1 //===- Linker.cpp - Module Linker Implementation --------------------------===//
3 // This file implements the LLVM module linker.
6 // * Merges global variables between the two modules
7 // * Uninit + Uninit = Init, Init + Uninit = Init, Init + Init = Error if !=
8 // * Merges functions between two modules
10 //===----------------------------------------------------------------------===//
12 #include "llvm/Transforms/Utils/Linker.h"
13 #include "llvm/Module.h"
14 #include "llvm/SymbolTable.h"
15 #include "llvm/DerivedTypes.h"
16 #include "llvm/iOther.h"
17 #include "llvm/Constants.h"
19 // Error - Simple wrapper function to conditionally assign to E and return true.
20 // This just makes error return conditions a little bit simpler...
22 static inline bool Error(std::string *E, std::string Message) {
27 // LinkTypes - Go through the symbol table of the Src module and see if any
28 // types are named in the src module that are not named in the Dst module.
29 // Make sure there are no type name conflicts.
31 static bool LinkTypes(Module *Dest, const Module *Src, std::string *Err) {
32 SymbolTable *DestST = &Dest->getSymbolTable();
33 const SymbolTable *SrcST = &Src->getSymbolTable();
35 // Look for a type plane for Type's...
36 SymbolTable::const_iterator PI = SrcST->find(Type::TypeTy);
37 if (PI == SrcST->end()) return false; // No named types, do nothing.
39 const SymbolTable::VarMap &VM = PI->second;
40 for (SymbolTable::type_const_iterator I = VM.begin(), E = VM.end();
42 const std::string &Name = I->first;
43 const Type *RHS = cast<Type>(I->second);
45 // Check to see if this type name is already in the dest module...
46 const Type *Entry = cast_or_null<Type>(DestST->lookup(Type::TypeTy, Name));
47 if (Entry && !isa<OpaqueType>(Entry)) { // Yup, the value already exists...
49 if (OpaqueType *OT = dyn_cast<OpaqueType>(const_cast<Type*>(RHS))) {
50 OT->refineAbstractTypeTo(Entry);
52 // If it's the same, noop. Otherwise, error.
53 return Error(Err, "Type named '" + Name +
54 "' of different shape in modules.\n Src='" +
55 Entry->getDescription() + "'.\n Dst='" +
56 RHS->getDescription() + "'");
59 } else { // Type not in dest module. Add it now.
61 OpaqueType *OT = cast<OpaqueType>(const_cast<Type*>(Entry));
62 OT->refineAbstractTypeTo(RHS);
65 // TODO: FIXME WHEN TYPES AREN'T CONST
66 DestST->insert(Name, const_cast<Type*>(RHS));
72 static void PrintMap(const std::map<const Value*, Value*> &M) {
73 for (std::map<const Value*, Value*>::const_iterator I = M.begin(), E =M.end();
75 std::cerr << " Fr: " << (void*)I->first << " ";
77 std::cerr << " To: " << (void*)I->second << " ";
84 // RemapOperand - Use LocalMap and GlobalMap to convert references from one
85 // module to another. This is somewhat sophisticated in that it can
86 // automatically handle constant references correctly as well...
88 static Value *RemapOperand(const Value *In,
89 std::map<const Value*, Value*> &LocalMap,
90 std::map<const Value*, Value*> *GlobalMap) {
91 std::map<const Value*,Value*>::const_iterator I = LocalMap.find(In);
92 if (I != LocalMap.end()) return I->second;
95 I = GlobalMap->find(In);
96 if (I != GlobalMap->end()) return I->second;
99 // Check to see if it's a constant that we are interesting in transforming...
100 if (const Constant *CPV = dyn_cast<Constant>(In)) {
101 if (!isa<DerivedType>(CPV->getType()) && !isa<ConstantExpr>(CPV))
102 return const_cast<Constant*>(CPV); // Simple constants stay identical...
