getNullValue(PT->getElementType()); 1369 new StoreInst(Null, FieldGlobals[i], SI); 1370 } 1371 // Erase the original store. 1372 SI->eraseFromParent(); 1373 } 1374 1375 // While we have PHIs that are interesting to rewrite, do it. 1376 while (!PHIsToRewrite.empty()) { 1377 PHINode *PN = PHIsToRewrite.back().first; 1378 unsigned FieldNo = PHIsToRewrite.back().second; 1379 PHIsToRewrite.pop_back(); 1380 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]); 1381 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi"); 1382 1383 // Add all the incoming values. This can materialize more phis. 1384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1385 Value *InVal = PN->getIncomingValue(i); 1386 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues, 1387 PHIsToRewrite); 1388 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i)); 1389 } 1390 } 1391 1392 // Drop all inter-phi links and any loads that made it this far. 1393 for (DenseMap<Value*, std::vector<Value*> >::iterator 1394 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1395 I != E; ++I) { 1396 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1397 PN->dropAllReferences(); 1398 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1399 LI->dropAllReferences(); 1400 } 1401 1402 // Delete all the phis and loads now that inter-references are dead. 1403 for (DenseMap<Value*, std::vector<Value*> >::iterator 1404 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1405 I != E; ++I) { 1406 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1407 PN->eraseFromParent(); 1408 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1409 LI->eraseFromParent(); 1410 } 1411 1412 // The old global is now dead, remove it. 1413 GV->eraseFromParent(); 1414 1415 ++NumHeapSRA; 1416 return cast<GlobalVariable>(FieldGlobals[0]); 1417} 1418 1419/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1420/// pointer global variable with a single value stored it that is a malloc or 1421/// cast of malloc. 1422static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, 1423 MallocInst *MI, 1424 Module::global_iterator &GVI, 1425 TargetData &TD) { 1426 // If this is a malloc of an abstract type, don't touch it. 1427 if (!MI->getAllocatedType()->isSized()) 1428 return false; 1429 1430 // We can't optimize this global unless all uses of it are *known* to be 1431 // of the malloc value, not of the null initializer value (consider a use 1432 // that compares the global's value against zero to see if the malloc has 1433 // been reached). To do this, we check to see if all uses of the global 1434 // would trap if the global were null: this proves that they must all 1435 // happen after the malloc. 1436 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1437 return false; 1438 1439 // We can't optimize this if the malloc itself is used in a complex way, 1440 // for example, being stored into multiple globals. This allows the 1441 // malloc to be stored into the specified global, loaded setcc'd, and 1442 // GEP'd. These are all things we could transform to using the global 1443 // for. 1444 { 1445 SmallPtrSet<PHINode*, 8> PHIs; 1446 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1447 return false; 1448 } 1449 1450 1451 // If we have a global that is only initialized with a fixed size malloc, 1452 // transform the program to use global memory instead of malloc'd memory. 1453 // This eliminates dynamic allocation, avoids an indirection accessing the 1454 // data, and exposes the resultant global to further GlobalOpt. 1455 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1456 // Restrict this transformation to only working on small allocations 1457 // (2048 bytes currently), as we don't want to introduce a 16M global or 1458 // something. 1459 if (NElements->getZExtValue()* 1460 TD.getTypeAllocSize(MI->getAllocatedType()) < 2048) { 1461 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1462 return true; 1463 } 1464 } 1465 1466 // If the allocation is an array of structures, consider transforming this 1467 // into multiple malloc'd arrays, one for each field. This is basically 1468 // SRoA for malloc'd memory. 1469 const Type *AllocTy = MI->getAllocatedType(); 1470 1471 // If this is an allocation of a fixed size array of structs, analyze as a 1472 // variable size array. malloc [100 x struct],1 -> malloc struct, 100 1473 if (!MI->isArrayAllocation()) 1474 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy)) 1475 AllocTy = AT->getElementType(); 1476 1477 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) { 1478 // This the structure has an unreasonable number of fields, leave it 1479 // alone. 1480 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 && 1481 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1482 1483 // If this is a fixed size array, transform the Malloc to be an alloc of 1484 // structs. malloc [100 x struct],1 -> malloc struct, 100 1485 if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) { 1486 MallocInst *NewMI = 1487 new MallocInst(AllocSTy, 1488 ConstantInt::get(Type::Int32Ty, AT->getNumElements()), 1489 "", MI); 1490 NewMI->takeName(MI); 1491 Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI); 1492 MI->replaceAllUsesWith(Cast); 1493 MI->eraseFromParent(); 1494 MI = NewMI; 1495 } 1496 1497 GVI = PerformHeapAllocSRoA(GV, MI); 1498 return true; 1499 } 1500 } 1501 1502 return false; 1503} 1504 1505// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1506// that only one value (besides its initializer) is ever stored to the global. 