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/PseudoProbe.h" 20 #include "llvm/IR/TypeFinder.h" 21 #include "llvm/Object/ModuleSymbolTable.h" 22 #include "llvm/Support/Error.h" 23 #include "llvm/Transforms/Utils/Cloning.h" 24 #include <utility> 25 using namespace llvm; 26 27 //===----------------------------------------------------------------------===// 28 // TypeMap implementation. 29 //===----------------------------------------------------------------------===// 30 31 namespace { 32 class TypeMapTy : public ValueMapTypeRemapper { 33 /// This is a mapping from a source type to a destination type to use. 34 DenseMap<Type *, Type *> MappedTypes; 35 36 /// When checking to see if two subgraphs are isomorphic, we speculatively 37 /// add types to MappedTypes, but keep track of them here in case we need to 38 /// roll back. 39 SmallVector<Type *, 16> SpeculativeTypes; 40 41 SmallVector<StructType *, 16> SpeculativeDstOpaqueTypes; 42 43 /// This is a list of non-opaque structs in the source module that are mapped 44 /// to an opaque struct in the destination module. 45 SmallVector<StructType *, 16> SrcDefinitionsToResolve; 46 47 /// This is the set of opaque types in the destination modules who are 48 /// getting a body from the source module. 49 SmallPtrSet<StructType *, 16> DstResolvedOpaqueTypes; 50 51 public: 52 TypeMapTy(IRMover::IdentifiedStructTypeSet &DstStructTypesSet) 53 : DstStructTypesSet(DstStructTypesSet) {} 54 55 IRMover::IdentifiedStructTypeSet &DstStructTypesSet; 56 /// Indicate that the specified type in the destination module is conceptually 57 /// equivalent to the specified type in the source module. 58 void addTypeMapping(Type *DstTy, Type *SrcTy); 59 60 /// Produce a body for an opaque type in the dest module from a type 61 /// definition in the source module. 62 void linkDefinedTypeBodies(); 63 64 /// Return the mapped type to use for the specified input type from the 65 /// source module. 66 Type *get(Type *SrcTy); 67 Type *get(Type *SrcTy, SmallPtrSet<StructType *, 8> &Visited); 68 69 void finishType(StructType *DTy, StructType *STy, ArrayRef<Type *> ETypes); 70 71 FunctionType *get(FunctionType *T) { 72 return cast<FunctionType>(get((Type *)T)); 73 } 74 75 private: 76 Type *remapType(Type *SrcTy) override { return get(SrcTy); } 77 78 bool areTypesIsomorphic(Type *DstTy, Type *SrcTy); 79 }; 80 } 81 82 void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) { 83 assert(SpeculativeTypes.empty()); 84 assert(SpeculativeDstOpaqueTypes.empty()); 85 86 // Check to see if these types are recursively isomorphic and establish a 87 // mapping between them if so. 88 if (!areTypesIsomorphic(DstTy, SrcTy)) { 89 // Oops, they aren't isomorphic. Just discard this request by rolling out 90 // any speculative mappings we've established. 91 for (Type *Ty : SpeculativeTypes) 92 MappedTypes.erase(Ty); 93 94 SrcDefinitionsToResolve.resize(SrcDefinitionsToResolve.size() - 95 SpeculativeDstOpaqueTypes.size()); 96 for (StructType *Ty : SpeculativeDstOpaqueTypes) 97 DstResolvedOpaqueTypes.erase(Ty); 98 } else { 99 // SrcTy and DstTy are recursively ismorphic. We clear names of SrcTy 100 // and all its descendants to lower amount of renaming in LLVM context 101 // Renaming occurs because we load all source modules to the same context 102 // and declaration with existing name gets renamed (i.e Foo -> Foo.42). 103 // As a result we may get several different types in the destination 104 // module, which are in fact the same. 105 for (Type *Ty : SpeculativeTypes) 106 if (auto *STy = dyn_cast<StructType>(Ty)) 107 if (STy->hasName()) 108 STy->setName(""); 109 } 110 SpeculativeTypes.clear(); 111 SpeculativeDstOpaqueTypes.clear(); 112 } 113 114 /// Recursively walk this pair of types, returning true if they are isomorphic, 115 /// false if they are not. 116 bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) { 117 // Two types with differing kinds are clearly not isomorphic. 118 if (DstTy->getTypeID() != SrcTy->getTypeID()) 119 return false; 120 121 // If we have an entry in the MappedTypes table, then we have our answer. 122 Type *&Entry = MappedTypes[SrcTy]; 123 if (Entry) 124 return Entry == DstTy; 125 126 // Two identical types are clearly isomorphic. Remember this 127 // non-speculatively. 128 if (DstTy == SrcTy) { 129 Entry = DstTy; 130 return true; 131 } 132 133 // Okay, we have two types with identical kinds that we haven't seen before. 134 135 // If this is an opaque struct type, special case it. 136 if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) { 137 // Mapping an opaque type to any struct, just keep the dest struct. 138 if (SSTy->isOpaque()) { 139 Entry = DstTy; 140 SpeculativeTypes.push_back(SrcTy); 141 return true; 142 } 143 144 // Mapping a non-opaque source type to an opaque dest. If this is the first 145 // type that we're mapping onto this destination type then we succeed. Keep 146 // the dest, but fill it in later. If this is the second (different) type 147 // that we're trying to map onto the same opaque type then we fail. 148 if (cast<StructType>(DstTy)->isOpaque()) { 149 // We can only map one source type onto the opaque destination type. 150 if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)).second) 151 return false; 152 SrcDefinitionsToResolve.push_back(SSTy); 153 SpeculativeTypes.push_back(SrcTy); 154 SpeculativeDstOpaqueTypes.push_back(cast<StructType>(DstTy)); 155 Entry = DstTy; 156 return true; 157 } 158 } 159 160 // If the number of subtypes disagree between the two types, then we fail. 161 if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes()) 162 return false; 163 164 // Fail if any of the extra properties (e.g. array size) of the type disagree. 165 if (isa<IntegerType>(DstTy)) 166 return false; // bitwidth disagrees. 167 if (PointerType *PT = dyn_cast<PointerType>(DstTy)) { 168 if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace()) 169 return false; 170 } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) { 171 if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg()) 172 return false; 173 } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) { 174 StructType *SSTy = cast<StructType>(SrcTy); 175 if (DSTy->isLiteral() != SSTy->isLiteral() || 176 DSTy->isPacked() != SSTy->isPacked()) 177 return false; 178 } else if (auto *DArrTy = dyn_cast<ArrayType>(DstTy)) { 179 if (DArrTy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements()) 180 return false; 181 } else if (auto *DVecTy = dyn_cast<VectorType>(DstTy)) { 182 if (DVecTy->getElementCount() != cast<VectorType>(SrcTy)->getElementCount()) 183 return false; 184 } 185 186 // Otherwise, we speculate that these two types will line up and recursively 187 // check the subelements. 188 Entry = DstTy; 189 SpeculativeTypes.push_back(SrcTy); 190 191 for (unsigned I = 0, E = SrcTy->getNumContainedTypes(); I != E; ++I) 192 if (!areTypesIsomorphic(DstTy->getContainedType(I), 193 SrcTy->getContainedType(I))) 194 return false; 195 196 // If everything seems to have lined up, then everything is great. 197 return true; 198 } 199 200 void TypeMapTy::linkDefinedTypeBodies() { 201 SmallVector<Type *, 16> Elements; 202 for (StructType *SrcSTy : SrcDefinitionsToResolve) { 203 StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]); 204 assert(DstSTy->isOpaque()); 205 206 // Map the body of the source type over to a new body for the dest type. 207 Elements.resize(SrcSTy->getNumElements()); 208 for (unsigned I = 0, E = Elements.size(); I != E; ++I) 209 Elements[I] = get(SrcSTy->getElementType(I)); 210 211 DstSTy->setBody(Elements, SrcSTy->isPacked()); 212 DstStructTypesSet.switchToNonOpaque(DstSTy); 213 } 214 SrcDefinitionsToResolve.clear(); 215 DstResolvedOpaqueTypes.clear(); 216 } 217 218 void TypeMapTy::finishType(StructType *DTy, StructType *STy, 219 ArrayRef<Type *> ETypes) { 220 DTy->setBody(ETypes, STy->isPacked()); 221 222 // Steal STy's name. 223 if (STy->hasName()) { 224 SmallString<16> TmpName = STy->getName(); 225 STy->setName(""); 226 DTy->setName(TmpName); 227 } 228 229 DstStructTypesSet.addNonOpaque(DTy); 230 } 231 232 Type *TypeMapTy::get(Type *Ty) { 233 SmallPtrSet<StructType *, 8> Visited; 234 return get(Ty, Visited); 235 } 236 237 Type *TypeMapTy::get(Type *Ty, SmallPtrSet<StructType *, 8> &Visited) { 238 // If we already have an entry for this type, return it. 239 Type **Entry = &MappedTypes[Ty]; 240 if (*Entry) 241 return *Entry; 242 243 // These are types that LLVM itself will unique. 