GlobalOpt.cpp revision 17311da4af2de2788d3b1e2f4cde1d97f92bfef4
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/Analysis/MemoryBuiltins.h" 27#include "llvm/Target/TargetData.h" 28#include "llvm/Support/CallSite.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/Support/ErrorHandling.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/MathExtras.h" 33#include "llvm/Support/raw_ostream.h" 34#include "llvm/ADT/DenseMap.h" 35#include "llvm/ADT/SmallPtrSet.h" 36#include "llvm/ADT/SmallVector.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/ADT/STLExtras.h" 39#include <algorithm> 40using namespace llvm; 41 42STATISTIC(NumMarked , "Number of globals marked constant"); 43STATISTIC(NumSRA , "Number of aggregate globals broken into scalars"); 44STATISTIC(NumHeapSRA , "Number of heap objects SRA'd"); 45STATISTIC(NumSubstitute,"Number of globals with initializers stored into them"); 46STATISTIC(NumDeleted , "Number of globals deleted"); 47STATISTIC(NumFnDeleted , "Number of functions deleted"); 48STATISTIC(NumGlobUses , "Number of global uses devirtualized"); 49STATISTIC(NumLocalized , "Number of globals localized"); 50STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans"); 51STATISTIC(NumFastCallFns , "Number of functions converted to fastcc"); 52STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated"); 53STATISTIC(NumNestRemoved , "Number of nest attributes removed"); 54STATISTIC(NumAliasesResolved, "Number of global aliases resolved"); 55STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated"); 56 57namespace { 58 struct GlobalOpt : public ModulePass { 59 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 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 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 const 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// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used 140// by constants itself. Note that constants cannot be cyclic, so this test is 141// pretty easy to implement recursively. 142// 143static bool SafeToDestroyConstant(const Constant *C) { 144 if (isa<GlobalValue>(C)) return false; 145 146 for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; 147 ++UI) 148 if (const Constant *CU = dyn_cast<Constant>(*UI)) { 149 if (!SafeToDestroyConstant(CU)) return false; 150 } else 151 return false; 152 return true; 153} 154 155 156/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus 157/// structure. If the global has its address taken, return true to indicate we 158/// can't do anything with it. 159/// 160static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, 161 SmallPtrSet<const PHINode*, 16> &PHIUsers) { 162 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; 163 ++UI) { 164 const User *U = *UI; 165 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 166 GS.HasNonInstructionUser = true; 167 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; 168 } else if (const Instruction *I = dyn_cast<Instruction>(U)) { 169 if (!GS.HasMultipleAccessingFunctions) { 170 const Function *F = I->getParent()->getParent(); 171 if (GS.AccessingFunction == 0) 172 GS.AccessingFunction = F; 173 else if (GS.AccessingFunction != F) 174 GS.HasMultipleAccessingFunctions = true; 175 } 176 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) { 177 GS.isLoaded = true; 178 if (LI->isVolatile()) return true; // Don't hack on volatile loads. 179 } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) { 180 // Don't allow a store OF the address, only stores TO the address. 181 if (SI->getOperand(0) == V) return true; 182 183 if (SI->isVolatile()) return true; // Don't hack on volatile stores. 184 185 // If this is a direct store to the global (i.e., the global is a scalar 186 // value, not an aggregate), keep more specific information about 187 // stores. 188 if (GS.StoredType != GlobalStatus::isStored) { 189 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>( 190 SI->getOperand(1))) { 191 Value *StoredVal = SI->getOperand(0); 192 if (StoredVal == GV->getInitializer()) { 193 if (GS.StoredType < GlobalStatus::isInitializerStored) 194 GS.StoredType = GlobalStatus::isInitializerStored; 195 } else if (isa<LoadInst>(StoredVal) && 196 cast<LoadInst>(StoredVal)->getOperand(0) == GV) { 197 if (GS.StoredType < GlobalStatus::isInitializerStored) 198 GS.StoredType = GlobalStatus::isInitializerStored; 199 } else if (GS.StoredType < GlobalStatus::isStoredOnce) { 200 GS.StoredType = GlobalStatus::isStoredOnce; 201 GS.StoredOnceValue = StoredVal; 202 } else if (GS.StoredType == GlobalStatus::isStoredOnce && 203 GS.StoredOnceValue == StoredVal) { 204 // noop. 205 } else { 206 GS.StoredType = GlobalStatus::isStored; 207 } 208 } else { 209 GS.StoredType = GlobalStatus::isStored; 210 } 211 } 212 } else if (isa<GetElementPtrInst>(I)) { 213 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 214 } else if (isa<SelectInst>(I)) { 215 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 216 } else if (const PHINode *PN = dyn_cast<PHINode>(I)) { 217 // PHI nodes we can check just like select or GEP instructions, but we 218 // have to be careful about infinite recursion. 219 if (PHIUsers.insert(PN)) // Not already visited. 220 if (AnalyzeGlobal(I, GS, PHIUsers)) return true; 221 GS.HasPHIUser = true; 222 } else if (isa<CmpInst>(I)) { 223 // Nothing to analyse. 224 } else if (isa<MemTransferInst>(I)) { 225 const MemTransferInst *MTI = cast<MemTransferInst>(I); 226 if (MTI->getArgOperand(0) == V) 227 GS.StoredType = GlobalStatus::isStored; 228 if (MTI->getArgOperand(1) == V) 229 GS.isLoaded = true; 230 } else if (isa<MemSetInst>(I)) { 231 assert(cast<MemSetInst>(I)->getArgOperand(0) == V && 232 "Memset only takes one pointer!"); 233 GS.StoredType = GlobalStatus::isStored; 234 } else { 235 return true; // Any other non-load instruction might take address! 236 } 237 } else if (const Constant *C = dyn_cast<Constant>(U)) { 238 GS.HasNonInstructionUser = true; 239 // We might have a dead and dangling constant hanging off of here. 240 if (!SafeToDestroyConstant(C)) 241 return true; 242 } else { 243 GS.HasNonInstructionUser = true; 244 // Otherwise must be some other user. 245 return true; 246 } 247 } 248 249 return false; 250} 251 252static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { 253 ConstantInt *CI = dyn_cast<ConstantInt>(Idx); 254 if (!CI) return 0; 255 unsigned IdxV = CI->getZExtValue(); 256 257 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) { 258 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); 259 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) { 260 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); 261 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) { 262 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); 263 } else if (isa<ConstantAggregateZero>(Agg)) { 264 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 265 if (IdxV < STy->getNumElements()) 266 return Constant::getNullValue(STy->getElementType(IdxV)); 267 } else if (const SequentialType *STy = 268 dyn_cast<SequentialType>(Agg->getType())) { 269 return Constant::getNullValue(STy->getElementType()); 270 } 271 } else if (isa<UndefValue>(Agg)) { 272 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) { 273 if (IdxV < STy->getNumElements()) 274 return UndefValue::get(STy->getElementType(IdxV)); 275 } else if (const SequentialType *STy = 276 dyn_cast<SequentialType>(Agg->getType())) { 277 return UndefValue::get(STy->getElementType()); 278 } 279 } 280 return 0; 281} 282 283 284/// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all 285/// users of the global, cleaning up the obvious ones. This is largely just a 286/// quick scan over the use list to clean up the easy and obvious cruft. This 287/// returns true if it made a change. 288static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { 289 bool Changed = false; 290 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { 291 User *U = *UI++; 292 293 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 294 if (Init) { 295 // Replace the load with the initializer. 296 LI->replaceAllUsesWith(Init); 297 LI->eraseFromParent(); 298 Changed = true; 299 } 300 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 301 // Store must be unreachable or storing Init into the global. 302 SI->eraseFromParent(); 303 Changed = true; 304 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) { 305 if (CE->getOpcode() == Instruction::GetElementPtr) { 306 Constant *SubInit = 0; 307 if (Init) 308 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 309 Changed |= CleanupConstantGlobalUsers(CE, SubInit); 310 } else if (CE->getOpcode() == Instruction::BitCast && 311 CE->getType()->isPointerTy()) { 312 // Pointer cast, delete any stores and memsets to the global. 313 Changed |= CleanupConstantGlobalUsers(CE, 0); 314 } 315 316 if (CE->use_empty()) { 317 CE->destroyConstant(); 318 Changed = true; 319 } 320 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 321 // Do not transform "gepinst (gep constexpr (GV))" here, because forming 322 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold 323 // and will invalidate our notion of what Init is. 324 Constant *SubInit = 0; 325 if (!isa<ConstantExpr>(GEP->getOperand(0))) { 326 ConstantExpr *CE = 327 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP)); 328 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) 329 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); 330 } 331 Changed |= CleanupConstantGlobalUsers(GEP, SubInit); 332 333 if (GEP->use_empty()) { 334 GEP->eraseFromParent(); 335 Changed = true; 336 } 337 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv 338 if (MI->getRawDest() == V) { 339 MI->eraseFromParent(); 340 Changed = true; 341 } 342 343 } else if (Constant *C = dyn_cast<Constant>(U)) { 344 // If we have a chain of dead constantexprs or other things dangling from 345 // us, and if they are all dead, nuke them without remorse. 346 if (SafeToDestroyConstant(C)) { 347 C->destroyConstant(); 348 // This could have invalidated UI, start over from scratch. 349 CleanupConstantGlobalUsers(V, Init); 350 return true; 351 } 352 } 353 } 354 return Changed; 355} 356 357/// isSafeSROAElementUse - Return true if the specified instruction is a safe 358/// user of a derived expression from a global that we want to SROA. 359static bool isSafeSROAElementUse(Value *V) { 360 // We might have a dead and dangling constant hanging off of here. 361 if (Constant *C = dyn_cast<Constant>(V)) 362 return SafeToDestroyConstant(C); 363 364 Instruction *I = dyn_cast<Instruction>(V); 365 if (!I) return false; 366 367 // Loads are ok. 368 if (isa<LoadInst>(I)) return true; 369 370 // Stores *to* the pointer are ok. 371 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 372 return SI->getOperand(0) != V; 373 374 // Otherwise, it must be a GEP. 375 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I); 376 if (GEPI == 0) return false; 377 378 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) || 379 !cast<Constant>(GEPI->getOperand(1))->isNullValue()) 380 return false; 381 382 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end(); 383 I != E; ++I) 384 if (!isSafeSROAElementUse(*I)) 385 return false; 386 return true; 387} 388 389 390/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value. 391/// Look at it and its uses and decide whether it is safe to SROA this global. 392/// 393static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) { 394 // The user of the global must be a GEP Inst or a ConstantExpr GEP. 395 if (!isa<GetElementPtrInst>(U) && 396 (!isa<ConstantExpr>(U) || 397 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr)) 398 return false; 399 400 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we 401 // don't like < 3 operand CE's, and we don't like non-constant integer 402 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some 403 // value of C. 404 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) || 405 !cast<Constant>(U->getOperand(1))->isNullValue() || 406 !isa<ConstantInt>(U->getOperand(2))) 407 return false; 408 409 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U); 410 ++GEPI; // Skip over the pointer index. 