142418abaSMehdi Amini //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===// 242418abaSMehdi Amini // 342418abaSMehdi Amini // The LLVM Compiler Infrastructure 442418abaSMehdi Amini // 542418abaSMehdi Amini // This file is distributed under the University of Illinois Open Source 642418abaSMehdi Amini // License. See LICENSE.TXT for details. 742418abaSMehdi Amini // 842418abaSMehdi Amini //===----------------------------------------------------------------------===// 942418abaSMehdi Amini // 1042418abaSMehdi Amini // This file implements Function import based on summaries. 1142418abaSMehdi Amini // 1242418abaSMehdi Amini //===----------------------------------------------------------------------===// 1342418abaSMehdi Amini 1442418abaSMehdi Amini #include "llvm/Transforms/IPO/FunctionImport.h" 1542418abaSMehdi Amini 1642418abaSMehdi Amini #include "llvm/ADT/StringSet.h" 1742418abaSMehdi Amini #include "llvm/IR/AutoUpgrade.h" 1842418abaSMehdi Amini #include "llvm/IR/DiagnosticPrinter.h" 1942418abaSMehdi Amini #include "llvm/IR/IntrinsicInst.h" 2042418abaSMehdi Amini #include "llvm/IR/Module.h" 2142418abaSMehdi Amini #include "llvm/IRReader/IRReader.h" 2242418abaSMehdi Amini #include "llvm/Linker/Linker.h" 2342418abaSMehdi Amini #include "llvm/Object/FunctionIndexObjectFile.h" 2442418abaSMehdi Amini #include "llvm/Support/CommandLine.h" 2542418abaSMehdi Amini #include "llvm/Support/Debug.h" 2642418abaSMehdi Amini #include "llvm/Support/SourceMgr.h" 2742418abaSMehdi Amini using namespace llvm; 2842418abaSMehdi Amini 2942418abaSMehdi Amini #define DEBUG_TYPE "function-import" 3042418abaSMehdi Amini 3139303619STeresa Johnson /// Limit on instruction count of imported functions. 3239303619STeresa Johnson static cl::opt<unsigned> ImportInstrLimit( 3339303619STeresa Johnson "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"), 3439303619STeresa Johnson cl::desc("Only import functions with less than N instructions")); 3539303619STeresa Johnson 3642418abaSMehdi Amini // Load lazily a module from \p FileName in \p Context. 3742418abaSMehdi Amini static std::unique_ptr<Module> loadFile(const std::string &FileName, 3842418abaSMehdi Amini LLVMContext &Context) { 3942418abaSMehdi Amini SMDiagnostic Err; 4042418abaSMehdi Amini DEBUG(dbgs() << "Loading '" << FileName << "'\n"); 4142418abaSMehdi Amini std::unique_ptr<Module> Result = getLazyIRFileModule(FileName, Err, Context); 4242418abaSMehdi Amini if (!Result) { 4342418abaSMehdi Amini Err.print("function-import", errs()); 4442418abaSMehdi Amini return nullptr; 4542418abaSMehdi Amini } 4642418abaSMehdi Amini 4742418abaSMehdi Amini Result->materializeMetadata(); 4842418abaSMehdi Amini UpgradeDebugInfo(*Result); 4942418abaSMehdi Amini 5042418abaSMehdi Amini return Result; 5142418abaSMehdi Amini } 5242418abaSMehdi Amini 5342418abaSMehdi Amini // Get a Module for \p FileName from the cache, or load it lazily. 54a11bdc8eSMehdi Amini Module &ModuleLazyLoaderCache::operator()(StringRef FileName) { 5542418abaSMehdi Amini auto &Module = ModuleMap[FileName]; 5642418abaSMehdi Amini if (!Module) 5742418abaSMehdi Amini Module = loadFile(FileName, Context); 5842418abaSMehdi Amini return *Module; 5942418abaSMehdi Amini } 6042418abaSMehdi Amini 61d450da32STeresa Johnson /// Walk through the instructions in \p F looking for external 62d450da32STeresa Johnson /// calls not already in the \p CalledFunctions set. If any are 63d450da32STeresa Johnson /// found they are added to the \p Worklist for importing. 