LinkModules.cpp revision ed74a4ef3b813bed0c0f4bb5af42229d50133da2
1//===- lib/Linker/LinkModules.cpp - Module Linker Implementation ----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the LLVM module linker.
11//
12// Specifically, this:
13//  * Merges global variables between the two modules
14//    * Uninit + Uninit = Init, Init + Uninit = Init, Init + Init = Error if !=
15//  * Merges functions between two modules
16//
17//===----------------------------------------------------------------------===//
18
19#include "llvm/Linker.h"
20#include "llvm/Constants.h"
21#include "llvm/DerivedTypes.h"
22#include "llvm/Module.h"
23#include "llvm/SymbolTable.h"
24#include "llvm/Instructions.h"
25#include "llvm/Assembly/Writer.h"
26#include "llvm/System/Path.h"
27#include <iostream>
28#include <sstream>
29using namespace llvm;
30
31// Error - Simple wrapper function to conditionally assign to E and return true.
32// This just makes error return conditions a little bit simpler...
33static inline bool Error(std::string *E, const std::string &Message) {
34  if (E) *E = Message;
35  return true;
36}
37
38// ToStr - Simple wrapper function to convert a type to a string.
39static std::string ToStr(const Type *Ty, const Module *M) {
40  std::ostringstream OS;
41  WriteTypeSymbolic(OS, Ty, M);
42  return OS.str();
43}
44
45//
46// Function: ResolveTypes()
47//
48// Description:
49//  Attempt to link the two specified types together.
50//
51// Inputs:
52//  DestTy - The type to which we wish to resolve.
53//  SrcTy  - The original type which we want to resolve.
54//  Name   - The name of the type.
55//
56// Outputs:
57//  DestST - The symbol table in which the new type should be placed.
58//
59// Return value:
60//  true  - There is an error and the types cannot yet be linked.
61//  false - No errors.
62//
63static bool ResolveTypes(const Type *DestTy, const Type *SrcTy,
64                         SymbolTable *DestST, const std::string &Name) {
65  if (DestTy == SrcTy) return false;       // If already equal, noop
66
67  // Does the type already exist in the module?
68  if (DestTy && !isa<OpaqueType>(DestTy)) {  // Yup, the type already exists...
69    if (const OpaqueType *OT = dyn_cast<OpaqueType>(SrcTy)) {
70      const_cast<OpaqueType*>(OT)->refineAbstractTypeTo(DestTy);
71    } else {
72      return true;  // Cannot link types... neither is opaque and not-equal
73    }
74  } else {                       // Type not in dest module.  Add it now.
75    if (DestTy)                  // Type _is_ in module, just opaque...
76      const_cast<OpaqueType*>(cast<OpaqueType>(DestTy))
77                           ->refineAbstractTypeTo(SrcTy);
78    else if (!Name.empty())
79      DestST->insert(Name, const_cast<Type*>(SrcTy));
80  }
81  return false;
82}
83
84static const FunctionType *getFT(const PATypeHolder &TH) {
85  return cast<FunctionType>(TH.get());
86}
87static const StructType *getST(const PATypeHolder &TH) {
88  return cast<StructType>(TH.get());
89}
90
91// RecursiveResolveTypes - This is just like ResolveTypes, except that it
92// recurses down into derived types, merging the used types if the parent types
93// are compatible.
94static bool RecursiveResolveTypesI(const PATypeHolder &DestTy,
95                                   const PATypeHolder &SrcTy,
96                                   SymbolTable *DestST, const std::string &Name,
97                std::vector<std::pair<PATypeHolder, PATypeHolder> > &Pointers) {
98  const Type *SrcTyT = SrcTy.get();
99  const Type *DestTyT = DestTy.get();
100  if (DestTyT == SrcTyT) return false;       // If already equal, noop
101
102  // If we found our opaque type, resolve it now!
103  if (isa<OpaqueType>(DestTyT) || isa<OpaqueType>(SrcTyT))
104    return ResolveTypes(DestTyT, SrcTyT, DestST, Name);
105
106  // Two types cannot be resolved together if they are of different primitive
107  // type.  For example, we cannot resolve an int to a float.
108  if (DestTyT->getTypeID() != SrcTyT->getTypeID()) return true;
109
110  // Otherwise, resolve the used type used by this derived type...