411 412 // If this is a use of an array allocation, do a bit more checking for sanity. 413 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) { 414 uint64_t NumElements = AT->getNumElements(); 415 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2)); 416 417 // Check to make sure that index falls within the array. If not, 418 // something funny is going on, so we won't do the optimization. 419 // 420 if (Idx->getZExtValue() >= NumElements) 421 return false; 422 423 // We cannot scalar repl this level of the array unless any array 424 // sub-indices are in-range constants. In particular, consider: 425 // A[0][i]. We cannot know that the user isn't doing invalid things like 426 // allowing i to index an out-of-range subscript that accesses A[1]. 427 // 428 // Scalar replacing *just* the outer index of the array is probably not 429 // going to be a win anyway, so just give up. 430 for (++GEPI; // Skip array index. 431 GEPI != E; 432 ++GEPI) { 433 uint64_t NumElements; 434 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI)) 435 NumElements = SubArrayTy->getNumElements(); 436 else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI)) 437 NumElements = SubVectorTy->getNumElements(); 438 else { 439 assert((*GEPI)->isStructTy() && 440 "Indexed GEP type is not array, vector, or struct!"); 441 continue; 442 } 443 444 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand()); 445 if (!IdxVal || IdxVal->getZExtValue() >= NumElements) 446 return false; 447 } 448 } 449 450 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) 451 if (!isSafeSROAElementUse(*I)) 452 return false; 453 return true; 454} 455 456/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it 457/// is safe for us to perform this transformation. 458/// 459static bool GlobalUsersSafeToSRA(GlobalValue *GV) { 460 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 461 UI != E; ++UI) { 462 if (!IsUserOfGlobalSafeForSRA(*UI, GV)) 463 return false; 464 } 465 return true; 466} 467 468 469/// SRAGlobal - Perform scalar replacement of aggregates on the specified global 470/// variable. This opens the door for other optimizations by exposing the 471/// behavior of the program in a more fine-grained way. We have determined that 472/// this transformation is safe already. We return the first global variable we 473/// insert so that the caller can reprocess it. 474static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { 475 // Make sure this global only has simple uses that we can SRA. 476 if (!GlobalUsersSafeToSRA(GV)) 477 return 0; 478 479 assert(GV->hasLocalLinkage() && !GV->isConstant()); 480 Constant *Init = GV->getInitializer(); 481 const Type *Ty = Init->getType(); 482 483 std::vector<GlobalVariable*> NewGlobals; 484 Module::GlobalListType &Globals = GV->getParent()->getGlobalList(); 485 486 // Get the alignment of the global, either explicit or target-specific. 487 unsigned StartAlignment = GV->getAlignment(); 488 if (StartAlignment == 0) 489 StartAlignment = TD.getABITypeAlignment(GV->getType()); 490 491 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 492 NewGlobals.reserve(STy->getNumElements()); 493 const StructLayout &Layout = *TD.getStructLayout(STy); 494 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 495 Constant *In = getAggregateConstantElement(Init, 496 ConstantInt::get(Type::getInt32Ty(STy->getContext()), i)); 497 assert(In && "Couldn't get element of initializer?"); 498 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, 499 GlobalVariable::InternalLinkage, 500 In, GV->getName()+"."+Twine(i), 501 GV->isThreadLocal(), 502 GV->getType()->getAddressSpace()); 503 Globals.insert(GV, NGV); 504 NewGlobals.push_back(NGV); 505 506 // Calculate the known alignment of the field. If the original aggregate 507 // had 256 byte alignment for example, something might depend on that: 508 // propagate info to each field. 509 uint64_t FieldOffset = Layout.getElementOffset(i); 510 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset); 511 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i))) 512 NGV->setAlignment(NewAlign); 513 } 514 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) { 515 unsigned NumElements = 0; 516 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 517 NumElements = ATy->getNumElements(); 518 else 519 NumElements = cast<VectorType>(STy)->getNumElements(); 520 521 if (NumElements > 16 && GV->hasNUsesOrMore(16)) 522 return 0; // It's not worth it. 523 NewGlobals.reserve(NumElements); 524 525 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType()); 526 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); 527 for (unsigned i = 0, e = NumElements; i != e; ++i) { 528 Constant *In = getAggregateConstantElement(Init, 529 ConstantInt::get(Type::getInt32Ty(Init->getContext()), i)); 530 assert(In && "Couldn't get element of initializer?"); 531 532 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, 533 GlobalVariable::InternalLinkage, 534 In, GV->getName()+"."+Twine(i), 535 GV->isThreadLocal(), 536 GV->getType()->getAddressSpace()); 537 Globals.insert(GV, NGV); 538 NewGlobals.push_back(NGV); 539 540 // Calculate the known alignment of the field. If the original aggregate 541 // had 256 byte alignment for example, something might depend on that: 542 // propagate info to each field. 543 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i); 544 if (NewAlign > EltAlign) 545 NGV->setAlignment(NewAlign); 546 } 547 } 548 549 if (NewGlobals.empty()) 550 return 0; 551 552 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV); 553 554 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext())); 555 556 // Loop over all of the uses of the global, replacing the constantexpr geps, 557 // with smaller constantexpr geps or direct references. 558 while (!GV->use_empty()) { 559 User *GEP = GV->use_back(); 560 assert(((isa<ConstantExpr>(GEP) && 561 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)|| 562 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!"); 563 564 // Ignore the 1th operand, which has to be zero or else the program is quite 565 // broken (undefined). Get the 2nd operand, which is the structure or array 566 // index. 567 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 568 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access. 569 570 Value *NewPtr = NewGlobals[Val]; 571 572 // Form a shorter GEP if needed. 573 if (GEP->getNumOperands() > 3) { 574 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) { 575 SmallVector<Constant*, 8> Idxs; 576 Idxs.push_back(NullInt); 577 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i) 578 Idxs.push_back(CE->getOperand(i)); 579 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), 580 &Idxs[0], Idxs.size()); 581 } else { 582 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP); 583 SmallVector<Value*, 8> Idxs; 584 Idxs.push_back(NullInt); 585 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) 586 Idxs.push_back(GEPI->getOperand(i)); 587 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(), 588 GEPI->getName()+"."+Twine(Val),GEPI); 589 } 590 } 591 GEP->replaceAllUsesWith(NewPtr); 592 593 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP)) 594 GEPI->eraseFromParent(); 595 else 596 cast<ConstantExpr>(GEP)->destroyConstant(); 597 } 598 599 // Delete the old global, now that it is dead. 600 Globals.erase(GV); 601 ++NumSRA; 602 603 // Loop over the new globals array deleting any globals that are obviously 604 // dead. This can arise due to scalarization of a structure or an array that 605 // has elements that are dead. 606 unsigned FirstGlobal = 0; 607 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i) 608 if (NewGlobals[i]->use_empty()) { 609 Globals.erase(NewGlobals[i]); 610 if (FirstGlobal == i) ++FirstGlobal; 611 } 612 613 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0; 614} 615 616/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified 617/// value will trap if the value is dynamically null. PHIs keeps track of any 618/// phi nodes we've seen to avoid reprocessing them. 619static bool AllUsesOfValueWillTrapIfNull(const Value *V, 620 SmallPtrSet<const PHINode*, 8> &PHIs) { 621 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; 622 ++UI) { 623 const User *U = *UI; 624 625 if (isa<LoadInst>(U)) { 626 // Will trap. 627 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) { 628 if (SI->getOperand(0) == V) { 629 //cerr << "NONTRAPPING USE: " << *U; 630 return false; // Storing the value. 631 } 632 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) { 633 if (CI->getCalledValue() != V) { 634 //cerr << "NONTRAPPING USE: " << *U; 635 return false; // Not calling the ptr 636 } 637 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) { 638 if (II->getCalledValue() != V) { 639 //cerr << "NONTRAPPING USE: " << *U; 640 return false; // Not calling the ptr 641 } 642 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) { 643 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false; 644 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 645 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false; 646 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) { 647 // If we've already seen this phi node, ignore it, it has already been 648 // checked. 649 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs)) 650 return false; 651 } else if (isa<ICmpInst>(U) && 652 isa<ConstantPointerNull>(UI->getOperand(1))) { 653 // Ignore icmp X, null 654 } else { 655 //cerr << "NONTRAPPING USE: " << *U; 656 return false; 657 } 658 } 659 return true; 660} 661 662/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads 663/// from GV will trap if the loaded value is null. Note that this also permits 664/// comparisons of the loaded value against null, as a special case. 665static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) { 666 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end(); 667 UI != E; ++UI) { 668 const User *U = *UI; 669 670 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 671 SmallPtrSet<const PHINode*, 8> PHIs; 672 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs)) 673 return false; 674 } else if (isa<StoreInst>(U)) { 675 // Ignore stores to the global. 676 } else { 677 // We don't know or understand this user, bail out. 678 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U; 679 return false; 680 } 681 } 682 return true; 683} 684 685static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { 686 bool Changed = false; 687 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) { 688 Instruction *I = cast<Instruction>(*UI++); 689 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 690 LI->setOperand(0, NewV); 691 Changed = true; 692 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 693 if (SI->getOperand(1) == V) { 694 SI->setOperand(1, NewV); 695 Changed = true; 696 } 697 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 698 CallSite CS(I); 699 if (CS.getCalledValue() == V) { 700 // Calling through the pointer! Turn into a direct call, but be careful 701 // that the pointer is not also being passed as an argument. 