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/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringSet.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/IR/AutoUpgrade.h"
21 #include "llvm/IR/DiagnosticPrinter.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/IR/Verifier.h"
25 #include "llvm/IRReader/IRReader.h"
26 #include "llvm/Linker/Linker.h"
27 #include "llvm/Object/IRObjectFile.h"
28 #include "llvm/Object/ModuleSummaryIndexObjectFile.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/SourceMgr.h"
32 #include "llvm/Transforms/IPO/Internalize.h"
33 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
35 #define DEBUG_TYPE "function-import"
39 STATISTIC(NumImportedFunctions, "Number of functions imported");
40 STATISTIC(NumImportedModules, "Number of modules imported from");
41 STATISTIC(NumDeadSymbols, "Number of dead stripped symbols in index");
42 STATISTIC(NumLiveSymbols, "Number of live symbols in index");
44 /// Limit on instruction count of imported functions.
45 static cl::opt<unsigned> ImportInstrLimit(
46 "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
47 cl::desc("Only import functions with less than N instructions"));
50 ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7),
51 cl::Hidden, cl::value_desc("x"),
52 cl::desc("As we import functions, multiply the "
53 "`import-instr-limit` threshold by this factor "
54 "before processing newly imported functions"));
56 static cl::opt<float> ImportHotInstrFactor(
57 "import-hot-evolution-factor", cl::init(1.0), cl::Hidden,
59 cl::desc("As we import functions called from hot callsite, multiply the "
60 "`import-instr-limit` threshold by this factor "
61 "before processing newly imported functions"));
63 static cl::opt<float> ImportHotMultiplier(
64 "import-hot-multiplier", cl::init(3.0), cl::Hidden, cl::value_desc("x"),
65 cl::desc("Multiply the `import-instr-limit` threshold for hot callsites"));
67 // FIXME: This multiplier was not really tuned up.
68 static cl::opt<float> ImportColdMultiplier(
69 "import-cold-multiplier", cl::init(0), cl::Hidden, cl::value_desc("N"),
70 cl::desc("Multiply the `import-instr-limit` threshold for cold callsites"));
72 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden,
73 cl::desc("Print imported functions"));
75 static cl::opt<bool> ComputeDead("compute-dead", cl::init(true), cl::Hidden,
76 cl::desc("Compute dead symbols"));
78 static cl::opt<bool> EnableImportMetadata(
79 "enable-import-metadata", cl::init(
81 true /*Enabled with asserts.*/
86 cl::Hidden, cl::desc("Enable import metadata like 'thinlto_src_module'"));
88 // Load lazily a module from \p FileName in \p Context.
89 static std::unique_ptr<Module> loadFile(const std::string &FileName,
90 LLVMContext &Context) {
92 DEBUG(dbgs() << "Loading '" << FileName << "'\n");
93 // Metadata isn't loaded until functions are imported, to minimize
94 // the memory overhead.
95 std::unique_ptr<Module> Result =
96 getLazyIRFileModule(FileName, Err, Context,
97 /* ShouldLazyLoadMetadata = */ true);
99 Err.print("function-import", errs());
100 report_fatal_error("Abort");
108 /// Given a list of possible callee implementation for a call site, select one
109 /// that fits the \p Threshold.
111 /// FIXME: select "best" instead of first that fits. But what is "best"?
112 /// - The smallest: more likely to be inlined.
113 /// - The one with the least outgoing edges (already well optimized).
114 /// - One from a module already being imported from in order to reduce the
115 /// number of source modules parsed/linked.
116 /// - One that has PGO data attached.
117 /// - [insert you fancy metric here]
118 static const GlobalValueSummary *
119 selectCallee(const ModuleSummaryIndex &Index,
120 const GlobalValueSummaryList &CalleeSummaryList,
121 unsigned Threshold, StringRef CallerModulePath) {
122 auto It = llvm::find_if(
124 [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
125 auto *GVSummary = SummaryPtr.get();
126 if (GlobalValue::isInterposableLinkage(GVSummary->linkage()))
127 // There is no point in importing these, we can't inline them
129 if (auto *AS = dyn_cast<AliasSummary>(GVSummary)) {
130 GVSummary = &AS->getAliasee();
131 // Alias can't point to "available_externally". However when we import
132 // linkOnceODR the linkage does not change. So we import the alias
133 // and aliasee only in this case.