1507static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1508 Module::global_iterator &GVI, 1509 TargetData &TD) { 1510 // Ignore no-op GEPs and bitcasts. 1511 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1512 1513 // If we are dealing with a pointer global that is initialized to null and 1514 // only has one (non-null) value stored into it, then we can optimize any 1515 // users of the loaded value (often calls and loads) that would trap if the 1516 // value was null. 1517 if (isa<PointerType>(GV->getInitializer()->getType()) && 1518 GV->getInitializer()->isNullValue()) { 1519 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1520 if (GV->getInitializer()->getType() != SOVC->getType()) 1521 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1522 1523 // Optimize away any trapping uses of the loaded value. 1524 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1525 return true; 1526 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1527 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD)) 1528 return true; 1529 } 1530 } 1531 1532 return false; 1533} 1534 1535/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1536/// two values ever stored into GV are its initializer and OtherVal. See if we 1537/// can shrink the global into a boolean and select between the two values 1538/// whenever it is used. This exposes the values to other scalar optimizations. 1539static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1540 const Type *GVElType = GV->getType()->getElementType(); 1541 1542 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1543 // an FP value, pointer or vector, don't do this optimization because a select 1544 // between them is very expensive and unlikely to lead to later 1545 // simplification. In these cases, we typically end up with "cond ? v1 : v2" 1546 // where v1 and v2 both require constant pool loads, a big loss. 1547 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1548 isa<PointerType>(GVElType) || isa<VectorType>(GVElType)) 1549 return false; 1550 1551 // Walk the use list of the global seeing if all the uses are load or store. 1552 // If there is anything else, bail out. 1553 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1554 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1555 return false; 1556 1557 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1558 1559 // Create the new global, initializing it to false. 1560 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1561 GlobalValue::InternalLinkage, ConstantInt::getFalse(), 1562 GV->getName()+".b", 1563 (Module *)NULL, 1564 GV->isThreadLocal()); 1565 GV->getParent()->getGlobalList().insert(GV, NewGV); 1566 1567 Constant *InitVal = GV->getInitializer(); 1568 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1569 1570 // If initialized to zero and storing one into the global, we can use a cast 1571 // instead of a select to synthesize the desired value. 1572 bool IsOneZero = false; 1573 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1574 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1575 1576 while (!GV->use_empty()) { 1577 Instruction *UI = cast<Instruction>(GV->use_back()); 1578 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1579 // Change the store into a boolean store. 1580 bool StoringOther = SI->getOperand(0) == OtherVal; 1581 // Only do this if we weren't storing a loaded value. 1582 Value *StoreVal; 1583 if (StoringOther || SI->getOperand(0) == InitVal) 1584 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1585 else { 1586 // Otherwise, we are storing a previously loaded copy. To do this, 1587 // change the copy from copying the original value to just copying the 1588 // bool. 1589 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1590 1591 // If we're already replaced the input, StoredVal will be a cast or 1592 // select instruction. If not, it will be a load of the original 1593 // global. 1594 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1595 assert(LI->getOperand(0) == GV && "Not a copy!"); 1596 // Insert a new load, to preserve the saved value. 1597 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1598 } else { 1599 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1600 "This is not a form that we understand!"); 1601 StoreVal = StoredVal->getOperand(0); 1602 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1603 } 1604 } 1605 new StoreInst(StoreVal, NewGV, SI); 1606 } else { 1607 // Change the load into a load of bool then a select. 