64d450da32STeresa Johnson static void findExternalCalls(const Function &F, StringSet<> &CalledFunctions, 65d450da32STeresa Johnson SmallVector<StringRef, 64> &Worklist) { 66d450da32STeresa Johnson for (auto &BB : F) { 67d450da32STeresa Johnson for (auto &I : BB) { 68d450da32STeresa Johnson if (isa<CallInst>(I)) { 69d450da32STeresa Johnson auto CalledFunction = cast<CallInst>(I).getCalledFunction(); 70d450da32STeresa Johnson // Insert any new external calls that have not already been 71d450da32STeresa Johnson // added to set/worklist. 72d450da32STeresa Johnson if (CalledFunction && CalledFunction->hasName() && 73d450da32STeresa Johnson CalledFunction->isDeclaration() && 74d450da32STeresa Johnson !CalledFunctions.count(CalledFunction->getName())) { 75d450da32STeresa Johnson CalledFunctions.insert(CalledFunction->getName()); 76d450da32STeresa Johnson Worklist.push_back(CalledFunction->getName()); 77d450da32STeresa Johnson } 78d450da32STeresa Johnson } 79d450da32STeresa Johnson } 80d450da32STeresa Johnson } 81d450da32STeresa Johnson } 82d450da32STeresa Johnson 83c8c55170SMehdi Amini // Helper function: given a worklist and an index, will process all the worklist 84c8c55170SMehdi Amini // and import them based on the summary information 85311fef6eSMehdi Amini static unsigned ProcessImportWorklist( 86311fef6eSMehdi Amini Module &DestModule, SmallVector<StringRef, 64> &Worklist, 87311fef6eSMehdi Amini StringSet<> &CalledFunctions, Linker &TheLinker, 88311fef6eSMehdi Amini const FunctionInfoIndex &Index, 89c8c55170SMehdi Amini std::function<Module &(StringRef FileName)> &LazyModuleLoader) { 90c8c55170SMehdi Amini unsigned ImportCount = 0; 9142418abaSMehdi Amini while (!Worklist.empty()) { 9242418abaSMehdi Amini auto CalledFunctionName = Worklist.pop_back_val(); 9342418abaSMehdi Amini DEBUG(dbgs() << "Process import for " << CalledFunctionName << "\n"); 9442418abaSMehdi Amini 9542418abaSMehdi Amini // Try to get a summary for this function call. 9642418abaSMehdi Amini auto InfoList = Index.findFunctionInfoList(CalledFunctionName); 9742418abaSMehdi Amini if (InfoList == Index.end()) { 9842418abaSMehdi Amini DEBUG(dbgs() << "No summary for " << CalledFunctionName 9942418abaSMehdi Amini << " Ignoring.\n"); 10042418abaSMehdi Amini continue; 10142418abaSMehdi Amini } 10242418abaSMehdi Amini assert(!InfoList->second.empty() && "No summary, error at import?"); 10342418abaSMehdi Amini 10442418abaSMehdi Amini // Comdat can have multiple entries, FIXME: what do we do with them? 10542418abaSMehdi Amini auto &Info = InfoList->second[0]; 10642418abaSMehdi Amini assert(Info && "Nullptr in list, error importing summaries?\n"); 10742418abaSMehdi Amini 10842418abaSMehdi Amini auto *Summary = Info->functionSummary(); 10942418abaSMehdi Amini if (!Summary) { 11042418abaSMehdi Amini // FIXME: in case we are lazyloading summaries, we can do it now. 111430110ccSTeresa Johnson DEBUG(dbgs() << "Missing summary for " << CalledFunctionName 112430110ccSTeresa Johnson << ", error at import?\n"); 11342418abaSMehdi Amini llvm_unreachable("Missing summary"); 11442418abaSMehdi Amini } 11542418abaSMehdi Amini 11639303619STeresa Johnson if (Summary->instCount() > ImportInstrLimit) { 117430110ccSTeresa Johnson DEBUG(dbgs() << "Skip import of " << CalledFunctionName << " with " 11839303619STeresa Johnson << Summary->instCount() << " instructions (limit " 119430110ccSTeresa Johnson << ImportInstrLimit << ")\n"); 12039303619STeresa Johnson continue; 12139303619STeresa Johnson } 12239303619STeresa Johnson 12342418abaSMehdi Amini // Get the module path from the summary. 