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/Support/Error.h"
21#include "llvm/Transforms/Utils/Cloning.h"
22#include <utility>
23using namespace llvm;
24
25//===----------------------------------------------------------------------===//
26// TypeMap implementation.
27//===----------------------------------------------------------------------===//
28
29namespace {
30class TypeMapTy : public ValueMapTypeRemapper {
31  /// This is a mapping from a source type to a destination type to use.
32  DenseMap<Type *, Type *> MappedTypes;
33
34  /// When checking to see if two subgraphs are isomorphic, we speculatively
35  /// add types to MappedTypes, but keep track of them here in case we need to
36  /// roll back.
37  SmallVector<Type *, 16> SpeculativeTypes;
38
39  SmallVector<StructType *, 16> SpeculativeDstOpaqueTypes;
40
41  /// This is a list of non-opaque structs in the source module that are mapped
42  /// to an opaque struct in the destination module.
43  SmallVector<StructType *, 16> SrcDefinitionsToResolve;
44
45  /// This is the set of opaque types in the destination modules who are
46  /// getting a body from the source module.
47  SmallPtrSet<StructType *, 16> DstResolvedOpaqueTypes;
48
49public:
50  TypeMapTy(IRMover::IdentifiedStructTypeSet &DstStructTypesSet)
51      : DstStructTypesSet(DstStructTypesSet) {}
52
53  IRMover::IdentifiedStructTypeSet &DstStructTypesSet;
54  /// Indicate that the specified type in the destination module is conceptually
55  /// equivalent to the specified type in the source module.
56  void addTypeMapping(Type *DstTy, Type *SrcTy);
57
58  /// Produce a body for an opaque type in the dest module from a type
59  /// definition in the source module.
60  void linkDefinedTypeBodies();
61
62  /// Return the mapped type to use for the specified input type from the
63  /// source module.
64  Type *get(Type *SrcTy);
65  Type *get(Type *SrcTy, SmallPtrSet<StructType *, 8> &Visited);
66
67  void finishType(StructType *DTy, StructType *STy, ArrayRef<Type *> ETypes);
68
69  FunctionType *get(FunctionType *T) {
70    return cast<FunctionType>(get((Type *)T));
71  }
72
73private:
74  Type *remapType(Type *SrcTy) override { return get(SrcTy); }
75
76  bool areTypesIsomorphic(Type *DstTy, Type *SrcTy);
77};
78}
79
80void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) {
81  assert(SpeculativeTypes.empty());
82  assert(SpeculativeDstOpaqueTypes.empty());
83
84  // Check to see if these types are recursively isomorphic and establish a
85  // mapping between them if so.
86  if (!areTypesIsomorphic(DstTy, SrcTy)) {
87    // Oops, they aren't isomorphic.  Just discard this request by rolling out
88    // any speculative mappings we've established.
89    for (Type *Ty : SpeculativeTypes)
90      MappedTypes.erase(Ty);
91
92    SrcDefinitionsToResolve.resize(SrcDefinitionsToResolve.size() -
93                                   SpeculativeDstOpaqueTypes.size());
94    for (StructType *Ty : SpeculativeDstOpaqueTypes)
95      DstResolvedOpaqueTypes.erase(Ty);
96  } else {
97    // SrcTy and DstTy are recursively ismorphic. We clear names of SrcTy
98    // and all its descendants to lower amount of renaming in LLVM context
99    // Renaming occurs because we load all source modules to the same context
100    // and declaration with existing name gets renamed (i.e Foo -> Foo.42).
101    // As a result we may get several different types in the destination
102    // module, which are in fact the same.
103    for (Type *Ty : SpeculativeTypes)
104      if (auto *STy = dyn_cast<StructType>(Ty))
105        if (STy->hasName())
106          STy->setName("");
107  }
108  SpeculativeTypes.clear();
109  SpeculativeDstOpaqueTypes.clear();
110}
111
112/// Recursively walk this pair of types, returning true if they are isomorphic,
113/// false if they are not.
114bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) {
115  // Two types with differing kinds are clearly not isomorphic.
116  if (DstTy->getTypeID() != SrcTy->getTypeID())
117    return false;
118
119  // If we have an entry in the MappedTypes table, then we have our answer.
120  Type *&Entry = MappedTypes[SrcTy];
121  if (Entry)
122    return Entry == DstTy;
123
124  // Two identical types are clearly isomorphic.  Remember this
125  // non-speculatively.
126  if (DstTy == SrcTy) {
127    Entry = DstTy;
128    return true;
129  }
130
131  // Okay, we have two types with identical kinds that we haven't seen before.
132
133  // If this is an opaque struct type, special case it.
134  if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) {
135    // Mapping an opaque type to any struct, just keep the dest struct.
136    if (SSTy->isOpaque()) {
137      Entry = DstTy;
138      SpeculativeTypes.push_back(SrcTy);
139      return true;
140    }
141
142    // Mapping a non-opaque source type to an opaque dest.  If this is the first
143    // type that we're mapping onto this destination type then we succeed.  Keep
144    // the dest, but fill it in later. If this is the second (different) type
145    // that we're trying to map onto the same opaque type then we fail.
146    if (cast<StructType>(DstTy)->isOpaque()) {
147      // We can only map one source type onto the opaque destination type.
148      if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)).second)
149        return false;
150      SrcDefinitionsToResolve.push_back(SSTy);
151      SpeculativeTypes.push_back(SrcTy);
152      SpeculativeDstOpaqueTypes.push_back(cast<StructType>(DstTy));
153      Entry = DstTy;
154      return true;
155    }
156  }
157
158  // If the number of subtypes disagree between the two types, then we fail.
159  if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes())
160    return false;
161
162  // Fail if any of the extra properties (e.g. array size) of the type disagree.
163  if (isa<IntegerType>(DstTy))
164    return false; // bitwidth disagrees.
165  if (PointerType *PT = dyn_cast<PointerType>(DstTy)) {
166    if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace())
167      return false;
168  } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) {
169    if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg())
170      return false;
171  } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) {
172    StructType *SSTy = cast<StructType>(SrcTy);
173    if (DSTy->isLiteral() != SSTy->isLiteral() ||
174        DSTy->isPacked() != SSTy->isPacked())
175      return false;
176  } else if (auto *DArrTy = dyn_cast<ArrayType>(DstTy)) {
177    if (DArrTy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
178      return false;
179  } else if (auto *DVecTy = dyn_cast<VectorType>(DstTy)) {
180    if (DVecTy->getElementCount() != cast<VectorType>(SrcTy)->getElementCount())
181      return false;
182  }
183
184  // Otherwise, we speculate that these two types will line up and recursively
185  // check the subelements.
186  Entry = DstTy;
187  SpeculativeTypes.push_back(SrcTy);
188
189  for (unsigned I = 0, E = SrcTy->getNumContainedTypes(); I != E; ++I)
190    if (!areTypesIsomorphic(DstTy->getContainedType(I),
191                            SrcTy->getContainedType(I)))
192      return false;
193
194  // If everything seems to have lined up, then everything is great.
195  return true;
196}
197
198void TypeMapTy::linkDefinedTypeBodies() {
199  SmallVector<Type *, 16> Elements;
200  for (StructType *SrcSTy : SrcDefinitionsToResolve) {
201    StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]);
202    assert(DstSTy->isOpaque());
203
204    // Map the body of the source type over to a new body for the dest type.
205    Elements.resize(SrcSTy->getNumElements());
206    for (unsigned I = 0, E = Elements.size(); I != E; ++I)
207      Elements[I] = get(SrcSTy->getElementType(I));
208
209    DstSTy->setBody(Elements, SrcSTy->isPacked());
210    DstStructTypesSet.switchToNonOpaque(DstSTy);
211  }
212  SrcDefinitionsToResolve.clear();
213  DstResolvedOpaqueTypes.clear();
214}
215
216void TypeMapTy::finishType(StructType *DTy, StructType *STy,
217                           ArrayRef<Type *> ETypes) {
218  DTy->setBody(ETypes, STy->isPacked());
219
220  // Steal STy's name.
