1 //===- lib/Linker/IRMover.cpp ---------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "llvm/Linker/IRMover.h" 10 #include "LinkDiagnosticInfo.h" 11 #include "llvm/ADT/SetVector.h" 12 #include "llvm/ADT/SmallString.h" 13 #include "llvm/ADT/Triple.h" 14 #include "llvm/IR/Constants.h" 15 #include "llvm/IR/DebugInfo.h" 16 #include "llvm/IR/DiagnosticPrinter.h" 17 #include "llvm/IR/GVMaterializer.h" 18 #include "llvm/IR/Intrinsics.h" 19 #include "llvm/IR/TypeFinder.h" 20 #include "llvm/Object/ModuleSymbolTable.h" 21 #include "llvm/Support/Error.h" 22 #include "llvm/Transforms/Utils/Cloning.h" 23 #include <utility> 24 using namespace llvm; 25 26 //===----------------------------------------------------------------------===// 27 // TypeMap implementation. 28 //===----------------------------------------------------------------------===// 29 30 namespace { 31 class TypeMapTy : public ValueMapTypeRemapper { 32 /// This is a mapping from a source type to a destination type to use. 33 DenseMap<Type *, Type *> MappedTypes; 34 35 /// When checking to see if two subgraphs are isomorphic, we speculatively 36 /// add types to MappedTypes, but keep track of them here in case we need to 37 /// roll back. 38 SmallVector<Type *, 16> SpeculativeTypes; 39 40 SmallVector<StructType *, 16> SpeculativeDstOpaqueTypes; 41 42 /// This is a list of non-opaque structs in the source module that are mapped 43 /// to an opaque struct in the destination module. 44 SmallVector<StructType *, 16> SrcDefinitionsToResolve; 45 46 /// This is the set of opaque types in the destination modules who are 47 /// getting a body from the source module. 48 SmallPtrSet<StructType *, 16> DstResolvedOpaqueTypes; 49 50 public: 51 TypeMapTy(IRMover::IdentifiedStructTypeSet &DstStructTypesSet) 52 : DstStructTypesSet(DstStructTypesSet) {} 53 54 IRMover::IdentifiedStructTypeSet &DstStructTypesSet; 55 /// Indicate that the specified type in the destination module is conceptually 56 /// equivalent to the specified type in the source module. 57 void addTypeMapping(Type *DstTy, Type *SrcTy); 58 59 /// Produce a body for an opaque type in the dest module from a type 60 /// definition in the source module. 61 void linkDefinedTypeBodies(); 62 63 /// Return the mapped type to use for the specified input type from the 64 /// source module. 65 Type *get(Type *SrcTy); 66 Type *get(Type *SrcTy, SmallPtrSet<StructType *, 8> &Visited); 67 68 void finishType(StructType *DTy, StructType *STy, ArrayRef<Type *> ETypes); 69 70 FunctionType *get(FunctionType *T) { 71 return cast<FunctionType>(get((Type *)T)); 72 } 73 74 private: 75 Type *remapType(Type *SrcTy) override { return get(SrcTy); } 76 77 bool areTypesIsomorphic(Type *DstTy, Type *SrcTy); 78 }; 79 } 80 81 void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) { 82 assert(SpeculativeTypes.empty()); 83 assert(SpeculativeDstOpaqueTypes.empty()); 84 85 // Check to see if these types are recursively isomorphic and establish a 86 // mapping between them if so. 87 if (!areTypesIsomorphic(DstTy, SrcTy)) { 88 // Oops, they aren't isomorphic. Just discard this request by rolling out 89 // any speculative mappings we've established. 90 for (Type *Ty : SpeculativeTypes) 91 MappedTypes.erase(Ty); 92 93 SrcDefinitionsToResolve.resize(SrcDefinitionsToResolve.size() - 94 SpeculativeDstOpaqueTypes.size()); 95 for (StructType *Ty : SpeculativeDstOpaqueTypes) 96 DstResolvedOpaqueTypes.erase(Ty); 97 } else { 98 // SrcTy and DstTy are recursively ismorphic. We clear names of SrcTy 99 // and all its descendants to lower amount of renaming in LLVM context 100 // Renaming occurs because we load all source modules to the same context 101 // and declaration with existing name gets renamed (i.e Foo -> Foo.42). 102 // As a result we may get several different types in the destination 103 // module, which are in fact the same. 104 for (Type *Ty : SpeculativeTypes) 105 if (auto *STy = dyn_cast<StructType>(Ty)) 106 if (STy->hasName()) 107 STy->setName(""); 108 } 109 SpeculativeTypes.clear(); 110 SpeculativeDstOpaqueTypes.clear(); 111 } 112 113 /// Recursively walk this pair of types, returning true if they are isomorphic, 114 /// false if they are not. 115 bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) { 116 // Two types with differing kinds are clearly not isomorphic. 117 if (DstTy->getTypeID() != SrcTy->getTypeID()) 118 return false; 119 120 // If we have an entry in the MappedTypes table, then we have our answer. 121 Type *&Entry = MappedTypes[SrcTy]; 122 if (Entry) 123 return Entry == DstTy; 124 125 // Two identical types are clearly isomorphic. Remember this 126 // non-speculatively. 127 if (DstTy == SrcTy) { 128 Entry = DstTy; 129 return true; 130 } 131 132 // Okay, we have two types with identical kinds that we haven't seen before. 133 134 // If this is an opaque struct type, special case it. 135 if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) { 136 // Mapping an opaque type to any struct, just keep the dest struct. 137 if (SSTy->isOpaque()) { 138 Entry = DstTy; 139 SpeculativeTypes.push_back(SrcTy); 140 return true; 141 } 142 143 // Mapping a non-opaque source type to an opaque dest. If this is the first 144 // type that we're mapping onto this destination type then we succeed. Keep 145 // the dest, but fill it in later. If this is the second (different) type 146 // that we're trying to map onto the same opaque type then we fail. 147 if (cast<StructType>(DstTy)->isOpaque()) { 148 // We can only map one source type onto the opaque destination type. 149 if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)).second) 150 return false; 151 SrcDefinitionsToResolve.push_back(SSTy); 152 SpeculativeTypes.push_back(SrcTy); 153 SpeculativeDstOpaqueTypes.push_back(cast<StructType>(DstTy)); 154 Entry = DstTy; 155 return true; 156 } 157 } 158 159 // If the number of subtypes disagree between the two types, then we fail. 160 if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes()) 161 return false; 162 163 // Fail if any of the extra properties (e.g. array size) of the type disagree. 164 if (isa<IntegerType>(DstTy)) 165 return false; // bitwidth disagrees. 166 if (PointerType *PT = dyn_cast<PointerType>(DstTy)) { 167 if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace()) 168 return false; 169 } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) { 170 if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg()) 171 return false; 172 } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) { 173 StructType *SSTy = cast<StructType>(SrcTy); 174 if (DSTy->isLiteral() != SSTy->isLiteral() || 175 DSTy->isPacked() != SSTy->isPacked()) 176 return false; 177 } else if (auto *DArrTy = dyn_cast<ArrayType>(DstTy)) { 178 if (DArrTy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements()) 179 return false; 180 } else if (auto *DVecTy = dyn_cast<VectorType>(DstTy)) { 181 if (DVecTy->getElementCount() != cast<VectorType>(SrcTy)->getElementCount()) 182 return false; 183 } 184 185 // Otherwise, we speculate that these two types will line up and recursively 186 // check the subelements. 187 Entry = DstTy; 188 SpeculativeTypes.push_back(SrcTy); 189 190 for (unsigned I = 0, E = SrcTy->getNumContainedTypes(); I != E; ++I) 191 if (!areTypesIsomorphic(DstTy->getContainedType(I), 192 SrcTy->getContainedType(I))) 193 return false; 194 195 // If everything seems to have lined up, then everything is great. 