1608 LoadInst *LI = cast<LoadInst>(UI); 1609 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1610 Value *NSI; 1611 if (IsOneZero) 1612 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1613 else 1614 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1615 NSI->takeName(LI); 1616 LI->replaceAllUsesWith(NSI); 1617 } 1618 UI->eraseFromParent(); 1619 } 1620 1621 GV->eraseFromParent(); 1622 return true; 1623} 1624 1625 1626/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1627/// it if possible. If we make a change, return true. 1628bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1629 Module::global_iterator &GVI) { 1630 SmallPtrSet<PHINode*, 16> PHIUsers; 1631 GlobalStatus GS; 1632 GV->removeDeadConstantUsers(); 1633 1634 if (GV->use_empty()) { 1635 DOUT << "GLOBAL DEAD: " << *GV; 1636 GV->eraseFromParent(); 1637 ++NumDeleted; 1638 return true; 1639 } 1640 1641 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1642#if 0 1643 cerr << "Global: " << *GV; 1644 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1645 cerr << " StoredType = "; 1646 switch (GS.StoredType) { 1647 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1648 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1649 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1650 case GlobalStatus::isStored: cerr << "stored\n"; break; 1651 } 1652 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1653 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1654 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1655 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1656 << "\n"; 1657 cerr << " HasMultipleAccessingFunctions = " 1658 << GS.HasMultipleAccessingFunctions << "\n"; 1659 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1660 cerr << "\n"; 1661#endif 1662 1663 // If this is a first class global and has only one accessing function 1664 // and this function is main (which we know is not recursive we can make 1665 // this global a local variable) we replace the global with a local alloca 1666 // in this function. 1667 // 1668 // NOTE: It doesn't make sense to promote non single-value types since we 1669 // are just replacing static memory to stack memory. 1670 if (!GS.HasMultipleAccessingFunctions && 1671 GS.AccessingFunction && !GS.HasNonInstructionUser && 1672 GV->getType()->getElementType()->isSingleValueType() && 1673 GS.AccessingFunction->getName() == "main" && 1674 GS.AccessingFunction->hasExternalLinkage()) { 1675 DOUT << "LOCALIZING GLOBAL: " << *GV; 1676 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1677 const Type* ElemTy = GV->getType()->getElementType(); 1678 // FIXME: Pass Global's alignment when globals have alignment 1679 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1680 if (!isa<UndefValue>(GV->getInitializer())) 1681 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1682 1683 GV->replaceAllUsesWith(Alloca); 1684 GV->eraseFromParent(); 1685 ++NumLocalized; 1686 return true; 1687 } 1688 1689 // If the global is never loaded (but may be stored to), it is dead. 1690 // Delete it now. 1691 if (!GS.isLoaded) { 1692 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1693 1694 // Delete any stores we can find to the global. We may not be able to 1695 // make it completely dead though. 1696 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1697 1698 // If the global is dead now, delete it. 1699 if (GV->use_empty()) { 1700 GV->eraseFromParent(); 1701 ++NumDeleted; 1702 Changed = true; 1703 } 1704 return Changed; 1705 1706 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1707 DOUT << "MARKING CONSTANT: " << *GV; 1708 GV->setConstant(true); 1709 1710 // Clean up any obviously simplifiable users now. 1711 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1712 1713 // If the global is dead now, just nuke it. 1714 if (GV->use_empty()) { 1715 DOUT << " *** Marking constant allowed us to simplify " 1716 << "all users and delete global!\n"; 1717 GV->eraseFromParent(); 1718 ++NumDeleted; 1719 } 1720 1721 ++NumMarked; 1722 return true; 1723 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1724 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1725 getAnalysis<TargetData>())) { 1726 GVI = FirstNewGV; // Don't skip the newly produced globals! 1727 return true; 1728 } 1729 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1730 // If the initial value for the global was an undef value, and if only 1731 // one other value was stored into it, we can just change the 1732 // initializer to be the stored value, then delete all stores to the 1733 // global. This allows us to mark it constant. 1734 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1735 if (isa<UndefValue>(GV->getInitializer())) { 1736 // Change the initial value here. 1737 GV->setInitializer(SOVConstant); 1738 1739 // Clean up any obviously simplifiable users now. 1740 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1741 1742 if (GV->use_empty()) { 1743 DOUT << " *** Substituting initializer allowed us to " 1744 << "simplify all users and delete global!\n"; 1745 GV->eraseFromParent(); 1746 ++NumDeleted; 1747 } else { 1748 GVI = GV; 1749 } 1750 ++NumSubstitute; 1751 return true; 1752 } 1753 1754 // Try to optimize globals based on the knowledge that only one value 1755 // (besides its initializer) is ever stored to the global. 1756 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1757 getAnalysis<TargetData>())) 1758 return true; 1759 1760 // Otherwise, if the global was not a boolean, we can shrink it to be a 1761 // boolean. 