104 Constant *Result = 0;
106 if (const ConstantArray *CPA = dyn_cast<ConstantArray>(CPV)) {
107 const std::vector<Use> &Ops = CPA->getValues();
108 std::vector<Constant*> Operands(Ops.size());
109 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
111 cast<Constant>(RemapOperand(Ops[i], LocalMap, GlobalMap));
112 Result = ConstantArray::get(cast<ArrayType>(CPA->getType()), Operands);
113 } else if (const ConstantStruct *CPS = dyn_cast<ConstantStruct>(CPV)) {
114 const std::vector<Use> &Ops = CPS->getValues();
115 std::vector<Constant*> Operands(Ops.size());
116 for (unsigned i = 0; i < Ops.size(); ++i)
118 cast<Constant>(RemapOperand(Ops[i], LocalMap, GlobalMap));
119 Result = ConstantStruct::get(cast<StructType>(CPS->getType()), Operands);
120 } else if (isa<ConstantPointerNull>(CPV)) {
121 Result = const_cast<Constant*>(CPV);
122 } else if (const ConstantPointerRef *CPR =
123 dyn_cast<ConstantPointerRef>(CPV)) {
124 Value *V = RemapOperand(CPR->getValue(), LocalMap, GlobalMap);
125 Result = ConstantPointerRef::get(cast<GlobalValue>(V));
126 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(CPV)) {
127 if (CE->getOpcode() == Instruction::GetElementPtr) {
128 Value *Ptr = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap);
129 std::vector<Constant*> Indices;
130 Indices.reserve(CE->getNumOperands()-1);
131 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
132 Indices.push_back(cast<Constant>(RemapOperand(CE->getOperand(i),
133 LocalMap, GlobalMap)));
135 Result = ConstantExpr::getGetElementPtr(cast<Constant>(Ptr), Indices);
136 } else if (CE->getNumOperands() == 1) {
138 assert(CE->getOpcode() == Instruction::Cast);
139 Value *V = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap);
140 Result = ConstantExpr::getCast(cast<Constant>(V), CE->getType());
141 } else if (CE->getNumOperands() == 2) {
142 // Binary operator...
143 Value *V1 = RemapOperand(CE->getOperand(0), LocalMap, GlobalMap);
144 Value *V2 = RemapOperand(CE->getOperand(1), LocalMap, GlobalMap);
146 Result = ConstantExpr::get(CE->getOpcode(), cast<Constant>(V1),
149 assert(0 && "Unknown constant expr type!");
153 assert(0 && "Unknown type of derived type constant value!");
156 // Cache the mapping in our local map structure...
158 GlobalMap->insert(std::make_pair(In, Result));
160 LocalMap.insert(std::make_pair(In, Result));
164 std::cerr << "XXX LocalMap: \n";
168 std::cerr << "XXX GlobalMap: \n";
169 PrintMap(*GlobalMap);
172 std::cerr << "Couldn't remap value: " << (void*)In << " " << *In << "\n";
173 assert(0 && "Couldn't remap value!");
178 // LinkGlobals - Loop through the global variables in the src module and merge
179 // them into the dest module...
181 static bool LinkGlobals(Module *Dest, const Module *Src,
182 std::map<const Value*, Value*> &ValueMap,
184 // We will need a module level symbol table if the src module has a module
185 // level symbol table...
186 SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable();
188 // Loop over all of the globals in the src module, mapping them over as we go
190 for (Module::const_giterator I = Src->gbegin(), E = Src->gend(); I != E; ++I){
191 const GlobalVariable *SGV = I;
192 GlobalVariable *DGV = 0;
193 if (SGV->hasName()) {
194 // A same named thing is a global variable, because the only two things
195 // that may be in a module level symbol table are Global Vars and
196 // Functions, and they both have distinct, nonoverlapping, possible types.
198 DGV = cast_or_null<GlobalVariable>(ST->lookup(SGV->getType(),
202 assert(SGV->hasInitializer() || SGV->hasExternalLinkage() &&
203 "Global must either be external or have an initializer!");
205 bool SGExtern = SGV->isExternal();
206 bool DGExtern = DGV ? DGV->isExternal() : false;
208 if (!DGV || DGV->hasInternalLinkage() || SGV->hasInternalLinkage()) {
209 // No linking to be performed, simply create an identical version of the
210 // symbol over in the dest module... the initializer will be filled in
211 // later by LinkGlobalInits...