196 return true; 197 } 198 199 void TypeMapTy::linkDefinedTypeBodies() { 200 SmallVector<Type *, 16> Elements; 201 for (StructType *SrcSTy : SrcDefinitionsToResolve) { 202 StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]); 203 assert(DstSTy->isOpaque()); 204 205 // Map the body of the source type over to a new body for the dest type. 206 Elements.resize(SrcSTy->getNumElements()); 207 for (unsigned I = 0, E = Elements.size(); I != E; ++I) 208 Elements[I] = get(SrcSTy->getElementType(I)); 209 210 DstSTy->setBody(Elements, SrcSTy->isPacked()); 211 DstStructTypesSet.switchToNonOpaque(DstSTy); 212 } 213 SrcDefinitionsToResolve.clear(); 214 DstResolvedOpaqueTypes.clear(); 215 } 216 217 void TypeMapTy::finishType(StructType *DTy, StructType *STy, 218 ArrayRef<Type *> ETypes) { 219 DTy->setBody(ETypes, STy->isPacked()); 220 221 // Steal STy's name. 222 if (STy->hasName()) { 223 SmallString<16> TmpName = STy->getName(); 224 STy->setName(""); 225 DTy->setName(TmpName); 226 } 227 228 DstStructTypesSet.addNonOpaque(DTy); 229 } 230 231 Type *TypeMapTy::get(Type *Ty) { 232 SmallPtrSet<StructType *, 8> Visited; 233 return get(Ty, Visited); 234 } 235 236 Type *TypeMapTy::get(Type *Ty, SmallPtrSet<StructType *, 8> &Visited) { 237 // If we already have an entry for this type, return it. 238 Type **Entry = &MappedTypes[Ty]; 239 if (*Entry) 240 return *Entry; 241 242 // These are types that LLVM itself will unique. 243 bool IsUniqued = !isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral(); 244 245 if (!IsUniqued) { 246 #ifndef NDEBUG 247 for (auto &Pair : MappedTypes) { 248 assert(!(Pair.first != Ty && Pair.second == Ty) && 249 "mapping to a source type"); 250 } 251 #endif 252 253 if (!Visited.insert(cast<StructType>(Ty)).second) { 254 StructType *DTy = StructType::create(Ty->getContext()); 255 return *Entry = DTy; 256 } 257 } 258 259 // If this is not a recursive type, then just map all of the elements and 260 // then rebuild the type from inside out. 261 SmallVector<Type *, 4> ElementTypes; 262 263 // If there are no element types to map, then the type is itself. This is 264 // true for the anonymous {} struct, things like 'float', integers, etc. 265 if (Ty->getNumContainedTypes() == 0 && IsUniqued) 266 return *Entry = Ty; 267 268 // Remap all of the elements, keeping track of whether any of them change. 269 bool AnyChange = false; 270 ElementTypes.resize(Ty->getNumContainedTypes()); 271 for (unsigned I = 0, E = Ty->getNumContainedTypes(); I != E; ++I) { 272 ElementTypes[I] = get(Ty->getContainedType(I), Visited); 273 AnyChange |= ElementTypes[I] != Ty->getContainedType(I); 274 } 275 276 // If we found our type while recursively processing stuff, just use it. 277 Entry = &MappedTypes[Ty]; 278 if (*Entry) { 279 if (auto *DTy = dyn_cast<StructType>(*Entry)) { 280 if (DTy->isOpaque()) { 281 auto *STy = cast<StructType>(Ty); 282 finishType(DTy, STy, ElementTypes); 283 } 284 } 285 return *Entry; 286 } 287 288 // If all of the element types mapped directly over and the type is not 289 // a named struct, then the type is usable as-is. 290 if (!AnyChange && IsUniqued) 291 return *Entry = Ty; 292 293 // Otherwise, rebuild a modified type. 294 switch (Ty->getTypeID()) { 295 default: 296 llvm_unreachable("unknown derived type to remap"); 297 case Type::ArrayTyID: 298 return *Entry = ArrayType::get(ElementTypes[0], 299 cast<ArrayType>(Ty)->getNumElements()); 300 case Type::ScalableVectorTyID: 301 case Type::FixedVectorTyID: 302 return *Entry = VectorType::get(ElementTypes[0], 303 cast<VectorType>(Ty)->getElementCount()); 304 case Type::PointerTyID: 305 return *Entry = PointerType::get(ElementTypes[0], 306 cast<PointerType>(Ty)->getAddressSpace()); 307 case Type::FunctionTyID: 308 return *Entry = FunctionType::get(ElementTypes[0], 309 makeArrayRef(ElementTypes).slice(1), 310 cast<FunctionType>(Ty)->isVarArg()); 311 case Type::StructTyID: { 312 auto *STy = cast<StructType>(Ty); 313 bool IsPacked = STy->isPacked(); 314 if (IsUniqued) 315 return *Entry = StructType::get(Ty->getContext(), ElementTypes, IsPacked); 316 317 // If the type is opaque, we can just use it directly. 318 if (STy->isOpaque()) { 319 DstStructTypesSet.addOpaque(STy); 320 return *Entry = Ty; 321 } 322 323 if (StructType *OldT = 324 DstStructTypesSet.findNonOpaque(ElementTypes, IsPacked)) { 325 STy->setName(""); 326 return *Entry = OldT; 327 } 328 329 if (!AnyChange) { 330 DstStructTypesSet.addNonOpaque(STy); 331 return *Entry = Ty; 332 } 333 334 StructType *DTy = StructType::create(Ty->getContext()); 335 finishType(DTy, STy, ElementTypes); 336 return *Entry = DTy; 337 } 338 } 339 } 340 341 LinkDiagnosticInfo::LinkDiagnosticInfo(DiagnosticSeverity Severity, 342 const Twine &Msg) 343 : DiagnosticInfo(DK_Linker, Severity), Msg(Msg) {} 344 void LinkDiagnosticInfo::print(DiagnosticPrinter &DP) const { DP << Msg; } 345 346 //===----------------------------------------------------------------------===// 347 // IRLinker implementation. 348 //===----------------------------------------------------------------------===// 349 350 namespace { 351 class IRLinker; 352 353 /// Creates prototypes for functions that are lazily linked on the fly. This 354 /// speeds up linking for modules with many/ lazily linked functions of which 355 /// few get used. 356 class GlobalValueMaterializer final : public ValueMaterializer { 357 IRLinker &TheIRLinker; 358 359 public: 360 GlobalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {} 361 Value *materialize(Value *V) override; 362 }; 363 364 class LocalValueMaterializer final : public ValueMaterializer { 365 IRLinker &TheIRLinker; 366 367 public: 368 LocalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {} 369 Value *materialize(Value *V) override; 370 }; 371 372 /// Type of the Metadata map in \a ValueToValueMapTy. 373 typedef DenseMap<const Metadata *, TrackingMDRef> MDMapT; 374 375 /// This is responsible for keeping track of the state used for moving data 376 /// from SrcM to DstM. 377 class IRLinker { 378 Module &DstM; 379 std::unique_ptr<Module> SrcM; 380 381 /// See IRMover::move(). 382 std::function<void(GlobalValue &, IRMover::ValueAdder)> AddLazyFor; 383 384 TypeMapTy TypeMap; 385 GlobalValueMaterializer GValMaterializer; 386 LocalValueMaterializer LValMaterializer; 387 388 /// A metadata map that's shared between IRLinker instances. 389 MDMapT &SharedMDs; 390 391 /// Mapping of values from what they used to be in Src, to what they are now 392 /// in DstM. ValueToValueMapTy is a ValueMap, which involves some overhead 393 /// due to the use of Value handles which the Linker doesn't actually need, 394 /// but this allows us to reuse the ValueMapper code. 395 ValueToValueMapTy ValueMap; 396 ValueToValueMapTy IndirectSymbolValueMap; 397 398 DenseSet<GlobalValue *> ValuesToLink; 399 std::vector<GlobalValue *> Worklist; 400 std::vector<std::pair<GlobalValue *, Value*>> RAUWWorklist; 401 402 void maybeAdd(GlobalValue *GV) { 403 if (ValuesToLink.insert(GV).second) 404 Worklist.push_back(GV); 405 } 406 407 /// Whether we are importing globals for ThinLTO, as opposed to linking the 408 /// source module. If this flag is set, it means that we can rely on some 409 /// other object file to define any non-GlobalValue entities defined by the 410 /// source module. This currently causes us to not link retained types in 411 /// debug info metadata and module inline asm. 412 bool IsPerformingImport; 413 414 /// Set to true when all global value body linking is complete (including 415 /// lazy linking). Used to prevent metadata linking from creating new 416 /// references. 417 bool DoneLinkingBodies = false; 418 419 /// The Error encountered during materialization. We use an Optional here to 420 /// avoid needing to manage an unconsumed success value. 421 Optional<Error> FoundError; 422 void setError(Error E) { 423 if (E) 424 FoundError = std::move(E); 425 } 426 427 /// Most of the errors produced by this module are inconvertible StringErrors. 428 /// This convenience function lets us return one of those more easily. 