LinkModules.cpp revision 1a31f3b90c012b067f8509546e1e037051e6482d
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.h"
15#include "llvm/Constants.h"
16#include "llvm/DerivedTypes.h"
17#include "llvm/Instructions.h"
18#include "llvm/Module.h"
19#include "llvm/ADT/SmallPtrSet.h"
20#include "llvm/Support/raw_ostream.h"
21#include "llvm/Support/Path.h"
22#include "llvm/Transforms/Utils/Cloning.h"
23#include "llvm/Transforms/Utils/ValueMapper.h"
24using namespace llvm;
25
26//===----------------------------------------------------------------------===//
27// TypeMap implementation.
28//===----------------------------------------------------------------------===//
29
30namespace {
31class TypeMapTy : public ValueMapTypeRemapper {
32  /// MappedTypes - This is a mapping from a source type to a destination type
33  /// to use.
34  DenseMap<Type*, Type*> MappedTypes;
35
36  /// SpeculativeTypes - When checking to see if two subgraphs are isomorphic,
37  /// we speculatively add types to MappedTypes, but keep track of them here in
38  /// case we need to roll back.
39  SmallVector<Type*, 16> SpeculativeTypes;
40
41  /// SrcDefinitionsToResolve - This is a list of non-opaque structs in the
42  /// source module that are mapped to an opaque struct in the destination
43  /// module.
44  SmallVector<StructType*, 16> SrcDefinitionsToResolve;
45
46  /// DstResolvedOpaqueTypes - This is the set of opaque types in the
47  /// destination modules who are getting a body from the source module.
48  SmallPtrSet<StructType*, 16> DstResolvedOpaqueTypes;
49public:
50
51  /// addTypeMapping - Indicate that the specified type in the destination
52  /// module is conceptually equivalent to the specified type in the source
53  /// module.
54  void addTypeMapping(Type *DstTy, Type *SrcTy);
55
56  /// linkDefinedTypeBodies - Produce a body for an opaque type in the dest
57  /// module from a type definition in the source module.
58  void linkDefinedTypeBodies();
59
60  /// get - Return the mapped type to use for the specified input type from the
61  /// source module.
62  Type *get(Type *SrcTy);
63
64  FunctionType *get(FunctionType *T) {return cast<FunctionType>(get((Type*)T));}
65
66private:
67  Type *getImpl(Type *T);
68  /// remapType - Implement the ValueMapTypeRemapper interface.
69  Type *remapType(Type *SrcTy) {
70    return get(SrcTy);
71  }
72
73  bool areTypesIsomorphic(Type *DstTy, Type *SrcTy);
74};
75}
76
77void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) {
78  Type *&Entry = MappedTypes[SrcTy];
79  if (Entry) return;
80
81  if (DstTy == SrcTy) {
82    Entry = DstTy;
83    return;
84  }
85
86  // Check to see if these types are recursively isomorphic and establish a
87  // mapping between them if so.
88  if (!areTypesIsomorphic(DstTy, SrcTy)) {
89    // Oops, they aren't isomorphic.  Just discard this request by rolling out
90    // any speculative mappings we've established.
91    for (unsigned i = 0, e = SpeculativeTypes.size(); i != e; ++i)
92      MappedTypes.erase(SpeculativeTypes[i]);
93  }
94  SpeculativeTypes.clear();
95}
96
97/// areTypesIsomorphic - Recursively walk this pair of types, returning true
98/// if they are isomorphic, false if they are not.
99bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) {
100  // Two types with differing kinds are clearly not isomorphic.
101  if (DstTy->getTypeID() != SrcTy->getTypeID()) return false;
102
103  // If we have an entry in the MappedTypes table, then we have our answer.
104  Type *&Entry = MappedTypes[SrcTy];
105  if (Entry)
106    return Entry == DstTy;
107
108  // Two identical types are clearly isomorphic.  Remember this
109  // non-speculatively.
110  if (DstTy == SrcTy) {
111    Entry = DstTy;
112    return true;
113  }
114
115  // Okay, we have two types with identical kinds that we haven't seen before.
116
117  // If this is an opaque struct type, special case it.
118  if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) {
119    // Mapping an opaque type to any struct, just keep the dest struct.
120    if (SSTy->isOpaque()) {
121      Entry = DstTy;
122      SpeculativeTypes.push_back(SrcTy);
123      return true;
124    }
125
126    // Mapping a non-opaque source type to an opaque dest.  If this is the first
127    // type that we're mapping onto this destination type then we succeed.  Keep
128    // the dest, but fill it in later.  This doesn't need to be speculative.  If
129    // this is the second (different) type that we're trying to map onto the
130    // same opaque type then we fail.
131    if (cast<StructType>(DstTy)->isOpaque()) {
132      // We can only map one source type onto the opaque destination type.
133      if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)))
134        return false;
135      SrcDefinitionsToResolve.push_back(SSTy);
136      Entry = DstTy;
137      return true;
138    }
139  }
140
141  // If the number of subtypes disagree between the two types, then we fail.
142  if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes())
143    return false;
144
145  // Fail if any of the extra properties (e.g. array size) of the type disagree.
146  if (isa<IntegerType>(DstTy))
147    return false;  // bitwidth disagrees.
148  if (PointerType *PT = dyn_cast<PointerType>(DstTy)) {
149    if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace())
150      return false;
151
152  } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) {
153    if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg())
154      return false;
155  } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) {
156    StructType *SSTy = cast<StructType>(SrcTy);
157    if (DSTy->isLiteral() != SSTy->isLiteral() ||
158        DSTy->isPacked() != SSTy->isPacked())
159      return false;
160  } else if (ArrayType *DATy = dyn_cast<ArrayType>(DstTy)) {
161    if (DATy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
162      return false;
163  } else if (VectorType *DVTy = dyn_cast<VectorType>(DstTy)) {
164    if (DVTy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
165      return false;
166  }
167
168  // Otherwise, we speculate that these two types will line up and recursively
169  // check the subelements.
