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. 54*a11bdc8eSMehdi 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 DEBUG(dbgs() << "Found a call: '" << I << "'\n"); 70d450da32STeresa Johnson auto CalledFunction = cast<CallInst>(I).getCalledFunction(); 71d450da32STeresa Johnson // Insert any new external calls that have not already been 72d450da32STeresa Johnson // added to set/worklist. 73d450da32STeresa Johnson if (CalledFunction && CalledFunction->hasName() && 74d450da32STeresa Johnson CalledFunction->isDeclaration() && 75d450da32STeresa Johnson !CalledFunctions.count(CalledFunction->getName())) { 76d450da32STeresa Johnson CalledFunctions.insert(CalledFunction->getName()); 77d450da32STeresa Johnson Worklist.push_back(CalledFunction->getName()); 78d450da32STeresa Johnson } 79d450da32STeresa Johnson } 80d450da32STeresa Johnson } 81d450da32STeresa Johnson } 82d450da32STeresa Johnson } 83d450da32STeresa Johnson 8442418abaSMehdi Amini // Automatically import functions in Module \p M based on the summaries index. 8542418abaSMehdi Amini // 8642418abaSMehdi Amini // The current implementation imports every called functions that exists in the 8742418abaSMehdi Amini // summaries index. 8842418abaSMehdi Amini bool FunctionImporter::importFunctions(Module &M) { 8942418abaSMehdi Amini 9042418abaSMehdi Amini bool Changed = false; 9142418abaSMehdi Amini 92d450da32STeresa Johnson /// First step is collecting the called external functions. 9342418abaSMehdi Amini StringSet<> CalledFunctions; 94d450da32STeresa Johnson SmallVector<StringRef, 64> Worklist; 9542418abaSMehdi Amini for (auto &F : M) { 9642418abaSMehdi Amini if (F.isDeclaration() || F.hasFnAttribute(Attribute::OptimizeNone)) 9742418abaSMehdi Amini continue; 98d450da32STeresa Johnson findExternalCalls(F, CalledFunctions, Worklist); 9942418abaSMehdi Amini } 10042418abaSMehdi Amini 10142418abaSMehdi Amini /// Second step: for every call to an external function, try to import it. 10242418abaSMehdi Amini 10342418abaSMehdi Amini // Linker that will be used for importing function 1040e309fe8SRafael Espindola Linker L(M, DiagnosticHandler); 10542418abaSMehdi Amini 10642418abaSMehdi Amini while (!Worklist.empty()) { 10742418abaSMehdi Amini auto CalledFunctionName = Worklist.pop_back_val(); 10842418abaSMehdi Amini DEBUG(dbgs() << "Process import for " << CalledFunctionName << "\n"); 10942418abaSMehdi Amini 11042418abaSMehdi Amini // Try to get a summary for this function call. 11142418abaSMehdi Amini auto InfoList = Index.findFunctionInfoList(CalledFunctionName); 11242418abaSMehdi Amini if (InfoList == Index.end()) { 11342418abaSMehdi Amini DEBUG(dbgs() << "No summary for " << CalledFunctionName 11442418abaSMehdi Amini << " Ignoring.\n"); 11542418abaSMehdi Amini continue; 11642418abaSMehdi Amini } 11742418abaSMehdi Amini assert(!InfoList->second.empty() && "No summary, error at import?"); 11842418abaSMehdi Amini 11942418abaSMehdi Amini // Comdat can have multiple entries, FIXME: what do we do with them? 12042418abaSMehdi Amini auto &Info = InfoList->second[0]; 12142418abaSMehdi Amini assert(Info && "Nullptr in list, error importing summaries?\n"); 12242418abaSMehdi Amini 12342418abaSMehdi Amini auto *Summary = Info->functionSummary(); 12442418abaSMehdi Amini if (!Summary) { 12542418abaSMehdi Amini // FIXME: in case we are lazyloading summaries, we can do it now. 126430110ccSTeresa Johnson DEBUG(dbgs() << "Missing summary for " << CalledFunctionName 127430110ccSTeresa Johnson << ", error at import?\n"); 12842418abaSMehdi Amini llvm_unreachable("Missing summary"); 12942418abaSMehdi Amini } 13042418abaSMehdi Amini 13139303619STeresa Johnson if (Summary->instCount() > ImportInstrLimit) { 132430110ccSTeresa Johnson DEBUG(dbgs() << "Skip import of " << CalledFunctionName << " with " 13339303619STeresa Johnson << Summary->instCount() << " instructions (limit " 134430110ccSTeresa Johnson << ImportInstrLimit << ")\n"); 13539303619STeresa Johnson continue; 13639303619STeresa Johnson } 13739303619STeresa Johnson 13842418abaSMehdi Amini // Get the module path from the summary. 