GlobalOpt.cpp revision 3e5206d4aa51829b72fd7b2b42652a484cfb9316
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.getTypePaddedSize(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){ 935 Instruction *Inst = dyn_cast<Instruction>(*UI); 936 if (Inst == 0) return false; 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 // 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 we have already recursively analyzed this PHI, then it is safe. 1048 if (LoadUsingPHIs.insert(PN)) 1049 continue; 1050 1051 // Make sure all uses of the PHI are simple enough to transform. 1052 if (!LoadUsesSimpleEnoughForHeapSRA(PN, LoadUsingPHIs)) 1053 return false; 1054 1055 continue; 1056 } 1057 1058 // Otherwise we don't know what this is, not ok. 1059 return false; 1060 } 1061 1062 return true; 1063} 1064 1065 1066/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1067/// GV are simple enough to perform HeapSRA, return true. 1068static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV, 1069 MallocInst *MI) { 1070 SmallPtrSet<PHINode*, 32> LoadUsingPHIs; 1071 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E; 1072 ++UI) 1073 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1074 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs)) 1075 return false; 1076 1077 // If we reach here, we know that all uses of the loads and transitive uses 1078 // (through PHI nodes) are simple enough to transform. However, we don't know 1079 // that all inputs the to the PHI nodes are in the same equivalence sets. 1080 // Check to verify that all operands of the PHIs are either PHIS that can be 1081 // transformed, loads from GV, or MI itself. 1082 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(), 1083 E = LoadUsingPHIs.end(); I != E; ++I) { 1084 PHINode *PN = *I; 1085 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) { 1086 Value *InVal = PN->getIncomingValue(op); 1087 1088 // PHI of the stored value itself is ok. 1089 if (InVal == MI) continue; 1090 1091 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) { 1092 // One of the PHIs in our set is (optimistically) ok. 1093 if (LoadUsingPHIs.count(InPN)) 1094 continue; 1095 return false; 1096 } 1097 1098 // Load from GV is ok. 1099 if (LoadInst *LI = dyn_cast<LoadInst>(InVal)) 1100 if (LI->getOperand(0) == GV) 1101 continue; 1102 1103 // UNDEF? NULL? 1104 1105 // Anything else is rejected. 1106 return false; 1107 } 1108 } 1109 1110 return true; 1111} 1112 1113static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, 1114 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1115 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1116 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V]; 1117 1118 if (FieldNo >= FieldVals.size()) 1119 FieldVals.resize(FieldNo+1); 1120 1121 // If we already have this value, just reuse the previously scalarized 1122 // version. 1123 if (Value *FieldVal = FieldVals[FieldNo]) 1124 return FieldVal; 1125 1126 // Depending on what instruction this is, we have several cases. 1127 Value *Result; 1128 if (LoadInst *LI = dyn_cast<LoadInst>(V)) { 1129 // This is a scalarized version of the load from the global. Just create 1130 // a new Load of the scalarized global. 1131 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo, 1132 InsertedScalarizedValues, 1133 PHIsToRewrite), 1134 LI->getName()+".f" + utostr(FieldNo), LI); 1135 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 1136 // PN's type is pointer to struct. Make a new PHI of pointer to struct 1137 // field. 1138 const StructType *ST = 1139 cast<StructType>(cast<PointerType>(PN->getType())->getElementType()); 1140 1141 Result =PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)), 1142 PN->getName()+".f"+utostr(FieldNo), PN); 1143 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); 1144 } else { 1145 assert(0 && "Unknown usable value"); 1146 Result = 0; 1147 } 1148 1149 return FieldVals[FieldNo] = Result; 1150} 1151 1152/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1153/// the load, rewrite the derived value to use the HeapSRoA'd load. 1154static void RewriteHeapSROALoadUser(Instruction *LoadUser, 1155 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1156 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1157 // If this is a comparison against null, handle it. 1158 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1159 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1160 // If we have a setcc of the loaded pointer, we can use a setcc of any 1161 // field. 1162 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0, 1163 InsertedScalarizedValues, PHIsToRewrite); 1164 1165 Value *New = new ICmpInst(SCI->getPredicate(), NPtr, 1166 Constant::getNullValue(NPtr->getType()), 1167 SCI->getName(), SCI); 1168 SCI->replaceAllUsesWith(New); 1169 SCI->eraseFromParent(); 1170 return; 1171 } 1172 1173 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...' 1174 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1175 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1176 && "Unexpected GEPI!"); 1177 1178 // Load the pointer for this field. 1179 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1180 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo, 1181 InsertedScalarizedValues, PHIsToRewrite); 1182 1183 // Create the new GEP idx vector. 1184 SmallVector<Value*, 8> GEPIdx; 1185 GEPIdx.push_back(GEPI->getOperand(1)); 1186 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1187 1188 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1189 GEPIdx.begin(), GEPIdx.end(), 1190 GEPI->getName(), GEPI); 1191 GEPI->replaceAllUsesWith(NGEPI); 1192 GEPI->eraseFromParent(); 1193 return; 1194 } 1195 1196 // Recursively transform the users of PHI nodes. This will lazily create the 1197 // PHIs that are needed for individual elements. Keep track of what PHIs we 1198 // see in InsertedScalarizedValues so that we don't get infinite loops (very 1199 // antisocial). If the PHI is already in InsertedScalarizedValues, it has 1200 // already been seen first by another load, so its uses have already been 1201 // processed. 1202 PHINode *PN = cast<PHINode>(LoadUser); 1203 bool Inserted; 1204 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos; 1205 tie(InsertPos, Inserted) = 1206 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>())); 1207 if (!Inserted) return; 1208 1209 // If this is the first time we've seen this PHI, recursively process all 1210 // users. 1211 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) { 1212 Instruction *User = cast<Instruction>(*UI++); 1213 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1214 } 1215} 1216 1217/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1218/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1219/// use FieldGlobals instead. All uses of loaded values satisfy 1220/// AllGlobalLoadUsesSimpleEnoughForHeapSRA. 