170  Entry = DstTy;
171  SpeculativeTypes.push_back(SrcTy);
172
173  for (unsigned i = 0, e = SrcTy->getNumContainedTypes(); i != e; ++i)
174    if (!areTypesIsomorphic(DstTy->getContainedType(i),
175                            SrcTy->getContainedType(i)))
176      return false;
177
178  // If everything seems to have lined up, then everything is great.
179  return true;
180}
181
182/// linkDefinedTypeBodies - Produce a body for an opaque type in the dest
183/// module from a type definition in the source module.
184void TypeMapTy::linkDefinedTypeBodies() {
185  SmallVector<Type*, 16> Elements;
186  SmallString<16> TmpName;
187
188  // Note that processing entries in this loop (calling 'get') can add new
189  // entries to the SrcDefinitionsToResolve vector.
190  while (!SrcDefinitionsToResolve.empty()) {
191    StructType *SrcSTy = SrcDefinitionsToResolve.pop_back_val();
192    StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]);
193
194    // TypeMap is a many-to-one mapping, if there were multiple types that
195    // provide a body for DstSTy then previous iterations of this loop may have
196    // already handled it.  Just ignore this case.
197    if (!DstSTy->isOpaque()) continue;
198    assert(!SrcSTy->isOpaque() && "Not resolving a definition?");
199
200    // Map the body of the source type over to a new body for the dest type.
201    Elements.resize(SrcSTy->getNumElements());
202    for (unsigned i = 0, e = Elements.size(); i != e; ++i)
203      Elements[i] = getImpl(SrcSTy->getElementType(i));
204
205    DstSTy->setBody(Elements, SrcSTy->isPacked());
206
207    // If DstSTy has no name or has a longer name than STy, then viciously steal
208    // STy's name.
209    if (!SrcSTy->hasName()) continue;
210    StringRef SrcName = SrcSTy->getName();
211
212    if (!DstSTy->hasName() || DstSTy->getName().size() > SrcName.size()) {
213      TmpName.insert(TmpName.end(), SrcName.begin(), SrcName.end());
214      SrcSTy->setName("");
215      DstSTy->setName(TmpName.str());
216      TmpName.clear();
217    }
218  }
219
220  DstResolvedOpaqueTypes.clear();
221}
222
223
224/// get - Return the mapped type to use for the specified input type from the
225/// source module.
226Type *TypeMapTy::get(Type *Ty) {
227  Type *Result = getImpl(Ty);
228
229  // If this caused a reference to any struct type, resolve it before returning.
230  if (!SrcDefinitionsToResolve.empty())
231    linkDefinedTypeBodies();
232  return Result;
233}
234
235/// getImpl - This is the recursive version of get().
236Type *TypeMapTy::getImpl(Type *Ty) {
237  // If we already have an entry for this type, return it.
238  Type **Entry = &MappedTypes[Ty];
239  if (*Entry) return *Entry;
240
241  // If this is not a named struct type, then just map all of the elements and
242  // then rebuild the type from inside out.
243  if (!isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral()) {
244    // If there are no element types to map, then the type is itself.  This is
245    // true for the anonymous {} struct, things like 'float', integers, etc.
246    if (Ty->getNumContainedTypes() == 0)
247      return *Entry = Ty;
248
249    // Remap all of the elements, keeping track of whether any of them change.
250    bool AnyChange = false;
251    SmallVector<Type*, 4> ElementTypes;
252    ElementTypes.resize(Ty->getNumContainedTypes());
253    for (unsigned i = 0, e = Ty->getNumContainedTypes(); i != e; ++i) {
254      ElementTypes[i] = getImpl(Ty->getContainedType(i));
255      AnyChange |= ElementTypes[i] != Ty->getContainedType(i);
256    }
257
258    // If we found our type while recursively processing stuff, just use it.
259    Entry = &MappedTypes[Ty];
260    if (*Entry) return *Entry;
261
262    // If all of the element types mapped directly over, then the type is usable
263    // as-is.
264    if (!AnyChange)
265      return *Entry = Ty;
266
267    // Otherwise, rebuild a modified type.
268    switch (Ty->getTypeID()) {
269    default: assert(0 && "unknown derived type to remap");
270    case Type::ArrayTyID:
271      return *Entry = ArrayType::get(ElementTypes[0],
272                                     cast<ArrayType>(Ty)->getNumElements());
273    case Type::VectorTyID:
274      return *Entry = VectorType::get(ElementTypes[0],
275                                      cast<VectorType>(Ty)->getNumElements());
276    case Type::PointerTyID:
277      return *Entry = PointerType::get(ElementTypes[0],
278                                      cast<PointerType>(Ty)->getAddressSpace());
279    case Type::FunctionTyID:
280      return *Entry = FunctionType::get(ElementTypes[0],
281                                        makeArrayRef(ElementTypes).slice(1),
282                                        cast<FunctionType>(Ty)->isVarArg());
283    case Type::StructTyID:
284      // Note that this is only reached for anonymous structs.
285      return *Entry = StructType::get(Ty->getContext(), ElementTypes,
286                                      cast<StructType>(Ty)->isPacked());
287    }
288  }
289
290  // Otherwise, this is an unmapped named struct.  If the struct can be directly
291  // mapped over, just use it as-is.  This happens in a case when the linked-in
292  // module has something like:
293  //   %T = type {%T*, i32}
294  //   @GV = global %T* null
295  // where T does not exist at all in the destination module.