111  switch (DestTyT->getTypeID()) {
112  case Type::FunctionTyID: {
113    if (cast<FunctionType>(DestTyT)->isVarArg() !=
114        cast<FunctionType>(SrcTyT)->isVarArg() ||
115        cast<FunctionType>(DestTyT)->getNumContainedTypes() !=
116        cast<FunctionType>(SrcTyT)->getNumContainedTypes())
117      return true;
118    for (unsigned i = 0, e = getFT(DestTy)->getNumContainedTypes(); i != e; ++i)
119      if (RecursiveResolveTypesI(getFT(DestTy)->getContainedType(i),
120                                 getFT(SrcTy)->getContainedType(i), DestST, "",
121                                 Pointers))
122        return true;
123    return false;
124  }
125  case Type::StructTyID: {
126    if (getST(DestTy)->getNumContainedTypes() !=
127        getST(SrcTy)->getNumContainedTypes()) return 1;
128    for (unsigned i = 0, e = getST(DestTy)->getNumContainedTypes(); i != e; ++i)
129      if (RecursiveResolveTypesI(getST(DestTy)->getContainedType(i),
130                                 getST(SrcTy)->getContainedType(i), DestST, "",
131                                 Pointers))
132        return true;
133    return false;
134  }
135  case Type::ArrayTyID: {
136    const ArrayType *DAT = cast<ArrayType>(DestTy.get());
137    const ArrayType *SAT = cast<ArrayType>(SrcTy.get());
138    if (DAT->getNumElements() != SAT->getNumElements()) return true;
139    return RecursiveResolveTypesI(DAT->getElementType(), SAT->getElementType(),
140                                  DestST, "", Pointers);
141  }
142  case Type::PointerTyID: {
143    // If this is a pointer type, check to see if we have already seen it.  If
144    // so, we are in a recursive branch.  Cut off the search now.  We cannot use
145    // an associative container for this search, because the type pointers (keys
146    // in the container) change whenever types get resolved...
147    for (unsigned i = 0, e = Pointers.size(); i != e; ++i)
148      if (Pointers[i].first == DestTy)
149        return Pointers[i].second != SrcTy;
150
151    // Otherwise, add the current pointers to the vector to stop recursion on
152    // this pair.
153    Pointers.push_back(std::make_pair(DestTyT, SrcTyT));
154    bool Result =
155      RecursiveResolveTypesI(cast<PointerType>(DestTy.get())->getElementType(),
156                             cast<PointerType>(SrcTy.get())->getElementType(),
157                             DestST, "", Pointers);
158    Pointers.pop_back();
159    return Result;
160  }
161  default: assert(0 && "Unexpected type!"); return true;
162  }
163}
164
165static bool RecursiveResolveTypes(const PATypeHolder &DestTy,
166                                  const PATypeHolder &SrcTy,
167                                  SymbolTable *DestST, const std::string &Name){
168  std::vector<std::pair<PATypeHolder, PATypeHolder> > PointerTypes;
169  return RecursiveResolveTypesI(DestTy, SrcTy, DestST, Name, PointerTypes);
170}
171
172
173// LinkTypes - Go through the symbol table of the Src module and see if any
174// types are named in the src module that are not named in the Dst module.
175// Make sure there are no type name conflicts.
176static bool LinkTypes(Module *Dest, const Module *Src, std::string *Err) {
177  SymbolTable       *DestST = &Dest->getSymbolTable();
178  const SymbolTable *SrcST  = &Src->getSymbolTable();
179
180  // Look for a type plane for Type's...
181  SymbolTable::type_const_iterator TI = SrcST->type_begin();
182  SymbolTable::type_const_iterator TE = SrcST->type_end();
183  if (TI == TE) return false;  // No named types, do nothing.
184
185  // Some types cannot be resolved immediately because they depend on other
186  // types being resolved to each other first.  This contains a list of types we
187  // are waiting to recheck.
188  std::vector<std::string> DelayedTypesToResolve;
189
190  for ( ; TI != TE; ++TI ) {
191    const std::string &Name = TI->first;
192    const Type *RHS = TI->second;
193
194    // Check to see if this type name is already in the dest module...
195    Type *Entry = DestST->lookupType(Name);
196
197    if (ResolveTypes(Entry, RHS, DestST, Name)) {
198      // They look different, save the types 'till later to resolve.
199      DelayedTypesToResolve.push_back(Name);
200    }
201  }
202
203  // Iteratively resolve types while we can...
204  while (!DelayedTypesToResolve.empty()) {
205    // Loop over all of the types, attempting to resolve them if possible...
206    unsigned OldSize = DelayedTypesToResolve.size();
207
208    // Try direct resolution by name...
209    for (unsigned i = 0; i != DelayedTypesToResolve.size(); ++i) {
210      const std::string &Name = DelayedTypesToResolve[i];
211      Type *T1 = SrcST->lookupType(Name);
212      Type *T2 = DestST->lookupType(Name);
213      if (!ResolveTypes(T2, T1, DestST, Name)) {
214        // We are making progress!
215        DelayedTypesToResolve.erase(DelayedTypesToResolve.begin()+i);
216        --i;
217      }
218    }
219
220    // Did we not eliminate any types?
221    if (DelayedTypesToResolve.size() == OldSize) {
222      // Attempt to resolve subelements of types.  This allows us to merge these
223      // two types: { int* } and { opaque* }
224      for (unsigned i = 0, e = DelayedTypesToResolve.size(); i != e; ++i) {
225        const std::string &Name = DelayedTypesToResolve[i];
226        PATypeHolder T1(SrcST->lookupType(Name));
227        PATypeHolder T2(DestST->lookupType(Name));
228
229        if (!RecursiveResolveTypes(T2, T1, DestST, Name)) {
230          // We are making progress!
231          DelayedTypesToResolve.erase(DelayedTypesToResolve.begin()+i);
232
233          // Go back to the main loop, perhaps we can resolve directly by name
234          // now...
235          break;
236        }
237      }
238
239      // If we STILL cannot resolve the types, then there is something wrong.