12442418abaSMehdi Amini auto FileName = Summary->modulePath(); 12542418abaSMehdi Amini DEBUG(dbgs() << "Importing " << CalledFunctionName << " from " << FileName 12642418abaSMehdi Amini << "\n"); 12742418abaSMehdi Amini 12842418abaSMehdi Amini // Get the module for the import (potentially from the cache). 129c8c55170SMehdi Amini auto &Module = LazyModuleLoader(FileName); 130c8c55170SMehdi Amini assert(&Module.getContext() == &DestModule.getContext()); 13142418abaSMehdi Amini 13242418abaSMehdi Amini // The function that we will import! 13342418abaSMehdi Amini GlobalValue *SGV = Module.getNamedValue(CalledFunctionName); 134130de7afSTeresa Johnson StringRef ImportFunctionName = CalledFunctionName; 135130de7afSTeresa Johnson if (!SGV) { 136c8c55170SMehdi Amini // Might be local in source Module, promoted/renamed in DestModule. 137130de7afSTeresa Johnson std::pair<StringRef, StringRef> Split = 138130de7afSTeresa Johnson CalledFunctionName.split(".llvm."); 139130de7afSTeresa Johnson SGV = Module.getNamedValue(Split.first); 140130de7afSTeresa Johnson #ifndef NDEBUG 141130de7afSTeresa Johnson // Assert that Split.second is module id 142130de7afSTeresa Johnson uint64_t ModuleId; 143130de7afSTeresa Johnson assert(!Split.second.getAsInteger(10, ModuleId)); 144130de7afSTeresa Johnson assert(ModuleId == Index.getModuleId(FileName)); 145130de7afSTeresa Johnson #endif 146130de7afSTeresa Johnson } 14742418abaSMehdi Amini Function *F = dyn_cast<Function>(SGV); 14842418abaSMehdi Amini if (!F && isa<GlobalAlias>(SGV)) { 14942418abaSMehdi Amini auto *SGA = dyn_cast<GlobalAlias>(SGV); 15042418abaSMehdi Amini F = dyn_cast<Function>(SGA->getBaseObject()); 151130de7afSTeresa Johnson ImportFunctionName = F->getName(); 15242418abaSMehdi Amini } 15342418abaSMehdi Amini if (!F) { 15442418abaSMehdi Amini errs() << "Can't load function '" << CalledFunctionName << "' in Module '" 15542418abaSMehdi Amini << FileName << "', error in the summary?\n"; 15642418abaSMehdi Amini llvm_unreachable("Can't load function in Module"); 15742418abaSMehdi Amini } 15842418abaSMehdi Amini 15917626654STeresa Johnson // We cannot import weak_any functions/aliases without possibly affecting 16017626654STeresa Johnson // the order they are seen and selected by the linker, changing program 16142418abaSMehdi Amini // semantics. 16217626654STeresa Johnson if (SGV->hasWeakAnyLinkage()) { 16317626654STeresa Johnson DEBUG(dbgs() << "Ignoring import request for weak-any " 16417626654STeresa Johnson << (isa<Function>(SGV) ? "function " : "alias ") 16542418abaSMehdi Amini << CalledFunctionName << " from " << FileName << "\n"); 16642418abaSMehdi Amini continue; 16742418abaSMehdi Amini } 16842418abaSMehdi Amini 16942418abaSMehdi Amini // Link in the specified function. 170ffe2e4aaSMehdi Amini DenseSet<const GlobalValue *> FunctionsToImport; 171ffe2e4aaSMehdi Amini FunctionsToImport.insert(F); 172c8c55170SMehdi Amini if (TheLinker.linkInModule(Module, Linker::Flags::None, &Index, 173ffe2e4aaSMehdi Amini &FunctionsToImport)) 17442418abaSMehdi Amini report_fatal_error("Function Import: link error"); 17542418abaSMehdi Amini 176130de7afSTeresa Johnson // Process the newly imported function and add callees to the worklist. 177c8c55170SMehdi Amini GlobalValue *NewGV = DestModule.getNamedValue(ImportFunctionName); 178130de7afSTeresa Johnson assert(NewGV); 179130de7afSTeresa Johnson Function *NewF = dyn_cast<Function>(NewGV); 180130de7afSTeresa Johnson assert(NewF); 181d450da32STeresa Johnson findExternalCalls(*NewF, CalledFunctions, Worklist); 182c8c55170SMehdi Amini ++ImportCount; 183c8c55170SMehdi Amini } 184c8c55170SMehdi Amini return ImportCount; 18542418abaSMehdi Amini } 186ffe2e4aaSMehdi Amini 187c8c55170SMehdi Amini // Automatically import functions in Module \p DestModule based on the summaries 188c8c55170SMehdi Amini // index. 