221  if (STy->hasName()) {
222    SmallString<16> TmpName = STy->getName();
223    STy->setName("");
224    DTy->setName(TmpName);
225  }
226
227  DstStructTypesSet.addNonOpaque(DTy);
228}
229
230Type *TypeMapTy::get(Type *Ty) {
231  SmallPtrSet<StructType *, 8> Visited;
232  return get(Ty, Visited);
233}
234
235Type *TypeMapTy::get(Type *Ty, SmallPtrSet<StructType *, 8> &Visited) {
236  // If we already have an entry for this type, return it.
237  Type **Entry = &MappedTypes[Ty];
238  if (*Entry)
239    return *Entry;
240
241  // These are types that LLVM itself will unique.
242  bool IsUniqued = !isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral();
243
244  if (!IsUniqued) {
245    StructType *STy = cast<StructType>(Ty);
246    // This is actually a type from the destination module, this can be reached
247    // when this type is loaded in another module, added to DstStructTypesSet,
248    // and then we reach the same type in another module where it has not been
249    // added to MappedTypes. (PR37684)
250    if (STy->getContext().isODRUniquingDebugTypes() && !STy->isOpaque() &&
251        DstStructTypesSet.hasType(STy))
252      return *Entry = STy;
253
254#ifndef NDEBUG
255    for (auto &Pair : MappedTypes) {
256      assert(!(Pair.first != Ty && Pair.second == Ty) &&
257             "mapping to a source type");
258    }
259#endif
260
261    if (!Visited.insert(STy).second) {
262      StructType *DTy = StructType::create(Ty->getContext());
263      return *Entry = DTy;
264    }
265  }
266
267  // If this is not a recursive type, then just map all of the elements and
268  // then rebuild the type from inside out.
269  SmallVector<Type *, 4> ElementTypes;
270
271  // If there are no element types to map, then the type is itself.  This is
272  // true for the anonymous {} struct, things like 'float', integers, etc.
273  if (Ty->getNumContainedTypes() == 0 && IsUniqued)
274    return *Entry = Ty;
275
276  // Remap all of the elements, keeping track of whether any of them change.
277  bool AnyChange = false;
278  ElementTypes.resize(Ty->getNumContainedTypes());
279  for (unsigned I = 0, E = Ty->getNumContainedTypes(); I != E; ++I) {
280    ElementTypes[I] = get(Ty->getContainedType(I), Visited);
281    AnyChange |= ElementTypes[I] != Ty->getContainedType(I);
282  }
283
284  // If we found our type while recursively processing stuff, just use it.
285  Entry = &MappedTypes[Ty];
286  if (*Entry) {
287    if (auto *DTy = dyn_cast<StructType>(*Entry)) {
288      if (DTy->isOpaque()) {
289        auto *STy = cast<StructType>(Ty);
290        finishType(DTy, STy, ElementTypes);
291      }
292    }
293    return *Entry;
294  }
295
296  // If all of the element types mapped directly over and the type is not
297  // a named struct, then the type is usable as-is.
298  if (!AnyChange && IsUniqued)
299    return *Entry = Ty;
300
301  // Otherwise, rebuild a modified type.
302  switch (Ty->getTypeID()) {
303  default:
304    llvm_unreachable("unknown derived type to remap");
305  case Type::ArrayTyID:
306    return *Entry = ArrayType::get(ElementTypes[0],
307                                   cast<ArrayType>(Ty)->getNumElements());
308  case Type::ScalableVectorTyID:
309    // FIXME: handle scalable vectors
310  case Type::FixedVectorTyID:
311    return *Entry = FixedVectorType::get(
312               ElementTypes[0], cast<FixedVectorType>(Ty)->getNumElements());
313  case Type::PointerTyID:
314    return *Entry = PointerType::get(ElementTypes[0],
315                                     cast<PointerType>(Ty)->getAddressSpace());
316  case Type::FunctionTyID:
317    return *Entry = FunctionType::get(ElementTypes[0],
318                                      makeArrayRef(ElementTypes).slice(1),
319                                      cast<FunctionType>(Ty)->isVarArg());
320  case Type::StructTyID: {
321    auto *STy = cast<StructType>(Ty);
322    bool IsPacked = STy->isPacked();
323    if (IsUniqued)
324      return *Entry = StructType::get(Ty->getContext(), ElementTypes, IsPacked);
325
326    // If the type is opaque, we can just use it directly.
327    if (STy->isOpaque()) {
328      DstStructTypesSet.addOpaque(STy);
329      return *Entry = Ty;
330    }
331
332    if (StructType *OldT =
333            DstStructTypesSet.findNonOpaque(ElementTypes, IsPacked)) {
334      STy->setName("");
335      return *Entry = OldT;
336    }
337
338    if (!AnyChange) {
339      DstStructTypesSet.addNonOpaque(STy);
340      return *Entry = Ty;
341    }
342
343    StructType *DTy = StructType::create(Ty->getContext());
344    finishType(DTy, STy, ElementTypes);
345    return *Entry = DTy;
346  }
347  }
348}
349
350LinkDiagnosticInfo::LinkDiagnosticInfo(DiagnosticSeverity Severity,
351                                       const Twine &Msg)
352    : DiagnosticInfo(DK_Linker, Severity), Msg(Msg) {}
353void LinkDiagnosticInfo::print(DiagnosticPrinter &DP) const { DP << Msg; }
354
355//===----------------------------------------------------------------------===//
356// IRLinker implementation.
357//===----------------------------------------------------------------------===//
358
359namespace {
360class IRLinker;
361
362/// Creates prototypes for functions that are lazily linked on the fly. This
363/// speeds up linking for modules with many/ lazily linked functions of which
364/// few get used.
365class GlobalValueMaterializer final : public ValueMaterializer {
366  IRLinker &TheIRLinker;
367
368public:
369  GlobalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {}
370  Value *materialize(Value *V) override;
371};
372
373class LocalValueMaterializer final : public ValueMaterializer {
374  IRLinker &TheIRLinker;
375
376public:
377  LocalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {}
378  Value *materialize(Value *V) override;
379};
380
381/// Type of the Metadata map in \a ValueToValueMapTy.
382typedef DenseMap<const Metadata *, TrackingMDRef> MDMapT;
383
384/// This is responsible for keeping track of the state used for moving data
385/// from SrcM to DstM.
386class IRLinker {
387  Module &DstM;
388  std::unique_ptr<Module> SrcM;
389
390  /// See IRMover::move().
391  std::function<void(GlobalValue &, IRMover::ValueAdder)> AddLazyFor;
392
393  TypeMapTy TypeMap;
394  GlobalValueMaterializer GValMaterializer;
395  LocalValueMaterializer LValMaterializer;
396
397  /// A metadata map that's shared between IRLinker instances.
398  MDMapT &SharedMDs;
399
400  /// Mapping of values from what they used to be in Src, to what they are now
401  /// in DstM.  ValueToValueMapTy is a ValueMap, which involves some overhead
402  /// due to the use of Value handles which the Linker doesn't actually need,
403  /// but this allows us to reuse the ValueMapper code.
404  ValueToValueMapTy ValueMap;
405  ValueToValueMapTy IndirectSymbolValueMap;
406
407  DenseSet<GlobalValue *> ValuesToLink;
408  std::vector<GlobalValue *> Worklist;
409  std::vector<std::pair<GlobalValue *, Value*>> RAUWWorklist;
410
411  void maybeAdd(GlobalValue *GV) {
412    if (ValuesToLink.insert(GV).second)
413      Worklist.push_back(GV);
414  }
415
416  /// Whether we are importing globals for ThinLTO, as opposed to linking the
417  /// source module. If this flag is set, it means that we can rely on some
418  /// other object file to define any non-GlobalValue entities defined by the
419  /// source module. This currently causes us to not link retained types in
420  /// debug info metadata and module inline asm.