134 // FIXME: we should import alias as available_externally *function*,
135 // the destination module does need to know it is an alias.
136 if (!GlobalValue::isLinkOnceODRLinkage(GVSummary->linkage()))
140 auto *Summary = cast<FunctionSummary>(GVSummary);
142 // If this is a local function, make sure we import the copy
143 // in the caller's module. The only time a local function can
144 // share an entry in the index is if there is a local with the same name
145 // in another module that had the same source file name (in a different
146 // directory), where each was compiled in their own directory so there
147 // was not distinguishing path.
148 // However, do the import from another module if there is only one
149 // entry in the list - in that case this must be a reference due
150 // to indirect call profile data, since a function pointer can point to
151 // a local in another module.
152 if (GlobalValue::isLocalLinkage(Summary->linkage()) &&
153 CalleeSummaryList.size() > 1 &&
154 Summary->modulePath() != CallerModulePath)
157 if (Summary->instCount() > Threshold)
160 if (Summary->notEligibleToImport())
165 if (It == CalleeSummaryList.end())
168 return cast<GlobalValueSummary>(It->get());
171 /// Return the summary for the function \p GUID that fits the \p Threshold, or
172 /// null if there's no match.
173 static const GlobalValueSummary *selectCallee(GlobalValue::GUID GUID,
175 const ModuleSummaryIndex &Index,
176 StringRef CallerModulePath) {
177 auto CalleeSummaryList = Index.findGlobalValueSummaryList(GUID);
178 if (CalleeSummaryList == Index.end())
179 return nullptr; // This function does not have a summary
180 return selectCallee(Index, CalleeSummaryList->second, Threshold,
184 using EdgeInfo = std::tuple<const FunctionSummary *, unsigned /* Threshold */,
187 /// Compute the list of functions to import for a given caller. Mark these
188 /// imported functions and the symbols they reference in their source module as
189 /// exported from their source module.
190 static void computeImportForFunction(
191 const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
192 const unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries,
193 SmallVectorImpl<EdgeInfo> &Worklist,
194 FunctionImporter::ImportMapTy &ImportList,
195 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
196 for (auto &Edge : Summary.calls()) {
197 auto GUID = Edge.first.getGUID();
198 DEBUG(dbgs() << " edge -> " << GUID << " Threshold:" << Threshold << "\n");
200 if (DefinedGVSummaries.count(GUID)) {
201 DEBUG(dbgs() << "ignored! Target already in destination module.\n");
205 auto GetBonusMultiplier = [](CalleeInfo::HotnessType Hotness) -> float {
206 if (Hotness == CalleeInfo::HotnessType::Hot)
207 return ImportHotMultiplier;
208 if (Hotness == CalleeInfo::HotnessType::Cold)
209 return ImportColdMultiplier;
213 const auto NewThreshold =
214 Threshold * GetBonusMultiplier(Edge.second.Hotness);
216 auto *CalleeSummary =
217 selectCallee(GUID, NewThreshold, Index, Summary.modulePath());
218 if (!CalleeSummary) {
219 DEBUG(dbgs() << "ignored! No qualifying callee with summary found.\n");
222 // "Resolve" the summary, traversing alias,
223 const FunctionSummary *ResolvedCalleeSummary;
224 if (isa<AliasSummary>(CalleeSummary)) {
225 ResolvedCalleeSummary = cast<FunctionSummary>(
226 &cast<AliasSummary>(CalleeSummary)->getAliasee());
228 GlobalValue::isLinkOnceODRLinkage(ResolvedCalleeSummary->linkage()) &&
229 "Unexpected alias to a non-linkonceODR in import list");
231 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
233 assert(ResolvedCalleeSummary->instCount() <= NewThreshold &&
234 "selectCallee() didn't honor the threshold");
236 auto GetAdjustedThreshold = [](unsigned Threshold, bool IsHotCallsite) {
237 // Adjust the threshold for next level of imported functions.