244 bool IsUniqued = !isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral(); 245 246 if (!IsUniqued) { 247 #ifndef NDEBUG 248 for (auto &Pair : MappedTypes) { 249 assert(!(Pair.first != Ty && Pair.second == Ty) && 250 "mapping to a source type"); 251 } 252 #endif 253 254 if (!Visited.insert(cast<StructType>(Ty)).second) { 255 StructType *DTy = StructType::create(Ty->getContext()); 256 return *Entry = DTy; 257 } 258 } 259 260 // If this is not a recursive type, then just map all of the elements and 261 // then rebuild the type from inside out. 262 SmallVector<Type *, 4> ElementTypes; 263 264 // If there are no element types to map, then the type is itself. This is 265 // true for the anonymous {} struct, things like 'float', integers, etc. 266 if (Ty->getNumContainedTypes() == 0 && IsUniqued) 267 return *Entry = Ty; 268 269 // Remap all of the elements, keeping track of whether any of them change. 270 bool AnyChange = false; 271 ElementTypes.resize(Ty->getNumContainedTypes()); 272 for (unsigned I = 0, E = Ty->getNumContainedTypes(); I != E; ++I) { 273 ElementTypes[I] = get(Ty->getContainedType(I), Visited); 274 AnyChange |= ElementTypes[I] != Ty->getContainedType(I); 275 } 276 277 // If we found our type while recursively processing stuff, just use it. 278 Entry = &MappedTypes[Ty]; 279 if (*Entry) { 280 if (auto *DTy = dyn_cast<StructType>(*Entry)) { 281 if (DTy->isOpaque()) { 282 auto *STy = cast<StructType>(Ty); 283 finishType(DTy, STy, ElementTypes); 284 } 285 } 286 return *Entry; 287 } 288 289 // If all of the element types mapped directly over and the type is not 290 // a named struct, then the type is usable as-is. 291 if (!AnyChange && IsUniqued) 292 return *Entry = Ty; 293 294 // Otherwise, rebuild a modified type. 295 switch (Ty->getTypeID()) { 296 default: 297 llvm_unreachable("unknown derived type to remap"); 298 case Type::ArrayTyID: 299 return *Entry = ArrayType::get(ElementTypes[0], 300 cast<ArrayType>(Ty)->getNumElements()); 301 case Type::ScalableVectorTyID: 302 case Type::FixedVectorTyID: 303 return *Entry = VectorType::get(ElementTypes[0], 304 cast<VectorType>(Ty)->getElementCount()); 305 case Type::PointerTyID: 306 return *Entry = PointerType::get(ElementTypes[0], 307 cast<PointerType>(Ty)->getAddressSpace()); 308 case Type::FunctionTyID: 309 return *Entry = FunctionType::get(ElementTypes[0], 310 makeArrayRef(ElementTypes).slice(1), 311 cast<FunctionType>(Ty)->isVarArg()); 312 case Type::StructTyID: { 313 auto *STy = cast<StructType>(Ty); 314 bool IsPacked = STy->isPacked(); 315 if (IsUniqued) 316 return *Entry = StructType::get(Ty->getContext(), ElementTypes, IsPacked); 317 318 // If the type is opaque, we can just use it directly. 319 if (STy->isOpaque()) { 320 DstStructTypesSet.addOpaque(STy); 321 return *Entry = Ty; 322 } 323 324 if (StructType *OldT = 325 DstStructTypesSet.findNonOpaque(ElementTypes, IsPacked)) { 326 STy->setName(""); 327 return *Entry = OldT; 328 } 329 330 if (!AnyChange) { 331 DstStructTypesSet.addNonOpaque(STy); 332 return *Entry = Ty; 333 } 334 335 StructType *DTy = StructType::create(Ty->getContext()); 336 finishType(DTy, STy, ElementTypes); 337 return *Entry = DTy; 338 } 339 } 340 } 341 342 LinkDiagnosticInfo::LinkDiagnosticInfo(DiagnosticSeverity Severity, 343 const Twine &Msg) 344 : DiagnosticInfo(DK_Linker, Severity), Msg(Msg) {} 345 void LinkDiagnosticInfo::print(DiagnosticPrinter &DP) const { DP << Msg; } 346 347 //===----------------------------------------------------------------------===// 348 // IRLinker implementation. 349 //===----------------------------------------------------------------------===// 350 351 namespace { 352 class IRLinker; 353 354 /// Creates prototypes for functions that are lazily linked on the fly. This 355 /// speeds up linking for modules with many/ lazily linked functions of which 356 /// few get used. 357 class GlobalValueMaterializer final : public ValueMaterializer { 358 IRLinker &TheIRLinker; 359 360 public: 361 GlobalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {} 362 Value *materialize(Value *V) override; 363 }; 364 365 class LocalValueMaterializer final : public ValueMaterializer { 366 IRLinker &TheIRLinker; 367 368 public: 369 LocalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {} 370 Value *materialize(Value *V) override; 371 }; 372 373 /// Type of the Metadata map in \a ValueToValueMapTy. 374 typedef DenseMap<const Metadata *, TrackingMDRef> MDMapT; 375 376 /// This is responsible for keeping track of the state used for moving data 377 /// from SrcM to DstM. 378 class IRLinker { 379 Module &DstM; 380 std::unique_ptr<Module> SrcM; 381 382 /// See IRMover::move(). 383 std::function<void(GlobalValue &, IRMover::ValueAdder)> AddLazyFor; 384 385 TypeMapTy TypeMap; 386 GlobalValueMaterializer GValMaterializer; 387 LocalValueMaterializer LValMaterializer; 388 389 /// A metadata map that's shared between IRLinker instances. 390 MDMapT &SharedMDs; 391 392 /// Mapping of values from what they used to be in Src, to what they are now 393 /// in DstM. ValueToValueMapTy is a ValueMap, which involves some overhead 394 /// due to the use of Value handles which the Linker doesn't actually need, 395 /// but this allows us to reuse the ValueMapper code. 396 ValueToValueMapTy ValueMap; 397 ValueToValueMapTy IndirectSymbolValueMap; 398 399 DenseSet<GlobalValue *> ValuesToLink; 400 std::vector<GlobalValue *> Worklist; 401 std::vector<std::pair<GlobalValue *, Value*>> RAUWWorklist; 402 403 void maybeAdd(GlobalValue *GV) { 404 if (ValuesToLink.insert(GV).second) 405 Worklist.push_back(GV); 406 } 407 408 /// Whether we are importing globals for ThinLTO, as opposed to linking the 409 /// source module. If this flag is set, it means that we can rely on some 410 /// other object file to define any non-GlobalValue entities defined by the 411 /// source module. This currently causes us to not link retained types in 412 /// debug info metadata and module inline asm. 413 bool IsPerformingImport; 414 415 /// Set to true when all global value body linking is complete (including 416 /// lazy linking). Used to prevent metadata linking from creating new 417 /// references. 418 bool DoneLinkingBodies = false; 419 420 /// The Error encountered during materialization. We use an Optional here to 421 /// avoid needing to manage an unconsumed success value. 422 Optional<Error> FoundError; 423 void setError(Error E) { 424 if (E) 425 FoundError = std::move(E); 426 } 427 428 /// Most of the errors produced by this module are inconvertible StringErrors. 429 /// This convenience function lets us return one of those more easily. 430 Error stringErr(const Twine &T) { 431 return make_error<StringError>(T, inconvertibleErrorCode()); 432 } 433 434 /// Entry point for mapping values and alternate context for mapping aliases. 435 ValueMapper Mapper; 436 unsigned IndirectSymbolMCID; 437 438 /// Handles cloning of a global values from the source module into 439 /// the destination module, including setting the attributes and visibility. 