LinkModules.cpp revision cd81d94322a39503e4a3e87b6ee03d4fcb3465fb
1//===- lib/Linker/LinkModules.cpp - Module Linker Implementation ----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the LLVM module linker.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Linker/Linker.h"
15#include "llvm-c/Linker.h"
16#include "llvm/ADT/Optional.h"
17#include "llvm/ADT/SetVector.h"
18#include "llvm/ADT/SmallString.h"
19#include "llvm/IR/Constants.h"
20#include "llvm/IR/Module.h"
21#include "llvm/IR/TypeFinder.h"
22#include "llvm/Support/CommandLine.h"
23#include "llvm/Support/Debug.h"
24#include "llvm/Support/raw_ostream.h"
25#include "llvm/Transforms/Utils/Cloning.h"
26#include <cctype>
27#include <tuple>
28using namespace llvm;
29
30
31//===----------------------------------------------------------------------===//
32// TypeMap implementation.
33//===----------------------------------------------------------------------===//
34
35namespace {
36  typedef SmallPtrSet<StructType*, 32> TypeSet;
37
38class TypeMapTy : public ValueMapTypeRemapper {
39  /// MappedTypes - This is a mapping from a source type to a destination type
40  /// to use.
41  DenseMap<Type*, Type*> MappedTypes;
42
43  /// SpeculativeTypes - When checking to see if two subgraphs are isomorphic,
44  /// we speculatively add types to MappedTypes, but keep track of them here in
45  /// case we need to roll back.
46  SmallVector<Type*, 16> SpeculativeTypes;
47
48  /// SrcDefinitionsToResolve - This is a list of non-opaque structs in the
49  /// source module that are mapped to an opaque struct in the destination
50  /// module.
51  SmallVector<StructType*, 16> SrcDefinitionsToResolve;
52
53  /// DstResolvedOpaqueTypes - This is the set of opaque types in the
54  /// destination modules who are getting a body from the source module.
55  SmallPtrSet<StructType*, 16> DstResolvedOpaqueTypes;
56
57public:
58  TypeMapTy(TypeSet &Set) : DstStructTypesSet(Set) {}
59
60  TypeSet &DstStructTypesSet;
61  /// addTypeMapping - Indicate that the specified type in the destination
62  /// module is conceptually equivalent to the specified type in the source
63  /// module.
64  void addTypeMapping(Type *DstTy, Type *SrcTy);
65
66  /// linkDefinedTypeBodies - Produce a body for an opaque type in the dest
67  /// module from a type definition in the source module.
68  void linkDefinedTypeBodies();
69
70  /// get - Return the mapped type to use for the specified input type from the
71  /// source module.
72  Type *get(Type *SrcTy);
73
74  FunctionType *get(FunctionType *T) {return cast<FunctionType>(get((Type*)T));}
75
76  /// dump - Dump out the type map for debugging purposes.
77  void dump() const {
78    for (DenseMap<Type*, Type*>::const_iterator
79           I = MappedTypes.begin(), E = MappedTypes.end(); I != E; ++I) {
80      dbgs() << "TypeMap: ";
81      I->first->dump();
82      dbgs() << " => ";
83      I->second->dump();
84      dbgs() << '\n';
85    }
86  }
87
88private:
89  Type *getImpl(Type *T);
90  /// remapType - Implement the ValueMapTypeRemapper interface.
91  Type *remapType(Type *SrcTy) override {
92    return get(SrcTy);
93  }
94
95  bool areTypesIsomorphic(Type *DstTy, Type *SrcTy);
96};
97}
98
99void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) {
100  Type *&Entry = MappedTypes[SrcTy];
101  if (Entry) return;
102
103  if (DstTy == SrcTy) {
104    Entry = DstTy;
105    return;
106  }
107
108  // Check to see if these types are recursively isomorphic and establish a
109  // mapping between them if so.
110  if (!areTypesIsomorphic(DstTy, SrcTy)) {
111    // Oops, they aren't isomorphic.  Just discard this request by rolling out
112    // any speculative mappings we've established.
113    for (unsigned i = 0, e = SpeculativeTypes.size(); i != e; ++i)
114      MappedTypes.erase(SpeculativeTypes[i]);
115  }
116  SpeculativeTypes.clear();
117}
118
119/// areTypesIsomorphic - Recursively walk this pair of types, returning true
120/// if they are isomorphic, false if they are not.
121bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) {
122  // Two types with differing kinds are clearly not isomorphic.
123  if (DstTy->getTypeID() != SrcTy->getTypeID()) return false;
124
125  // If we have an entry in the MappedTypes table, then we have our answer.
126  Type *&Entry = MappedTypes[SrcTy];
127  if (Entry)
128    return Entry == DstTy;
129
130  // Two identical types are clearly isomorphic.  Remember this
131  // non-speculatively.
132  if (DstTy == SrcTy) {
133    Entry = DstTy;
134    return true;
135  }
136
137  // Okay, we have two types with identical kinds that we haven't seen before.
138
139  // If this is an opaque struct type, special case it.
140  if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) {
141    // Mapping an opaque type to any struct, just keep the dest struct.
142    if (SSTy->isOpaque()) {
143      Entry = DstTy;
144      SpeculativeTypes.push_back(SrcTy);
145      return true;
146    }
147
148    // Mapping a non-opaque source type to an opaque dest.  If this is the first
149    // type that we're mapping onto this destination type then we succeed.  Keep
150    // the dest, but fill it in later.  This doesn't need to be speculative.  If
151    // this is the second (different) type that we're trying to map onto the
152    // same opaque type then we fail.
153    if (cast<StructType>(DstTy)->isOpaque()) {
154      // We can only map one source type onto the opaque destination type.
155      if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)))
156        return false;
157      SrcDefinitionsToResolve.push_back(SSTy);
158      Entry = DstTy;
159      return true;
160    }
161  }
162
163  // If the number of subtypes disagree between the two types, then we fail.
164  if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes())
165    return false;
166
167  // Fail if any of the extra properties (e.g. array size) of the type disagree.
168  if (isa<IntegerType>(DstTy))
169    return false;  // bitwidth disagrees.
170  if (PointerType *PT = dyn_cast<PointerType>(DstTy)) {
171    if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace())
172      return false;
173
174  } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) {
175    if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg())
176      return false;
177  } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) {
178    StructType *SSTy = cast<StructType>(SrcTy);
179    if (DSTy->isLiteral() != SSTy->isLiteral() ||
180        DSTy->isPacked() != SSTy->isPacked())
181      return false;
182  } else if (ArrayType *DATy = dyn_cast<ArrayType>(DstTy)) {
183    if (DATy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
184      return false;
185  } else if (VectorType *DVTy = dyn_cast<VectorType>(DstTy)) {
186    if (DVTy->getNumElements() != cast<VectorType>(SrcTy)->getNumElements())
187      return false;
188  }
189
190  // Otherwise, we speculate that these two types will line up and recursively
191  // check the subelements.
192  Entry = DstTy;
193  SpeculativeTypes.push_back(SrcTy);
194
195  for (unsigned i = 0, e = SrcTy->getNumContainedTypes(); i != e; ++i)
196    if (!areTypesIsomorphic(DstTy->getContainedType(i),
197                            SrcTy->getContainedType(i)))
198      return false;
199
200  // If everything seems to have lined up, then everything is great.
201  return true;
202}
203
204/// linkDefinedTypeBodies - Produce a body for an opaque type in the dest
205/// module from a type definition in the source module.
206void TypeMapTy::linkDefinedTypeBodies() {
207  SmallVector<Type*, 16> Elements;
208  SmallString<16> TmpName;
209
210  // Note that processing entries in this loop (calling 'get') can add new
211  // entries to the SrcDefinitionsToResolve vector.
212  while (!SrcDefinitionsToResolve.empty()) {
213    StructType *SrcSTy = SrcDefinitionsToResolve.pop_back_val();
214    StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]);
215
216    // TypeMap is a many-to-one mapping, if there were multiple types that
217    // provide a body for DstSTy then previous iterations of this loop may have
218    // already handled it.  Just ignore this case.
219    if (!DstSTy->isOpaque()) continue;
220    assert(!SrcSTy->isOpaque() && "Not resolving a definition?");
221
222    // Map the body of the source type over to a new body for the dest type.
223    Elements.resize(SrcSTy->getNumElements());
224    for (unsigned i = 0, e = Elements.size(); i != e; ++i)
225      Elements[i] = getImpl(SrcSTy->getElementType(i));
226
227    DstSTy->setBody(Elements, SrcSTy->isPacked());
228
229    // If DstSTy has no name or has a longer name than STy, then viciously steal
230    // STy's name.
