/// CleanupAndPrepareModules - Get the specified modules ready for code /// generator testing. /// static void CleanupAndPrepareModules(BugDriver &BD, Module *&Test, Module *Safe) { // Clean up the modules, removing extra cruft that we don't need anymore... Test = BD.performFinalCleanups(Test); // If we are executing the JIT, we have several nasty issues to take care of. if (!BD.isExecutingJIT()) return; // First, if the main function is in the Safe module, we must add a stub to // the Test module to call into it. Thus, we create a new function `main' // which just calls the old one. if (Function *oldMain = Safe->getFunction("main")) if (!oldMain->isDeclaration()) { // Rename it oldMain->setName("llvm_bugpoint_old_main"); // Create a NEW `main' function with same type in the test module. Function *newMain = Function::Create(oldMain->getFunctionType(), GlobalValue::ExternalLinkage, "main", Test); // Create an `oldmain' prototype in the test module, which will // corresponds to the real main function in the same module. Function *oldMainProto = Function::Create(oldMain->getFunctionType(), GlobalValue::ExternalLinkage, oldMain->getName(), Test); // Set up and remember the argument list for the main function. std::vector<Value*> args; for (Function::arg_iterator I = newMain->arg_begin(), E = newMain->arg_end(), OI = oldMain->arg_begin(); I != E; ++I, ++OI) { I->setName(OI->getName()); // Copy argument names from oldMain args.push_back(I); } // Call the old main function and return its result BasicBlock *BB = BasicBlock::Create(Safe->getContext(), "entry", newMain); CallInst *call = CallInst::Create(oldMainProto, args.begin(), args.end(), "", BB); // If the type of old function wasn't void, return value of call ReturnInst::Create(Safe->getContext(), call, BB); } // The second nasty issue we must deal with in the JIT is that the Safe // module cannot directly reference any functions defined in the test // module. Instead, we use a JIT API call to dynamically resolve the // symbol. // Add the resolver to the Safe module. // Prototype: void *getPointerToNamedFunction(const char* Name) Constant *resolverFunc = Safe->getOrInsertFunction("getPointerToNamedFunction", Type::getInt8PtrTy(Safe->getContext()), Type::getInt8PtrTy(Safe->getContext()), (Type *)0); // Use the function we just added to get addresses of functions we need. for (Module::iterator F = Safe->begin(), E = Safe->end(); F != E; ++F) { if (F->isDeclaration() && !F->use_empty() && &*F != resolverFunc && !F->isIntrinsic() /* ignore intrinsics */) { Function *TestFn = Test->getFunction(F->getName()); // Don't forward functions which are external in the test module too. if (TestFn && !TestFn->isDeclaration()) { // 1. Add a string constant with its name to the global file Constant *InitArray = ConstantArray::get(F->getContext(), F->getName()); GlobalVariable *funcName = new GlobalVariable(*Safe, InitArray->getType(), true /*isConstant*/, GlobalValue::InternalLinkage, InitArray, F->getName() + "_name"); // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an // sbyte* so it matches the signature of the resolver function. // GetElementPtr *funcName, ulong 0, ulong 0 std::vector<Constant*> GEPargs(2, Constant::getNullValue(Type::getInt32Ty(F->getContext()))); Value *GEP = ConstantExpr::getGetElementPtr(funcName, &GEPargs[0], 2); std::vector<Value*> ResolverArgs; ResolverArgs.push_back(GEP); // Rewrite uses of F in global initializers, etc. to uses of a wrapper // function that dynamically resolves the calls to F via our JIT API if (!F->use_empty()) { // Create a new global to hold the cached function pointer. Constant *NullPtr = ConstantPointerNull::get(F->getType()); GlobalVariable *Cache = new GlobalVariable(*F->getParent(), F->getType(), false, GlobalValue::InternalLinkage, NullPtr,F->getName()+".fpcache"); // Construct a new stub function that will re-route calls to F const FunctionType *FuncTy = F->getFunctionType(); Function *FuncWrapper = Function::Create(FuncTy, GlobalValue::InternalLinkage, F->getName() + "_wrapper", F->getParent()); BasicBlock *EntryBB = BasicBlock::Create(F->getContext(), "entry", FuncWrapper); BasicBlock *DoCallBB = BasicBlock::Create(F->getContext(), "usecache", FuncWrapper); BasicBlock *LookupBB = BasicBlock::Create(F->getContext(), "lookupfp", FuncWrapper); // Check to see if we already looked up the value. Value *CachedVal = new LoadInst(Cache, "fpcache", EntryBB); Value *IsNull = new ICmpInst(*EntryBB, ICmpInst::ICMP_EQ, CachedVal, NullPtr, "isNull"); BranchInst::Create(LookupBB, DoCallBB, IsNull, EntryBB); // Resolve the call to function F via the JIT API: // // call resolver(GetElementPtr...) CallInst *Resolver = CallInst::Create(resolverFunc, ResolverArgs.begin(), ResolverArgs.end(), "resolver", LookupBB); // Cast the result from the resolver to correctly-typed function. CastInst *CastedResolver = new BitCastInst(Resolver, PointerType::getUnqual(F->getFunctionType()), "resolverCast", LookupBB); // Save the value in our cache. new StoreInst(CastedResolver, Cache, LookupBB); BranchInst::Create(DoCallBB, LookupBB); PHINode *FuncPtr = PHINode::Create(NullPtr->getType(), "fp", DoCallBB); FuncPtr->addIncoming(CastedResolver, LookupBB); FuncPtr->addIncoming(CachedVal, EntryBB); // Save the argument list. std::vector<Value*> Args; for (Function::arg_iterator i = FuncWrapper->arg_begin(), e = FuncWrapper->arg_end(); i != e; ++i) Args.push_back(i); // Pass on the arguments to the real function, return its result if (F->getReturnType()->isVoidTy()) { CallInst::Create(FuncPtr, Args.begin(), Args.end(), "", DoCallBB); ReturnInst::Create(F->getContext(), DoCallBB); } else { CallInst *Call = CallInst::Create(FuncPtr, Args.begin(), Args.end(), "retval", DoCallBB); ReturnInst::Create(F->getContext(),Call, DoCallBB); } // Use the wrapper function instead of the old function F->replaceAllUsesWith(FuncWrapper); } } } } if (verifyModule(*Test) || verifyModule(*Safe)) { errs() << "Bugpoint has a bug, which corrupted a module!!\n"; abort(); } }
/// ExtractLoops - Given a reduced list of functions that still exposed the bug, /// check to see if we can extract the loops in the region without obscuring the /// bug. If so, it reduces the amount of code identified. /// static bool ExtractLoops(BugDriver &BD, bool (*TestFn)(BugDriver &, Module *, Module *, std::string &), std::vector<Function*> &MiscompiledFunctions, std::string &Error) { bool MadeChange = false; while (1) { if (BugpointIsInterrupted) return MadeChange; DenseMap<const Value*, Value*> ValueMap; Module *ToNotOptimize = CloneModule(BD.getProgram(), ValueMap); Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, MiscompiledFunctions, ValueMap); Module *ToOptimizeLoopExtracted = BD.ExtractLoop(ToOptimize); if (!ToOptimizeLoopExtracted) { // If the loop extractor crashed or if there were no extractible loops, // then this chapter of our odyssey is over with. delete ToNotOptimize; delete ToOptimize; return MadeChange; } errs() << "Extracted a loop from the breaking portion of the program.\n"; // Bugpoint is intentionally not very trusting of LLVM transformations. In // particular, we're not going to assume that the loop extractor works, so // we're going to test the newly loop extracted program to make sure nothing // has broken. If something broke, then we'll inform the user and stop // extraction. AbstractInterpreter *AI = BD.switchToSafeInterpreter(); bool Failure = TestMergedProgram(BD, ToOptimizeLoopExtracted, ToNotOptimize, false, Error); if (!Error.empty()) return false; if (Failure) { BD.switchToInterpreter(AI); // Merged program doesn't work anymore! errs() << " *** ERROR: Loop extraction broke the program. :(" << " Please report a bug!\n"; errs() << " Continuing on with un-loop-extracted version.\n"; BD.writeProgramToFile(OutputPrefix + "-loop-extract-fail-tno.bc", ToNotOptimize); BD.writeProgramToFile(OutputPrefix + "-loop-extract-fail-to.bc", ToOptimize); BD.writeProgramToFile(OutputPrefix + "-loop-extract-fail-to-le.bc", ToOptimizeLoopExtracted); errs() << "Please submit the " << OutputPrefix << "-loop-extract-fail-*.bc files.