1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements Function import based on summaries.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/IPO/FunctionImport.h"
16 #include "llvm/ADT/StringSet.h"
17 #include "llvm/IR/AutoUpgrade.h"
18 #include "llvm/IR/DiagnosticPrinter.h"
19 #include "llvm/IR/IntrinsicInst.h"
20 #include "llvm/IR/Module.h"
21 #include "llvm/IRReader/IRReader.h"
22 #include "llvm/Linker/Linker.h"
23 #include "llvm/Object/FunctionIndexObjectFile.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/SourceMgr.h"
32 #define DEBUG_TYPE "function-import"
34 /// Limit on instruction count of imported functions.
35 static cl::opt<unsigned> ImportInstrLimit(
36 "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
37 cl::desc("Only import functions with less than N instructions"));
39 // Load lazily a module from \p FileName in \p Context.
40 static std::unique_ptr<Module> loadFile(const std::string &FileName,
41 LLVMContext &Context) {
43 DEBUG(dbgs() << "Loading '" << FileName << "'\n");
44 // Metadata isn't loaded or linked until after all functions are
45 // imported, after which it will be materialized and linked.
46 std::unique_ptr<Module> Result =
47 getLazyIRFileModule(FileName, Err, Context,
48 /* ShouldLazyLoadMetadata = */ true);
50 Err.print("function-import", errs());
58 /// Helper to load on demand a Module from file and cache it for subsequent
59 /// queries. It can be used with the FunctionImporter.
60 class ModuleLazyLoaderCache {
61 /// Cache of lazily loaded module for import.
62 StringMap<std::unique_ptr<Module>> ModuleMap;
64 /// Retrieve a Module from the cache or lazily load it on demand.
65 std::function<std::unique_ptr<Module>(StringRef FileName)> createLazyModule;
68 /// Create the loader, Module will be initialized in \p Context.
69 ModuleLazyLoaderCache(std::function<
70 std::unique_ptr<Module>(StringRef FileName)> createLazyModule)
71 : createLazyModule(createLazyModule) {}
73 /// Retrieve a Module from the cache or lazily load it on demand.
74 Module &operator()(StringRef FileName);
76 std::unique_ptr<Module> takeModule(StringRef FileName) {
77 auto I = ModuleMap.find(FileName);
78 assert(I != ModuleMap.end());
79 std::unique_ptr<Module> Ret = std::move(I->second);
85 // Get a Module for \p FileName from the cache, or load it lazily.
86 Module &ModuleLazyLoaderCache::operator()(StringRef Identifier) {
87 auto &Module = ModuleMap[Identifier];
89 Module = createLazyModule(Identifier);
92 } // anonymous namespace
94 /// Walk through the instructions in \p F looking for external
95 /// calls not already in the \p CalledFunctions set. If any are
96 /// found they are added to the \p Worklist for importing.
97 static void findExternalCalls(const Module &DestModule, Function &F,
98 const FunctionInfoIndex &Index,
99 StringSet<> &CalledFunctions,
100 SmallVector<StringRef, 64> &Worklist) {
101 // We need to suffix internal function calls imported from other modules,
102 // prepare the suffix ahead of time.
104 if (F.getParent() != &DestModule)
107 Twine(Index.getModuleId(F.getParent()->getModuleIdentifier()))).str();
111 if (isa<CallInst>(I)) {
112 auto CalledFunction = cast<CallInst>(I).getCalledFunction();
113 // Insert any new external calls that have not already been
114 // added to set/worklist.
115 if (!CalledFunction || !CalledFunction->hasName())
117 // Ignore intrinsics early
118 if (CalledFunction->isIntrinsic()) {
119 assert(CalledFunction->getIntrinsicID() != 0);
122 auto ImportedName = CalledFunction->getName();
123 auto Renamed = (ImportedName + Suffix).str();
124 // Rename internal functions
125 if (CalledFunction->hasInternalLinkage()) {
126 ImportedName = Renamed;
128 auto It = CalledFunctions.insert(ImportedName);
130 // This is a call to a function we already considered, skip.