1221static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1222 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1223 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1224 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end(); 1225 UI != E; ) { 1226 Instruction *User = cast<Instruction>(*UI++); 1227 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1228 } 1229 1230 if (Load->use_empty()) { 1231 Load->eraseFromParent(); 1232 InsertedScalarizedValues.erase(Load); 1233 } 1234} 1235 1236/// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break 1237/// it up into multiple allocations of arrays of the fields. 1238static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){ 1239 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI; 1240 const StructType *STy = cast<StructType>(MI->getAllocatedType()); 1241 1242 // There is guaranteed to be at least one use of the malloc (storing 1243 // it into GV). If there are other uses, change them to be uses of 1244 // the global to simplify later code. This also deletes the store 1245 // into GV. 1246 ReplaceUsesOfMallocWithGlobal(MI, GV); 1247 1248 // Okay, at this point, there are no users of the malloc. Insert N 1249 // new mallocs at the same place as MI, and N globals. 1250 std::vector<Value*> FieldGlobals; 1251 std::vector<MallocInst*> FieldMallocs; 1252 1253 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1254 const Type *FieldTy = STy->getElementType(FieldNo); 1255 const Type *PFieldTy = PointerType::getUnqual(FieldTy); 1256 1257 GlobalVariable *NGV = 1258 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage, 1259 Constant::getNullValue(PFieldTy), 1260 GV->getName() + ".f" + utostr(FieldNo), GV, 1261 GV->isThreadLocal()); 1262 FieldGlobals.push_back(NGV); 1263 1264 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(), 1265 MI->getName() + ".f" + utostr(FieldNo),MI); 1266 FieldMallocs.push_back(NMI); 1267 new StoreInst(NMI, NGV, MI); 1268 } 1269 1270 // The tricky aspect of this transformation is handling the case when malloc 1271 // fails. In the original code, malloc failing would set the result pointer 1272 // of malloc to null. In this case, some mallocs could succeed and others 1273 // could fail. As such, we emit code that looks like this: 1274 // F0 = malloc(field0) 1275 // F1 = malloc(field1) 1276 // F2 = malloc(field2) 1277 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1278 // if (F0) { free(F0); F0 = 0; } 1279 // if (F1) { free(F1); F1 = 0; } 1280 // if (F2) { free(F2); F2 = 0; } 1281 // } 1282 Value *RunningOr = 0; 1283 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1284 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i], 1285 Constant::getNullValue(FieldMallocs[i]->getType()), 1286 "isnull", MI); 1287 if (!RunningOr) 1288 RunningOr = Cond; // First seteq 1289 else 1290 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI); 1291 } 1292 1293 // Split the basic block at the old malloc. 1294 BasicBlock *OrigBB = MI->getParent(); 1295 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont"); 1296 1297 // Create the block to check the first condition. Put all these blocks at the 1298 // end of the function as they are unlikely to be executed. 1299 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null", 1300 OrigBB->getParent()); 1301 1302 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1303 // branch on RunningOr. 1304 OrigBB->getTerminator()->eraseFromParent(); 1305 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1306 1307 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1308 // pointer, because some may be null while others are not. 1309 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1310 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1311 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal, 1312 Constant::getNullValue(GVVal->getType()), 1313 "tmp", NullPtrBlock); 1314 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent()); 1315 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent()); 1316 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock); 1317 1318 // Fill in FreeBlock. 1319 new FreeInst(GVVal, FreeBlock); 1320 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1321 FreeBlock); 1322 BranchInst::Create(NextBlock, FreeBlock); 1323 1324 NullPtrBlock = NextBlock; 1325 } 1326 1327 BranchInst::Create(ContBB, NullPtrBlock); 1328 1329 // MI is no longer needed, remove it. 1330 MI->eraseFromParent(); 1331 1332 /// InsertedScalarizedLoads - As we process loads, if we can't immediately 1333 /// update all uses of the load, keep track of what scalarized loads are 1334 /// inserted for a given load. 1335 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues; 1336 InsertedScalarizedValues[GV] = FieldGlobals; 1337 1338 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite; 1339 1340 // Okay, the malloc site is completely handled. All of the uses of GV are now 1341 // loads, and all uses of those loads are simple. Rewrite them to use loads 1342 // of the per-field globals instead. 1343 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) { 1344 Instruction *User = cast<Instruction>(*UI++); 1345 1346 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1347 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite); 1348 continue; 1349 } 1350 1351 // Must be a store of null. 1352 StoreInst *SI = cast<StoreInst>(User); 1353 assert(isa<ConstantPointerNull>(SI->getOperand(0)) && 1354 "Unexpected heap-sra user!"); 1355 1356 // Insert a store of null into each global. 1357 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1358 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType()); 1359 Constant *Null = Constant::getNullValue(PT->getElementType()); 1360 new StoreInst(Null, FieldGlobals[i], SI); 1361 } 1362 // Erase the original store. 1363 SI->eraseFromParent(); 1364 } 1365 1366 // While we have PHIs that are interesting to rewrite, do it. 1367 while (!PHIsToRewrite.empty()) { 1368 PHINode *PN = PHIsToRewrite.back().first; 1369 unsigned FieldNo = PHIsToRewrite.back().second; 1370 PHIsToRewrite.pop_back(); 1371 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]); 1372 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi"); 1373 1374 // Add all the incoming values. This can materialize more phis. 1375 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1376 Value *InVal = PN->getIncomingValue(i); 1377 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues, 1378 PHIsToRewrite); 1379 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i)); 1380 } 1381 } 1382 1383 // Drop all inter-phi links and any loads that made it this far. 1384 for (DenseMap<Value*, std::vector<Value*> >::iterator 1385 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1386 I != E; ++I) { 1387 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1388 PN->dropAllReferences(); 1389 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1390 LI->dropAllReferences(); 1391 } 1392 1393 // Delete all the phis and loads now that inter-references are dead. 