213 GlobalVariable *NewDGV =
214 new GlobalVariable(SGV->getType()->getElementType(),
215 SGV->isConstant(), SGV->getLinkage(), /*init*/0,
216 SGV->getName(), Dest);
218 // If the LLVM runtime renamed the global, but it is an externally visible
219 // symbol, DGV must be an existing global with internal linkage. Rename
221 if (NewDGV->getName() != SGV->getName() && !NewDGV->hasInternalLinkage()){
222 assert(DGV && DGV->getName() == SGV->getName() &&
223 DGV->hasInternalLinkage());
225 NewDGV->setName(SGV->getName()); // Force the name back
226 DGV->setName(SGV->getName()); // This will cause a renaming
227 assert(NewDGV->getName() == SGV->getName() &&
228 DGV->getName() != SGV->getName());
231 // Make sure to remember this mapping...
232 ValueMap.insert(std::make_pair(SGV, NewDGV));
233 } else if (SGV->isExternal()) {
234 // If SGV is external or if both SGV & DGV are external.. Just link the
235 // external globals, we aren't adding anything.
236 ValueMap.insert(std::make_pair(SGV, DGV));
238 } else if (DGV->isExternal()) { // If DGV is external but SGV is not...
239 ValueMap.insert(std::make_pair(SGV, DGV));
240 DGV->setLinkage(SGV->getLinkage()); // Inherit linkage!
241 } else if (SGV->getLinkage() != DGV->getLinkage()) {
242 return Error(Err, "Global variables named '" + SGV->getName() +
243 "' have different linkage specifiers!");
244 } else if (SGV->hasExternalLinkage()) {
245 // Allow linking two exactly identical external global variables...
246 if (SGV->isConstant() != DGV->isConstant() ||
247 SGV->getInitializer() != DGV->getInitializer())
248 return Error(Err, "Global Variable Collision on '" +
249 SGV->getType()->getDescription() + " %" + SGV->getName() +
250 "' - Global variables differ in const'ness");
251 ValueMap.insert(std::make_pair(SGV, DGV));
252 } else if (SGV->hasLinkOnceLinkage()) {
253 // If the global variable has a name, and that name is already in use in
254 // the Dest module, make sure that the name is a compatible global
257 // Check to see if the two GV's have the same Const'ness...
258 if (SGV->isConstant() != DGV->isConstant())
259 return Error(Err, "Global Variable Collision on '" +
260 SGV->getType()->getDescription() + " %" + SGV->getName() +
261 "' - Global variables differ in const'ness");
263 // Okay, everything is cool, remember the mapping...
264 ValueMap.insert(std::make_pair(SGV, DGV));
265 } else if (SGV->hasAppendingLinkage()) {
266 assert(0 && "FIXME: Appending linkage unimplemented!");
268 assert(0 && "Unknown linkage!");
275 // LinkGlobalInits - Update the initializers in the Dest module now that all
276 // globals that may be referenced are in Dest.
278 static bool LinkGlobalInits(Module *Dest, const Module *Src,
279 std::map<const Value*, Value*> &ValueMap,
282 // Loop over all of the globals in the src module, mapping them over as we go
284 for (Module::const_giterator I = Src->gbegin(), E = Src->gend(); I != E; ++I){
285 const GlobalVariable *SGV = I;
287 if (SGV->hasInitializer()) { // Only process initialized GV's
288 // Figure out what the initializer looks like in the dest module...
290 cast<Constant>(RemapOperand(SGV->getInitializer(), ValueMap, 0));
292 GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[SGV]);
293 if (DGV->hasInitializer()) {
294 assert(SGV->getLinkage() == DGV->getLinkage());
295 if (SGV->hasExternalLinkage()) {
296 if (DGV->getInitializer() != SInit)
297 return Error(Err, "Global Variable Collision on '" +
298 SGV->getType()->getDescription() +"':%"+SGV->getName()+
299 " - Global variables have different initializers");
300 } else if (DGV->hasLinkOnceLinkage()) {
301 // Nothing is required, mapped values will take the new global
303 } else if (DGV->hasAppendingLinkage()) {
304 assert(0 && "Appending linkage unimplemented!");
306 assert(0 && "Unknown linkage!");
309 // Copy the initializer over now...