13942418abaSMehdi Amini auto FileName = Summary->modulePath(); 14042418abaSMehdi Amini DEBUG(dbgs() << "Importing " << CalledFunctionName << " from " << FileName 14142418abaSMehdi Amini << "\n"); 14242418abaSMehdi Amini 14342418abaSMehdi Amini // Get the module for the import (potentially from the cache). 144*a11bdc8eSMehdi Amini auto &Module = getLazyModule(FileName); 145*a11bdc8eSMehdi Amini assert(&Module.getContext() == &M.getContext()); 14642418abaSMehdi Amini 14742418abaSMehdi Amini // The function that we will import! 14842418abaSMehdi Amini GlobalValue *SGV = Module.getNamedValue(CalledFunctionName); 149130de7afSTeresa Johnson StringRef ImportFunctionName = CalledFunctionName; 150130de7afSTeresa Johnson if (!SGV) { 151130de7afSTeresa Johnson // Might be local in source Module, promoted/renamed in dest Module M. 152130de7afSTeresa Johnson std::pair<StringRef, StringRef> Split = 153130de7afSTeresa Johnson CalledFunctionName.split(".llvm."); 154130de7afSTeresa Johnson SGV = Module.getNamedValue(Split.first); 155130de7afSTeresa Johnson #ifndef NDEBUG 156130de7afSTeresa Johnson // Assert that Split.second is module id 157130de7afSTeresa Johnson uint64_t ModuleId; 158130de7afSTeresa Johnson assert(!Split.second.getAsInteger(10, ModuleId)); 159130de7afSTeresa Johnson assert(ModuleId == Index.getModuleId(FileName)); 160130de7afSTeresa Johnson #endif 161130de7afSTeresa Johnson } 16242418abaSMehdi Amini Function *F = dyn_cast<Function>(SGV); 16342418abaSMehdi Amini if (!F && isa<GlobalAlias>(SGV)) { 16442418abaSMehdi Amini auto *SGA = dyn_cast<GlobalAlias>(SGV); 16542418abaSMehdi Amini F = dyn_cast<Function>(SGA->getBaseObject()); 166130de7afSTeresa Johnson ImportFunctionName = F->getName(); 16742418abaSMehdi Amini } 16842418abaSMehdi Amini if (!F) { 16942418abaSMehdi Amini errs() << "Can't load function '" << CalledFunctionName << "' in Module '" 17042418abaSMehdi Amini << FileName << "', error in the summary?\n"; 17142418abaSMehdi Amini llvm_unreachable("Can't load function in Module"); 17242418abaSMehdi Amini } 17342418abaSMehdi Amini 17417626654STeresa Johnson // We cannot import weak_any functions/aliases without possibly affecting 17517626654STeresa Johnson // the order they are seen and selected by the linker, changing program 17642418abaSMehdi Amini // semantics. 17717626654STeresa Johnson if (SGV->hasWeakAnyLinkage()) { 17817626654STeresa Johnson DEBUG(dbgs() << "Ignoring import request for weak-any " 17917626654STeresa Johnson << (isa<Function>(SGV) ? "function " : "alias ") 18042418abaSMehdi Amini << CalledFunctionName << " from " << FileName << "\n"); 18142418abaSMehdi Amini continue; 18242418abaSMehdi Amini } 18342418abaSMehdi Amini 18442418abaSMehdi Amini // Link in the specified function. 1850e309fe8SRafael Espindola if (L.linkInModule(Module, Linker::Flags::None, &Index, F)) 18642418abaSMehdi Amini report_fatal_error("Function Import: link error"); 18742418abaSMehdi Amini 188130de7afSTeresa Johnson // Process the newly imported function and add callees to the worklist. 189130de7afSTeresa Johnson GlobalValue *NewGV = M.getNamedValue(ImportFunctionName); 190130de7afSTeresa Johnson assert(NewGV); 191130de7afSTeresa Johnson Function *NewF = dyn_cast<Function>(NewGV); 192130de7afSTeresa Johnson assert(NewF); 193d450da32STeresa Johnson findExternalCalls(*NewF, CalledFunctions, Worklist); 19442418abaSMehdi Amini 19542418abaSMehdi Amini Changed = true; 19642418abaSMehdi Amini } 19742418abaSMehdi Amini return Changed; 19842418abaSMehdi Amini } 19942418abaSMehdi Amini 20042418abaSMehdi Amini /// Summary file to use for function importing when using -function-import from 20142418abaSMehdi Amini /// the command line. 20242418abaSMehdi Amini static cl::opt<std::string> 20342418abaSMehdi Amini SummaryFile("summary-file", 20442418abaSMehdi