LinkModules.cpp revision b976e668165e1875a8f1eb7af800e33bb1e4393d
1//===- Linker.cpp - Module Linker Implementation --------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the LLVM module linker. 11// 12// Specifically, this: 13// * Merges global variables between the two modules 14// * Uninit + Uninit = Init, Init + Uninit = Init, Init + Init = Error if != 15// * Merges functions between two modules 16// 17//===----------------------------------------------------------------------===// 18 19#include "llvm/Support/Linker.h" 20#include "llvm/Constants.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Module.h" 23#include "llvm/SymbolTable.h" 24#include "llvm/Instructions.h" 25#include "llvm/Assembly/Writer.h" 26#include "llvm/System/Path.h" 27#include <iostream> 28#include <sstream> 29using namespace llvm; 30 31// Error - Simple wrapper function to conditionally assign to E and return true. 32// This just makes error return conditions a little bit simpler... 33// 34static inline bool Error(std::string *E, const std::string &Message) { 35 if (E) *E = Message; 36 return true; 37} 38 39static std::string ToStr(const Type *Ty, const Module *M) { 40 std::ostringstream OS; 41 WriteTypeSymbolic(OS, Ty, M); 42 return OS.str(); 43} 44 45// 46// Function: ResolveTypes() 47// 48// Description: 49// Attempt to link the two specified types together. 50// 51// Inputs: 52// DestTy - The type to which we wish to resolve. 53// SrcTy - The original type which we want to resolve. 54// Name - The name of the type. 55// 56// Outputs: 57// DestST - The symbol table in which the new type should be placed. 58// 59// Return value: 60// true - There is an error and the types cannot yet be linked. 61// false - No errors. 62// 63static bool ResolveTypes(const Type *DestTy, const Type *SrcTy, 64 SymbolTable *DestST, const std::string &Name) { 65 if (DestTy == SrcTy) return false; // If already equal, noop 66 67 // Does the type already exist in the module? 68 if (DestTy && !isa<OpaqueType>(DestTy)) { // Yup, the type already exists... 69 if (const OpaqueType *OT = dyn_cast<OpaqueType>(SrcTy)) { 70 const_cast<OpaqueType*>(OT)->refineAbstractTypeTo(DestTy); 71 } else { 72 return true; // Cannot link types... neither is opaque and not-equal 73 } 74 } else { // Type not in dest module. Add it now. 75 if (DestTy) // Type _is_ in module, just opaque... 76 const_cast<OpaqueType*>(cast<OpaqueType>(DestTy)) 77 ->refineAbstractTypeTo(SrcTy); 78 else if (!Name.empty()) 79 DestST->insert(Name, const_cast<Type*>(SrcTy)); 80 } 81 return false; 82} 83 84static const FunctionType *getFT(const PATypeHolder &TH) { 85 return cast<FunctionType>(TH.get()); 86} 87static const StructType *getST(const PATypeHolder &TH) { 88 return cast<StructType>(TH.get()); 89} 90 91// RecursiveResolveTypes - This is just like ResolveTypes, except that it 92// recurses down into derived types, merging the used types if the parent types 93// are compatible. 94// 95static bool RecursiveResolveTypesI(const PATypeHolder &DestTy, 96 const PATypeHolder &SrcTy, 97 SymbolTable *DestST, const std::string &Name, 98 std::vector<std::pair<PATypeHolder, PATypeHolder> > &Pointers) { 99 const Type *SrcTyT = SrcTy.get(); 100 const Type *DestTyT = DestTy.get(); 101 if (DestTyT == SrcTyT) return false; // If already equal, noop 102 103 // If we found our opaque type, resolve it now! 104 if (isa<OpaqueType>(DestTyT) || isa<OpaqueType>(SrcTyT)) 105 return ResolveTypes(DestTyT, SrcTyT, DestST, Name); 106 107 // Two types cannot be resolved together if they are of different primitive 108 // type. For example, we cannot resolve an int to a float. 109 if (DestTyT->getTypeID() != SrcTyT->getTypeID()) return true; 110 111 // Otherwise, resolve the used type used by this derived type... 112 switch (DestTyT->getTypeID()) { 113 case Type::FunctionTyID: { 114 if (cast<FunctionType>(DestTyT)->isVarArg() != 115 cast<FunctionType>(SrcTyT)->isVarArg() || 116 cast<FunctionType>(DestTyT)->getNumContainedTypes() != 117 cast<FunctionType>(SrcTyT)->getNumContainedTypes()) 118 return true; 119 for (unsigned i = 0, e = getFT(DestTy)->getNumContainedTypes(); i != e; ++i) 120 if (RecursiveResolveTypesI(getFT(DestTy)->getContainedType(i), 121 getFT(SrcTy)->getContainedType(i), DestST, "", 122 Pointers)) 123 return true; 124 return false; 125 } 126 case Type::StructTyID: { 127 if (getST(DestTy)->getNumContainedTypes() != 128 getST(SrcTy)->getNumContainedTypes()) return 1; 129 for (unsigned i = 0, e = getST(DestTy)->getNumContainedTypes(); i != e; ++i) 130 if (RecursiveResolveTypesI(getST(DestTy)->getContainedType(i), 131 getST(SrcTy)->getContainedType(i), DestST, "", 132 Pointers)) 133 return true; 134 return false; 135 } 136 case Type::ArrayTyID: { 137 const ArrayType *DAT = cast<ArrayType>(DestTy.get()); 138 const ArrayType *SAT = cast<ArrayType>(SrcTy.get()); 139 if (DAT->getNumElements() != SAT->getNumElements()) return true; 140 return RecursiveResolveTypesI(DAT->getElementType(), SAT->getElementType(), 141 DestST, "", Pointers); 142 } 143 case Type::PointerTyID: { 144 // If this is a pointer type, check to see if we have already seen it. If 145 // so, we are in a recursive branch. Cut off the search now. We cannot use 146 // an associative container for this search, because the type pointers (keys 147 // in the container) change whenever types get resolved... 