240      if (DelayedTypesToResolve.size() == OldSize) {
241        // Remove the symbol name from the destination.
242        DelayedTypesToResolve.pop_back();
243      }
244    }
245  }
246
247
248  return false;
249}
250
251static void PrintMap(const std::map<const Value*, Value*> &M) {
252  for (std::map<const Value*, Value*>::const_iterator I = M.begin(), E =M.end();
253       I != E; ++I) {
254    std::cerr << " Fr: " << (void*)I->first << " ";
255    I->first->dump();
256    std::cerr << " To: " << (void*)I->second << " ";
257    I->second->dump();
258    std::cerr << "\n";
259  }
260}
261
262
263// RemapOperand - Use ValueMap to convert references from one module to another.
264// This is somewhat sophisticated in that it can automatically handle constant
265// references correctly as well...
266static Value *RemapOperand(const Value *In,
267                           std::map<const Value*, Value*> &ValueMap) {
268  std::map<const Value*,Value*>::const_iterator I = ValueMap.find(In);
269  if (I != ValueMap.end()) return I->second;
270
271  // Check to see if it's a constant that we are interesting in transforming.
272  if (const Constant *CPV = dyn_cast<Constant>(In)) {
273    if ((!isa<DerivedType>(CPV->getType()) && !isa<ConstantExpr>(CPV)) ||
274        isa<ConstantAggregateZero>(CPV))
275      return const_cast<Constant*>(CPV);   // Simple constants stay identical.
276
277    Constant *Result = 0;
278
279    if (const ConstantArray *CPA = dyn_cast<ConstantArray>(CPV)) {
280      std::vector<Constant*> Operands(CPA->getNumOperands());
281      for (unsigned i = 0, e = CPA->getNumOperands(); i != e; ++i)
282        Operands[i] =cast<Constant>(RemapOperand(CPA->getOperand(i), ValueMap));
283      Result = ConstantArray::get(cast<ArrayType>(CPA->getType()), Operands);
284    } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(CPV)) {
285      std::vector<Constant*> Operands(CPS->getNumOperands());
286      for (unsigned i = 0, e = CPS->getNumOperands(); i != e; ++i)
287        Operands[i] =cast<Constant>(RemapOperand(CPS->getOperand(i), ValueMap));
288      Result = ConstantStruct::get(cast<StructType>(CPS->getType()), Operands);
289    } else if (isa<ConstantPointerNull>(CPV) || isa<UndefValue>(CPV)) {
290      Result = const_cast<Constant*>(CPV);
291    } else if (isa<GlobalValue>(CPV)) {
292      Result = cast<Constant>(RemapOperand(CPV, ValueMap));
293    } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
294      if (CE->getOpcode() == Instruction::GetElementPtr) {
295        Value *Ptr = RemapOperand(CE->getOperand(0), ValueMap);
296        std::vector<Constant*> Indices;
297        Indices.reserve(CE->getNumOperands()-1);
298        for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
299          Indices.push_back(cast<Constant>(RemapOperand(CE->getOperand(i),
300                                                        ValueMap)));
301
302        Result = ConstantExpr::getGetElementPtr(cast<Constant>(Ptr), Indices);
303      } else if (CE->getNumOperands() == 1) {
304        // Cast instruction
305        assert(CE->getOpcode() == Instruction::Cast);
306        Value *V = RemapOperand(CE->getOperand(0), ValueMap);
307        Result = ConstantExpr::getCast(cast<Constant>(V), CE->getType());
308      } else if (CE->getNumOperands() == 3) {
309        // Select instruction
310        assert(CE->getOpcode() == Instruction::Select);
311        Value *V1 = RemapOperand(CE->getOperand(0), ValueMap);
312        Value *V2 = RemapOperand(CE->getOperand(1), ValueMap);
313        Value *V3 = RemapOperand(CE->getOperand(2), ValueMap);
314        Result = ConstantExpr::getSelect(cast<Constant>(V1), cast<Constant>(V2),
315                                         cast<Constant>(V3));
316      } else if (CE->getNumOperands() == 2) {
317        // Binary operator...
318        Value *V1 = RemapOperand(CE->getOperand(0), ValueMap);
319        Value *V2 = RemapOperand(CE->getOperand(1), ValueMap);
320
321        Result = ConstantExpr::get(CE->getOpcode(), cast<Constant>(V1),
322                                   cast<Constant>(V2));
323      } else {
324        assert(0 && "Unknown constant expr type!");
325      }
326
327    } else {
328      assert(0 && "Unknown type of derived type constant value!");
329    }
330
331    // Cache the mapping in our local map structure...
332    ValueMap.insert(std::make_pair(In, Result));
333    return Result;
334  }
335
336  std::cerr << "LinkModules ValueMap: \n";
337  PrintMap(ValueMap);
338
339  std::cerr << "Couldn't remap value: " << (void*)In << " " << *In << "\n";
340  assert(0 && "Couldn't remap value!");
341  return 0;
342}
343
344/// ForceRenaming - The LLVM SymbolTable class autorenames globals that conflict
345/// in the symbol table.  This is good for all clients except for us.  Go
346/// through the trouble to force this back.