702 CS.setCalledFunction(NewV); 703 Changed = true; 704 bool PassedAsArg = false; 705 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 706 if (CS.getArgument(i) == V) { 707 PassedAsArg = true; 708 CS.setArgument(i, NewV); 709 } 710 711 if (PassedAsArg) { 712 // Being passed as an argument also. Be careful to not invalidate UI! 713 UI = V->use_begin(); 714 } 715 } 716 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 717 Changed |= OptimizeAwayTrappingUsesOfValue(CI, 718 ConstantExpr::getCast(CI->getOpcode(), 719 NewV, CI->getType())); 720 if (CI->use_empty()) { 721 Changed = true; 722 CI->eraseFromParent(); 723 } 724 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 725 // Should handle GEP here. 726 SmallVector<Constant*, 8> Idxs; 727 Idxs.reserve(GEPI->getNumOperands()-1); 728 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); 729 i != e; ++i) 730 if (Constant *C = dyn_cast<Constant>(*i)) 731 Idxs.push_back(C); 732 else 733 break; 734 if (Idxs.size() == GEPI->getNumOperands()-1) 735 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI, 736 ConstantExpr::getGetElementPtr(NewV, &Idxs[0], 737 Idxs.size())); 738 if (GEPI->use_empty()) { 739 Changed = true; 740 GEPI->eraseFromParent(); 741 } 742 } 743 } 744 745 return Changed; 746} 747 748 749/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null 750/// value stored into it. If there are uses of the loaded value that would trap 751/// if the loaded value is dynamically null, then we know that they cannot be 752/// reachable with a null optimize away the load. 753static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { 754 bool Changed = false; 755 756 // Keep track of whether we are able to remove all the uses of the global 757 // other than the store that defines it. 758 bool AllNonStoreUsesGone = true; 759 760 // Replace all uses of loads with uses of uses of the stored value. 761 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){ 762 User *GlobalUser = *GUI++; 763 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) { 764 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV); 765 // If we were able to delete all uses of the loads 766 if (LI->use_empty()) { 767 LI->eraseFromParent(); 768 Changed = true; 769 } else { 770 AllNonStoreUsesGone = false; 771 } 772 } else if (isa<StoreInst>(GlobalUser)) { 773 // Ignore the store that stores "LV" to the global. 774 assert(GlobalUser->getOperand(1) == GV && 775 "Must be storing *to* the global"); 776 } else { 777 AllNonStoreUsesGone = false; 778 779 // If we get here we could have other crazy uses that are transitively 780 // loaded. 781 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || 782 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!"); 783 } 784 } 785 786 if (Changed) { 787 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV); 788 ++NumGlobUses; 789 } 790 791 // If we nuked all of the loads, then none of the stores are needed either, 792 // nor is the global. 793 if (AllNonStoreUsesGone) { 794 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n"); 795 CleanupConstantGlobalUsers(GV, 0); 796 if (GV->use_empty()) { 797 GV->eraseFromParent(); 798 ++NumDeleted; 799 } 800 Changed = true; 801 } 802 return Changed; 803} 804 805/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the 806/// instructions that are foldable. 807static void ConstantPropUsersOf(Value *V) { 808 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) 809 if (Instruction *I = dyn_cast<Instruction>(*UI++)) 810 if (Constant *NewC = ConstantFoldInstruction(I)) { 811 I->replaceAllUsesWith(NewC); 812 813 // Advance UI to the next non-I use to avoid invalidating it! 814 // Instructions could multiply use V. 815 while (UI != E && *UI == I) 816 ++UI; 817 I->eraseFromParent(); 818 } 819} 820 821/// OptimizeGlobalAddressOfMalloc - This function takes the specified global 822/// variable, and transforms the program as if it always contained the result of 823/// the specified malloc. Because it is always the result of the specified 824/// malloc, there is no reason to actually DO the malloc. Instead, turn the 825/// malloc into a global, and any loads of GV as uses of the new global. 826static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, 827 CallInst *CI, 828 const Type *AllocTy, 829 ConstantInt *NElements, 830 TargetData* TD) { 831 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n'); 832 833 const Type *GlobalType; 834 if (NElements->getZExtValue() == 1) 835 GlobalType = AllocTy; 836 else 837 // If we have an array allocation, the global variable is of an array. 838 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue()); 839 840 // Create the new global variable. The contents of the malloc'd memory is 841 // undefined, so initialize with an undef value. 842 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(), 843 GlobalType, false, 844 GlobalValue::InternalLinkage, 845 UndefValue::get(GlobalType), 846 GV->getName()+".body", 847 GV, 848 GV->isThreadLocal()); 849 850 // If there are bitcast users of the malloc (which is typical, usually we have 851 // a malloc + bitcast) then replace them with uses of the new global. Update 852 // other users to use the global as well. 853 BitCastInst *TheBC = 0; 854 while (!CI->use_empty()) { 855 Instruction *User = cast<Instruction>(CI->use_back()); 856 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 857 if (BCI->getType() == NewGV->getType()) { 858 BCI->replaceAllUsesWith(NewGV); 859 BCI->eraseFromParent(); 860 } else { 861 BCI->setOperand(0, NewGV); 862 } 863 } else { 864 if (TheBC == 0) 865 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI); 866 User->replaceUsesOfWith(CI, TheBC); 867 } 868 } 869 870 Constant *RepValue = NewGV; 871 if (NewGV->getType() != GV->getType()->getElementType()) 872 RepValue = ConstantExpr::getBitCast(RepValue, 873 GV->getType()->getElementType()); 874 875 // If there is a comparison against null, we will insert a global bool to 876 // keep track of whether the global was initialized yet or not. 877 GlobalVariable *InitBool = 878 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false, 879 GlobalValue::InternalLinkage, 880 ConstantInt::getFalse(GV->getContext()), 881 GV->getName()+".init", GV->isThreadLocal()); 882 bool InitBoolUsed = false; 883 884 // Loop over all uses of GV, processing them in turn. 885 while (!GV->use_empty()) { 886 if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) { 887 // The global is initialized when the store to it occurs. 888 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI); 889 SI->eraseFromParent(); 890 continue; 891 } 892 893 LoadInst *LI = cast<LoadInst>(GV->use_back()); 894 while (!LI->use_empty()) { 895 Use &LoadUse = LI->use_begin().getUse(); 896 if (!isa<ICmpInst>(LoadUse.getUser())) { 897 LoadUse = RepValue; 898 continue; 899 } 900 901 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser()); 902 // Replace the cmp X, 0 with a use of the bool value. 903 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI); 904 InitBoolUsed = true; 905 switch (ICI->getPredicate()) { 906 default: llvm_unreachable("Unknown ICmp Predicate!"); 907 case ICmpInst::ICMP_ULT: 908 case ICmpInst::ICMP_SLT: // X < null -> always false 909 LV = ConstantInt::getFalse(GV->getContext()); 910 break; 911 case ICmpInst::ICMP_ULE: 912 case ICmpInst::ICMP_SLE: 913 case ICmpInst::ICMP_EQ: 914 LV = BinaryOperator::CreateNot(LV, "notinit", ICI); 915 break; 916 case ICmpInst::ICMP_NE: 917 case ICmpInst::ICMP_UGE: 918 case ICmpInst::ICMP_SGE: 919 case ICmpInst::ICMP_UGT: 920 case ICmpInst::ICMP_SGT: 921 break; // no change. 922 } 923 ICI->replaceAllUsesWith(LV); 924 ICI->eraseFromParent(); 925 } 926 LI->eraseFromParent(); 927 } 928 929 // If the initialization boolean was used, insert it, otherwise delete it. 930 if (!InitBoolUsed) { 931 while (!InitBool->use_empty()) // Delete initializations 932 cast<StoreInst>(InitBool->use_back())->eraseFromParent(); 933 delete InitBool; 934 } else 935 GV->getParent()->getGlobalList().insert(GV, InitBool); 936 937 // Now the GV is dead, nuke it and the malloc.. 938 GV->eraseFromParent(); 939 CI->eraseFromParent(); 940 941 // To further other optimizations, loop over all users of NewGV and try to 942 // constant prop them. This will promote GEP instructions with constant 943 // indices into GEP constant-exprs, which will allow global-opt to hack on it. 944 ConstantPropUsersOf(NewGV); 945 if (RepValue != NewGV) 946 ConstantPropUsersOf(RepValue); 947 948 return NewGV; 949} 950 951/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking 952/// to make sure that there are no complex uses of V. We permit simple things 953/// like dereferencing the pointer, but not storing through the address, unless 954/// it is to the specified global. 955static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V, 956 const GlobalVariable *GV, 957 SmallPtrSet<const PHINode*, 8> &PHIs) { 958 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); 959 UI != E; ++UI) { 960 const Instruction *Inst = cast<Instruction>(*UI); 961 962 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { 963 continue; // Fine, ignore. 964 } 965 966 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 967 if (SI->getOperand(0) == V && SI->getOperand(1) != GV) 968 return false; // Storing the pointer itself... bad. 969 continue; // Otherwise, storing through it, or storing into GV... fine. 970 } 971 972 // Must index into the array and into the struct. 973 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) { 974 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs)) 975 return false; 976 continue; 977 } 978 979 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) { 980 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI 981 // cycles. 982 if (PHIs.insert(PN)) 983 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs)) 984 return false; 985 continue; 986 } 987 988 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { 989 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) 990 return false; 991 continue; 992 } 993 994 return false; 995 } 996 return true; 997} 998 999/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV 1000/// somewhere. Transform all uses of the allocation into loads from the 1001/// global and uses of the resultant pointer. Further, delete the store into 1002/// GV. This assumes that these value pass the 1003/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate. 1004static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc, 1005 GlobalVariable *GV) { 1006 while (!Alloc->use_empty()) { 1007 Instruction *U = cast<Instruction>(*Alloc->use_begin()); 1008 Instruction *InsertPt = U; 1009 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1010 // If this is the store of the allocation into the global, remove it. 1011 if (SI->getOperand(1) == GV) { 1012 SI->eraseFromParent(); 1013 continue; 1014 } 1015 } else if (PHINode *PN = dyn_cast<PHINode>(U)) { 1016 // Insert the load in the corresponding predecessor, not right before the 1017 // PHI. 1018 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator(); 1019 } else if (isa<BitCastInst>(U)) { 1020 // Must be bitcast between the malloc and store to initialize the global. 1021 ReplaceUsesOfMallocWithGlobal(U, GV); 1022 U->eraseFromParent(); 1023 continue; 1024 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 1025 // If this is a "GEP bitcast" and the user is a store to the global, then 1026 // just process it as a bitcast. 1027 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse()) 1028 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back())) 1029 if (SI->getOperand(1) == GV) { 1030 // Must be bitcast GEP between the malloc and store to initialize 1031 // the global. 1032 ReplaceUsesOfMallocWithGlobal(GEPI, GV); 1033 GEPI->eraseFromParent(); 1034 continue; 1035 } 1036 } 1037 1038 // Insert a load from the global, and use it instead of the malloc. 