1 //===- FunctionResolution.cpp - Resolve declarations to implementations ---===//
3 // Loop over the functions that are in the module and look for functions that
4 // have the same name. More often than not, there will be things like:
6 // declare void %foo(...)
7 // void %foo(int, int) { ... }
9 // because of the way things are declared in C. If this is the case, patch
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/IPO.h"
15 #include "llvm/Module.h"
16 #include "llvm/SymbolTable.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Pass.h"
19 #include "llvm/iOther.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Assembly/Writer.h" // FIXME: remove when varargs implemented
22 #include "Support/Statistic.h"
26 Statistic<>NumResolved("funcresolve", "Number of varargs functions resolved");
27 Statistic<> NumGlobals("funcresolve", "Number of global variables resolved");
29 struct FunctionResolvingPass : public Pass {
32 RegisterOpt<FunctionResolvingPass> X("funcresolve", "Resolve Functions");
35 Pass *createFunctionResolvingPass() {
36 return new FunctionResolvingPass();
39 // ConvertCallTo - Convert a call to a varargs function with no arg types
40 // specified to a concrete nonvarargs function.
42 static void ConvertCallTo(CallInst *CI, Function *Dest) {
43 const FunctionType::ParamTypes &ParamTys =
44 Dest->getFunctionType()->getParamTypes();
45 BasicBlock *BB = CI->getParent();
47 // Keep an iterator to where we want to insert cast instructions if the
48 // argument types don't agree.
50 BasicBlock::iterator BBI = CI;
51 unsigned NumArgsToCopy = CI->getNumOperands()-1;
52 if (CI->getNumOperands()-1 != ParamTys.size() &&
53 !(CI->getNumOperands()-1 > ParamTys.size() &&
54 Dest->getFunctionType()->isVarArg())) {
55 std::cerr << "WARNING: Call arguments do not match expected number of"
57 std::cerr << "WARNING: In function '"
58 << CI->getParent()->getParent()->getName() << "': call: " << *CI;
59 std::cerr << "Function resolved to: ";
60 WriteAsOperand(std::cerr, Dest);
64 std::vector<Value*> Params;
66 // Convert all of the call arguments over... inserting cast instructions if
67 // the types are not compatible.
68 for (unsigned i = 1; i <= NumArgsToCopy; ++i) {
69 Value *V = CI->getOperand(i);
71 if (i-1 < ParamTys.size() && V->getType() != ParamTys[i-1]) {
72 // Must insert a cast...
73 V = new CastInst(V, ParamTys[i-1], "argcast", BBI);
79 // Replace the old call instruction with a new call instruction that calls
82 Instruction *NewCall = new CallInst(Dest, Params, "", BBI);
84 // Remove the old call instruction from the program...
85 BB->getInstList().remove(BBI);
87 // Transfer the name over...
88 if (NewCall->getType() != Type::VoidTy)
89 NewCall->setName(CI->getName());
91 // Replace uses of the old instruction with the appropriate values...
93 if (NewCall->getType() == CI->getType()) {
94 CI->replaceAllUsesWith(NewCall);
95 NewCall->setName(CI->getName());
97 } else if (NewCall->getType() == Type::VoidTy) {
98 // Resolved function does not return a value but the prototype does. This
99 // often occurs because undefined functions default to returning integers.
100 // Just replace uses of the call (which are broken anyway) with dummy
102 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
103 } else if (CI->getType() == Type::VoidTy) {
104 // If we are gaining a new return value, we don't have to do anything
105 // special here, because it will automatically be ignored.
107 // Insert a cast instruction to convert the return value of the function
108 // into it's new type. Of course we only need to do this if the return
109 // value of the function is actually USED.
111 if (!CI->use_empty()) {
112 // Insert the new cast instruction...
113 CastInst *NewCast = new CastInst(NewCall, CI->getType(),
114 NewCall->getName(), BBI);
115 CI->replaceAllUsesWith(NewCast);
119 // The old instruction is no longer needed, destroy it!