440 GlobalValue *copyGlobalValueProto(const GlobalValue *SGV, bool ForDefinition); 441 442 void emitWarning(const Twine &Message) { 443 SrcM->getContext().diagnose(LinkDiagnosticInfo(DS_Warning, Message)); 444 } 445 446 /// Given a global in the source module, return the global in the 447 /// destination module that is being linked to, if any. 448 GlobalValue *getLinkedToGlobal(const GlobalValue *SrcGV) { 449 // If the source has no name it can't link. If it has local linkage, 450 // there is no name match-up going on. 451 if (!SrcGV->hasName() || SrcGV->hasLocalLinkage()) 452 return nullptr; 453 454 // Otherwise see if we have a match in the destination module's symtab. 455 GlobalValue *DGV = DstM.getNamedValue(SrcGV->getName()); 456 if (!DGV) 457 return nullptr; 458 459 // If we found a global with the same name in the dest module, but it has 460 // internal linkage, we are really not doing any linkage here. 461 if (DGV->hasLocalLinkage()) 462 return nullptr; 463 464 // If we found an intrinsic declaration with mismatching prototypes, we 465 // probably had a nameclash. Don't use that version. 466 if (auto *FDGV = dyn_cast<Function>(DGV)) 467 if (FDGV->isIntrinsic()) 468 if (const auto *FSrcGV = dyn_cast<Function>(SrcGV)) 469 if (FDGV->getFunctionType() != TypeMap.get(FSrcGV->getFunctionType())) 470 return nullptr; 471 472 // Otherwise, we do in fact link to the destination global. 473 return DGV; 474 } 475 476 void computeTypeMapping(); 477 478 Expected<Constant *> linkAppendingVarProto(GlobalVariable *DstGV, 479 const GlobalVariable *SrcGV); 480 481 /// Given the GlobaValue \p SGV in the source module, and the matching 482 /// GlobalValue \p DGV (if any), return true if the linker will pull \p SGV 483 /// into the destination module. 484 /// 485 /// Note this code may call the client-provided \p AddLazyFor. 486 bool shouldLink(GlobalValue *DGV, GlobalValue &SGV); 487 Expected<Constant *> linkGlobalValueProto(GlobalValue *GV, 488 bool ForIndirectSymbol); 489 490 Error linkModuleFlagsMetadata(); 491 492 void linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src); 493 Error linkFunctionBody(Function &Dst, Function &Src); 494 void linkIndirectSymbolBody(GlobalIndirectSymbol &Dst, 495 GlobalIndirectSymbol &Src); 496 Error linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src); 497 498 /// Replace all types in the source AttributeList with the 499 /// corresponding destination type. 500 AttributeList mapAttributeTypes(LLVMContext &C, AttributeList Attrs); 501 502 /// Functions that take care of cloning a specific global value type 503 /// into the destination module. 504 GlobalVariable *copyGlobalVariableProto(const GlobalVariable *SGVar); 505 Function *copyFunctionProto(const Function *SF); 506 GlobalValue *copyGlobalIndirectSymbolProto(const GlobalIndirectSymbol *SGIS); 507 508 /// Perform "replace all uses with" operations. These work items need to be 509 /// performed as part of materialization, but we postpone them to happen after 510 /// materialization is done. The materializer called by ValueMapper is not 511 /// expected to delete constants, as ValueMapper is holding pointers to some 512 /// of them, but constant destruction may be indirectly triggered by RAUW. 513 /// Hence, the need to move this out of the materialization call chain. 514 void flushRAUWWorklist(); 515 516 /// When importing for ThinLTO, prevent importing of types listed on 517 /// the DICompileUnit that we don't need a copy of in the importing 518 /// module. 519 void prepareCompileUnitsForImport(); 520 void linkNamedMDNodes(); 521 522 public: 523 IRLinker(Module &DstM, MDMapT &SharedMDs, 524 IRMover::IdentifiedStructTypeSet &Set, std::unique_ptr<Module> SrcM, 525 ArrayRef<GlobalValue *> ValuesToLink, 526 std::function<void(GlobalValue &, IRMover::ValueAdder)> AddLazyFor, 527 bool IsPerformingImport) 528 : DstM(DstM), SrcM(std::move(SrcM)), AddLazyFor(std::move(AddLazyFor)), 529 TypeMap(Set), GValMaterializer(*this), LValMaterializer(*this), 530 SharedMDs(SharedMDs), IsPerformingImport(IsPerformingImport), 531 Mapper(ValueMap, RF_ReuseAndMutateDistinctMDs | RF_IgnoreMissingLocals, 532 &TypeMap, &GValMaterializer), 533 IndirectSymbolMCID(Mapper.registerAlternateMappingContext( 534 IndirectSymbolValueMap, &LValMaterializer)) { 535 ValueMap.getMDMap() = std::move(SharedMDs); 536 for (GlobalValue *GV : ValuesToLink) 537 maybeAdd(GV); 538 if (IsPerformingImport) 539 prepareCompileUnitsForImport(); 540 } 541 ~IRLinker() { SharedMDs = std::move(*ValueMap.getMDMap()); } 542 543 Error run(); 544 Value *materialize(Value *V, bool ForIndirectSymbol); 545 }; 546 } 547 548 /// The LLVM SymbolTable class autorenames globals that conflict in the symbol 549 /// table. This is good for all clients except for us. Go through the trouble 550 /// to force this back. 551 static void forceRenaming(GlobalValue *GV, StringRef Name) { 552 // If the global doesn't force its name or if it already has the right name, 553 // there is nothing for us to do. 554 if (GV->hasLocalLinkage() || GV->getName() == Name) 555 return; 556 557 Module *M = GV->getParent(); 558 559 // If there is a conflict, rename the conflict. 560 if (GlobalValue *ConflictGV = M->getNamedValue(Name)) { 561 GV->takeName(ConflictGV); 562 ConflictGV->setName(Name); // This will cause ConflictGV to get renamed 563 assert(ConflictGV->getName() != Name && "forceRenaming didn't work"); 564 } else { 565 GV->setName(Name); // Force the name back 566 } 567 } 568 569 Value *GlobalValueMaterializer::materialize(Value *SGV) { 570 return TheIRLinker.materialize(SGV, false); 571 } 572 573 Value *LocalValueMaterializer::materialize(Value *SGV) { 574 return TheIRLinker.materialize(SGV, true); 575 } 576 577 Value *IRLinker::materialize(Value *V, bool ForIndirectSymbol) { 578 auto *SGV = dyn_cast<GlobalValue>(V); 579 if (!SGV) 580 return nullptr; 581 582 // When linking a global from other modules than source & dest, skip 583 // materializing it because it would be mapped later when its containing 584 // module is linked. Linking it now would potentially pull in many types that 585 // may not be mapped properly. 586 if (SGV->getParent() != &DstM && SGV->getParent() != SrcM.get()) 587 return nullptr; 588 589 Expected<Constant *> NewProto = linkGlobalValueProto(SGV, ForIndirectSymbol); 590 if (!NewProto) { 591 setError(NewProto.takeError()); 592 return nullptr; 593 } 594 if (!*NewProto) 595 return nullptr; 596 597 GlobalValue *New = dyn_cast<GlobalValue>(*NewProto); 598 if (!New) 599 return *NewProto; 600 601 // If we already created the body, just return. 602 if (auto *F = dyn_cast<Function>(New)) { 603 if (!F->isDeclaration()) 604 return New; 605 } else if (auto *V = dyn_cast<GlobalVariable>(New)) { 606 if (V->hasInitializer() || V->hasAppendingLinkage()) 607 return New; 608 } else { 609 auto *IS = cast<GlobalIndirectSymbol>(New); 610 if (IS->getIndirectSymbol()) 611 return New; 612 } 613 614 // If the global is being linked for an indirect symbol, it may have already 615 // been scheduled to satisfy a regular symbol. Similarly, a global being linked 616 // for a regular symbol may have already been scheduled for an indirect 617 // symbol. Check for these cases by looking in the other value map and 618 // confirming the same value has been scheduled. If there is an entry in the 619 // ValueMap but the value is different, it means that the value already had a 620 // definition in the destination module (linkonce for instance), but we need a 621 // new definition for the indirect symbol ("New" will be different). 622 if ((ForIndirectSymbol && ValueMap.lookup(SGV) == New) || 623 (!ForIndirectSymbol && IndirectSymbolValueMap.lookup(SGV) == New)) 624 return New; 625 626 if (ForIndirectSymbol || shouldLink(New, *SGV)) 627 setError(linkGlobalValueBody(*New, *SGV)); 628 629 return New; 630 } 631 632 /// Loop through the global variables in the src module and merge them into the 633 /// dest module. 