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->eraseFromParent(); 1399 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1400 LI->eraseFromParent(); 1401 } 1402 1403 // The old global is now dead, remove it. 1404 GV->eraseFromParent(); 1405 1406 ++NumHeapSRA; 1407 return cast<GlobalVariable>(FieldGlobals[0]); 1408} 1409 1410/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1411/// pointer global variable with a single value stored it that is a malloc or 1412/// cast of malloc. 1413static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, 1414 MallocInst *MI, 1415 Module::global_iterator &GVI, 1416 TargetData &TD) { 1417 // If this is a malloc of an abstract type, don't touch it. 1418 if (!MI->getAllocatedType()->isSized()) 1419 return false; 1420 1421 // We can't optimize this global unless all uses of it are *known* to be 1422 // of the malloc value, not of the null initializer value (consider a use 1423 // that compares the global's value against zero to see if the malloc has 1424 // been reached). To do this, we check to see if all uses of the global 1425 // would trap if the global were null: this proves that they must all 1426 // happen after the malloc. 1427 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1428 return false; 1429 1430 // We can't optimize this if the malloc itself is used in a complex way, 1431 // for example, being stored into multiple globals. This allows the 1432 // malloc to be stored into the specified global, loaded setcc'd, and 1433 // GEP'd. These are all things we could transform to using the global 1434 // for. 1435 { 1436 SmallPtrSet<PHINode*, 8> PHIs; 1437 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs)) 1438 return false; 1439 } 1440 1441 1442 // If we have a global that is only initialized with a fixed size malloc, 1443 // transform the program to use global memory instead of malloc'd memory. 1444 // This eliminates dynamic allocation, avoids an indirection accessing the 1445 // data, and exposes the resultant global to further GlobalOpt. 1446 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) { 1447 // Restrict this transformation to only working on small allocations 1448 // (2048 bytes currently), as we don't want to introduce a 16M global or 1449 // something. 1450 if (NElements->getZExtValue()* 1451 TD.getTypePaddedSize(MI->getAllocatedType()) < 2048) { 1452 GVI = OptimizeGlobalAddressOfMalloc(GV, MI); 1453 return true; 1454 } 1455 } 1456 1457 // If the allocation is an array of structures, consider transforming this 1458 // into multiple malloc'd arrays, one for each field. This is basically 1459 // SRoA for malloc'd memory. 1460 const Type *AllocTy = MI->getAllocatedType(); 1461 1462 // If this is an allocation of a fixed size array of structs, analyze as a 1463 // variable size array. malloc [100 x struct],1 -> malloc struct, 100 1464 if (!MI->isArrayAllocation()) 1465 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy)) 1466 AllocTy = AT->getElementType(); 1467 1468 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) { 1469 // This the structure has an unreasonable number of fields, leave it 1470 // alone. 1471 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 && 1472 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) { 1473 1474 // If this is a fixed size array, transform the Malloc to be an alloc of 1475 // structs. malloc [100 x struct],1 -> malloc struct, 100 1476 if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) { 1477 MallocInst *NewMI = 1478 new MallocInst(AllocSTy, 1479 ConstantInt::get(Type::Int32Ty, AT->getNumElements()), 1480 "", MI); 1481 NewMI->takeName(MI); 1482 Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI); 1483 MI->replaceAllUsesWith(Cast); 1484 MI->eraseFromParent(); 1485 MI = NewMI; 1486 } 1487 1488 GVI = PerformHeapAllocSRoA(GV, MI); 1489 return true; 1490 } 1491 } 1492 1493 return false; 1494} 1495 1496// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1497// that only one value (besides its initializer) is ever stored to the global. 1498static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1499 Module::global_iterator &GVI, 1500 TargetData &TD) { 1501 // Ignore no-op GEPs and bitcasts. 1502 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1503 1504 // If we are dealing with a pointer global that is initialized to null and 1505 // only has one (non-null) value stored into it, then we can optimize any 1506 // users of the loaded value (often calls and loads) that would trap if the 1507 // value was null. 1508 if (isa<PointerType>(GV->getInitializer()->getType()) && 1509 GV->getInitializer()->isNullValue()) { 1510 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1511 if (GV->getInitializer()->getType() != SOVC->getType()) 1512 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1513 1514 // Optimize away any trapping uses of the loaded value. 1515 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1516 return true; 1517 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) { 1518 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD)) 1519 return true; 1520 } 1521 } 1522 1523 return false; 1524} 1525 1526/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1527/// two values ever stored into GV are its initializer and OtherVal. See if we 1528/// can shrink the global into a boolean and select between the two values 1529/// whenever it is used. This exposes the values to other scalar optimizations. 1530static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1531 const Type *GVElType = GV->getType()->getElementType(); 1532 1533 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1534 // an FP value, pointer or vector, don't do this optimization because a select 1535 // between them is very expensive and unlikely to lead to later 1536 // simplification. In these cases, we typically end up with "cond ? v1 : v2" 1537 // where v1 and v2 both require constant pool loads, a big loss. 1538 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() || 1539 isa<PointerType>(GVElType) || isa<VectorType>(GVElType)) 1540 return false; 1541 1542 // Walk the use list of the global seeing if all the uses are load or store. 1543 // If there is anything else, bail out. 1544 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1545 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1546 return false; 1547 1548 DOUT << " *** SHRINKING TO BOOL: " << *GV; 1549 1550 // Create the new global, initializing it to false. 