238 // The threshold is different for hot callsites because we can then
239 // inline chains of hot calls.
241 return Threshold * ImportHotInstrFactor;
242 return Threshold * ImportInstrFactor;
245 bool IsHotCallsite = Edge.second.Hotness == CalleeInfo::HotnessType::Hot;
246 const auto AdjThreshold = GetAdjustedThreshold(Threshold, IsHotCallsite);
248 auto ExportModulePath = ResolvedCalleeSummary->modulePath();
249 auto &ProcessedThreshold = ImportList[ExportModulePath][GUID];
250 /// Since the traversal of the call graph is DFS, we can revisit a function
251 /// a second time with a higher threshold. In this case, it is added back to
252 /// the worklist with the new threshold.
253 if (ProcessedThreshold && ProcessedThreshold >= AdjThreshold) {
254 DEBUG(dbgs() << "ignored! Target was already seen with Threshold "
255 << ProcessedThreshold << "\n");
258 bool PreviouslyImported = ProcessedThreshold != 0;
259 // Mark this function as imported in this module, with the current Threshold
260 ProcessedThreshold = AdjThreshold;
262 // Make exports in the source module.
264 auto &ExportList = (*ExportLists)[ExportModulePath];
265 ExportList.insert(GUID);
266 if (!PreviouslyImported) {
267 // This is the first time this function was exported from its source
268 // module, so mark all functions and globals it references as exported
269 // to the outside if they are defined in the same source module.
270 // For efficiency, we unconditionally add all the referenced GUIDs
271 // to the ExportList for this module, and will prune out any not
272 // defined in the module later in a single pass.
273 for (auto &Edge : ResolvedCalleeSummary->calls()) {
274 auto CalleeGUID = Edge.first.getGUID();
275 ExportList.insert(CalleeGUID);
277 for (auto &Ref : ResolvedCalleeSummary->refs()) {
278 auto GUID = Ref.getGUID();
279 ExportList.insert(GUID);
284 // Insert the newly imported function to the worklist.
285 Worklist.emplace_back(ResolvedCalleeSummary, AdjThreshold, GUID);
289 /// Given the list of globals defined in a module, compute the list of imports
290 /// as well as the list of "exports", i.e. the list of symbols referenced from
291 /// another module (that may require promotion).
292 static void ComputeImportForModule(
293 const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
294 FunctionImporter::ImportMapTy &ImportList,
295 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr,
296 const DenseSet<GlobalValue::GUID> *DeadSymbols = nullptr) {
297 // Worklist contains the list of function imported in this module, for which
298 // we will analyse the callees and may import further down the callgraph.
299 SmallVector<EdgeInfo, 128> Worklist;
301 // Populate the worklist with the import for the functions in the current
303 for (auto &GVSummary : DefinedGVSummaries) {
304 if (DeadSymbols && DeadSymbols->count(GVSummary.first)) {
305 DEBUG(dbgs() << "Ignores Dead GUID: " << GVSummary.first << "\n");
308 auto *Summary = GVSummary.second;
309 if (auto *AS = dyn_cast<AliasSummary>(Summary))
310 Summary = &AS->getAliasee();
311 auto *FuncSummary = dyn_cast<FunctionSummary>(Summary);
313 // Skip import for global variables
315 DEBUG(dbgs() << "Initalize import for " << GVSummary.first << "\n");
316 computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
317 DefinedGVSummaries, Worklist, ImportList,
321 // Process the newly imported functions and add callees to the worklist.
322 while (!Worklist.empty()) {
323 auto FuncInfo = Worklist.pop_back_val();
324 auto *Summary = std::get<0>(FuncInfo);
325 auto Threshold = std::get<1>(FuncInfo);
326 auto GUID = std::get<2>(FuncInfo);
328 // Check if we later added this summary with a higher threshold.