634 GlobalVariable *IRLinker::copyGlobalVariableProto(const GlobalVariable *SGVar) { 635 // No linking to be performed or linking from the source: simply create an 636 // identical version of the symbol over in the dest module... the 637 // initializer will be filled in later by LinkGlobalInits. 638 GlobalVariable *NewDGV = 639 new GlobalVariable(DstM, TypeMap.get(SGVar->getValueType()), 640 SGVar->isConstant(), GlobalValue::ExternalLinkage, 641 /*init*/ nullptr, SGVar->getName(), 642 /*insertbefore*/ nullptr, SGVar->getThreadLocalMode(), 643 SGVar->getAddressSpace()); 644 NewDGV->setAlignment(MaybeAlign(SGVar->getAlignment())); 645 NewDGV->copyAttributesFrom(SGVar); 646 return NewDGV; 647 } 648 649 AttributeList IRLinker::mapAttributeTypes(LLVMContext &C, AttributeList Attrs) { 650 for (unsigned i = 0; i < Attrs.getNumAttrSets(); ++i) { 651 for (Attribute::AttrKind TypedAttr : 652 {Attribute::ByVal, Attribute::StructRet, Attribute::ByRef, 653 Attribute::InAlloca}) { 654 if (Attrs.hasAttribute(i, TypedAttr)) { 655 if (Type *Ty = Attrs.getAttribute(i, TypedAttr).getValueAsType()) { 656 Attrs = Attrs.replaceAttributeType(C, i, TypedAttr, TypeMap.get(Ty)); 657 break; 658 } 659 } 660 } 661 } 662 return Attrs; 663 } 664 665 /// Link the function in the source module into the destination module if 666 /// needed, setting up mapping information. 667 Function *IRLinker::copyFunctionProto(const Function *SF) { 668 // If there is no linkage to be performed or we are linking from the source, 669 // bring SF over. 670 auto *F = Function::Create(TypeMap.get(SF->getFunctionType()), 671 GlobalValue::ExternalLinkage, 672 SF->getAddressSpace(), SF->getName(), &DstM); 673 F->copyAttributesFrom(SF); 674 F->setAttributes(mapAttributeTypes(F->getContext(), F->getAttributes())); 675 return F; 676 } 677 678 /// Set up prototypes for any indirect symbols that come over from the source 679 /// module. 680 GlobalValue * 681 IRLinker::copyGlobalIndirectSymbolProto(const GlobalIndirectSymbol *SGIS) { 682 // If there is no linkage to be performed or we're linking from the source, 683 // bring over SGA. 684 auto *Ty = TypeMap.get(SGIS->getValueType()); 685 GlobalIndirectSymbol *GIS; 686 if (isa<GlobalAlias>(SGIS)) 687 GIS = GlobalAlias::create(Ty, SGIS->getAddressSpace(), 688 GlobalValue::ExternalLinkage, SGIS->getName(), 689 &DstM); 690 else 691 GIS = GlobalIFunc::create(Ty, SGIS->getAddressSpace(), 692 GlobalValue::ExternalLinkage, SGIS->getName(), 693 nullptr, &DstM); 694 GIS->copyAttributesFrom(SGIS); 695 return GIS; 696 } 697 698 GlobalValue *IRLinker::copyGlobalValueProto(const GlobalValue *SGV, 699 bool ForDefinition) { 700 GlobalValue *NewGV; 701 if (auto *SGVar = dyn_cast<GlobalVariable>(SGV)) { 702 NewGV = copyGlobalVariableProto(SGVar); 703 } else if (auto *SF = dyn_cast<Function>(SGV)) { 704 NewGV = copyFunctionProto(SF); 705 } else { 706 if (ForDefinition) 707 NewGV = copyGlobalIndirectSymbolProto(cast<GlobalIndirectSymbol>(SGV)); 708 else if (SGV->getValueType()->isFunctionTy()) 709 NewGV = 710 Function::Create(cast<FunctionType>(TypeMap.get(SGV->getValueType())), 711 GlobalValue::ExternalLinkage, SGV->getAddressSpace(), 712 SGV->getName(), &DstM); 713 else 714 NewGV = 715 new GlobalVariable(DstM, TypeMap.get(SGV->getValueType()), 716 /*isConstant*/ false, GlobalValue::ExternalLinkage, 717 /*init*/ nullptr, SGV->getName(), 718 /*insertbefore*/ nullptr, 719 SGV->getThreadLocalMode(), SGV->getAddressSpace()); 720 } 721 722 if (ForDefinition) 723 NewGV->setLinkage(SGV->getLinkage()); 724 else if (SGV->hasExternalWeakLinkage()) 725 NewGV->setLinkage(GlobalValue::ExternalWeakLinkage); 726 727 if (auto *NewGO = dyn_cast<GlobalObject>(NewGV)) { 728 // Metadata for global variables and function declarations is copied eagerly. 729 if (isa<GlobalVariable>(SGV) || SGV->isDeclaration()) 730 NewGO->copyMetadata(cast<GlobalObject>(SGV), 0); 731 } 732 733 // Remove these copied constants in case this stays a declaration, since 734 // they point to the source module. If the def is linked the values will 735 // be mapped in during linkFunctionBody. 736 if (auto *NewF = dyn_cast<Function>(NewGV)) { 737 NewF->setPersonalityFn(nullptr); 738 NewF->setPrefixData(nullptr); 739 NewF->setPrologueData(nullptr); 740 } 741 742 return NewGV; 743 } 744 745 static StringRef getTypeNamePrefix(StringRef Name) { 746 size_t DotPos = Name.rfind('.'); 747 return (DotPos == 0 || DotPos == StringRef::npos || Name.back() == '.' || 748 !isdigit(static_cast<unsigned char>(Name[DotPos + 1]))) 749 ? Name 750 : Name.substr(0, DotPos); 751 } 752 753 /// Loop over all of the linked values to compute type mappings. For example, 754 /// if we link "extern Foo *x" and "Foo *x = NULL", then we have two struct 755 /// types 'Foo' but one got renamed when the module was loaded into the same 756 /// LLVMContext. 757 void IRLinker::computeTypeMapping() { 758 for (GlobalValue &SGV : SrcM->globals()) { 759 GlobalValue *DGV = getLinkedToGlobal(&SGV); 760 if (!DGV) 761 continue; 762 763 if (!DGV->hasAppendingLinkage() || !SGV.hasAppendingLinkage()) { 764 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 765 continue; 766 } 767 768 // Unify the element type of appending arrays. 769 ArrayType *DAT = cast<ArrayType>(DGV->getValueType()); 770 ArrayType *SAT = cast<ArrayType>(SGV.getValueType()); 771 TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType()); 772 } 773 774 for (GlobalValue &SGV : *SrcM) 775 if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) { 776 if (DGV->getType() == SGV.getType()) { 777 // If the types of DGV and SGV are the same, it means that DGV is from 778 // the source module and got added to DstM from a shared metadata. We 779 // shouldn't map this type to itself in case the type's components get 780 // remapped to a new type from DstM (for instance, during the loop over 781 // SrcM->getIdentifiedStructTypes() below). 782 continue; 783 } 784 785 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 786 } 787 788 for (GlobalValue &SGV : SrcM->aliases()) 789 if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) 790 TypeMap.addTypeMapping(DGV->getType(), SGV.getType()); 791 792 // Incorporate types by name, scanning all the types in the source module. 793 // At this point, the destination module may have a type "%foo = { i32 }" for 794 // example. When the source module got loaded into the same LLVMContext, if 795 // it had the same type, it would have been renamed to "%foo.42 = { i32 }". 796 std::vector<StructType *> Types = SrcM->getIdentifiedStructTypes(); 797 for (StructType *ST : Types) { 798 if (!ST->hasName()) 799 continue; 800 801 if (TypeMap.DstStructTypesSet.hasType(ST)) { 802 // This is actually a type from the destination module. 803 // getIdentifiedStructTypes() can have found it by walking debug info 804 // metadata nodes, some of which get linked by name when ODR Type Uniquing 805 // is enabled on the Context, from the source to the destination module. 806 continue; 807 } 808 809 auto STTypePrefix = getTypeNamePrefix(ST->getName()); 810 if (STTypePrefix.size() == ST->getName().size()) 811 continue; 812 813 // Check to see if the destination module has a struct with the prefix name. 814 StructType *DST = StructType::getTypeByName(ST->getContext(), STTypePrefix); 815 if (!DST) 816 continue; 817 818 // Don't use it if this actually came from the source module. They're in 819 // the same LLVMContext after all. Also don't use it unless the type is 820 // actually used in the destination module. This can happen in situations 821 // like this: 822 // 823 // Module A Module B 824 // -------- -------- 825 // %Z = type { %A } %B = type { %C.1 } 826 // %A = type { %B.1, [7 x i8] } %C.1 = type { i8* } 827 // %B.1 = type { %C } %A.2 = type { %B.3, [5 x i8] } 828 // %C = type { i8* } %B.3 = type { %C.1 } 829 // 830 // When we link Module B with Module A, the '%B' in Module B is 831 // used. However, that would then use '%C.1'. But when we process '%C.1', 832 // we prefer to take the '%C' version. So we are then left with both 833 // '%C.1' and '%C' being used for the same types. This leads to some 834 // variables using one type and some using the other. 