347static void ForceRenaming(GlobalValue *GV, const std::string &Name) {
348  assert(GV->getName() != Name && "Can't force rename to self");
349  SymbolTable &ST = GV->getParent()->getSymbolTable();
350
351  // If there is a conflict, rename the conflict.
352  Value *ConflictVal = ST.lookup(GV->getType(), Name);
353  assert(ConflictVal&&"Why do we have to force rename if there is no conflic?");
354  GlobalValue *ConflictGV = cast<GlobalValue>(ConflictVal);
355  assert(ConflictGV->hasInternalLinkage() &&
356         "Not conflicting with a static global, should link instead!");
357
358  ConflictGV->setName("");          // Eliminate the conflict
359  GV->setName(Name);                // Force the name back
360  ConflictGV->setName(Name);        // This will cause ConflictGV to get renamed
361  assert(GV->getName() == Name && ConflictGV->getName() != Name &&
362         "ForceRenaming didn't work");
363}
364
365/// GetLinkageResult - This analyzes the two global values and determines what
366/// the result will look like in the destination module.  In particular, it
367/// computes the resultant linkage type, computes whether the global in the
368/// source should be copied over to the destination (replacing the existing
369/// one), and computes whether this linkage is an error or not.
370static bool GetLinkageResult(GlobalValue *Dest, GlobalValue *Src,
371                             GlobalValue::LinkageTypes &LT, bool &LinkFromSrc,
372                             std::string *Err) {
373  assert((!Dest || !Src->hasInternalLinkage()) &&
374         "If Src has internal linkage, Dest shouldn't be set!");
375  if (!Dest) {
376    // Linking something to nothing.
377    LinkFromSrc = true;
378    LT = Src->getLinkage();
379  } else if (Src->isExternal()) {
380    // If Src is external or if both Src & Drc are external..  Just link the
381    // external globals, we aren't adding anything.
382    LinkFromSrc = false;
383    LT = Dest->getLinkage();
384  } else if (Dest->isExternal()) {
385    // If Dest is external but Src is not:
386    LinkFromSrc = true;
387    LT = Src->getLinkage();
388  } else if (Src->hasAppendingLinkage() || Dest->hasAppendingLinkage()) {
389    if (Src->getLinkage() != Dest->getLinkage())
390      return Error(Err, "Linking globals named '" + Src->getName() +
391            "': can only link appending global with another appending global!");
392    LinkFromSrc = true; // Special cased.
393    LT = Src->getLinkage();
394  } else if (Src->hasWeakLinkage() || Src->hasLinkOnceLinkage()) {
395    // At this point we know that Dest has LinkOnce, External or Weak linkage.
396    if (Dest->hasLinkOnceLinkage() && Src->hasWeakLinkage()) {
397      LinkFromSrc = true;
398      LT = Src->getLinkage();
399    } else {
400      LinkFromSrc = false;
401      LT = Dest->getLinkage();
402    }
403  } else if (Dest->hasWeakLinkage() || Dest->hasLinkOnceLinkage()) {
404    // At this point we know that Src has External linkage.
405    LinkFromSrc = true;
406    LT = GlobalValue::ExternalLinkage;
407  } else {
408    assert(Dest->hasExternalLinkage() && Src->hasExternalLinkage() &&
409           "Unexpected linkage type!");
410    return Error(Err, "Linking globals named '" + Src->getName() +
411                 "': symbol multiply defined!");
412  }
413  return false;
414}
415
416// LinkGlobals - Loop through the global variables in the src module and merge
417// them into the dest module.
418static bool LinkGlobals(Module *Dest, Module *Src,
419                        std::map<const Value*, Value*> &ValueMap,
420                    std::multimap<std::string, GlobalVariable *> &AppendingVars,
421                        std::map<std::string, GlobalValue*> &GlobalsByName,
422                        std::string *Err) {
423  // We will need a module level symbol table if the src module has a module
424  // level symbol table...
425  SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable();
426
427  // Loop over all of the globals in the src module, mapping them over as we go
428  for (Module::global_iterator I = Src->global_begin(), E = Src->global_end(); I != E; ++I) {
429    GlobalVariable *SGV = I;
430    GlobalVariable *DGV = 0;
431    // Check to see if may have to link the global.
432    if (SGV->hasName() && !SGV->hasInternalLinkage())
433      if (!(DGV = Dest->getGlobalVariable(SGV->getName(),
434                                          SGV->getType()->getElementType()))) {
435        std::map<std::string, GlobalValue*>::iterator EGV =
436          GlobalsByName.find(SGV->getName());
437        if (EGV != GlobalsByName.end())
438          DGV = dyn_cast<GlobalVariable>(EGV->second);
439        if (DGV)
440          // If types don't agree due to opaque types, try to resolve them.
441          RecursiveResolveTypes(SGV->getType(), DGV->getType(),ST, "");
442      }
443
444    if (DGV && DGV->hasInternalLinkage())
445      DGV = 0;
446
447    assert(SGV->hasInitializer() || SGV->hasExternalLinkage() &&
448           "Global must either be external or have an initializer!");
449
450    GlobalValue::LinkageTypes NewLinkage;
451    bool LinkFromSrc;
452    if (GetLinkageResult(DGV, SGV, NewLinkage, LinkFromSrc, Err))
453      return true;
454
455    if (!DGV) {
456      // No linking to be performed, simply create an identical version of the
457      // symbol over in the dest module... the initializer will be filled in
458      // later by LinkGlobalInits...