124 static bool ResolveFunctions(Module &M, std::vector<GlobalValue*> &Globals,
125 Function *Concrete) {
126 bool Changed = false;
127 for (unsigned i = 0; i != Globals.size(); ++i)
128 if (Globals[i] != Concrete) {
129 Function *Old = cast<Function>(Globals[i]);
130 const FunctionType *OldMT = Old->getFunctionType();
131 const FunctionType *ConcreteMT = Concrete->getFunctionType();
133 assert(OldMT->getParamTypes().size() <=
134 ConcreteMT->getParamTypes().size() &&
135 "Concrete type must have more specified parameters!");
137 // Check to make sure that if there are specified types, that they
140 for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
141 if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
142 std::cerr << "funcresolve: Function [" << Old->getName()
143 << "]: Parameter types conflict for: '" << OldMT
144 << "' and '" << ConcreteMT << "'\n";
148 // Attempt to convert all of the uses of the old function to the
149 // concrete form of the function. If there is a use of the fn that
150 // we don't understand here we punt to avoid making a bad
153 // At this point, we know that the return values are the same for
154 // our two functions and that the Old function has no varargs fns
155 // specified. In otherwords it's just <retty> (...)
157 for (unsigned i = 0; i < Old->use_size(); ) {
158 User *U = *(Old->use_begin()+i);
159 if (CastInst *CI = dyn_cast<CastInst>(U)) {
160 // Convert casts directly
161 assert(CI->getOperand(0) == Old);
162 CI->setOperand(0, Concrete);
165 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
166 // Can only fix up calls TO the argument, not args passed in.
167 if (CI->getCalledValue() == Old) {
168 ConvertCallTo(CI, Concrete);
172 std::cerr << "Couldn't cleanup this function call, must be an"
173 << " argument or something!" << CI;
177 std::cerr << "Cannot convert use of function: " << U << "\n";
186 static bool ResolveGlobalVariables(Module &M,
187 std::vector<GlobalValue*> &Globals,
188 GlobalVariable *Concrete) {
189 bool Changed = false;
190 assert(isa<ArrayType>(Concrete->getType()->getElementType()) &&
191 "Concrete version should be an array type!");
193 // Get the type of the things that may be resolved to us...
195 cast<ArrayType>(Concrete->getType()->getElementType())->getElementType();
197 std::vector<Constant*> Args;
198 Args.push_back(Constant::getNullValue(Type::LongTy));
199 Args.push_back(Constant::getNullValue(Type::LongTy));
200 ConstantExpr *Replacement =
201 ConstantExpr::getGetElementPtr(ConstantPointerRef::get(Concrete), Args);
203 for (unsigned i = 0; i != Globals.size(); ++i)
204 if (Globals[i] != Concrete) {
205 GlobalVariable *Old = cast<GlobalVariable>(Globals[i]);
206 if (Old->getType()->getElementType() != AETy) {
207 std::cerr << "WARNING: Two global variables exist with the same name "
208 << "that cannot be resolved!\n";
212 // In this case, Old is a pointer to T, Concrete is a pointer to array of
213 // T. Because of this, replace all uses of Old with a constantexpr
214 // getelementptr that returns the address of the first element of the
217 Old->replaceAllUsesWith(Replacement);
218 // Since there are no uses of Old anymore, remove it from the module.
219 M.getGlobalList().erase(Old);
227 static bool ProcessGlobalsWithSameName(Module &M,
228 std::vector<GlobalValue*> &Globals) {
229 assert(!Globals.empty() && "Globals list shouldn't be empty here!");
231 bool isFunction = isa<Function>(Globals[0]); // Is this group all functions?
232 bool Changed = false;
233 GlobalValue *Concrete = 0; // The most concrete implementation to resolve to
235 assert((isFunction ^ isa<GlobalVariable>(Globals[0])) &&
236 "Should either be function or gvar!");
238 for (unsigned i = 0; i != Globals.size(); ) {
239 if (isa<Function>(Globals[i]) != isFunction) {
240 std::cerr << "WARNING: Found function and global variable with the "
241 << "same name: '" << Globals[i]->getName() << "'.\n";
242 return false; // Don't know how to handle this, bail out!