231    if (!SrcSTy->hasName()) continue;
232    StringRef SrcName = SrcSTy->getName();
233
234    if (!DstSTy->hasName() || DstSTy->getName().size() > SrcName.size()) {
235      TmpName.insert(TmpName.end(), SrcName.begin(), SrcName.end());
236      SrcSTy->setName("");
237      DstSTy->setName(TmpName.str());
238      TmpName.clear();
239    }
240  }
241
242  DstResolvedOpaqueTypes.clear();
243}
244
245/// get - Return the mapped type to use for the specified input type from the
246/// source module.
247Type *TypeMapTy::get(Type *Ty) {
248  Type *Result = getImpl(Ty);
249
250  // If this caused a reference to any struct type, resolve it before returning.
251  if (!SrcDefinitionsToResolve.empty())
252    linkDefinedTypeBodies();
253  return Result;
254}
255
256/// getImpl - This is the recursive version of get().
257Type *TypeMapTy::getImpl(Type *Ty) {
258  // If we already have an entry for this type, return it.
259  Type **Entry = &MappedTypes[Ty];
260  if (*Entry) return *Entry;
261
262  // If this is not a named struct type, then just map all of the elements and
263  // then rebuild the type from inside out.
264  if (!isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral()) {
265    // If there are no element types to map, then the type is itself.  This is
266    // true for the anonymous {} struct, things like 'float', integers, etc.
267    if (Ty->getNumContainedTypes() == 0)
268      return *Entry = Ty;
269
270    // Remap all of the elements, keeping track of whether any of them change.
271    bool AnyChange = false;
272    SmallVector<Type*, 4> ElementTypes;
273    ElementTypes.resize(Ty->getNumContainedTypes());
274    for (unsigned i = 0, e = Ty->getNumContainedTypes(); i != e; ++i) {
275      ElementTypes[i] = getImpl(Ty->getContainedType(i));
276      AnyChange |= ElementTypes[i] != Ty->getContainedType(i);
277    }
278
279    // If we found our type while recursively processing stuff, just use it.
280    Entry = &MappedTypes[Ty];
281    if (*Entry) return *Entry;
282
283    // If all of the element types mapped directly over, then the type is usable
284    // as-is.
285    if (!AnyChange)
286      return *Entry = Ty;
287
288    // Otherwise, rebuild a modified type.
289    switch (Ty->getTypeID()) {
290    default: llvm_unreachable("unknown derived type to remap");
291    case Type::ArrayTyID:
292      return *Entry = ArrayType::get(ElementTypes[0],
293                                     cast<ArrayType>(Ty)->getNumElements());
294    case Type::VectorTyID:
295      return *Entry = VectorType::get(ElementTypes[0],
296                                      cast<VectorType>(Ty)->getNumElements());
297    case Type::PointerTyID:
298      return *Entry = PointerType::get(ElementTypes[0],
299                                      cast<PointerType>(Ty)->getAddressSpace());
300    case Type::FunctionTyID:
301      return *Entry = FunctionType::get(ElementTypes[0],
302                                        makeArrayRef(ElementTypes).slice(1),
303                                        cast<FunctionType>(Ty)->isVarArg());
304    case Type::StructTyID:
305      // Note that this is only reached for anonymous structs.
306      return *Entry = StructType::get(Ty->getContext(), ElementTypes,
307                                      cast<StructType>(Ty)->isPacked());
308    }
309  }
310
311  // Otherwise, this is an unmapped named struct.  If the struct can be directly
312  // mapped over, just use it as-is.  This happens in a case when the linked-in
313  // module has something like:
314  //   %T = type {%T*, i32}
315  //   @GV = global %T* null
316  // where T does not exist at all in the destination module.
317  //
318  // The other case we watch for is when the type is not in the destination
319  // module, but that it has to be rebuilt because it refers to something that
320  // is already mapped.  For example, if the destination module has:
321  //  %A = type { i32 }
322  // and the source module has something like
323  //  %A' = type { i32 }
324  //  %B = type { %A'* }
325  //  @GV = global %B* null
326  // then we want to create a new type: "%B = type { %A*}" and have it take the
327  // pristine "%B" name from the source module.
328  //
329  // To determine which case this is, we have to recursively walk the type graph
330  // speculating that we'll be able to reuse it unmodified.  Only if this is
331  // safe would we map the entire thing over.  Because this is an optimization,
332  // and is not required for the prettiness of the linked module, we just skip
333  // it and always rebuild a type here.
334  StructType *STy = cast<StructType>(Ty);
335
336  // If the type is opaque, we can just use it directly.
337  if (STy->isOpaque()) {
338    // A named structure type from src module is used. Add it to the Set of
339    // identified structs in the destination module.
340    DstStructTypesSet.insert(STy);
341    return *Entry = STy;
342  }
343
344  // Otherwise we create a new type and resolve its body later.  This will be
345  // resolved by the top level of get().
346  SrcDefinitionsToResolve.push_back(STy);
347  StructType *DTy = StructType::create(STy->getContext());
348  // A new identified structure type was created. Add it to the set of
349  // identified structs in the destination module.
350  DstStructTypesSet.insert(DTy);
351  DstResolvedOpaqueTypes.insert(DTy);
352  return *Entry = DTy;
353}
354
355//===----------------------------------------------------------------------===//
356// ModuleLinker implementation.
357//===----------------------------------------------------------------------===//
358
359namespace {
360  class ModuleLinker;
361
362  /// ValueMaterializerTy - Creates prototypes for functions that are lazily
363  /// linked on the fly. This speeds up linking for modules with many
364  /// lazily linked functions of which few get used.
365  class ValueMaterializerTy : public ValueMaterializer {
366    TypeMapTy &TypeMap;
367    Module *DstM;
368    std::vector<Function*> &LazilyLinkFunctions;
369  public:
370    ValueMaterializerTy(TypeMapTy &TypeMap, Module *DstM,
371                        std::vector<Function*> &LazilyLinkFunctions) :
372      ValueMaterializer(), TypeMap(TypeMap), DstM(DstM),
373      LazilyLinkFunctions(LazilyLinkFunctions) {
374    }
375
376    Value *materializeValueFor(Value *V) override;
377  };
378
379  /// ModuleLinker - This is an implementation class for the LinkModules
380  /// function, which is the entrypoint for this file.
381  class ModuleLinker {
382    Module *DstM, *SrcM;
383
384    TypeMapTy TypeMap;
385    ValueMaterializerTy ValMaterializer;
386
387    /// ValueMap - Mapping of values from what they used to be in Src, to what
388    /// they are now in DstM.  ValueToValueMapTy is a ValueMap, which involves
389    /// some overhead due to the use of Value handles which the Linker doesn't
390    /// actually need, but this allows us to reuse the ValueMapper code.
391    ValueToValueMapTy ValueMap;
392
393    struct AppendingVarInfo {
394      GlobalVariable *NewGV;  // New aggregate global in dest module.
395      Constant *DstInit;      // Old initializer from dest module.
396      Constant *SrcInit;      // Old initializer from src module.
397    };
398
399    std::vector<AppendingVarInfo> AppendingVars;
400
401    unsigned Mode; // Mode to treat source module.
402
403    // Set of items not to link in from source.
404    SmallPtrSet<const Value*, 16> DoNotLinkFromSource;
405
406    // Vector of functions to lazily link in.
407    std::vector<Function*> LazilyLinkFunctions;
408
409    bool SuppressWarnings;
410
411  public:
412    std::string ErrorMsg;
413
414    ModuleLinker(Module *dstM, TypeSet &Set, Module *srcM, unsigned mode,
415                 bool SuppressWarnings=false)
416        : DstM(dstM), SrcM(srcM), TypeMap(Set),
417          ValMaterializer(TypeMap, DstM, LazilyLinkFunctions), Mode(mode),
418          SuppressWarnings(SuppressWarnings) {}
419
420    bool run();
421
422  private:
423    /// emitError - Helper method for setting a message and returning an error
424    /// code.