148 // 149 for (unsigned i = 0, e = Pointers.size(); i != e; ++i) 150 if (Pointers[i].first == DestTy) 151 return Pointers[i].second != SrcTy; 152 153 // Otherwise, add the current pointers to the vector to stop recursion on 154 // this pair. 155 Pointers.push_back(std::make_pair(DestTyT, SrcTyT)); 156 bool Result = 157 RecursiveResolveTypesI(cast<PointerType>(DestTy.get())->getElementType(), 158 cast<PointerType>(SrcTy.get())->getElementType(), 159 DestST, "", Pointers); 160 Pointers.pop_back(); 161 return Result; 162 } 163 default: assert(0 && "Unexpected type!"); return true; 164 } 165} 166 167static bool RecursiveResolveTypes(const PATypeHolder &DestTy, 168 const PATypeHolder &SrcTy, 169 SymbolTable *DestST, const std::string &Name){ 170 std::vector<std::pair<PATypeHolder, PATypeHolder> > PointerTypes; 171 return RecursiveResolveTypesI(DestTy, SrcTy, DestST, Name, PointerTypes); 172} 173 174 175// LinkTypes - Go through the symbol table of the Src module and see if any 176// types are named in the src module that are not named in the Dst module. 177// Make sure there are no type name conflicts. 178// 179static bool LinkTypes(Module *Dest, const Module *Src, std::string *Err) { 180 SymbolTable *DestST = &Dest->getSymbolTable(); 181 const SymbolTable *SrcST = &Src->getSymbolTable(); 182 183 // Look for a type plane for Type's... 184 SymbolTable::type_const_iterator TI = SrcST->type_begin(); 185 SymbolTable::type_const_iterator TE = SrcST->type_end(); 186 if (TI == TE) return false; // No named types, do nothing. 187 188 // Some types cannot be resolved immediately because they depend on other 189 // types being resolved to each other first. This contains a list of types we 190 // are waiting to recheck. 191 std::vector<std::string> DelayedTypesToResolve; 192 193 for ( ; TI != TE; ++TI ) { 194 const std::string &Name = TI->first; 195 const Type *RHS = TI->second; 196 197 // Check to see if this type name is already in the dest module... 198 Type *Entry = DestST->lookupType(Name); 199 200 if (ResolveTypes(Entry, RHS, DestST, Name)) { 201 // They look different, save the types 'till later to resolve. 202 DelayedTypesToResolve.push_back(Name); 203 } 204 } 205 206 // Iteratively resolve types while we can... 207 while (!DelayedTypesToResolve.empty()) { 208 // Loop over all of the types, attempting to resolve them if possible... 209 unsigned OldSize = DelayedTypesToResolve.size(); 210 211 // Try direct resolution by name... 212 for (unsigned i = 0; i != DelayedTypesToResolve.size(); ++i) { 213 const std::string &Name = DelayedTypesToResolve[i]; 214 Type *T1 = SrcST->lookupType(Name); 215 Type *T2 = DestST->lookupType(Name); 216 if (!ResolveTypes(T2, T1, DestST, Name)) { 217 // We are making progress! 218 DelayedTypesToResolve.erase(DelayedTypesToResolve.begin()+i); 219 --i; 220 } 221 } 222 223 // Did we not eliminate any types? 224 if (DelayedTypesToResolve.size() == OldSize) { 225 // Attempt to resolve subelements of types. This allows us to merge these 226 // two types: { int* } and { opaque* } 227 for (unsigned i = 0, e = DelayedTypesToResolve.size(); i != e; ++i) { 228 const std::string &Name = DelayedTypesToResolve[i]; 229 PATypeHolder T1(SrcST->lookupType(Name)); 230 PATypeHolder T2(DestST->lookupType(Name)); 231 232 if (!RecursiveResolveTypes(T2, T1, DestST, Name)) { 233 // We are making progress! 234 DelayedTypesToResolve.erase(DelayedTypesToResolve.begin()+i); 235 236 // Go back to the main loop, perhaps we can resolve directly by name 237 // now... 238 break; 239 } 240 } 241 242 // If we STILL cannot resolve the types, then there is something wrong. 243 // Report the warning and delete one of the names. 244 if (DelayedTypesToResolve.size() == OldSize) { 245 const std::string &Name = DelayedTypesToResolve.back(); 246 247 const Type *T1 = SrcST->lookupType(Name); 248 const Type *T2 = DestST->lookupType(Name); 249 std::cerr << "WARNING: Type conflict between types named '" << Name 250 << "'.\n Src='"; 251 WriteTypeSymbolic(std::cerr, T1, Src); 252 std::cerr << "'.\n Dest='"; 253 WriteTypeSymbolic(std::cerr, T2, Dest); 254 std::cerr << "'\n"; 255 256 // Remove the symbol name from the destination. 