459      GlobalVariable *NewDGV =
460        new GlobalVariable(SGV->getType()->getElementType(),
461                           SGV->isConstant(), SGV->getLinkage(), /*init*/0,
462                           SGV->getName(), Dest);
463
464      // If the LLVM runtime renamed the global, but it is an externally visible
465      // symbol, DGV must be an existing global with internal linkage.  Rename
466      // it.
467      if (NewDGV->getName() != SGV->getName() && !NewDGV->hasInternalLinkage())
468        ForceRenaming(NewDGV, SGV->getName());
469
470      // Make sure to remember this mapping...
471      ValueMap.insert(std::make_pair(SGV, NewDGV));
472      if (SGV->hasAppendingLinkage())
473        // Keep track that this is an appending variable...
474        AppendingVars.insert(std::make_pair(SGV->getName(), NewDGV));
475    } else if (DGV->hasAppendingLinkage()) {
476      // No linking is performed yet.  Just insert a new copy of the global, and
477      // keep track of the fact that it is an appending variable in the
478      // AppendingVars map.  The name is cleared out so that no linkage is
479      // performed.
480      GlobalVariable *NewDGV =
481        new GlobalVariable(SGV->getType()->getElementType(),
482                           SGV->isConstant(), SGV->getLinkage(), /*init*/0,
483                           "", Dest);
484
485      // Make sure to remember this mapping...
486      ValueMap.insert(std::make_pair(SGV, NewDGV));
487
488      // Keep track that this is an appending variable...
489      AppendingVars.insert(std::make_pair(SGV->getName(), NewDGV));
490    } else {
491      // Otherwise, perform the mapping as instructed by GetLinkageResult.  If
492      // the types don't match, and if we are to link from the source, nuke DGV
493      // and create a new one of the appropriate type.
494      if (SGV->getType() != DGV->getType() && LinkFromSrc) {
495        GlobalVariable *NewDGV =
496          new GlobalVariable(SGV->getType()->getElementType(),
497                             DGV->isConstant(), DGV->getLinkage());
498        Dest->getGlobalList().insert(DGV, NewDGV);
499        DGV->replaceAllUsesWith(ConstantExpr::getCast(NewDGV, DGV->getType()));
500        DGV->eraseFromParent();
501        NewDGV->setName(SGV->getName());
502        DGV = NewDGV;
503      }
504
505      DGV->setLinkage(NewLinkage);
506
507      if (LinkFromSrc) {
508        // Inherit const as appropriate
509        DGV->setConstant(SGV->isConstant());
510        DGV->setInitializer(0);
511      } else {
512        if (SGV->isConstant() && !DGV->isConstant()) {
513          if (DGV->isExternal())
514            DGV->setConstant(true);
515        }
516        SGV->setLinkage(GlobalValue::ExternalLinkage);
517        SGV->setInitializer(0);
518      }
519
520      ValueMap.insert(std::make_pair(SGV,
521                                     ConstantExpr::getCast(DGV,
522                                                           SGV->getType())));
523    }
524  }
525  return false;
526}
527
528
529// LinkGlobalInits - Update the initializers in the Dest module now that all
530// globals that may be referenced are in Dest.
531static bool LinkGlobalInits(Module *Dest, const Module *Src,
532                            std::map<const Value*, Value*> &ValueMap,
533                            std::string *Err) {
534
535  // Loop over all of the globals in the src module, mapping them over as we go
536  for (Module::const_global_iterator I = Src->global_begin(), E = Src->global_end(); I != E; ++I){
537    const GlobalVariable *SGV = I;
538
539    if (SGV->hasInitializer()) {      // Only process initialized GV's
540      // Figure out what the initializer looks like in the dest module...
541      Constant *SInit =
542        cast<Constant>(RemapOperand(SGV->getInitializer(), ValueMap));
543
544      GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[SGV]);
545      if (DGV->hasInitializer()) {
546        if (SGV->hasExternalLinkage()) {
547          if (DGV->getInitializer() != SInit)
548            return Error(Err, "Global Variable Collision on '" +
549                         ToStr(SGV->getType(), Src) +"':%"+SGV->getName()+
550                         " - Global variables have different initializers");
551        } else if (DGV->hasLinkOnceLinkage() || DGV->hasWeakLinkage()) {
552          // Nothing is required, mapped values will take the new global
553          // automatically.
554        } else if (SGV->hasLinkOnceLinkage() || SGV->hasWeakLinkage()) {
555          // Nothing is required, mapped values will take the new global
556          // automatically.
557        } else if (DGV->hasAppendingLinkage()) {
558          assert(0 && "Appending linkage unimplemented!");
559        } else {
560          assert(0 && "Unknown linkage!");
561        }
562      } else {
563        // Copy the initializer over now...