421  bool IsPerformingImport;
422
423  /// Set to true when all global value body linking is complete (including
424  /// lazy linking). Used to prevent metadata linking from creating new
425  /// references.
426  bool DoneLinkingBodies = false;
427
428  /// The Error encountered during materialization. We use an Optional here to
429  /// avoid needing to manage an unconsumed success value.
430  Optional<Error> FoundError;
431  void setError(Error E) {
432    if (E)
433      FoundError = std::move(E);
434  }
435
436  /// Most of the errors produced by this module are inconvertible StringErrors.
437  /// This convenience function lets us return one of those more easily.
438  Error stringErr(const Twine &T) {
439    return make_error<StringError>(T, inconvertibleErrorCode());
440  }
441
442  /// Entry point for mapping values and alternate context for mapping aliases.
443  ValueMapper Mapper;
444  unsigned IndirectSymbolMCID;
445
446  /// Handles cloning of a global values from the source module into
447  /// the destination module, including setting the attributes and visibility.
448  GlobalValue *copyGlobalValueProto(const GlobalValue *SGV, bool ForDefinition);
449
450  void emitWarning(const Twine &Message) {
451    SrcM->getContext().diagnose(LinkDiagnosticInfo(DS_Warning, Message));
452  }
453
454  /// Given a global in the source module, return the global in the
455  /// destination module that is being linked to, if any.
456  GlobalValue *getLinkedToGlobal(const GlobalValue *SrcGV) {
457    // If the source has no name it can't link.  If it has local linkage,
458    // there is no name match-up going on.
459    if (!SrcGV->hasName() || SrcGV->hasLocalLinkage())
460      return nullptr;
461
462    // Otherwise see if we have a match in the destination module's symtab.
463    GlobalValue *DGV = DstM.getNamedValue(SrcGV->getName());
464    if (!DGV)
465      return nullptr;
466
467    // If we found a global with the same name in the dest module, but it has
468    // internal linkage, we are really not doing any linkage here.
469    if (DGV->hasLocalLinkage())
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
522public:
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_MoveDistinctMDs | RF_IgnoreMissingLocals, &TypeMap,
532               &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.
551static 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
569Value *GlobalValueMaterializer::materialize(Value *SGV) {
570  return TheIRLinker.materialize(SGV, false);
571}
572
573Value *LocalValueMaterializer::materialize(Value *SGV) {
574  return TheIRLinker.materialize(SGV, true);
575}
576
577Value *IRLinker::materialize(Value *V, bool ForIndirectSymbol) {
578  auto *SGV = dyn_cast<GlobalValue>(V);
579  if (!SGV)
580    return nullptr;
581
582  Expected<Constant *> NewProto = linkGlobalValueProto(SGV, ForIndirectSymbol);
583  if (!NewProto) {
584    setError(NewProto.takeError());
585    return nullptr;
586  }
587  if (!*NewProto)
588    return nullptr;
589
590  GlobalValue *New = dyn_cast<GlobalValue>(*NewProto);
591  if (!New)
592    return *NewProto;
593
594  // If we already created the body, just return.
595  if (auto *F = dyn_cast<Function>(New)) {
596    if (!F->isDeclaration())
597      return New;
598  } else if (auto *V = dyn_cast<GlobalVariable>(New)) {
599    if (V->hasInitializer() || V->hasAppendingLinkage())
600      return New;
601  } else {
602    auto *IS = cast<GlobalIndirectSymbol>(New);
603    if (IS->getIndirectSymbol())
604      return New;
605  }
606
607  // When linking a global for an indirect symbol, it will always be linked.
608  // However we need to check if it was not already scheduled to satisfy a
609  // reference from a regular global value initializer. We know if it has been
610  // schedule if the "New" GlobalValue that is mapped here for the indirect
611  // symbol is the same as the one already mapped. If there is an entry in the
612  // ValueMap but the value is different, it means that the value already had a
613  // definition in the destination module (linkonce for instance), but we need a
614  // new definition for the indirect symbol ("New" will be different.
615  if (ForIndirectSymbol && ValueMap.lookup(SGV) == New)
616    return New;
617
618  if (ForIndirectSymbol || shouldLink(New, *SGV))
619    setError(linkGlobalValueBody(*New, *SGV));
620
621  return New;
622}
623
624/// Loop through the global variables in the src module and merge them into the
625/// dest module.
626GlobalVariable *IRLinker::copyGlobalVariableProto(const GlobalVariable *SGVar) {
627  // No linking to be performed or linking from the source: simply create an
628  // identical version of the symbol over in the dest module... the
629  // initializer will be filled in later by LinkGlobalInits.
630  GlobalVariable *NewDGV =
631      new GlobalVariable(DstM, TypeMap.get(SGVar->getValueType()),
632                         SGVar->isConstant(), GlobalValue::ExternalLinkage,
633                         /*init*/ nullptr, SGVar->getName(),
634                         /*insertbefore*/ nullptr, SGVar->getThreadLocalMode(),
635                         SGVar->getAddressSpace());
636  NewDGV->setAlignment(MaybeAlign(SGVar->getAlignment()));
637  NewDGV->copyAttributesFrom(SGVar);
638  return NewDGV;
639}
640
641AttributeList IRLinker::mapAttributeTypes(LLVMContext &C, AttributeList Attrs) {
642  for (unsigned i = 0; i < Attrs.getNumAttrSets(); ++i) {
643    if (Attrs.hasAttribute(i, Attribute::ByVal)) {
644      Type *Ty = Attrs.getAttribute(i, Attribute::ByVal).getValueAsType();
645      if (!Ty)
646        continue;
647
648      Attrs = Attrs.removeAttribute(C, i, Attribute::ByVal);
649      Attrs = Attrs.addAttribute(
650          C, i, Attribute::getWithByValType(C, TypeMap.get(Ty)));
651    }
652  }
653  return Attrs;
654}
655
656/// Link the function in the source module into the destination module if
657/// needed, setting up mapping information.
658Function *IRLinker::copyFunctionProto(const Function *SF) {
659  // If there is no linkage to be performed or we are linking from the source,
660  // bring SF over.
661  auto *F = Function::Create(TypeMap.get(SF->getFunctionType()),
662                             GlobalValue::ExternalLinkage,
663                             SF->getAddressSpace(), SF->getName(), &DstM);
664  F->copyAttributesFrom(SF);
665  F->setAttributes(mapAttributeTypes(F->getContext(), F->getAttributes()));
666  return F;
667}
668
669/// Set up prototypes for any indirect symbols that come over from the source
670/// module.
671GlobalValue *
672IRLinker::copyGlobalIndirectSymbolProto(const GlobalIndirectSymbol *SGIS) {
673  // If there is no linkage to be performed or we're linking from the source,
674  // bring over SGA.
675  auto *Ty = TypeMap.get(SGIS->getValueType());
676  GlobalIndirectSymbol *GIS;
677  if (isa<GlobalAlias>(SGIS))
678    GIS = GlobalAlias::create(Ty, SGIS->getAddressSpace(),
679                              GlobalValue::ExternalLinkage, SGIS->getName(),
680                              &DstM);
681  else
682    GIS = GlobalIFunc::create(Ty, SGIS->getAddressSpace(),
683                              GlobalValue::ExternalLinkage, SGIS->getName(),
684                              nullptr, &DstM);
685  GIS->copyAttributesFrom(SGIS);
686  return GIS;
687}
688
689GlobalValue *IRLinker::copyGlobalValueProto(const GlobalValue *SGV,
690                                            bool ForDefinition) {
691  GlobalValue *NewGV;
692  if (auto *SGVar = dyn_cast<GlobalVariable>(SGV)) {
693    NewGV = copyGlobalVariableProto(SGVar);
694  } else if (auto *SF = dyn_cast<Function>(SGV)) {
695    NewGV = copyFunctionProto(SF);
696  } else {
697    if (ForDefinition)
698      NewGV = copyGlobalIndirectSymbolProto(cast<GlobalIndirectSymbol>(SGV));
699    else if (SGV->getValueType()->isFunctionTy())
700      NewGV =
701          Function::Create(cast<FunctionType>(TypeMap.get(SGV->getValueType())),
702                           GlobalValue::ExternalLinkage, SGV->getAddressSpace(),
703                           SGV->getName(), &DstM);
704    else
705      NewGV =
706          new GlobalVariable(DstM, TypeMap.get(SGV->getValueType()),
707                             /*isConstant*/ false, GlobalValue::ExternalLinkage,
708                             /*init*/ nullptr, SGV->getName(),
709                             /*insertbefore*/ nullptr,
710                             SGV->getThreadLocalMode(), SGV->getAddressSpace());
711  }
712
713  if (ForDefinition)
714    NewGV->setLinkage(SGV->getLinkage());
715  else if (SGV->hasExternalWeakLinkage())
716    NewGV->setLinkage(GlobalValue::ExternalWeakLinkage);
717
718  if (auto *NewGO = dyn_cast<GlobalObject>(NewGV)) {
719    // Metadata for global variables and function declarations is copied eagerly.
