//===- FunctionResolution.cpp - Resolve declarations to implementations ---===//
//
+// The LLVM Compiler Infrastructure
+//
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
// Loop over the functions that are in the module and look for functions that
// have the same name. More often than not, there will be things like:
//
#include "llvm/Transforms/IPO.h"
#include "llvm/Module.h"
-#include "llvm/SymbolTable.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Pass.h"
-#include "llvm/iOther.h"
-#include "llvm/Constant.h"
-#include "Support/StatisticReporter.h"
+#include "llvm/Instructions.h"
+#include "llvm/Constants.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/ADT/Statistic.h"
#include <algorithm>
-
-using std::vector;
-using std::string;
-using std::cerr;
+using namespace llvm;
namespace {
- Statistic<>NumResolved("funcresolve\t- Number of varargs functions resolved");
+ Statistic NumResolved("funcresolve", "Number of varargs functions resolved");
+ Statistic NumGlobals("funcresolve", "Number of global variables resolved");
- struct FunctionResolvingPass : public Pass {
- bool run(Module &M);
+ struct FunctionResolvingPass : public ModulePass {
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<TargetData>();
+ }
+
+ bool runOnModule(Module &M);
};
- RegisterOpt<FunctionResolvingPass> X("funcresolve", "Resolve Functions");
+ RegisterPass<FunctionResolvingPass> X("funcresolve", "Resolve Functions");
}
-Pass *createFunctionResolvingPass() {
+ModulePass *llvm::createFunctionResolvingPass() {
return new FunctionResolvingPass();
}
-// ConvertCallTo - Convert a call to a varargs function with no arg types
-// specified to a concrete nonvarargs function.
-//
-static void ConvertCallTo(CallInst *CI, Function *Dest) {
- const FunctionType::ParamTypes &ParamTys =
- Dest->getFunctionType()->getParamTypes();
- BasicBlock *BB = CI->getParent();
+static bool ResolveFunctions(Module &M, std::vector<GlobalValue*> &Globals,
+ Function *Concrete) {
+ bool Changed = false;
+ for (unsigned i = 0; i != Globals.size(); ++i)
+ if (Globals[i] != Concrete) {
+ Function *Old = cast<Function>(Globals[i]);
+ const FunctionType *OldFT = Old->getFunctionType();
+ const FunctionType *ConcreteFT = Concrete->getFunctionType();
- // Keep an iterator to where we want to insert cast instructions if the
- // argument types don't agree.
- //
- BasicBlock::iterator BBI = CI;
- assert(CI->getNumOperands()-1 == ParamTys.size() &&
- "Function calls resolved funny somehow, incompatible number of args");
+ if (OldFT->getNumParams() > ConcreteFT->getNumParams() &&
+ !ConcreteFT->isVarArg())
+ if (!Old->use_empty()) {
+ cerr << "WARNING: Linking function '" << Old->getName()
+ << "' is causing arguments to be dropped.\n";
+ cerr << "WARNING: Prototype: ";
+ WriteAsOperand(*cerr.stream(), Old);
+ cerr << " resolved to ";
+ WriteAsOperand(*cerr.stream(), Concrete);
+ cerr << "\n";
+ }
- vector<Value*> Params;
+ // Check to make sure that if there are specified types, that they
+ // match...
+ //
+ unsigned NumArguments = std::min(OldFT->getNumParams(),
+ ConcreteFT->getNumParams());
- // Convert all of the call arguments over... inserting cast instructions if
- // the types are not compatible.
- for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
- Value *V = CI->getOperand(i);
+ if (!Old->use_empty() && !Concrete->use_empty())
+ for (unsigned i = 0; i < NumArguments; ++i)
+ if (OldFT->getParamType(i) != ConcreteFT->getParamType(i))
+ if (OldFT->getParamType(i)->getTypeID() !=
+ ConcreteFT->getParamType(i)->getTypeID()) {
+ cerr << "WARNING: Function [" << Old->getName()
+ << "]: Parameter types conflict for: '";
+ WriteTypeSymbolic(*cerr.stream(), OldFT, &M);
+ cerr << "' (in "
+ << Old->getParent()->getModuleIdentifier() << ") and '";
+ WriteTypeSymbolic(*cerr.stream(), ConcreteFT, &M);
+ cerr << "'(in "
+ << Concrete->getParent()->getModuleIdentifier() << ")\n";
+ return Changed;
+ }
- if (V->getType() != ParamTys[i-1]) // Must insert a cast...