257 DelayedTypesToResolve.pop_back(); 258 } 259 } 260 } 261 262 263 return false; 264} 265 266static void PrintMap(const std::map<const Value*, Value*> &M) { 267 for (std::map<const Value*, Value*>::const_iterator I = M.begin(), E =M.end(); 268 I != E; ++I) { 269 std::cerr << " Fr: " << (void*)I->first << " "; 270 I->first->dump(); 271 std::cerr << " To: " << (void*)I->second << " "; 272 I->second->dump(); 273 std::cerr << "\n"; 274 } 275} 276 277 278// RemapOperand - Use LocalMap and GlobalMap to convert references from one 279// module to another. This is somewhat sophisticated in that it can 280// automatically handle constant references correctly as well... 281// 282static Value *RemapOperand(const Value *In, 283 std::map<const Value*, Value*> &LocalMap, 284 std::map<const Value*, Value*> *GlobalMap) { 285 std::map<const Value*,Value*>::const_iterator I = LocalMap.find(In); 286 if (I != LocalMap.end()) return I->second; 287 288 if (GlobalMap) { 289 I = GlobalMap->find(In); 290 if (I != GlobalMap->end()) return I->second; 291 } 292 293 // Check to see if it's a constant that we are interesting in transforming... 294 if (const Constant *CPV = dyn_cast<Constant>(In)) { 295 if ((!isa<DerivedType>(CPV->getType()) && !isa<ConstantExpr>(CPV)) || 296 isa<ConstantAggregateZero>(CPV)) 297 return const_cast<Constant*>(CPV); // Simple constants stay identical... 298 299 Constant *Result = 0; 300 301 if (const ConstantArray *CPA = dyn_cast<ConstantArray>(CPV)) { 302 std::vector<Constant*> Operands(CPA->getNumOperands()); 303 for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i) 304 Operands[i] = 305 cast<Constant>(RemapOperand(CPA->getOperand(i), LocalMap, GlobalMap)); 306 Result = ConstantArray::get(cast<ArrayType>(CPA->getType()), Operands); 307 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(CPV)) { 308 std::vector<Constant*> Operands(CPS->getNumOperands()); 309 for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i) 310 Operands[i] = 311 cast<Constant>(RemapOperand(CPS->getOperand(i), LocalMap, GlobalMap)); 312 Result = ConstantStruct::get(cast<StructType>(CPS->getType()), Operands); 313 } else if (isa<ConstantPointerNull>(CPV) || isa<UndefValue>(CPV)) { 314 Result = const_cast<Constant*>(CPV); 315 } else if (isa<GlobalValue>(CPV)) { 316 Result = cast<Constant>(RemapOperand(CPV, LocalMap, GlobalMap)); 317 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) { 318 if (CE->getOpcode() == Instruction::GetElementPtr) { 319 Value *Ptr = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap); 320 std::vector<Constant*> Indices; 321 Indices.reserve(CE->getNumOperands()-1); 322 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 323 Indices.push_back(cast<Constant>(RemapOperand(CE->getOperand(i), 324 LocalMap, GlobalMap))); 325 326 Result = ConstantExpr::getGetElementPtr(cast<Constant>(Ptr), Indices); 327 } else if (CE->getNumOperands() == 1) { 328 // Cast instruction 329 assert(CE->getOpcode() == Instruction::Cast); 330 Value *V = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap); 331 Result = ConstantExpr::getCast(cast<Constant>(V), CE->getType()); 332 } else if (CE->getNumOperands() == 3) { 333 // Select instruction 334 assert(CE->getOpcode() == Instruction::Select); 335 Value *V1 = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap); 336 Value *V2 = RemapOperand(CE->getOperand(1), LocalMap, GlobalMap); 337 Value *V3 = RemapOperand(CE->getOperand(2), LocalMap, GlobalMap); 338 Result = ConstantExpr::getSelect(cast<Constant>(V1), cast<Constant>(V2), 339 cast<Constant>(V3)); 340 } else if (CE->getNumOperands() == 2) { 341 // Binary operator... 342 Value *V1 = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap); 343 Value *V2 = RemapOperand(CE->getOperand(1), LocalMap, GlobalMap); 344 345 Result = ConstantExpr::get(CE->getOpcode(), cast<Constant>(V1), 346 cast<Constant>(V2)); 347 } else { 348 assert(0 && "Unknown constant expr type!"); 349 } 350 351 } else { 352 assert(0 && "Unknown type of derived type constant value!"); 353 } 354 355 // Cache the mapping in our local map structure... 356 if (GlobalMap) 357 GlobalMap->insert(std::make_pair(In, Result)); 358 else 359 LocalMap.insert(std::make_pair(In, Result)); 360 return Result; 361 } 362 363 std::cerr << "XXX LocalMap: \n"; 364 PrintMap(LocalMap); 365 366 if (GlobalMap) { 367 std::cerr << "XXX GlobalMap: \n"; 368 PrintMap(*GlobalMap); 369 } 370 371 std::cerr << "Couldn't remap value: " << (void*)In << " " << *In << "\n"; 372 assert(0 && "Couldn't remap value!"); 373 return 0; 374} 375 376/// ForceRenaming - The LLVM SymbolTable class autorenames globals that conflict 377/// in the symbol table. This is good for all clients except for us. Go 378/// through the trouble to force this back. 379static void ForceRenaming(GlobalValue *GV, const std::string &Name) { 380 assert(GV->getName() != Name && "Can't force rename to self"); 381 SymbolTable &ST = GV->getParent()->getSymbolTable(); 382 383 // If there is a conflict, rename the conflict. 