1039 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt); 1040 U->replaceUsesOfWith(Alloc, NL); 1041 } 1042} 1043 1044/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi 1045/// of a load) are simple enough to perform heap SRA on. This permits GEP's 1046/// that index through the array and struct field, icmps of null, and PHIs. 1047static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V, 1048 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs, 1049 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) { 1050 // We permit two users of the load: setcc comparing against the null 1051 // pointer, and a getelementptr of a specific form. 1052 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; 1053 ++UI) { 1054 const Instruction *User = cast<Instruction>(*UI); 1055 1056 // Comparison against null is ok. 1057 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) { 1058 if (!isa<ConstantPointerNull>(ICI->getOperand(1))) 1059 return false; 1060 continue; 1061 } 1062 1063 // getelementptr is also ok, but only a simple form. 1064 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1065 // Must index into the array and into the struct. 1066 if (GEPI->getNumOperands() < 3) 1067 return false; 1068 1069 // Otherwise the GEP is ok. 1070 continue; 1071 } 1072 1073 if (const PHINode *PN = dyn_cast<PHINode>(User)) { 1074 if (!LoadUsingPHIsPerLoad.insert(PN)) 1075 // This means some phi nodes are dependent on each other. 1076 // Avoid infinite looping! 1077 return false; 1078 if (!LoadUsingPHIs.insert(PN)) 1079 // If we have already analyzed this PHI, then it is safe. 1080 continue; 1081 1082 // Make sure all uses of the PHI are simple enough to transform. 1083 if (!LoadUsesSimpleEnoughForHeapSRA(PN, 1084 LoadUsingPHIs, LoadUsingPHIsPerLoad)) 1085 return false; 1086 1087 continue; 1088 } 1089 1090 // Otherwise we don't know what this is, not ok. 1091 return false; 1092 } 1093 1094 return true; 1095} 1096 1097 1098/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from 1099/// GV are simple enough to perform HeapSRA, return true. 1100static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV, 1101 Instruction *StoredVal) { 1102 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs; 1103 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad; 1104 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end(); 1105 UI != E; ++UI) 1106 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 1107 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs, 1108 LoadUsingPHIsPerLoad)) 1109 return false; 1110 LoadUsingPHIsPerLoad.clear(); 1111 } 1112 1113 // If we reach here, we know that all uses of the loads and transitive uses 1114 // (through PHI nodes) are simple enough to transform. However, we don't know 1115 // that all inputs the to the PHI nodes are in the same equivalence sets. 1116 // Check to verify that all operands of the PHIs are either PHIS that can be 1117 // transformed, loads from GV, or MI itself. 1118 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin() 1119 , E = LoadUsingPHIs.end(); I != E; ++I) { 1120 const PHINode *PN = *I; 1121 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) { 1122 Value *InVal = PN->getIncomingValue(op); 1123 1124 // PHI of the stored value itself is ok. 1125 if (InVal == StoredVal) continue; 1126 1127 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) { 1128 // One of the PHIs in our set is (optimistically) ok. 1129 if (LoadUsingPHIs.count(InPN)) 1130 continue; 1131 return false; 1132 } 1133 1134 // Load from GV is ok. 1135 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal)) 1136 if (LI->getOperand(0) == GV) 1137 continue; 1138 1139 // UNDEF? NULL? 1140 1141 // Anything else is rejected. 1142 return false; 1143 } 1144 } 1145 1146 return true; 1147} 1148 1149static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, 1150 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1151 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1152 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V]; 1153 1154 if (FieldNo >= FieldVals.size()) 1155 FieldVals.resize(FieldNo+1); 1156 1157 // If we already have this value, just reuse the previously scalarized 1158 // version. 1159 if (Value *FieldVal = FieldVals[FieldNo]) 1160 return FieldVal; 1161 1162 // Depending on what instruction this is, we have several cases. 1163 Value *Result; 1164 if (LoadInst *LI = dyn_cast<LoadInst>(V)) { 1165 // This is a scalarized version of the load from the global. Just create 1166 // a new Load of the scalarized global. 1167 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo, 1168 InsertedScalarizedValues, 1169 PHIsToRewrite), 1170 LI->getName()+".f"+Twine(FieldNo), LI); 1171 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 1172 // PN's type is pointer to struct. Make a new PHI of pointer to struct 1173 // field. 1174 const StructType *ST = 1175 cast<StructType>(cast<PointerType>(PN->getType())->getElementType()); 1176 1177 Result = 1178 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)), 1179 PN->getName()+".f"+Twine(FieldNo), PN); 1180 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); 1181 } else { 1182 llvm_unreachable("Unknown usable value"); 1183 Result = 0; 1184 } 1185 1186 return FieldVals[FieldNo] = Result; 1187} 1188 1189/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from 1190/// the load, rewrite the derived value to use the HeapSRoA'd load. 1191static void RewriteHeapSROALoadUser(Instruction *LoadUser, 1192 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1193 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1194 // If this is a comparison against null, handle it. 1195 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) { 1196 assert(isa<ConstantPointerNull>(SCI->getOperand(1))); 1197 // If we have a setcc of the loaded pointer, we can use a setcc of any 1198 // field. 1199 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0, 1200 InsertedScalarizedValues, PHIsToRewrite); 1201 1202 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr, 1203 Constant::getNullValue(NPtr->getType()), 1204 SCI->getName()); 1205 SCI->replaceAllUsesWith(New); 1206 SCI->eraseFromParent(); 1207 return; 1208 } 1209 1210 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...' 1211 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) { 1212 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2)) 1213 && "Unexpected GEPI!"); 1214 1215 // Load the pointer for this field. 1216 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 1217 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo, 1218 InsertedScalarizedValues, PHIsToRewrite); 1219 1220 // Create the new GEP idx vector. 1221 SmallVector<Value*, 8> GEPIdx; 1222 GEPIdx.push_back(GEPI->getOperand(1)); 1223 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end()); 1224 1225 Value *NGEPI = GetElementPtrInst::Create(NewPtr, 1226 GEPIdx.begin(), GEPIdx.end(), 1227 GEPI->getName(), GEPI); 1228 GEPI->replaceAllUsesWith(NGEPI); 1229 GEPI->eraseFromParent(); 1230 return; 1231 } 1232 1233 // Recursively transform the users of PHI nodes. This will lazily create the 1234 // PHIs that are needed for individual elements. Keep track of what PHIs we 1235 // see in InsertedScalarizedValues so that we don't get infinite loops (very 1236 // antisocial). If the PHI is already in InsertedScalarizedValues, it has 1237 // already been seen first by another load, so its uses have already been 1238 // processed. 1239 PHINode *PN = cast<PHINode>(LoadUser); 1240 bool Inserted; 1241 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos; 1242 tie(InsertPos, Inserted) = 1243 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>())); 1244 if (!Inserted) return; 1245 1246 // If this is the first time we've seen this PHI, recursively process all 1247 // users. 1248 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) { 1249 Instruction *User = cast<Instruction>(*UI++); 1250 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1251 } 1252} 1253 1254/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr 1255/// is a value loaded from the global. Eliminate all uses of Ptr, making them 1256/// use FieldGlobals instead. All uses of loaded values satisfy 1257/// AllGlobalLoadUsesSimpleEnoughForHeapSRA. 1258static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, 1259 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues, 1260 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) { 1261 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end(); 1262 UI != E; ) { 1263 Instruction *User = cast<Instruction>(*UI++); 1264 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite); 1265 } 1266 1267 if (Load->use_empty()) { 1268 Load->eraseFromParent(); 1269 InsertedScalarizedValues.erase(Load); 1270 } 1271} 1272 1273/// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break 1274/// it up into multiple allocations of arrays of the fields. 1275static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, 1276 Value* NElems, TargetData *TD) { 1277 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n'); 1278 const Type* MAT = getMallocAllocatedType(CI); 1279 const StructType *STy = cast<StructType>(MAT); 1280 1281 // There is guaranteed to be at least one use of the malloc (storing 1282 // it into GV). If there are other uses, change them to be uses of 1283 // the global to simplify later code. This also deletes the store 1284 // into GV. 1285 ReplaceUsesOfMallocWithGlobal(CI, GV); 1286 1287 // Okay, at this point, there are no users of the malloc. Insert N 1288 // new mallocs at the same place as CI, and N globals. 1289 std::vector<Value*> FieldGlobals; 1290 std::vector<Value*> FieldMallocs; 1291 1292 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){ 1293 const Type *FieldTy = STy->getElementType(FieldNo); 1294 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy); 1295 1296 GlobalVariable *NGV = 1297 new GlobalVariable(*GV->getParent(), 1298 PFieldTy, false, GlobalValue::InternalLinkage, 1299 Constant::getNullValue(PFieldTy), 1300 GV->getName() + ".f" + Twine(FieldNo), GV, 1301 GV->isThreadLocal()); 1302 FieldGlobals.push_back(NGV); 1303 1304 unsigned TypeSize = TD->getTypeAllocSize(FieldTy); 1305 if (const StructType *ST = dyn_cast<StructType>(FieldTy)) 1306 TypeSize = TD->getStructLayout(ST)->getSizeInBytes(); 1307 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext()); 1308 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy, 1309 ConstantInt::get(IntPtrTy, TypeSize), 1310 NElems, 1311 CI->getName() + ".f" + Twine(FieldNo)); 1312 FieldMallocs.push_back(NMI); 1313 new StoreInst(NMI, NGV, CI); 1314 } 1315 1316 // The tricky aspect of this transformation is handling the case when malloc 1317 // fails. In the original code, malloc failing would set the result pointer 1318 // of malloc to null. In this case, some mallocs could succeed and others 1319 // could fail. As such, we emit code that looks like this: 1320 // F0 = malloc(field0) 1321 // F1 = malloc(field1) 1322 // F2 = malloc(field2) 1323 // if (F0 == 0 || F1 == 0 || F2 == 0) { 1324 // if (F0) { free(F0); F0 = 0; } 1325 // if (F1) { free(F1); F1 = 0; } 1326 // if (F2) { free(F2); F2 = 0; } 1327 // } 1328 // The malloc can also fail if its argument is too large. 1329 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0); 1330 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0), 1331 ConstantZero, "isneg"); 1332 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) { 1333 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i], 1334 Constant::getNullValue(FieldMallocs[i]->getType()), 1335 "isnull"); 1336 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI); 1337 } 1338 1339 // Split the basic block at the old malloc. 1340 BasicBlock *OrigBB = CI->getParent(); 1341 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont"); 1342 1343 // Create the block to check the first condition. Put all these blocks at the 1344 // end of the function as they are unlikely to be executed. 