329 // If so, skip this entry.
330 auto ExportModulePath = Summary->modulePath();
331 auto &LatestProcessedThreshold = ImportList[ExportModulePath][GUID];
332 if (LatestProcessedThreshold > Threshold)
335 computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries,
336 Worklist, ImportList, ExportLists);
340 } // anonymous namespace
342 /// Compute all the import and export for every module using the Index.
343 void llvm::ComputeCrossModuleImport(
344 const ModuleSummaryIndex &Index,
345 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
346 StringMap<FunctionImporter::ImportMapTy> &ImportLists,
347 StringMap<FunctionImporter::ExportSetTy> &ExportLists,
348 const DenseSet<GlobalValue::GUID> *DeadSymbols) {
349 // For each module that has function defined, compute the import/export lists.
350 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
351 auto &ImportList = ImportLists[DefinedGVSummaries.first()];
352 DEBUG(dbgs() << "Computing import for Module '"
353 << DefinedGVSummaries.first() << "'\n");
354 ComputeImportForModule(DefinedGVSummaries.second, Index, ImportList,
355 &ExportLists, DeadSymbols);
358 // When computing imports we added all GUIDs referenced by anything
359 // imported from the module to its ExportList. Now we prune each ExportList
360 // of any not defined in that module. This is more efficient than checking
361 // while computing imports because some of the summary lists may be long
362 // due to linkonce (comdat) copies.
363 for (auto &ELI : ExportLists) {
364 const auto &DefinedGVSummaries =
365 ModuleToDefinedGVSummaries.lookup(ELI.first());
366 for (auto EI = ELI.second.begin(); EI != ELI.second.end();) {
367 if (!DefinedGVSummaries.count(*EI))
368 EI = ELI.second.erase(EI);
375 DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
377 for (auto &ModuleImports : ImportLists) {
378 auto ModName = ModuleImports.first();
379 auto &Exports = ExportLists[ModName];
380 DEBUG(dbgs() << "* Module " << ModName << " exports " << Exports.size()
381 << " functions. Imports from " << ModuleImports.second.size()
383 for (auto &Src : ModuleImports.second) {
384 auto SrcModName = Src.first();
385 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
386 << SrcModName << "\n");
392 /// Compute all the imports for the given module in the Index.
393 void llvm::ComputeCrossModuleImportForModule(
394 StringRef ModulePath, const ModuleSummaryIndex &Index,
395 FunctionImporter::ImportMapTy &ImportList) {
397 // Collect the list of functions this module defines.
399 GVSummaryMapTy FunctionSummaryMap;
400 Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
402 // Compute the import list for this module.
403 DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
404 ComputeImportForModule(FunctionSummaryMap, Index, ImportList);
407 DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
408 << ImportList.size() << " modules.\n");
409 for (auto &Src : ImportList) {
410 auto SrcModName = Src.first();
411 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
412 << SrcModName << "\n");
417 DenseSet<GlobalValue::GUID> llvm::computeDeadSymbols(
418 const ModuleSummaryIndex &Index,
419 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
421 return DenseSet<GlobalValue::GUID>();
422 if (GUIDPreservedSymbols.empty())
423 // Don't do anything when nothing is live, this is friendly with tests.
424 return DenseSet<GlobalValue::GUID>();
425 DenseSet<GlobalValue::GUID> LiveSymbols = GUIDPreservedSymbols;
426 SmallVector<GlobalValue::GUID, 128> Worklist;
427 Worklist.reserve(LiveSymbols.size() * 2);
428 for (auto GUID : LiveSymbols) {
429 DEBUG(dbgs() << "Live root: " << GUID << "\n");
430 Worklist.push_back(GUID);
432 // Add values flagged in the index as live roots to the worklist.