- V = new CastInst(V, ParamTys[i-1], "argcast", BBI);
+ // Attempt to convert all of the uses of the old function to the concrete
+ // form of the function. If there is a use of the fn that we don't
+ // understand here we punt to avoid making a bad transformation.
+ //
+ // At this point, we know that the return values are the same for our two
+ // functions and that the Old function has no varargs fns specified. In
+ // otherwords it's just <retty> (...)
+ //
+ if (!Old->use_empty()) {
+ Value *Replacement = Concrete;
+ if (Concrete->getType() != Old->getType())
+ Replacement = ConstantExpr::getBitCast(Concrete, Old->getType());
+ NumResolved += Old->getNumUses();
+ Old->replaceAllUsesWith(Replacement);
+ }
- Params.push_back(V);
- }
+ // Since there are no uses of Old anymore, remove it from the module.
+ M.getFunctionList().erase(Old);
+ }
+ return Changed;
+}
- // Replace the old call instruction with a new call instruction that calls
- // the real function.
- //
- Instruction *NewCall = new CallInst(Dest, Params, "", BBI);
- // Remove the old call instruction from the program...
- BB->getInstList().remove(BBI);
+static bool ResolveGlobalVariables(Module &M,
+ std::vector<GlobalValue*> &Globals,
+ GlobalVariable *Concrete) {
+ bool Changed = false;
- // Transfer the name over...
- if (NewCall->getType() != Type::VoidTy)
- NewCall->setName(CI->getName());
+ for (unsigned i = 0; i != Globals.size(); ++i)
+ if (Globals[i] != Concrete) {
+ Constant *Cast = ConstantExpr::getBitCast(Concrete,Globals[i]->getType());
+ Globals[i]->replaceAllUsesWith(Cast);
- // Replace uses of the old instruction with the appropriate values...
- //
- if (NewCall->getType() == CI->getType()) {
- CI->replaceAllUsesWith(NewCall);
- NewCall->setName(CI->getName());
-
- } else if (NewCall->getType() == Type::VoidTy) {
- // Resolved function does not return a value but the prototype does. This
- // often occurs because undefined functions default to returning integers.
- // Just replace uses of the call (which are broken anyway) with dummy
- // values.
- CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
- } else if (CI->getType() == Type::VoidTy) {
- // If we are gaining a new return value, we don't have to do anything
- // special here, because it will automatically be ignored.
- } else {
- // Insert a cast instruction to convert the return value of the function
- // into it's new type. Of course we only need to do this if the return
- // value of the function is actually USED.
- //
- if (!CI->use_empty()) {
- // Insert the new cast instruction...
- CastInst *NewCast = new CastInst(NewCall, CI->getType(),
- NewCall->getName(), BBI);
- CI->replaceAllUsesWith(NewCast);
+ // Since there are no uses of Old anymore, remove it from the module.
+ M.getGlobalList().erase(cast<GlobalVariable>(Globals[i]));
+
+ ++NumGlobals;
+ Changed = true;
}
- }
+ return Changed;
+}
- // The old instruction is no longer needed, destroy it!
- delete CI;
+// Check to see if all of the callers of F ignore the return value.
+static bool CallersAllIgnoreReturnValue(Function &F) {
+ if (F.getReturnType() == Type::VoidTy) return true;
+ for (Value::use_iterator I = F.use_begin(), E = F.use_end(); I != E; ++I) {
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(*I)) {
+ for (Value::use_iterator I = GV->use_begin(), E = GV->use_end();
+ I != E; ++I) {
+ CallSite CS = CallSite::get(*I);
+ if (!CS.getInstruction() || !CS.getInstruction()->use_empty())
+ return false;
+ }
+ } else {
+ CallSite CS = CallSite::get(*I);
+ if (!CS.getInstruction() || !CS.getInstruction()->use_empty())
+ return false;
+ }
+ }
+ return true;
}
+static bool ProcessGlobalsWithSameName(Module &M, TargetData &TD,
+ std::vector<GlobalValue*> &Globals) {
+ assert(!Globals.empty() && "Globals list shouldn't be empty here!");
-bool FunctionResolvingPass::run(Module &M) {
- SymbolTable *ST = M.getSymbolTable();
- if (!ST) return false;
+ bool isFunction = isa<Function>(Globals[0]); // Is this group all functions?