1345 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(), 1346 "malloc_ret_null", 1347 OrigBB->getParent()); 1348 1349 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond 1350 // branch on RunningOr. 1351 OrigBB->getTerminator()->eraseFromParent(); 1352 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB); 1353 1354 // Within the NullPtrBlock, we need to emit a comparison and branch for each 1355 // pointer, because some may be null while others are not. 1356 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1357 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock); 1358 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal, 1359 Constant::getNullValue(GVVal->getType()), 1360 "tmp"); 1361 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it", 1362 OrigBB->getParent()); 1363 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next", 1364 OrigBB->getParent()); 1365 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock, 1366 Cmp, NullPtrBlock); 1367 1368 // Fill in FreeBlock. 1369 CallInst::CreateFree(GVVal, BI); 1370 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i], 1371 FreeBlock); 1372 BranchInst::Create(NextBlock, FreeBlock); 1373 1374 NullPtrBlock = NextBlock; 1375 } 1376 1377 BranchInst::Create(ContBB, NullPtrBlock); 1378 1379 // CI is no longer needed, remove it. 1380 CI->eraseFromParent(); 1381 1382 /// InsertedScalarizedLoads - As we process loads, if we can't immediately 1383 /// update all uses of the load, keep track of what scalarized loads are 1384 /// inserted for a given load. 1385 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues; 1386 InsertedScalarizedValues[GV] = FieldGlobals; 1387 1388 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite; 1389 1390 // Okay, the malloc site is completely handled. All of the uses of GV are now 1391 // loads, and all uses of those loads are simple. Rewrite them to use loads 1392 // of the per-field globals instead. 1393 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) { 1394 Instruction *User = cast<Instruction>(*UI++); 1395 1396 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1397 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite); 1398 continue; 1399 } 1400 1401 // Must be a store of null. 1402 StoreInst *SI = cast<StoreInst>(User); 1403 assert(isa<ConstantPointerNull>(SI->getOperand(0)) && 1404 "Unexpected heap-sra user!"); 1405 1406 // Insert a store of null into each global. 1407 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) { 1408 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType()); 1409 Constant *Null = Constant::getNullValue(PT->getElementType()); 1410 new StoreInst(Null, FieldGlobals[i], SI); 1411 } 1412 // Erase the original store. 1413 SI->eraseFromParent(); 1414 } 1415 1416 // While we have PHIs that are interesting to rewrite, do it. 1417 while (!PHIsToRewrite.empty()) { 1418 PHINode *PN = PHIsToRewrite.back().first; 1419 unsigned FieldNo = PHIsToRewrite.back().second; 1420 PHIsToRewrite.pop_back(); 1421 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]); 1422 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi"); 1423 1424 // Add all the incoming values. This can materialize more phis. 1425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1426 Value *InVal = PN->getIncomingValue(i); 1427 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues, 1428 PHIsToRewrite); 1429 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i)); 1430 } 1431 } 1432 1433 // Drop all inter-phi links and any loads that made it this far. 1434 for (DenseMap<Value*, std::vector<Value*> >::iterator 1435 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1436 I != E; ++I) { 1437 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1438 PN->dropAllReferences(); 1439 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1440 LI->dropAllReferences(); 1441 } 1442 1443 // Delete all the phis and loads now that inter-references are dead. 1444 for (DenseMap<Value*, std::vector<Value*> >::iterator 1445 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end(); 1446 I != E; ++I) { 1447 if (PHINode *PN = dyn_cast<PHINode>(I->first)) 1448 PN->eraseFromParent(); 1449 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first)) 1450 LI->eraseFromParent(); 1451 } 1452 1453 // The old global is now dead, remove it. 1454 GV->eraseFromParent(); 1455 1456 ++NumHeapSRA; 1457 return cast<GlobalVariable>(FieldGlobals[0]); 1458} 1459 1460/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a 1461/// pointer global variable with a single value stored it that is a malloc or 1462/// cast of malloc. 1463static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, 1464 CallInst *CI, 1465 const Type *AllocTy, 1466 Module::global_iterator &GVI, 1467 TargetData *TD) { 1468 if (!TD) 1469 return false; 1470 1471 // If this is a malloc of an abstract type, don't touch it. 1472 if (!AllocTy->isSized()) 1473 return false; 1474 1475 // We can't optimize this global unless all uses of it are *known* to be 1476 // of the malloc value, not of the null initializer value (consider a use 1477 // that compares the global's value against zero to see if the malloc has 1478 // been reached). To do this, we check to see if all uses of the global 1479 // would trap if the global were null: this proves that they must all 1480 // happen after the malloc. 1481 if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) 1482 return false; 1483 1484 // We can't optimize this if the malloc itself is used in a complex way, 1485 // for example, being stored into multiple globals. This allows the 1486 // malloc to be stored into the specified global, loaded setcc'd, and 1487 // GEP'd. These are all things we could transform to using the global 1488 // for. 1489 SmallPtrSet<const PHINode*, 8> PHIs; 1490 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs)) 1491 return false; 1492 1493 // If we have a global that is only initialized with a fixed size malloc, 1494 // transform the program to use global memory instead of malloc'd memory. 1495 // This eliminates dynamic allocation, avoids an indirection accessing the 1496 // data, and exposes the resultant global to further GlobalOpt. 1497 // We cannot optimize the malloc if we cannot determine malloc array size. 1498 Value *NElems = getMallocArraySize(CI, TD, true); 1499 if (!NElems) 1500 return false; 1501 1502 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 1503 // Restrict this transformation to only working on small allocations 1504 // (2048 bytes currently), as we don't want to introduce a 16M global or 1505 // something. 1506 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) { 1507 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD); 1508 return true; 1509 } 1510 1511 // If the allocation is an array of structures, consider transforming this 1512 // into multiple malloc'd arrays, one for each field. This is basically 1513 // SRoA for malloc'd memory. 1514 1515 // If this is an allocation of a fixed size array of structs, analyze as a 1516 // variable size array. malloc [100 x struct],1 -> malloc struct, 100 1517 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1)) 1518 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy)) 1519 AllocTy = AT->getElementType(); 1520 1521 const StructType *AllocSTy = dyn_cast<StructType>(AllocTy); 1522 if (!AllocSTy) 1523 return false; 1524 1525 // This the structure has an unreasonable number of fields, leave it 1526 // alone. 1527 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 && 1528 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) { 1529 1530 // If this is a fixed size array, transform the Malloc to be an alloc of 1531 // structs. malloc [100 x struct],1 -> malloc struct, 100 1532 if (const ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) { 1533 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext()); 1534 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes(); 1535 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize); 1536 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements()); 1537 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy, 1538 AllocSize, NumElements, 1539 CI->getName()); 1540 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI); 1541 CI->replaceAllUsesWith(Cast); 1542 CI->eraseFromParent(); 1543 CI = dyn_cast<BitCastInst>(Malloc) ? 1544 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc); 1545 } 1546 1547 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD); 1548 return true; 1549 } 1550 1551 return false; 1552} 1553 1554// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge 1555// that only one value (besides its initializer) is ever stored to the global. 1556static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, 1557 Module::global_iterator &GVI, 1558 TargetData *TD) { 1559 // Ignore no-op GEPs and bitcasts. 1560 StoredOnceVal = StoredOnceVal->stripPointerCasts(); 1561 1562 // If we are dealing with a pointer global that is initialized to null and 1563 // only has one (non-null) value stored into it, then we can optimize any 1564 // users of the loaded value (often calls and loads) that would trap if the 1565 // value was null. 1566 if (GV->getInitializer()->getType()->isPointerTy() && 1567 GV->getInitializer()->isNullValue()) { 1568 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) { 1569 if (GV->getInitializer()->getType() != SOVC->getType()) 1570 SOVC = 1571 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); 1572 1573 // Optimize away any trapping uses of the loaded value. 1574 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) 1575 return true; 1576 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) { 1577 const Type* MallocType = getMallocAllocatedType(CI); 1578 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, 1579 GVI, TD)) 1580 return true; 1581 } 1582 } 1583 1584 return false; 1585} 1586 1587/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only 1588/// two values ever stored into GV are its initializer and OtherVal. See if we 1589/// can shrink the global into a boolean and select between the two values 1590/// whenever it is used. This exposes the values to other scalar optimizations. 1591static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { 1592 const Type *GVElType = GV->getType()->getElementType(); 1593 1594 // If GVElType is already i1, it is already shrunk. If the type of the GV is 1595 // an FP value, pointer or vector, don't do this optimization because a select 1596 // between them is very expensive and unlikely to lead to later 1597 // simplification. In these cases, we typically end up with "cond ? v1 : v2" 1598 // where v1 and v2 both require constant pool loads, a big loss. 1599 if (GVElType == Type::getInt1Ty(GV->getContext()) || 1600 GVElType->isFloatingPointTy() || 1601 GVElType->isPointerTy() || GVElType->isVectorTy()) 1602 return false; 1603 1604 // Walk the use list of the global seeing if all the uses are load or store. 1605 // If there is anything else, bail out. 1606 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I) 1607 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 1608 return false; 1609 1610 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV); 1611 1612 // Create the new global, initializing it to false. 1613 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()), 1614 false, 1615 GlobalValue::InternalLinkage, 1616 ConstantInt::getFalse(GV->getContext()), 1617 GV->getName()+".b", 1618 GV->isThreadLocal()); 1619 GV->getParent()->getGlobalList().insert(GV, NewGV); 1620 1621 Constant *InitVal = GV->getInitializer(); 1622 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) && 1623 "No reason to shrink to bool!"); 1624 1625 // If initialized to zero and storing one into the global, we can use a cast 1626 // instead of a select to synthesize the desired value. 