433 for (const auto &Entry : Index) {
434 bool IsLiveRoot = llvm::any_of(
436 [&](const std::unique_ptr<llvm::GlobalValueSummary> &Summary) {
437 return Summary->liveRoot();
441 DEBUG(dbgs() << "Live root (summary): " << Entry.first << "\n");
442 Worklist.push_back(Entry.first);
445 while (!Worklist.empty()) {
446 auto GUID = Worklist.pop_back_val();
447 auto It = Index.findGlobalValueSummaryList(GUID);
448 if (It == Index.end()) {
449 DEBUG(dbgs() << "Not in index: " << GUID << "\n");
453 // FIXME: we should only make the prevailing copy live here
454 for (auto &Summary : It->second) {
455 for (auto Ref : Summary->refs()) {
456 auto RefGUID = Ref.getGUID();
457 if (LiveSymbols.insert(RefGUID).second) {
458 DEBUG(dbgs() << "Marking live (ref): " << RefGUID << "\n");
459 Worklist.push_back(RefGUID);
462 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) {
463 for (auto Call : FS->calls()) {
464 auto CallGUID = Call.first.getGUID();
465 if (LiveSymbols.insert(CallGUID).second) {
466 DEBUG(dbgs() << "Marking live (call): " << CallGUID << "\n");
467 Worklist.push_back(CallGUID);
471 if (auto *AS = dyn_cast<AliasSummary>(Summary.get())) {
472 auto AliaseeGUID = AS->getAliasee().getOriginalName();
473 if (LiveSymbols.insert(AliaseeGUID).second) {
474 DEBUG(dbgs() << "Marking live (alias): " << AliaseeGUID << "\n");
475 Worklist.push_back(AliaseeGUID);
480 DenseSet<GlobalValue::GUID> DeadSymbols;
482 std::min(Index.size(), Index.size() - LiveSymbols.size()));
483 for (auto &Entry : Index) {
484 auto GUID = Entry.first;
485 if (!LiveSymbols.count(GUID)) {
486 DEBUG(dbgs() << "Marking dead: " << GUID << "\n");
487 DeadSymbols.insert(GUID);
490 DEBUG(dbgs() << LiveSymbols.size() << " symbols Live, and "
491 << DeadSymbols.size() << " symbols Dead \n");
492 NumDeadSymbols += DeadSymbols.size();
493 NumLiveSymbols += LiveSymbols.size();
497 /// Compute the set of summaries needed for a ThinLTO backend compilation of
499 void llvm::gatherImportedSummariesForModule(
500 StringRef ModulePath,
501 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
502 const FunctionImporter::ImportMapTy &ImportList,
503 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
504 // Include all summaries from the importing module.
505 ModuleToSummariesForIndex[ModulePath] =
506 ModuleToDefinedGVSummaries.lookup(ModulePath);
507 // Include summaries for imports.
508 for (auto &ILI : ImportList) {
509 auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()];
510 const auto &DefinedGVSummaries =
511 ModuleToDefinedGVSummaries.lookup(ILI.first());
512 for (auto &GI : ILI.second) {
513 const auto &DS = DefinedGVSummaries.find(GI.first);
514 assert(DS != DefinedGVSummaries.end() &&
515 "Expected a defined summary for imported global value");
516 SummariesForIndex[GI.first] = DS->second;
521 /// Emit the files \p ModulePath will import from into \p OutputFilename.
523 llvm::EmitImportsFiles(StringRef ModulePath, StringRef OutputFilename,
524 const FunctionImporter::ImportMapTy &ModuleImports) {
526 raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::F_None);
529 for (auto &ILI : ModuleImports)
530 ImportsOS << ILI.first() << "\n";
531 return std::error_code();
534 /// Fixup WeakForLinker linkages in \p TheModule based on summary analysis.
535 void llvm::thinLTOResolveWeakForLinkerModule(
536 Module &TheModule, const GVSummaryMapTy &DefinedGlobals) {
537 auto ConvertToDeclaration = [](GlobalValue &GV) {
538 DEBUG(dbgs() << "Converting to a declaration: `" << GV.getName() << "\n");
539 if (Function *F = dyn_cast<Function>(&GV)) {
542 } else if (GlobalVariable *V = dyn_cast<GlobalVariable>(&GV)) {
543 V->setInitializer(nullptr);
544 V->setLinkage(GlobalValue::ExternalLinkage);
547 // For now we don't resolve or drop aliases. Once we do we'll
548 // need to add support here for creating either a function or
549 // variable declaration, and return the new GlobalValue* for
550 // the caller to use.