296  //
297  // The other case we watch for is when the type is not in the destination
298  // module, but that it has to be rebuilt because it refers to something that
299  // is already mapped.  For example, if the destination module has:
300  //  %A = type { i32 }
301  // and the source module has something like
302  //  %A' = type { i32 }
303  //  %B = type { %A'* }
304  //  @GV = global %B* null
305  // then we want to create a new type: "%B = type { %A*}" and have it take the
306  // pristine "%B" name from the source module.
307  //
308  // To determine which case this is, we have to recursively walk the type graph
309  // speculating that we'll be able to reuse it unmodified.  Only if this is
310  // safe would we map the entire thing over.  Because this is an optimization,
311  // and is not required for the prettiness of the linked module, we just skip
312  // it and always rebuild a type here.
313  StructType *STy = cast<StructType>(Ty);
314
315  // If the type is opaque, we can just use it directly.
316  if (STy->isOpaque())
317    return *Entry = STy;
318
319  // Otherwise we create a new type and resolve its body later.  This will be
320  // resolved by the top level of get().
321  SrcDefinitionsToResolve.push_back(STy);
322  StructType *DTy = StructType::create(STy->getContext());
323  DstResolvedOpaqueTypes.insert(DTy);
324  return *Entry = DTy;
325}
326
327
328
329//===----------------------------------------------------------------------===//
330// ModuleLinker implementation.
331//===----------------------------------------------------------------------===//
332
333namespace {
334  /// ModuleLinker - This is an implementation class for the LinkModules
335  /// function, which is the entrypoint for this file.
336  class ModuleLinker {
337    Module *DstM, *SrcM;
338
339    TypeMapTy TypeMap;
340
341    /// ValueMap - Mapping of values from what they used to be in Src, to what
342    /// they are now in DstM.  ValueToValueMapTy is a ValueMap, which involves
343    /// some overhead due to the use of Value handles which the Linker doesn't
344    /// actually need, but this allows us to reuse the ValueMapper code.
345    ValueToValueMapTy ValueMap;
346
347    struct AppendingVarInfo {
348      GlobalVariable *NewGV;  // New aggregate global in dest module.
349      Constant *DstInit;      // Old initializer from dest module.
350      Constant *SrcInit;      // Old initializer from src module.
351    };
352
353    std::vector<AppendingVarInfo> AppendingVars;
354
355    unsigned Mode; // Mode to treat source module.
356
357    // Set of items not to link in from source.
358    SmallPtrSet<const Value*, 16> DoNotLinkFromSource;
359
360    // Vector of functions to lazily link in.
361    std::vector<Function*> LazilyLinkFunctions;
362
363  public:
364    std::string ErrorMsg;
365
366    ModuleLinker(Module *dstM, Module *srcM, unsigned mode)
367      : DstM(dstM), SrcM(srcM), Mode(mode) { }
368
369    bool run();
370
371  private:
372    /// emitError - Helper method for setting a message and returning an error
373    /// code.
374    bool emitError(const Twine &Message) {
375      ErrorMsg = Message.str();
376      return true;
377    }
378
379    /// getLinkageResult - This analyzes the two global values and determines
380    /// what the result will look like in the destination module.
381    bool getLinkageResult(GlobalValue *Dest, const GlobalValue *Src,
382                          GlobalValue::LinkageTypes &LT, bool &LinkFromSrc);
383
384    /// getLinkedToGlobal - Given a global in the source module, return the
385    /// global in the destination module that is being linked to, if any.
386    GlobalValue *getLinkedToGlobal(GlobalValue *SrcGV) {
387      // If the source has no name it can't link.  If it has local linkage,
388      // there is no name match-up going on.
389      if (!SrcGV->hasName() || SrcGV->hasLocalLinkage())
390        return 0;
391
392      // Otherwise see if we have a match in the destination module's symtab.
393      GlobalValue *DGV = DstM->getNamedValue(SrcGV->getName());
394      if (DGV == 0) return 0;
395
396      // If we found a global with the same name in the dest module, but it has
397      // internal linkage, we are really not doing any linkage here.
398      if (DGV->hasLocalLinkage())
399        return 0;
400
401      // Otherwise, we do in fact link to the destination global.
402      return DGV;
403    }
404
405    void computeTypeMapping();
406
407    bool linkAppendingVarProto(GlobalVariable *DstGV, GlobalVariable *SrcGV);
408    bool linkGlobalProto(GlobalVariable *SrcGV);
409    bool linkFunctionProto(Function *SrcF);
410    bool linkAliasProto(GlobalAlias *SrcA);
411
412    void linkAppendingVarInit(const AppendingVarInfo &AVI);
413    void linkGlobalInits();
414    void linkFunctionBody(Function *Dst, Function *Src);
415    void linkAliasBodies();
416    void linkNamedMDNodes();
417  };
418}
419
420
421
422/// forceRenaming - The LLVM SymbolTable class autorenames globals that conflict
423/// in the symbol table.  This is good for all clients except for us.  Go
424/// through the trouble to force this back.
425static void forceRenaming(GlobalValue *GV, StringRef Name) {
426  // If the global doesn't force its name or if it already has the right name,
427  // there is nothing for us to do.
428  if (GV->hasLocalLinkage() || GV->getName() == Name)
429    return;
430
431  Module *M = GV->getParent();
432
433  // If there is a conflict, rename the conflict.
434  if (GlobalValue *ConflictGV = M->getNamedValue(Name)) {
435    GV->takeName(ConflictGV);
436    ConflictGV->setName(Name);    // This will cause ConflictGV to get renamed
437    assert(ConflictGV->getName() != Name && "forceRenaming didn't work");
438  } else {
439    GV->setName(Name);              // Force the name back
440  }
441}
442
443/// CopyGVAttributes - copy additional attributes (those not needed to construct
444/// a GlobalValue) from the SrcGV to the DestGV.