720    if (isa<GlobalVariable>(SGV) || SGV->isDeclaration())
721      NewGO->copyMetadata(cast<GlobalObject>(SGV), 0);
722  }
723
724  // Remove these copied constants in case this stays a declaration, since
725  // they point to the source module. If the def is linked the values will
726  // be mapped in during linkFunctionBody.
727  if (auto *NewF = dyn_cast<Function>(NewGV)) {
728    NewF->setPersonalityFn(nullptr);
729    NewF->setPrefixData(nullptr);
730    NewF->setPrologueData(nullptr);
731  }
732
733  return NewGV;
734}
735
736static StringRef getTypeNamePrefix(StringRef Name) {
737  size_t DotPos = Name.rfind('.');
738  return (DotPos == 0 || DotPos == StringRef::npos || Name.back() == '.' ||
739          !isdigit(static_cast<unsigned char>(Name[DotPos + 1])))
740             ? Name
741             : Name.substr(0, DotPos);
742}
743
744/// Loop over all of the linked values to compute type mappings.  For example,
745/// if we link "extern Foo *x" and "Foo *x = NULL", then we have two struct
746/// types 'Foo' but one got renamed when the module was loaded into the same
747/// LLVMContext.
748void IRLinker::computeTypeMapping() {
749  for (GlobalValue &SGV : SrcM->globals()) {
750    GlobalValue *DGV = getLinkedToGlobal(&SGV);
751    if (!DGV)
752      continue;
753
754    if (!DGV->hasAppendingLinkage() || !SGV.hasAppendingLinkage()) {
755      TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
756      continue;
757    }
758
759    // Unify the element type of appending arrays.
760    ArrayType *DAT = cast<ArrayType>(DGV->getValueType());
761    ArrayType *SAT = cast<ArrayType>(SGV.getValueType());
762    TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
763  }
764
765  for (GlobalValue &SGV : *SrcM)
766    if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) {
767      if (DGV->getType() == SGV.getType()) {
768        // If the types of DGV and SGV are the same, it means that DGV is from
769        // the source module and got added to DstM from a shared metadata.  We
770        // shouldn't map this type to itself in case the type's components get
771        // remapped to a new type from DstM (for instance, during the loop over
772        // SrcM->getIdentifiedStructTypes() below).
773        continue;
774      }
775
776      TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
777    }
778
779  for (GlobalValue &SGV : SrcM->aliases())
780    if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
781      TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
782
783  // Incorporate types by name, scanning all the types in the source module.
784  // At this point, the destination module may have a type "%foo = { i32 }" for
785  // example.  When the source module got loaded into the same LLVMContext, if
786  // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
787  std::vector<StructType *> Types = SrcM->getIdentifiedStructTypes();
788  for (StructType *ST : Types) {
789    if (!ST->hasName())
790      continue;
791
792    if (TypeMap.DstStructTypesSet.hasType(ST)) {
793      // This is actually a type from the destination module.
794      // getIdentifiedStructTypes() can have found it by walking debug info
795      // metadata nodes, some of which get linked by name when ODR Type Uniquing
796      // is enabled on the Context, from the source to the destination module.
797      continue;
798    }
799
800    auto STTypePrefix = getTypeNamePrefix(ST->getName());
801    if (STTypePrefix.size()== ST->getName().size())
802      continue;
803
804    // Check to see if the destination module has a struct with the prefix name.
805    StructType *DST = DstM.getTypeByName(STTypePrefix);
806    if (!DST)
807      continue;
808
809    // Don't use it if this actually came from the source module. They're in
810    // the same LLVMContext after all. Also don't use it unless the type is
811    // actually used in the destination module. This can happen in situations
812    // like this:
813    //
814    //      Module A                         Module B
815    //      --------                         --------
816    //   %Z = type { %A }                %B = type { %C.1 }
817    //   %A = type { %B.1, [7 x i8] }    %C.1 = type { i8* }
818    //   %B.1 = type { %C }              %A.2 = type { %B.3, [5 x i8] }
819    //   %C = type { i8* }               %B.3 = type { %C.1 }
820    //
821    // When we link Module B with Module A, the '%B' in Module B is
822    // used. However, that would then use '%C.1'. But when we process '%C.1',
823    // we prefer to take the '%C' version. So we are then left with both
824    // '%C.1' and '%C' being used for the same types. This leads to some
825    // variables using one type and some using the other.
826    if (TypeMap.DstStructTypesSet.hasType(DST))
827      TypeMap.addTypeMapping(DST, ST);
828  }
829
830  // Now that we have discovered all of the type equivalences, get a body for
831  // any 'opaque' types in the dest module that are now resolved.
832  TypeMap.linkDefinedTypeBodies();
833}
834
835static void getArrayElements(const Constant *C,
836                             SmallVectorImpl<Constant *> &Dest) {
837  unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements();
838
839  for (unsigned i = 0; i != NumElements; ++i)
840    Dest.push_back(C->getAggregateElement(i));
841}
842
843/// If there were any appending global variables, link them together now.
844Expected<Constant *>
845IRLinker::linkAppendingVarProto(GlobalVariable *DstGV,
846                                const GlobalVariable *SrcGV) {
847  Type *EltTy = cast<ArrayType>(TypeMap.get(SrcGV->getValueType()))
848                    ->getElementType();
849
850  // FIXME: This upgrade is done during linking to support the C API.  Once the
851  // old form is deprecated, we should move this upgrade to
852  // llvm::UpgradeGlobalVariable() and simplify the logic here and in
853  // Mapper::mapAppendingVariable() in ValueMapper.cpp.
854  StringRef Name = SrcGV->getName();
855  bool IsNewStructor = false;
856  bool IsOldStructor = false;
857  if (Name == "llvm.global_ctors" || Name == "llvm.global_dtors") {
858    if (cast<StructType>(EltTy)->getNumElements() == 3)
859      IsNewStructor = true;
860    else
861      IsOldStructor = true;
862  }
863
864  PointerType *VoidPtrTy = Type::getInt8Ty(SrcGV->getContext())->getPointerTo();
865  if (IsOldStructor) {
866    auto &ST = *cast<StructType>(EltTy);
867    Type *Tys[3] = {ST.getElementType(0), ST.getElementType(1), VoidPtrTy};
868    EltTy = StructType::get(SrcGV->getContext(), Tys, false);
869  }
870
871  uint64_t DstNumElements = 0;
872  if (DstGV) {
873    ArrayType *DstTy = cast<ArrayType>(DstGV->getValueType());
874    DstNumElements = DstTy->getNumElements();
875
876    if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage())
877      return stringErr(
878          "Linking globals named '" + SrcGV->getName() +
879          "': can only link appending global with another appending "
880          "global!");
881
882    // Check to see that they two arrays agree on type.