384 Value *ConflictVal = ST.lookup(GV->getType(), Name); 385 assert(ConflictVal&&"Why do we have to force rename if there is no conflic?"); 386 GlobalValue *ConflictGV = cast<GlobalValue>(ConflictVal); 387 assert(ConflictGV->hasInternalLinkage() && 388 "Not conflicting with a static global, should link instead!"); 389 390 ConflictGV->setName(""); // Eliminate the conflict 391 GV->setName(Name); // Force the name back 392 ConflictGV->setName(Name); // This will cause ConflictGV to get renamed 393 assert(GV->getName() == Name && ConflictGV->getName() != Name && 394 "ForceRenaming didn't work"); 395} 396 397 398// LinkGlobals - Loop through the global variables in the src module and merge 399// them into the dest module. 400// 401static bool LinkGlobals(Module *Dest, const Module *Src, 402 std::map<const Value*, Value*> &ValueMap, 403 std::multimap<std::string, GlobalVariable *> &AppendingVars, 404 std::map<std::string, GlobalValue*> &GlobalsByName, 405 std::string *Err) { 406 // We will need a module level symbol table if the src module has a module 407 // level symbol table... 408 SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable(); 409 410 // Loop over all of the globals in the src module, mapping them over as we go 411 // 412 for (Module::const_giterator I = Src->gbegin(), E = Src->gend(); I != E; ++I){ 413 const GlobalVariable *SGV = I; 414 GlobalVariable *DGV = 0; 415 // Check to see if may have to link the global. 416 if (SGV->hasName() && !SGV->hasInternalLinkage()) 417 if (!(DGV = Dest->getGlobalVariable(SGV->getName(), 418 SGV->getType()->getElementType()))) { 419 std::map<std::string, GlobalValue*>::iterator EGV = 420 GlobalsByName.find(SGV->getName()); 421 if (EGV != GlobalsByName.end()) 422 DGV = dyn_cast<GlobalVariable>(EGV->second); 423 if (DGV && RecursiveResolveTypes(SGV->getType(), DGV->getType(), ST, "")) 424 DGV = 0; // FIXME: gross. 425 } 426 427 assert(SGV->hasInitializer() || SGV->hasExternalLinkage() && 428 "Global must either be external or have an initializer!"); 429 430 bool SGExtern = SGV->isExternal(); 431 bool DGExtern = DGV ? DGV->isExternal() : false; 432 433 if (!DGV || DGV->hasInternalLinkage() || SGV->hasInternalLinkage()) { 434 // No linking to be performed, simply create an identical version of the 435 // symbol over in the dest module... the initializer will be filled in 436 // later by LinkGlobalInits... 437 // 438 GlobalVariable *NewDGV = 439 new GlobalVariable(SGV->getType()->getElementType(), 440 SGV->isConstant(), SGV->getLinkage(), /*init*/0, 441 SGV->getName(), Dest); 442 443 // If the LLVM runtime renamed the global, but it is an externally visible 444 // symbol, DGV must be an existing global with internal linkage. Rename 445 // it. 446 if (NewDGV->getName() != SGV->getName() && !NewDGV->hasInternalLinkage()) 447 ForceRenaming(NewDGV, SGV->getName()); 448 449 // Make sure to remember this mapping... 450 ValueMap.insert(std::make_pair(SGV, NewDGV)); 451 if (SGV->hasAppendingLinkage()) 452 // Keep track that this is an appending variable... 453 AppendingVars.insert(std::make_pair(SGV->getName(), NewDGV)); 454 455 } else if (SGV->isExternal()) { 456 // If SGV is external or if both SGV & DGV are external.. Just link the 457 // external globals, we aren't adding anything. 458 ValueMap.insert(std::make_pair(SGV, DGV)); 459 460 // Inherit 'const' information. 461 if (SGV->isConstant()) DGV->setConstant(true); 462 463 } else if (DGV->isExternal()) { // If DGV is external but SGV is not... 464 ValueMap.insert(std::make_pair(SGV, DGV)); 465 DGV->setLinkage(SGV->getLinkage()); // Inherit linkage! 466 467 if (DGV->isConstant() && !SGV->isConstant()) 468 return Error(Err, "Linking globals named '" + SGV->getName() + 469 "': declaration is const but definition is not!"); 470 471 // Inherit 'const' information. 472 if (SGV->isConstant()) DGV->setConstant(true); 473 474 } else if (SGV->hasWeakLinkage() || SGV->hasLinkOnceLinkage()) { 475 // At this point we know that DGV has LinkOnce, Appending, Weak, or 476 // External linkage. If DGV is Appending, this is an error. 477 if (DGV->hasAppendingLinkage()) 478 return Error(Err, "Linking globals named '" + SGV->getName() + 479 "' with 'weak' and 'appending' linkage is not allowed!"); 480 481 if (SGV->isConstant() != DGV->isConstant()) 482 return Error(Err, "Global Variable Collision on '" + 483 ToStr(SGV->getType(), Src) + " %" + SGV->getName() + 484 "' - Global variables differ in const'ness"); 485 486 // Otherwise, just perform the link. 487 ValueMap.insert(std::make_pair(SGV, DGV)); 488 489 // Linkonce+Weak = Weak 490 if (DGV->hasLinkOnceLinkage() && SGV->hasWeakLinkage()) 491 DGV->setLinkage(SGV->getLinkage()); 492 493 } else if (DGV->hasWeakLinkage() || DGV->hasLinkOnceLinkage()) { 494 // At this point we know that SGV has LinkOnce, Appending, or External 495 // linkage. If SGV is Appending, this is an error. 