+ GlobalValue *Concrete = 0; // The most concrete implementation to resolve to
- std::map<string, vector<Function*> > Functions;
+ for (unsigned i = 0; i != Globals.size(); ) {
+ if (isa<Function>(Globals[i]) != isFunction) {
+ cerr << "WARNING: Found function and global variable with the "
+ << "same name: '" << Globals[i]->getName() << "'.\n";
+ return false; // Don't know how to handle this, bail out!
+ }
- // Loop over the entries in the symbol table. If an entry is a func pointer,
- // then add it to the Functions map. We do a two pass algorithm here to avoid
- // problems with iterators getting invalidated if we did a one pass scheme.
- //
- for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
- if (const PointerType *PT = dyn_cast<PointerType>(I->first))
- if (isa<FunctionType>(PT->getElementType())) {
- SymbolTable::VarMap &Plane = I->second;
- for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
- PI != PE; ++PI) {
- Function *F = cast<Function>(PI->second);
- assert(PI->first == F->getName() &&
- "Function name and symbol table do not agree!");
- if (F->hasExternalLinkage()) // Only resolve decls to external fns
- Functions[PI->first].push_back(F);
- }
- }
+ if (isFunction) {
+ // For functions, we look to merge functions definitions of "int (...)"
+ // to 'int (int)' or 'int ()' or whatever else is not completely generic.
+ //
+ Function *F = cast<Function>(Globals[i]);
+ if (!F->isExternal()) {
+ if (Concrete && !Concrete->isExternal())
+ return false; // Found two different functions types. Can't choose!
- bool Changed = false;
+ Concrete = Globals[i];
+ } else if (Concrete) {
+ if (Concrete->isExternal()) // If we have multiple external symbols...
+ if (F->getFunctionType()->getNumParams() >
+ cast<Function>(Concrete)->getFunctionType()->getNumParams())
+ Concrete = F; // We are more concrete than "Concrete"!
- // Now we have a list of all functions with a particular name. If there is
- // more than one entry in a list, merge the functions together.
- //
- for (std::map<string, vector<Function*> >::iterator I = Functions.begin(),
- E = Functions.end(); I != E; ++I) {
- vector<Function*> &Functions = I->second;
- Function *Implementation = 0; // Find the implementation
- Function *Concrete = 0;
- for (unsigned i = 0; i < Functions.size(); ) {
- if (!Functions[i]->isExternal()) { // Found an implementation
- if (Implementation != 0)
- assert(Implementation == 0 && "Multiple definitions of the same"
- " function. Case not handled yet!");
- Implementation = Functions[i];
} else {
- // Ignore functions that are never used so they don't cause spurious
- // warnings... here we will actually DCE the function so that it isn't
- // used later.
- //
- if (Functions[i]->use_empty()) {
- M.getFunctionList().erase(Functions[i]);
- Functions.erase(Functions.begin()+i);
- Changed = true;
- ++NumResolved;
- continue;
- }
+ Concrete = F;
}
-
- if (Functions[i] && (!Functions[i]->getFunctionType()->isVarArg())) {
- if (Concrete) { // Found two different functions types. Can't choose
- Concrete = 0;
- break;
+ } else {
+ GlobalVariable *GV = cast<GlobalVariable>(Globals[i]);
+ if (!GV->isExternal()) {
+ if (Concrete) {
+ cerr << "WARNING: Two global variables with external linkage"
+ << " exist with the same name: '" << GV->getName()
+ << "'!\n";
+ return false;
}
- Concrete = Functions[i];
+ Concrete = GV;
}
- ++i;
}
+ ++i;
+ }
- if (Functions.size() > 1) { // Found a multiply defined function...
- // We should find exactly one non-vararg function definition, which is
- // probably the implementation. Change all of the function definitions
- // and uses to use it instead.
- //
- if (!Concrete) {
- cerr << "Warning: Found functions types that are not compatible:\n";
- for (unsigned i = 0; i < Functions.size(); ++i) {
- cerr << "\t" << Functions[i]->getType()->getDescription() << " %"
- << Functions[i]->getName() << "\n";
+ if (Globals.size() > 1) { // Found a multiply defined global...