1762 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1763 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1764 ++NumShrunkToBool; 1765 return true; 1766 } 1767 } 1768 } 1769 return false; 1770} 1771 1772/// OnlyCalledDirectly - Return true if the specified function is only called 1773/// directly. In other words, its address is never taken. 1774static bool OnlyCalledDirectly(Function *F) { 1775 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1776 Instruction *User = dyn_cast<Instruction>(*UI); 1777 if (!User) return false; 1778 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1779 1780 // See if the function address is passed as an argument. 1781 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end(); 1782 i != e; ++i) 1783 if (*i == F) return false; 1784 } 1785 return true; 1786} 1787 1788/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1789/// function, changing them to FastCC. 1790static void ChangeCalleesToFastCall(Function *F) { 1791 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1792 CallSite User(cast<Instruction>(*UI)); 1793 User.setCallingConv(CallingConv::Fast); 1794 } 1795} 1796 1797static AttrListPtr StripNest(const AttrListPtr &Attrs) { 1798 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1799 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) 1800 continue; 1801 1802 // There can be only one. 1803 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); 1804 } 1805 1806 return Attrs; 1807} 1808 1809static void RemoveNestAttribute(Function *F) { 1810 F->setAttributes(StripNest(F->getAttributes())); 1811 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1812 CallSite User(cast<Instruction>(*UI)); 1813 User.setAttributes(StripNest(User.getAttributes())); 1814 } 1815} 1816 1817bool GlobalOpt::OptimizeFunctions(Module &M) { 1818 bool Changed = false; 1819 // Optimize functions. 1820 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1821 Function *F = FI++; 1822 // Functions without names cannot be referenced outside this module. 1823 if (!F->hasName() && !F->isDeclaration()) 1824 F->setLinkage(GlobalValue::InternalLinkage); 1825 F->removeDeadConstantUsers(); 1826 if (F->use_empty() && (F->hasLocalLinkage() || 1827 F->hasLinkOnceLinkage())) { 1828 M.getFunctionList().erase(F); 1829 Changed = true; 1830 ++NumFnDeleted; 1831 } else if (F->hasLocalLinkage()) { 1832 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1833 OnlyCalledDirectly(F)) { 1834 // If this function has C calling conventions, is not a varargs 1835 // function, and is only called directly, promote it to use the Fast 1836 // calling convention. 1837 F->setCallingConv(CallingConv::Fast); 1838 ChangeCalleesToFastCall(F); 1839 ++NumFastCallFns; 1840 Changed = true; 1841 } 1842 1843 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1844 OnlyCalledDirectly(F)) { 1845 // The function is not used by a trampoline intrinsic, so it is safe 1846 // to remove the 'nest' attribute. 1847 RemoveNestAttribute(F); 1848 ++NumNestRemoved; 1849 Changed = true; 1850 } 1851 } 1852 } 1853 return Changed; 1854} 1855 1856bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1857 bool Changed = false; 1858 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1859 GVI != E; ) { 1860 GlobalVariable *GV = GVI++; 1861 // Global variables without names cannot be referenced outside this module. 1862 if (!GV->hasName() && !GV->isDeclaration()) 1863 GV->setLinkage(GlobalValue::InternalLinkage); 1864 if (!GV->isConstant() && GV->hasLocalLinkage() && 1865 GV->hasInitializer()) 1866 Changed |= ProcessInternalGlobal(GV, GVI); 1867 } 1868 return Changed; 1869} 1870 1871/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1872/// initializers have an init priority of 65535. 1873GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1874 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1875 I != E; ++I) 1876 if (I->getName() == "llvm.global_ctors") { 1877 // Found it, verify it's an array of { int, void()* }. 1878 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1879 if (!ATy) return 0; 1880 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1881 if (!STy || STy->getNumElements() != 2 || 1882 STy->getElementType(0) != Type::Int32Ty) return 0; 1883 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1884 if (!PFTy) return 0; 1885 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1886 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1887 FTy->getNumParams() != 0) 1888 return 0; 1889 1890 // Verify that the initializer is simple enough for us to handle. 1891 if (!I->hasInitializer()) return 0; 1892 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1893 if (!CA) return 0; 1894 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1895 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1896 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1897 continue; 1898 1899 // Must have a function or null ptr. 1900 if (!isa<Function>(CS->getOperand(1))) 1901 return 0; 1902 1903 // Init priority must be standard. 1904 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1905 if (!CI || CI->getZExtValue() != 65535) 1906 return 0; 1907 } else { 1908 return 0; 1909 } 1910 1911 return I; 1912 } 1913 return 0; 1914} 1915 1916/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1917/// return a list of the functions and null terminator as a vector. 