835 if (TypeMap.DstStructTypesSet.hasType(DST)) 836 TypeMap.addTypeMapping(DST, ST); 837 } 838 839 // Now that we have discovered all of the type equivalences, get a body for 840 // any 'opaque' types in the dest module that are now resolved. 841 TypeMap.linkDefinedTypeBodies(); 842 } 843 844 static void getArrayElements(const Constant *C, 845 SmallVectorImpl<Constant *> &Dest) { 846 unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements(); 847 848 for (unsigned i = 0; i != NumElements; ++i) 849 Dest.push_back(C->getAggregateElement(i)); 850 } 851 852 /// If there were any appending global variables, link them together now. 853 Expected<Constant *> 854 IRLinker::linkAppendingVarProto(GlobalVariable *DstGV, 855 const GlobalVariable *SrcGV) { 856 // Check that both variables have compatible properties. 857 if (DstGV && !DstGV->isDeclaration() && !SrcGV->isDeclaration()) { 858 if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage()) 859 return stringErr( 860 "Linking globals named '" + SrcGV->getName() + 861 "': can only link appending global with another appending " 862 "global!"); 863 864 if (DstGV->isConstant() != SrcGV->isConstant()) 865 return stringErr("Appending variables linked with different const'ness!"); 866 867 if (DstGV->getAlignment() != SrcGV->getAlignment()) 868 return stringErr( 869 "Appending variables with different alignment need to be linked!"); 870 871 if (DstGV->getVisibility() != SrcGV->getVisibility()) 872 return stringErr( 873 "Appending variables with different visibility need to be linked!"); 874 875 if (DstGV->hasGlobalUnnamedAddr() != SrcGV->hasGlobalUnnamedAddr()) 876 return stringErr( 877 "Appending variables with different unnamed_addr need to be linked!"); 878 879 if (DstGV->getSection() != SrcGV->getSection()) 880 return stringErr( 881 "Appending variables with different section name need to be linked!"); 882 } 883 884 // Do not need to do anything if source is a declaration. 885 if (SrcGV->isDeclaration()) 886 return DstGV; 887 888 Type *EltTy = cast<ArrayType>(TypeMap.get(SrcGV->getValueType())) 889 ->getElementType(); 890 891 // FIXME: This upgrade is done during linking to support the C API. Once the 892 // old form is deprecated, we should move this upgrade to 893 // llvm::UpgradeGlobalVariable() and simplify the logic here and in 894 // Mapper::mapAppendingVariable() in ValueMapper.cpp. 895 StringRef Name = SrcGV->getName(); 896 bool IsNewStructor = false; 897 bool IsOldStructor = false; 898 if (Name == "llvm.global_ctors" || Name == "llvm.global_dtors") { 899 if (cast<StructType>(EltTy)->getNumElements() == 3) 900 IsNewStructor = true; 901 else 902 IsOldStructor = true; 903 } 904 905 PointerType *VoidPtrTy = Type::getInt8Ty(SrcGV->getContext())->getPointerTo(); 906 if (IsOldStructor) { 907 auto &ST = *cast<StructType>(EltTy); 908 Type *Tys[3] = {ST.getElementType(0), ST.getElementType(1), VoidPtrTy}; 909 EltTy = StructType::get(SrcGV->getContext(), Tys, false); 910 } 911 912 uint64_t DstNumElements = 0; 913 if (DstGV && !DstGV->isDeclaration()) { 914 ArrayType *DstTy = cast<ArrayType>(DstGV->getValueType()); 915 DstNumElements = DstTy->getNumElements(); 916 917 // Check to see that they two arrays agree on type. 918 if (EltTy != DstTy->getElementType()) 919 return stringErr("Appending variables with different element types!"); 920 } 921 922 SmallVector<Constant *, 16> SrcElements; 923 getArrayElements(SrcGV->getInitializer(), SrcElements); 924 925 if (IsNewStructor) { 926 erase_if(SrcElements, [this](Constant *E) { 927 auto *Key = 928 dyn_cast<GlobalValue>(E->getAggregateElement(2)->stripPointerCasts()); 929 if (!Key) 930 return false; 931 GlobalValue *DGV = getLinkedToGlobal(Key); 932 return !shouldLink(DGV, *Key); 933 }); 934 } 935 uint64_t NewSize = DstNumElements + SrcElements.size(); 936 ArrayType *NewType = ArrayType::get(EltTy, NewSize); 937 938 // Create the new global variable. 939 GlobalVariable *NG = new GlobalVariable( 940 DstM, NewType, SrcGV->isConstant(), SrcGV->getLinkage(), 941 /*init*/ nullptr, /*name*/ "", DstGV, SrcGV->getThreadLocalMode(), 942 SrcGV->getAddressSpace()); 943 944 NG->copyAttributesFrom(SrcGV); 945 forceRenaming(NG, SrcGV->getName()); 946 947 Constant *Ret = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType())); 948 949 Mapper.scheduleMapAppendingVariable( 950 *NG, 951 (DstGV && !DstGV->isDeclaration()) ? DstGV->getInitializer() : nullptr, 952 IsOldStructor, SrcElements); 953 954 // Replace any uses of the two global variables with uses of the new 955 // global. 956 if (DstGV) { 957 RAUWWorklist.push_back( 958 std::make_pair(DstGV, ConstantExpr::getBitCast(NG, DstGV->getType()))); 959 } 960 961 return Ret; 962 } 963 964 bool IRLinker::shouldLink(GlobalValue *DGV, GlobalValue &SGV) { 965 if (ValuesToLink.count(&SGV) || SGV.hasLocalLinkage()) 966 return true; 967 968 if (DGV && !DGV->isDeclarationForLinker()) 969 return false; 970 971 if (SGV.isDeclaration() || DoneLinkingBodies) 972 return false; 973 974 // Callback to the client to give a chance to lazily add the Global to the 975 // list of value to link. 976 bool LazilyAdded = false; 977 AddLazyFor(SGV, [this, &LazilyAdded](GlobalValue &GV) { 978 maybeAdd(&GV); 979 LazilyAdded = true; 980 }); 981 return LazilyAdded; 982 } 983 984 Expected<Constant *> IRLinker::linkGlobalValueProto(GlobalValue *SGV, 985 bool ForIndirectSymbol) { 986 GlobalValue *DGV = getLinkedToGlobal(SGV); 987 988 bool ShouldLink = shouldLink(DGV, *SGV); 989 990 // just missing from map 991 if (ShouldLink) { 992 auto I = ValueMap.find(SGV); 993 if (I != ValueMap.end()) 994 return cast<Constant>(I->second); 995 996 I = IndirectSymbolValueMap.find(SGV); 997 if (I != IndirectSymbolValueMap.end()) 998 return cast<Constant>(I->second); 999 } 1000 1001 if (!ShouldLink && ForIndirectSymbol) 1002 DGV = nullptr; 1003 1004 // Handle the ultra special appending linkage case first. 1005 if (SGV->hasAppendingLinkage() || (DGV && DGV->hasAppendingLinkage())) 1006 return linkAppendingVarProto(cast_or_null<GlobalVariable>(DGV), 1007 cast<GlobalVariable>(SGV)); 1008 1009 bool NeedsRenaming = false; 1010 GlobalValue *NewGV; 1011 if (DGV && !ShouldLink) { 1012 NewGV = DGV; 1013 } else { 1014 // If we are done linking global value bodies (i.e. we are performing 1015 // metadata linking), don't link in the global value due to this 1016 // reference, simply map it to null. 1017 if (DoneLinkingBodies) 1018 return nullptr; 1019 1020 NewGV = copyGlobalValueProto(SGV, ShouldLink || ForIndirectSymbol); 1021 if (ShouldLink || !ForIndirectSymbol) 1022 NeedsRenaming = true; 1023 } 1024 1025 // Overloaded intrinsics have overloaded types names as part of their 1026 // names. If we renamed overloaded types we should rename the intrinsic 1027 // as well. 1028 if (Function *F = dyn_cast<Function>(NewGV)) 1029 if (auto Remangled = Intrinsic::remangleIntrinsicFunction(F)) { 1030 NewGV->eraseFromParent(); 1031 NewGV = Remangled.getValue(); 1032 NeedsRenaming = false; 1033 } 1034 1035 if (NeedsRenaming) 1036 forceRenaming(NewGV, SGV->getName()); 1037 1038 if (ShouldLink || ForIndirectSymbol) { 1039 if (const Comdat *SC = SGV->getComdat()) { 1040 if (auto *GO = dyn_cast<GlobalObject>(NewGV)) { 1041 Comdat *DC = DstM.getOrInsertComdat(SC->getName()); 1042 DC->setSelectionKind(SC->getSelectionKind()); 1043 GO->setComdat(DC); 1044 } 1045 } 1046 } 1047 1048 if (!ShouldLink && ForIndirectSymbol) 1049 NewGV->setLinkage(GlobalValue::InternalLinkage); 1050 1051 Constant *C = NewGV; 1052 // Only create a bitcast if necessary. In particular, with 1053 // DebugTypeODRUniquing we may reach metadata in the destination module 1054 // containing a GV from the source module, in which case SGV will be 1055 // the same as DGV and NewGV, and TypeMap.get() will assert since it 1056 // assumes it is being invoked on a type in the source module. 