883    if (EltTy != DstTy->getElementType())
884      return stringErr("Appending variables with different element types!");
885    if (DstGV->isConstant() != SrcGV->isConstant())
886      return stringErr("Appending variables linked with different const'ness!");
887
888    if (DstGV->getAlignment() != SrcGV->getAlignment())
889      return stringErr(
890          "Appending variables with different alignment need to be linked!");
891
892    if (DstGV->getVisibility() != SrcGV->getVisibility())
893      return stringErr(
894          "Appending variables with different visibility need to be linked!");
895
896    if (DstGV->hasGlobalUnnamedAddr() != SrcGV->hasGlobalUnnamedAddr())
897      return stringErr(
898          "Appending variables with different unnamed_addr need to be linked!");
899
900    if (DstGV->getSection() != SrcGV->getSection())
901      return stringErr(
902          "Appending variables with different section name need to be linked!");
903  }
904
905  SmallVector<Constant *, 16> SrcElements;
906  getArrayElements(SrcGV->getInitializer(), SrcElements);
907
908  if (IsNewStructor) {
909    auto It = remove_if(SrcElements, [this](Constant *E) {
910      auto *Key =
911          dyn_cast<GlobalValue>(E->getAggregateElement(2)->stripPointerCasts());
912      if (!Key)
913        return false;
914      GlobalValue *DGV = getLinkedToGlobal(Key);
915      return !shouldLink(DGV, *Key);
916    });
917    SrcElements.erase(It, SrcElements.end());
918  }
919  uint64_t NewSize = DstNumElements + SrcElements.size();
920  ArrayType *NewType = ArrayType::get(EltTy, NewSize);
921
922  // Create the new global variable.
923  GlobalVariable *NG = new GlobalVariable(
924      DstM, NewType, SrcGV->isConstant(), SrcGV->getLinkage(),
925      /*init*/ nullptr, /*name*/ "", DstGV, SrcGV->getThreadLocalMode(),
926      SrcGV->getAddressSpace());
927
928  NG->copyAttributesFrom(SrcGV);
929  forceRenaming(NG, SrcGV->getName());
930
931  Constant *Ret = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType()));
932
933  Mapper.scheduleMapAppendingVariable(*NG,
934                                      DstGV ? DstGV->getInitializer() : nullptr,
935                                      IsOldStructor, SrcElements);
936
937  // Replace any uses of the two global variables with uses of the new
938  // global.
939  if (DstGV) {
940    RAUWWorklist.push_back(
941        std::make_pair(DstGV, ConstantExpr::getBitCast(NG, DstGV->getType())));
942  }
943
944  return Ret;
945}
946
947bool IRLinker::shouldLink(GlobalValue *DGV, GlobalValue &SGV) {
948  if (ValuesToLink.count(&SGV) || SGV.hasLocalLinkage())
949    return true;
950
951  if (DGV && !DGV->isDeclarationForLinker())
952    return false;
953
954  if (SGV.isDeclaration() || DoneLinkingBodies)
955    return false;
956
957  // Callback to the client to give a chance to lazily add the Global to the
958  // list of value to link.
959  bool LazilyAdded = false;
960  AddLazyFor(SGV, [this, &LazilyAdded](GlobalValue &GV) {
961    maybeAdd(&GV);
962    LazilyAdded = true;
963  });
964  return LazilyAdded;
965}
966
967Expected<Constant *> IRLinker::linkGlobalValueProto(GlobalValue *SGV,
968                                                    bool ForIndirectSymbol) {
969  GlobalValue *DGV = getLinkedToGlobal(SGV);
970
971  bool ShouldLink = shouldLink(DGV, *SGV);
972
973  // just missing from map
974  if (ShouldLink) {
975    auto I = ValueMap.find(SGV);
976    if (I != ValueMap.end())
977      return cast<Constant>(I->second);
978
979    I = IndirectSymbolValueMap.find(SGV);
980    if (I != IndirectSymbolValueMap.end())
981      return cast<Constant>(I->second);
982  }
983
984  if (!ShouldLink && ForIndirectSymbol)
985    DGV = nullptr;
986
987  // Handle the ultra special appending linkage case first.
988  assert(!DGV || SGV->hasAppendingLinkage() == DGV->hasAppendingLinkage());
989  if (SGV->hasAppendingLinkage())
990    return linkAppendingVarProto(cast_or_null<GlobalVariable>(DGV),
991                                 cast<GlobalVariable>(SGV));
992
993  GlobalValue *NewGV;
994  if (DGV && !ShouldLink) {
995    NewGV = DGV;
996  } else {
997    // If we are done linking global value bodies (i.e. we are performing
998    // metadata linking), don't link in the global value due to this
999    // reference, simply map it to null.
1000    if (DoneLinkingBodies)
1001      return nullptr;
1002
1003    NewGV = copyGlobalValueProto(SGV, ShouldLink || ForIndirectSymbol);
1004    if (ShouldLink || !ForIndirectSymbol)
1005      forceRenaming(NewGV, SGV->getName());
1006  }
1007
1008  // Overloaded intrinsics have overloaded types names as part of their
1009  // names. If we renamed overloaded types we should rename the intrinsic
1010  // as well.
1011  if (Function *F = dyn_cast<Function>(NewGV))
1012    if (auto Remangled = Intrinsic::remangleIntrinsicFunction(F))
1013      NewGV = Remangled.getValue();
1014
1015  if (ShouldLink || ForIndirectSymbol) {
1016    if (const Comdat *SC = SGV->getComdat()) {
1017      if (auto *GO = dyn_cast<GlobalObject>(NewGV)) {
1018        Comdat *DC = DstM.getOrInsertComdat(SC->getName());
1019        DC->setSelectionKind(SC->getSelectionKind());
1020        GO->setComdat(DC);
1021      }
1022    }
1023  }
1024
1025  if (!ShouldLink && ForIndirectSymbol)
1026    NewGV->setLinkage(GlobalValue::InternalLinkage);
1027
1028  Constant *C = NewGV;
1029  // Only create a bitcast if necessary. In particular, with
1030  // DebugTypeODRUniquing we may reach metadata in the destination module
1031  // containing a GV from the source module, in which case SGV will be
1032  // the same as DGV and NewGV, and TypeMap.get() will assert since it
1033  // assumes it is being invoked on a type in the source module.
1034  if (DGV && NewGV != SGV) {
1035    C = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
1036      NewGV, TypeMap.get(SGV->getType()));
1037  }
1038
1039  if (DGV && NewGV != DGV) {
1040    // Schedule "replace all uses with" to happen after materializing is
1041    // done. It is not safe to do it now, since ValueMapper may be holding
1042    // pointers to constants that will get deleted if RAUW runs.
1043    RAUWWorklist.push_back(std::make_pair(
1044        DGV,
1045        ConstantExpr::getPointerBitCastOrAddrSpaceCast(NewGV, DGV->getType())));
1046  }
1047
1048  return C;
1049}
1050
1051/// Update the initializers in the Dest module now that all globals that may be
1052/// referenced are in Dest.
1053void IRLinker::linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src) {
1054  // Figure out what the initializer looks like in the dest module.
1055  Mapper.scheduleMapGlobalInitializer(Dst, *Src.getInitializer());
1056}
1057
1058/// Copy the source function over into the dest function and fix up references
1059/// to values. At this point we know that Dest is an external function, and
1060/// that Src is not.
1061Error IRLinker::linkFunctionBody(Function &Dst, Function &Src) {
1062  assert(Dst.isDeclaration() && !Src.isDeclaration());
1063
1064  // Materialize if needed.
1065  if (Error Err = Src.materialize())
1066    return Err;
1067
1068  // Link in the operands without remapping.
1069  if (Src.hasPrefixData())
1070    Dst.setPrefixData(Src.getPrefixData());
1071  if (Src.hasPrologueData())
1072    Dst.setPrologueData(Src.getPrologueData());
1073  if (Src.hasPersonalityFn())
1074    Dst.setPersonalityFn(Src.getPersonalityFn());
1075
1076  // Copy over the metadata attachments without remapping.