429 Error stringErr(const Twine &T) { 430 return make_error<StringError>(T, inconvertibleErrorCode()); 431 } 432 433 /// Entry point for mapping values and alternate context for mapping aliases. 434 ValueMapper Mapper; 435 unsigned IndirectSymbolMCID; 436 437 /// Handles cloning of a global values from the source module into 438 /// the destination module, including setting the attributes and visibility. 439 GlobalValue *copyGlobalValueProto(const GlobalValue *SGV, bool ForDefinition); 440 441 void emitWarning(const Twine &Message) { 442 SrcM->getContext().diagnose(LinkDiagnosticInfo(DS_Warning, Message)); 443 } 444 445 /// Given a global in the source module, return the global in the 446 /// destination module that is being linked to, if any. 447 GlobalValue *getLinkedToGlobal(const GlobalValue *SrcGV) { 448 // If the source has no name it can't link. If it has local linkage, 449 // there is no name match-up going on. 450 if (!SrcGV->hasName() || SrcGV->hasLocalLinkage()) 451 return nullptr; 452 453 // Otherwise see if we have a match in the destination module's symtab. 454 GlobalValue *DGV = DstM.getNamedValue(SrcGV->getName()); 455 if (!DGV) 456 return nullptr; 457 458 // If we found a global with the same name in the dest module, but it has 459 // internal linkage, we are really not doing any linkage here. 460 if (DGV->hasLocalLinkage()) 461 return nullptr; 462 463 // If we found an intrinsic declaration with mismatching prototypes, we 464 // probably had a nameclash. Don't use that version. 465 if (auto *FDGV = dyn_cast<Function>(DGV)) 466 if (FDGV->isIntrinsic()) 467 if (const auto *FSrcGV = dyn_cast<Function>(SrcGV)) 468 if (FDGV->getFunctionType() != TypeMap.get(FSrcGV->getFunctionType())) 469 return nullptr; 470 471 // Otherwise, we do in fact link to the destination global. 472 return DGV; 473 } 474 475 void computeTypeMapping(); 476 477 Expected<Constant *> linkAppendingVarProto(GlobalVariable *DstGV, 478 const GlobalVariable *SrcGV); 479 480 /// Given the GlobaValue \p SGV in the source module, and the matching 481 /// GlobalValue \p DGV (if any), return true if the linker will pull \p SGV 482 /// into the destination module. 483 /// 484 /// Note this code may call the client-provided \p AddLazyFor. 485 bool shouldLink(GlobalValue *DGV, GlobalValue &SGV); 486 Expected<Constant *> linkGlobalValueProto(GlobalValue *GV, 487 bool ForIndirectSymbol); 488 489 Error linkModuleFlagsMetadata(); 490 491 void linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src); 492 Error linkFunctionBody(Function &Dst, Function &Src); 493 void linkIndirectSymbolBody(GlobalIndirectSymbol &Dst, 494 GlobalIndirectSymbol &Src); 495 Error linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src); 496 497 /// Replace all types in the source AttributeList with the 498 /// corresponding destination type. 499 AttributeList mapAttributeTypes(LLVMContext &C, AttributeList Attrs); 500 501 /// Functions that take care of cloning a specific global value type 502 /// into the destination module. 503 GlobalVariable *copyGlobalVariableProto(const GlobalVariable *SGVar); 504 Function *copyFunctionProto(const Function *SF); 505 GlobalValue *copyGlobalIndirectSymbolProto(const GlobalIndirectSymbol *SGIS); 506 507 /// Perform "replace all uses with" operations. These work items need to be 508 /// performed as part of materialization, but we postpone them to happen after 509 /// materialization is done. The materializer called by ValueMapper is not 510 /// expected to delete constants, as ValueMapper is holding pointers to some 511 /// of them, but constant destruction may be indirectly triggered by RAUW. 512 /// Hence, the need to move this out of the materialization call chain. 513 void flushRAUWWorklist(); 514 515 /// When importing for ThinLTO, prevent importing of types listed on 516 /// the DICompileUnit that we don't need a copy of in the importing 517 /// module. 518 void prepareCompileUnitsForImport(); 519 void linkNamedMDNodes(); 520 521 public: 522 IRLinker(Module &DstM, MDMapT &SharedMDs, 523 IRMover::IdentifiedStructTypeSet &Set, std::unique_ptr<Module> SrcM, 524 ArrayRef<GlobalValue *> ValuesToLink, 525 std::function<void(GlobalValue &, IRMover::ValueAdder)> AddLazyFor, 526 bool IsPerformingImport) 527 : DstM(DstM), SrcM(std::move(SrcM)), AddLazyFor(std::move(AddLazyFor)), 528 TypeMap(Set), GValMaterializer(*this), LValMaterializer(*this), 529 SharedMDs(SharedMDs), IsPerformingImport(IsPerformingImport), 530 Mapper(ValueMap, RF_ReuseAndMutateDistinctMDs | RF_IgnoreMissingLocals, 531 &TypeMap, &GValMaterializer), 532 IndirectSymbolMCID(Mapper.registerAlternateMappingContext( 533 IndirectSymbolValueMap, &LValMaterializer)) { 534 ValueMap.getMDMap() = std::move(SharedMDs); 535 for (GlobalValue *GV : ValuesToLink) 536 maybeAdd(GV); 537 if (IsPerformingImport) 538 prepareCompileUnitsForImport(); 539 } 540 ~IRLinker() { SharedMDs = std::move(*ValueMap.getMDMap()); } 541 542 Error run(); 543 Value *materialize(Value *V, bool ForIndirectSymbol); 544 }; 545 } 546 547 /// The LLVM SymbolTable class autorenames globals that conflict in the symbol 548 /// table. This is good for all clients except for us. Go through the trouble 549 /// to force this back. 550 static void forceRenaming(GlobalValue *GV, StringRef Name) { 551 // If the global doesn't force its name or if it already has the right name, 552 // there is nothing for us to do. 553 if (GV->hasLocalLinkage() || GV->getName() == Name) 554 return; 555 556 Module *M = GV->getParent(); 557 558 // If there is a conflict, rename the conflict. 559 if (GlobalValue *ConflictGV = M->getNamedValue(Name)) { 560 GV->takeName(ConflictGV); 561 ConflictGV->setName(Name); // This will cause ConflictGV to get renamed 562 assert(ConflictGV->getName() != Name && "forceRenaming didn't work"); 563 } else { 564 GV->setName(Name); // Force the name back 565 } 566 } 567 568 Value *GlobalValueMaterializer::materialize(Value *SGV) { 569 return TheIRLinker.materialize(SGV, false); 570 } 571 572 Value *LocalValueMaterializer::materialize(Value *SGV) { 573 return TheIRLinker.materialize(SGV, true); 574 } 575 576 Value *IRLinker::materialize(Value *V, bool ForIndirectSymbol) { 577 auto *SGV = dyn_cast<GlobalValue>(V); 578 if (!SGV) 579 return nullptr; 580 581 // When linking a global from other modules than source & dest, skip 582 // materializing it because it would be mapped later when its containing 583 // module is linked. Linking it now would potentially pull in many types that 584 // may not be mapped properly. 585 if (SGV->getParent() != &DstM && SGV->getParent() != SrcM.get()) 586 return nullptr; 587 588 Expected<Constant *> NewProto = linkGlobalValueProto(SGV, ForIndirectSymbol); 589 if (!NewProto) { 590 setError(NewProto.takeError()); 591 return nullptr; 592 } 593 if (!*NewProto) 594 return nullptr; 595 596 GlobalValue *New = dyn_cast<GlobalValue>(*NewProto); 597 if (!New) 598 return *NewProto; 599 600 // If we already created the body, just return. 601 if (auto *F = dyn_cast<Function>(New)) { 602 if (!F->isDeclaration()) 603 return New; 604 } else if (auto *V = dyn_cast<GlobalVariable>(New)) { 605 if (V->hasInitializer() || V->hasAppendingLinkage()) 606 return New; 607 } else { 608 auto *IS = cast<GlobalIndirectSymbol>(New); 609 if (IS->getIndirectSymbol()) 610 return New; 611 } 612 613 // When linking a global for an indirect symbol, it will always be linked. 614 // However we need to check if it was not already scheduled to satisfy a 615 // reference from a regular global value initializer. We know if it has been 616 // schedule if the "New" GlobalValue that is mapped here for the indirect 617 // symbol is the same as the one already mapped. If there is an entry in the 618 // ValueMap but the value is different, it means that the value already had a 619 // definition in the destination module (linkonce for instance), but we need a 620 // new definition for the indirect symbol ("New" will be different. 