1551 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false, 1552 GlobalValue::InternalLinkage, 1553 ConstantInt::getFalse(), 1554 "", 1555 (Module *)NULL, 1556 GV->isThreadLocal()); 1557 NewGV->takeName(GV); 1558 GV->getParent()->getGlobalList().insert(GV, NewGV); 1559 1560 Constant *InitVal = GV->getInitializer(); 1561 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!"); 1562 1563 // If initialized to zero and storing one into the global, we can use a cast 1564 // instead of a select to synthesize the desired value. 1565 bool IsOneZero = false; 1566 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1567 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1568 1569 while (!GV->use_empty()) { 1570 Instruction *UI = cast<Instruction>(GV->use_back()); 1571 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1572 // Change the store into a boolean store. 1573 bool StoringOther = SI->getOperand(0) == OtherVal; 1574 // Only do this if we weren't storing a loaded value. 1575 Value *StoreVal; 1576 if (StoringOther || SI->getOperand(0) == InitVal) 1577 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther); 1578 else { 1579 // Otherwise, we are storing a previously loaded copy. To do this, 1580 // change the copy from copying the original value to just copying the 1581 // bool. 1582 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1583 1584 // If we're already replaced the input, StoredVal will be a cast or 1585 // select instruction. If not, it will be a load of the original 1586 // global. 1587 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1588 assert(LI->getOperand(0) == GV && "Not a copy!"); 1589 // Insert a new load, to preserve the saved value. 1590 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1591 } else { 1592 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1593 "This is not a form that we understand!"); 1594 StoreVal = StoredVal->getOperand(0); 1595 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1596 } 1597 } 1598 new StoreInst(StoreVal, NewGV, SI); 1599 } else { 1600 // Change the load into a load of bool then a select. 1601 LoadInst *LI = cast<LoadInst>(UI); 1602 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1603 Value *NSI; 1604 if (IsOneZero) 1605 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1606 else 1607 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1608 NSI->takeName(LI); 1609 LI->replaceAllUsesWith(NSI); 1610 } 1611 UI->eraseFromParent(); 1612 } 1613 1614 GV->eraseFromParent(); 1615 return true; 1616} 1617 1618 1619/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1620/// it if possible. If we make a change, return true. 1621bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1622 Module::global_iterator &GVI) { 1623 SmallPtrSet<PHINode*, 16> PHIUsers; 1624 GlobalStatus GS; 1625 GV->removeDeadConstantUsers(); 1626 1627 if (GV->use_empty()) { 1628 DOUT << "GLOBAL DEAD: " << *GV; 1629 GV->eraseFromParent(); 1630 ++NumDeleted; 1631 return true; 1632 } 1633 1634 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1635#if 0 1636 cerr << "Global: " << *GV; 1637 cerr << " isLoaded = " << GS.isLoaded << "\n"; 1638 cerr << " StoredType = "; 1639 switch (GS.StoredType) { 1640 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break; 1641 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break; 1642 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break; 1643 case GlobalStatus::isStored: cerr << "stored\n"; break; 1644 } 1645 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1646 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"; 1647 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1648 cerr << " AccessingFunction = " << GS.AccessingFunction->getName() 1649 << "\n"; 1650 cerr << " HasMultipleAccessingFunctions = " 1651 << GS.HasMultipleAccessingFunctions << "\n"; 1652 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n"; 1653 cerr << "\n"; 1654#endif 1655 1656 // If this is a first class global and has only one accessing function 1657 // and this function is main (which we know is not recursive we can make 1658 // this global a local variable) we replace the global with a local alloca 1659 // in this function. 1660 // 1661 // NOTE: It doesn't make sense to promote non single-value types since we 1662 // are just replacing static memory to stack memory. 1663 if (!GS.HasMultipleAccessingFunctions && 1664 GS.AccessingFunction && !GS.HasNonInstructionUser && 1665 GV->getType()->getElementType()->isSingleValueType() && 1666 GS.AccessingFunction->getName() == "main" && 1667 GS.AccessingFunction->hasExternalLinkage()) { 1668 DOUT << "LOCALIZING GLOBAL: " << *GV; 1669 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin(); 1670 const Type* ElemTy = GV->getType()->getElementType(); 1671 // FIXME: Pass Global's alignment when globals have alignment 1672 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI); 1673 if (!isa<UndefValue>(GV->getInitializer())) 1674 new StoreInst(GV->getInitializer(), Alloca, FirstI); 1675 1676 GV->replaceAllUsesWith(Alloca); 1677 GV->eraseFromParent(); 1678 ++NumLocalized; 1679 return true; 1680 } 1681 1682 // If the global is never loaded (but may be stored to), it is dead. 1683 // Delete it now. 1684 if (!GS.isLoaded) { 1685 DOUT << "GLOBAL NEVER LOADED: " << *GV; 1686 1687 // Delete any stores we can find to the global. We may not be able to 1688 // make it completely dead though. 1689 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1690 1691 // If the global is dead now, delete it. 1692 if (GV->use_empty()) { 1693 GV->eraseFromParent(); 1694 ++NumDeleted; 1695 Changed = true; 1696 } 1697 return Changed; 1698 1699 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1700 DOUT << "MARKING CONSTANT: " << *GV; 1701 GV->setConstant(true); 1702 1703 // Clean up any obviously simplifiable users now. 1704 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1705 1706 // If the global is dead now, just nuke it. 1707 if (GV->use_empty()) { 1708 DOUT << " *** Marking constant allowed us to simplify " 1709 << "all users and delete global!\n"; 1710 GV->eraseFromParent(); 1711 ++NumDeleted; 1712 } 1713 1714 ++NumMarked; 1715 return true; 1716 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1717 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, 1718 getAnalysis<TargetData>())) { 1719 GVI = FirstNewGV; // Don't skip the newly produced globals! 1720 return true; 1721 } 1722 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1723 // If the initial value for the global was an undef value, and if only 1724 // one other value was stored into it, we can just change the 1725 // initializer to be the stored value, then delete all stores to the 1726 // global. This allows us to mark it constant. 