1077  Dst.copyMetadata(&Src, 0);
1078
1079  // Steal arguments and splice the body of Src into Dst.
1080  Dst.stealArgumentListFrom(Src);
1081  Dst.getBasicBlockList().splice(Dst.end(), Src.getBasicBlockList());
1082
1083  // Everything has been moved over.  Remap it.
1084  Mapper.scheduleRemapFunction(Dst);
1085  return Error::success();
1086}
1087
1088void IRLinker::linkIndirectSymbolBody(GlobalIndirectSymbol &Dst,
1089                                      GlobalIndirectSymbol &Src) {
1090  Mapper.scheduleMapGlobalIndirectSymbol(Dst, *Src.getIndirectSymbol(),
1091                                         IndirectSymbolMCID);
1092}
1093
1094Error IRLinker::linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src) {
1095  if (auto *F = dyn_cast<Function>(&Src))
1096    return linkFunctionBody(cast<Function>(Dst), *F);
1097  if (auto *GVar = dyn_cast<GlobalVariable>(&Src)) {
1098    linkGlobalVariable(cast<GlobalVariable>(Dst), *GVar);
1099    return Error::success();
1100  }
1101  linkIndirectSymbolBody(cast<GlobalIndirectSymbol>(Dst), cast<GlobalIndirectSymbol>(Src));
1102  return Error::success();
1103}
1104
1105void IRLinker::flushRAUWWorklist() {
1106  for (const auto &Elem : RAUWWorklist) {
1107    GlobalValue *Old;
1108    Value *New;
1109    std::tie(Old, New) = Elem;
1110
1111    Old->replaceAllUsesWith(New);
1112    Old->eraseFromParent();
1113  }
1114  RAUWWorklist.clear();
1115}
1116
1117void IRLinker::prepareCompileUnitsForImport() {
1118  NamedMDNode *SrcCompileUnits = SrcM->getNamedMetadata("llvm.dbg.cu");
1119  if (!SrcCompileUnits)
1120    return;
1121  // When importing for ThinLTO, prevent importing of types listed on
1122  // the DICompileUnit that we don't need a copy of in the importing
1123  // module. They will be emitted by the originating module.
1124  for (unsigned I = 0, E = SrcCompileUnits->getNumOperands(); I != E; ++I) {
1125    auto *CU = cast<DICompileUnit>(SrcCompileUnits->getOperand(I));
1126    assert(CU && "Expected valid compile unit");
1127    // Enums, macros, and retained types don't need to be listed on the
1128    // imported DICompileUnit. This means they will only be imported
1129    // if reached from the mapped IR. Do this by setting their value map
1130    // entries to nullptr, which will automatically prevent their importing
1131    // when reached from the DICompileUnit during metadata mapping.
1132    ValueMap.MD()[CU->getRawEnumTypes()].reset(nullptr);
1133    ValueMap.MD()[CU->getRawMacros()].reset(nullptr);
1134    ValueMap.MD()[CU->getRawRetainedTypes()].reset(nullptr);
1135    // The original definition (or at least its debug info - if the variable is
1136    // internalized an optimized away) will remain in the source module, so
1137    // there's no need to import them.
1138    // If LLVM ever does more advanced optimizations on global variables
1139    // (removing/localizing write operations, for instance) that can track
1140    // through debug info, this decision may need to be revisited - but do so
1141    // with care when it comes to debug info size. Emitting small CUs containing
1142    // only a few imported entities into every destination module may be very
1143    // size inefficient.
1144    ValueMap.MD()[CU->getRawGlobalVariables()].reset(nullptr);
1145
1146    // Imported entities only need to be mapped in if they have local
1147    // scope, as those might correspond to an imported entity inside a
1148    // function being imported (any locally scoped imported entities that
1149    // don't end up referenced by an imported function will not be emitted
1150    // into the object). Imported entities not in a local scope
1151    // (e.g. on the namespace) only need to be emitted by the originating
1152    // module. Create a list of the locally scoped imported entities, and
1153    // replace the source CUs imported entity list with the new list, so
1154    // only those are mapped in.
1155    // FIXME: Locally-scoped imported entities could be moved to the
1156    // functions they are local to instead of listing them on the CU, and
1157    // we would naturally only link in those needed by function importing.
1158    SmallVector<TrackingMDNodeRef, 4> AllImportedModules;
1159    bool ReplaceImportedEntities = false;
1160    for (auto *IE : CU->getImportedEntities()) {
1161      DIScope *Scope = IE->getScope();
1162      assert(Scope && "Invalid Scope encoding!");
1163      if (isa<DILocalScope>(Scope))
1164        AllImportedModules.emplace_back(IE);
1165      else
1166        ReplaceImportedEntities = true;
1167    }
1168    if (ReplaceImportedEntities) {
1169      if (!AllImportedModules.empty())
1170        CU->replaceImportedEntities(MDTuple::get(
1171            CU->getContext(),
1172            SmallVector<Metadata *, 16>(AllImportedModules.begin(),
1173                                        AllImportedModules.end())));
1174      else
1175        // If there were no local scope imported entities, we can map
1176        // the whole list to nullptr.
1177        ValueMap.MD()[CU->getRawImportedEntities()].reset(nullptr);
1178    }
1179  }
1180}
1181
1182/// Insert all of the named MDNodes in Src into the Dest module.
1183void IRLinker::linkNamedMDNodes() {
1184  const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1185  for (const NamedMDNode &NMD : SrcM->named_metadata()) {
1186    // Don't link module flags here. Do them separately.
1187    if (&NMD == SrcModFlags)
1188      continue;
1189    NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName());
1190    // Add Src elements into Dest node.
1191    for (const MDNode *Op : NMD.operands())
1192      DestNMD->addOperand(Mapper.mapMDNode(*Op));
1193  }
1194}
1195
1196/// Merge the linker flags in Src into the Dest module.
1197Error IRLinker::linkModuleFlagsMetadata() {
1198  // If the source module has no module flags, we are done.
1199  const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1200  if (!SrcModFlags)
1201    return Error::success();
1202
1203  // If the destination module doesn't have module flags yet, then just copy
1204  // over the source module's flags.
1205  NamedMDNode *DstModFlags = DstM.getOrInsertModuleFlagsMetadata();
1206  if (DstModFlags->getNumOperands() == 0) {
1207    for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I)
1208      DstModFlags->addOperand(SrcModFlags->getOperand(I));
1209
1210    return Error::success();
1211  }
1212
1213  // First build a map of the existing module flags and requirements.
1214  DenseMap<MDString *, std::pair<MDNode *, unsigned>> Flags;
1215  SmallSetVector<MDNode *, 16> Requirements;
1216  for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) {
1217    MDNode *Op = DstModFlags->getOperand(I);
1218    ConstantInt *Behavior = mdconst::extract<ConstantInt>(Op->getOperand(0));
1219    MDString *ID = cast<MDString>(Op->getOperand(1));
1220
1221    if (Behavior->getZExtValue() == Module::Require) {
1222      Requirements.insert(cast<MDNode>(Op->getOperand(2)));
1223    } else {
1224      Flags[ID] = std::make_pair(Op, I);
1225    }
1226  }
1227
1228  // Merge in the flags from the source module, and also collect its set of
1229  // requirements.
1230  for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) {
1231    MDNode *SrcOp = SrcModFlags->getOperand(I);
1232    ConstantInt *SrcBehavior =
1233        mdconst::extract<ConstantInt>(SrcOp->getOperand(0));
1234    MDString *ID = cast<MDString>(SrcOp->getOperand(1));
1235    MDNode *DstOp;
1236    unsigned DstIndex;
1237    std::tie(DstOp, DstIndex) = Flags.lookup(ID);
1238    unsigned SrcBehaviorValue = SrcBehavior->getZExtValue();
1239
1240    // If this is a requirement, add it and continue.
1241    if (SrcBehaviorValue == Module::Require) {
1242      // If the destination module does not already have this requirement, add
1243      // it.