564        DGV->setInitializer(SInit);
565      }
566    }
567  }
568  return false;
569}
570
571// LinkFunctionProtos - Link the functions together between the two modules,
572// without doing function bodies... this just adds external function prototypes
573// to the Dest function...
574//
575static bool LinkFunctionProtos(Module *Dest, const Module *Src,
576                               std::map<const Value*, Value*> &ValueMap,
577                             std::map<std::string, GlobalValue*> &GlobalsByName,
578                               std::string *Err) {
579  SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable();
580
581  // Loop over all of the functions in the src module, mapping them over as we
582  // go
583  for (Module::const_iterator I = Src->begin(), E = Src->end(); I != E; ++I) {
584    const Function *SF = I;   // SrcFunction
585    Function *DF = 0;
586    if (SF->hasName() && !SF->hasInternalLinkage()) {
587      // Check to see if may have to link the function.
588      if (!(DF = Dest->getFunction(SF->getName(), SF->getFunctionType()))) {
589        std::map<std::string, GlobalValue*>::iterator EF =
590          GlobalsByName.find(SF->getName());
591        if (EF != GlobalsByName.end())
592          DF = dyn_cast<Function>(EF->second);
593        if (DF && RecursiveResolveTypes(SF->getType(), DF->getType(), ST, ""))
594          DF = 0;  // FIXME: gross.
595      }
596    }
597
598    if (!DF || SF->hasInternalLinkage() || DF->hasInternalLinkage()) {
599      // Function does not already exist, simply insert an function signature
600      // identical to SF into the dest module...
601      Function *NewDF = new Function(SF->getFunctionType(), SF->getLinkage(),
602                                     SF->getName(), Dest);
603      NewDF->setCallingConv(SF->getCallingConv());
604
605      // If the LLVM runtime renamed the function, but it is an externally
606      // visible symbol, DF must be an existing function with internal linkage.
607      // Rename it.
608      if (NewDF->getName() != SF->getName() && !NewDF->hasInternalLinkage())
609        ForceRenaming(NewDF, SF->getName());
610
611      // ... and remember this mapping...
612      ValueMap.insert(std::make_pair(SF, NewDF));
613    } else if (SF->isExternal()) {
614      // If SF is external or if both SF & DF are external..  Just link the
615      // external functions, we aren't adding anything.
616      ValueMap.insert(std::make_pair(SF, DF));
617    } else if (DF->isExternal()) {   // If DF is external but SF is not...
618      // Link the external functions, update linkage qualifiers
619      ValueMap.insert(std::make_pair(SF, DF));
620      DF->setLinkage(SF->getLinkage());
621
622    } else if (SF->hasWeakLinkage() || SF->hasLinkOnceLinkage()) {
623      // At this point we know that DF has LinkOnce, Weak, or External linkage.
624      ValueMap.insert(std::make_pair(SF, DF));
625
626      // Linkonce+Weak = Weak
627      if (DF->hasLinkOnceLinkage() && SF->hasWeakLinkage())
628        DF->setLinkage(SF->getLinkage());
629
630    } else if (DF->hasWeakLinkage() || DF->hasLinkOnceLinkage()) {
631      // At this point we know that SF has LinkOnce or External linkage.
632      ValueMap.insert(std::make_pair(SF, DF));
633      if (!SF->hasLinkOnceLinkage())   // Don't inherit linkonce linkage
634        DF->setLinkage(SF->getLinkage());
635
636    } else if (SF->getLinkage() != DF->getLinkage()) {
637      return Error(Err, "Functions named '" + SF->getName() +
638                   "' have different linkage specifiers!");
639    } else if (SF->hasExternalLinkage()) {
640      // The function is defined in both modules!!
641      return Error(Err, "Function '" +
642                   ToStr(SF->getFunctionType(), Src) + "':\"" +
643                   SF->getName() + "\" - Function is already defined!");
644    } else {
645      assert(0 && "Unknown linkage configuration found!");
646    }
647  }
648  return false;
649}
650
651// LinkFunctionBody - Copy the source function over into the dest function and
652// fix up references to values.  At this point we know that Dest is an external
653// function, and that Src is not.
654static bool LinkFunctionBody(Function *Dest, Function *Src,
655                             std::map<const Value*, Value*> &GlobalMap,
656                             std::string *Err) {
657  assert(Src && Dest && Dest->isExternal() && !Src->isExternal());
658
659  // Go through and convert function arguments over, remembering the mapping.
660  Function::arg_iterator DI = Dest->arg_begin();
661  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
662       I != E; ++I, ++DI) {
663    DI->setName(I->getName());  // Copy the name information over...
664
665    // Add a mapping to our local map
666    GlobalMap.insert(std::make_pair(I, DI));
667  }
668
669  // Splice the body of the source function into the dest function.
670  Dest->getBasicBlockList().splice(Dest->end(), Src->getBasicBlockList());
671
672  // At this point, all of the instructions and values of the function are now
673  // copied over.  The only problem is that they are still referencing values in
674  // the Source function as operands.  Loop through all of the operands of the
675  // functions and patch them up to point to the local versions...