189c8c55170SMehdi Amini // 190c8c55170SMehdi Amini // The current implementation imports every called functions that exists in the 191c8c55170SMehdi Amini // summaries index. 192c8c55170SMehdi Amini bool FunctionImporter::importFunctions(Module &DestModule) { 193311fef6eSMehdi Amini DEBUG(errs() << "Starting import for Module " 194311fef6eSMehdi Amini << DestModule.getModuleIdentifier() << "\n"); 195c8c55170SMehdi Amini unsigned ImportedCount = 0; 196c8c55170SMehdi Amini 197c8c55170SMehdi Amini /// First step is collecting the called external functions. 198c8c55170SMehdi Amini StringSet<> CalledFunctions; 199c8c55170SMehdi Amini SmallVector<StringRef, 64> Worklist; 200c8c55170SMehdi Amini for (auto &F : DestModule) { 201c8c55170SMehdi Amini if (F.isDeclaration() || F.hasFnAttribute(Attribute::OptimizeNone)) 202c8c55170SMehdi Amini continue; 203c8c55170SMehdi Amini findExternalCalls(F, CalledFunctions, Worklist); 204c8c55170SMehdi Amini } 205c8c55170SMehdi Amini if (Worklist.empty()) 206c8c55170SMehdi Amini return false; 207c8c55170SMehdi Amini 208c8c55170SMehdi Amini /// Second step: for every call to an external function, try to import it. 209c8c55170SMehdi Amini 210c8c55170SMehdi Amini // Linker that will be used for importing function 211c8c55170SMehdi Amini Linker TheLinker(DestModule, DiagnosticHandler); 212c8c55170SMehdi Amini 213311fef6eSMehdi Amini ImportedCount += ProcessImportWorklist(DestModule, Worklist, CalledFunctions, 214311fef6eSMehdi Amini TheLinker, Index, getLazyModule); 215c8c55170SMehdi Amini 216c8c55170SMehdi Amini DEBUG(errs() << "Imported " << ImportedCount << " functions for Module " 217c8c55170SMehdi Amini << DestModule.getModuleIdentifier() << "\n"); 218c8c55170SMehdi Amini return ImportedCount; 21942418abaSMehdi Amini } 22042418abaSMehdi Amini 22142418abaSMehdi Amini /// Summary file to use for function importing when using -function-import from 22242418abaSMehdi Amini /// the command line. 22342418abaSMehdi Amini static cl::opt<std::string> 22442418abaSMehdi Amini SummaryFile("summary-file", 22542418abaSMehdi Amini cl::desc("The summary file to use for function importing.")); 22642418abaSMehdi Amini 22742418abaSMehdi Amini static void diagnosticHandler(const DiagnosticInfo &DI) { 22842418abaSMehdi Amini raw_ostream &OS = errs(); 22942418abaSMehdi Amini DiagnosticPrinterRawOStream DP(OS); 23042418abaSMehdi Amini DI.print(DP); 23142418abaSMehdi Amini OS << '\n'; 23242418abaSMehdi Amini } 23342418abaSMehdi Amini 23442418abaSMehdi Amini /// Parse the function index out of an IR file and return the function 23542418abaSMehdi Amini /// index object if found, or nullptr if not. 23642418abaSMehdi Amini static std::unique_ptr<FunctionInfoIndex> 23742418abaSMehdi Amini getFunctionIndexForFile(StringRef Path, std::string &Error, 23842418abaSMehdi Amini DiagnosticHandlerFunction DiagnosticHandler) { 23942418abaSMehdi Amini std::unique_ptr<MemoryBuffer> Buffer; 24042418abaSMehdi Amini ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr = 24142418abaSMehdi Amini MemoryBuffer::getFile(Path); 24242418abaSMehdi Amini if (std::error_code EC = BufferOrErr.getError()) { 24342418abaSMehdi Amini Error = EC.message(); 24442418abaSMehdi Amini return nullptr; 24542418abaSMehdi Amini } 24642418abaSMehdi Amini Buffer = std::move(BufferOrErr.get()); 24742418abaSMehdi Amini ErrorOr<std::unique_ptr<object::FunctionIndexObjectFile>> ObjOrErr = 24842418abaSMehdi Amini