445static void CopyGVAttributes(GlobalValue *DestGV, const GlobalValue *SrcGV) {
446  // Use the maximum alignment, rather than just copying the alignment of SrcGV.
447  unsigned Alignment = std::max(DestGV->getAlignment(), SrcGV->getAlignment());
448  DestGV->copyAttributesFrom(SrcGV);
449  DestGV->setAlignment(Alignment);
450
451  forceRenaming(DestGV, SrcGV->getName());
452}
453
454/// getLinkageResult - This analyzes the two global values and determines what
455/// the result will look like in the destination module.  In particular, it
456/// computes the resultant linkage type, computes whether the global in the
457/// source should be copied over to the destination (replacing the existing
458/// one), and computes whether this linkage is an error or not. It also performs
459/// visibility checks: we cannot link together two symbols with different
460/// visibilities.
461bool ModuleLinker::getLinkageResult(GlobalValue *Dest, const GlobalValue *Src,
462                                    GlobalValue::LinkageTypes &LT,
463                                    bool &LinkFromSrc) {
464  assert(Dest && "Must have two globals being queried");
465  assert(!Src->hasLocalLinkage() &&
466         "If Src has internal linkage, Dest shouldn't be set!");
467
468  bool SrcIsDeclaration = Src->isDeclaration() && !Src->isMaterializable();
469  bool DestIsDeclaration = Dest->isDeclaration();
470
471  if (SrcIsDeclaration) {
472    // If Src is external or if both Src & Dest are external..  Just link the
473    // external globals, we aren't adding anything.
474    if (Src->hasDLLImportLinkage()) {
475      // If one of GVs has DLLImport linkage, result should be dllimport'ed.
476      if (DestIsDeclaration) {
477        LinkFromSrc = true;
478        LT = Src->getLinkage();
479      }
480    } else if (Dest->hasExternalWeakLinkage()) {
481      // If the Dest is weak, use the source linkage.
482      LinkFromSrc = true;
483      LT = Src->getLinkage();
484    } else {
485      LinkFromSrc = false;
486      LT = Dest->getLinkage();
487    }
488  } else if (DestIsDeclaration && !Dest->hasDLLImportLinkage()) {
489    // If Dest is external but Src is not:
490    LinkFromSrc = true;
491    LT = Src->getLinkage();
492  } else if (Src->isWeakForLinker()) {
493    // At this point we know that Dest has LinkOnce, External*, Weak, Common,
494    // or DLL* linkage.
495    if (Dest->hasExternalWeakLinkage() ||
496        Dest->hasAvailableExternallyLinkage() ||
497        (Dest->hasLinkOnceLinkage() &&
498         (Src->hasWeakLinkage() || Src->hasCommonLinkage()))) {
499      LinkFromSrc = true;
500      LT = Src->getLinkage();
501    } else {
502      LinkFromSrc = false;
503      LT = Dest->getLinkage();
504    }
505  } else if (Dest->isWeakForLinker()) {
506    // At this point we know that Src has External* or DLL* linkage.
507    if (Src->hasExternalWeakLinkage()) {
508      LinkFromSrc = false;
509      LT = Dest->getLinkage();
510    } else {
511      LinkFromSrc = true;
512      LT = GlobalValue::ExternalLinkage;
513    }
514  } else {
515    assert((Dest->hasExternalLinkage()  || Dest->hasDLLImportLinkage() ||
516            Dest->hasDLLExportLinkage() || Dest->hasExternalWeakLinkage()) &&
517           (Src->hasExternalLinkage()   || Src->hasDLLImportLinkage() ||
518            Src->hasDLLExportLinkage()  || Src->hasExternalWeakLinkage()) &&
519           "Unexpected linkage type!");
520    return emitError("Linking globals named '" + Src->getName() +
521                 "': symbol multiply defined!");
522  }
523
524  // Check visibility
525  if (Src->getVisibility() != Dest->getVisibility() &&
526      !SrcIsDeclaration && !DestIsDeclaration &&
527      !Src->hasAvailableExternallyLinkage() &&
528      !Dest->hasAvailableExternallyLinkage())
529    return emitError("Linking globals named '" + Src->getName() +
530                   "': symbols have different visibilities!");
531  return false;
532}
533
534/// computeTypeMapping - Loop over all of the linked values to compute type
535/// mappings.  For example, if we link "extern Foo *x" and "Foo *x = NULL", then
536/// we have two struct types 'Foo' but one got renamed when the module was
537/// loaded into the same LLVMContext.
538void ModuleLinker::computeTypeMapping() {
539  // Incorporate globals.
540  for (Module::global_iterator I = SrcM->global_begin(),
541       E = SrcM->global_end(); I != E; ++I) {
542    GlobalValue *DGV = getLinkedToGlobal(I);
543    if (DGV == 0) continue;
544
545    if (!DGV->hasAppendingLinkage() || !I->hasAppendingLinkage()) {
546      TypeMap.addTypeMapping(DGV->getType(), I->getType());
547      continue;
548    }
549
550    // Unify the element type of appending arrays.
551    ArrayType *DAT = cast<ArrayType>(DGV->getType()->getElementType());
552    ArrayType *SAT = cast<ArrayType>(I->getType()->getElementType());
553    TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
554  }
555
556  // Incorporate functions.
557  for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I) {
558    if (GlobalValue *DGV = getLinkedToGlobal(I))
559      TypeMap.addTypeMapping(DGV->getType(), I->getType());
560  }
561
562  // Incorporate types by name, scanning all the types in the source module.