621 if (ForIndirectSymbol && ValueMap.lookup(SGV) == New) 622 return New; 623 624 if (ForIndirectSymbol || shouldLink(New, *SGV)) 625 setError(linkGlobalValueBody(*New, *SGV)); 626 627 return New; 628 } 629 630 /// Loop through the global variables in the src module and merge them into the 631 /// dest module. 632 GlobalVariable *IRLinker::copyGlobalVariableProto(const GlobalVariable *SGVar) { 633 // No linking to be performed or linking from the source: simply create an 634 // identical version of the symbol over in the dest module... the 635 // initializer will be filled in later by LinkGlobalInits. 636 GlobalVariable *NewDGV = 637 new GlobalVariable(DstM, TypeMap.get(SGVar->getValueType()), 638 SGVar->isConstant(), GlobalValue::ExternalLinkage, 639 /*init*/ nullptr, SGVar->getName(), 640 /*insertbefore*/ nullptr, SGVar->getThreadLocalMode(), 641 SGVar->getAddressSpace()); 642 NewDGV->setAlignment(MaybeAlign(SGVar->getAlignment())); 643 NewDGV->copyAttributesFrom(SGVar); 644 return NewDGV; 645 } 646 647 AttributeList IRLinker::mapAttributeTypes(LLVMContext &C, AttributeList Attrs) { 648 for (unsigned i = 0; i < Attrs.getNumAttrSets(); ++i) { 649 for (Attribute::AttrKind TypedAttr : 650 {Attribute::ByVal, Attribute::StructRet, Attribute::ByRef, 651 Attribute::InAlloca}) { 652 if (Attrs.hasAttribute(i, TypedAttr)) { 653 if (Type *Ty = Attrs.getAttribute(i, TypedAttr).getValueAsType()) { 654 Attrs = Attrs.replaceAttributeType(C, i, TypedAttr, TypeMap.get(Ty)); 655 break; 656 } 657 } 658 } 659 } 660 return Attrs; 661 } 662 663 /// Link the function in the source module into the destination module if 664 /// needed, setting up mapping information. 665 Function *IRLinker::copyFunctionProto(const Function *SF) { 666 // If there is no linkage to be performed or we are linking from the source, 667 // bring SF over. 668 auto *F = Function::Create(TypeMap.get(SF->getFunctionType()), 669 GlobalValue::ExternalLinkage, 670 SF->getAddressSpace(), SF->getName(), &DstM); 671 F->copyAttributesFrom(SF); 672 F->setAttributes(mapAttributeTypes(F->getContext(), F->getAttributes())); 673 return F; 674 } 675 676 /// Set up prototypes for any indirect symbols that come over from the source 677 /// module. 678 GlobalValue * 679 IRLinker::copyGlobalIndirectSymbolProto(const GlobalIndirectSymbol *SGIS) { 680 // If there is no linkage to be performed or we're linking from the source, 681 // bring over SGA. 682 auto *Ty = TypeMap.get(SGIS->getValueType()); 683 GlobalIndirectSymbol *GIS; 684 if (isa<GlobalAlias>(SGIS)) 685 GIS = GlobalAlias::create(Ty, SGIS->getAddressSpace(), 686 GlobalValue::ExternalLinkage, SGIS->getName(), 687 &DstM); 688 else 689 GIS = GlobalIFunc::create(Ty, SGIS->getAddressSpace(), 690 GlobalValue::ExternalLinkage, SGIS->getName(), 691 nullptr, &DstM); 692 GIS->copyAttributesFrom(SGIS); 693 return GIS; 694 } 695 696 GlobalValue *IRLinker::copyGlobalValueProto(const GlobalValue *SGV, 697 bool ForDefinition) { 698 GlobalValue *NewGV; 699 if (auto *SGVar = dyn_cast<GlobalVariable>(SGV)) { 700 NewGV = copyGlobalVariableProto(SGVar); 701 } else if (auto *SF = dyn_cast<Function>(SGV)) { 702 NewGV = copyFunctionProto(SF); 703 } else { 704 if (ForDefinition) 705 NewGV = copyGlobalIndirectSymbolProto(cast<GlobalIndirectSymbol>(SGV)); 706 else if (SGV->getValueType()->isFunctionTy()) 707 NewGV = 708 Function::Create(cast<FunctionType>(TypeMap.get(SGV->getValueType())), 709 GlobalValue::ExternalLinkage, SGV->getAddressSpace(), 710 SGV->getName(), &DstM); 711 else 712 NewGV = 713 new GlobalVariable(DstM, TypeMap.get(SGV->getValueType()), 714 /*isConstant*/ false, GlobalValue::ExternalLinkage, 715 /*init*/ nullptr, SGV->getName(), 716 /*insertbefore*/ nullptr, 717 SGV->getThreadLocalMode(), SGV->getAddressSpace()); 718 } 719 720 if (ForDefinition) 721 NewGV->setLinkage(SGV->getLinkage()); 722 else if (SGV->hasExternalWeakLinkage()) 723 NewGV->setLinkage(GlobalValue::ExternalWeakLinkage); 724 725 if (auto *NewGO = dyn_cast<GlobalObject>(NewGV)) { 726 // Metadata for global variables and function declarations is copied eagerly. 727 if (isa<GlobalVariable>(SGV) || SGV->isDeclaration()) 728 NewGO->copyMetadata(cast<GlobalObject>(SGV), 0); 729 } 730 731 // Remove these copied constants in case this stays a declaration, since 732 // they point to the source module. If the def is linked the values will 733 // be mapped in during linkFunctionBody. 734 if (auto *NewF = dyn_cast<Function>(NewGV)) { 735 NewF->setPersonalityFn(nullptr); 736 NewF->setPrefixData(nullptr); 737 NewF->setPrologueData(nullptr); 738 } 739 740 return NewGV; 741 } 742 743 static StringRef getTypeNamePrefix(StringRef Name) { 744 size_t DotPos = Name.rfind('.'); 745 return (DotPos == 0 || DotPos == StringRef::npos || Name.back() == '.' || 746 !isdigit(static_cast<unsigned char>(Name[DotPos + 1]))) 747 ? Name 748 : Name.substr(0, DotPos); 749 } 750 751 /// Loop over all of the linked values to compute type mappings. For example, 752 /// if we link "extern Foo *x" and "Foo *x = NULL", then we have two struct 753 /// types 'Foo' but one got renamed when the module was loaded into the same 754 /// LLVMContext. 755 void IRLinker::computeTypeMapping() { 756 for (GlobalValue &SGV : SrcM->globals()) { 757 GlobalValue *DGV = getLinkedToGlobal(&SGV); 758 if (!DGV) 759 continue; 760 761 if (!DGV->hasAppendingLinkage() || !SGV.hasAppendingLinkage()) { 762 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 763 continue; 764 } 765 766 // Unify the element type of appending arrays. 767 ArrayType *DAT = cast<ArrayType>(DGV->getValueType()); 768 ArrayType *SAT = cast<ArrayType>(SGV.getValueType()); 769 TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType()); 770 } 771 772 for (GlobalValue &SGV : *SrcM) 773 if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) { 774 if (DGV->getType() == SGV.getType()) { 775 // If the types of DGV and SGV are the same, it means that DGV is from 776 // the source module and got added to DstM from a shared metadata. We 777 // shouldn't map this type to itself in case the type's components get 778 // remapped to a new type from DstM (for instance, during the loop over 779 // SrcM->getIdentifiedStructTypes() below). 780 continue; 781 } 782 783 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 784 } 785 786 for (GlobalValue &SGV : SrcM->aliases()) 787 if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) 788 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 789 790 // Incorporate types by name, scanning all the types in the source module. 791 // At this point, the destination module may have a type "%foo = { i32 }" for 792 // example. When the source module got loaded into the same LLVMContext, if 793 // it had the same type, it would have been renamed to "%foo.42 = { i32 }". 794 std::vector<StructType *> Types = SrcM->getIdentifiedStructTypes(); 795 for (StructType *ST : Types) { 796 if (!ST->hasName()) 797 continue; 798 799 if (TypeMap.DstStructTypesSet.hasType(ST)) { 800 // This is actually a type from the destination module. 801 // getIdentifiedStructTypes() can have found it by walking debug info 802 // metadata nodes, some of which get linked by name when ODR Type Uniquing 803 // is enabled on the Context, from the source to the destination module. 