1918static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1919 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1920 std::vector<Function*> Result; 1921 Result.reserve(CA->getNumOperands()); 1922 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1923 ConstantStruct *CS = cast<ConstantStruct>(*i); 1924 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1925 } 1926 return Result; 1927} 1928 1929/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1930/// specified array, returning the new global to use. 1931static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1932 const std::vector<Function*> &Ctors) { 1933 // If we made a change, reassemble the initializer list. 1934 std::vector<Constant*> CSVals; 1935 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1936 CSVals.push_back(0); 1937 1938 // Create the new init list. 1939 std::vector<Constant*> CAList; 1940 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1941 if (Ctors[i]) { 1942 CSVals[1] = Ctors[i]; 1943 } else { 1944 const Type *FTy = FunctionType::get(Type::VoidTy, 1945 std::vector<const Type*>(), false); 1946 const PointerType *PFTy = PointerType::getUnqual(FTy); 1947 CSVals[1] = Constant::getNullValue(PFTy); 1948 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1949 } 1950 CAList.push_back(ConstantStruct::get(CSVals)); 1951 } 1952 1953 // Create the array initializer. 1954 const Type *StructTy = 1955 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1956 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1957 CAList); 1958 1959 // If we didn't change the number of elements, don't create a new GV. 1960 if (CA->getType() == GCL->getInitializer()->getType()) { 1961 GCL->setInitializer(CA); 1962 return GCL; 1963 } 1964 1965 // Create the new global and insert it next to the existing list. 1966 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1967 GCL->getLinkage(), CA, "", 1968 (Module *)NULL, 1969 GCL->isThreadLocal()); 1970 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1971 NGV->takeName(GCL); 1972 1973 // Nuke the old list, replacing any uses with the new one. 1974 if (!GCL->use_empty()) { 1975 Constant *V = NGV; 1976 if (V->getType() != GCL->getType()) 1977 V = ConstantExpr::getBitCast(V, GCL->getType()); 1978 GCL->replaceAllUsesWith(V); 1979 } 1980 GCL->eraseFromParent(); 1981 1982 if (Ctors.size()) 1983 return NGV; 1984 else 1985 return 0; 1986} 1987 1988 1989static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, 1990 Value *V) { 1991 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1992 Constant *R = ComputedValues[V]; 1993 assert(R && "Reference to an uncomputed value!"); 1994 return R; 1995} 1996 1997/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1998/// enough for us to understand. In particular, if it is a cast of something, 1999/// we punt. We basically just support direct accesses to globals and GEP's of 2000/// globals. This should be kept up to date with CommitValueTo. 2001static bool isSimpleEnoughPointerToCommit(Constant *C) { 2002 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 2003 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 2004 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 2005 return !GV->isDeclaration(); // reject external globals. 2006 } 2007 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2008 // Handle a constantexpr gep. 2009 if (CE->getOpcode() == Instruction::GetElementPtr && 2010 isa<GlobalVariable>(CE->getOperand(0))) { 2011 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2012 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 2013 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 2014 return GV->hasInitializer() && 2015 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2016 } 2017 return false; 2018} 2019 2020/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 2021/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 2022/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 2023static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 2024 ConstantExpr *Addr, unsigned OpNo) { 2025 // Base case of the recursion. 2026 if (OpNo == Addr->getNumOperands()) { 2027 assert(Val->getType() == Init->getType() && "Type mismatch!"); 2028 return Val; 2029 } 2030 2031 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 2032 std::vector<Constant*> Elts; 2033 2034 // Break up the constant into its elements. 2035 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 2036 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 2037 Elts.push_back(cast<Constant>(*i)); 2038 } else if (isa<ConstantAggregateZero>(Init)) { 2039 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2040 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 2041 } else if (isa<UndefValue>(Init)) { 2042 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2043 Elts.push_back(UndefValue::get(STy->getElementType(i))); 2044 } else { 2045 assert(0 && "This code is out of sync with " 2046 " ConstantFoldLoadThroughGEPConstantExpr"); 2047 } 2048 2049 // Replace the element that we are supposed to. 2050 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 2051 unsigned Idx = CU->getZExtValue(); 2052 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 2053 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 2054 2055 // Return the modified struct. 