551 assert(false && "Expected function or variable");
554 auto updateLinkage = [&](GlobalValue &GV) {
555 if (!GlobalValue::isWeakForLinker(GV.getLinkage()))
557 // See if the global summary analysis computed a new resolved linkage.
558 const auto &GS = DefinedGlobals.find(GV.getGUID());
559 if (GS == DefinedGlobals.end())
561 auto NewLinkage = GS->second->linkage();
562 if (NewLinkage == GV.getLinkage())
564 // Check for a non-prevailing def that has interposable linkage
565 // (e.g. non-odr weak or linkonce). In that case we can't simply
566 // convert to available_externally, since it would lose the
567 // interposable property and possibly get inlined. Simply drop
568 // the definition in that case.
569 if (GlobalValue::isAvailableExternallyLinkage(NewLinkage) &&
570 GlobalValue::isInterposableLinkage(GV.getLinkage()))
571 ConvertToDeclaration(GV);
573 DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName() << "` from "
574 << GV.getLinkage() << " to " << NewLinkage << "\n");
575 GV.setLinkage(NewLinkage);
577 // Remove declarations from comdats, including available_externally
578 // as this is a declaration for the linker, and will be dropped eventually.
579 // It is illegal for comdats to contain declarations.
580 auto *GO = dyn_cast_or_null<GlobalObject>(&GV);
581 if (GO && GO->isDeclarationForLinker() && GO->hasComdat())
582 GO->setComdat(nullptr);
585 // Process functions and global now
586 for (auto &GV : TheModule)
588 for (auto &GV : TheModule.globals())
590 for (auto &GV : TheModule.aliases())
594 /// Run internalization on \p TheModule based on symmary analysis.
595 void llvm::thinLTOInternalizeModule(Module &TheModule,
596 const GVSummaryMapTy &DefinedGlobals) {
597 // Parse inline ASM and collect the list of symbols that are not defined in
598 // the current module.
599 StringSet<> AsmUndefinedRefs;
600 ModuleSymbolTable::CollectAsmSymbols(
601 Triple(TheModule.getTargetTriple()), TheModule.getModuleInlineAsm(),
602 [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
603 if (Flags & object::BasicSymbolRef::SF_Undefined)
604 AsmUndefinedRefs.insert(Name);
607 // Declare a callback for the internalize pass that will ask for every
608 // candidate GlobalValue if it can be internalized or not.
609 auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
610 // Can't be internalized if referenced in inline asm.
611 if (AsmUndefinedRefs.count(GV.getName()))
614 // Lookup the linkage recorded in the summaries during global analysis.
615 const auto &GS = DefinedGlobals.find(GV.getGUID());
616 GlobalValue::LinkageTypes Linkage;
617 if (GS == DefinedGlobals.end()) {
618 // Must have been promoted (possibly conservatively). Find original
619 // name so that we can access the correct summary and see if it can
620 // be internalized again.
621 // FIXME: Eventually we should control promotion instead of promoting
622 // and internalizing again.
624 ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
625 std::string OrigId = GlobalValue::getGlobalIdentifier(
626 OrigName, GlobalValue::InternalLinkage,
627 TheModule.getSourceFileName());
628 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
629 if (GS == DefinedGlobals.end()) {
630 // Also check the original non-promoted non-globalized name. In some
631 // cases a preempted weak value is linked in as a local copy because
632 // it is referenced by an alias (IRLinker::linkGlobalValueProto).
633 // In that case, since it was originally not a local value, it was
634 // recorded in the index using the original name.
635 // FIXME: This may not be needed once PR27866 is fixed.