563  // At this point, the destination module may have a type "%foo = { i32 }" for
564  // example.  When the source module got loaded into the same LLVMContext, if
565  // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
566  // Though it isn't required for correctness, attempt to link these up to clean
567  // up the IR.
568  std::vector<StructType*> SrcStructTypes;
569  SrcM->findUsedStructTypes(SrcStructTypes);
570
571  SmallPtrSet<StructType*, 32> SrcStructTypesSet(SrcStructTypes.begin(),
572                                                 SrcStructTypes.end());
573
574  for (unsigned i = 0, e = SrcStructTypes.size(); i != e; ++i) {
575    StructType *ST = SrcStructTypes[i];
576    if (!ST->hasName()) continue;
577
578    // Check to see if there is a dot in the name followed by a digit.
579    size_t DotPos = ST->getName().rfind('.');
580    if (DotPos == 0 || DotPos == StringRef::npos ||
581        ST->getName().back() == '.' || !isdigit(ST->getName()[DotPos+1]))
582      continue;
583
584    // Check to see if the destination module has a struct with the prefix name.
585    if (StructType *DST = DstM->getTypeByName(ST->getName().substr(0, DotPos)))
586      // Don't use it if this actually came from the source module.  They're in
587      // the same LLVMContext after all.
588      if (!SrcStructTypesSet.count(DST))
589        TypeMap.addTypeMapping(DST, ST);
590  }
591
592
593  // Don't bother incorporating aliases, they aren't generally typed well.
594
595  // Now that we have discovered all of the type equivalences, get a body for
596  // any 'opaque' types in the dest module that are now resolved.
597  TypeMap.linkDefinedTypeBodies();
598}
599
600/// linkAppendingVarProto - If there were any appending global variables, link
601/// them together now.  Return true on error.
602bool ModuleLinker::linkAppendingVarProto(GlobalVariable *DstGV,
603                                         GlobalVariable *SrcGV) {
604
605  if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage())
606    return emitError("Linking globals named '" + SrcGV->getName() +
607           "': can only link appending global with another appending global!");
608
609  ArrayType *DstTy = cast<ArrayType>(DstGV->getType()->getElementType());
610  ArrayType *SrcTy =
611    cast<ArrayType>(TypeMap.get(SrcGV->getType()->getElementType()));
612  Type *EltTy = DstTy->getElementType();
613
614  // Check to see that they two arrays agree on type.
615  if (EltTy != SrcTy->getElementType())
616    return emitError("Appending variables with different element types!");
617  if (DstGV->isConstant() != SrcGV->isConstant())
618    return emitError("Appending variables linked with different const'ness!");
619
620  if (DstGV->getAlignment() != SrcGV->getAlignment())
621    return emitError(
622             "Appending variables with different alignment need to be linked!");
623
624  if (DstGV->getVisibility() != SrcGV->getVisibility())
625    return emitError(
626            "Appending variables with different visibility need to be linked!");
627
628  if (DstGV->getSection() != SrcGV->getSection())
629    return emitError(
630          "Appending variables with different section name need to be linked!");
631
632  uint64_t NewSize = DstTy->getNumElements() + SrcTy->getNumElements();
633  ArrayType *NewType = ArrayType::get(EltTy, NewSize);
634
635  // Create the new global variable.
636  GlobalVariable *NG =
637    new GlobalVariable(*DstGV->getParent(), NewType, SrcGV->isConstant(),
638                       DstGV->getLinkage(), /*init*/0, /*name*/"", DstGV,
639                       DstGV->isThreadLocal(),
640                       DstGV->getType()->getAddressSpace());
641
642  // Propagate alignment, visibility and section info.
643  CopyGVAttributes(NG, DstGV);
644
645  AppendingVarInfo AVI;
646  AVI.NewGV = NG;
647  AVI.DstInit = DstGV->getInitializer();
648  AVI.SrcInit = SrcGV->getInitializer();
649  AppendingVars.push_back(AVI);
650
651  // Replace any uses of the two global variables with uses of the new
652  // global.
653  ValueMap[SrcGV] = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType()));
654
655  DstGV->replaceAllUsesWith(ConstantExpr::getBitCast(NG, DstGV->getType()));
656  DstGV->eraseFromParent();
657
658  // Track the source variable so we don't try to link it.
659  DoNotLinkFromSource.insert(SrcGV);
660
661  return false;
662}
663
664/// linkGlobalProto - Loop through the global variables in the src module and
665/// merge them into the dest module.
666bool ModuleLinker::linkGlobalProto(GlobalVariable *SGV) {
667  GlobalValue *DGV = getLinkedToGlobal(SGV);
668
669  if (DGV) {
670    // Concatenation of appending linkage variables is magic and handled later.
671    if (DGV->hasAppendingLinkage() || SGV->hasAppendingLinkage())
672      return linkAppendingVarProto(cast<GlobalVariable>(DGV), SGV);
673
674    // Determine whether linkage of these two globals follows the source
675    // module's definition or the destination module's definition.
676    GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
677    bool LinkFromSrc = false;
678    if (getLinkageResult(DGV, SGV, NewLinkage, LinkFromSrc))
679      return true;
680
681    // If we're not linking from the source, then keep the definition that we
682    // have.
683    if (!LinkFromSrc) {
684      // Special case for const propagation.
685      if (GlobalVariable *DGVar = dyn_cast<GlobalVariable>(DGV))
686        if (DGVar->isDeclaration() && SGV->isConstant() && !DGVar->isConstant())
687          DGVar->setConstant(true);
688
689      // Set calculated linkage.
690      DGV->setLinkage(NewLinkage);
691
692      // Make sure to remember this mapping.