804 continue; 805 } 806 807 auto STTypePrefix = getTypeNamePrefix(ST->getName()); 808 if (STTypePrefix.size() == ST->getName().size()) 809 continue; 810 811 // Check to see if the destination module has a struct with the prefix name. 812 StructType *DST = StructType::getTypeByName(ST->getContext(), STTypePrefix); 813 if (!DST) 814 continue; 815 816 // Don't use it if this actually came from the source module. They're in 817 // the same LLVMContext after all. Also don't use it unless the type is 818 // actually used in the destination module. This can happen in situations 819 // like this: 820 // 821 // Module A Module B 822 // -------- -------- 823 // %Z = type { %A } %B = type { %C.1 } 824 // %A = type { %B.1, [7 x i8] } %C.1 = type { i8* } 825 // %B.1 = type { %C } %A.2 = type { %B.3, [5 x i8] } 826 // %C = type { i8* } %B.3 = type { %C.1 } 827 // 828 // When we link Module B with Module A, the '%B' in Module B is 829 // used. However, that would then use '%C.1'. But when we process '%C.1', 830 // we prefer to take the '%C' version. So we are then left with both 831 // '%C.1' and '%C' being used for the same types. This leads to some 832 // variables using one type and some using the other. 833 if (TypeMap.DstStructTypesSet.hasType(DST)) 834 TypeMap.addTypeMapping(DST, ST); 835 } 836 837 // Now that we have discovered all of the type equivalences, get a body for 838 // any 'opaque' types in the dest module that are now resolved. 839 TypeMap.linkDefinedTypeBodies(); 840 } 841 842 static void getArrayElements(const Constant *C, 843 SmallVectorImpl<Constant *> &Dest) { 844 unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements(); 845 846 for (unsigned i = 0; i != NumElements; ++i) 847 Dest.push_back(C->getAggregateElement(i)); 848 } 849 850 /// If there were any appending global variables, link them together now. 851 Expected<Constant *> 852 IRLinker::linkAppendingVarProto(GlobalVariable *DstGV, 853 const GlobalVariable *SrcGV) { 854 // Check that both variables have compatible properties. 855 if (DstGV && !DstGV->isDeclaration() && !SrcGV->isDeclaration()) { 856 if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage()) 857 return stringErr( 858 "Linking globals named '" + SrcGV->getName() + 859 "': can only link appending global with another appending " 860 "global!"); 861 862 if (DstGV->isConstant() != SrcGV->isConstant()) 863 return stringErr("Appending variables linked with different const'ness!"); 864 865 if (DstGV->getAlignment() != SrcGV->getAlignment()) 866 return stringErr( 867 "Appending variables with different alignment need to be linked!"); 868 869 if (DstGV->getVisibility() != SrcGV->getVisibility()) 870 return stringErr( 871 "Appending variables with different visibility need to be linked!"); 872 873 if (DstGV->hasGlobalUnnamedAddr() != SrcGV->hasGlobalUnnamedAddr()) 874 return stringErr( 875 "Appending variables with different unnamed_addr need to be linked!"); 876 877 if (DstGV->getSection() != SrcGV->getSection()) 878 return stringErr( 879 "Appending variables with different section name need to be linked!"); 880 } 881 882 // Do not need to do anything if source is a declaration. 883 if (SrcGV->isDeclaration()) 884 return DstGV; 885 886 Type *EltTy = cast<ArrayType>(TypeMap.get(SrcGV->getValueType())) 887 ->getElementType(); 888 889 // FIXME: This upgrade is done during linking to support the C API. Once the 890 // old form is deprecated, we should move this upgrade to 891 // llvm::UpgradeGlobalVariable() and simplify the logic here and in 892 // Mapper::mapAppendingVariable() in ValueMapper.cpp. 893 StringRef Name = SrcGV->getName(); 894 bool IsNewStructor = false; 895 bool IsOldStructor = false; 896 if (Name == "llvm.global_ctors" || Name == "llvm.global_dtors") { 897 if (cast<StructType>(EltTy)->getNumElements() == 3) 898 IsNewStructor = true; 899 else 900 IsOldStructor = true; 901 } 902 903 PointerType *VoidPtrTy = Type::getInt8Ty(SrcGV->getContext())->getPointerTo(); 904 if (IsOldStructor) { 905 auto &ST = *cast<StructType>(EltTy); 906 Type *Tys[3] = {ST.getElementType(0), ST.getElementType(1), VoidPtrTy}; 907 EltTy = StructType::get(SrcGV->getContext(), Tys, false); 908 } 909 910 uint64_t DstNumElements = 0; 911 if (DstGV && !DstGV->isDeclaration()) { 912 ArrayType *DstTy = cast<ArrayType>(DstGV->getValueType()); 913 DstNumElements = DstTy->getNumElements(); 914 915 // Check to see that they two arrays agree on type. 916 if (EltTy != DstTy->getElementType()) 917 return stringErr("Appending variables with different element types!"); 918 } 919 920 SmallVector<Constant *, 16> SrcElements; 921 getArrayElements(SrcGV->getInitializer(), SrcElements); 922 923 if (IsNewStructor) { 924 erase_if(SrcElements, [this](Constant *E) { 925 auto *Key = 926 dyn_cast<GlobalValue>(E->getAggregateElement(2)->stripPointerCasts()); 927 if (!Key) 928 return false; 929 GlobalValue *DGV = getLinkedToGlobal(Key); 930 return !shouldLink(DGV, *Key); 931 }); 932 } 933 uint64_t NewSize = DstNumElements + SrcElements.size(); 934 ArrayType *NewType = ArrayType::get(EltTy, NewSize); 935 936 // Create the new global variable. 937 GlobalVariable *NG = new GlobalVariable( 938 DstM, NewType, SrcGV->isConstant(), SrcGV->getLinkage(), 939 /*init*/ nullptr, /*name*/ "", DstGV, SrcGV->getThreadLocalMode(), 940 SrcGV->getAddressSpace()); 941 942 NG->copyAttributesFrom(SrcGV); 943 forceRenaming(NG, SrcGV->getName()); 944 945 Constant *Ret = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType())); 946 947 Mapper.scheduleMapAppendingVariable( 948 *NG, 949 (DstGV && !DstGV->isDeclaration()) ? DstGV->getInitializer() : nullptr, 950 IsOldStructor, SrcElements); 951 952 // Replace any uses of the two global variables with uses of the new 953 // global. 954 if (DstGV) { 955 RAUWWorklist.push_back( 956 std::make_pair(DstGV, ConstantExpr::getBitCast(NG, DstGV->getType()))); 957 } 958 959 return Ret; 960 } 961 962 bool IRLinker::shouldLink(GlobalValue *DGV, GlobalValue &SGV) { 963 if (ValuesToLink.count(&SGV) || SGV.hasLocalLinkage()) 964 return true; 965 966 if (DGV && !DGV->isDeclarationForLinker()) 967 return false; 968 969 if (SGV.isDeclaration() || DoneLinkingBodies) 970 return false; 971 972 // Callback to the client to give a chance to lazily add the Global to the 973 // list of value to link. 974 bool LazilyAdded = false; 975 AddLazyFor(SGV, [this, &LazilyAdded](GlobalValue &GV) { 976 maybeAdd(&GV); 977 LazilyAdded = true; 978 }); 979 return LazilyAdded; 980 } 981 982 Expected<Constant *> IRLinker::linkGlobalValueProto(GlobalValue *SGV, 983 bool ForIndirectSymbol) { 984 GlobalValue *DGV = getLinkedToGlobal(SGV); 985 986 bool ShouldLink = shouldLink(DGV, *SGV); 987 988 // just missing from map 989 if (ShouldLink) { 990 auto I = ValueMap.find(SGV); 991 if (I != ValueMap.end()) 992 return cast<Constant>(I->second); 993 994 I = IndirectSymbolValueMap.find(SGV); 995 if (I != IndirectSymbolValueMap.end()) 996 return cast<Constant>(I->second); 997 } 998 999 if (!ShouldLink && ForIndirectSymbol) 1000 DGV = nullptr; 1001 1002 // Handle the ultra special appending linkage case first. 