636 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
637 assert(GS != DefinedGlobals.end());
638 Linkage = GS->second->linkage();
640 Linkage = GS->second->linkage();
643 Linkage = GS->second->linkage();
644 return !GlobalValue::isLocalLinkage(Linkage);
647 // FIXME: See if we can just internalize directly here via linkage changes
648 // based on the index, rather than invoking internalizeModule.
649 llvm::internalizeModule(TheModule, MustPreserveGV);
652 // Automatically import functions in Module \p DestModule based on the summaries
655 Expected<bool> FunctionImporter::importFunctions(
656 Module &DestModule, const FunctionImporter::ImportMapTy &ImportList) {
657 DEBUG(dbgs() << "Starting import for Module "
658 << DestModule.getModuleIdentifier() << "\n");
659 unsigned ImportedCount = 0;
661 // Linker that will be used for importing function
662 Linker TheLinker(DestModule);
663 // Do the actual import of functions now, one Module at a time
664 std::set<StringRef> ModuleNameOrderedList;
665 for (auto &FunctionsToImportPerModule : ImportList) {
666 ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
668 for (auto &Name : ModuleNameOrderedList) {
669 // Get the module for the import
670 const auto &FunctionsToImportPerModule = ImportList.find(Name);
671 assert(FunctionsToImportPerModule != ImportList.end());
672 Expected<std::unique_ptr<Module>> SrcModuleOrErr = ModuleLoader(Name);
674 return SrcModuleOrErr.takeError();
675 std::unique_ptr<Module> SrcModule = std::move(*SrcModuleOrErr);
676 assert(&DestModule.getContext() == &SrcModule->getContext() &&
679 // If modules were created with lazy metadata loading, materialize it
680 // now, before linking it (otherwise this will be a noop).
681 if (Error Err = SrcModule->materializeMetadata())
682 return std::move(Err);
684 auto &ImportGUIDs = FunctionsToImportPerModule->second;
685 // Find the globals to import
686 DenseSet<const GlobalValue *> GlobalsToImport;
687 for (Function &F : *SrcModule) {
690 auto GUID = F.getGUID();
691 auto Import = ImportGUIDs.count(GUID);
692 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function " << GUID
693 << " " << F.getName() << " from "
694 << SrcModule->getSourceFileName() << "\n");
696 if (Error Err = F.materialize())
697 return std::move(Err);
698 if (EnableImportMetadata) {
699 // Add 'thinlto_src_module' metadata for statistics and debugging.
701 "thinlto_src_module",
703 DestModule.getContext(),
704 {llvm::MDString::get(DestModule.getContext(),
705 SrcModule->getSourceFileName())}));
707 GlobalsToImport.insert(&F);
710 for (GlobalVariable &GV : SrcModule->globals()) {
713 auto GUID = GV.getGUID();
714 auto Import = ImportGUIDs.count(GUID);
715 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global " << GUID
716 << " " << GV.getName() << " from "
717 << SrcModule->getSourceFileName() << "\n");
719 if (Error Err = GV.materialize())
720 return std::move(Err);
721 GlobalsToImport.insert(&GV);
724 for (GlobalAlias &GA : SrcModule->aliases()) {
727 auto GUID = GA.getGUID();
728 auto Import = ImportGUIDs.count(GUID);
729 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias " << GUID
730 << " " << GA.getName() << " from "
731 << SrcModule->getSourceFileName() << "\n");
733 // Alias can't point to "available_externally". However when we import
734 // linkOnceODR the linkage does not change. So we import the alias
735 // and aliasee only in this case. This has been handled by
736 // computeImportForFunction()
737 GlobalObject *GO = GA.getBaseObject();
738 assert(GO->hasLinkOnceODRLinkage() &&
739 "Unexpected alias to a non-linkonceODR in import list");
741 if (!GlobalsToImport.count(GO))
742 DEBUG(dbgs() << " alias triggers importing aliasee " << GO->getGUID()
743 << " " << GO->getName() << " from "
744 << SrcModule->getSourceFileName() << "\n");
746 if (Error Err = GO->materialize())
747 return std::move(Err);
748 GlobalsToImport.insert(GO);
749 if (Error Err = GA.materialize())
750 return std::move(Err);
751 GlobalsToImport.insert(&GA);
755 // Upgrade debug info after we're done materializing all the globals and we
756 // have loaded all the required metadata!