133 // Ignore functions already present in the destination module
134 auto *SrcGV = DestModule.getNamedValue(ImportedName);
136 assert(isa<Function>(SrcGV) && "Name collision during import");
137 if (!cast<Function>(SrcGV)->isDeclaration()) {
138 DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Ignoring "
139 << ImportedName << " already in DestinationModule\n");
144 Worklist.push_back(It.first->getKey());
145 DEBUG(dbgs() << DestModule.getModuleIdentifier()
146 << ": Adding callee for : " << ImportedName << " : "
147 << F.getName() << "\n");
153 // Helper function: given a worklist and an index, will process all the worklist
154 // and decide what to import based on the summary information.
156 // Nothing is actually imported, functions are materialized in their source
157 // module and analyzed there.
159 // \p ModuleToFunctionsToImportMap is filled with the set of Function to import
161 static void GetImportList(Module &DestModule,
162 SmallVector<StringRef, 64> &Worklist,
163 StringSet<> &CalledFunctions,
164 std::map<StringRef, DenseSet<const GlobalValue *>>
165 &ModuleToFunctionsToImportMap,
166 const FunctionInfoIndex &Index,
167 ModuleLazyLoaderCache &ModuleLoaderCache) {
168 while (!Worklist.empty()) {
169 auto CalledFunctionName = Worklist.pop_back_val();
170 DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Process import for "
171 << CalledFunctionName << "\n");
173 // Try to get a summary for this function call.
174 auto InfoList = Index.findFunctionInfoList(CalledFunctionName);
175 if (InfoList == Index.end()) {
176 DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": No summary for "
177 << CalledFunctionName << " Ignoring.\n");
180 assert(!InfoList->second.empty() && "No summary, error at import?");
182 // Comdat can have multiple entries, FIXME: what do we do with them?
183 auto &Info = InfoList->second[0];
184 assert(Info && "Nullptr in list, error importing summaries?\n");
186 auto *Summary = Info->functionSummary();
188 // FIXME: in case we are lazyloading summaries, we can do it now.
189 DEBUG(dbgs() << DestModule.getModuleIdentifier()
190 << ": Missing summary for " << CalledFunctionName
191 << ", error at import?\n");
192 llvm_unreachable("Missing summary");
195 if (Summary->instCount() > ImportInstrLimit) {
196 DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Skip import of "
197 << CalledFunctionName << " with " << Summary->instCount()
198 << " instructions (limit " << ImportInstrLimit << ")\n");
202 // Get the module path from the summary.
203 auto ModuleIdentifier = Summary->modulePath();
204 DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Importing "
205 << CalledFunctionName << " from " << ModuleIdentifier << "\n");
207 auto &SrcModule = ModuleLoaderCache(ModuleIdentifier);
209 // The function that we will import!
210 GlobalValue *SGV = SrcModule.getNamedValue(CalledFunctionName);
213 // The destination module is referencing function using their renamed name
214 // when importing a function that was originally local in the source
215 // module. The source module we have might not have been renamed so we try
216 // to remove the suffix added during the renaming to recover the original
217 // name in the source module.