1727 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1728 if (isa<UndefValue>(GV->getInitializer())) { 1729 // Change the initial value here. 1730 GV->setInitializer(SOVConstant); 1731 1732 // Clean up any obviously simplifiable users now. 1733 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1734 1735 if (GV->use_empty()) { 1736 DOUT << " *** Substituting initializer allowed us to " 1737 << "simplify all users and delete global!\n"; 1738 GV->eraseFromParent(); 1739 ++NumDeleted; 1740 } else { 1741 GVI = GV; 1742 } 1743 ++NumSubstitute; 1744 return true; 1745 } 1746 1747 // Try to optimize globals based on the knowledge that only one value 1748 // (besides its initializer) is ever stored to the global. 1749 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1750 getAnalysis<TargetData>())) 1751 return true; 1752 1753 // Otherwise, if the global was not a boolean, we can shrink it to be a 1754 // boolean. 1755 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1756 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1757 ++NumShrunkToBool; 1758 return true; 1759 } 1760 } 1761 } 1762 return false; 1763} 1764 1765/// OnlyCalledDirectly - Return true if the specified function is only called 1766/// directly. In other words, its address is never taken. 1767static bool OnlyCalledDirectly(Function *F) { 1768 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1769 Instruction *User = dyn_cast<Instruction>(*UI); 1770 if (!User) return false; 1771 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false; 1772 1773 // See if the function address is passed as an argument. 1774 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end(); 1775 i != e; ++i) 1776 if (*i == F) return false; 1777 } 1778 return true; 1779} 1780 1781/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1782/// function, changing them to FastCC. 1783static void ChangeCalleesToFastCall(Function *F) { 1784 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1785 CallSite User(cast<Instruction>(*UI)); 1786 User.setCallingConv(CallingConv::Fast); 1787 } 1788} 1789 1790static AttrListPtr StripNest(const AttrListPtr &Attrs) { 1791 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1792 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) 1793 continue; 1794 1795 // There can be only one. 1796 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); 1797 } 1798 1799 return Attrs; 1800} 1801 1802static void RemoveNestAttribute(Function *F) { 1803 F->setAttributes(StripNest(F->getAttributes())); 1804 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1805 CallSite User(cast<Instruction>(*UI)); 1806 User.setAttributes(StripNest(User.getAttributes())); 1807 } 1808} 1809 1810bool GlobalOpt::OptimizeFunctions(Module &M) { 1811 bool Changed = false; 1812 // Optimize functions. 1813 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1814 Function *F = FI++; 1815 // Functions without names cannot be referenced outside this module. 1816 if (!F->hasName() && !F->isDeclaration()) 1817 F->setLinkage(GlobalValue::InternalLinkage); 1818 F->removeDeadConstantUsers(); 1819 if (F->use_empty() && (F->hasLocalLinkage() || 1820 F->hasLinkOnceLinkage())) { 1821 M.getFunctionList().erase(F); 1822 Changed = true; 1823 ++NumFnDeleted; 1824 } else if (F->hasLocalLinkage()) { 1825 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1826 OnlyCalledDirectly(F)) { 1827 // If this function has C calling conventions, is not a varargs 1828 // function, and is only called directly, promote it to use the Fast 1829 // calling convention. 1830 F->setCallingConv(CallingConv::Fast); 1831 ChangeCalleesToFastCall(F); 1832 ++NumFastCallFns; 1833 Changed = true; 1834 } 1835 1836 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1837 OnlyCalledDirectly(F)) { 1838 // The function is not used by a trampoline intrinsic, so it is safe 1839 // to remove the 'nest' attribute. 1840 RemoveNestAttribute(F); 1841 ++NumNestRemoved; 1842 Changed = true; 1843 } 1844 } 1845 } 1846 return Changed; 1847} 1848 1849bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1850 bool Changed = false; 1851 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1852 GVI != E; ) { 1853 GlobalVariable *GV = GVI++; 1854 // Global variables without names cannot be referenced outside this module. 1855 if (!GV->hasName() && !GV->isDeclaration()) 1856 GV->setLinkage(GlobalValue::InternalLinkage); 1857 if (!GV->isConstant() && GV->hasLocalLinkage() && 1858 GV->hasInitializer()) 1859 Changed |= ProcessInternalGlobal(GV, GVI); 1860 } 1861 return Changed; 1862} 1863 1864/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1865/// initializers have an init priority of 65535. 1866GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1867 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1868 I != E; ++I) 1869 if (I->getName() == "llvm.global_ctors") { 1870 // Found it, verify it's an array of { int, void()* }. 1871 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1872 if (!ATy) return 0; 1873 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1874 if (!STy || STy->getNumElements() != 2 || 1875 STy->getElementType(0) != Type::Int32Ty) return 0; 1876 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1877 if (!PFTy) return 0; 1878 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1879 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() || 1880 FTy->getNumParams() != 0) 1881 return 0; 1882 1883 // Verify that the initializer is simple enough for us to handle. 1884 if (!I->hasInitializer()) return 0; 1885 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1886 if (!CA) return 0; 1887 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1888 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1889 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1890 continue; 1891 1892 // Must have a function or null ptr. 1893 if (!isa<Function>(CS->getOperand(1))) 1894 return 0; 1895 1896 // Init priority must be standard. 1897 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1898 if (!CI || CI->getZExtValue() != 65535) 1899 return 0; 1900 } else { 1901 return 0; 1902 } 1903 1904 return I; 1905 } 1906 return 0; 1907} 1908 1909/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1910/// return a list of the functions and null terminator as a vector. 