676  //
677  for (Function::iterator BB = Dest->begin(), BE = Dest->end(); BB != BE; ++BB)
678    for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
679      for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
680           OI != OE; ++OI)
681        if (!isa<Instruction>(*OI) && !isa<BasicBlock>(*OI))
682          *OI = RemapOperand(*OI, GlobalMap);
683
684  // There is no need to map the arguments anymore.
685  for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end(); I != E; ++I)
686    GlobalMap.erase(I);
687
688  return false;
689}
690
691
692// LinkFunctionBodies - Link in the function bodies that are defined in the
693// source module into the DestModule.  This consists basically of copying the
694// function over and fixing up references to values.
695static bool LinkFunctionBodies(Module *Dest, Module *Src,
696                               std::map<const Value*, Value*> &ValueMap,
697                               std::string *Err) {
698
699  // Loop over all of the functions in the src module, mapping them over as we
700  // go
701  for (Module::iterator SF = Src->begin(), E = Src->end(); SF != E; ++SF) {
702    if (!SF->isExternal()) {                  // No body if function is external
703      Function *DF = cast<Function>(ValueMap[SF]); // Destination function
704
705      // DF not external SF external?
706      if (DF->isExternal()) {
707        // Only provide the function body if there isn't one already.
708        if (LinkFunctionBody(DF, SF, ValueMap, Err))
709          return true;
710      }
711    }
712  }
713  return false;
714}
715
716// LinkAppendingVars - If there were any appending global variables, link them
717// together now.  Return true on error.
718static bool LinkAppendingVars(Module *M,
719                  std::multimap<std::string, GlobalVariable *> &AppendingVars,
720                              std::string *ErrorMsg) {
721  if (AppendingVars.empty()) return false; // Nothing to do.
722
723  // Loop over the multimap of appending vars, processing any variables with the
724  // same name, forming a new appending global variable with both of the
725  // initializers merged together, then rewrite references to the old variables
726  // and delete them.
727  std::vector<Constant*> Inits;
728  while (AppendingVars.size() > 1) {
729    // Get the first two elements in the map...
730    std::multimap<std::string,
731      GlobalVariable*>::iterator Second = AppendingVars.begin(), First=Second++;
732
733    // If the first two elements are for different names, there is no pair...
734    // Otherwise there is a pair, so link them together...
735    if (First->first == Second->first) {
736      GlobalVariable *G1 = First->second, *G2 = Second->second;
737      const ArrayType *T1 = cast<ArrayType>(G1->getType()->getElementType());
738      const ArrayType *T2 = cast<ArrayType>(G2->getType()->getElementType());
739
740      // Check to see that they two arrays agree on type...
741      if (T1->getElementType() != T2->getElementType())
742        return Error(ErrorMsg,
743         "Appending variables with different element types need to be linked!");
744      if (G1->isConstant() != G2->isConstant())
745        return Error(ErrorMsg,
746                     "Appending variables linked with different const'ness!");
747
748      unsigned NewSize = T1->getNumElements() + T2->getNumElements();
749      ArrayType *NewType = ArrayType::get(T1->getElementType(), NewSize);
750
751      G1->setName("");   // Clear G1's name in case of a conflict!
752
753      // Create the new global variable...
754      GlobalVariable *NG =
755        new GlobalVariable(NewType, G1->isConstant(), G1->getLinkage(),
756                           /*init*/0, First->first, M);
757
758      // Merge the initializer...
759      Inits.reserve(NewSize);
760      if (ConstantArray *I = dyn_cast<ConstantArray>(G1->getInitializer())) {
761        for (unsigned i = 0, e = T1->getNumElements(); i != e; ++i)
762          Inits.push_back(I->getOperand(i));
763      } else {
764        assert(isa<ConstantAggregateZero>(G1->getInitializer()));
765        Constant *CV = Constant::getNullValue(T1->getElementType());
766        for (unsigned i = 0, e = T1->getNumElements(); i != e; ++i)
767          Inits.push_back(CV);
768      }
769      if (ConstantArray *I = dyn_cast<ConstantArray>(G2->getInitializer())) {
770        for (unsigned i = 0, e = T2->getNumElements(); i != e; ++i)
771          Inits.push_back(I->getOperand(i));
772      } else {
773        assert(isa<ConstantAggregateZero>(G2->getInitializer()));
774        Constant *CV = Constant::getNullValue(T2->getElementType());
775        for (unsigned i = 0, e = T2->getNumElements(); i != e; ++i)
776          Inits.push_back(CV);
777      }
778      NG->setInitializer(ConstantArray::get(NewType, Inits));
779      Inits.clear();
780
781      // Replace any uses of the two global variables with uses of the new
782      // global...
783
784      // FIXME: This should rewrite simple/straight-forward uses such as
785      // getelementptr instructions to not use the Cast!
786      G1->replaceAllUsesWith(ConstantExpr::getCast(NG, G1->getType()));
787      G2->replaceAllUsesWith(ConstantExpr::getCast(NG, G2->getType()));
788
789      // Remove the two globals from the module now...
790      M->getGlobalList().erase(G1);
791      M->getGlobalList().erase(G2);
792
793      // Put the new global into the AppendingVars map so that we can handle
794      // linking of more than two vars...