\n"; delete ToOptimize; delete ToNotOptimize; delete ToOptimizeLoopExtracted; return MadeChange; } delete ToOptimize; BD.switchToInterpreter(AI); outs() << " Testing after loop extraction:\n"; // Clone modules, the tester function will free them. Module *TOLEBackup = CloneModule(ToOptimizeLoopExtracted); Module *TNOBackup = CloneModule(ToNotOptimize); Failure = TestFn(BD, ToOptimizeLoopExtracted, ToNotOptimize, Error); if (!Error.empty()) return false; if (!Failure) { outs() << "*** Loop extraction masked the problem. Undoing.\n"; // If the program is not still broken, then loop extraction did something // that masked the error. Stop loop extraction now. delete TOLEBackup; delete TNOBackup; return MadeChange; } ToOptimizeLoopExtracted = TOLEBackup; ToNotOptimize = TNOBackup; outs() << "*** Loop extraction successful!\n"; std::vector<std::pair<std::string, const FunctionType*> > MisCompFunctions; for (Module::iterator I = ToOptimizeLoopExtracted->begin(), E = ToOptimizeLoopExtracted->end(); I != E; ++I) if (!I->isDeclaration()) MisCompFunctions.push_back(std::make_pair(I->getName(), I->getFunctionType())); // Okay, great! Now we know that we extracted a loop and that loop // extraction both didn't break the program, and didn't mask the problem. // Replace the current program with the loop extracted version, and try to // extract another loop. std::string ErrorMsg; if (Linker::LinkModules(ToNotOptimize, ToOptimizeLoopExtracted, &ErrorMsg)){ errs() << BD.getToolName() << ": Error linking modules together:" << ErrorMsg << '\n'; exit(1); } delete ToOptimizeLoopExtracted; // All of the Function*'s in the MiscompiledFunctions list are in the old // module. Update this list to include all of the functions in the // optimized and loop extracted module. MiscompiledFunctions.clear(); for (unsigned i = 0, e = MisCompFunctions.size(); i != e; ++i) { Function *NewF = ToNotOptimize->getFunction(MisCompFunctions[i].first); assert(NewF && "Function not found??"); assert(NewF->getFunctionType() == MisCompFunctions[i].second && "found wrong function type?"); MiscompiledFunctions.push_back(NewF); } BD.setNewProgram(ToNotOptimize); MadeChange = true; } }
/// ExtractBlocks - Given a reduced list of functions that still expose the bug, /// extract as many basic blocks from the region as possible without obscuring /// the bug. /// static bool ExtractBlocks(BugDriver &BD, bool (*TestFn)(BugDriver &, Module *, Module *, std::string &), std::vector<Function*> &MiscompiledFunctions, std::string &Error) { if (BugpointIsInterrupted) return false; std::vector<BasicBlock*> Blocks; for (unsigned i = 0, e = MiscompiledFunctions.size(); i != e; ++i) for (Function::iterator I = MiscompiledFunctions[i]->begin(), E = MiscompiledFunctions[i]->end(); I != E; ++I) Blocks.push_back(I); // Use the list reducer to identify blocks that can be extracted without // obscuring the bug. The Blocks list will end up containing blocks that must // be retained from the original program. unsigned OldSize = Blocks.size(); // Check to see if all blocks are extractible first. bool Ret = ReduceMiscompiledBlocks(BD, TestFn, MiscompiledFunctions) .TestFuncs(std::vector<BasicBlock*>(), Error); if (!Error.empty()) return false; if (Ret) { Blocks.clear(); } else { ReduceMiscompiledBlocks(BD, TestFn, MiscompiledFunctions).reduceList(Blocks, Error); if (!Error.empty()) return false; if (Blocks.size() == OldSize) return false; } DenseMap<const Value*, Value*> ValueMap; Module *ProgClone = CloneModule(BD.getProgram(), ValueMap); Module *ToExtract = SplitFunctionsOutOfModule(ProgClone, MiscompiledFunctions, ValueMap); Module *Extracted = BD.ExtractMappedBlocksFromModule(Blocks, ToExtract); if (Extracted == 0) { // Weird, extraction should have worked. errs() << "Nondeterministic problem extracting blocks??