1911static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1912 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1913 std::vector<Function*> Result; 1914 Result.reserve(CA->getNumOperands()); 1915 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1916 ConstantStruct *CS = cast<ConstantStruct>(*i); 1917 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1918 } 1919 return Result; 1920} 1921 1922/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1923/// specified array, returning the new global to use. 1924static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1925 const std::vector<Function*> &Ctors) { 1926 // If we made a change, reassemble the initializer list. 1927 std::vector<Constant*> CSVals; 1928 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535)); 1929 CSVals.push_back(0); 1930 1931 // Create the new init list. 1932 std::vector<Constant*> CAList; 1933 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1934 if (Ctors[i]) { 1935 CSVals[1] = Ctors[i]; 1936 } else { 1937 const Type *FTy = FunctionType::get(Type::VoidTy, 1938 std::vector<const Type*>(), false); 1939 const PointerType *PFTy = PointerType::getUnqual(FTy); 1940 CSVals[1] = Constant::getNullValue(PFTy); 1941 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647); 1942 } 1943 CAList.push_back(ConstantStruct::get(CSVals)); 1944 } 1945 1946 // Create the array initializer. 1947 const Type *StructTy = 1948 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 1949 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()), 1950 CAList); 1951 1952 // If we didn't change the number of elements, don't create a new GV. 1953 if (CA->getType() == GCL->getInitializer()->getType()) { 1954 GCL->setInitializer(CA); 1955 return GCL; 1956 } 1957 1958 // Create the new global and insert it next to the existing list. 1959 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 1960 GCL->getLinkage(), CA, "", 1961 (Module *)NULL, 1962 GCL->isThreadLocal()); 1963 GCL->getParent()->getGlobalList().insert(GCL, NGV); 1964 NGV->takeName(GCL); 1965 1966 // Nuke the old list, replacing any uses with the new one. 1967 if (!GCL->use_empty()) { 1968 Constant *V = NGV; 1969 if (V->getType() != GCL->getType()) 1970 V = ConstantExpr::getBitCast(V, GCL->getType()); 1971 GCL->replaceAllUsesWith(V); 1972 } 1973 GCL->eraseFromParent(); 1974 1975 if (Ctors.size()) 1976 return NGV; 1977 else 1978 return 0; 1979} 1980 1981 1982static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, 1983 Value *V) { 1984 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 1985 Constant *R = ComputedValues[V]; 1986 assert(R && "Reference to an uncomputed value!"); 1987 return R; 1988} 1989 1990/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 1991/// enough for us to understand. In particular, if it is a cast of something, 1992/// we punt. We basically just support direct accesses to globals and GEP's of 1993/// globals. This should be kept up to date with CommitValueTo. 1994static bool isSimpleEnoughPointerToCommit(Constant *C) { 1995 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) { 1996 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 1997 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 1998 return !GV->isDeclaration(); // reject external globals. 1999 } 2000 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2001 // Handle a constantexpr gep. 2002 if (CE->getOpcode() == Instruction::GetElementPtr && 2003 isa<GlobalVariable>(CE->getOperand(0))) { 2004 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2005 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage()) 2006 return false; // do not allow weak/linkonce/dllimport/dllexport linkage. 2007 return GV->hasInitializer() && 2008 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2009 } 2010 return false; 2011} 2012 2013/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 2014/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 2015/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 2016static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 2017 ConstantExpr *Addr, unsigned OpNo) { 2018 // Base case of the recursion. 2019 if (OpNo == Addr->getNumOperands()) { 2020 assert(Val->getType() == Init->getType() && "Type mismatch!"); 2021 return Val; 2022 } 2023 2024 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 2025 std::vector<Constant*> Elts; 2026 2027 // Break up the constant into its elements. 2028 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 2029 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 2030 Elts.push_back(cast<Constant>(*i)); 2031 } else if (isa<ConstantAggregateZero>(Init)) { 2032 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2033 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 2034 } else if (isa<UndefValue>(Init)) { 2035 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2036 Elts.push_back(UndefValue::get(STy->getElementType(i))); 2037 } else { 2038 assert(0 && "This code is out of sync with " 2039 " ConstantFoldLoadThroughGEPConstantExpr"); 2040 } 2041 2042 // Replace the element that we are supposed to. 2043 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 2044 unsigned Idx = CU->getZExtValue(); 2045 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 2046 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 2047 2048 // Return the modified struct. 2049 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked()); 2050 } else { 2051 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 2052 const ArrayType *ATy = cast<ArrayType>(Init->getType()); 2053 2054 // Break up the array into elements. 2055 std::vector<Constant*> Elts; 2056 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 2057 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 2058 Elts.push_back(cast<Constant>(*i)); 2059 } else if (isa<ConstantAggregateZero>(Init)) { 2060 Constant *Elt = Constant::getNullValue(ATy->getElementType()); 2061 Elts.assign(ATy->getNumElements(), Elt); 2062 } else if (isa<UndefValue>(Init)) { 2063 Constant *Elt = UndefValue::get(ATy->getElementType()); 2064 Elts.assign(ATy->getNumElements(), Elt); 2065 } else { 2066 assert(0 && "This code is out of sync with " 2067 " ConstantFoldLoadThroughGEPConstantExpr"); 2068 } 2069 2070 assert(CI->getZExtValue() < ATy->getNumElements()); 2071 Elts[CI->getZExtValue()] = 2072 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 2073 return ConstantArray::get(ATy, Elts); 2074 } 2075} 2076 2077/// CommitValueTo - We have decided that Addr (which satisfies the predicate 2078/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 2079static void CommitValueTo(Constant *Val, Constant *Addr) { 2080 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 2081 assert(GV->hasInitializer()); 2082 GV->setInitializer(Val); 2083 return; 2084 } 2085 2086 ConstantExpr *CE = cast<ConstantExpr>(Addr); 2087 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2088 2089 Constant *Init = GV->getInitializer(); 2090 Init = EvaluateStoreInto(Init, Val, CE, 2); 2091 GV->setInitializer(Init); 2092} 2093 2094/// ComputeLoadResult - Return the value that would be computed by a load from 2095/// P after the stores reflected by 'memory' have been performed. If we can't 2096/// decide, return null. 2097static Constant *ComputeLoadResult(Constant *P, 2098 const DenseMap<Constant*, Constant*> &Memory) { 2099 // If this memory location has been recently stored, use the stored value: it 2100 // is the most up-to-date. 2101 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P); 2102 if (I != Memory.end()) return I->second; 2103 2104 // Access it. 2105 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 2106 if (GV->hasInitializer()) 2107 return GV->getInitializer(); 2108 return 0; 2109 } 2110 2111 // Handle a constantexpr getelementptr. 