310 DGV->setInitializer(SInit);
317 // LinkFunctionProtos - Link the functions together between the two modules,
318 // without doing function bodies... this just adds external function prototypes
319 // to the Dest function...
321 static bool LinkFunctionProtos(Module *Dest, const Module *Src,
322 std::map<const Value*, Value*> &ValueMap,
324 SymbolTable *ST = (SymbolTable*)&Dest->getSymbolTable();
326 // Loop over all of the functions in the src module, mapping them over as we
329 for (Module::const_iterator I = Src->begin(), E = Src->end(); I != E; ++I) {
330 const Function *SF = I; // SrcFunction
333 // The same named thing is a Function, because the only two things
334 // that may be in a module level symbol table are Global Vars and
335 // Functions, and they both have distinct, nonoverlapping, possible types.
337 DF = cast_or_null<Function>(ST->lookup(SF->getType(), SF->getName()));
339 if (!DF || SF->hasInternalLinkage() || DF->hasInternalLinkage()) {
340 // Function does not already exist, simply insert an function signature
341 // identical to SF into the dest module...
342 Function *NewDF = new Function(SF->getFunctionType(), SF->getLinkage(),
343 SF->getName(), Dest);
345 // If the LLVM runtime renamed the function, but it is an externally
346 // visible symbol, DF must be an existing function with internal linkage.
348 if (NewDF->getName() != SF->getName() && !NewDF->hasInternalLinkage()) {
349 assert(DF && DF->getName() == SF->getName() &&DF->hasInternalLinkage());
351 NewDF->setName(SF->getName()); // Force the name back
352 DF->setName(SF->getName()); // This will cause a renaming
353 assert(NewDF->getName() == SF->getName() &&
354 DF->getName() != SF->getName());
357 // ... and remember this mapping...
358 ValueMap.insert(std::make_pair(SF, NewDF));
359 } else if (SF->isExternal()) {
360 // If SF is external or if both SF & DF are external.. Just link the
361 // external functions, we aren't adding anything.
362 ValueMap.insert(std::make_pair(SF, DF));
363 } else if (DF->isExternal()) { // If DF is external but SF is not...
364 // Link the external functions, update linkage qualifiers
365 ValueMap.insert(std::make_pair(SF, DF));
366 DF->setLinkage(SF->getLinkage());
368 } else if (SF->getLinkage() != DF->getLinkage()) {
369 return Error(Err, "Functions named '" + SF->getName() +
370 "' have different linkage specifiers!");
371 } else if (SF->hasExternalLinkage()) {
372 // The function is defined in both modules!!
373 return Error(Err, "Function '" +
374 SF->getFunctionType()->getDescription() + "':\"" +
375 SF->getName() + "\" - Function is already defined!");
376 } else if (SF->hasLinkOnceLinkage()) {
377 // Completely ignore the source function.
378 ValueMap.insert(std::make_pair(SF, DF));
380 assert(0 && "Unknown linkage configuration found!");
386 // LinkFunctionBody - Copy the source function over into the dest function and
387 // fix up references to values. At this point we know that Dest is an external
388 // function, and that Src is not.
390 static bool LinkFunctionBody(Function *Dest, const Function *Src,
391 std::map<const Value*, Value*> &GlobalMap,
393 assert(Src && Dest && Dest->isExternal() && !Src->isExternal());
394 std::map<const Value*, Value*> LocalMap; // Map for function local values
396 // Go through and convert function arguments over...
397 Function::aiterator DI = Dest->abegin();
398 for (Function::const_aiterator I = Src->abegin(), E = Src->aend();
400 DI->setName(I->getName()); // Copy the name information over...
402 // Add a mapping to our local map
403 LocalMap.insert(std::make_pair(I, DI));
406 // Loop over all of the basic blocks, copying the instructions over...
408 for (Function::const_iterator I = Src->begin(), E = Src->end(); I != E; ++I) {
409 // Create new basic block and add to mapping and the Dest function...