2056 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 2057 } else { 2058 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 2059 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 2060 2061 // Break up the array into elements. 2062 std::vector<Constant*> Elts; 2063 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 2064 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 2065 Elts.push_back(cast<Constant>(*i)); 2066 } else if (isa<ConstantAggregateZero>(Init)) { 2067 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 2068 Elts.assign(ATy->getNumElements(), Elt); 2069 } else if (isa<UndefValue>(Init)) { 2070 Constant *Elt = UndefValue::get(ATy->getElementType()); 2071 Elts.assign(ATy->getNumElements(), Elt); 2072 } else { 2073 assert(0 && "This code is out of sync with " 2074 " ConstantFoldLoadThroughGEPConstantExpr"); 2075 } 2076 2077 assert(CI->getZExtValue() < ATy->getNumElements()); 2078 Elts[CI->getZExtValue()] = 2079 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 2080 return ConstantArray::get(ATy, Elts); 2081 } 2082} 2083 2084/// CommitValueTo - We have decided that Addr (which satisfies the predicate 2085/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 2086static void CommitValueTo(Constant *Val, Constant *Addr) { 2087 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 2088 assert(GV->hasInitializer()); 2089 GV->setInitializer(Val); 2090 return; 2091 } 2092 2093 ConstantExpr *CE = cast<ConstantExpr>(Addr); 2094 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2095 2096 Constant *Init = GV->getInitializer(); 2097 Init = EvaluateStoreInto(Init, Val, CE, 2); 2098 GV->setInitializer(Init); 2099} 2100 2101/// ComputeLoadResult - Return the value that would be computed by a load from 2102/// P after the stores reflected by 'memory' have been performed. If we can't 2103/// decide, return null. 2104static Constant *ComputeLoadResult(Constant *P, 2105 const DenseMap<Constant*, Constant*> &Memory) { 2106 // If this memory location has been recently stored, use the stored value: it 2107 // is the most up-to-date. 2108 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P); 2109 if (I != Memory.end()) return I->second; 2110 2111 // Access it. 2112 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 2113 if (GV->hasInitializer()) 2114 return GV->getInitializer(); 2115 return 0; 2116 } 2117 2118 // Handle a constantexpr getelementptr. 2119 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 2120 if (CE->getOpcode() == Instruction::GetElementPtr && 2121 isa<GlobalVariable>(CE->getOperand(0))) { 2122 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2123 if (GV->hasInitializer()) 2124 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2125 } 2126 2127 return 0; // don't know how to evaluate. 2128} 2129 2130/// EvaluateFunction - Evaluate a call to function F, returning true if 2131/// successful, false if we can't evaluate it. ActualArgs contains the formal 2132/// arguments for the function. 2133static bool EvaluateFunction(Function *F, Constant *&RetVal, 2134 const std::vector<Constant*> &ActualArgs, 2135 std::vector<Function*> &CallStack, 2136 DenseMap<Constant*, Constant*> &MutatedMemory, 2137 std::vector<GlobalVariable*> &AllocaTmps) { 2138 // Check to see if this function is already executing (recursion). If so, 2139 // bail out. TODO: we might want to accept limited recursion. 2140 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 2141 return false; 2142 2143 CallStack.push_back(F); 2144 2145 /// Values - As we compute SSA register values, we store their contents here. 2146 DenseMap<Value*, Constant*> Values; 2147 2148 // Initialize arguments to the incoming values specified. 2149 unsigned ArgNo = 0; 2150 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 2151 ++AI, ++ArgNo) 2152 Values[AI] = ActualArgs[ArgNo]; 2153 2154 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 2155 /// we can only evaluate any one basic block at most once. This set keeps 2156 /// track of what we have executed so we can detect recursive cases etc. 2157 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 2158 2159 // CurInst - The current instruction we're evaluating. 2160 BasicBlock::iterator CurInst = F->begin()->begin(); 2161 2162 // This is the main evaluation loop. 2163 while (1) { 2164 Constant *InstResult = 0; 2165 2166 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2167 if (SI->isVolatile()) return false; // no volatile accesses. 