1003 if (SGV->hasAppendingLinkage() || (DGV && DGV->hasAppendingLinkage())) 1004 return linkAppendingVarProto(cast_or_null<GlobalVariable>(DGV), 1005 cast<GlobalVariable>(SGV)); 1006 1007 bool NeedsRenaming = false; 1008 GlobalValue *NewGV; 1009 if (DGV && !ShouldLink) { 1010 NewGV = DGV; 1011 } else { 1012 // If we are done linking global value bodies (i.e. we are performing 1013 // metadata linking), don't link in the global value due to this 1014 // reference, simply map it to null. 1015 if (DoneLinkingBodies) 1016 return nullptr; 1017 1018 NewGV = copyGlobalValueProto(SGV, ShouldLink || ForIndirectSymbol); 1019 if (ShouldLink || !ForIndirectSymbol) 1020 NeedsRenaming = true; 1021 } 1022 1023 // Overloaded intrinsics have overloaded types names as part of their 1024 // names. If we renamed overloaded types we should rename the intrinsic 1025 // as well. 1026 if (Function *F = dyn_cast<Function>(NewGV)) 1027 if (auto Remangled = Intrinsic::remangleIntrinsicFunction(F)) { 1028 NewGV->eraseFromParent(); 1029 NewGV = Remangled.getValue(); 1030 NeedsRenaming = false; 1031 } 1032 1033 if (NeedsRenaming) 1034 forceRenaming(NewGV, SGV->getName()); 1035 1036 if (ShouldLink || ForIndirectSymbol) { 1037 if (const Comdat *SC = SGV->getComdat()) { 1038 if (auto *GO = dyn_cast<GlobalObject>(NewGV)) { 1039 Comdat *DC = DstM.getOrInsertComdat(SC->getName()); 1040 DC->setSelectionKind(SC->getSelectionKind()); 1041 GO->setComdat(DC); 1042 } 1043 } 1044 } 1045 1046 if (!ShouldLink && ForIndirectSymbol) 1047 NewGV->setLinkage(GlobalValue::InternalLinkage); 1048 1049 Constant *C = NewGV; 1050 // Only create a bitcast if necessary. In particular, with 1051 // DebugTypeODRUniquing we may reach metadata in the destination module 1052 // containing a GV from the source module, in which case SGV will be 1053 // the same as DGV and NewGV, and TypeMap.get() will assert since it 1054 // assumes it is being invoked on a type in the source module. 1055 if (DGV && NewGV != SGV) { 1056 C = ConstantExpr::getPointerBitCastOrAddrSpaceCast( 1057 NewGV, TypeMap.get(SGV->getType())); 1058 } 1059 1060 if (DGV && NewGV != DGV) { 1061 // Schedule "replace all uses with" to happen after materializing is 1062 // done. It is not safe to do it now, since ValueMapper may be holding 1063 // pointers to constants that will get deleted if RAUW runs. 1064 RAUWWorklist.push_back(std::make_pair( 1065 DGV, 1066 ConstantExpr::getPointerBitCastOrAddrSpaceCast(NewGV, DGV->getType()))); 1067 } 1068 1069 return C; 1070 } 1071 1072 /// Update the initializers in the Dest module now that all globals that may be 1073 /// referenced are in Dest. 1074 void IRLinker::linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src) { 1075 // Figure out what the initializer looks like in the dest module. 1076 Mapper.scheduleMapGlobalInitializer(Dst, *Src.getInitializer()); 1077 } 1078 1079 /// Copy the source function over into the dest function and fix up references 1080 /// to values. At this point we know that Dest is an external function, and 1081 /// that Src is not. 1082 Error IRLinker::linkFunctionBody(Function &Dst, Function &Src) { 1083 assert(Dst.isDeclaration() && !Src.isDeclaration()); 1084 1085 // Materialize if needed. 1086 if (Error Err = Src.materialize()) 1087 return Err; 1088 1089 // Link in the operands without remapping. 1090 if (Src.hasPrefixData()) 1091 Dst.setPrefixData(Src.getPrefixData()); 1092 if (Src.hasPrologueData()) 1093 Dst.setPrologueData(Src.getPrologueData()); 1094 if (Src.hasPersonalityFn()) 1095 Dst.setPersonalityFn(Src.getPersonalityFn()); 1096 1097 // Copy over the metadata attachments without remapping. 1098 Dst.copyMetadata(&Src, 0); 1099 1100 // Steal arguments and splice the body of Src into Dst. 1101 Dst.stealArgumentListFrom(Src); 1102 Dst.getBasicBlockList().splice(Dst.end(), Src.getBasicBlockList()); 1103 1104 // Everything has been moved over. Remap it. 1105 Mapper.scheduleRemapFunction(Dst); 1106 return Error::success(); 1107 } 1108 1109 void IRLinker::linkIndirectSymbolBody(GlobalIndirectSymbol &Dst, 1110 GlobalIndirectSymbol &Src) { 1111 Mapper.scheduleMapGlobalIndirectSymbol(Dst, *Src.getIndirectSymbol(), 1112 IndirectSymbolMCID); 1113 } 1114 1115 Error IRLinker::linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src) { 1116 if (auto *F = dyn_cast<Function>(&Src)) 1117 return linkFunctionBody(cast<Function>(Dst), *F); 1118 if (auto *GVar = dyn_cast<GlobalVariable>(&Src)) { 1119 linkGlobalVariable(cast<GlobalVariable>(Dst), *GVar); 1120 return Error::success(); 1121 } 1122 linkIndirectSymbolBody(cast<GlobalIndirectSymbol>(Dst), cast<GlobalIndirectSymbol>(Src)); 1123 return Error::success(); 1124 } 1125 1126 void IRLinker::flushRAUWWorklist() { 1127 for (const auto &Elem : RAUWWorklist) { 1128 GlobalValue *Old; 1129 Value *New; 1130 std::tie(Old, New) = Elem; 1131 1132 Old->replaceAllUsesWith(New); 1133 Old->eraseFromParent(); 1134 } 1135 RAUWWorklist.clear(); 1136 } 1137 1138 void IRLinker::prepareCompileUnitsForImport() { 1139 NamedMDNode *SrcCompileUnits = SrcM->getNamedMetadata("llvm.dbg.cu"); 1140 if (!SrcCompileUnits) 1141 return; 1142 // When importing for ThinLTO, prevent importing of types listed on 1143 // the DICompileUnit that we don't need a copy of in the importing 1144 // module. They will be emitted by the originating module. 1145 for (unsigned I = 0, E = SrcCompileUnits->getNumOperands(); I != E; ++I) { 1146 auto *CU = cast<DICompileUnit>(SrcCompileUnits->getOperand(I)); 1147 assert(CU && "Expected valid compile unit"); 1148 // Enums, macros, and retained types don't need to be listed on the 1149 // imported DICompileUnit. This means they will only be imported 1150 // if reached from the mapped IR. 1151 CU->replaceEnumTypes(nullptr); 1152 CU->replaceMacros(nullptr); 1153 CU->replaceRetainedTypes(nullptr); 1154 1155 // The original definition (or at least its debug info - if the variable is 1156 // internalized and optimized away) will remain in the source module, so 1157 // there's no need to import them. 1158 // If LLVM ever does more advanced optimizations on global variables 1159 // (removing/localizing write operations, for instance) that can track 1160 // through debug info, this decision may need to be revisited - but do so 1161 // with care when it comes to debug info size. Emitting small CUs containing 1162 // only a few imported entities into every destination module may be very 1163 // size inefficient. 1164 CU->replaceGlobalVariables(nullptr); 1165 1166 // Imported entities only need to be mapped in if they have local 1167 // scope, as those might correspond to an imported entity inside a 1168 // function being imported (any locally scoped imported entities that 1169 // don't end up referenced by an imported function will not be emitted 1170 // into the object). Imported entities not in a local scope 1171 // (e.g. on the namespace) only need to be emitted by the originating 1172 // module. Create a list of the locally scoped imported entities, and 1173 // replace the source CUs imported entity list with the new list, so 1174 // only those are mapped in. 1175 // FIXME: Locally-scoped imported entities could be moved to the 1176 // functions they are local to instead of listing them on the CU, and 1177 // we would naturally only link in those needed by function importing. 1178 SmallVector<TrackingMDNodeRef, 4> AllImportedModules; 1179 bool ReplaceImportedEntities = false; 1180 for (auto *IE : CU->getImportedEntities()) { 1181 DIScope *Scope = IE->getScope(); 1182 assert(Scope && "Invalid Scope encoding!"); 1183 if (isa<DILocalScope>(Scope)) 1184 AllImportedModules.emplace_back(IE); 1185 else 1186 ReplaceImportedEntities = true; 1187 } 1188 if (ReplaceImportedEntities) { 1189 if (!AllImportedModules.empty()) 1190 CU->replaceImportedEntities(MDTuple::get( 1191 CU->getContext(), 1192 SmallVector<Metadata *, 16>(AllImportedModules.begin(), 1193 AllImportedModules.end()))); 1194 else 1195 // If there were no local scope imported entities, we can map 1196 // the whole list to nullptr. 1197 CU->replaceImportedEntities(nullptr); 1198 } 1199 } 1200 } 1201 1202 /// Insert all of the named MDNodes in Src into the Dest module. 1203 void IRLinker::linkNamedMDNodes() { 1204 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata(); 1205 for (const NamedMDNode &NMD : SrcM->named_metadata()) { 1206 // Don't link module flags here. Do them separately. 