425    bool emitError(const Twine &Message) {
426      ErrorMsg = Message.str();
427      return true;
428    }
429
430    bool getComdatLeader(Module *M, StringRef ComdatName,
431                         const GlobalVariable *&GVar);
432    bool computeResultingSelectionKind(StringRef ComdatName,
433                                       Comdat::SelectionKind Src,
434                                       Comdat::SelectionKind Dst,
435                                       Comdat::SelectionKind &Result,
436                                       bool &LinkFromSrc);
437    std::map<const Comdat *, std::pair<Comdat::SelectionKind, bool>>
438        ComdatsChosen;
439    bool getComdatResult(const Comdat *SrcC, Comdat::SelectionKind &SK,
440                         bool &LinkFromSrc);
441
442    /// getLinkageResult - This analyzes the two global values and determines
443    /// what the result will look like in the destination module.
444    bool getLinkageResult(GlobalValue *Dest, const GlobalValue *Src,
445                          GlobalValue::LinkageTypes &LT,
446                          GlobalValue::VisibilityTypes &Vis,
447                          bool &LinkFromSrc);
448
449    /// getLinkedToGlobal - Given a global in the source module, return the
450    /// global in the destination module that is being linked to, if any.
451    GlobalValue *getLinkedToGlobal(GlobalValue *SrcGV) {
452      // If the source has no name it can't link.  If it has local linkage,
453      // there is no name match-up going on.
454      if (!SrcGV->hasName() || SrcGV->hasLocalLinkage())
455        return nullptr;
456
457      // Otherwise see if we have a match in the destination module's symtab.
458      GlobalValue *DGV = DstM->getNamedValue(SrcGV->getName());
459      if (!DGV) return nullptr;
460
461      // If we found a global with the same name in the dest module, but it has
462      // internal linkage, we are really not doing any linkage here.
463      if (DGV->hasLocalLinkage())
464        return nullptr;
465
466      // Otherwise, we do in fact link to the destination global.
467      return DGV;
468    }
469
470    void computeTypeMapping();
471
472    bool linkAppendingVarProto(GlobalVariable *DstGV, GlobalVariable *SrcGV);
473    bool linkGlobalProto(GlobalVariable *SrcGV);
474    bool linkFunctionProto(Function *SrcF);
475    bool linkAliasProto(GlobalAlias *SrcA);
476    bool linkModuleFlagsMetadata();
477
478    void linkAppendingVarInit(const AppendingVarInfo &AVI);
479    void linkGlobalInits();
480    void linkFunctionBody(Function *Dst, Function *Src);
481    void linkAliasBodies();
482    void linkNamedMDNodes();
483  };
484}
485
486/// forceRenaming - The LLVM SymbolTable class autorenames globals that conflict
487/// in the symbol table.  This is good for all clients except for us.  Go
488/// through the trouble to force this back.
489static void forceRenaming(GlobalValue *GV, StringRef Name) {
490  // If the global doesn't force its name or if it already has the right name,
491  // there is nothing for us to do.
492  if (GV->hasLocalLinkage() || GV->getName() == Name)
493    return;
494
495  Module *M = GV->getParent();
496
497  // If there is a conflict, rename the conflict.
498  if (GlobalValue *ConflictGV = M->getNamedValue(Name)) {
499    GV->takeName(ConflictGV);
500    ConflictGV->setName(Name);    // This will cause ConflictGV to get renamed
501    assert(ConflictGV->getName() != Name && "forceRenaming didn't work");
502  } else {
503    GV->setName(Name);              // Force the name back
504  }
505}
506
507/// copyGVAttributes - copy additional attributes (those not needed to construct
508/// a GlobalValue) from the SrcGV to the DestGV.
509static void copyGVAttributes(GlobalValue *DestGV, const GlobalValue *SrcGV) {
510  // Use the maximum alignment, rather than just copying the alignment of SrcGV.
511  auto *DestGO = dyn_cast<GlobalObject>(DestGV);
512  unsigned Alignment;
513  if (DestGO)
514    Alignment = std::max(DestGO->getAlignment(), SrcGV->getAlignment());
515
516  DestGV->copyAttributesFrom(SrcGV);
517
518  if (DestGO)
519    DestGO->setAlignment(Alignment);
520
521  forceRenaming(DestGV, SrcGV->getName());
522}
523
524static bool isLessConstraining(GlobalValue::VisibilityTypes a,
525                               GlobalValue::VisibilityTypes b) {
526  if (a == GlobalValue::HiddenVisibility)
527    return false;
528  if (b == GlobalValue::HiddenVisibility)
529    return true;
530  if (a == GlobalValue::ProtectedVisibility)
531    return false;
532  if (b == GlobalValue::ProtectedVisibility)
533    return true;
534  return false;
535}
536
537Value *ValueMaterializerTy::materializeValueFor(Value *V) {
538  Function *SF = dyn_cast<Function>(V);
539  if (!SF)
540    return nullptr;
541
542  Function *DF = Function::Create(TypeMap.get(SF->getFunctionType()),
543                                  SF->getLinkage(), SF->getName(), DstM);
544  copyGVAttributes(DF, SF);
545
546  LazilyLinkFunctions.push_back(SF);
547  return DF;
548}
549
550bool ModuleLinker::getComdatLeader(Module *M, StringRef ComdatName,
551                                   const GlobalVariable *&GVar) {
552  const GlobalValue *GVal = M->getNamedValue(ComdatName);
553  if (const auto *GA = dyn_cast_or_null<GlobalAlias>(GVal)) {
554    GVal = GA->getBaseObject();
555    if (!GVal)
556      // We cannot resolve the size of the aliasee yet.
557      return emitError("Linking COMDATs named '" + ComdatName +
558                       "': COMDAT key involves incomputable alias size.");
559  }
560
561  GVar = dyn_cast_or_null<GlobalVariable>(GVal);
562  if (!GVar)
563    return emitError(
564        "Linking COMDATs named '" + ComdatName +
565        "': GlobalVariable required for data dependent selection!");
566
567  return false;
568}
569
570bool ModuleLinker::computeResultingSelectionKind(StringRef ComdatName,
571                                                 Comdat::SelectionKind Src,
572                                                 Comdat::SelectionKind Dst,
573                                                 Comdat::SelectionKind &Result,
574                                                 bool &LinkFromSrc) {
575  // The ability to mix Comdat::SelectionKind::Any with
576  // Comdat::SelectionKind::Largest is a behavior that comes from COFF.
577  bool DstAnyOrLargest = Dst == Comdat::SelectionKind::Any ||
578                         Dst == Comdat::SelectionKind::Largest;
579  bool SrcAnyOrLargest = Src == Comdat::SelectionKind::Any ||
580                         Src == Comdat::SelectionKind::Largest;
581  if (DstAnyOrLargest && SrcAnyOrLargest) {
582    if (Dst == Comdat::SelectionKind::Largest ||
583        Src == Comdat::SelectionKind::Largest)
584      Result = Comdat::SelectionKind::Largest;
585    else
586      Result = Comdat::SelectionKind::Any;
587  } else if (Src == Dst) {
588    Result = Dst;
589  } else {
590    return emitError("Linking COMDATs named '" + ComdatName +
591                     "': invalid selection kinds!");
592  }
593
594  switch (Result) {
595  case Comdat::SelectionKind::Any:
596    // Go with Dst.