1057 if (DGV && NewGV != SGV) { 1058 C = ConstantExpr::getPointerBitCastOrAddrSpaceCast( 1059 NewGV, TypeMap.get(SGV->getType())); 1060 } 1061 1062 if (DGV && NewGV != DGV) { 1063 // Schedule "replace all uses with" to happen after materializing is 1064 // done. It is not safe to do it now, since ValueMapper may be holding 1065 // pointers to constants that will get deleted if RAUW runs. 1066 RAUWWorklist.push_back(std::make_pair( 1067 DGV, 1068 ConstantExpr::getPointerBitCastOrAddrSpaceCast(NewGV, DGV->getType()))); 1069 } 1070 1071 return C; 1072 } 1073 1074 /// Update the initializers in the Dest module now that all globals that may be 1075 /// referenced are in Dest. 1076 void IRLinker::linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src) { 1077 // Figure out what the initializer looks like in the dest module. 1078 Mapper.scheduleMapGlobalInitializer(Dst, *Src.getInitializer()); 1079 } 1080 1081 /// Copy the source function over into the dest function and fix up references 1082 /// to values. At this point we know that Dest is an external function, and 1083 /// that Src is not. 1084 Error IRLinker::linkFunctionBody(Function &Dst, Function &Src) { 1085 assert(Dst.isDeclaration() && !Src.isDeclaration()); 1086 1087 // Materialize if needed. 1088 if (Error Err = Src.materialize()) 1089 return Err; 1090 1091 // Link in the operands without remapping. 1092 if (Src.hasPrefixData()) 1093 Dst.setPrefixData(Src.getPrefixData()); 1094 if (Src.hasPrologueData()) 1095 Dst.setPrologueData(Src.getPrologueData()); 1096 if (Src.hasPersonalityFn()) 1097 Dst.setPersonalityFn(Src.getPersonalityFn()); 1098 1099 // Copy over the metadata attachments without remapping. 1100 Dst.copyMetadata(&Src, 0); 1101 1102 // Steal arguments and splice the body of Src into Dst. 1103 Dst.stealArgumentListFrom(Src); 1104 Dst.getBasicBlockList().splice(Dst.end(), Src.getBasicBlockList()); 1105 1106 // Everything has been moved over. Remap it. 1107 Mapper.scheduleRemapFunction(Dst); 1108 return Error::success(); 1109 } 1110 1111 void IRLinker::linkIndirectSymbolBody(GlobalIndirectSymbol &Dst, 1112 GlobalIndirectSymbol &Src) { 1113 Mapper.scheduleMapGlobalIndirectSymbol(Dst, *Src.getIndirectSymbol(), 1114 IndirectSymbolMCID); 1115 } 1116 1117 Error IRLinker::linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src) { 1118 if (auto *F = dyn_cast<Function>(&Src)) 1119 return linkFunctionBody(cast<Function>(Dst), *F); 1120 if (auto *GVar = dyn_cast<GlobalVariable>(&Src)) { 1121 linkGlobalVariable(cast<GlobalVariable>(Dst), *GVar); 1122 return Error::success(); 1123 } 1124 linkIndirectSymbolBody(cast<GlobalIndirectSymbol>(Dst), cast<GlobalIndirectSymbol>(Src)); 1125 return Error::success(); 1126 } 1127 1128 void IRLinker::flushRAUWWorklist() { 1129 for (const auto &Elem : RAUWWorklist) { 1130 GlobalValue *Old; 1131 Value *New; 1132 std::tie(Old, New) = Elem; 1133 1134 Old->replaceAllUsesWith(New); 1135 Old->eraseFromParent(); 1136 } 1137 RAUWWorklist.clear(); 1138 } 1139 1140 void IRLinker::prepareCompileUnitsForImport() { 1141 NamedMDNode *SrcCompileUnits = SrcM->getNamedMetadata("llvm.dbg.cu"); 1142 if (!SrcCompileUnits) 1143 return; 1144 // When importing for ThinLTO, prevent importing of types listed on 1145 // the DICompileUnit that we don't need a copy of in the importing 1146 // module. They will be emitted by the originating module. 1147 for (unsigned I = 0, E = SrcCompileUnits->getNumOperands(); I != E; ++I) { 1148 auto *CU = cast<DICompileUnit>(SrcCompileUnits->getOperand(I)); 1149 assert(CU && "Expected valid compile unit"); 1150 // Enums, macros, and retained types don't need to be listed on the 1151 // imported DICompileUnit. This means they will only be imported 1152 // if reached from the mapped IR. 1153 CU->replaceEnumTypes(nullptr); 1154 CU->replaceMacros(nullptr); 1155 CU->replaceRetainedTypes(nullptr); 1156 1157 // The original definition (or at least its debug info - if the variable is 1158 // internalized and optimized away) will remain in the source module, so 1159 // there's no need to import them. 1160 // If LLVM ever does more advanced optimizations on global variables 1161 // (removing/localizing write operations, for instance) that can track 1162 // through debug info, this decision may need to be revisited - but do so 1163 // with care when it comes to debug info size. Emitting small CUs containing 1164 // only a few imported entities into every destination module may be very 1165 // size inefficient. 1166 CU->replaceGlobalVariables(nullptr); 1167 1168 // Imported entities only need to be mapped in if they have local 1169 // scope, as those might correspond to an imported entity inside a 1170 // function being imported (any locally scoped imported entities that 1171 // don't end up referenced by an imported function will not be emitted 1172 // into the object). Imported entities not in a local scope 1173 // (e.g. on the namespace) only need to be emitted by the originating 1174 // module. Create a list of the locally scoped imported entities, and 1175 // replace the source CUs imported entity list with the new list, so 1176 // only those are mapped in. 1177 // FIXME: Locally-scoped imported entities could be moved to the 1178 // functions they are local to instead of listing them on the CU, and 1179 // we would naturally only link in those needed by function importing. 1180 SmallVector<TrackingMDNodeRef, 4> AllImportedModules; 1181 bool ReplaceImportedEntities = false; 1182 for (auto *IE : CU->getImportedEntities()) { 1183 DIScope *Scope = IE->getScope(); 1184 assert(Scope && "Invalid Scope encoding!"); 1185 if (isa<DILocalScope>(Scope)) 1186 AllImportedModules.emplace_back(IE); 1187 else 1188 ReplaceImportedEntities = true; 1189 } 1190 if (ReplaceImportedEntities) { 1191 if (!AllImportedModules.empty()) 1192 CU->replaceImportedEntities(MDTuple::get( 1193 CU->getContext(), 1194 SmallVector<Metadata *, 16>(AllImportedModules.begin(), 1195 AllImportedModules.end()))); 1196 else 1197 // If there were no local scope imported entities, we can map 1198 // the whole list to nullptr. 1199 CU->replaceImportedEntities(nullptr); 1200 } 1201 } 1202 } 1203 1204 /// Insert all of the named MDNodes in Src into the Dest module. 1205 void IRLinker::linkNamedMDNodes() { 1206 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata(); 1207 for (const NamedMDNode &NMD : SrcM->named_metadata()) { 1208 // Don't link module flags here. Do them separately. 1209 if (&NMD == SrcModFlags) 1210 continue; 1211 // Don't import pseudo probe descriptors here for thinLTO. They will be 1212 // emitted by the originating module. 1213 if (IsPerformingImport && NMD.getName() == PseudoProbeDescMetadataName) 1214 continue; 1215 NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName()); 1216 // Add Src elements into Dest node. 1217 for (const MDNode *Op : NMD.operands()) 1218 DestNMD->addOperand(Mapper.mapMDNode(*Op)); 1219 } 1220 } 1221 1222 /// Merge the linker flags in Src into the Dest module. 1223 Error IRLinker::linkModuleFlagsMetadata() { 1224 // If the source module has no module flags, we are done. 1225 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata(); 1226 if (!SrcModFlags) 1227 return Error::success(); 1228 1229 // If the destination module doesn't have module flags yet, then just copy 1230 // over the source module's flags. 