1207 if (&NMD == SrcModFlags) 1208 continue; 1209 NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName()); 1210 // Add Src elements into Dest node. 1211 for (const MDNode *Op : NMD.operands()) 1212 DestNMD->addOperand(Mapper.mapMDNode(*Op)); 1213 } 1214 } 1215 1216 /// Merge the linker flags in Src into the Dest module. 1217 Error IRLinker::linkModuleFlagsMetadata() { 1218 // If the source module has no module flags, we are done. 1219 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata(); 1220 if (!SrcModFlags) 1221 return Error::success(); 1222 1223 // If the destination module doesn't have module flags yet, then just copy 1224 // over the source module's flags. 1225 NamedMDNode *DstModFlags = DstM.getOrInsertModuleFlagsMetadata(); 1226 if (DstModFlags->getNumOperands() == 0) { 1227 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) 1228 DstModFlags->addOperand(SrcModFlags->getOperand(I)); 1229 1230 return Error::success(); 1231 } 1232 1233 // First build a map of the existing module flags and requirements. 1234 DenseMap<MDString *, std::pair<MDNode *, unsigned>> Flags; 1235 SmallSetVector<MDNode *, 16> Requirements; 1236 for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) { 1237 MDNode *Op = DstModFlags->getOperand(I); 1238 ConstantInt *Behavior = mdconst::extract<ConstantInt>(Op->getOperand(0)); 1239 MDString *ID = cast<MDString>(Op->getOperand(1)); 1240 1241 if (Behavior->getZExtValue() == Module::Require) { 1242 Requirements.insert(cast<MDNode>(Op->getOperand(2))); 1243 } else { 1244 Flags[ID] = std::make_pair(Op, I); 1245 } 1246 } 1247 1248 // Merge in the flags from the source module, and also collect its set of 1249 // requirements. 1250 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) { 1251 MDNode *SrcOp = SrcModFlags->getOperand(I); 1252 ConstantInt *SrcBehavior = 1253 mdconst::extract<ConstantInt>(SrcOp->getOperand(0)); 1254 MDString *ID = cast<MDString>(SrcOp->getOperand(1)); 1255 MDNode *DstOp; 1256 unsigned DstIndex; 1257 std::tie(DstOp, DstIndex) = Flags.lookup(ID); 1258 unsigned SrcBehaviorValue = SrcBehavior->getZExtValue(); 1259 1260 // If this is a requirement, add it and continue. 1261 if (SrcBehaviorValue == Module::Require) { 1262 // If the destination module does not already have this requirement, add 1263 // it. 1264 if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) { 1265 DstModFlags->addOperand(SrcOp); 1266 } 1267 continue; 1268 } 1269 1270 // If there is no existing flag with this ID, just add it. 1271 if (!DstOp) { 1272 Flags[ID] = std::make_pair(SrcOp, DstModFlags->getNumOperands()); 1273 DstModFlags->addOperand(SrcOp); 1274 continue; 1275 } 1276 1277 // Otherwise, perform a merge. 1278 ConstantInt *DstBehavior = 1279 mdconst::extract<ConstantInt>(DstOp->getOperand(0)); 1280 unsigned DstBehaviorValue = DstBehavior->getZExtValue(); 1281 1282 auto overrideDstValue = [&]() { 1283 DstModFlags->setOperand(DstIndex, SrcOp); 1284 Flags[ID].first = SrcOp; 1285 }; 1286 1287 // If either flag has override behavior, handle it first. 1288 if (DstBehaviorValue == Module::Override) { 1289 // Diagnose inconsistent flags which both have override behavior. 1290 if (SrcBehaviorValue == Module::Override && 1291 SrcOp->getOperand(2) != DstOp->getOperand(2)) 1292 return stringErr("linking module flags '" + ID->getString() + 1293 "': IDs have conflicting override values in '" + 1294 SrcM->getModuleIdentifier() + "' and '" + 1295 DstM.getModuleIdentifier() + "'"); 1296 continue; 1297 } else if (SrcBehaviorValue == Module::Override) { 1298 // Update the destination flag to that of the source. 1299 overrideDstValue(); 1300 continue; 1301 } 1302 1303 // Diagnose inconsistent merge behavior types. 1304 if (SrcBehaviorValue != DstBehaviorValue) { 1305 bool MaxAndWarn = (SrcBehaviorValue == Module::Max && 1306 DstBehaviorValue == Module::Warning) || 1307 (DstBehaviorValue == Module::Max && 1308 SrcBehaviorValue == Module::Warning); 1309 if (!MaxAndWarn) 1310 return stringErr("linking module flags '" + ID->getString() + 1311 "': IDs have conflicting behaviors in '" + 1312 SrcM->getModuleIdentifier() + "' and '" + 1313 DstM.getModuleIdentifier() + "'"); 1314 } 1315 1316 auto replaceDstValue = [&](MDNode *New) { 1317 Metadata *FlagOps[] = {DstOp->getOperand(0), ID, New}; 1318 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps); 1319 DstModFlags->setOperand(DstIndex, Flag); 1320 Flags[ID].first = Flag; 1321 }; 1322 1323 // Emit a warning if the values differ and either source or destination 1324 // request Warning behavior. 1325 if ((DstBehaviorValue == Module::Warning || 1326 SrcBehaviorValue == Module::Warning) && 1327 SrcOp->getOperand(2) != DstOp->getOperand(2)) { 1328 std::string Str; 1329 raw_string_ostream(Str) 1330 << "linking module flags '" << ID->getString() 1331 << "': IDs have conflicting values ('" << *SrcOp->getOperand(2) 1332 << "' from " << SrcM->getModuleIdentifier() << " with '" 1333 << *DstOp->getOperand(2) << "' from " << DstM.getModuleIdentifier() 1334 << ')'; 1335 emitWarning(Str); 1336 } 1337 1338 // Choose the maximum if either source or destination request Max behavior. 1339 if (DstBehaviorValue == Module::Max || SrcBehaviorValue == Module::Max) { 1340 ConstantInt *DstValue = 1341 mdconst::extract<ConstantInt>(DstOp->getOperand(2)); 1342 ConstantInt *SrcValue = 1343 mdconst::extract<ConstantInt>(SrcOp->getOperand(2)); 1344 1345 // The resulting flag should have a Max behavior, and contain the maximum 1346 // value from between the source and destination values. 1347 Metadata *FlagOps[] = { 1348 (DstBehaviorValue != Module::Max ? SrcOp : DstOp)->getOperand(0), ID, 1349 (SrcValue->getZExtValue() > DstValue->getZExtValue() ? SrcOp : DstOp) 1350 ->getOperand(2)}; 1351 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps); 1352 DstModFlags->setOperand(DstIndex, Flag); 1353 Flags[ID].first = Flag; 1354 continue; 1355 } 1356 1357 // Perform the merge for standard behavior types. 1358 switch (SrcBehaviorValue) { 1359 case Module::Require: 1360 case Module::Override: 1361 llvm_unreachable("not possible"); 1362 case Module::Error: { 1363 // Emit an error if the values differ. 1364 if (SrcOp->getOperand(2) != DstOp->getOperand(2)) 1365 return stringErr("linking module flags '" + ID->getString() + 1366 "': IDs have conflicting values in '" + 1367 SrcM->getModuleIdentifier() + "' and '" + 1368 DstM.getModuleIdentifier() + "'"); 1369 continue; 1370 } 1371 case Module::Warning: { 1372 break; 1373 } 1374 case Module::Max: { 1375 break; 1376 } 1377 case Module::Append: { 1378 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2)); 1379 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2)); 1380 SmallVector<Metadata *, 8> MDs; 1381 MDs.reserve(DstValue->getNumOperands() + SrcValue->getNumOperands()); 1382 MDs.append(DstValue->op_begin(), DstValue->op_end()); 1383 MDs.append(SrcValue->op_begin(), SrcValue->op_end()); 1384 1385 replaceDstValue(MDNode::get(DstM.getContext(), MDs)); 1386 break; 1387 } 1388 case Module::AppendUnique: { 1389 SmallSetVector<Metadata *, 16> Elts; 1390 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2)); 1391 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2)); 1392 Elts.insert(DstValue->op_begin(), DstValue->op_end()); 1393 Elts.insert(SrcValue->op_begin(), SrcValue->op_end()); 1394 1395 replaceDstValue(MDNode::get(DstM.getContext(), 1396 makeArrayRef(Elts.begin(), Elts.end()))); 1397 break; 1398 } 1399 } 1400 1401 } 1402 1403 // Check all of the requirements. 1404 for (unsigned I = 0, E = Requirements.size(); I != E; ++I) { 1405 MDNode *Requirement = Requirements[I]; 1406 MDString *Flag = cast<MDString>(Requirement->getOperand(0)); 1407 Metadata *ReqValue = Requirement->getOperand(1); 1408 1409 MDNode *Op = Flags[Flag].first; 1410 if (!Op || Op->getOperand(2) != ReqValue) 1411 return stringErr("linking module flags '" + Flag->getString() + 1412 "': does not have the required value"); 1413 } 1414 return Error::success(); 1415 } 1416 1417 /// Return InlineAsm adjusted with target-specific directives if required. 