+ // If there are no external declarations, and there is at most one
+ // externally visible instance of the global, then there is nothing to do.
+ //
+ bool HasExternal = false;
+ unsigned NumInstancesWithExternalLinkage = 0;
+
+ for (unsigned i = 0, e = Globals.size(); i != e; ++i) {
+ if (Globals[i]->isExternal())
+ HasExternal = true;
+ else if (!Globals[i]->hasInternalLinkage())
+ NumInstancesWithExternalLinkage++;
+ }
+
+ if (!HasExternal && NumInstancesWithExternalLinkage <= 1)
+ return false; // Nothing to do? Must have multiple internal definitions.
+
+ // There are a couple of special cases we don't want to print the warning
+ // for, check them now.
+ bool DontPrintWarning = false;
+ if (Concrete && Globals.size() == 2) {
+ GlobalValue *Other = Globals[Globals[0] == Concrete];
+ // If the non-concrete global is a function which takes (...) arguments,
+ // and the return values match (or was never used), do not warn.
+ if (Function *ConcreteF = dyn_cast<Function>(Concrete))
+ if (Function *OtherF = dyn_cast<Function>(Other))
+ if ((ConcreteF->getReturnType() == OtherF->getReturnType() ||
+ CallersAllIgnoreReturnValue(*OtherF)) &&
+ OtherF->getFunctionType()->isVarArg() &&
+ OtherF->getFunctionType()->getNumParams() == 0)
+ DontPrintWarning = true;
+
+ // Otherwise, if the non-concrete global is a global array variable with a
+ // size of 0, and the concrete global is an array with a real size, don't
+ // warn. This occurs due to declaring 'extern int A[];'.
+ if (GlobalVariable *ConcreteGV = dyn_cast<GlobalVariable>(Concrete))
+ if (GlobalVariable *OtherGV = dyn_cast<GlobalVariable>(Other)) {
+ const Type *CTy = ConcreteGV->getType();
+ const Type *OTy = OtherGV->getType();
+
+ if (CTy->isSized())
+ if (!OTy->isSized() || !TD.getTypeSize(OTy) ||
+ TD.getTypeSize(OTy) == TD.getTypeSize(CTy))
+ DontPrintWarning = true;
}
- cerr << " No linkage of functions named '" << Functions[0]->getName()
- << "' performed!\n";
- } else {
- for (unsigned i = 0; i < Functions.size(); ++i)
- if (Functions[i] != Concrete) {
- Function *Old = Functions[i];
- const FunctionType *OldMT = Old->getFunctionType();
- const FunctionType *ConcreteMT = Concrete->getFunctionType();
- bool Broken = false;
-
- assert(OldMT->getParamTypes().size() <=
- ConcreteMT->getParamTypes().size() &&
- "Concrete type must have more specified parameters!");
-
- // Check to make sure that if there are specified types, that they
- // match...
- //
- for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
- if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
- cerr << "Parameter types conflict for" << OldMT
- << " and " << ConcreteMT;
- Broken = true;
- }
- if (Broken) break; // Can't process this one!
-
-
- // Attempt to convert all of the uses of the old function to the
- // concrete form of the function. If there is a use of the fn that
- // we don't understand here we punt to avoid making a bad
- // transformation.
- //
- // At this point, we know that the return values are the same for
- // our two functions and that the Old function has no varargs fns
- // specified. In otherwords it's just <retty> (...)
- //
- for (unsigned i = 0; i < Old->use_size(); ) {
- User *U = *(Old->use_begin()+i);
- if (CastInst *CI = dyn_cast<CastInst>(U)) {
- // Convert casts directly
- assert(CI->getOperand(0) == Old);
- CI->setOperand(0, Concrete);
- Changed = true;
- ++NumResolved;
- } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
- // Can only fix up calls TO the argument, not args passed in.