1244      if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) {
1245        DstModFlags->addOperand(SrcOp);
1246      }
1247      continue;
1248    }
1249
1250    // If there is no existing flag with this ID, just add it.
1251    if (!DstOp) {
1252      Flags[ID] = std::make_pair(SrcOp, DstModFlags->getNumOperands());
1253      DstModFlags->addOperand(SrcOp);
1254      continue;
1255    }
1256
1257    // Otherwise, perform a merge.
1258    ConstantInt *DstBehavior =
1259        mdconst::extract<ConstantInt>(DstOp->getOperand(0));
1260    unsigned DstBehaviorValue = DstBehavior->getZExtValue();
1261
1262    auto overrideDstValue = [&]() {
1263      DstModFlags->setOperand(DstIndex, SrcOp);
1264      Flags[ID].first = SrcOp;
1265    };
1266
1267    // If either flag has override behavior, handle it first.
1268    if (DstBehaviorValue == Module::Override) {
1269      // Diagnose inconsistent flags which both have override behavior.
1270      if (SrcBehaviorValue == Module::Override &&
1271          SrcOp->getOperand(2) != DstOp->getOperand(2))
1272        return stringErr("linking module flags '" + ID->getString() +
1273                         "': IDs have conflicting override values in '" +
1274                         SrcM->getModuleIdentifier() + "' and '" +
1275                         DstM.getModuleIdentifier() + "'");
1276      continue;
1277    } else if (SrcBehaviorValue == Module::Override) {
1278      // Update the destination flag to that of the source.
1279      overrideDstValue();
1280      continue;
1281    }
1282
1283    // Diagnose inconsistent merge behavior types.
1284    if (SrcBehaviorValue != DstBehaviorValue) {
1285      bool MaxAndWarn = (SrcBehaviorValue == Module::Max &&
1286                         DstBehaviorValue == Module::Warning) ||
1287                        (DstBehaviorValue == Module::Max &&
1288                         SrcBehaviorValue == Module::Warning);
1289      if (!MaxAndWarn)
1290        return stringErr("linking module flags '" + ID->getString() +
1291                         "': IDs have conflicting behaviors in '" +
1292                         SrcM->getModuleIdentifier() + "' and '" +
1293                         DstM.getModuleIdentifier() + "'");
1294    }
1295
1296    auto replaceDstValue = [&](MDNode *New) {
1297      Metadata *FlagOps[] = {DstOp->getOperand(0), ID, New};
1298      MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps);
1299      DstModFlags->setOperand(DstIndex, Flag);
1300      Flags[ID].first = Flag;
1301    };
1302
1303    // Emit a warning if the values differ and either source or destination
1304    // request Warning behavior.
1305    if ((DstBehaviorValue == Module::Warning ||
1306         SrcBehaviorValue == Module::Warning) &&
1307        SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1308      std::string Str;
1309      raw_string_ostream(Str)
1310          << "linking module flags '" << ID->getString()
1311          << "': IDs have conflicting values ('" << *SrcOp->getOperand(2)
1312          << "' from " << SrcM->getModuleIdentifier() << " with '"
1313          << *DstOp->getOperand(2) << "' from " << DstM.getModuleIdentifier()
1314          << ')';
1315      emitWarning(Str);
1316    }
1317
1318    // Choose the maximum if either source or destination request Max behavior.
1319    if (DstBehaviorValue == Module::Max || SrcBehaviorValue == Module::Max) {
1320      ConstantInt *DstValue =
1321          mdconst::extract<ConstantInt>(DstOp->getOperand(2));
1322      ConstantInt *SrcValue =
1323          mdconst::extract<ConstantInt>(SrcOp->getOperand(2));
1324
1325      // The resulting flag should have a Max behavior, and contain the maximum
1326      // value from between the source and destination values.
1327      Metadata *FlagOps[] = {
1328          (DstBehaviorValue != Module::Max ? SrcOp : DstOp)->getOperand(0), ID,
1329          (SrcValue->getZExtValue() > DstValue->getZExtValue() ? SrcOp : DstOp)
1330              ->getOperand(2)};
1331      MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps);
1332      DstModFlags->setOperand(DstIndex, Flag);
1333      Flags[ID].first = Flag;
1334      continue;
1335    }
1336
1337    // Perform the merge for standard behavior types.
1338    switch (SrcBehaviorValue) {
1339    case Module::Require:
1340    case Module::Override:
1341      llvm_unreachable("not possible");
1342    case Module::Error: {
1343      // Emit an error if the values differ.
1344      if (SrcOp->getOperand(2) != DstOp->getOperand(2))
1345        return stringErr("linking module flags '" + ID->getString() +
1346                         "': IDs have conflicting values in '" +
1347                         SrcM->getModuleIdentifier() + "' and '" +
1348                         DstM.getModuleIdentifier() + "'");
1349      continue;
1350    }
1351    case Module::Warning: {
1352      break;
1353    }
1354    case Module::Max: {
1355      break;
1356    }
1357    case Module::Append: {
1358      MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1359      MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1360      SmallVector<Metadata *, 8> MDs;
1361      MDs.reserve(DstValue->getNumOperands() + SrcValue->getNumOperands());
1362      MDs.append(DstValue->op_begin(), DstValue->op_end());
1363      MDs.append(SrcValue->op_begin(), SrcValue->op_end());
1364
1365      replaceDstValue(MDNode::get(DstM.getContext(), MDs));
1366      break;
1367    }
1368    case Module::AppendUnique: {
1369      SmallSetVector<Metadata *, 16> Elts;
1370      MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1371      MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1372      Elts.insert(DstValue->op_begin(), DstValue->op_end());
1373      Elts.insert(SrcValue->op_begin(), SrcValue->op_end());
1374
1375      replaceDstValue(MDNode::get(DstM.getContext(),
1376                                  makeArrayRef(Elts.begin(), Elts.end())));
1377      break;
1378    }
1379    }
1380
1381  }
1382
1383  // Check all of the requirements.
1384  for (unsigned I = 0, E = Requirements.size(); I != E; ++I) {
1385    MDNode *Requirement = Requirements[I];
1386    MDString *Flag = cast<MDString>(Requirement->getOperand(0));
1387    Metadata *ReqValue = Requirement->getOperand(1);
1388
1389    MDNode *Op = Flags[Flag].first;
1390    if (!Op || Op->getOperand(2) != ReqValue)
1391      return stringErr("linking module flags '" + Flag->getString() +
1392                       "': does not have the required value");
1393  }
1394  return Error::success();
1395}
1396
1397/// Return InlineAsm adjusted with target-specific directives if required.
1398/// For ARM and Thumb, we have to add directives to select the appropriate ISA
1399/// to support mixing module-level inline assembly from ARM and Thumb modules.
1400static std::string adjustInlineAsm(const std::string &InlineAsm,
1401                                   const Triple &Triple) {
1402  if (Triple.getArch() == Triple::thumb || Triple.getArch() == Triple::thumbeb)
1403    return ".text\n.balign 2\n.thumb\n" + InlineAsm;
1404  if (Triple.getArch() == Triple::arm || Triple.getArch() == Triple::armeb)
1405    return ".text\n.balign 4\n.arm\n" + InlineAsm;
1406  return InlineAsm;
1407}
1408
1409Error IRLinker::run() {
1410  // Ensure metadata materialized before value mapping.
1411  if (SrcM->getMaterializer())
1412    if (Error Err = SrcM->getMaterializer()->materializeMetadata())
1413      return Err;
1414
1415  // Inherit the target data from the source module if the destination module
1416  // doesn't have one already.
1417  if (DstM.getDataLayout().isDefault())
1418    DstM.setDataLayout(SrcM->getDataLayout());
1419
1420  if (SrcM->getDataLayout() != DstM.getDataLayout()) {
1421    emitWarning("Linking two modules of different data layouts: '" +
1422                SrcM->getModuleIdentifier() + "' is '" +
1423                SrcM->getDataLayoutStr() + "' whereas '" +
1424                DstM.getModuleIdentifier() + "' is '" +
1425                DstM.getDataLayoutStr() + "'\n");
1426  }
1427
1428  // Copy the target triple from the source to dest if the dest's is empty.