496 if (SGV->hasAppendingLinkage()) 497 return Error(Err, "Linking globals named '" + SGV->getName() + 498 " ' with 'weak' and 'appending' linkage is not allowed!"); 499 500 if (SGV->isConstant() != DGV->isConstant()) 501 return Error(Err, "Global Variable Collision on '" + 502 ToStr(SGV->getType(), Src) + " %" + SGV->getName() + 503 "' - Global variables differ in const'ness"); 504 505 if (!SGV->hasLinkOnceLinkage()) 506 DGV->setLinkage(SGV->getLinkage()); // Inherit linkage! 507 ValueMap.insert(std::make_pair(SGV, DGV)); 508 509 } else if (SGV->getLinkage() != DGV->getLinkage()) { 510 return Error(Err, "Global variables named '" + SGV->getName() + 511 "' have different linkage specifiers!"); 512 // Inherit 'const' information. 513 if (SGV->isConstant()) DGV->setConstant(true); 514 515 } else if (SGV->hasExternalLinkage()) { 516 // Allow linking two exactly identical external global variables... 517 if (SGV->isConstant() != DGV->isConstant()) 518 return Error(Err, "Global Variable Collision on '" + 519 ToStr(SGV->getType(), Src) + " %" + SGV->getName() + 520 "' - Global variables differ in const'ness"); 521 522 if (SGV->getInitializer() != DGV->getInitializer()) 523 return Error(Err, "Global Variable Collision on '" + 524 ToStr(SGV->getType(), Src) + " %" + SGV->getName() + 525 "' - External linkage globals have different initializers"); 526 527 ValueMap.insert(std::make_pair(SGV, DGV)); 528 } else if (SGV->hasAppendingLinkage()) { 529 // No linking is performed yet. Just insert a new copy of the global, and 530 // keep track of the fact that it is an appending variable in the 531 // AppendingVars map. The name is cleared out so that no linkage is 532 // performed. 533 GlobalVariable *NewDGV = 534 new GlobalVariable(SGV->getType()->getElementType(), 535 SGV->isConstant(), SGV->getLinkage(), /*init*/0, 536 "", Dest); 537 538 // Make sure to remember this mapping... 539 ValueMap.insert(std::make_pair(SGV, NewDGV)); 540 541 // Keep track that this is an appending variable... 542 AppendingVars.insert(std::make_pair(SGV->getName(), NewDGV)); 543 } else { 544 assert(0 && "Unknown linkage!"); 545 } 546 } 547 return false; 548} 549 550 551// LinkGlobalInits - Update the initializers in the Dest module now that all 552// globals that may be referenced are in Dest. 553// 554static bool LinkGlobalInits(Module *Dest, const Module *Src, 555 std::map<const Value*, Value*> &ValueMap, 556 std::string *Err) { 557 558 // Loop over all of the globals in the src module, mapping them over as we go 559 // 560 for (Module::const_giterator I = Src->gbegin(), E = Src->gend(); I != E; ++I){ 561 const GlobalVariable *SGV = I; 562 563 if (SGV->hasInitializer()) { // Only process initialized GV's 564 // Figure out what the initializer looks like in the dest module... 565 Constant *SInit = 566 cast<Constant>(RemapOperand(SGV->getInitializer(), ValueMap, 0)); 567 568 GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[SGV]); 569 if (DGV->hasInitializer()) { 570 if (SGV->hasExternalLinkage()) { 571 if (DGV->getInitializer() != SInit) 572 return Error(Err, "Global Variable Collision on '" + 573 ToStr(SGV->getType(), Src) +"':%"+SGV->getName()+ 574 " - Global variables have different initializers"); 575 } else if (DGV->hasLinkOnceLinkage() || DGV->hasWeakLinkage()) { 576 // Nothing is required, mapped values will take the new global 577 // automatically. 578 } else if (SGV->hasLinkOnceLinkage() || SGV->hasWeakLinkage()) { 579 // Nothing is required, mapped values will take the new global 580 // automatically. 581 } else if (DGV->hasAppendingLinkage()) { 582 assert(0 && "Appending linkage unimplemented!"); 583 } else { 584 assert(0 && "Unknown linkage!"); 585 } 586 } else { 587 // Copy the initializer over now... 588 DGV->setInitializer(SInit); 589 } 590 } 591 } 592 return false; 593} 594 595// LinkFunctionProtos - Link the functions together between the two modules, 596// without doing function bodies... this just adds external function prototypes 597// to the Dest function... 598// 599static bool LinkFunctionProtos(Module *Dest, const Module *Src, 600 std::map<const Value*, Value*> &ValueMap, 601 std::map<std::string, GlobalValue*> &GlobalsByName, 602 std::string *Err) { 603 SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable(); 604 605 // Loop over all of the functions in the src module, mapping them over as we 606 // go 607 // 608 for (Module::const_iterator I = Src->begin(), E = Src->end(); I != E; ++I) { 609 const Function *SF = I; // SrcFunction 610 Function *DF = 0; 611 if (SF->hasName() && !SF->hasInternalLinkage()) { 612 // Check to see if may have to link the function. 