218 std::pair<StringRef, StringRef> Split =
219 CalledFunctionName.split(".llvm.");
220 SGV = SrcModule.getNamedValue(Split.first);
221 assert(SGV && "Can't find function to import in source module");
224 report_fatal_error(Twine("Can't load function '") + CalledFunctionName +
225 "' in Module '" + SrcModule.getModuleIdentifier() +
226 "', error in the summary?\n");
229 Function *F = dyn_cast<Function>(SGV);
230 if (!F && isa<GlobalAlias>(SGV)) {
231 auto *SGA = dyn_cast<GlobalAlias>(SGV);
232 F = dyn_cast<Function>(SGA->getBaseObject());
233 CalledFunctionName = F->getName();
235 assert(F && "Imported Function is ... not a Function");
237 // We cannot import weak_any functions/aliases without possibly affecting
238 // the order they are seen and selected by the linker, changing program
240 if (SGV->hasWeakAnyLinkage()) {
241 DEBUG(dbgs() << DestModule.getModuleIdentifier()
242 << ": Ignoring import request for weak-any "
243 << (isa<Function>(SGV) ? "function " : "alias ")
244 << CalledFunctionName << " from "
245 << SrcModule.getModuleIdentifier() << "\n");
249 // Add the function to the import list
250 auto &Entry = ModuleToFunctionsToImportMap[SrcModule.getModuleIdentifier()];
253 // Process the newly imported functions and add callees to the worklist.
255 findExternalCalls(DestModule, *F, Index, CalledFunctions, Worklist);
259 // Automatically import functions in Module \p DestModule based on the summaries
262 // The current implementation imports every called functions that exists in the
264 bool FunctionImporter::importFunctions(Module &DestModule) {
265 DEBUG(dbgs() << "Starting import for Module "
266 << DestModule.getModuleIdentifier() << "\n");
267 unsigned ImportedCount = 0;
269 /// First step is collecting the called external functions.
270 StringSet<> CalledFunctions;
271 SmallVector<StringRef, 64> Worklist;
272 for (auto &F : DestModule) {
273 if (F.isDeclaration() || F.hasFnAttribute(Attribute::OptimizeNone))
275 findExternalCalls(DestModule, F, Index, CalledFunctions, Worklist);
277 if (Worklist.empty())
280 /// Second step: for every call to an external function, try to import it.
282 // Linker that will be used for importing function
283 Linker TheLinker(DestModule);
285 // Map of Module -> List of Function to import from the Module
286 std::map<StringRef, DenseSet<const GlobalValue *>>
287 ModuleToFunctionsToImportMap;
289 // Analyze the summaries and get the list of functions to import by
290 // populating ModuleToFunctionsToImportMap
291 ModuleLazyLoaderCache ModuleLoaderCache(ModuleLoader);
292 GetImportList(DestModule, Worklist, CalledFunctions,
293 ModuleToFunctionsToImportMap, Index, ModuleLoaderCache);
294 assert(Worklist.empty() && "Worklist hasn't been flushed in GetImportList");
296 StringMap<std::unique_ptr<DenseMap<unsigned, MDNode *>>>
297 ModuleToTempMDValsMap;
299 // Do the actual import of functions now, one Module at a time
300 for (auto &FunctionsToImportPerModule : ModuleToFunctionsToImportMap) {
301 // Get the module for the import
302 auto &FunctionsToImport = FunctionsToImportPerModule.second;
303 std::unique_ptr<Module> SrcModule =
304 ModuleLoaderCache.takeModule(FunctionsToImportPerModule.first);
305 assert(&DestModule.getContext() == &SrcModule->getContext() &&
308 // Save the mapping of value ids to temporary metadata created when
309 // importing this function. If we have already imported from this module,
310 // add new temporary metadata to the existing mapping.
311 auto &TempMDVals = ModuleToTempMDValsMap[SrcModule->getModuleIdentifier()];
313 TempMDVals = llvm::make_unique<DenseMap<unsigned, MDNode *>>();
315 // Link in the specified functions.
316 if (TheLinker.linkInModule(std::move(SrcModule), Linker::Flags::None,
317 &Index, &FunctionsToImport, TempMDVals.get()))
318 report_fatal_error("Function Import: link error");
320 ImportedCount += FunctionsToImport.size();
323 // Now link in metadata for all modules from which we imported functions.
324 for (StringMapEntry<std::unique_ptr<DenseMap<unsigned, MDNode *>>> &SME :
325 ModuleToTempMDValsMap) {
326 // Load the specified source module.