1627 bool IsOneZero = false; 1628 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) 1629 IsOneZero = InitVal->isNullValue() && CI->isOne(); 1630 1631 while (!GV->use_empty()) { 1632 Instruction *UI = cast<Instruction>(GV->use_back()); 1633 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1634 // Change the store into a boolean store. 1635 bool StoringOther = SI->getOperand(0) == OtherVal; 1636 // Only do this if we weren't storing a loaded value. 1637 Value *StoreVal; 1638 if (StoringOther || SI->getOperand(0) == InitVal) 1639 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()), 1640 StoringOther); 1641 else { 1642 // Otherwise, we are storing a previously loaded copy. To do this, 1643 // change the copy from copying the original value to just copying the 1644 // bool. 1645 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0)); 1646 1647 // If we've already replaced the input, StoredVal will be a cast or 1648 // select instruction. If not, it will be a load of the original 1649 // global. 1650 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 1651 assert(LI->getOperand(0) == GV && "Not a copy!"); 1652 // Insert a new load, to preserve the saved value. 1653 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); 1654 } else { 1655 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && 1656 "This is not a form that we understand!"); 1657 StoreVal = StoredVal->getOperand(0); 1658 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!"); 1659 } 1660 } 1661 new StoreInst(StoreVal, NewGV, SI); 1662 } else { 1663 // Change the load into a load of bool then a select. 1664 LoadInst *LI = cast<LoadInst>(UI); 1665 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); 1666 Value *NSI; 1667 if (IsOneZero) 1668 NSI = new ZExtInst(NLI, LI->getType(), "", LI); 1669 else 1670 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI); 1671 NSI->takeName(LI); 1672 LI->replaceAllUsesWith(NSI); 1673 } 1674 UI->eraseFromParent(); 1675 } 1676 1677 GV->eraseFromParent(); 1678 return true; 1679} 1680 1681 1682/// ProcessInternalGlobal - Analyze the specified global variable and optimize 1683/// it if possible. If we make a change, return true. 1684bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 1685 Module::global_iterator &GVI) { 1686 SmallPtrSet<const PHINode*, 16> PHIUsers; 1687 GlobalStatus GS; 1688 GV->removeDeadConstantUsers(); 1689 1690 if (GV->use_empty()) { 1691 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV); 1692 GV->eraseFromParent(); 1693 ++NumDeleted; 1694 return true; 1695 } 1696 1697 if (!AnalyzeGlobal(GV, GS, PHIUsers)) { 1698#if 0 1699 DEBUG(dbgs() << "Global: " << *GV); 1700 DEBUG(dbgs() << " isLoaded = " << GS.isLoaded << "\n"); 1701 DEBUG(dbgs() << " StoredType = "); 1702 switch (GS.StoredType) { 1703 case GlobalStatus::NotStored: DEBUG(dbgs() << "NEVER STORED\n"); break; 1704 case GlobalStatus::isInitializerStored: DEBUG(dbgs() << "INIT STORED\n"); 1705 break; 1706 case GlobalStatus::isStoredOnce: DEBUG(dbgs() << "STORED ONCE\n"); break; 1707 case GlobalStatus::isStored: DEBUG(dbgs() << "stored\n"); break; 1708 } 1709 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue) 1710 DEBUG(dbgs() << " StoredOnceValue = " << *GS.StoredOnceValue << "\n"); 1711 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions) 1712 DEBUG(dbgs() << " AccessingFunction = " 1713 << GS.AccessingFunction->getName() << "\n"); 1714 DEBUG(dbgs() << " HasMultipleAccessingFunctions = " 1715 << GS.HasMultipleAccessingFunctions << "\n"); 1716 DEBUG(dbgs() << " HasNonInstructionUser = " 1717 << GS.HasNonInstructionUser<<"\n"); 1718 DEBUG(dbgs() << "\n"); 1719#endif 1720 1721 // If this is a first class global and has only one accessing function 1722 // and this function is main (which we know is not recursive we can make 1723 // this global a local variable) we replace the global with a local alloca 1724 // in this function. 1725 // 1726 // NOTE: It doesn't make sense to promote non single-value types since we 1727 // are just replacing static memory to stack memory. 1728 // 1729 // If the global is in different address space, don't bring it to stack. 1730 if (!GS.HasMultipleAccessingFunctions && 1731 GS.AccessingFunction && !GS.HasNonInstructionUser && 1732 GV->getType()->getElementType()->isSingleValueType() && 1733 GS.AccessingFunction->getName() == "main" && 1734 GS.AccessingFunction->hasExternalLinkage() && 1735 GV->getType()->getAddressSpace() == 0) { 1736 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV); 1737 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction 1738 ->getEntryBlock().begin()); 1739 const Type* ElemTy = GV->getType()->getElementType(); 1740 // FIXME: Pass Global's alignment when globals have alignment 1741 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI); 1742 if (!isa<UndefValue>(GV->getInitializer())) 1743 new StoreInst(GV->getInitializer(), Alloca, &FirstI); 1744 1745 GV->replaceAllUsesWith(Alloca); 1746 GV->eraseFromParent(); 1747 ++NumLocalized; 1748 return true; 1749 } 1750 1751 // If the global is never loaded (but may be stored to), it is dead. 1752 // Delete it now. 1753 if (!GS.isLoaded) { 1754 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV); 1755 1756 // Delete any stores we can find to the global. We may not be able to 1757 // make it completely dead though. 1758 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1759 1760 // If the global is dead now, delete it. 1761 if (GV->use_empty()) { 1762 GV->eraseFromParent(); 1763 ++NumDeleted; 1764 Changed = true; 1765 } 1766 return Changed; 1767 1768 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { 1769 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV); 1770 GV->setConstant(true); 1771 1772 // Clean up any obviously simplifiable users now. 1773 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1774 1775 // If the global is dead now, just nuke it. 1776 if (GV->use_empty()) { 1777 DEBUG(dbgs() << " *** Marking constant allowed us to simplify " 1778 << "all users and delete global!\n"); 1779 GV->eraseFromParent(); 1780 ++NumDeleted; 1781 } 1782 1783 ++NumMarked; 1784 return true; 1785 } else if (!GV->getInitializer()->getType()->isSingleValueType()) { 1786 if (TargetData *TD = getAnalysisIfAvailable<TargetData>()) 1787 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) { 1788 GVI = FirstNewGV; // Don't skip the newly produced globals! 1789 return true; 1790 } 1791 } else if (GS.StoredType == GlobalStatus::isStoredOnce) { 1792 // If the initial value for the global was an undef value, and if only 1793 // one other value was stored into it, we can just change the 1794 // initializer to be the stored value, then delete all stores to the 1795 // global. This allows us to mark it constant. 1796 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1797 if (isa<UndefValue>(GV->getInitializer())) { 1798 // Change the initial value here. 1799 GV->setInitializer(SOVConstant); 1800 1801 // Clean up any obviously simplifiable users now. 1802 CleanupConstantGlobalUsers(GV, GV->getInitializer()); 1803 1804 if (GV->use_empty()) { 1805 DEBUG(dbgs() << " *** Substituting initializer allowed us to " 1806 << "simplify all users and delete global!\n"); 1807 GV->eraseFromParent(); 1808 ++NumDeleted; 1809 } else { 1810 GVI = GV; 1811 } 1812 ++NumSubstitute; 1813 return true; 1814 } 1815 1816 // Try to optimize globals based on the knowledge that only one value 1817 // (besides its initializer) is ever stored to the global. 1818 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI, 1819 getAnalysisIfAvailable<TargetData>())) 1820 return true; 1821 1822 // Otherwise, if the global was not a boolean, we can shrink it to be a 1823 // boolean. 1824 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) 1825 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) { 1826 ++NumShrunkToBool; 1827 return true; 1828 } 1829 } 1830 } 1831 return false; 1832} 1833 1834/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified 1835/// function, changing them to FastCC. 1836static void ChangeCalleesToFastCall(Function *F) { 1837 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1838 CallSite User(cast<Instruction>(*UI)); 1839 User.setCallingConv(CallingConv::Fast); 1840 } 1841} 1842 1843static AttrListPtr StripNest(const AttrListPtr &Attrs) { 1844 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { 1845 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) 1846 continue; 1847 1848 // There can be only one. 1849 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); 1850 } 1851 1852 return Attrs; 1853} 1854 1855static void RemoveNestAttribute(Function *F) { 1856 F->setAttributes(StripNest(F->getAttributes())); 1857 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ 1858 CallSite User(cast<Instruction>(*UI)); 1859 User.setAttributes(StripNest(User.getAttributes())); 1860 } 1861} 1862 1863bool GlobalOpt::OptimizeFunctions(Module &M) { 1864 bool Changed = false; 1865 // Optimize functions. 1866 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) { 1867 Function *F = FI++; 1868 // Functions without names cannot be referenced outside this module. 1869 if (!F->hasName() && !F->isDeclaration()) 1870 F->setLinkage(GlobalValue::InternalLinkage); 1871 F->removeDeadConstantUsers(); 1872 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) { 1873 F->eraseFromParent(); 1874 Changed = true; 1875 ++NumFnDeleted; 1876 } else if (F->hasLocalLinkage()) { 1877 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() && 1878 !F->hasAddressTaken()) { 1879 // If this function has C calling conventions, is not a varargs 1880 // function, and is only called directly, promote it to use the Fast 1881 // calling convention. 1882 F->setCallingConv(CallingConv::Fast); 1883 ChangeCalleesToFastCall(F); 1884 ++NumFastCallFns; 1885 Changed = true; 1886 } 1887 1888 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && 1889 !F->hasAddressTaken()) { 1890 // The function is not used by a trampoline intrinsic, so it is safe 1891 // to remove the 'nest' attribute. 1892 RemoveNestAttribute(F); 1893 ++NumNestRemoved; 1894 Changed = true; 1895 } 1896 } 1897 } 1898 return Changed; 1899} 1900 1901bool GlobalOpt::OptimizeGlobalVars(Module &M) { 1902 bool Changed = false; 1903 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end(); 1904 GVI != E; ) { 1905 GlobalVariable *GV = GVI++; 1906 // Global variables without names cannot be referenced outside this module. 1907 if (!GV->hasName() && !GV->isDeclaration()) 1908 GV->setLinkage(GlobalValue::InternalLinkage); 1909 // Simplify the initializer. 1910 if (GV->hasInitializer()) 1911 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) { 1912 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 1913 Constant *New = ConstantFoldConstantExpression(CE, TD); 1914 if (New && New != CE) 1915 GV->setInitializer(New); 1916 } 1917 // Do more involved optimizations if the global is internal. 1918 if (!GV->isConstant() && GV->hasLocalLinkage() && 1919 GV->hasInitializer()) 1920 Changed |= ProcessInternalGlobal(GV, GVI); 1921 } 1922 return Changed; 1923} 1924 1925/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all 1926/// initializers have an init priority of 65535. 1927GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { 1928 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 1929 I != E; ++I) 1930 if (I->getName() == "llvm.global_ctors") { 1931 // Found it, verify it's an array of { int, void()* }. 1932 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType()); 1933 if (!ATy) return 0; 1934 const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); 1935 if (!STy || STy->getNumElements() != 2 || 1936 !STy->getElementType(0)->isIntegerTy(32)) return 0; 1937 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); 1938 if (!PFTy) return 0; 1939 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); 1940 if (!FTy || !FTy->getReturnType()->isVoidTy() || 1941 FTy->isVarArg() || FTy->getNumParams() != 0) 1942 return 0; 1943 1944 // Verify that the initializer is simple enough for us to handle. 