1231 NamedMDNode *DstModFlags = DstM.getOrInsertModuleFlagsMetadata(); 1232 if (DstModFlags->getNumOperands() == 0) { 1233 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) 1234 DstModFlags->addOperand(SrcModFlags->getOperand(I)); 1235 1236 return Error::success(); 1237 } 1238 1239 // First build a map of the existing module flags and requirements. 1240 DenseMap<MDString *, std::pair<MDNode *, unsigned>> Flags; 1241 SmallSetVector<MDNode *, 16> Requirements; 1242 for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) { 1243 MDNode *Op = DstModFlags->getOperand(I); 1244 ConstantInt *Behavior = mdconst::extract<ConstantInt>(Op->getOperand(0)); 1245 MDString *ID = cast<MDString>(Op->getOperand(1)); 1246 1247 if (Behavior->getZExtValue() == Module::Require) { 1248 Requirements.insert(cast<MDNode>(Op->getOperand(2))); 1249 } else { 1250 Flags[ID] = std::make_pair(Op, I); 1251 } 1252 } 1253 1254 // Merge in the flags from the source module, and also collect its set of 1255 // requirements. 1256 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) { 1257 MDNode *SrcOp = SrcModFlags->getOperand(I); 1258 ConstantInt *SrcBehavior = 1259 mdconst::extract<ConstantInt>(SrcOp->getOperand(0)); 1260 MDString *ID = cast<MDString>(SrcOp->getOperand(1)); 1261 MDNode *DstOp; 1262 unsigned DstIndex; 1263 std::tie(DstOp, DstIndex) = Flags.lookup(ID); 1264 unsigned SrcBehaviorValue = SrcBehavior->getZExtValue(); 1265 1266 // If this is a requirement, add it and continue. 1267 if (SrcBehaviorValue == Module::Require) { 1268 // If the destination module does not already have this requirement, add 1269 // it. 1270 if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) { 1271 DstModFlags->addOperand(SrcOp); 1272 } 1273 continue; 1274 } 1275 1276 // If there is no existing flag with this ID, just add it. 1277 if (!DstOp) { 1278 Flags[ID] = std::make_pair(SrcOp, DstModFlags->getNumOperands()); 1279 DstModFlags->addOperand(SrcOp); 1280 continue; 1281 } 1282 1283 // Otherwise, perform a merge. 1284 ConstantInt *DstBehavior = 1285 mdconst::extract<ConstantInt>(DstOp->getOperand(0)); 1286 unsigned DstBehaviorValue = DstBehavior->getZExtValue(); 1287 1288 auto overrideDstValue = [&]() { 1289 DstModFlags->setOperand(DstIndex, SrcOp); 1290 Flags[ID].first = SrcOp; 1291 }; 1292 1293 // If either flag has override behavior, handle it first. 1294 if (DstBehaviorValue == Module::Override) { 1295 // Diagnose inconsistent flags which both have override behavior. 1296 if (SrcBehaviorValue == Module::Override && 1297 SrcOp->getOperand(2) != DstOp->getOperand(2)) 1298 return stringErr("linking module flags '" + ID->getString() + 1299 "': IDs have conflicting override values in '" + 1300 SrcM->getModuleIdentifier() + "' and '" + 1301 DstM.getModuleIdentifier() + "'"); 1302 continue; 1303 } else if (SrcBehaviorValue == Module::Override) { 1304 // Update the destination flag to that of the source. 1305 overrideDstValue(); 1306 continue; 1307 } 1308 1309 // Diagnose inconsistent merge behavior types. 1310 if (SrcBehaviorValue != DstBehaviorValue) { 1311 bool MaxAndWarn = (SrcBehaviorValue == Module::Max && 1312 DstBehaviorValue == Module::Warning) || 1313 (DstBehaviorValue == Module::Max && 1314 SrcBehaviorValue == Module::Warning); 1315 if (!MaxAndWarn) 1316 return stringErr("linking module flags '" + ID->getString() + 1317 "': IDs have conflicting behaviors in '" + 1318 SrcM->getModuleIdentifier() + "' and '" + 1319 DstM.getModuleIdentifier() + "'"); 1320 } 1321 1322 auto replaceDstValue = [&](MDNode *New) { 1323 Metadata *FlagOps[] = {DstOp->getOperand(0), ID, New}; 1324 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps); 1325 DstModFlags->setOperand(DstIndex, Flag); 1326 Flags[ID].first = Flag; 1327 }; 1328 1329 // Emit a warning if the values differ and either source or destination 1330 // request Warning behavior. 1331 if ((DstBehaviorValue == Module::Warning || 1332 SrcBehaviorValue == Module::Warning) && 1333 SrcOp->getOperand(2) != DstOp->getOperand(2)) { 1334 std::string Str; 1335 raw_string_ostream(Str) 1336 << "linking module flags '" << ID->getString() 1337 << "': IDs have conflicting values ('" << *SrcOp->getOperand(2) 1338 << "' from " << SrcM->getModuleIdentifier() << " with '" 1339 << *DstOp->getOperand(2) << "' from " << DstM.getModuleIdentifier() 1340 << ')'; 1341 emitWarning(Str); 1342 } 1343 1344 // Choose the maximum if either source or destination request Max behavior. 1345 if (DstBehaviorValue == Module::Max || SrcBehaviorValue == Module::Max) { 1346 ConstantInt *DstValue = 1347 mdconst::extract<ConstantInt>(DstOp->getOperand(2)); 1348 ConstantInt *SrcValue = 1349 mdconst::extract<ConstantInt>(SrcOp->getOperand(2)); 1350 1351 // The resulting flag should have a Max behavior, and contain the maximum 1352 // value from between the source and destination values. 1353 Metadata *FlagOps[] = { 1354 (DstBehaviorValue != Module::Max ? SrcOp : DstOp)->getOperand(0), ID, 1355 (SrcValue->getZExtValue() > DstValue->getZExtValue() ? SrcOp : DstOp) 1356 ->getOperand(2)}; 1357 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps); 1358 DstModFlags->setOperand(DstIndex, Flag); 1359 Flags[ID].first = Flag; 1360 continue; 1361 } 1362 1363 // Perform the merge for standard behavior types. 1364 switch (SrcBehaviorValue) { 1365 case Module::Require: 1366 case Module::Override: 1367 llvm_unreachable("not possible"); 1368 case Module::Error: { 1369 // Emit an error if the values differ. 1370 if (SrcOp->getOperand(2) != DstOp->getOperand(2)) 1371 return stringErr("linking module flags '" + ID->getString() + 1372 "': IDs have conflicting values in '" + 1373 SrcM->getModuleIdentifier() + "' and '" + 1374 DstM.getModuleIdentifier() + "'"); 1375 continue; 1376 } 1377 case Module::Warning: { 1378 break; 1379 } 1380 case Module::Max: { 1381 break; 1382 } 1383 case Module::Append: { 1384 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2)); 1385 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2)); 1386 SmallVector<Metadata *, 8> MDs; 1387 MDs.reserve(DstValue->getNumOperands() + SrcValue->getNumOperands()); 1388 MDs.append(DstValue->op_begin(), DstValue->op_end()); 1389 MDs.append(SrcValue->op_begin(), SrcValue->op_end()); 1390 1391 replaceDstValue(MDNode::get(DstM.getContext(), MDs)); 1392 break; 1393 } 1394 case Module::AppendUnique: { 1395 SmallSetVector<Metadata *, 16> Elts; 1396 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2)); 1397 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2)); 1398 Elts.insert(DstValue->op_begin(), DstValue->op_end()); 1399 Elts.insert(SrcValue->op_begin(), SrcValue->op_end()); 1400 1401 replaceDstValue(MDNode::get(DstM.getContext(), 1402 makeArrayRef(Elts.begin(), Elts.end()))); 1403 break; 1404 } 1405 } 1406 1407 } 1408 1409 // Check all of the requirements. 1410 for (unsigned I = 0, E = Requirements.size(); I != E; ++I) { 1411 MDNode *Requirement = Requirements[I]; 1412 MDString *Flag = cast<MDString>(Requirement->getOperand(0)); 1413 Metadata *ReqValue = Requirement->getOperand(1); 1414 1415 MDNode *Op = Flags[Flag].first; 1416 if (!Op || Op->getOperand(2) != ReqValue) 1417 return stringErr("linking module flags '" + Flag->getString() + 1418 "': does not have the required value"); 1419 } 1420 return Error::success(); 1421 } 1422 1423 /// Return InlineAsm adjusted with target-specific directives if required. 1424 /// For ARM and Thumb, we have to add directives to select the appropriate ISA 1425 /// to support mixing module-level inline assembly from ARM and Thumb modules. 