2112 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 2113 if (CE->getOpcode() == Instruction::GetElementPtr && 2114 isa<GlobalVariable>(CE->getOperand(0))) { 2115 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2116 if (GV->hasInitializer()) 2117 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2118 } 2119 2120 return 0; // don't know how to evaluate. 2121} 2122 2123/// EvaluateFunction - Evaluate a call to function F, returning true if 2124/// successful, false if we can't evaluate it. ActualArgs contains the formal 2125/// arguments for the function. 2126static bool EvaluateFunction(Function *F, Constant *&RetVal, 2127 const std::vector<Constant*> &ActualArgs, 2128 std::vector<Function*> &CallStack, 2129 DenseMap<Constant*, Constant*> &MutatedMemory, 2130 std::vector<GlobalVariable*> &AllocaTmps) { 2131 // Check to see if this function is already executing (recursion). If so, 2132 // bail out. TODO: we might want to accept limited recursion. 2133 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 2134 return false; 2135 2136 CallStack.push_back(F); 2137 2138 /// Values - As we compute SSA register values, we store their contents here. 2139 DenseMap<Value*, Constant*> Values; 2140 2141 // Initialize arguments to the incoming values specified. 2142 unsigned ArgNo = 0; 2143 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 2144 ++AI, ++ArgNo) 2145 Values[AI] = ActualArgs[ArgNo]; 2146 2147 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 2148 /// we can only evaluate any one basic block at most once. This set keeps 2149 /// track of what we have executed so we can detect recursive cases etc. 2150 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 2151 2152 // CurInst - The current instruction we're evaluating. 2153 BasicBlock::iterator CurInst = F->begin()->begin(); 2154 2155 // This is the main evaluation loop. 2156 while (1) { 2157 Constant *InstResult = 0; 2158 2159 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2160 if (SI->isVolatile()) return false; // no volatile accesses. 2161 Constant *Ptr = getVal(Values, SI->getOperand(1)); 2162 if (!isSimpleEnoughPointerToCommit(Ptr)) 2163 // If this is too complex for us to commit, reject it. 2164 return false; 2165 Constant *Val = getVal(Values, SI->getOperand(0)); 2166 MutatedMemory[Ptr] = Val; 2167 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2168 InstResult = ConstantExpr::get(BO->getOpcode(), 2169 getVal(Values, BO->getOperand(0)), 2170 getVal(Values, BO->getOperand(1))); 2171 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2172 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2173 getVal(Values, CI->getOperand(0)), 2174 getVal(Values, CI->getOperand(1))); 2175 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2176 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2177 getVal(Values, CI->getOperand(0)), 2178 CI->getType()); 2179 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2180 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2181 getVal(Values, SI->getOperand(1)), 2182 getVal(Values, SI->getOperand(2))); 2183 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2184 Constant *P = getVal(Values, GEP->getOperand(0)); 2185 SmallVector<Constant*, 8> GEPOps; 2186 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2187 i != e; ++i) 2188 GEPOps.push_back(getVal(Values, *i)); 2189 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2190 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2191 if (LI->isVolatile()) return false; // no volatile accesses. 2192 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2193 MutatedMemory); 2194 if (InstResult == 0) return false; // Could not evaluate load. 2195 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2196 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2197 const Type *Ty = AI->getType()->getElementType(); 2198 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2199 GlobalValue::InternalLinkage, 2200 UndefValue::get(Ty), 2201 AI->getName())); 2202 InstResult = AllocaTmps.back(); 2203 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2204 2205 // Debug info can safely be ignored here. 2206 if (isa<DbgInfoIntrinsic>(CI)) { 2207 ++CurInst; 2208 continue; 2209 } 2210 2211 // Cannot handle inline asm. 2212 if (isa<InlineAsm>(CI->getOperand(0))) return false; 2213 2214 // Resolve function pointers. 2215 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0))); 2216 if (!Callee) return false; // Cannot resolve. 2217 2218 std::vector<Constant*> Formals; 2219 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end(); 2220 i != e; ++i) 2221 Formals.push_back(getVal(Values, *i)); 2222 2223 if (Callee->isDeclaration()) { 2224 // If this is a function we can constant fold, do it. 2225 if (Constant *C = ConstantFoldCall(Callee, &Formals[0], 2226 Formals.size())) { 2227 InstResult = C; 2228 } else { 2229 return false; 2230 } 2231 } else { 2232 if (Callee->getFunctionType()->isVarArg()) 2233 return false; 2234 2235 Constant *RetVal; 2236 // Execute the call, if successful, use the return value. 2237 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2238 MutatedMemory, AllocaTmps)) 2239 return false; 2240 InstResult = RetVal; 2241 } 2242 } else if (isa<TerminatorInst>(CurInst)) { 2243 BasicBlock *NewBB = 0; 2244 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2245 if (BI->isUnconditional()) { 2246 NewBB = BI->getSuccessor(0); 2247 } else { 2248 ConstantInt *Cond = 2249 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2250 if (!Cond) return false; // Cannot determine. 2251 2252 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2253 } 2254 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2255 ConstantInt *Val = 2256 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2257 if (!Val) return false; // Cannot determine. 2258 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2259 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2260 if (RI->getNumOperands()) 2261 RetVal = getVal(Values, RI->getOperand(0)); 2262 2263 CallStack.pop_back(); // return from fn. 2264 return true; // We succeeded at evaluating this ctor! 2265 } else { 2266 // invoke, unwind, unreachable. 2267 return false; // Cannot handle this terminator. 