597    LinkFromSrc = false;
598    break;
599  case Comdat::SelectionKind::NoDuplicates:
600    return emitError("Linking COMDATs named '" + ComdatName +
601                     "': noduplicates has been violated!");
602  case Comdat::SelectionKind::ExactMatch:
603  case Comdat::SelectionKind::Largest:
604  case Comdat::SelectionKind::SameSize: {
605    const GlobalVariable *DstGV;
606    const GlobalVariable *SrcGV;
607    if (getComdatLeader(DstM, ComdatName, DstGV) ||
608        getComdatLeader(SrcM, ComdatName, SrcGV))
609      return true;
610
611    const DataLayout *DstDL = DstM->getDataLayout();
612    const DataLayout *SrcDL = SrcM->getDataLayout();
613    if (!DstDL || !SrcDL) {
614      return emitError(
615          "Linking COMDATs named '" + ComdatName +
616          "': can't do size dependent selection without DataLayout!");
617    }
618    uint64_t DstSize =
619        DstDL->getTypeAllocSize(DstGV->getType()->getPointerElementType());
620    uint64_t SrcSize =
621        SrcDL->getTypeAllocSize(SrcGV->getType()->getPointerElementType());
622    if (Result == Comdat::SelectionKind::ExactMatch) {
623      if (SrcGV->getInitializer() != DstGV->getInitializer())
624        return emitError("Linking COMDATs named '" + ComdatName +
625                         "': ExactMatch violated!");
626      LinkFromSrc = false;
627    } else if (Result == Comdat::SelectionKind::Largest) {
628      LinkFromSrc = SrcSize > DstSize;
629    } else if (Result == Comdat::SelectionKind::SameSize) {
630      if (SrcSize != DstSize)
631        return emitError("Linking COMDATs named '" + ComdatName +
632                         "': SameSize violated!");
633      LinkFromSrc = false;
634    } else {
635      llvm_unreachable("unknown selection kind");
636    }
637    break;
638  }
639  }
640
641  return false;
642}
643
644bool ModuleLinker::getComdatResult(const Comdat *SrcC,
645                                   Comdat::SelectionKind &Result,
646                                   bool &LinkFromSrc) {
647  StringRef ComdatName = SrcC->getName();
648  Module::ComdatSymTabType &ComdatSymTab = DstM->getComdatSymbolTable();
649  Module::ComdatSymTabType::iterator DstCI = ComdatSymTab.find(ComdatName);
650  if (DstCI != ComdatSymTab.end()) {
651    const Comdat *DstC = &DstCI->second;
652    Comdat::SelectionKind SSK = SrcC->getSelectionKind();
653    Comdat::SelectionKind DSK = DstC->getSelectionKind();
654    if (computeResultingSelectionKind(ComdatName, SSK, DSK, Result, LinkFromSrc))
655      return true;
656  }
657  return false;
658}
659
660/// getLinkageResult - This analyzes the two global values and determines what
661/// the result will look like in the destination module.  In particular, it
662/// computes the resultant linkage type and visibility, computes whether the
663/// global in the source should be copied over to the destination (replacing
664/// the existing one), and computes whether this linkage is an error or not.
665bool ModuleLinker::getLinkageResult(GlobalValue *Dest, const GlobalValue *Src,
666                                    GlobalValue::LinkageTypes &LT,
667                                    GlobalValue::VisibilityTypes &Vis,
668                                    bool &LinkFromSrc) {
669  assert(Dest && "Must have two globals being queried");
670  assert(!Src->hasLocalLinkage() &&
671         "If Src has internal linkage, Dest shouldn't be set!");
672
673  bool SrcIsDeclaration = Src->isDeclaration() && !Src->isMaterializable();
674  bool DestIsDeclaration = Dest->isDeclaration();
675
676  if (SrcIsDeclaration) {
677    // If Src is external or if both Src & Dest are external..  Just link the
678    // external globals, we aren't adding anything.
679    if (Src->hasDLLImportStorageClass()) {
680      // If one of GVs is marked as DLLImport, result should be dllimport'ed.
681      if (DestIsDeclaration) {
682        LinkFromSrc = true;
683        LT = Src->getLinkage();
684      }
685    } else if (Dest->hasExternalWeakLinkage()) {
686      // If the Dest is weak, use the source linkage.
687      LinkFromSrc = true;
688      LT = Src->getLinkage();
689    } else {
690      LinkFromSrc = false;
691      LT = Dest->getLinkage();
692    }
693  } else if (DestIsDeclaration && !Dest->hasDLLImportStorageClass()) {
694    // If Dest is external but Src is not:
695    LinkFromSrc = true;
696    LT = Src->getLinkage();
697  } else if (Src->isWeakForLinker()) {
698    // At this point we know that Dest has LinkOnce, External*, Weak, Common,
699    // or DLL* linkage.
700    if (Dest->hasExternalWeakLinkage() ||
701        Dest->hasAvailableExternallyLinkage() ||
702        (Dest->hasLinkOnceLinkage() &&
703         (Src->hasWeakLinkage() || Src->hasCommonLinkage()))) {
704      LinkFromSrc = true;
705      LT = Src->getLinkage();
706    } else {
707      LinkFromSrc = false;
708      LT = Dest->getLinkage();
709    }
710  } else if (Dest->isWeakForLinker()) {
711    // At this point we know that Src has External* or DLL* linkage.
712    if (Src->hasExternalWeakLinkage()) {
713      LinkFromSrc = false;
714      LT = Dest->getLinkage();
715    } else {
716      LinkFromSrc = true;
717      LT = GlobalValue::ExternalLinkage;
718    }
719  } else {
720    assert((Dest->hasExternalLinkage()  || Dest->hasExternalWeakLinkage()) &&
721           (Src->hasExternalLinkage()   || Src->hasExternalWeakLinkage()) &&
722           "Unexpected linkage type!");
723    return emitError("Linking globals named '" + Src->getName() +
724                 "': symbol multiply defined!");
725  }
726
727  // Compute the visibility. We follow the rules in the System V Application
728  // Binary Interface.
729  assert(!GlobalValue::isLocalLinkage(LT) &&
730         "Symbols with local linkage should not be merged");
731  Vis = isLessConstraining(Src->getVisibility(), Dest->getVisibility()) ?
732    Dest->getVisibility() : Src->getVisibility();
733  return false;
734}
735
736/// computeTypeMapping - Loop over all of the linked values to compute type
737/// mappings.  For example, if we link "extern Foo *x" and "Foo *x = NULL", then
738/// we have two struct types 'Foo' but one got renamed when the module was
739/// loaded into the same LLVMContext.
740void ModuleLinker::computeTypeMapping() {
741  // Incorporate globals.
742  for (Module::global_iterator I = SrcM->global_begin(),
743       E = SrcM->global_end(); I != E; ++I) {
744    GlobalValue *DGV = getLinkedToGlobal(I);
745    if (!DGV) continue;
746
747    if (!DGV->hasAppendingLinkage() || !I->hasAppendingLinkage()) {
748      TypeMap.addTypeMapping(DGV->getType(), I->getType());
749      continue;
750    }
751
752    // Unify the element type of appending arrays.
753    ArrayType *DAT = cast<ArrayType>(DGV->getType()->getElementType());
754    ArrayType *SAT = cast<ArrayType>(I->getType()->getElementType());
755    TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
756  }
757
758  // Incorporate functions.
759  for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I) {
760    if (GlobalValue *DGV = getLinkedToGlobal(I))
761      TypeMap.addTypeMapping(DGV->getType(), I->getType());
762  }
763
764  // Incorporate types by name, scanning all the types in the source module.
765  // At this point, the destination module may have a type "%foo = { i32 }" for
766  // example.  When the source module got loaded into the same LLVMContext, if
767  // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
768  TypeFinder SrcStructTypes;
769  SrcStructTypes.run(*SrcM, true);
770  SmallPtrSet<StructType*, 32> SrcStructTypesSet(SrcStructTypes.begin(),
771                                                 SrcStructTypes.end());
772
773  for (unsigned i = 0, e = SrcStructTypes.size(); i != e; ++i) {
774    StructType *ST = SrcStructTypes[i];
775    if (!ST->hasName()) continue;
776
777    // Check to see if there is a dot in the name followed by a digit.
778    size_t DotPos = ST->getName().rfind('.');
779    if (DotPos == 0 || DotPos == StringRef::npos ||
780        ST->getName().back() == '.' ||
781        !isdigit(static_cast<unsigned char>(ST->getName()[DotPos+1])))
782      continue;
783
784    // Check to see if the destination module has a struct with the prefix name.
785    if (StructType *DST = DstM->getTypeByName(ST->getName().substr(0, DotPos)))
786      // Don't use it if this actually came from the source module. They're in
787      // the same LLVMContext after all. Also don't use it unless the type is
788      // actually used in the destination module. This can happen in situations
789      // like this:
790      //
791      //      Module A                         Module B
792      //      --------                         --------
793      //   %Z = type { %A }                %B = type { %C.1 }
794      //   %A = type { %B.1, [7 x i8] }    %C.1 = type { i8* }
795      //   %B.1 = type { %C }              %A.2 = type { %B.3, [5 x i8] }
796      //   %C = type { i8* }               %B.3 = type { %C.1 }
797      //
798      // When we link Module B with Module A, the '%B' in Module B is
799      // used. However, that would then use '%C.1'. But when we process '%C.1',
800      // we prefer to take the '%C' version. So we are then left with both
801      // '%C.1' and '%C' being used for the same types. This leads to some
802      // variables using one type and some using the other.
803      if (!SrcStructTypesSet.count(DST) && TypeMap.DstStructTypesSet.count(DST))
804        TypeMap.addTypeMapping(DST, ST);
805  }
806
807  // Don't bother incorporating aliases, they aren't generally typed well.
808
809  // Now that we have discovered all of the type equivalences, get a body for
810  // any 'opaque' types in the dest module that are now resolved.