757 UpgradeDebugInfo(*SrcModule);
759 // Link in the specified functions.
760 if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport))
764 for (const auto *GV : GlobalsToImport)
765 dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
766 << " from " << SrcModule->getSourceFileName() << "\n";
769 // Instruct the linker that the client will take care of linkonce resolution
770 unsigned Flags = Linker::Flags::DontForceLinkLinkonceODR;
772 if (TheLinker.linkInModule(std::move(SrcModule), Flags, &GlobalsToImport))
773 report_fatal_error("Function Import: link error");
775 ImportedCount += GlobalsToImport.size();
776 NumImportedModules++;
779 NumImportedFunctions += ImportedCount;
781 DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
782 << DestModule.getModuleIdentifier() << "\n");
783 return ImportedCount;
786 /// Summary file to use for function importing when using -function-import from
787 /// the command line.
788 static cl::opt<std::string>
789 SummaryFile("summary-file",
790 cl::desc("The summary file to use for function importing."));
792 static bool doImportingForModule(Module &M) {
793 if (SummaryFile.empty())
794 report_fatal_error("error: -function-import requires -summary-file\n");
795 Expected<std::unique_ptr<ModuleSummaryIndex>> IndexPtrOrErr =
796 getModuleSummaryIndexForFile(SummaryFile);
797 if (!IndexPtrOrErr) {
798 logAllUnhandledErrors(IndexPtrOrErr.takeError(), errs(),
799 "Error loading file '" + SummaryFile + "': ");
802 std::unique_ptr<ModuleSummaryIndex> Index = std::move(*IndexPtrOrErr);
804 // First step is collecting the import list.
805 FunctionImporter::ImportMapTy ImportList;
806 ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
809 // Conservatively mark all internal values as promoted. This interface is
810 // only used when doing importing via the function importing pass. The pass
811 // is only enabled when testing importing via the 'opt' tool, which does
812 // not do the ThinLink that would normally determine what values to promote.
813 for (auto &I : *Index) {
814 for (auto &S : I.second) {
815 if (GlobalValue::isLocalLinkage(S->linkage()))
816 S->setLinkage(GlobalValue::ExternalLinkage);
820 // Next we need to promote to global scope and rename any local values that
821 // are potentially exported to other modules.
822 if (renameModuleForThinLTO(M, *Index, nullptr)) {
823 errs() << "Error renaming module\n";
827 // Perform the import now.
828 auto ModuleLoader = [&M](StringRef Identifier) {
829 return loadFile(Identifier, M.getContext());
831 FunctionImporter Importer(*Index, ModuleLoader);
832 Expected<bool> Result = Importer.importFunctions(M, ImportList);
834 // FIXME: Probably need to propagate Errors through the pass manager.
836 logAllUnhandledErrors(Result.takeError(), errs(),
837 "Error importing module: ");
845 /// Pass that performs cross-module function import provided a summary file.
846 class FunctionImportLegacyPass : public ModulePass {
848 /// Pass identification, replacement for typeid
851 /// Specify pass name for debug output
852 StringRef getPassName() const override { return "Function Importing"; }
854 explicit FunctionImportLegacyPass() : ModulePass(ID) {}
856 bool runOnModule(Module &M) override {
860 return doImportingForModule(M);
863 } // anonymous namespace
865 PreservedAnalyses FunctionImportPass::run(Module &M,
866 ModuleAnalysisManager &AM) {
867 if (!doImportingForModule(M))
868 return PreservedAnalyses::all();
870 return PreservedAnalyses::none();
873 char FunctionImportLegacyPass::ID = 0;
874 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import",
875 "Summary Based Function Import", false, false)
878 Pass *createFunctionImportPass() {
879 return new FunctionImportLegacyPass();