613 if (!(DF = Dest->getFunction(SF->getName(), SF->getFunctionType()))) { 614 std::map<std::string, GlobalValue*>::iterator EF = 615 GlobalsByName.find(SF->getName()); 616 if (EF != GlobalsByName.end()) 617 DF = dyn_cast<Function>(EF->second); 618 if (DF && RecursiveResolveTypes(SF->getType(), DF->getType(), ST, "")) 619 DF = 0; // FIXME: gross. 620 } 621 } 622 623 if (!DF || SF->hasInternalLinkage() || DF->hasInternalLinkage()) { 624 // Function does not already exist, simply insert an function signature 625 // identical to SF into the dest module... 626 Function *NewDF = new Function(SF->getFunctionType(), SF->getLinkage(), 627 SF->getName(), Dest); 628 629 // If the LLVM runtime renamed the function, but it is an externally 630 // visible symbol, DF must be an existing function with internal linkage. 631 // Rename it. 632 if (NewDF->getName() != SF->getName() && !NewDF->hasInternalLinkage()) 633 ForceRenaming(NewDF, SF->getName()); 634 635 // ... and remember this mapping... 636 ValueMap.insert(std::make_pair(SF, NewDF)); 637 } else if (SF->isExternal()) { 638 // If SF is external or if both SF & DF are external.. Just link the 639 // external functions, we aren't adding anything. 640 ValueMap.insert(std::make_pair(SF, DF)); 641 } else if (DF->isExternal()) { // If DF is external but SF is not... 642 // Link the external functions, update linkage qualifiers 643 ValueMap.insert(std::make_pair(SF, DF)); 644 DF->setLinkage(SF->getLinkage()); 645 646 } else if (SF->hasWeakLinkage() || SF->hasLinkOnceLinkage()) { 647 // At this point we know that DF has LinkOnce, Weak, or External linkage. 648 ValueMap.insert(std::make_pair(SF, DF)); 649 650 // Linkonce+Weak = Weak 651 if (DF->hasLinkOnceLinkage() && SF->hasWeakLinkage()) 652 DF->setLinkage(SF->getLinkage()); 653 654 } else if (DF->hasWeakLinkage() || DF->hasLinkOnceLinkage()) { 655 // At this point we know that SF has LinkOnce or External linkage. 656 ValueMap.insert(std::make_pair(SF, DF)); 657 if (!SF->hasLinkOnceLinkage()) // Don't inherit linkonce linkage 658 DF->setLinkage(SF->getLinkage()); 659 660 } else if (SF->getLinkage() != DF->getLinkage()) { 661 return Error(Err, "Functions named '" + SF->getName() + 662 "' have different linkage specifiers!"); 663 } else if (SF->hasExternalLinkage()) { 664 // The function is defined in both modules!! 665 return Error(Err, "Function '" + 666 ToStr(SF->getFunctionType(), Src) + "':\"" + 667 SF->getName() + "\" - Function is already defined!"); 668 } else { 669 assert(0 && "Unknown linkage configuration found!"); 670 } 671 } 672 return false; 673} 674 675// LinkFunctionBody - Copy the source function over into the dest function and 676// fix up references to values. At this point we know that Dest is an external 677// function, and that Src is not. 678// 679static bool LinkFunctionBody(Function *Dest, const Function *Src, 680 std::map<const Value*, Value*> &GlobalMap, 681 std::string *Err) { 682 assert(Src && Dest && Dest->isExternal() && !Src->isExternal()); 683 std::map<const Value*, Value*> LocalMap; // Map for function local values 684 685 // Go through and convert function arguments over... 686 Function::aiterator DI = Dest->abegin(); 687 for (Function::const_aiterator I = Src->abegin(), E = Src->aend(); 688 I != E; ++I, ++DI) { 689 DI->setName(I->getName()); // Copy the name information over... 690 691 // Add a mapping to our local map 692 LocalMap.insert(std::make_pair(I, DI)); 693 } 694 695 // Loop over all of the basic blocks, copying the instructions over... 696 // 697 for (Function::const_iterator I = Src->begin(), E = Src->end(); I != E; ++I) { 698 // Create new basic block and add to mapping and the Dest function... 699 BasicBlock *DBB = new BasicBlock(I->getName(), Dest); 700 LocalMap.insert(std::make_pair(I, DBB)); 701 702 // Loop over all of the instructions in the src basic block, copying them 703 // over. Note that this is broken in a strict sense because the cloned 704 // instructions will still be referencing values in the Src module, not 705 // the remapped values. In our case, however, we will not get caught and 706 // so we can delay patching the values up until later... 707 // 708 for (BasicBlock::const_iterator II = I->begin(), IE = I->end(); 709 II != IE; ++II) { 710 Instruction *DI = II->clone(); 711 DI->setName(II->getName()); 712 DBB->getInstList().push_back(DI); 713 LocalMap.insert(std::make_pair(II, DI)); 714 } 715 } 716 717 // At this point, all of the instructions and values of the function are now 718 // copied over. The only problem is that they are still referencing values in 719 // the Source function as operands. Loop through all of the operands of the 720 // functions and patch them up to point to the local versions... 