1418 /// For ARM and Thumb, we have to add directives to select the appropriate ISA 1419 /// to support mixing module-level inline assembly from ARM and Thumb modules. 1420 static std::string adjustInlineAsm(const std::string &InlineAsm, 1421 const Triple &Triple) { 1422 if (Triple.getArch() == Triple::thumb || Triple.getArch() == Triple::thumbeb) 1423 return ".text\n.balign 2\n.thumb\n" + InlineAsm; 1424 if (Triple.getArch() == Triple::arm || Triple.getArch() == Triple::armeb) 1425 return ".text\n.balign 4\n.arm\n" + InlineAsm; 1426 return InlineAsm; 1427 } 1428 1429 Error IRLinker::run() { 1430 // Ensure metadata materialized before value mapping. 1431 if (SrcM->getMaterializer()) 1432 if (Error Err = SrcM->getMaterializer()->materializeMetadata()) 1433 return Err; 1434 1435 // Inherit the target data from the source module if the destination module 1436 // doesn't have one already. 1437 if (DstM.getDataLayout().isDefault()) 1438 DstM.setDataLayout(SrcM->getDataLayout()); 1439 1440 if (SrcM->getDataLayout() != DstM.getDataLayout()) { 1441 emitWarning("Linking two modules of different data layouts: '" + 1442 SrcM->getModuleIdentifier() + "' is '" + 1443 SrcM->getDataLayoutStr() + "' whereas '" + 1444 DstM.getModuleIdentifier() + "' is '" + 1445 DstM.getDataLayoutStr() + "'\n"); 1446 } 1447 1448 // Copy the target triple from the source to dest if the dest's is empty. 1449 if (DstM.getTargetTriple().empty() && !SrcM->getTargetTriple().empty()) 1450 DstM.setTargetTriple(SrcM->getTargetTriple()); 1451 1452 Triple SrcTriple(SrcM->getTargetTriple()), DstTriple(DstM.getTargetTriple()); 1453 1454 if (!SrcM->getTargetTriple().empty()&& 1455 !SrcTriple.isCompatibleWith(DstTriple)) 1456 emitWarning("Linking two modules of different target triples: '" + 1457 SrcM->getModuleIdentifier() + "' is '" + 1458 SrcM->getTargetTriple() + "' whereas '" + 1459 DstM.getModuleIdentifier() + "' is '" + DstM.getTargetTriple() + 1460 "'\n"); 1461 1462 DstM.setTargetTriple(SrcTriple.merge(DstTriple)); 1463 1464 // Loop over all of the linked values to compute type mappings. 1465 computeTypeMapping(); 1466 1467 std::reverse(Worklist.begin(), Worklist.end()); 1468 while (!Worklist.empty()) { 1469 GlobalValue *GV = Worklist.back(); 1470 Worklist.pop_back(); 1471 1472 // Already mapped. 1473 if (ValueMap.find(GV) != ValueMap.end() || 1474 IndirectSymbolValueMap.find(GV) != IndirectSymbolValueMap.end()) 1475 continue; 1476 1477 assert(!GV->isDeclaration()); 1478 Mapper.mapValue(*GV); 1479 if (FoundError) 1480 return std::move(*FoundError); 1481 flushRAUWWorklist(); 1482 } 1483 1484 // Note that we are done linking global value bodies. This prevents 1485 // metadata linking from creating new references. 1486 DoneLinkingBodies = true; 1487 Mapper.addFlags(RF_NullMapMissingGlobalValues); 1488 1489 // Remap all of the named MDNodes in Src into the DstM module. We do this 1490 // after linking GlobalValues so that MDNodes that reference GlobalValues 1491 // are properly remapped. 1492 linkNamedMDNodes(); 1493 1494 if (!IsPerformingImport && !SrcM->getModuleInlineAsm().empty()) { 1495 // Append the module inline asm string. 1496 DstM.appendModuleInlineAsm(adjustInlineAsm(SrcM->getModuleInlineAsm(), 1497 SrcTriple)); 1498 } else if (IsPerformingImport) { 1499 // Import any symver directives for symbols in DstM. 1500 ModuleSymbolTable::CollectAsmSymvers(*SrcM, 1501 [&](StringRef Name, StringRef Alias) { 1502 if (DstM.getNamedValue(Name)) { 1503 SmallString<256> S(".symver "); 1504 S += Name; 1505 S += ", "; 1506 S += Alias; 1507 DstM.appendModuleInlineAsm(S); 1508 } 1509 }); 1510 } 1511 1512 // Merge the module flags into the DstM module. 1513 return linkModuleFlagsMetadata(); 1514 } 1515 1516 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(ArrayRef<Type *> E, bool P) 1517 : ETypes(E), IsPacked(P) {} 1518 1519 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(const StructType *ST) 1520 : ETypes(ST->elements()), IsPacked(ST->isPacked()) {} 1521 1522 bool IRMover::StructTypeKeyInfo::KeyTy::operator==(const KeyTy &That) const { 1523 return IsPacked == That.IsPacked && ETypes == That.ETypes; 1524 } 1525 1526 bool IRMover::StructTypeKeyInfo::KeyTy::operator!=(const KeyTy &That) const { 1527 return !this->operator==(That); 1528 } 1529 1530 StructType *IRMover::StructTypeKeyInfo::getEmptyKey() { 1531 return DenseMapInfo<StructType *>::getEmptyKey(); 1532 } 1533 1534 StructType *IRMover::StructTypeKeyInfo::getTombstoneKey() { 1535 return DenseMapInfo<StructType *>::getTombstoneKey(); 1536 } 1537 1538 unsigned IRMover::StructTypeKeyInfo::getHashValue(const KeyTy &Key) { 1539 return hash_combine(hash_combine_range(Key.ETypes.begin(), Key.ETypes.end()), 1540 Key.IsPacked); 1541 } 1542 1543 unsigned IRMover::StructTypeKeyInfo::getHashValue(const StructType *ST) { 1544 return getHashValue(KeyTy(ST)); 1545 } 1546 1547 bool IRMover::StructTypeKeyInfo::isEqual(const KeyTy &LHS, 1548 const StructType *RHS) { 1549 if (RHS == getEmptyKey() || RHS == getTombstoneKey()) 1550 return false; 1551 return LHS == KeyTy(RHS); 1552 } 1553 1554 bool IRMover::StructTypeKeyInfo::isEqual(const StructType *LHS, 1555 const StructType *RHS) { 1556 if (RHS == getEmptyKey() || RHS == getTombstoneKey()) 1557 return LHS == RHS; 1558 return KeyTy(LHS) == KeyTy(RHS); 1559 } 1560 1561 void IRMover::IdentifiedStructTypeSet::addNonOpaque(StructType *Ty) { 1562 assert(!Ty->isOpaque()); 1563 NonOpaqueStructTypes.insert(Ty); 1564 } 1565 1566 void IRMover::IdentifiedStructTypeSet::switchToNonOpaque(StructType *Ty) { 1567 assert(!Ty->isOpaque()); 1568 NonOpaqueStructTypes.insert(Ty); 1569 bool Removed = OpaqueStructTypes.erase(Ty); 1570 (void)Removed; 1571 assert(Removed); 1572 } 1573 1574 void IRMover::IdentifiedStructTypeSet::addOpaque(StructType *Ty) { 1575 assert(Ty->isOpaque()); 1576 OpaqueStructTypes.insert(Ty); 1577 } 1578 1579 StructType * 1580 IRMover::IdentifiedStructTypeSet::findNonOpaque(ArrayRef<Type *> ETypes, 1581 bool IsPacked) { 1582 IRMover::StructTypeKeyInfo::KeyTy Key(ETypes, IsPacked); 1583 auto I = NonOpaqueStructTypes.find_as(Key); 1584 return I == NonOpaqueStructTypes.end() ? nullptr : *I; 1585 } 1586 1587 bool IRMover::IdentifiedStructTypeSet::hasType(StructType *Ty) { 1588 if (Ty->isOpaque()) 1589 return OpaqueStructTypes.count(Ty); 1590 auto I = NonOpaqueStructTypes.find(Ty); 1591 return I == NonOpaqueStructTypes.end() ? false : *I == Ty; 1592 } 1593 1594 IRMover::IRMover(Module &M) : Composite(M) { 1595 TypeFinder StructTypes; 1596 StructTypes.run(M, /* OnlyNamed */ false); 1597 for (StructType *Ty : StructTypes) { 1598 if (Ty->isOpaque()) 1599 IdentifiedStructTypes.addOpaque(Ty); 1600 else 1601 IdentifiedStructTypes.addNonOpaque(Ty); 1602 } 1603 // Self-map metadatas in the destination module. This is needed when 1604 // DebugTypeODRUniquing is enabled on the LLVMContext, since metadata in the 1605 // destination module may be reached from the source module. 1606 for (auto *MD : StructTypes.getVisitedMetadata()) { 1607 SharedMDs[MD].reset(const_cast<MDNode *>(MD)); 1608 } 1609 } 1610 1611 Error IRMover::move( 1612 std::unique_ptr<Module> Src, ArrayRef<GlobalValue *> ValuesToLink, 1613 std::function<void(GlobalValue &, ValueAdder Add)> AddLazyFor, 1614 bool IsPerformingImport) { 1615 IRLinker TheIRLinker(Composite, SharedMDs, IdentifiedStructTypes, 1616 std::move(Src), ValuesToLink, std::move(AddLazyFor), 1617 IsPerformingImport); 1618 Error E = TheIRLinker.run(); 1619 Composite.dropTriviallyDeadConstantArrays(); 1620 return E; 1621 } 1622