\n"; delete ProgClone; delete ToExtract; return false; } // Otherwise, block extraction succeeded. Link the two program fragments back // together. delete ToExtract; std::vector<std::pair<std::string, const FunctionType*> > MisCompFunctions; for (Module::iterator I = Extracted->begin(), E = Extracted->end(); I != E; ++I) if (!I->isDeclaration()) MisCompFunctions.push_back(std::make_pair(I->getName(), I->getFunctionType())); std::string ErrorMsg; if (Linker::LinkModules(ProgClone, Extracted, &ErrorMsg)) { errs() << BD.getToolName() << ": Error linking modules together:" << ErrorMsg << '\n'; exit(1); } delete Extracted; // Set the new program and delete the old one. BD.setNewProgram(ProgClone); // Update the list of miscompiled functions. MiscompiledFunctions.clear(); for (unsigned i = 0, e = MisCompFunctions.size(); i != e; ++i) { Function *NewF = ProgClone->getFunction(MisCompFunctions[i].first); assert(NewF && "Function not found??"); assert(NewF->getFunctionType() == MisCompFunctions[i].second && "Function has wrong type??"); MiscompiledFunctions.push_back(NewF); } return true; }
/*! * This method identify which is value sym and which is object sym */ void SymbolTableInfo::buildMemModel(llvm::Module& module) { analysisUtil::increaseStackSize(); prePassSchedule(module); mod = &module; maxFieldLimit = maxFieldNumLimit; // Object #0 is black hole the object that may point to any object assert(totalSymNum == BlackHole && "Something changed!"); symTyMap.insert(std::make_pair(totalSymNum++, BlackHole)); createBlkOrConstantObj(BlackHole); // Object #1 always represents the constant assert(totalSymNum == ConstantObj && "Something changed!"); symTyMap.insert(std::make_pair(totalSymNum++, ConstantObj)); createBlkOrConstantObj(ConstantObj); // Pointer #2 always represents the pointer points-to black hole. assert(totalSymNum == BlkPtr && "Something changed!"); symTyMap.insert(std::make_pair(totalSymNum++, BlkPtr)); // Pointer #3 always represents the null pointer. assert(totalSymNum == NullPtr && "Something changed!"); symTyMap.insert(std::make_pair(totalSymNum, NullPtr)); // Add symbols for all the globals . for (Module::global_iterator I = module.global_begin(), E = module.global_end(); I != E; ++I) { collectSym(&*I); } // Add symbols for all the global aliases for (Module::alias_iterator I = module.alias_begin(), E = module.alias_end(); I != E; I++) { collectSym(&*I); } // Add symbols for all of the functions and the instructions in them. for (Module::iterator F = module.begin(), E = module.end(); F != E; ++F) { collectSym(&*F); collectRet(&*F); if (F->getFunctionType()->isVarArg()) collectVararg(&*F); // Add symbols for all formal parameters. for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) { collectSym(&*I); } // collect and create symbols inside the function body for (inst_iterator II = inst_begin(&*F), E = inst_end(&*F); II != E; ++II) { const Instruction *inst = &*II; collectSym(inst); // initialization for some special instructions //{@ if (const StoreInst *st = dyn_cast<StoreInst>(inst)) { collectSym(st->getPointerOperand()); collectSym(st->getValueOperand()); } else if (const LoadInst *ld = dyn_cast<LoadInst>(inst)) { collectSym(ld->getPointerOperand()); } else if (const PHINode *phi = dyn_cast<PHINode>(inst)) { for (u32_t i = 0; i < phi->getNumIncomingValues(); ++i) { collectSym(phi->getIncomingValue(i)); } } else if (const GetElementPtrInst *gep = dyn_cast<GetElementPtrInst>( inst)) { collectSym(gep->getPointerOperand()); } else if (const SelectInst *sel = dyn_cast<SelectInst>(inst)) { collectSym(sel->getTrueValue()); collectSym(sel->getFalseValue()); } else if (const CastInst *cast = dyn_cast<CastInst>(inst)) { collectSym(cast->getOperand(0)); } else if (const ReturnInst *ret = dyn_cast<ReturnInst>(inst)) { if(ret->getReturnValue()) collectSym(ret->getReturnValue()); } else if (isCallSite(inst) && isInstrinsicDbgInst(inst)==false) { CallSite cs = analysisUtil::getLLVMCallSite(inst); callSiteSet.insert(cs); for (CallSite::arg_iterator it = cs.arg_begin(); it != cs.arg_end(); ++it) { collectSym(*it); } // Calls to inline asm need to be added as well because the callee isn't // referenced anywhere else. const Value *Callee = cs.getCalledValue(); collectSym(Callee); //TODO handle inlineAsm ///if (isa<InlineAsm>(Callee)) } //@} } } }