1945 if (!I->hasDefinitiveInitializer()) return 0; 1946 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer()); 1947 if (!CA) return 0; 1948 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 1949 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) { 1950 if (isa<ConstantPointerNull>(CS->getOperand(1))) 1951 continue; 1952 1953 // Must have a function or null ptr. 1954 if (!isa<Function>(CS->getOperand(1))) 1955 return 0; 1956 1957 // Init priority must be standard. 1958 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 1959 if (!CI || CI->getZExtValue() != 65535) 1960 return 0; 1961 } else { 1962 return 0; 1963 } 1964 1965 return I; 1966 } 1967 return 0; 1968} 1969 1970/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand, 1971/// return a list of the functions and null terminator as a vector. 1972static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) { 1973 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1974 std::vector<Function*> Result; 1975 Result.reserve(CA->getNumOperands()); 1976 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) { 1977 ConstantStruct *CS = cast<ConstantStruct>(*i); 1978 Result.push_back(dyn_cast<Function>(CS->getOperand(1))); 1979 } 1980 return Result; 1981} 1982 1983/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the 1984/// specified array, returning the new global to use. 1985static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, 1986 const std::vector<Function*> &Ctors) { 1987 // If we made a change, reassemble the initializer list. 1988 std::vector<Constant*> CSVals; 1989 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535)); 1990 CSVals.push_back(0); 1991 1992 // Create the new init list. 1993 std::vector<Constant*> CAList; 1994 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) { 1995 if (Ctors[i]) { 1996 CSVals[1] = Ctors[i]; 1997 } else { 1998 const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()), 1999 false); 2000 const PointerType *PFTy = PointerType::getUnqual(FTy); 2001 CSVals[1] = Constant::getNullValue(PFTy); 2002 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 2003 2147483647); 2004 } 2005 CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false)); 2006 } 2007 2008 // Create the array initializer. 2009 const Type *StructTy = 2010 cast<ArrayType>(GCL->getType()->getElementType())->getElementType(); 2011 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, 2012 CAList.size()), CAList); 2013 2014 // If we didn't change the number of elements, don't create a new GV. 2015 if (CA->getType() == GCL->getInitializer()->getType()) { 2016 GCL->setInitializer(CA); 2017 return GCL; 2018 } 2019 2020 // Create the new global and insert it next to the existing list. 2021 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(), 2022 GCL->getLinkage(), CA, "", 2023 GCL->isThreadLocal()); 2024 GCL->getParent()->getGlobalList().insert(GCL, NGV); 2025 NGV->takeName(GCL); 2026 2027 // Nuke the old list, replacing any uses with the new one. 2028 if (!GCL->use_empty()) { 2029 Constant *V = NGV; 2030 if (V->getType() != GCL->getType()) 2031 V = ConstantExpr::getBitCast(V, GCL->getType()); 2032 GCL->replaceAllUsesWith(V); 2033 } 2034 GCL->eraseFromParent(); 2035 2036 if (Ctors.size()) 2037 return NGV; 2038 else 2039 return 0; 2040} 2041 2042 2043static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, 2044 Value *V) { 2045 if (Constant *CV = dyn_cast<Constant>(V)) return CV; 2046 Constant *R = ComputedValues[V]; 2047 assert(R && "Reference to an uncomputed value!"); 2048 return R; 2049} 2050 2051/// isSimpleEnoughPointerToCommit - Return true if this constant is simple 2052/// enough for us to understand. In particular, if it is a cast of something, 2053/// we punt. We basically just support direct accesses to globals and GEP's of 2054/// globals. This should be kept up to date with CommitValueTo. 2055static bool isSimpleEnoughPointerToCommit(Constant *C) { 2056 // Conservatively, avoid aggregate types. This is because we don't 2057 // want to worry about them partially overlapping other stores. 2058 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType()) 2059 return false; 2060 2061 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 2062 // Do not allow weak/linkonce/dllimport/dllexport linkage or 2063 // external globals. 2064 return GV->hasDefinitiveInitializer(); 2065 2066 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2067 // Handle a constantexpr gep. 2068 if (CE->getOpcode() == Instruction::GetElementPtr && 2069 isa<GlobalVariable>(CE->getOperand(0)) && 2070 cast<GEPOperator>(CE)->isInBounds()) { 2071 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2072 // Do not allow weak/linkonce/dllimport/dllexport linkage or 2073 // external globals. 2074 if (!GV->hasDefinitiveInitializer()) 2075 return false; 2076 2077 // The first index must be zero. 2078 ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin())); 2079 if (!CI || !CI->isZero()) return false; 2080 2081 // The remaining indices must be compile-time known integers within the 2082 // notional bounds of the corresponding static array types. 2083 if (!CE->isGEPWithNoNotionalOverIndexing()) 2084 return false; 2085 2086 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2087 } 2088 return false; 2089} 2090 2091/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global 2092/// initializer. This returns 'Init' modified to reflect 'Val' stored into it. 2093/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into. 2094static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, 2095 ConstantExpr *Addr, unsigned OpNo) { 2096 // Base case of the recursion. 2097 if (OpNo == Addr->getNumOperands()) { 2098 assert(Val->getType() == Init->getType() && "Type mismatch!"); 2099 return Val; 2100 } 2101 2102 std::vector<Constant*> Elts; 2103 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 2104 2105 // Break up the constant into its elements. 2106 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 2107 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) 2108 Elts.push_back(cast<Constant>(*i)); 2109 } else if (isa<ConstantAggregateZero>(Init)) { 2110 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2111 Elts.push_back(Constant::getNullValue(STy->getElementType(i))); 2112 } else if (isa<UndefValue>(Init)) { 2113 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 2114 Elts.push_back(UndefValue::get(STy->getElementType(i))); 2115 } else { 2116 llvm_unreachable("This code is out of sync with " 2117 " ConstantFoldLoadThroughGEPConstantExpr"); 2118 } 2119 2120 // Replace the element that we are supposed to. 2121 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo)); 2122 unsigned Idx = CU->getZExtValue(); 2123 assert(Idx < STy->getNumElements() && "Struct index out of range!"); 2124 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1); 2125 2126 // Return the modified struct. 2127 return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(), 2128 STy->isPacked()); 2129 } else { 2130 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo)); 2131 const SequentialType *InitTy = cast<SequentialType>(Init->getType()); 2132 2133 uint64_t NumElts; 2134 if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy)) 2135 NumElts = ATy->getNumElements(); 2136 else 2137 NumElts = cast<VectorType>(InitTy)->getNumElements(); 2138 2139 2140 // Break up the array into elements. 2141 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 2142 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) 2143 Elts.push_back(cast<Constant>(*i)); 2144 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) { 2145 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i) 2146 Elts.push_back(cast<Constant>(*i)); 2147 } else if (isa<ConstantAggregateZero>(Init)) { 2148 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType())); 2149 } else { 2150 assert(isa<UndefValue>(Init) && "This code is out of sync with " 2151 " ConstantFoldLoadThroughGEPConstantExpr"); 2152 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType())); 2153 } 2154 2155 assert(CI->getZExtValue() < NumElts); 2156 Elts[CI->getZExtValue()] = 2157 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1); 2158 2159 if (Init->getType()->isArrayTy()) 2160 return ConstantArray::get(cast<ArrayType>(InitTy), Elts); 2161 else 2162 return ConstantVector::get(&Elts[0], Elts.size()); 2163 } 2164} 2165 2166/// CommitValueTo - We have decided that Addr (which satisfies the predicate 2167/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen. 2168static void CommitValueTo(Constant *Val, Constant *Addr) { 2169 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 2170 assert(GV->hasInitializer()); 2171 GV->setInitializer(Val); 2172 return; 2173 } 2174 2175 ConstantExpr *CE = cast<ConstantExpr>(Addr); 2176 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2177 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2)); 2178} 2179 2180/// ComputeLoadResult - Return the value that would be computed by a load from 2181/// P after the stores reflected by 'memory' have been performed. If we can't 2182/// decide, return null. 2183static Constant *ComputeLoadResult(Constant *P, 2184 const DenseMap<Constant*, Constant*> &Memory) { 2185 // If this memory location has been recently stored, use the stored value: it 2186 // is the most up-to-date. 2187 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P); 2188 if (I != Memory.end()) return I->second; 2189 2190 // Access it. 2191 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 2192 if (GV->hasDefinitiveInitializer()) 2193 return GV->getInitializer(); 2194 return 0; 2195 } 2196 2197 // Handle a constantexpr getelementptr. 2198 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) 2199 if (CE->getOpcode() == Instruction::GetElementPtr && 2200 isa<GlobalVariable>(CE->getOperand(0))) { 2201 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 2202 if (GV->hasDefinitiveInitializer()) 2203 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 2204 } 2205 2206 return 0; // don't know how to evaluate. 2207} 2208 2209/// EvaluateFunction - Evaluate a call to function F, returning true if 2210/// successful, false if we can't evaluate it. ActualArgs contains the formal 2211/// arguments for the function. 2212static bool EvaluateFunction(Function *F, Constant *&RetVal, 2213 const SmallVectorImpl<Constant*> &ActualArgs, 2214 std::vector<Function*> &CallStack, 2215 DenseMap<Constant*, Constant*> &MutatedMemory, 2216 std::vector<GlobalVariable*> &AllocaTmps) { 2217 // Check to see if this function is already executing (recursion). If so, 2218 // bail out. TODO: we might want to accept limited recursion. 2219 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end()) 2220 return false; 2221 2222 CallStack.push_back(F); 2223 2224 /// Values - As we compute SSA register values, we store their contents here. 2225 DenseMap<Value*, Constant*> Values; 2226 2227 // Initialize arguments to the incoming values specified. 2228 unsigned ArgNo = 0; 2229 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 2230 ++AI, ++ArgNo) 2231 Values[AI] = ActualArgs[ArgNo]; 2232 2233 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 2234 /// we can only evaluate any one basic block at most once. This set keeps 2235 /// track of what we have executed so we can detect recursive cases etc. 2236 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 2237 2238 // CurInst - The current instruction we're evaluating. 