1429  if (DstM.getTargetTriple().empty() && !SrcM->getTargetTriple().empty())
1430    DstM.setTargetTriple(SrcM->getTargetTriple());
1431
1432  Triple SrcTriple(SrcM->getTargetTriple()), DstTriple(DstM.getTargetTriple());
1433
1434  if (!SrcM->getTargetTriple().empty()&&
1435      !SrcTriple.isCompatibleWith(DstTriple))
1436    emitWarning("Linking two modules of different target triples: " +
1437                SrcM->getModuleIdentifier() + "' is '" +
1438                SrcM->getTargetTriple() + "' whereas '" +
1439                DstM.getModuleIdentifier() + "' is '" + DstM.getTargetTriple() +
1440                "'\n");
1441
1442  DstM.setTargetTriple(SrcTriple.merge(DstTriple));
1443
1444  // Append the module inline asm string.
1445  if (!IsPerformingImport && !SrcM->getModuleInlineAsm().empty()) {
1446    std::string SrcModuleInlineAsm = adjustInlineAsm(SrcM->getModuleInlineAsm(),
1447                                                     SrcTriple);
1448    if (DstM.getModuleInlineAsm().empty())
1449      DstM.setModuleInlineAsm(SrcModuleInlineAsm);
1450    else
1451      DstM.setModuleInlineAsm(DstM.getModuleInlineAsm() + "\n" +
1452                              SrcModuleInlineAsm);
1453  }
1454
1455  // Loop over all of the linked values to compute type mappings.
1456  computeTypeMapping();
1457
1458  std::reverse(Worklist.begin(), Worklist.end());
1459  while (!Worklist.empty()) {
1460    GlobalValue *GV = Worklist.back();
1461    Worklist.pop_back();
1462
1463    // Already mapped.
1464    if (ValueMap.find(GV) != ValueMap.end() ||
1465        IndirectSymbolValueMap.find(GV) != IndirectSymbolValueMap.end())
1466      continue;
1467
1468    assert(!GV->isDeclaration());
1469    Mapper.mapValue(*GV);
1470    if (FoundError)
1471      return std::move(*FoundError);
1472    flushRAUWWorklist();
1473  }
1474
1475  // Note that we are done linking global value bodies. This prevents
1476  // metadata linking from creating new references.
1477  DoneLinkingBodies = true;
1478  Mapper.addFlags(RF_NullMapMissingGlobalValues);
1479
1480  // Remap all of the named MDNodes in Src into the DstM module. We do this
1481  // after linking GlobalValues so that MDNodes that reference GlobalValues
1482  // are properly remapped.
1483  linkNamedMDNodes();
1484
1485  // Merge the module flags into the DstM module.
1486  return linkModuleFlagsMetadata();
1487}
1488
1489IRMover::StructTypeKeyInfo::KeyTy::KeyTy(ArrayRef<Type *> E, bool P)
1490    : ETypes(E), IsPacked(P) {}
1491
1492IRMover::StructTypeKeyInfo::KeyTy::KeyTy(const StructType *ST)
1493    : ETypes(ST->elements()), IsPacked(ST->isPacked()) {}
1494
1495bool IRMover::StructTypeKeyInfo::KeyTy::operator==(const KeyTy &That) const {
1496  return IsPacked == That.IsPacked && ETypes == That.ETypes;
1497}
1498
1499bool IRMover::StructTypeKeyInfo::KeyTy::operator!=(const KeyTy &That) const {
1500  return !this->operator==(That);
1501}
1502
1503StructType *IRMover::StructTypeKeyInfo::getEmptyKey() {
1504  return DenseMapInfo<StructType *>::getEmptyKey();
1505}
1506
1507StructType *IRMover::StructTypeKeyInfo::getTombstoneKey() {
1508  return DenseMapInfo<StructType *>::getTombstoneKey();
1509}
1510
1511unsigned IRMover::StructTypeKeyInfo::getHashValue(const KeyTy &Key) {
1512  return hash_combine(hash_combine_range(Key.ETypes.begin(), Key.ETypes.end()),
1513                      Key.IsPacked);
1514}
1515
1516unsigned IRMover::StructTypeKeyInfo::getHashValue(const StructType *ST) {
1517  return getHashValue(KeyTy(ST));
1518}
1519
1520bool IRMover::StructTypeKeyInfo::isEqual(const KeyTy &LHS,
1521                                         const StructType *RHS) {
1522  if (RHS == getEmptyKey() || RHS == getTombstoneKey())
1523    return false;
1524  return LHS == KeyTy(RHS);
1525}
1526
1527bool IRMover::StructTypeKeyInfo::isEqual(const StructType *LHS,
1528                                         const StructType *RHS) {
1529  if (RHS == getEmptyKey() || RHS == getTombstoneKey())
1530    return LHS == RHS;
1531  return KeyTy(LHS) == KeyTy(RHS);
1532}
1533
1534void IRMover::IdentifiedStructTypeSet::addNonOpaque(StructType *Ty) {
1535  assert(!Ty->isOpaque());
1536  NonOpaqueStructTypes.insert(Ty);
1537}
1538
1539void IRMover::IdentifiedStructTypeSet::switchToNonOpaque(StructType *Ty) {
1540  assert(!Ty->isOpaque());
1541  NonOpaqueStructTypes.insert(Ty);
1542  bool Removed = OpaqueStructTypes.erase(Ty);
1543  (void)Removed;
1544  assert(Removed);
1545}
1546
1547void IRMover::IdentifiedStructTypeSet::addOpaque(StructType *Ty) {
1548  assert(Ty->isOpaque());
1549  OpaqueStructTypes.insert(Ty);
1550}
1551
1552StructType *
1553IRMover::IdentifiedStructTypeSet::findNonOpaque(ArrayRef<Type *> ETypes,
1554                                                bool IsPacked) {
1555  IRMover::StructTypeKeyInfo::KeyTy Key(ETypes, IsPacked);
1556  auto I = NonOpaqueStructTypes.find_as(Key);
1557  return I == NonOpaqueStructTypes.end() ? nullptr : *I;
1558}
1559
1560bool IRMover::IdentifiedStructTypeSet::hasType(StructType *Ty) {
1561  if (Ty->isOpaque())
1562    return OpaqueStructTypes.count(Ty);
1563  auto I = NonOpaqueStructTypes.find(Ty);
1564  return I == NonOpaqueStructTypes.end() ? false : *I == Ty;
1565}
1566
1567IRMover::IRMover(Module &M) : Composite(M) {
1568  TypeFinder StructTypes;
1569  StructTypes.run(M, /* OnlyNamed */ false);
1570  for (StructType *Ty : StructTypes) {
1571    if (Ty->isOpaque())
1572      IdentifiedStructTypes.addOpaque(Ty);
1573    else
1574      IdentifiedStructTypes.addNonOpaque(Ty);
1575  }
1576  // Self-map metadatas in the destination module. This is needed when
1577  // DebugTypeODRUniquing is enabled on the LLVMContext, since metadata in the
1578  // destination module may be reached from the source module.
1579  for (auto *MD : StructTypes.getVisitedMetadata()) {
1580    SharedMDs[MD].reset(const_cast<MDNode *>(MD));
1581  }
1582}
1583
1584Error IRMover::move(
1585    std::unique_ptr<Module> Src, ArrayRef<GlobalValue *> ValuesToLink,
1586    std::function<void(GlobalValue &, ValueAdder Add)> AddLazyFor,
1587    bool IsPerformingImport) {
1588  IRLinker TheIRLinker(Composite, SharedMDs, IdentifiedStructTypes,
1589                       std::move(Src), ValuesToLink, std::move(AddLazyFor),
1590                       IsPerformingImport);
1591  Error E = TheIRLinker.run();
1592  Composite.dropTriviallyDeadConstantArrays();
1593  return E;
1594}
1595