795      Second->second = NG;
796    }
797    AppendingVars.erase(First);
798  }
799
800  return false;
801}
802
803
804// LinkModules - This function links two modules together, with the resulting
805// left module modified to be the composite of the two input modules.  If an
806// error occurs, true is returned and ErrorMsg (if not null) is set to indicate
807// the problem.  Upon failure, the Dest module could be in a modified state, and
808// shouldn't be relied on to be consistent.
809bool
810Linker::LinkModules(Module *Dest, Module *Src, std::string *ErrorMsg) {
811  assert(Dest != 0 && "Invalid Destination module");
812  assert(Src  != 0 && "Invalid Source Module");
813
814  if (Dest->getEndianness() == Module::AnyEndianness)
815    Dest->setEndianness(Src->getEndianness());
816  if (Dest->getPointerSize() == Module::AnyPointerSize)
817    Dest->setPointerSize(Src->getPointerSize());
818  if (Dest->getTargetTriple().empty())
819    Dest->setTargetTriple(Src->getTargetTriple());
820
821  if (Src->getEndianness() != Module::AnyEndianness &&
822      Dest->getEndianness() != Src->getEndianness())
823    std::cerr << "WARNING: Linking two modules of different endianness!\n";
824  if (Src->getPointerSize() != Module::AnyPointerSize &&
825      Dest->getPointerSize() != Src->getPointerSize())
826    std::cerr << "WARNING: Linking two modules of different pointer size!\n";
827  if (!Src->getTargetTriple().empty() &&
828      Dest->getTargetTriple() != Src->getTargetTriple())
829    std::cerr << "WARNING: Linking two modules of different target triples!\n";
830
831  // Update the destination module's dependent libraries list with the libraries
832  // from the source module. There's no opportunity for duplicates here as the
833  // Module ensures that duplicate insertions are discarded.
834  Module::lib_iterator SI = Src->lib_begin();
835  Module::lib_iterator SE = Src->lib_end();
836  while ( SI != SE ) {
837    Dest->addLibrary(*SI);
838    ++SI;
839  }
840
841  // LinkTypes - Go through the symbol table of the Src module and see if any
842  // types are named in the src module that are not named in the Dst module.
843  // Make sure there are no type name conflicts.
844  if (LinkTypes(Dest, Src, ErrorMsg)) return true;
845
846  // ValueMap - Mapping of values from what they used to be in Src, to what they
847  // are now in Dest.
848  std::map<const Value*, Value*> ValueMap;
849
850  // AppendingVars - Keep track of global variables in the destination module
851  // with appending linkage.  After the module is linked together, they are
852  // appended and the module is rewritten.
853  std::multimap<std::string, GlobalVariable *> AppendingVars;
854
855  // GlobalsByName - The LLVM SymbolTable class fights our best efforts at
856  // linking by separating globals by type.  Until PR411 is fixed, we replicate
857  // it's functionality here.
858  std::map<std::string, GlobalValue*> GlobalsByName;
859
860  for (Module::global_iterator I = Dest->global_begin(), E = Dest->global_end(); I != E; ++I) {
861    // Add all of the appending globals already in the Dest module to
862    // AppendingVars.
863    if (I->hasAppendingLinkage())
864      AppendingVars.insert(std::make_pair(I->getName(), I));
865
866    // Keep track of all globals by name.
867    if (!I->hasInternalLinkage() && I->hasName())
868      GlobalsByName[I->getName()] = I;
869  }
870
871  // Keep track of all globals by name.
872  for (Module::iterator I = Dest->begin(), E = Dest->end(); I != E; ++I)
873    if (!I->hasInternalLinkage() && I->hasName())
874      GlobalsByName[I->getName()] = I;
875
876  // Insert all of the globals in src into the Dest module... without linking
877  // initializers (which could refer to functions not yet mapped over).
878  if (LinkGlobals(Dest, Src, ValueMap, AppendingVars, GlobalsByName, ErrorMsg))
879    return true;
880
881  // Link the functions together between the two modules, without doing function
882  // bodies... this just adds external function prototypes to the Dest
883  // function...  We do this so that when we begin processing function bodies,
884  // all of the global values that may be referenced are available in our
885  // ValueMap.
886  if (LinkFunctionProtos(Dest, Src, ValueMap, GlobalsByName, ErrorMsg))
887    return true;
888
889  // Update the initializers in the Dest module now that all globals that may
890  // be referenced are in Dest.
891  if (LinkGlobalInits(Dest, Src, ValueMap, ErrorMsg)) return true;
892
893  // Link in the function bodies that are defined in the source module into the
894  // DestModule.  This consists basically of copying the function over and
895  // fixing up references to values.
896  if (LinkFunctionBodies(Dest, Src, ValueMap, ErrorMsg)) return true;
897
898  // If there were any appending global variables, link them together now.
899  if (LinkAppendingVars(Dest, AppendingVars, ErrorMsg)) return true;
900
901  // If the source library's module id is in the dependent library list of the
902  // destination library, remove it since that module is now linked in.
903  sys::Path modId;
904  modId.set(Src->getModuleIdentifier());
905  if (!modId.isEmpty())
906    Dest->removeLibrary(modId.getBasename());
907
908  return false;
909}
910
911// vim: sw=2
912