811  TypeMap.linkDefinedTypeBodies();
812}
813
814/// linkAppendingVarProto - If there were any appending global variables, link
815/// them together now.  Return true on error.
816bool ModuleLinker::linkAppendingVarProto(GlobalVariable *DstGV,
817                                         GlobalVariable *SrcGV) {
818
819  if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage())
820    return emitError("Linking globals named '" + SrcGV->getName() +
821           "': can only link appending global with another appending global!");
822
823  ArrayType *DstTy = cast<ArrayType>(DstGV->getType()->getElementType());
824  ArrayType *SrcTy =
825    cast<ArrayType>(TypeMap.get(SrcGV->getType()->getElementType()));
826  Type *EltTy = DstTy->getElementType();
827
828  // Check to see that they two arrays agree on type.
829  if (EltTy != SrcTy->getElementType())
830    return emitError("Appending variables with different element types!");
831  if (DstGV->isConstant() != SrcGV->isConstant())
832    return emitError("Appending variables linked with different const'ness!");
833
834  if (DstGV->getAlignment() != SrcGV->getAlignment())
835    return emitError(
836             "Appending variables with different alignment need to be linked!");
837
838  if (DstGV->getVisibility() != SrcGV->getVisibility())
839    return emitError(
840            "Appending variables with different visibility need to be linked!");
841
842  if (DstGV->hasUnnamedAddr() != SrcGV->hasUnnamedAddr())
843    return emitError(
844        "Appending variables with different unnamed_addr need to be linked!");
845
846  if (StringRef(DstGV->getSection()) != SrcGV->getSection())
847    return emitError(
848          "Appending variables with different section name need to be linked!");
849
850  uint64_t NewSize = DstTy->getNumElements() + SrcTy->getNumElements();
851  ArrayType *NewType = ArrayType::get(EltTy, NewSize);
852
853  // Create the new global variable.
854  GlobalVariable *NG =
855    new GlobalVariable(*DstGV->getParent(), NewType, SrcGV->isConstant(),
856                       DstGV->getLinkage(), /*init*/nullptr, /*name*/"", DstGV,
857                       DstGV->getThreadLocalMode(),
858                       DstGV->getType()->getAddressSpace());
859
860  // Propagate alignment, visibility and section info.
861  copyGVAttributes(NG, DstGV);
862
863  AppendingVarInfo AVI;
864  AVI.NewGV = NG;
865  AVI.DstInit = DstGV->getInitializer();
866  AVI.SrcInit = SrcGV->getInitializer();
867  AppendingVars.push_back(AVI);
868
869  // Replace any uses of the two global variables with uses of the new
870  // global.
871  ValueMap[SrcGV] = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType()));
872
873  DstGV->replaceAllUsesWith(ConstantExpr::getBitCast(NG, DstGV->getType()));
874  DstGV->eraseFromParent();
875
876  // Track the source variable so we don't try to link it.
877  DoNotLinkFromSource.insert(SrcGV);
878
879  return false;
880}
881
882/// linkGlobalProto - Loop through the global variables in the src module and
883/// merge them into the dest module.
884bool ModuleLinker::linkGlobalProto(GlobalVariable *SGV) {
885  GlobalValue *DGV = getLinkedToGlobal(SGV);
886  llvm::Optional<GlobalValue::VisibilityTypes> NewVisibility;
887  bool HasUnnamedAddr = SGV->hasUnnamedAddr();
888
889  bool LinkFromSrc = false;
890  Comdat *DC = nullptr;
891  if (const Comdat *SC = SGV->getComdat()) {
892    Comdat::SelectionKind SK;
893    std::tie(SK, LinkFromSrc) = ComdatsChosen[SC];
894    DC = DstM->getOrInsertComdat(SC->getName());
895    DC->setSelectionKind(SK);
896  }
897
898  if (DGV) {
899    if (!DC) {
900      // Concatenation of appending linkage variables is magic and handled later.
901      if (DGV->hasAppendingLinkage() || SGV->hasAppendingLinkage())
902        return linkAppendingVarProto(cast<GlobalVariable>(DGV), SGV);
903
904      // Determine whether linkage of these two globals follows the source
905      // module's definition or the destination module's definition.
906      GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
907      GlobalValue::VisibilityTypes NV;
908      if (getLinkageResult(DGV, SGV, NewLinkage, NV, LinkFromSrc))
909        return true;
910      NewVisibility = NV;
911      HasUnnamedAddr = HasUnnamedAddr && DGV->hasUnnamedAddr();
912
913      // If we're not linking from the source, then keep the definition that we
914      // have.
915      if (!LinkFromSrc) {
916        // Special case for const propagation.
917        if (GlobalVariable *DGVar = dyn_cast<GlobalVariable>(DGV))
918          if (DGVar->isDeclaration() && SGV->isConstant() &&
919              !DGVar->isConstant())
920            DGVar->setConstant(true);
921
922        // Set calculated linkage, visibility and unnamed_addr.
923        DGV->setLinkage(NewLinkage);
924        DGV->setVisibility(*NewVisibility);
925        DGV->setUnnamedAddr(HasUnnamedAddr);
926      }
927    }
928
929    if (!LinkFromSrc) {
930      // Make sure to remember this mapping.
931      ValueMap[SGV] = ConstantExpr::getBitCast(DGV,TypeMap.get(SGV->getType()));
932
933      // Track the source global so that we don't attempt to copy it over when
934      // processing global initializers.
935      DoNotLinkFromSource.insert(SGV);
936
937      return false;
938    }
939  }
940
941  // If the Comdat this variable was inside of wasn't selected, skip it.
942  if (DC && !DGV && !LinkFromSrc) {
943    DoNotLinkFromSource.insert(SGV);
944    return false;
945  }
946
947  // No linking to be performed or linking from the source: simply create an
948  // identical version of the symbol over in the dest module... the
949  // initializer will be filled in later by LinkGlobalInits.
950  GlobalVariable *NewDGV =
951    new GlobalVariable(*DstM, TypeMap.get(SGV->getType()->getElementType()),
952                       SGV->isConstant(), SGV->getLinkage(), /*init*/nullptr,
953                       SGV->getName(), /*insertbefore*/nullptr,
954                       SGV->getThreadLocalMode(),
955                       SGV->getType()->getAddressSpace());
956  // Propagate alignment, visibility and section info.
957  copyGVAttributes(NewDGV, SGV);
958  if (NewVisibility)
959    NewDGV->setVisibility(*NewVisibility);
960  NewDGV->setUnnamedAddr(HasUnnamedAddr);
961
962  if (DC)
963    NewDGV->setComdat(DC);
964
965  if (DGV) {
966    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDGV, DGV->getType()));
967    DGV->eraseFromParent();
968  }
969
970  // Make sure to remember this mapping.
971  ValueMap[SGV] = NewDGV;
972  return false;
973}
974
975/// linkFunctionProto - Link the function in the source module into the
976/// destination module if needed, setting up mapping information.
977bool ModuleLinker::linkFunctionProto(Function *SF) {
978  GlobalValue *DGV = getLinkedToGlobal(SF);
979  llvm::Optional<GlobalValue::VisibilityTypes> NewVisibility;
980  bool HasUnnamedAddr = SF->hasUnnamedAddr();
981
982  bool LinkFromSrc = false;
983  Comdat *DC = nullptr;
984  if (const Comdat *SC = SF->getComdat()) {
985    Comdat::SelectionKind SK;
986    std::tie(SK, LinkFromSrc) = ComdatsChosen[SC];
987    DC = DstM->getOrInsertComdat(SC->getName());
988    DC->setSelectionKind(SK);
989  }
990
991  if (DGV) {
992    if (!DC) {
993      GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
994      GlobalValue::VisibilityTypes NV;
995      if (getLinkageResult(DGV, SF, NewLinkage, NV, LinkFromSrc))
996        return true;
997      NewVisibility = NV;
998      HasUnnamedAddr = HasUnnamedAddr && DGV->hasUnnamedAddr();
999
1000      if (!LinkFromSrc) {
1001        // Set calculated linkage
1002        DGV->setLinkage(NewLinkage);
1003        DGV->setVisibility(*NewVisibility);
1004        DGV->setUnnamedAddr(HasUnnamedAddr);
1005      }
1006    }
1007
1008    if (!LinkFromSrc) {
1009      // Make sure to remember this mapping.
1010      ValueMap[SF] = ConstantExpr::getBitCast(DGV, TypeMap.get(SF->getType()));
1011
1012      // Track the function from the source module so we don't attempt to remap
1013      // it.