693      ValueMap[SGV] = ConstantExpr::getBitCast(DGV,TypeMap.get(SGV->getType()));
694
695      // Track the source global so that we don't attempt to copy it over when
696      // processing global initializers.
697      DoNotLinkFromSource.insert(SGV);
698
699      return false;
700    }
701  }
702
703  // No linking to be performed or linking from the source: simply create an
704  // identical version of the symbol over in the dest module... the
705  // initializer will be filled in later by LinkGlobalInits.
706  GlobalVariable *NewDGV =
707    new GlobalVariable(*DstM, TypeMap.get(SGV->getType()->getElementType()),
708                       SGV->isConstant(), SGV->getLinkage(), /*init*/0,
709                       SGV->getName(), /*insertbefore*/0,
710                       SGV->isThreadLocal(),
711                       SGV->getType()->getAddressSpace());
712  // Propagate alignment, visibility and section info.
713  CopyGVAttributes(NewDGV, SGV);
714
715  if (DGV) {
716    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDGV, DGV->getType()));
717    DGV->eraseFromParent();
718  }
719
720  // Make sure to remember this mapping.
721  ValueMap[SGV] = NewDGV;
722  return false;
723}
724
725/// linkFunctionProto - Link the function in the source module into the
726/// destination module if needed, setting up mapping information.
727bool ModuleLinker::linkFunctionProto(Function *SF) {
728  GlobalValue *DGV = getLinkedToGlobal(SF);
729
730  if (DGV) {
731    GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
732    bool LinkFromSrc = false;
733    if (getLinkageResult(DGV, SF, NewLinkage, LinkFromSrc))
734      return true;
735
736    if (!LinkFromSrc) {
737      // Set calculated linkage
738      DGV->setLinkage(NewLinkage);
739
740      // Make sure to remember this mapping.
741      ValueMap[SF] = ConstantExpr::getBitCast(DGV, TypeMap.get(SF->getType()));
742
743      // Track the function from the source module so we don't attempt to remap
744      // it.
745      DoNotLinkFromSource.insert(SF);
746
747      return false;
748    }
749  }
750
751  // If there is no linkage to be performed or we are linking from the source,
752  // bring SF over.
753  Function *NewDF = Function::Create(TypeMap.get(SF->getFunctionType()),
754                                     SF->getLinkage(), SF->getName(), DstM);
755  CopyGVAttributes(NewDF, SF);
756
757  if (DGV) {
758    // Any uses of DF need to change to NewDF, with cast.
759    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDF, DGV->getType()));
760    DGV->eraseFromParent();
761  } else {
762    // Internal, LO_ODR, or LO linkage - stick in set to ignore and lazily link.
763    if (SF->hasLocalLinkage() || SF->hasLinkOnceLinkage() ||
764        SF->hasAvailableExternallyLinkage()) {
765      DoNotLinkFromSource.insert(SF);
766      LazilyLinkFunctions.push_back(SF);
767    }
768  }
769
770  ValueMap[SF] = NewDF;
771  return false;
772}
773
774/// LinkAliasProto - Set up prototypes for any aliases that come over from the
775/// source module.
776bool ModuleLinker::linkAliasProto(GlobalAlias *SGA) {
777  GlobalValue *DGV = getLinkedToGlobal(SGA);
778
779  if (DGV) {
780    GlobalValue::LinkageTypes NewLinkage = GlobalValue::InternalLinkage;
781    bool LinkFromSrc = false;
782    if (getLinkageResult(DGV, SGA, NewLinkage, LinkFromSrc))
783      return true;
784
785    if (!LinkFromSrc) {
786      // Set calculated linkage.
787      DGV->setLinkage(NewLinkage);
788
789      // Make sure to remember this mapping.
790      ValueMap[SGA] = ConstantExpr::getBitCast(DGV,TypeMap.get(SGA->getType()));
791
792      // Track the alias from the source module so we don't attempt to remap it.
793      DoNotLinkFromSource.insert(SGA);
794
795      return false;
796    }
797  }
798
799  // If there is no linkage to be performed or we're linking from the source,
800  // bring over SGA.
801  GlobalAlias *NewDA = new GlobalAlias(TypeMap.get(SGA->getType()),
802                                       SGA->getLinkage(), SGA->getName(),
803                                       /*aliasee*/0, DstM);
804  CopyGVAttributes(NewDA, SGA);
805
806  if (DGV) {
807    // Any uses of DGV need to change to NewDA, with cast.
808    DGV->replaceAllUsesWith(ConstantExpr::getBitCast(NewDA, DGV->getType()));
809    DGV->eraseFromParent();
810  }
811
812  ValueMap[SGA] = NewDA;
813  return false;
814}
815
816void ModuleLinker::linkAppendingVarInit(const AppendingVarInfo &AVI) {
817  // Merge the initializer.
818  SmallVector<Constant*, 16> Elements;
819  if (ConstantArray *I = dyn_cast<ConstantArray>(AVI.DstInit)) {
820    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
821      Elements.push_back(I->getOperand(i));
822  } else {
823    assert(isa<ConstantAggregateZero>(AVI.DstInit));
824    ArrayType *DstAT = cast<ArrayType>(AVI.DstInit->getType());
825    Type *EltTy = DstAT->getElementType();
826    Elements.append(DstAT->getNumElements(), Constant::getNullValue(EltTy));
827  }
828
829  Constant *SrcInit = MapValue(AVI.SrcInit, ValueMap, RF_None, &TypeMap);
830  if (const ConstantArray *I = dyn_cast<ConstantArray>(SrcInit)) {
831    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
832      Elements.push_back(I->getOperand(i));
833  } else {
834    assert(isa<ConstantAggregateZero>(SrcInit));
835    ArrayType *SrcAT = cast<ArrayType>(SrcInit->getType());
836    Type *EltTy = SrcAT->getElementType();
837    Elements.append(SrcAT->getNumElements(), Constant::getNullValue(EltTy));
838  }
839  ArrayType *NewType = cast<ArrayType>(AVI.NewGV->getType()->getElementType());
840  AVI.NewGV->setInitializer(ConstantArray::get(NewType, Elements));
841}
842
843
844// linkGlobalInits - Update the initializers in the Dest module now that all
845// globals that may be referenced are in Dest.