721 // 722 for (Function::iterator BB = Dest->begin(), BE = Dest->end(); BB != BE; ++BB) 723 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 724 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 725 OI != OE; ++OI) 726 *OI = RemapOperand(*OI, LocalMap, &GlobalMap); 727 728 return false; 729} 730 731 732// LinkFunctionBodies - Link in the function bodies that are defined in the 733// source module into the DestModule. This consists basically of copying the 734// function over and fixing up references to values. 735// 736static bool LinkFunctionBodies(Module *Dest, const Module *Src, 737 std::map<const Value*, Value*> &ValueMap, 738 std::string *Err) { 739 740 // Loop over all of the functions in the src module, mapping them over as we 741 // go 742 // 743 for (Module::const_iterator SF = Src->begin(), E = Src->end(); SF != E; ++SF){ 744 if (!SF->isExternal()) { // No body if function is external 745 Function *DF = cast<Function>(ValueMap[SF]); // Destination function 746 747 // DF not external SF external? 748 if (DF->isExternal()) { 749 // Only provide the function body if there isn't one already. 750 if (LinkFunctionBody(DF, SF, ValueMap, Err)) 751 return true; 752 } 753 } 754 } 755 return false; 756} 757 758// LinkAppendingVars - If there were any appending global variables, link them 759// together now. Return true on error. 760// 761static bool LinkAppendingVars(Module *M, 762 std::multimap<std::string, GlobalVariable *> &AppendingVars, 763 std::string *ErrorMsg) { 764 if (AppendingVars.empty()) return false; // Nothing to do. 765 766 // Loop over the multimap of appending vars, processing any variables with the 767 // same name, forming a new appending global variable with both of the 768 // initializers merged together, then rewrite references to the old variables 769 // and delete them. 770 // 771 std::vector<Constant*> Inits; 772 while (AppendingVars.size() > 1) { 773 // Get the first two elements in the map... 774 std::multimap<std::string, 775 GlobalVariable*>::iterator Second = AppendingVars.begin(), First=Second++; 776 777 // If the first two elements are for different names, there is no pair... 778 // Otherwise there is a pair, so link them together... 779 if (First->first == Second->first) { 780 GlobalVariable *G1 = First->second, *G2 = Second->second; 781 const ArrayType *T1 = cast<ArrayType>(G1->getType()->getElementType()); 782 const ArrayType *T2 = cast<ArrayType>(G2->getType()->getElementType()); 783 784 // Check to see that they two arrays agree on type... 785 if (T1->getElementType() != T2->getElementType()) 786 return Error(ErrorMsg, 787 "Appending variables with different element types need to be linked!"); 788 if (G1->isConstant() != G2->isConstant()) 789 return Error(ErrorMsg, 790 "Appending variables linked with different const'ness!"); 791 792 unsigned NewSize = T1->getNumElements() + T2->getNumElements(); 793 ArrayType *NewType = ArrayType::get(T1->getElementType(), NewSize); 794 795 // Create the new global variable... 796 GlobalVariable *NG = 797 new GlobalVariable(NewType, G1->isConstant(), G1->getLinkage(), 798 /*init*/0, First->first, M); 799 800 // Merge the initializer... 801 Inits.reserve(NewSize); 802 if (ConstantArray *I = dyn_cast<ConstantArray>(G1->getInitializer())) { 803 for (unsigned i = 0, e = T1->getNumElements(); i != e; ++i) 804 Inits.push_back(I->getOperand(i)); 805 } else { 806 assert(isa<ConstantAggregateZero>(G1->getInitializer())); 807 Constant *CV = Constant::getNullValue(T1->getElementType()); 808 for (unsigned i = 0, e = T1->getNumElements(); i != e; ++i) 809 Inits.push_back(CV); 810 } 811 if (ConstantArray *I = dyn_cast<ConstantArray>(G2->getInitializer())) { 812 for (unsigned i = 0, e = T2->getNumElements(); i != e; ++i) 813 Inits.push_back(I->getOperand(i)); 814 } else { 815 assert(isa<ConstantAggregateZero>(G2->getInitializer())); 816 Constant *CV = Constant::getNullValue(T2->getElementType()); 817 for (unsigned i = 0, e = T2->getNumElements(); i != e; ++i) 818 Inits.push_back(CV); 819 } 820 NG->setInitializer(ConstantArray::get(NewType, Inits)); 821 Inits.clear(); 822 823 // Replace any uses of the two global variables with uses of the new 824 // global... 825 826 // FIXME: This should rewrite simple/straight-forward uses such as 827 // getelementptr instructions to not use the Cast! 828 G1->replaceAllUsesWith(ConstantExpr::getCast(NG, G1->getType())); 829 G2->replaceAllUsesWith(ConstantExpr::getCast(NG, G2->getType())); 830 831 // Remove the two globals from the module now... 832 M->getGlobalList().erase(G1); 833 M->getGlobalList().erase(G2); 834 835 // Put the new global into the AppendingVars map so that we can handle 836 // linking of more than two vars... 837 Second->second = NG; 838 } 839 AppendingVars.erase(First); 840 } 841 842 return false; 843} 844 845 846// LinkModules - This function links two modules together, with the resulting 847// left module modified to be the composite of the two input modules. If an 848// error occurs, true is returned and ErrorMsg (if not null) is set to indicate 849// the problem. Upon failure, the Dest module could be in a modified state, and 850// shouldn't be relied on to be consistent. 