Amini cl::desc("The summary file to use for function importing.")); 20542418abaSMehdi Amini 20642418abaSMehdi Amini static void diagnosticHandler(const DiagnosticInfo &DI) { 20742418abaSMehdi Amini raw_ostream &OS = errs(); 20842418abaSMehdi Amini DiagnosticPrinterRawOStream DP(OS); 20942418abaSMehdi Amini DI.print(DP); 21042418abaSMehdi Amini OS << '\n'; 21142418abaSMehdi Amini } 21242418abaSMehdi Amini 21342418abaSMehdi Amini /// Parse the function index out of an IR file and return the function 21442418abaSMehdi Amini /// index object if found, or nullptr if not. 21542418abaSMehdi Amini static std::unique_ptr<FunctionInfoIndex> 21642418abaSMehdi Amini getFunctionIndexForFile(StringRef Path, std::string &Error, 21742418abaSMehdi Amini DiagnosticHandlerFunction DiagnosticHandler) { 21842418abaSMehdi Amini std::unique_ptr<MemoryBuffer> Buffer; 21942418abaSMehdi Amini ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr = 22042418abaSMehdi Amini MemoryBuffer::getFile(Path); 22142418abaSMehdi Amini if (std::error_code EC = BufferOrErr.getError()) { 22242418abaSMehdi Amini Error = EC.message(); 22342418abaSMehdi Amini return nullptr; 22442418abaSMehdi Amini } 22542418abaSMehdi Amini Buffer = std::move(BufferOrErr.get()); 22642418abaSMehdi Amini ErrorOr<std::unique_ptr<object::FunctionIndexObjectFile>> ObjOrErr = 22742418abaSMehdi Amini object::FunctionIndexObjectFile::create(Buffer->getMemBufferRef(), 22842418abaSMehdi Amini DiagnosticHandler); 22942418abaSMehdi Amini if (std::error_code EC = ObjOrErr.getError()) { 23042418abaSMehdi Amini Error = EC.message(); 23142418abaSMehdi Amini return nullptr; 23242418abaSMehdi Amini } 23342418abaSMehdi Amini return (*ObjOrErr)->takeIndex(); 23442418abaSMehdi Amini } 23542418abaSMehdi Amini 23642418abaSMehdi Amini /// Pass that performs cross-module function import provided a summary file. 23742418abaSMehdi Amini class FunctionImportPass : public ModulePass { 23842418abaSMehdi Amini 23942418abaSMehdi Amini public: 24042418abaSMehdi Amini /// Pass identification, replacement for typeid 24142418abaSMehdi Amini static char ID; 24242418abaSMehdi Amini 24342418abaSMehdi Amini explicit FunctionImportPass() : ModulePass(ID) {} 24442418abaSMehdi Amini 24542418abaSMehdi Amini bool runOnModule(Module &M) override { 24642418abaSMehdi Amini if (SummaryFile.empty()) { 24742418abaSMehdi Amini report_fatal_error("error: -function-import requires -summary-file\n"); 24842418abaSMehdi Amini } 24942418abaSMehdi Amini std::string Error; 25042418abaSMehdi Amini std::unique_ptr<FunctionInfoIndex> Index = 25142418abaSMehdi Amini getFunctionIndexForFile(SummaryFile, Error, diagnosticHandler); 25242418abaSMehdi Amini if (!Index) { 25342418abaSMehdi Amini errs() << "Error loading file '" << SummaryFile << "': " << Error << "\n"; 25442418abaSMehdi Amini return false; 25542418abaSMehdi Amini } 25642418abaSMehdi Amini 25742418abaSMehdi Amini // Perform the import now. 258*a11bdc8eSMehdi Amini ModuleLazyLoaderCache Loader(M.getContext()); 259*a11bdc8eSMehdi Amini FunctionImporter Importer(*Index, diagnosticHandler, 260*a11bdc8eSMehdi Amini [&](StringRef Name) 261*a11bdc8eSMehdi Amini -> Module &{ return Loader(Name); }); 26242418abaSMehdi Amini return Importer.importFunctions(M); 26342418abaSMehdi Amini 26442418abaSMehdi Amini return false; 26542418abaSMehdi Amini } 26642418abaSMehdi Amini }; 26742418abaSMehdi Amini 26842418abaSMehdi Amini char FunctionImportPass::ID = 0; 26942418abaSMehdi Amini INITIALIZE_PASS_BEGIN(FunctionImportPass, "function-import", 27042418abaSMehdi Amini "Summary Based Function Import", false, false) 27142418abaSMehdi Amini INITIALIZE_PASS_END(FunctionImportPass, "function-import", 27242418abaSMehdi Amini "Summary Based Function Import", false, false) 27342418abaSMehdi Amini 27442418abaSMehdi Amini namespace llvm { 27542418abaSMehdi Amini Pass *createFunctionImportPass() { return new FunctionImportPass(); } 27642418abaSMehdi Amini } 277