846void ModuleLinker::linkGlobalInits() {
847  // Loop over all of the globals in the src module, mapping them over as we go
848  for (Module::const_global_iterator I = SrcM->global_begin(),
849       E = SrcM->global_end(); I != E; ++I) {
850
851    // Only process initialized GV's or ones not already in dest.
852    if (!I->hasInitializer() || DoNotLinkFromSource.count(I)) continue;
853
854    // Grab destination global variable.
855    GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[I]);
856    // Figure out what the initializer looks like in the dest module.
857    DGV->setInitializer(MapValue(I->getInitializer(), ValueMap,
858                                 RF_None, &TypeMap));
859  }
860}
861
862// linkFunctionBody - Copy the source function over into the dest function and
863// fix up references to values.  At this point we know that Dest is an external
864// function, and that Src is not.
865void ModuleLinker::linkFunctionBody(Function *Dst, Function *Src) {
866  assert(Src && Dst && Dst->isDeclaration() && !Src->isDeclaration());
867
868  // Go through and convert function arguments over, remembering the mapping.
869  Function::arg_iterator DI = Dst->arg_begin();
870  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
871       I != E; ++I, ++DI) {
872    DI->setName(I->getName());  // Copy the name over.
873
874    // Add a mapping to our mapping.
875    ValueMap[I] = DI;
876  }
877
878  if (Mode == Linker::DestroySource) {
879    // Splice the body of the source function into the dest function.
880    Dst->getBasicBlockList().splice(Dst->end(), Src->getBasicBlockList());
881
882    // At this point, all of the instructions and values of the function are now
883    // copied over.  The only problem is that they are still referencing values in
884    // the Source function as operands.  Loop through all of the operands of the
885    // functions and patch them up to point to the local versions.
886    for (Function::iterator BB = Dst->begin(), BE = Dst->end(); BB != BE; ++BB)
887      for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
888        RemapInstruction(I, ValueMap, RF_IgnoreMissingEntries, &TypeMap);
889
890  } else {
891    // Clone the body of the function into the dest function.
892    SmallVector<ReturnInst*, 8> Returns; // Ignore returns.
893    CloneFunctionInto(Dst, Src, ValueMap, false, Returns);
894  }
895
896  // There is no need to map the arguments anymore.
897  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
898       I != E; ++I)
899    ValueMap.erase(I);
900
901}
902
903
904void ModuleLinker::linkAliasBodies() {
905  for (Module::alias_iterator I = SrcM->alias_begin(), E = SrcM->alias_end();
906       I != E; ++I) {
907    if (DoNotLinkFromSource.count(I))
908      continue;
909    if (Constant *Aliasee = I->getAliasee()) {
910      GlobalAlias *DA = cast<GlobalAlias>(ValueMap[I]);
911      DA->setAliasee(MapValue(Aliasee, ValueMap, RF_None, &TypeMap));
912    }
913  }
914}
915
916/// linkNamedMDNodes - Insert all of the named mdnodes in Src into the Dest
917/// module.
918void ModuleLinker::linkNamedMDNodes() {
919  for (Module::const_named_metadata_iterator I = SrcM->named_metadata_begin(),
920       E = SrcM->named_metadata_end(); I != E; ++I) {
921    NamedMDNode *DestNMD = DstM->getOrInsertNamedMetadata(I->getName());
922    // Add Src elements into Dest node.
923    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
924      DestNMD->addOperand(MapValue(I->getOperand(i), ValueMap,
925                                   RF_None, &TypeMap));
926  }
927}
928
929bool ModuleLinker::run() {
930  assert(DstM && "Null Destination module");
931  assert(SrcM && "Null Source Module");
932
933  // Inherit the target data from the source module if the destination module
934  // doesn't have one already.
935  if (DstM->getDataLayout().empty() && !SrcM->getDataLayout().empty())
936    DstM->setDataLayout(SrcM->getDataLayout());
937
938  // Copy the target triple from the source to dest if the dest's is empty.
939  if (DstM->getTargetTriple().empty() && !SrcM->getTargetTriple().empty())
940    DstM->setTargetTriple(SrcM->getTargetTriple());
941
942  if (!SrcM->getDataLayout().empty() && !DstM->getDataLayout().empty() &&
943      SrcM->getDataLayout() != DstM->getDataLayout())
944    errs() << "WARNING: Linking two modules of different data layouts!\n";
945  if (!SrcM->getTargetTriple().empty() &&
946      DstM->getTargetTriple() != SrcM->getTargetTriple()) {
947    errs() << "WARNING: Linking two modules of different target triples: ";
948    if (!SrcM->getModuleIdentifier().empty())
949      errs() << SrcM->getModuleIdentifier() << ": ";
950    errs() << "'" << SrcM->getTargetTriple() << "' and '"
951           << DstM->getTargetTriple() << "'\n";
952  }
953
954  // Append the module inline asm string.