410 BasicBlock *DBB = new BasicBlock(I->getName(), Dest);
411 LocalMap.insert(std::make_pair(I, DBB));
413 // Loop over all of the instructions in the src basic block, copying them
414 // over. Note that this is broken in a strict sense because the cloned
415 // instructions will still be referencing values in the Src module, not
416 // the remapped values. In our case, however, we will not get caught and
417 // so we can delay patching the values up until later...
419 for (BasicBlock::const_iterator II = I->begin(), IE = I->end();
421 Instruction *DI = II->clone();
422 DI->setName(II->getName());
423 DBB->getInstList().push_back(DI);
424 LocalMap.insert(std::make_pair(II, DI));
428 // At this point, all of the instructions and values of the function are now
429 // copied over. The only problem is that they are still referencing values in
430 // the Source function as operands. Loop through all of the operands of the
431 // functions and patch them up to point to the local versions...
433 for (Function::iterator BB = Dest->begin(), BE = Dest->end(); BB != BE; ++BB)
434 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
435 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
437 *OI = RemapOperand(*OI, LocalMap, &GlobalMap);
443 // LinkFunctionBodies - Link in the function bodies that are defined in the
444 // source module into the DestModule. This consists basically of copying the
445 // function over and fixing up references to values.
447 static bool LinkFunctionBodies(Module *Dest, const Module *Src,
448 std::map<const Value*, Value*> &ValueMap,
451 // Loop over all of the functions in the src module, mapping them over as we
454 for (Module::const_iterator SF = Src->begin(), E = Src->end(); SF != E; ++SF){
455 if (!SF->isExternal()) { // No body if function is external
456 Function *DF = cast<Function>(ValueMap[SF]); // Destination function
458 // DF not external SF external?
459 if (!DF->isExternal()) {
460 if (DF->hasLinkOnceLinkage()) continue; // No relinkage for link-once!
462 *Err = "Function '" + (SF->hasName() ? SF->getName() :std::string(""))
463 + "' body multiply defined!";
467 if (LinkFunctionBody(DF, SF, ValueMap, Err)) return true;
475 // LinkModules - This function links two modules together, with the resulting
476 // left module modified to be the composite of the two input modules. If an
477 // error occurs, true is returned and ErrorMsg (if not null) is set to indicate
478 // the problem. Upon failure, the Dest module could be in a modified state, and
479 // shouldn't be relied on to be consistent.
481 bool LinkModules(Module *Dest, const Module *Src, std::string *ErrorMsg) {
482 if (Dest->getEndianness() != Src->getEndianness())
483 std::cerr << "WARNING: Linking two modules of different endianness!\n";
484 if (Dest->getPointerSize() != Src->getPointerSize())
485 std::cerr << "WARNING: Linking two modules of different pointer size!\n";
487 // LinkTypes - Go through the symbol table of the Src module and see if any
488 // types are named in the src module that are not named in the Dst module.
489 // Make sure there are no type name conflicts.
491 if (LinkTypes(Dest, Src, ErrorMsg)) return true;
493 // ValueMap - Mapping of values from what they used to be in Src, to what they
496 std::map<const Value*, Value*> ValueMap;
498 // Insert all of the globals in src into the Dest module... without
500 if (LinkGlobals(Dest, Src, ValueMap, ErrorMsg)) return true;
502 // Link the functions together between the two modules, without doing function
503 // bodies... this just adds external function prototypes to the Dest
504 // function... We do this so that when we begin processing function bodies,
505 // all of the global values that may be referenced are available in our
508 if (LinkFunctionProtos(Dest, Src, ValueMap, ErrorMsg)) return true;
510 // Update the initializers in the Dest module now that all globals that may
511 // be referenced are in Dest.
513 if (LinkGlobalInits(Dest, Src, ValueMap, ErrorMsg)) return true;
515 // Link in the function bodies that are defined in the source module into the
516 // DestModule. This consists basically of copying the function over and
517 // fixing up references to values.
519 if (LinkFunctionBodies(Dest, Src, ValueMap, ErrorMsg)) return true;