2239 BasicBlock::iterator CurInst = F->begin()->begin(); 2240 2241 // This is the main evaluation loop. 2242 while (1) { 2243 Constant *InstResult = 0; 2244 2245 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 2246 if (SI->isVolatile()) return false; // no volatile accesses. 2247 Constant *Ptr = getVal(Values, SI->getOperand(1)); 2248 if (!isSimpleEnoughPointerToCommit(Ptr)) 2249 // If this is too complex for us to commit, reject it. 2250 return false; 2251 Constant *Val = getVal(Values, SI->getOperand(0)); 2252 MutatedMemory[Ptr] = Val; 2253 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 2254 InstResult = ConstantExpr::get(BO->getOpcode(), 2255 getVal(Values, BO->getOperand(0)), 2256 getVal(Values, BO->getOperand(1))); 2257 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 2258 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 2259 getVal(Values, CI->getOperand(0)), 2260 getVal(Values, CI->getOperand(1))); 2261 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 2262 InstResult = ConstantExpr::getCast(CI->getOpcode(), 2263 getVal(Values, CI->getOperand(0)), 2264 CI->getType()); 2265 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 2266 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), 2267 getVal(Values, SI->getOperand(1)), 2268 getVal(Values, SI->getOperand(2))); 2269 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 2270 Constant *P = getVal(Values, GEP->getOperand(0)); 2271 SmallVector<Constant*, 8> GEPOps; 2272 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 2273 i != e; ++i) 2274 GEPOps.push_back(getVal(Values, *i)); 2275 InstResult = cast<GEPOperator>(GEP)->isInBounds() ? 2276 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) : 2277 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size()); 2278 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 2279 if (LI->isVolatile()) return false; // no volatile accesses. 2280 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), 2281 MutatedMemory); 2282 if (InstResult == 0) return false; // Could not evaluate load. 2283 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 2284 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. 2285 const Type *Ty = AI->getType()->getElementType(); 2286 AllocaTmps.push_back(new GlobalVariable(Ty, false, 2287 GlobalValue::InternalLinkage, 2288 UndefValue::get(Ty), 2289 AI->getName())); 2290 InstResult = AllocaTmps.back(); 2291 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) { 2292 2293 // Debug info can safely be ignored here. 2294 if (isa<DbgInfoIntrinsic>(CI)) { 2295 ++CurInst; 2296 continue; 2297 } 2298 2299 // Cannot handle inline asm. 2300 if (isa<InlineAsm>(CI->getCalledValue())) return false; 2301 2302 // Resolve function pointers. 2303 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getCalledValue())); 2304 if (!Callee) return false; // Cannot resolve. 2305 2306 SmallVector<Constant*, 8> Formals; 2307 CallSite CS(CI); 2308 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); 2309 i != e; ++i) 2310 Formals.push_back(getVal(Values, *i)); 2311 2312 if (Callee->isDeclaration()) { 2313 // If this is a function we can constant fold, do it. 2314 if (Constant *C = ConstantFoldCall(Callee, Formals.data(), 2315 Formals.size())) { 2316 InstResult = C; 2317 } else { 2318 return false; 2319 } 2320 } else { 2321 if (Callee->getFunctionType()->isVarArg()) 2322 return false; 2323 2324 Constant *RetVal; 2325 // Execute the call, if successful, use the return value. 2326 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, 2327 MutatedMemory, AllocaTmps)) 2328 return false; 2329 InstResult = RetVal; 2330 } 2331 } else if (isa<TerminatorInst>(CurInst)) { 2332 BasicBlock *NewBB = 0; 2333 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 2334 if (BI->isUnconditional()) { 2335 NewBB = BI->getSuccessor(0); 2336 } else { 2337 ConstantInt *Cond = 2338 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition())); 2339 if (!Cond) return false; // Cannot determine. 2340 2341 NewBB = BI->getSuccessor(!Cond->getZExtValue()); 2342 } 2343 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 2344 ConstantInt *Val = 2345 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition())); 2346 if (!Val) return false; // Cannot determine. 2347 NewBB = SI->getSuccessor(SI->findCaseValue(Val)); 2348 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) { 2349 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts(); 2350 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val)) 2351 NewBB = BA->getBasicBlock(); 2352 else 2353 return false; // Cannot determine. 2354 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) { 2355 if (RI->getNumOperands()) 2356 RetVal = getVal(Values, RI->getOperand(0)); 2357 2358 CallStack.pop_back(); // return from fn. 2359 return true; // We succeeded at evaluating this ctor! 2360 } else { 2361 // invoke, unwind, unreachable. 2362 return false; // Cannot handle this terminator. 2363 } 2364 2365 // Okay, we succeeded in evaluating this control flow. See if we have 2366 // executed the new block before. If so, we have a looping function, 2367 // which we cannot evaluate in reasonable time. 2368 if (!ExecutedBlocks.insert(NewBB)) 2369 return false; // looped! 2370 2371 // Okay, we have never been in this block before. Check to see if there 2372 // are any PHI nodes. If so, evaluate them with information about where 2373 // we came from. 2374 BasicBlock *OldBB = CurInst->getParent(); 2375 CurInst = NewBB->begin(); 2376 PHINode *PN; 2377 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 2378 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); 2379 2380 // Do NOT increment CurInst. We know that the terminator had no value. 2381 continue; 2382 } else { 2383 // Did not know how to evaluate this! 2384 return false; 2385 } 2386 2387 if (!CurInst->use_empty()) 2388 Values[CurInst] = InstResult; 2389 2390 // Advance program counter. 2391 ++CurInst; 2392 } 2393} 2394 2395/// EvaluateStaticConstructor - Evaluate static constructors in the function, if 2396/// we can. Return true if we can, false otherwise. 2397static bool EvaluateStaticConstructor(Function *F) { 2398 /// MutatedMemory - For each store we execute, we update this map. Loads 2399 /// check this to get the most up-to-date value. If evaluation is successful, 2400 /// this state is committed to the process. 2401 DenseMap<Constant*, Constant*> MutatedMemory; 2402 2403 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable 2404 /// to represent its body. This vector is needed so we can delete the 2405 /// temporary globals when we are done. 2406 std::vector<GlobalVariable*> AllocaTmps; 2407 2408 /// CallStack - This is used to detect recursion. In pathological situations 2409 /// we could hit exponential behavior, but at least there is nothing 2410 /// unbounded. 2411 std::vector<Function*> CallStack; 2412 2413 // Call the function. 2414 Constant *RetValDummy; 2415 bool EvalSuccess = EvaluateFunction(F, RetValDummy, 2416 SmallVector<Constant*, 0>(), CallStack, 2417 MutatedMemory, AllocaTmps); 2418 if (EvalSuccess) { 2419 // We succeeded at evaluation: commit the result. 2420 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" 2421 << F->getName() << "' to " << MutatedMemory.size() 2422 << " stores.\n"); 2423 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(), 2424 E = MutatedMemory.end(); I != E; ++I) 2425 CommitValueTo(I->second, I->first); 2426 } 2427 2428 // At this point, we are done interpreting. If we created any 'alloca' 2429 // temporaries, release them now. 2430 while (!AllocaTmps.empty()) { 2431 GlobalVariable *Tmp = AllocaTmps.back(); 2432 AllocaTmps.pop_back(); 2433 2434 // If there are still users of the alloca, the program is doing something 2435 // silly, e.g. storing the address of the alloca somewhere and using it 2436 // later. Since this is undefined, we'll just make it be null. 2437 if (!Tmp->use_empty()) 2438 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); 2439 delete Tmp; 2440 } 2441 2442 return EvalSuccess; 2443} 2444 2445 2446 2447/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. 2448/// Return true if anything changed. 2449bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { 2450 std::vector<Function*> Ctors = ParseGlobalCtors(GCL); 2451 bool MadeChange = false; 2452 if (Ctors.empty()) return false; 2453 2454 // Loop over global ctors, optimizing them when we can. 2455 for (unsigned i = 0; i != Ctors.size(); ++i) { 2456 Function *F = Ctors[i]; 2457 // Found a null terminator in the middle of the list, prune off the rest of 2458 // the list. 2459 if (F == 0) { 2460 if (i != Ctors.size()-1) { 2461 Ctors.resize(i+1); 2462 MadeChange = true; 2463 } 2464 break; 2465 } 2466 2467 // We cannot simplify external ctor functions. 2468 if (F->empty()) continue; 2469 2470 // If we can evaluate the ctor at compile time, do. 2471 if (EvaluateStaticConstructor(F)) { 2472 Ctors.erase(Ctors.begin()+i); 2473 MadeChange = true; 2474 --i; 2475 ++NumCtorsEvaluated; 2476 continue; 2477 } 2478 } 2479 2480 if (!MadeChange) return false; 2481 2482 GCL = InstallGlobalCtors(GCL, Ctors); 2483 return true; 2484} 2485 2486bool GlobalOpt::OptimizeGlobalAliases(Module &M) { 2487 bool Changed = false; 2488 2489 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 2490 I != E;) { 2491 Module::alias_iterator J = I++; 2492 // Aliases without names cannot be referenced outside this module. 2493 if (!J->hasName() && !J->isDeclaration()) 2494 J->setLinkage(GlobalValue::InternalLinkage); 2495 // If the aliasee may change at link time, nothing can be done - bail out. 2496 if (J->mayBeOverridden()) 2497 continue; 2498 2499 Constant *Aliasee = J->getAliasee(); 2500 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts()); 2501 Target->removeDeadConstantUsers(); 2502 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse(); 2503 2504 // Make all users of the alias use the aliasee instead. 2505 if (!J->use_empty()) { 2506 J->replaceAllUsesWith(Aliasee); 2507 ++NumAliasesResolved; 2508 Changed = true; 2509 } 2510 2511 // If the alias is externally visible, we may still be able to simplify it. 2512 if (!J->hasLocalLinkage()) { 2513 // If the aliasee has internal linkage, give it the name and linkage 2514 // of the alias, and delete the alias. This turns: 2515 // define internal ... @f(...) 2516 // @a = alias ... @f 2517 // into: 2518 // define ... @a(...) 2519 if (!Target->hasLocalLinkage()) 2520 continue; 2521 2522 // Do not perform the transform if multiple aliases potentially target the 2523 // aliasee. This check also ensures that it is safe to replace the section 2524 // and other attributes of the aliasee with those of the alias. 2525 if (!hasOneUse) 2526 continue; 2527 2528 // Give the aliasee the name, linkage and other attributes of the alias. 2529 Target->takeName(J); 2530 Target->setLinkage(J->getLinkage()); 2531 Target->GlobalValue::copyAttributesFrom(J); 2532 } 2533 2534 // Delete the alias. 2535 M.getAliasList().erase(J); 2536 ++NumAliasesRemoved; 2537 Changed = true; 2538 } 2539 2540 return Changed; 2541} 2542 2543bool GlobalOpt::runOnModule(Module &M) { 2544 bool Changed = false; 2545 2546 // Try to find the llvm.globalctors list. 2547 GlobalVariable *GlobalCtors = FindGlobalCtors(M); 2548 2549 bool LocalChange = true; 2550 while (LocalChange) { 2551 LocalChange = false; 2552 2553 // Delete functions that are trivially dead, ccc -> fastcc 2554 LocalChange |= OptimizeFunctions(M); 2555 2556 // Optimize global_ctors list. 2557 if (GlobalCtors) 2558 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors); 2559 2560 // Optimize non-address-taken globals. 2561 LocalChange |= OptimizeGlobalVars(M); 2562 2563 // Resolve aliases, when possible. 2564 LocalChange |= OptimizeGlobalAliases(M); 2565 Changed |= LocalChange; 2566 } 2567 2568 // TODO: Move all global ctors functions to the end of the module for code 2569 // layout. 2570 2571 return Changed; 2572} 2573