- if (CI->getCalledValue() == Old) {
- ConvertCallTo(CI, Concrete);
- Changed = true;
- ++NumResolved;
- } else {
- cerr << "Couldn't cleanup this function call, must be an"
- << " argument or something!" << CI;
- ++i;
- }
- } else {
- cerr << "Cannot convert use of function: " << U << "\n";
- ++i;
- }
- }
+ }
+
+ if (0 && !DontPrintWarning) {
+ cerr << "WARNING: Found global types that are not compatible:\n";
+ for (unsigned i = 0; i < Globals.size(); ++i) {
+ cerr << "\t";
+ WriteTypeSymbolic(*cerr.stream(), Globals[i]->getType(), &M);
+ cerr << " %" << Globals[i]->getName() << "\n";
+ }
+ }
+
+ if (!Concrete)
+ Concrete = Globals[0];
+ else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Concrete)) {
+ // Handle special case hack to change globals if it will make their types
+ // happier in the long run. The situation we do this is intentionally
+ // extremely limited.
+ if (GV->use_empty() && GV->hasInitializer() &&
+ GV->getInitializer()->isNullValue()) {
+ // Check to see if there is another (external) global with the same size
+ // and a non-empty use-list. If so, we will make IT be the real
+ // implementation.
+ unsigned TS = TD.getTypeSize(Concrete->getType()->getElementType());
+ for (unsigned i = 0, e = Globals.size(); i != e; ++i)
+ if (Globals[i] != Concrete && !Globals[i]->use_empty() &&
+ isa<GlobalVariable>(Globals[i]) &&
+ TD.getTypeSize(Globals[i]->getType()->getElementType()) == TS) {
+ // At this point we want to replace Concrete with Globals[i]. Make
+ // concrete external, and Globals[i] have an initializer.
+ GlobalVariable *NGV = cast<GlobalVariable>(Globals[i]);
+ const Type *ElTy = NGV->getType()->getElementType();
+ NGV->setInitializer(Constant::getNullValue(ElTy));
+ cast<GlobalVariable>(Concrete)->setInitializer(0);
+ Concrete = NGV;
+ break;
}
- }
+ }
}
+
+ if (isFunction)
+ return ResolveFunctions(M, Globals, cast<Function>(Concrete));
+ else
+ return ResolveGlobalVariables(M, Globals,
+ cast<GlobalVariable>(Concrete));
+ }
+ return false;
+}
+
+bool FunctionResolvingPass::runOnModule(Module &M) {
+ std::map<std::string, std::vector<GlobalValue*> > Globals;
+
+ // Loop over the globals, adding them to the Globals map. We use a two pass
+ // algorithm here to avoid problems with iterators getting invalidated if we
+ // did a one pass scheme.
+ //
+ bool Changed = false;
+ for (Module::iterator I = M.begin(), E = M.end(); I != E; ) {
+ Function *F = I++;
+ if (F->use_empty() && F->isExternal()) {
+ M.getFunctionList().erase(F);
+ Changed = true;
+ } else if (!F->hasInternalLinkage() && !F->getName().empty() &&
+ !F->getIntrinsicID())
+ Globals[F->getName()].push_back(F);
}
+ for (Module::global_iterator I = M.global_begin(), E = M.global_end();
+ I != E; ) {
+ GlobalVariable *GV = I++;
+ if (GV->use_empty() && GV->isExternal()) {
+ M.getGlobalList().erase(GV);
+ Changed = true;
+ } else if (!GV->hasInternalLinkage() && !GV->getName().empty())
+ Globals[GV->getName()].push_back(GV);
+ }
+
+ TargetData &TD = getAnalysis<TargetData>();
+
+ // Now we have a list of all functions with a particular name. If there is
+ // more than one entry in a list, merge the functions together.
+ //
+ for (std::map<std::string, std::vector<GlobalValue*> >::iterator
+ I = Globals.begin(), E = Globals.end(); I != E; ++I)
+ Changed |= ProcessGlobalsWithSameName(M, TD, I->second);
+
+ // Now loop over all of the globals, checking to see if any are trivially
+ // dead. If so, remove them now.
+
+ for (Module::iterator I = M.begin(), E = M.end(); I != E; )
+ if (I->isExternal() && I->use_empty()) {
+ Function *F = I;
+ ++I;
+ M.getFunctionList().erase(F);
+ ++NumResolved;
+ Changed = true;
+ } else {
+ ++I;
+ }
+
+ for (Module::global_iterator I = M.global_begin(), E = M.global_end();
+ I != E; )
+ if (I->isExternal() && I->use_empty()) {
+ GlobalVariable *GV = I;
+ ++I;
+ M.getGlobalList().erase(GV);
+ ++NumGlobals;
+ Changed = true;
+ } else {
+ ++I;
+ }
+
return Changed;
}