327 auto &SrcModule = ModuleLoaderCache(SME.getKey());
328 // The modules were created with lazy metadata loading. Materialize it
329 // now, before linking it.
330 SrcModule.materializeMetadata();
331 UpgradeDebugInfo(SrcModule);
333 // Link in all necessary metadata from this module.
334 if (TheLinker.linkInMetadata(SrcModule, SME.getValue().get()))
338 DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
339 << DestModule.getModuleIdentifier() << "\n");
340 return ImportedCount;
343 /// Summary file to use for function importing when using -function-import from
344 /// the command line.
345 static cl::opt<std::string>
346 SummaryFile("summary-file",
347 cl::desc("The summary file to use for function importing."));
349 static void diagnosticHandler(const DiagnosticInfo &DI) {
350 raw_ostream &OS = errs();
351 DiagnosticPrinterRawOStream DP(OS);
356 /// Parse the function index out of an IR file and return the function
357 /// index object if found, or nullptr if not.
358 static std::unique_ptr<FunctionInfoIndex>
359 getFunctionIndexForFile(StringRef Path, std::string &Error,
360 DiagnosticHandlerFunction DiagnosticHandler) {
361 std::unique_ptr<MemoryBuffer> Buffer;
362 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
363 MemoryBuffer::getFile(Path);
364 if (std::error_code EC = BufferOrErr.getError()) {
365 Error = EC.message();
368 Buffer = std::move(BufferOrErr.get());
369 ErrorOr<std::unique_ptr<object::FunctionIndexObjectFile>> ObjOrErr =
370 object::FunctionIndexObjectFile::create(Buffer->getMemBufferRef(),
372 if (std::error_code EC = ObjOrErr.getError()) {
373 Error = EC.message();
376 return (*ObjOrErr)->takeIndex();
380 /// Pass that performs cross-module function import provided a summary file.
381 class FunctionImportPass : public ModulePass {
382 /// Optional function summary index to use for importing, otherwise
383 /// the summary-file option must be specified.
384 const FunctionInfoIndex *Index;
387 /// Pass identification, replacement for typeid
390 /// Specify pass name for debug output
391 const char *getPassName() const override {
392 return "Function Importing";
395 explicit FunctionImportPass(const FunctionInfoIndex *Index = nullptr)
396 : ModulePass(ID), Index(Index) {}
398 bool runOnModule(Module &M) override {
399 if (SummaryFile.empty() && !Index)
400 report_fatal_error("error: -function-import requires -summary-file or "
401 "file from frontend\n");
402 std::unique_ptr<FunctionInfoIndex> IndexPtr;
403 if (!SummaryFile.empty()) {
405 report_fatal_error("error: -summary-file and index from frontend\n");
407 IndexPtr = getFunctionIndexForFile(SummaryFile, Error, diagnosticHandler);
409 errs() << "Error loading file '" << SummaryFile << "': " << Error
413 Index = IndexPtr.get();
416 // First we need to promote to global scope and rename any local values that
417 // are potentially exported to other modules.
418 if (renameModuleForThinLTO(M, Index)) {
419 errs() << "Error renaming module\n";
423 // Perform the import now.
424 auto ModuleLoader = [&M](StringRef Identifier) {
425 return loadFile(Identifier, M.getContext());
427 FunctionImporter Importer(*Index, ModuleLoader);
428 return Importer.importFunctions(M);
431 } // anonymous namespace
433 char FunctionImportPass::ID = 0;
434 INITIALIZE_PASS_BEGIN(FunctionImportPass, "function-import",
435 "Summary Based Function Import", false, false)
436 INITIALIZE_PASS_END(FunctionImportPass, "function-import",
437 "Summary Based Function Import", false, false)
440 Pass *createFunctionImportPass(const FunctionInfoIndex *Index = nullptr) {
441 return new FunctionImportPass(Index);