1426 static std::string adjustInlineAsm(const std::string &InlineAsm, 1427 const Triple &Triple) { 1428 if (Triple.getArch() == Triple::thumb || Triple.getArch() == Triple::thumbeb) 1429 return ".text\n.balign 2\n.thumb\n" + InlineAsm; 1430 if (Triple.getArch() == Triple::arm || Triple.getArch() == Triple::armeb) 1431 return ".text\n.balign 4\n.arm\n" + InlineAsm; 1432 return InlineAsm; 1433 } 1434 1435 Error IRLinker::run() { 1436 // Ensure metadata materialized before value mapping. 1437 if (SrcM->getMaterializer()) 1438 if (Error Err = SrcM->getMaterializer()->materializeMetadata()) 1439 return Err; 1440 1441 // Inherit the target data from the source module if the destination module 1442 // doesn't have one already. 1443 if (DstM.getDataLayout().isDefault()) 1444 DstM.setDataLayout(SrcM->getDataLayout()); 1445 1446 if (SrcM->getDataLayout() != DstM.getDataLayout()) { 1447 emitWarning("Linking two modules of different data layouts: '" + 1448 SrcM->getModuleIdentifier() + "' is '" + 1449 SrcM->getDataLayoutStr() + "' whereas '" + 1450 DstM.getModuleIdentifier() + "' is '" + 1451 DstM.getDataLayoutStr() + "'\n"); 1452 } 1453 1454 // Copy the target triple from the source to dest if the dest's is empty. 1455 if (DstM.getTargetTriple().empty() && !SrcM->getTargetTriple().empty()) 1456 DstM.setTargetTriple(SrcM->getTargetTriple()); 1457 1458 Triple SrcTriple(SrcM->getTargetTriple()), DstTriple(DstM.getTargetTriple()); 1459 1460 if (!SrcM->getTargetTriple().empty()&& 1461 !SrcTriple.isCompatibleWith(DstTriple)) 1462 emitWarning("Linking two modules of different target triples: '" + 1463 SrcM->getModuleIdentifier() + "' is '" + 1464 SrcM->getTargetTriple() + "' whereas '" + 1465 DstM.getModuleIdentifier() + "' is '" + DstM.getTargetTriple() + 1466 "'\n"); 1467 1468 DstM.setTargetTriple(SrcTriple.merge(DstTriple)); 1469 1470 // Loop over all of the linked values to compute type mappings. 1471 computeTypeMapping(); 1472 1473 std::reverse(Worklist.begin(), Worklist.end()); 1474 while (!Worklist.empty()) { 1475 GlobalValue *GV = Worklist.back(); 1476 Worklist.pop_back(); 1477 1478 // Already mapped. 1479 if (ValueMap.find(GV) != ValueMap.end() || 1480 IndirectSymbolValueMap.find(GV) != IndirectSymbolValueMap.end()) 1481 continue; 1482 1483 assert(!GV->isDeclaration()); 1484 Mapper.mapValue(*GV); 1485 if (FoundError) 1486 return std::move(*FoundError); 1487 flushRAUWWorklist(); 1488 } 1489 1490 // Note that we are done linking global value bodies. This prevents 1491 // metadata linking from creating new references. 1492 DoneLinkingBodies = true; 1493 Mapper.addFlags(RF_NullMapMissingGlobalValues); 1494 1495 // Remap all of the named MDNodes in Src into the DstM module. We do this 1496 // after linking GlobalValues so that MDNodes that reference GlobalValues 1497 // are properly remapped. 1498 linkNamedMDNodes(); 1499 1500 if (!IsPerformingImport && !SrcM->getModuleInlineAsm().empty()) { 1501 // Append the module inline asm string. 1502 DstM.appendModuleInlineAsm(adjustInlineAsm(SrcM->getModuleInlineAsm(), 1503 SrcTriple)); 1504 } else if (IsPerformingImport) { 1505 // Import any symver directives for symbols in DstM. 1506 ModuleSymbolTable::CollectAsmSymvers(*SrcM, 1507 [&](StringRef Name, StringRef Alias) { 1508 if (DstM.getNamedValue(Name)) { 1509 SmallString<256> S(".symver "); 1510 S += Name; 1511 S += ", "; 1512 S += Alias; 1513 DstM.appendModuleInlineAsm(S); 1514 } 1515 }); 1516 } 1517 1518 // Reorder the globals just added to the destination module to match their 1519 // original order in the source module. 1520 Module::GlobalListType &Globals = DstM.getGlobalList(); 1521 for (GlobalVariable &GV : SrcM->globals()) { 1522 if (GV.hasAppendingLinkage()) 1523 continue; 1524 Value *NewValue = Mapper.mapValue(GV); 1525 if (NewValue) { 1526 auto *NewGV = dyn_cast<GlobalVariable>(NewValue->stripPointerCasts()); 1527 if (NewGV) 1528 Globals.splice(Globals.end(), Globals, NewGV->getIterator()); 1529 } 1530 } 1531 1532 // Merge the module flags into the DstM module. 1533 return linkModuleFlagsMetadata(); 1534 } 1535 1536 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(ArrayRef<Type *> E, bool P) 1537 : ETypes(E), IsPacked(P) {} 1538 1539 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(const StructType *ST) 1540 : ETypes(ST->elements()), IsPacked(ST->isPacked()) {} 1541 1542 bool IRMover::StructTypeKeyInfo::KeyTy::operator==(const KeyTy &That) const { 1543 return IsPacked == That.IsPacked && ETypes == That.ETypes; 1544 } 1545 1546 bool IRMover::StructTypeKeyInfo::KeyTy::operator!=(const KeyTy &That) const { 1547 return !this->operator==(That); 1548 } 1549 1550 StructType *IRMover::StructTypeKeyInfo::getEmptyKey() { 1551 return DenseMapInfo<StructType *>::getEmptyKey(); 1552 } 1553 1554 StructType *IRMover::StructTypeKeyInfo::getTombstoneKey() { 1555 return DenseMapInfo<StructType *>::getTombstoneKey(); 1556 } 1557 1558 unsigned IRMover::StructTypeKeyInfo::getHashValue(const KeyTy &Key) { 1559 return hash_combine(hash_combine_range(Key.ETypes.begin(), Key.ETypes.end()), 1560 Key.IsPacked); 1561 } 1562 1563 unsigned IRMover::StructTypeKeyInfo::getHashValue(const StructType *ST) { 1564 return getHashValue(KeyTy(ST)); 1565 } 1566 1567 bool IRMover::StructTypeKeyInfo::isEqual(const KeyTy &LHS, 1568 const StructType *RHS) { 1569 if (RHS == getEmptyKey() || RHS == getTombstoneKey()) 1570 return false; 1571 return LHS == KeyTy(RHS); 1572 } 1573 1574 bool IRMover::StructTypeKeyInfo::isEqual(const StructType *LHS, 1575 const StructType *RHS) { 1576 if (RHS == getEmptyKey() || RHS == getTombstoneKey()) 1577 return LHS == RHS; 1578 return KeyTy(LHS) == KeyTy(RHS); 1579 } 1580 1581 void IRMover::IdentifiedStructTypeSet::addNonOpaque(StructType *Ty) { 1582 assert(!Ty->isOpaque()); 1583 NonOpaqueStructTypes.insert(Ty); 1584 } 1585 1586 void IRMover::IdentifiedStructTypeSet::switchToNonOpaque(StructType *Ty) { 1587 assert(!Ty->isOpaque()); 1588 NonOpaqueStructTypes.insert(Ty); 1589 bool Removed = OpaqueStructTypes.erase(Ty); 1590 (void)Removed; 1591 assert(Removed); 1592 } 1593 1594 void IRMover::IdentifiedStructTypeSet::addOpaque(StructType *Ty) { 1595 assert(Ty->isOpaque()); 1596 OpaqueStructTypes.insert(Ty); 1597 } 1598 1599 StructType * 1600 IRMover::IdentifiedStructTypeSet::findNonOpaque(ArrayRef<Type *> ETypes, 1601 bool IsPacked) { 1602 IRMover::StructTypeKeyInfo::KeyTy Key(ETypes, IsPacked); 1603 auto I = NonOpaqueStructTypes.find_as(Key); 1604 return I == NonOpaqueStructTypes.end() ? nullptr : *I; 1605 } 1606 1607 bool IRMover::IdentifiedStructTypeSet::hasType(StructType *Ty) { 1608 if (Ty->isOpaque()) 1609 return OpaqueStructTypes.count(Ty); 1610 auto I = NonOpaqueStructTypes.find(Ty); 1611 return I == NonOpaqueStructTypes.end() ? false : *I == Ty; 1612 } 1613 1614 IRMover::IRMover(Module &M) : Composite(M) { 1615 TypeFinder StructTypes; 1616 StructTypes.run(M, /* OnlyNamed */ false); 1617 for (StructType *Ty : StructTypes) { 1618 if (Ty->isOpaque()) 1619 IdentifiedStructTypes.addOpaque(Ty); 1620 else 1621 IdentifiedStructTypes.addNonOpaque(Ty); 1622 } 1623 // Self-map metadatas in the destination module. This is needed when 1624 // DebugTypeODRUniquing is enabled on the LLVMContext, since metadata in the 1625 // destination module may be reached from the source module. 1626 for (auto *MD : StructTypes.getVisitedMetadata()) { 1627 SharedMDs[MD].reset(const_cast<MDNode *>(MD)); 1628 } 1629 } 1630 1631 Error IRMover::move( 1632 std::unique_ptr<Module> Src, ArrayRef<GlobalValue *> ValuesToLink, 1633 std::function<void(GlobalValue &, ValueAdder Add)> AddLazyFor, 1634 bool IsPerformingImport) { 1635 IRLinker TheIRLinker(Composite, SharedMDs, IdentifiedStructTypes, 1636 std::move(Src), ValuesToLink, std::move(AddLazyFor), 1637 IsPerformingImport); 1638 Error E = TheIRLinker.run(); 1639 Composite.dropTriviallyDeadConstantArrays(); 1640 return E; 1641 } 1642