1014      DoNotLinkFromSource.insert(SF);
1015
1016      return false;
1017    }
1018  }
1019
1020  // If the function is to be lazily linked, don't create it just yet.
1021  // The ValueMaterializerTy will deal with creating it if it's used.
1022  if (!DGV && (SF->hasLocalLinkage() || SF->hasLinkOnceLinkage() ||
1023               SF->hasAvailableExternallyLinkage())) {
1024    DoNotLinkFromSource.insert(SF);
1025    return false;
1026  }
1027
1028  // If the Comdat this function was inside of wasn't selected, skip it.
1029  if (DC && !DGV && !LinkFromSrc) {
1030    DoNotLinkFromSource.insert(SF);
1031    return false;
1032  }
1033
1034  // If there is no linkage to be performed or we are linking from the source,
1035  // bring SF over.
1036  Function *NewDF = Function::Create(TypeMap.get(SF->getFunctionType()),
1037                                     SF->getLinkage(), SF->getName(), DstM);
1038  copyGVAttributes(NewDF, SF);
1039  if (NewVisibility)
1040    NewDF->setVisibility(*NewVisibility);
1041  NewDF->setUnnamedAddr(HasUnnamedAddr);
1042
1043  if (DC)
1044    NewDF->setComdat(DC);
1045
1046  if (DGV) {
1047    // Any uses of DF need to change to NewDF, with cast.
1048    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDF, DGV->getType()));
1049    DGV->eraseFromParent();
1050  }
1051
1052  ValueMap[SF] = NewDF;
1053  return false;
1054}
1055
1056/// LinkAliasProto - Set up prototypes for any aliases that come over from the
1057/// source module.
1058bool ModuleLinker::linkAliasProto(GlobalAlias *SGA) {
1059  GlobalValue *DGV = getLinkedToGlobal(SGA);
1060  llvm::Optional<GlobalValue::VisibilityTypes> NewVisibility;
1061  bool HasUnnamedAddr = SGA->hasUnnamedAddr();
1062
1063  bool LinkFromSrc = false;
1064  Comdat *DC = nullptr;
1065  if (const Comdat *SC = SGA->getComdat()) {
1066    Comdat::SelectionKind SK;
1067    std::tie(SK, LinkFromSrc) = ComdatsChosen[SC];
1068    DC = DstM->getOrInsertComdat(SC->getName());
1069    DC->setSelectionKind(SK);
1070  }
1071
1072  if (DGV) {
1073    if (!DC) {
1074      GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
1075      GlobalValue::VisibilityTypes NV;
1076      if (getLinkageResult(DGV, SGA, NewLinkage, NV, LinkFromSrc))
1077        return true;
1078      NewVisibility = NV;
1079      HasUnnamedAddr = HasUnnamedAddr && DGV->hasUnnamedAddr();
1080
1081      if (!LinkFromSrc) {
1082        // Set calculated linkage.
1083        DGV->setLinkage(NewLinkage);
1084        DGV->setVisibility(*NewVisibility);
1085        DGV->setUnnamedAddr(HasUnnamedAddr);
1086      }
1087    }
1088
1089    if (!LinkFromSrc) {
1090      // Make sure to remember this mapping.
1091      ValueMap[SGA] = ConstantExpr::getBitCast(DGV,TypeMap.get(SGA->getType()));
1092
1093      // Track the alias from the source module so we don't attempt to remap it.
1094      DoNotLinkFromSource.insert(SGA);
1095
1096      return false;
1097    }
1098  }
1099
1100  // If the Comdat this alias was inside of wasn't selected, skip it.
1101  if (DC && !DGV && !LinkFromSrc) {
1102    DoNotLinkFromSource.insert(SGA);
1103    return false;
1104  }
1105
1106  // If there is no linkage to be performed or we're linking from the source,
1107  // bring over SGA.
1108  auto *PTy = cast<PointerType>(TypeMap.get(SGA->getType()));
1109  auto *NewDA =
1110      GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
1111                          SGA->getLinkage(), SGA->getName(), DstM);
1112  copyGVAttributes(NewDA, SGA);
1113  if (NewVisibility)
1114    NewDA->setVisibility(*NewVisibility);
1115  NewDA->setUnnamedAddr(HasUnnamedAddr);
1116
1117  if (DGV) {
1118    // Any uses of DGV need to change to NewDA, with cast.
1119    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDA, DGV->getType()));
1120    DGV->eraseFromParent();
1121  }
1122
1123  ValueMap[SGA] = NewDA;
1124  return false;
1125}
1126
1127static void getArrayElements(Constant *C, SmallVectorImpl<Constant*> &Dest) {
1128  unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements();
1129
1130  for (unsigned i = 0; i != NumElements; ++i)
1131    Dest.push_back(C->getAggregateElement(i));
1132}
1133
1134void ModuleLinker::linkAppendingVarInit(const AppendingVarInfo &AVI) {
1135  // Merge the initializer.
1136  SmallVector<Constant*, 16> Elements;
1137  getArrayElements(AVI.DstInit, Elements);
1138
1139  Constant *SrcInit = MapValue(AVI.SrcInit, ValueMap, RF_None, &TypeMap, &ValMaterializer);
1140  getArrayElements(SrcInit, Elements);
1141
1142  ArrayType *NewType = cast<ArrayType>(AVI.NewGV->getType()->getElementType());
1143  AVI.NewGV->setInitializer(ConstantArray::get(NewType, Elements));
1144}
1145
1146/// linkGlobalInits - Update the initializers in the Dest module now that all
1147/// globals that may be referenced are in Dest.
1148void ModuleLinker::linkGlobalInits() {
1149  // Loop over all of the globals in the src module, mapping them over as we go
1150  for (Module::const_global_iterator I = SrcM->global_begin(),
1151       E = SrcM->global_end(); I != E; ++I) {
1152
1153    // Only process initialized GV's or ones not already in dest.
1154    if (!I->hasInitializer() || DoNotLinkFromSource.count(I)) continue;
1155
1156    // Grab destination global variable.
1157    GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[I]);
1158    // Figure out what the initializer looks like in the dest module.
1159    DGV->setInitializer(MapValue(I->getInitializer(), ValueMap,
1160                                 RF_None, &TypeMap, &ValMaterializer));
1161  }
1162}
1163
1164/// linkFunctionBody - Copy the source function over into the dest function and
1165/// fix up references to values.  At this point we know that Dest is an external
1166/// function, and that Src is not.
1167void ModuleLinker::linkFunctionBody(Function *Dst, Function *Src) {
1168  assert(Src && Dst && Dst->isDeclaration() && !Src->isDeclaration());
1169
1170  // Go through and convert function arguments over, remembering the mapping.
1171  Function::arg_iterator DI = Dst->arg_begin();
1172  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
1173       I != E; ++I, ++DI) {
1174    DI->setName(I->getName());  // Copy the name over.
1175
1176    // Add a mapping to our mapping.
1177    ValueMap[I] = DI;
1178  }
1179
1180  if (Mode == Linker::DestroySource) {
1181    // Splice the body of the source function into the dest function.
1182    Dst->getBasicBlockList().splice(Dst->end(), Src->getBasicBlockList());
1183
1184    // At this point, all of the instructions and values of the function are now
1185    // copied over.  The only problem is that they are still referencing values in
1186    // the Source function as operands.  Loop through all of the operands of the
1187    // functions and patch them up to point to the local versions.
1188    for (Function::iterator BB = Dst->begin(), BE = Dst->end(); BB != BE; ++BB)
1189      for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1190        RemapInstruction(I, ValueMap, RF_IgnoreMissingEntries,
1191                         &TypeMap, &ValMaterializer);
1192
1193  } else {
1194    // Clone the body of the function into the dest function.
1195    SmallVector<ReturnInst*, 8> Returns; // Ignore returns.
1196    CloneFunctionInto(Dst, Src, ValueMap, false, Returns, "", nullptr,
1197                      &TypeMap, &ValMaterializer);
1198  }
1199
1200  // There is no need to map the arguments anymore.
1201  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
1202       I != E; ++I)
1203    ValueMap.erase(I);
1204
1205}
1206
1207/// linkAliasBodies - Insert all of the aliases in Src into the Dest module.