2168 Constant *Ptr = getVal(Values, SI->getOperand(1)); 2169 if (!isSimpleEnoughPointerToCommit(Ptr)) 2170 // If this is too complex for us to commit, reject it. 2171 return false; 2172 Constant *Val = getVal(Values, SI->getOperand(0)); 2173 MutatedMemory[Ptr] = Val; 2174 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2175 InstResult = ConstantExpr::get(BO->getOpcode(), 2176 getVal(Values, BO->getOperand(0)), 2177 getVal(Values, BO->getOperand(1))); 2178 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2179 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2180 getVal(Values, CI->getOperand(0)), 2181 getVal(Values, CI->getOperand(1))); 2182 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2183 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2184 getVal(Values, CI->getOperand(0)), 2185 CI->getType()); 2186 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2187 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2188 getVal(Values, SI->getOperand(1)), 2189 getVal(Values, SI->getOperand(2))); 2190 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2191 Constant *P = getVal(Values, GEP->getOperand(0)); 2192 SmallVector<Constant*, 8> GEPOps; 2193 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2194 i != e; ++i) 2195 GEPOps.push_back(getVal(Values, *i)); 2196 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2197 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2198 if (LI->isVolatile()) return false; // no volatile accesses. 2199 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2200 MutatedMemory); 2201 if (InstResult == 0) return false; // Could not evaluate load. 2202 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2203 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2204 const Type *Ty = AI->getType()->getElementType(); 2205 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2206 GlobalValue::InternalLinkage, 2207 UndefValue::get(Ty), 2208 AI->getName())); 2209 InstResult = AllocaTmps.back(); 2210 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2211 2212 // Debug info can safely be ignored here. 2213 if (isa<DbgInfoIntrinsic>(CI)) { 2214 ++CurInst; 2215 continue; 2216 } 2217 2218 // Cannot handle inline asm. 2219 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2220 2221 // Resolve function pointers. 2222 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2223 if (!Callee) return false; // Cannot resolve. 2224 2225 std::vector<Constant*> Formals; 2226 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end(); 2227 i != e; ++i) 2228 Formals.push_back(getVal(Values, *i)); 2229 2230 if (Callee->isDeclaration()) { 2231 // If this is a function we can constant fold, do it. 2232 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2233 Formals.size())) { 2234 InstResult = C; 2235 } else { 2236 return false; 2237 } 2238 } else { 2239 if (Callee->getFunctionType()->isVarArg()) 2240 return false; 2241 2242 Constant *RetVal; 2243 // Execute the call, if successful, use the return value. 2244 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2245 MutatedMemory, AllocaTmps)) 2246 return false; 2247 InstResult = RetVal; 2248 } 2249 } else if (isa<TerminatorInst>(CurInst)) { 2250 BasicBlock *NewBB = 0; 2251 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2252 if (BI->isUnconditional()) { 2253 NewBB = BI->getSuccessor(0); 2254 } else { 2255 ConstantInt *Cond = 2256 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2257 if (!Cond) return false; // Cannot determine. 2258 2259 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2260 } 2261 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2262 ConstantInt *Val = 2263 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2264 if (!Val) return false; // Cannot determine. 2265 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2266 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2267 if (RI->getNumOperands()) 2268 RetVal = getVal(Values, RI->getOperand(0)); 2269 2270 CallStack.pop_back(); // return from fn. 2271 return true; // We succeeded at evaluating this ctor! 2272 } else { 2273 // invoke, unwind, unreachable. 2274 return false; // Cannot handle this terminator. 2275 } 2276 2277 // Okay, we succeeded in evaluating this control flow. See if we have 2278 // executed the new block before. If so, we have a looping function, 2279 // which we cannot evaluate in reasonable time. 2280 if (!ExecutedBlocks.insert(NewBB)) 2281 return false; // looped! 2282 2283 // Okay, we have never been in this block before. Check to see if there 2284 // are any PHI nodes. If so, evaluate them with information about where 2285 // we came from. 2286 BasicBlock *OldBB = CurInst->getParent(); 2287 CurInst = NewBB->begin(); 2288 PHINode *PN; 2289 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2290 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2291 2292 // Do NOT increment CurInst. We know that the terminator had no value. 2293 continue; 2294 } else { 2295 // Did not know how to evaluate this! 2296 return false; 2297 } 2298 2299 if (!CurInst->use_empty()) 2300 Values[CurInst] = InstResult; 2301 2302 // Advance program counter. 2303 ++CurInst; 2304 } 2305} 2306 2307/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2308/// we can. Return true if we can, false otherwise. 2309static bool EvaluateStaticConstructor(Function *F) { 2310 /// MutatedMemory - For each store we execute, we update this map. Loads 2311 /// check this to get the most up-to-date value. If evaluation is successful, 2312 /// this state is committed to the process. 2313 DenseMap<Constant*, Constant*> MutatedMemory; 2314 2315 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2316 /// to represent its body. This vector is needed so we can delete the 2317 /// temporary globals when we are done. 2318 std::vector<GlobalVariable*> AllocaTmps; 2319 2320 /// CallStack - This is used to detect recursion. In pathological situations 2321 /// we could hit exponential behavior, but at least there is nothing 2322 /// unbounded. 