2268 } 2269 2270 // Okay, we succeeded in evaluating this control flow. See if we have 2271 // executed the new block before. If so, we have a looping function, 2272 // which we cannot evaluate in reasonable time. 2273 if (!ExecutedBlocks.insert(NewBB)) 2274 return false; // looped! 2275 2276 // Okay, we have never been in this block before. Check to see if there 2277 // are any PHI nodes. If so, evaluate them with information about where 2278 // we came from. 2279 BasicBlock *OldBB = CurInst->getParent(); 2280 CurInst = NewBB->begin(); 2281 PHINode *PN; 2282 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2283 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2284 2285 // Do NOT increment CurInst. We know that the terminator had no value. 2286 continue; 2287 } else { 2288 // Did not know how to evaluate this! 2289 return false; 2290 } 2291 2292 if (!CurInst->use_empty()) 2293 Values[CurInst] = InstResult; 2294 2295 // Advance program counter. 2296 ++CurInst; 2297 } 2298} 2299 2300/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2301/// we can. Return true if we can, false otherwise. 2302static bool EvaluateStaticConstructor(Function *F) { 2303 /// MutatedMemory - For each store we execute, we update this map. Loads 2304 /// check this to get the most up-to-date value. If evaluation is successful, 2305 /// this state is committed to the process. 2306 DenseMap<Constant*, Constant*> MutatedMemory; 2307 2308 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2309 /// to represent its body. This vector is needed so we can delete the 2310 /// temporary globals when we are done. 2311 std::vector<GlobalVariable*> AllocaTmps; 2312 2313 /// CallStack - This is used to detect recursion. In pathological situations 2314 /// we could hit exponential behavior, but at least there is nothing 2315 /// unbounded. 2316 std::vector<Function*> CallStack; 2317 2318 // Call the function. 2319 Constant *RetValDummy; 2320 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(), 2321 CallStack, MutatedMemory, AllocaTmps); 2322 if (EvalSuccess) { 2323 // We succeeded at evaluation: commit the result. 2324 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2325 << F->getName() << "' to " << MutatedMemory.size() 2326 << " stores.\n"; 2327 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2328 E = MutatedMemory.end(); I != E; ++I) 2329 CommitValueTo(I->second, I->first); 2330 } 2331 2332 // At this point, we are done interpreting. If we created any 'alloca' 2333 // temporaries, release them now. 2334 while (!AllocaTmps.empty()) { 2335 GlobalVariable *Tmp = AllocaTmps.back(); 2336 AllocaTmps.pop_back(); 2337 2338 // If there are still users of the alloca, the program is doing something 2339 // silly, e.g. storing the address of the alloca somewhere and using it 2340 // later. Since this is undefined, we'll just make it be null. 2341 if (!Tmp->use_empty()) 2342 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2343 delete Tmp; 2344 } 2345 2346 return EvalSuccess; 2347} 2348 2349 2350 2351/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2352/// Return true if anything changed. 2353bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2354 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2355 bool MadeChange = false; 2356 if (Ctors.empty()) return false; 2357 2358 // Loop over global ctors, optimizing them when we can. 2359 for (unsigned i = 0; i != Ctors.size(); ++i) { 2360 Function *F = Ctors[i]; 2361 // Found a null terminator in the middle of the list, prune off the rest of 2362 // the list. 2363 if (F == 0) { 2364 if (i != Ctors.size()-1) { 2365 Ctors.resize(i+1); 2366 MadeChange = true; 2367 } 2368 break; 2369 } 2370 2371 // We cannot simplify external ctor functions. 2372 if (F->empty()) continue; 2373 2374 // If we can evaluate the ctor at compile time, do. 2375 if (EvaluateStaticConstructor(F)) { 2376 Ctors.erase(Ctors.begin()+i); 2377 MadeChange = true; 2378 --i; 2379 ++NumCtorsEvaluated; 2380 continue; 2381 } 2382 } 2383 2384 if (!MadeChange) return false; 2385 2386 GCL = InstallGlobalCtors(GCL, Ctors); 2387 return true; 2388} 2389 2390bool GlobalOpt::OptimizeGlobalAliases(Module &M) { 2391 bool Changed = false; 2392 2393 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 2394 I != E;) { 2395 Module::alias_iterator J = I++; 2396 // Aliases without names cannot be referenced outside this module. 2397 if (!J->hasName() && !J->isDeclaration()) 2398 J->setLinkage(GlobalValue::InternalLinkage); 2399 // If the aliasee may change at link time, nothing can be done - bail out. 2400 if (J->mayBeOverridden()) 2401 continue; 2402 2403 Constant *Aliasee = J->getAliasee(); 2404 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts()); 2405 Target->removeDeadConstantUsers(); 2406 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse(); 2407 2408 // Make all users of the alias use the aliasee instead. 2409 if (!J->use_empty()) { 2410 J->replaceAllUsesWith(Aliasee); 2411 ++NumAliasesResolved; 2412 Changed = true; 2413 } 2414 2415 // If the aliasee has internal linkage, give it the name and linkage 2416 // of the alias, and delete the alias. This turns: 2417 // define internal ... @f(...) 2418 // @a = alias ... @f 2419 // into: 2420 // define ... @a(...) 2421 if (!Target->hasLocalLinkage()) 2422 continue; 2423 2424 // The transform is only useful if the alias does not have internal linkage. 2425 if (J->hasLocalLinkage()) 2426 continue; 2427 2428 // Do not perform the transform if multiple aliases potentially target the 2429 // aliasee. This check also ensures that it is safe to replace the section 2430 // and other attributes of the aliasee with those of the alias. 2431 if (!hasOneUse) 2432 continue; 2433 2434 // Give the aliasee the name, linkage and other attributes of the alias. 2435 Target->takeName(J); 2436 Target->setLinkage(J->getLinkage()); 2437 Target->GlobalValue::copyAttributesFrom(J); 2438 2439 // Delete the alias. 2440 M.getAliasList().erase(J); 2441 ++NumAliasesRemoved; 2442 Changed = true; 2443 } 2444 2445 return Changed; 2446} 2447 2448bool GlobalOpt::runOnModule(Module &M) { 2449 bool Changed = false; 2450 2451 // Try to find the llvm.globalctors list. 2452 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2453 2454 bool LocalChange = true; 2455 while (LocalChange) { 2456 LocalChange = false; 2457 2458 // Delete functions that are trivially dead, ccc -> fastcc 2459 LocalChange |= OptimizeFunctions(M); 2460 2461 // Optimize global_ctors list. 2462 if (GlobalCtors) 2463 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2464 2465 // Optimize non-address-taken globals. 2466 LocalChange |= OptimizeGlobalVars(M); 2467 2468 // Resolve aliases, when possible. 2469 LocalChange |= OptimizeGlobalAliases(M); 2470 Changed |= LocalChange; 2471 } 2472 2473 // TODO: Move all global ctors functions to the end of the module for code 2474 // layout. 2475 2476 return Changed; 2477} 2478