1208void ModuleLinker::linkAliasBodies() {
1209  for (Module::alias_iterator I = SrcM->alias_begin(), E = SrcM->alias_end();
1210       I != E; ++I) {
1211    if (DoNotLinkFromSource.count(I))
1212      continue;
1213    if (Constant *Aliasee = I->getAliasee()) {
1214      GlobalAlias *DA = cast<GlobalAlias>(ValueMap[I]);
1215      Constant *Val =
1216          MapValue(Aliasee, ValueMap, RF_None, &TypeMap, &ValMaterializer);
1217      DA->setAliasee(Val);
1218    }
1219  }
1220}
1221
1222/// linkNamedMDNodes - Insert all of the named MDNodes in Src into the Dest
1223/// module.
1224void ModuleLinker::linkNamedMDNodes() {
1225  const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1226  for (Module::const_named_metadata_iterator I = SrcM->named_metadata_begin(),
1227       E = SrcM->named_metadata_end(); I != E; ++I) {
1228    // Don't link module flags here. Do them separately.
1229    if (&*I == SrcModFlags) continue;
1230    NamedMDNode *DestNMD = DstM->getOrInsertNamedMetadata(I->getName());
1231    // Add Src elements into Dest node.
1232    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1233      DestNMD->addOperand(MapValue(I->getOperand(i), ValueMap,
1234                                   RF_None, &TypeMap, &ValMaterializer));
1235  }
1236}
1237
1238/// linkModuleFlagsMetadata - Merge the linker flags in Src into the Dest
1239/// module.
1240bool ModuleLinker::linkModuleFlagsMetadata() {
1241  // If the source module has no module flags, we are done.
1242  const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1243  if (!SrcModFlags) return false;
1244
1245  // If the destination module doesn't have module flags yet, then just copy
1246  // over the source module's flags.
1247  NamedMDNode *DstModFlags = DstM->getOrInsertModuleFlagsMetadata();
1248  if (DstModFlags->getNumOperands() == 0) {
1249    for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I)
1250      DstModFlags->addOperand(SrcModFlags->getOperand(I));
1251
1252    return false;
1253  }
1254
1255  // First build a map of the existing module flags and requirements.
1256  DenseMap<MDString*, MDNode*> Flags;
1257  SmallSetVector<MDNode*, 16> Requirements;
1258  for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) {
1259    MDNode *Op = DstModFlags->getOperand(I);
1260    ConstantInt *Behavior = cast<ConstantInt>(Op->getOperand(0));
1261    MDString *ID = cast<MDString>(Op->getOperand(1));
1262
1263    if (Behavior->getZExtValue() == Module::Require) {
1264      Requirements.insert(cast<MDNode>(Op->getOperand(2)));
1265    } else {
1266      Flags[ID] = Op;
1267    }
1268  }
1269
1270  // Merge in the flags from the source module, and also collect its set of
1271  // requirements.
1272  bool HasErr = false;
1273  for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) {
1274    MDNode *SrcOp = SrcModFlags->getOperand(I);
1275    ConstantInt *SrcBehavior = cast<ConstantInt>(SrcOp->getOperand(0));
1276    MDString *ID = cast<MDString>(SrcOp->getOperand(1));
1277    MDNode *DstOp = Flags.lookup(ID);
1278    unsigned SrcBehaviorValue = SrcBehavior->getZExtValue();
1279
1280    // If this is a requirement, add it and continue.
1281    if (SrcBehaviorValue == Module::Require) {
1282      // If the destination module does not already have this requirement, add
1283      // it.
1284      if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) {
1285        DstModFlags->addOperand(SrcOp);
1286      }
1287      continue;
1288    }
1289
1290    // If there is no existing flag with this ID, just add it.
1291    if (!DstOp) {
1292      Flags[ID] = SrcOp;
1293      DstModFlags->addOperand(SrcOp);
1294      continue;
1295    }
1296
1297    // Otherwise, perform a merge.
1298    ConstantInt *DstBehavior = cast<ConstantInt>(DstOp->getOperand(0));
1299    unsigned DstBehaviorValue = DstBehavior->getZExtValue();
1300
1301    // If either flag has override behavior, handle it first.
1302    if (DstBehaviorValue == Module::Override) {
1303      // Diagnose inconsistent flags which both have override behavior.
1304      if (SrcBehaviorValue == Module::Override &&
1305          SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1306        HasErr |= emitError("linking module flags '" + ID->getString() +
1307                            "': IDs have conflicting override values");
1308      }
1309      continue;
1310    } else if (SrcBehaviorValue == Module::Override) {
1311      // Update the destination flag to that of the source.
1312      DstOp->replaceOperandWith(0, SrcBehavior);
1313      DstOp->replaceOperandWith(2, SrcOp->getOperand(2));
1314      continue;
1315    }
1316
1317    // Diagnose inconsistent merge behavior types.
1318    if (SrcBehaviorValue != DstBehaviorValue) {
1319      HasErr |= emitError("linking module flags '" + ID->getString() +
1320                          "': IDs have conflicting behaviors");
1321      continue;
1322    }
1323
1324    // Perform the merge for standard behavior types.
1325    switch (SrcBehaviorValue) {
1326    case Module::Require:
1327    case Module::Override: llvm_unreachable("not possible");
1328    case Module::Error: {
1329      // Emit an error if the values differ.
1330      if (SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1331        HasErr |= emitError("linking module flags '" + ID->getString() +
1332                            "': IDs have conflicting values");
1333      }
1334      continue;
1335    }
1336    case Module::Warning: {
1337      // Emit a warning if the values differ.
1338      if (SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1339        if (!SuppressWarnings) {
1340          errs() << "WARNING: linking module flags '" << ID->getString()
1341                 << "': IDs have conflicting values";
1342        }
1343      }
1344      continue;
1345    }
1346    case Module::Append: {
1347      MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1348      MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1349      unsigned NumOps = DstValue->getNumOperands() + SrcValue->getNumOperands();
1350      Value **VP, **Values = VP = new Value*[NumOps];
1351      for (unsigned i = 0, e = DstValue->getNumOperands(); i != e; ++i, ++VP)
1352        *VP = DstValue->getOperand(i);
1353      for (unsigned i = 0, e = SrcValue->getNumOperands(); i != e; ++i, ++VP)
1354        *VP = SrcValue->getOperand(i);
1355      DstOp->replaceOperandWith(2, MDNode::get(DstM->getContext(),
1356                                               ArrayRef<Value*>(Values,
1357                                                                NumOps)));
1358      delete[] Values;
1359      break;
1360    }
1361    case Module::AppendUnique: {
1362      SmallSetVector<Value*, 16> Elts;
1363      MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1364      MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1365      for (unsigned i = 0, e = DstValue->getNumOperands(); i != e; ++i)
1366        Elts.insert(DstValue->getOperand(i));
1367      for (unsigned i = 0, e = SrcValue->getNumOperands(); i != e; ++i)
1368        Elts.insert(SrcValue->getOperand(i));
1369      DstOp->replaceOperandWith(2, MDNode::get(DstM->getContext(),
1370                                               ArrayRef<Value*>(Elts.begin(),
1371                                                                Elts.end())));
1372      break;
1373    }
1374    }
1375  }
1376
1377  // Check all of the requirements.
1378  for (unsigned I = 0, E = Requirements.size(); I != E; ++I) {
1379    MDNode *Requirement = Requirements[I];
1380    MDString *Flag = cast<MDString>(Requirement->getOperand(0));
1381    Value *ReqValue = Requirement->getOperand(1);
1382
1383    MDNode *Op = Flags[Flag];
1384    if (!Op || Op->getOperand(2) != ReqValue) {
1385      HasErr |= emitError("linking module flags '" + Flag->getString() +
1386                          "': does not have the required value");
1387      continue;
1388    }
1389  }
1390
1391  return HasErr;
1392}
1393
1394bool ModuleLinker::run() {
1395  assert(DstM && "Null destination module");
1396  assert(SrcM && "Null source module");
1397
1398  // Inherit the target data from the source module if the destination module
1399  // doesn't have one already.
1400  if (!DstM->getDataLayout() && SrcM->getDataLayout())
1401    DstM->setDataLayout(SrcM->getDataLayout());
1402
1403  // Copy the target triple from the source to dest if the dest's is empty.