955  if (!SrcM->getModuleInlineAsm().empty()) {
956    if (DstM->getModuleInlineAsm().empty())
957      DstM->setModuleInlineAsm(SrcM->getModuleInlineAsm());
958    else
959      DstM->setModuleInlineAsm(DstM->getModuleInlineAsm()+"\n"+
960                               SrcM->getModuleInlineAsm());
961  }
962
963  // Update the destination module's dependent libraries list with the libraries
964  // from the source module. There's no opportunity for duplicates here as the
965  // Module ensures that duplicate insertions are discarded.
966  for (Module::lib_iterator SI = SrcM->lib_begin(), SE = SrcM->lib_end();
967       SI != SE; ++SI)
968    DstM->addLibrary(*SI);
969
970  // If the source library's module id is in the dependent library list of the
971  // destination library, remove it since that module is now linked in.
972  StringRef ModuleId = SrcM->getModuleIdentifier();
973  if (!ModuleId.empty())
974    DstM->removeLibrary(sys::path::stem(ModuleId));
975
976  // Loop over all of the linked values to compute type mappings.
977  computeTypeMapping();
978
979  // Insert all of the globals in src into the DstM module... without linking
980  // initializers (which could refer to functions not yet mapped over).
981  for (Module::global_iterator I = SrcM->global_begin(),
982       E = SrcM->global_end(); I != E; ++I)
983    if (linkGlobalProto(I))
984      return true;
985
986  // Link the functions together between the two modules, without doing function
987  // bodies... this just adds external function prototypes to the DstM
988  // function...  We do this so that when we begin processing function bodies,
989  // all of the global values that may be referenced are available in our
990  // ValueMap.
991  for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I)
992    if (linkFunctionProto(I))
993      return true;
994
995  // If there were any aliases, link them now.
996  for (Module::alias_iterator I = SrcM->alias_begin(),
997       E = SrcM->alias_end(); I != E; ++I)
998    if (linkAliasProto(I))
999      return true;
1000
1001  for (unsigned i = 0, e = AppendingVars.size(); i != e; ++i)
1002    linkAppendingVarInit(AppendingVars[i]);
1003
1004  // Update the initializers in the DstM module now that all globals that may
1005  // be referenced are in DstM.
1006  linkGlobalInits();
1007
1008  // Link in the function bodies that are defined in the source module into
1009  // DstM.
1010  for (Module::iterator SF = SrcM->begin(), E = SrcM->end(); SF != E; ++SF) {
1011
1012    // Skip if not linking from source.
1013    if (DoNotLinkFromSource.count(SF)) continue;
1014
1015    // Skip if no body (function is external) or materialize.
1016    if (SF->isDeclaration()) {
1017      if (!SF->isMaterializable())
1018        continue;
1019      if (SF->Materialize(&ErrorMsg))
1020        return true;
1021    }
1022
1023    linkFunctionBody(cast<Function>(ValueMap[SF]), SF);
1024  }
1025
1026  // Resolve all uses of aliases with aliasees.
1027  linkAliasBodies();
1028
1029  // Remap all of the named mdnoes in Src into the DstM module. We do this
1030  // after linking GlobalValues so that MDNodes that reference GlobalValues
1031  // are properly remapped.
1032  linkNamedMDNodes();
1033
1034  // Process vector of lazily linked in functions.
1035  bool LinkedInAnyFunctions;
1036  do {
1037    LinkedInAnyFunctions = false;
1038
1039    for(std::vector<Function*>::iterator I = LazilyLinkFunctions.begin(),
1040        E = LazilyLinkFunctions.end(); I != E; ++I) {
1041      if (!*I)
1042        continue;
1043
1044      Function *SF = *I;
1045      Function *DF = cast<Function>(ValueMap[SF]);
1046
1047      if (!DF->use_empty()) {
1048
1049        // Materialize if necessary.
1050        if (SF->isDeclaration()) {
1051          if (!SF->isMaterializable())
1052            continue;
1053          if (SF->Materialize(&ErrorMsg))
1054            return true;
1055        }
1056
1057        // Link in function body.
1058        linkFunctionBody(DF, SF);
1059
1060        // "Remove" from vector by setting the element to 0.
1061        *I = 0;
1062
1063        // Set flag to indicate we may have more functions to lazily link in
1064        // since we linked in a function.
1065        LinkedInAnyFunctions = true;
1066      }
1067    }
1068  } while (LinkedInAnyFunctions);
1069
1070  // Remove any prototypes of functions that were not actually linked in.
1071  for(std::vector<Function*>::iterator I = LazilyLinkFunctions.begin(),
1072      E = LazilyLinkFunctions.end(); I != E; ++I) {
1073    if (!*I)
1074      continue;
1075
1076    Function *SF = *I;
1077    Function *DF = cast<Function>(ValueMap[SF]);
1078    if (DF->use_empty())
1079      DF->eraseFromParent();
1080  }
1081
1082  // Now that all of the types from the source are used, resolve any structs
1083  // copied over to the dest that didn't exist there.
1084  TypeMap.linkDefinedTypeBodies();
1085
1086  return false;
1087}
1088
1089//===----------------------------------------------------------------------===//
1090// LinkModules entrypoint.
1091//===----------------------------------------------------------------------===//
1092
1093// LinkModules - This function links two modules together, with the resulting
1094// left module modified to be the composite of the two input modules.  If an
1095// error occurs, true is returned and ErrorMsg (if not null) is set to indicate
1096// the problem.  Upon failure, the Dest module could be in a modified state, and
1097// shouldn't be relied on to be consistent.
1098bool Linker::LinkModules(Module *Dest, Module *Src, unsigned Mode,
1099                         std::string *ErrorMsg) {
1100  ModuleLinker TheLinker(Dest, Src, Mode);
1101  if (TheLinker.run()) {
1102    if (ErrorMsg) *ErrorMsg = TheLinker.ErrorMsg;
1103    return true;
1104  }
1105
1106  return false;
1107}
1108