2323 std::vector<Function*> CallStack; 2324 2325 // Call the function. 2326 Constant *RetValDummy; 2327 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2328 CallStack, MutatedMemory, AllocaTmps); 2329 if (EvalSuccess) { 2330 // We succeeded at evaluation: commit the result. 2331 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2332 << F->getName() << "' to " << MutatedMemory.size() 2333 << " stores.\n"; 2334 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2335 E = MutatedMemory.end(); I != E; ++I) 2336 CommitValueTo(I->second, I->first); 2337 } 2338 2339 // At this point, we are done interpreting. If we created any 'alloca' 2340 // temporaries, release them now. 2341 while (!AllocaTmps.empty()) { 2342 GlobalVariable *Tmp = AllocaTmps.back(); 2343 AllocaTmps.pop_back(); 2344 2345 // If there are still users of the alloca, the program is doing something 2346 // silly, e.g. storing the address of the alloca somewhere and using it 2347 // later. Since this is undefined, we'll just make it be null. 2348 if (!Tmp->use_empty()) 2349 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2350 delete Tmp; 2351 } 2352 2353 return EvalSuccess; 2354} 2355 2356 2357 2358/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2359/// Return true if anything changed. 2360bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2361 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2362 bool MadeChange = false; 2363 if (Ctors.empty()) return false; 2364 2365 // Loop over global ctors, optimizing them when we can. 2366 for (unsigned i = 0; i != Ctors.size(); ++i) { 2367 Function *F = Ctors[i]; 2368 // Found a null terminator in the middle of the list, prune off the rest of 2369 // the list. 2370 if (F == 0) { 2371 if (i != Ctors.size()-1) { 2372 Ctors.resize(i+1); 2373 MadeChange = true; 2374 } 2375 break; 2376 } 2377 2378 // We cannot simplify external ctor functions. 2379 if (F->empty()) continue; 2380 2381 // If we can evaluate the ctor at compile time, do. 2382 if (EvaluateStaticConstructor(F)) { 2383 Ctors.erase(Ctors.begin()+i); 2384 MadeChange = true; 2385 --i; 2386 ++NumCtorsEvaluated; 2387 continue; 2388 } 2389 } 2390 2391 if (!MadeChange) return false; 2392 2393 GCL = InstallGlobalCtors(GCL, Ctors); 2394 return true; 2395} 2396 2397bool GlobalOpt::OptimizeGlobalAliases(Module &M) { 2398 bool Changed = false; 2399 2400 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 2401 I != E;) { 2402 Module::alias_iterator J = I++; 2403 // Aliases without names cannot be referenced outside this module. 2404 if (!J->hasName() && !J->isDeclaration()) 2405 J->setLinkage(GlobalValue::InternalLinkage); 2406 // If the aliasee may change at link time, nothing can be done - bail out. 2407 if (J->mayBeOverridden()) 2408 continue; 2409 2410 Constant *Aliasee = J->getAliasee(); 2411 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts()); 2412 Target->removeDeadConstantUsers(); 2413 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse(); 2414 2415 // Make all users of the alias use the aliasee instead. 2416 if (!J->use_empty()) { 2417 J->replaceAllUsesWith(Aliasee); 2418 ++NumAliasesResolved; 2419 Changed = true; 2420 } 2421 2422 // If the aliasee has internal linkage, give it the name and linkage 2423 // of the alias, and delete the alias. This turns: 2424 // define internal ... @f(...) 2425 // @a = alias ... @f 2426 // into: 2427 // define ... @a(...) 2428 if (!Target->hasLocalLinkage()) 2429 continue; 2430 2431 // The transform is only useful if the alias does not have internal linkage. 2432 if (J->hasLocalLinkage()) 2433 continue; 2434 2435 // Do not perform the transform if multiple aliases potentially target the 2436 // aliasee. This check also ensures that it is safe to replace the section 2437 // and other attributes of the aliasee with those of the alias. 2438 if (!hasOneUse) 2439 continue; 2440 2441 // Give the aliasee the name, linkage and other attributes of the alias. 2442 Target->takeName(J); 2443 Target->setLinkage(J->getLinkage()); 2444 Target->GlobalValue::copyAttributesFrom(J); 2445 2446 // Delete the alias. 2447 M.getAliasList().erase(J); 2448 ++NumAliasesRemoved; 2449 Changed = true; 2450 } 2451 2452 return Changed; 2453} 2454 2455bool GlobalOpt::runOnModule(Module &M) { 2456 bool Changed = false; 2457 2458 // Try to find the llvm.globalctors list. 2459 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2460 2461 bool LocalChange = true; 2462 while (LocalChange) { 2463 LocalChange = false; 2464 2465 // Delete functions that are trivially dead, ccc -> fastcc 2466 LocalChange |= OptimizeFunctions(M); 2467 2468 // Optimize global_ctors list. 2469 if (GlobalCtors) 2470 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2471 2472 // Optimize non-address-taken globals. 2473 LocalChange |= OptimizeGlobalVars(M); 2474 2475 // Resolve aliases, when possible. 2476 LocalChange |= OptimizeGlobalAliases(M); 2477 Changed |= LocalChange; 2478 } 2479 2480 // TODO: Move all global ctors functions to the end of the module for code 2481 // layout. 2482 2483 return Changed; 2484}
|