851// 852bool llvm::LinkModules(Module *Dest, const Module *Src, std::string *ErrorMsg) { 853 assert(Dest != 0 && "Invalid Destination module"); 854 assert(Src != 0 && "Invalid Source Module"); 855 856 if (Dest->getEndianness() == Module::AnyEndianness) 857 Dest->setEndianness(Src->getEndianness()); 858 if (Dest->getPointerSize() == Module::AnyPointerSize) 859 Dest->setPointerSize(Src->getPointerSize()); 860 861 if (Src->getEndianness() != Module::AnyEndianness && 862 Dest->getEndianness() != Src->getEndianness()) 863 std::cerr << "WARNING: Linking two modules of different endianness!\n"; 864 if (Src->getPointerSize() != Module::AnyPointerSize && 865 Dest->getPointerSize() != Src->getPointerSize()) 866 std::cerr << "WARNING: Linking two modules of different pointer size!\n"; 867 868 // Update the destination module's dependent libraries list with the libraries 869 // from the source module. There's no opportunity for duplicates here as the 870 // Module ensures that duplicate insertions are discarded. 871 Module::lib_iterator SI = Src->lib_begin(); 872 Module::lib_iterator SE = Src->lib_end(); 873 while ( SI != SE ) { 874 Dest->addLibrary(*SI); 875 ++SI; 876 } 877 878 // LinkTypes - Go through the symbol table of the Src module and see if any 879 // types are named in the src module that are not named in the Dst module. 880 // Make sure there are no type name conflicts. 881 // 882 if (LinkTypes(Dest, Src, ErrorMsg)) return true; 883 884 // ValueMap - Mapping of values from what they used to be in Src, to what they 885 // are now in Dest. 886 // 887 std::map<const Value*, Value*> ValueMap; 888 889 // AppendingVars - Keep track of global variables in the destination module 890 // with appending linkage. After the module is linked together, they are 891 // appended and the module is rewritten. 892 // 893 std::multimap<std::string, GlobalVariable *> AppendingVars; 894 895 // GlobalsByName - The LLVM SymbolTable class fights our best efforts at 896 // linking by separating globals by type. Until PR411 is fixed, we replicate 897 // it's functionality here. 898 std::map<std::string, GlobalValue*> GlobalsByName; 899 900 for (Module::giterator I = Dest->gbegin(), E = Dest->gend(); I != E; ++I) { 901 // Add all of the appending globals already in the Dest module to 902 // AppendingVars. 903 if (I->hasAppendingLinkage()) 904 AppendingVars.insert(std::make_pair(I->getName(), I)); 905 906 // Keep track of all globals by name. 907 if (!I->hasInternalLinkage() && I->hasName()) 908 GlobalsByName[I->getName()] = I; 909 } 910 911 // Keep track of all globals by name. 912 for (Module::iterator I = Dest->begin(), E = Dest->end(); I != E; ++I) 913 if (!I->hasInternalLinkage() && I->hasName()) 914 GlobalsByName[I->getName()] = I; 915 916 // Insert all of the globals in src into the Dest module... without linking 917 // initializers (which could refer to functions not yet mapped over). 918 // 919 if (LinkGlobals(Dest, Src, ValueMap, AppendingVars, GlobalsByName, ErrorMsg)) 920 return true; 921 922 // Link the functions together between the two modules, without doing function 923 // bodies... this just adds external function prototypes to the Dest 924 // function... We do this so that when we begin processing function bodies, 925 // all of the global values that may be referenced are available in our 926 // ValueMap. 927 // 928 if (LinkFunctionProtos(Dest, Src, ValueMap, GlobalsByName, ErrorMsg)) 929 return true; 930 931 // Update the initializers in the Dest module now that all globals that may 932 // be referenced are in Dest. 933 // 934 if (LinkGlobalInits(Dest, Src, ValueMap, ErrorMsg)) return true; 935 936 // Link in the function bodies that are defined in the source module into the 937 // DestModule. This consists basically of copying the function over and 938 // fixing up references to values. 939 // 940 if (LinkFunctionBodies(Dest, Src, ValueMap, ErrorMsg)) return true; 941 942 // If there were any appending global variables, link them together now. 943 // 944 if (LinkAppendingVars(Dest, AppendingVars, ErrorMsg)) return true; 945 946 // If the source library's module id is in the dependent library list of the 947 // destination library, remove it since that module is now linked in. 948 sys::Path modId; 949 modId.set_file(Src->getModuleIdentifier()); 950 if (!modId.is_empty()) 951 Dest->removeLibrary(modId.get_basename()); 952 953 return false; 954} 955 956// vim: sw=2 957