object::FunctionIndexObjectFile::create(Buffer->getMemBufferRef(), 24942418abaSMehdi Amini DiagnosticHandler); 25042418abaSMehdi Amini if (std::error_code EC = ObjOrErr.getError()) { 25142418abaSMehdi Amini Error = EC.message(); 25242418abaSMehdi Amini return nullptr; 25342418abaSMehdi Amini } 25442418abaSMehdi Amini return (*ObjOrErr)->takeIndex(); 25542418abaSMehdi Amini } 25642418abaSMehdi Amini 25742418abaSMehdi Amini /// Pass that performs cross-module function import provided a summary file. 25842418abaSMehdi Amini class FunctionImportPass : public ModulePass { 259*5fcbdb71STeresa Johnson /// Optional function summary index to use for importing, otherwise 260*5fcbdb71STeresa Johnson /// the summary-file option must be specified. 261*5fcbdb71STeresa Johnson FunctionInfoIndex *Index; 26242418abaSMehdi Amini 26342418abaSMehdi Amini public: 26442418abaSMehdi Amini /// Pass identification, replacement for typeid 26542418abaSMehdi Amini static char ID; 26642418abaSMehdi Amini 267*5fcbdb71STeresa Johnson /// Specify pass name for debug output 268*5fcbdb71STeresa Johnson const char *getPassName() const override { 269*5fcbdb71STeresa Johnson return "Function Importing"; 270*5fcbdb71STeresa Johnson } 271*5fcbdb71STeresa Johnson 272*5fcbdb71STeresa Johnson explicit FunctionImportPass(FunctionInfoIndex *Index = nullptr) 273*5fcbdb71STeresa Johnson : ModulePass(ID), Index(Index) {} 27442418abaSMehdi Amini 27542418abaSMehdi Amini bool runOnModule(Module &M) override { 276*5fcbdb71STeresa Johnson if (SummaryFile.empty() && !Index) 277*5fcbdb71STeresa Johnson report_fatal_error("error: -function-import requires -summary-file or " 278*5fcbdb71STeresa Johnson "file from frontend\n"); 279*5fcbdb71STeresa Johnson std::unique_ptr<FunctionInfoIndex> IndexPtr; 280*5fcbdb71STeresa Johnson if (!SummaryFile.empty()) { 281*5fcbdb71STeresa Johnson if (Index) 282*5fcbdb71STeresa Johnson report_fatal_error("error: -summary-file and index from frontend\n"); 28342418abaSMehdi Amini std::string Error; 284*5fcbdb71STeresa Johnson IndexPtr = getFunctionIndexForFile(SummaryFile, Error, diagnosticHandler); 285*5fcbdb71STeresa Johnson if (!IndexPtr) { 286*5fcbdb71STeresa Johnson errs() << "Error loading file '" << SummaryFile << "': " << Error 287*5fcbdb71STeresa Johnson << "\n"; 28842418abaSMehdi Amini return false; 28942418abaSMehdi Amini } 290*5fcbdb71STeresa Johnson Index = IndexPtr.get(); 291*5fcbdb71STeresa Johnson } 29242418abaSMehdi Amini 29342418abaSMehdi Amini // Perform the import now. 294a11bdc8eSMehdi Amini ModuleLazyLoaderCache Loader(M.getContext()); 295a11bdc8eSMehdi Amini FunctionImporter Importer(*Index, diagnosticHandler, 296a11bdc8eSMehdi Amini [&](StringRef Name) 297a11bdc8eSMehdi Amini -> Module &{ return Loader(Name); }); 29842418abaSMehdi Amini return Importer.importFunctions(M); 29942418abaSMehdi Amini 30042418abaSMehdi Amini return false; 30142418abaSMehdi Amini } 30242418abaSMehdi Amini }; 30342418abaSMehdi Amini 30442418abaSMehdi Amini char FunctionImportPass::ID = 0; 30542418abaSMehdi Amini INITIALIZE_PASS_BEGIN(FunctionImportPass, "function-import", 30642418abaSMehdi Amini "Summary Based Function Import", false, false) 30742418abaSMehdi Amini INITIALIZE_PASS_END(FunctionImportPass, "function-import", 30842418abaSMehdi Amini "Summary Based Function Import", false, false) 30942418abaSMehdi Amini 31042418abaSMehdi Amini namespace llvm { 311*5fcbdb71STeresa Johnson Pass *createFunctionImportPass(FunctionInfoIndex *Index = nullptr) { 312*5fcbdb71STeresa Johnson return new FunctionImportPass(Index); 313*5fcbdb71STeresa Johnson } 31442418abaSMehdi Amini } 315