1404  if (DstM->getTargetTriple().empty() && !SrcM->getTargetTriple().empty())
1405    DstM->setTargetTriple(SrcM->getTargetTriple());
1406
1407  if (SrcM->getDataLayout() && DstM->getDataLayout() &&
1408      *SrcM->getDataLayout() != *DstM->getDataLayout()) {
1409    if (!SuppressWarnings) {
1410      errs() << "WARNING: Linking two modules of different data layouts: '"
1411             << SrcM->getModuleIdentifier() << "' is '"
1412             << SrcM->getDataLayoutStr() << "' whereas '"
1413             << DstM->getModuleIdentifier() << "' is '"
1414             << DstM->getDataLayoutStr() << "'\n";
1415    }
1416  }
1417  if (!SrcM->getTargetTriple().empty() &&
1418      DstM->getTargetTriple() != SrcM->getTargetTriple()) {
1419    if (!SuppressWarnings) {
1420      errs() << "WARNING: Linking two modules of different target triples: "
1421             << SrcM->getModuleIdentifier() << "' is '"
1422             << SrcM->getTargetTriple() << "' whereas '"
1423             << DstM->getModuleIdentifier() << "' is '"
1424             << DstM->getTargetTriple() << "'\n";
1425    }
1426  }
1427
1428  // Append the module inline asm string.
1429  if (!SrcM->getModuleInlineAsm().empty()) {
1430    if (DstM->getModuleInlineAsm().empty())
1431      DstM->setModuleInlineAsm(SrcM->getModuleInlineAsm());
1432    else
1433      DstM->setModuleInlineAsm(DstM->getModuleInlineAsm()+"\n"+
1434                               SrcM->getModuleInlineAsm());
1435  }
1436
1437  // Loop over all of the linked values to compute type mappings.
1438  computeTypeMapping();
1439
1440  ComdatsChosen.clear();
1441  for (const StringMapEntry<llvm::Comdat> &SMEC : SrcM->getComdatSymbolTable()) {
1442    const Comdat &C = SMEC.getValue();
1443    if (ComdatsChosen.count(&C))
1444      continue;
1445    Comdat::SelectionKind SK;
1446    bool LinkFromSrc;
1447    if (getComdatResult(&C, SK, LinkFromSrc))
1448      return true;
1449    ComdatsChosen[&C] = std::make_pair(SK, LinkFromSrc);
1450  }
1451
1452  // Insert all of the globals in src into the DstM module... without linking
1453  // initializers (which could refer to functions not yet mapped over).
1454  for (Module::global_iterator I = SrcM->global_begin(),
1455       E = SrcM->global_end(); I != E; ++I)
1456    if (linkGlobalProto(I))
1457      return true;
1458
1459  // Link the functions together between the two modules, without doing function
1460  // bodies... this just adds external function prototypes to the DstM
1461  // function...  We do this so that when we begin processing function bodies,
1462  // all of the global values that may be referenced are available in our
1463  // ValueMap.
1464  for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I)
1465    if (linkFunctionProto(I))
1466      return true;
1467
1468  // If there were any aliases, link them now.
1469  for (Module::alias_iterator I = SrcM->alias_begin(),
1470       E = SrcM->alias_end(); I != E; ++I)
1471    if (linkAliasProto(I))
1472      return true;
1473
1474  for (unsigned i = 0, e = AppendingVars.size(); i != e; ++i)
1475    linkAppendingVarInit(AppendingVars[i]);
1476
1477  // Link in the function bodies that are defined in the source module into
1478  // DstM.
1479  for (Module::iterator SF = SrcM->begin(), E = SrcM->end(); SF != E; ++SF) {
1480    // Skip if not linking from source.
1481    if (DoNotLinkFromSource.count(SF)) continue;
1482
1483    Function *DF = cast<Function>(ValueMap[SF]);
1484    if (SF->hasPrefixData()) {
1485      // Link in the prefix data.
1486      DF->setPrefixData(MapValue(
1487          SF->getPrefixData(), ValueMap, RF_None, &TypeMap, &ValMaterializer));
1488    }
1489
1490    // Skip if no body (function is external) or materialize.
1491    if (SF->isDeclaration()) {
1492      if (!SF->isMaterializable())
1493        continue;
1494      if (SF->Materialize(&ErrorMsg))
1495        return true;
1496    }
1497
1498    linkFunctionBody(DF, SF);
1499    SF->Dematerialize();
1500  }
1501
1502  // Resolve all uses of aliases with aliasees.
1503  linkAliasBodies();
1504
1505  // Remap all of the named MDNodes in Src into the DstM module. We do this
1506  // after linking GlobalValues so that MDNodes that reference GlobalValues
1507  // are properly remapped.
1508  linkNamedMDNodes();
1509
1510  // Merge the module flags into the DstM module.
1511  if (linkModuleFlagsMetadata())
1512    return true;
1513
1514  // Update the initializers in the DstM module now that all globals that may
1515  // be referenced are in DstM.
1516  linkGlobalInits();
1517
1518  // Process vector of lazily linked in functions.
1519  bool LinkedInAnyFunctions;
1520  do {
1521    LinkedInAnyFunctions = false;
1522
1523    for(std::vector<Function*>::iterator I = LazilyLinkFunctions.begin(),
1524        E = LazilyLinkFunctions.end(); I != E; ++I) {
1525      Function *SF = *I;
1526      if (!SF)
1527        continue;
1528
1529      Function *DF = cast<Function>(ValueMap[SF]);
1530      if (SF->hasPrefixData()) {
1531        // Link in the prefix data.
1532        DF->setPrefixData(MapValue(SF->getPrefixData(),
1533                                   ValueMap,
1534                                   RF_None,
1535                                   &TypeMap,
1536                                   &ValMaterializer));
1537      }
1538
1539      // Materialize if necessary.
1540      if (SF->isDeclaration()) {
1541        if (!SF->isMaterializable())
1542          continue;
1543        if (SF->Materialize(&ErrorMsg))
1544          return true;
1545      }
1546
1547      // Erase from vector *before* the function body is linked - linkFunctionBody could
1548      // invalidate I.
1549      LazilyLinkFunctions.erase(I);
1550
1551      // Link in function body.
1552      linkFunctionBody(DF, SF);
1553      SF->Dematerialize();
1554
1555      // Set flag to indicate we may have more functions to lazily link in
1556      // since we linked in a function.
1557      LinkedInAnyFunctions = true;
1558      break;
1559    }
1560  } while (LinkedInAnyFunctions);
1561
1562  // Now that all of the types from the source are used, resolve any structs
1563  // copied over to the dest that didn't exist there.
1564  TypeMap.linkDefinedTypeBodies();
1565
1566  return false;
1567}
1568
1569Linker::Linker(Module *M, bool SuppressWarnings)
1570    : Composite(M), SuppressWarnings(SuppressWarnings) {
1571  TypeFinder StructTypes;
1572  StructTypes.run(*M, true);
1573  IdentifiedStructTypes.insert(StructTypes.begin(), StructTypes.end());
1574}
1575
1576Linker::~Linker() {
1577}
1578
1579void Linker::deleteModule() {
1580  delete Composite;
1581  Composite = nullptr;
1582}
1583
1584bool Linker::linkInModule(Module *Src, unsigned Mode, std::string *ErrorMsg) {
1585  ModuleLinker TheLinker(Composite, IdentifiedStructTypes, Src, Mode,
1586                         SuppressWarnings);
1587  if (TheLinker.run()) {
1588    if (ErrorMsg)
1589      *ErrorMsg = TheLinker.ErrorMsg;
1590    return true;
1591  }
1592  return false;
1593}
1594
1595//===----------------------------------------------------------------------===//
1596// LinkModules entrypoint.
1597//===----------------------------------------------------------------------===//
1598
1599/// LinkModules - This function links two modules together, with the resulting
1600/// Dest module modified to be the composite of the two input modules.  If an
1601/// error occurs, true is returned and ErrorMsg (if not null) is set to indicate
1602/// the problem.  Upon failure, the Dest module could be in a modified state,
1603/// and shouldn't be relied on to be consistent.
1604bool Linker::LinkModules(Module *Dest, Module *Src, unsigned Mode,
1605                         std::string *ErrorMsg) {
1606  Linker L(Dest);
1607  return L.linkInModule(Src, Mode, ErrorMsg);
1608}
1609
1610//===----------------------------------------------------------------------===//
1611// C API.
1612//===----------------------------------------------------------------------===//
1613
1614LLVMBool LLVMLinkModules(LLVMModuleRef Dest, LLVMModuleRef Src,
1615                         LLVMLinkerMode Mode, char **OutMessages) {
1616  std::string Messages;
1617  LLVMBool Result = Linker::LinkModules(unwrap(Dest), unwrap(Src),
1618                                        Mode, OutMessages? &Messages : nullptr);
1619  if (OutMessages)
1620    *OutMessages = strdup(Messages.c_str());
1621  return Result;
1622}
1623