246 // For functions, we look to merge functions definitions of "int (...)"
247 // to 'int (int)' or 'int ()' or whatever else is not completely generic.
249 Function *F = cast<Function>(Globals[i]);
250 if (!F->isExternal()) {
251 if (Concrete && !Concrete->isExternal())
252 return false; // Found two different functions types. Can't choose!
254 Concrete = Globals[i];
255 } else if (Concrete) {
256 if (Concrete->isExternal()) // If we have multiple external symbols...x
257 if (F->getFunctionType()->getNumParams() >
258 cast<Function>(Concrete)->getFunctionType()->getNumParams())
259 Concrete = F; // We are more concrete than "Concrete"!
266 // For global variables, we have to merge C definitions int A[][4] with
268 GlobalVariable *GV = cast<GlobalVariable>(Globals[i]);
270 if (isa<ArrayType>(GV->getType()->getElementType()))
272 } else { // Must have different types... one is an array of the other?
273 const ArrayType *AT =
274 dyn_cast<ArrayType>(GV->getType()->getElementType());
276 // If GV is an array of Concrete, then GV is the array.
277 if (AT && AT->getElementType() == Concrete->getType()->getElementType())
280 // Concrete must be an array type, check to see if the element type of
281 // concrete is already GV.
282 AT = cast<ArrayType>(Concrete->getType()->getElementType());
283 if (AT->getElementType() != GV->getType()->getElementType())
284 Concrete = 0; // Don't know how to handle it!
292 if (Globals.size() > 1) { // Found a multiply defined global...
293 // We should find exactly one concrete function definition, which is
294 // probably the implementation. Change all of the function definitions and
295 // uses to use it instead.
298 std::cerr << "WARNING: Found function types that are not compatible:\n";
299 for (unsigned i = 0; i < Globals.size(); ++i) {
300 std::cerr << "\t" << Globals[i]->getType()->getDescription() << " %"
301 << Globals[i]->getName() << "\n";
303 std::cerr << " No linkage of globals named '" << Globals[0]->getName()
309 return Changed | ResolveFunctions(M, Globals, cast<Function>(Concrete));
311 return Changed | ResolveGlobalVariables(M, Globals,
312 cast<GlobalVariable>(Concrete));
317 bool FunctionResolvingPass::run(Module &M) {
318 SymbolTable &ST = M.getSymbolTable();
320 std::map<std::string, std::vector<GlobalValue*> > Globals;
322 // Loop over the entries in the symbol table. If an entry is a func pointer,
323 // then add it to the Functions map. We do a two pass algorithm here to avoid
324 // problems with iterators getting invalidated if we did a one pass scheme.
326 for (SymbolTable::iterator I = ST.begin(), E = ST.end(); I != E; ++I)
327 if (const PointerType *PT = dyn_cast<PointerType>(I->first)) {
328 SymbolTable::VarMap &Plane = I->second;
329 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
331 GlobalValue *GV = cast<GlobalValue>(PI->second);
332 assert(PI->first == GV->getName() &&
333 "Global name and symbol table do not agree!");
334 if (GV->hasExternalLinkage()) // Only resolve decls to external fns
335 Globals[PI->first].push_back(GV);
339 bool Changed = false;
341 // Now we have a list of all functions with a particular name. If there is
342 // more than one entry in a list, merge the functions together.
344 for (std::map<std::string, std::vector<GlobalValue*> >::iterator
345 I = Globals.begin(), E = Globals.end(); I != E; ++I)
346 Changed |= ProcessGlobalsWithSameName(M, I->second);
348 // Now loop over all of the globals, checking to see if any are trivially
349 // dead. If so, remove them now.
351 for (Module::iterator I = M.begin(), E = M.end(); I != E; )
352 if (I->isExternal() && I->use_empty()) {
355 M.getFunctionList().erase(F);
362 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; )
363 if (I->isExternal() && I->use_empty()) {
364 GlobalVariable *GV = I;
366 M.getGlobalList().erase(GV);