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 (Index.findGlobalValueSummaryList(GUID) == Index.end()) {
201 // For SamplePGO, the indirect call targets for local functions will
202 // have its original name annotated in profile. We try to find the
203 // corresponding PGOFuncName as the GUID.
204 GUID = Index.getGUIDFromOriginalID(GUID);
209 if (DefinedGVSummaries.count(GUID)) {
210 DEBUG(dbgs() << "ignored! Target already in destination module.\n");
214 auto GetBonusMultiplier = [](CalleeInfo::HotnessType Hotness) -> float {
215 if (Hotness == CalleeInfo::HotnessType::Hot)
216 return ImportHotMultiplier;
217 if (Hotness == CalleeInfo::HotnessType::Cold)
218 return ImportColdMultiplier;
222 const auto NewThreshold =
223 Threshold * GetBonusMultiplier(Edge.second.Hotness);
225 auto *CalleeSummary =
226 selectCallee(GUID, NewThreshold, Index, Summary.modulePath());
227 if (!CalleeSummary) {
228 DEBUG(dbgs() << "ignored! No qualifying callee with summary found.\n");
231 // "Resolve" the summary, traversing alias,
232 const FunctionSummary *ResolvedCalleeSummary;
233 if (isa<AliasSummary>(CalleeSummary)) {
234 ResolvedCalleeSummary = cast<FunctionSummary>(
235 &cast<AliasSummary>(CalleeSummary)->getAliasee());
237 GlobalValue::isLinkOnceODRLinkage(ResolvedCalleeSummary->linkage()) &&
238 "Unexpected alias to a non-linkonceODR in import list");
240 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
242 assert(ResolvedCalleeSummary->instCount() <= NewThreshold &&
243 "selectCallee() didn't honor the threshold");
245 auto GetAdjustedThreshold = [](unsigned Threshold, bool IsHotCallsite) {
246 // Adjust the threshold for next level of imported functions.
247 // The threshold is different for hot callsites because we can then
248 // inline chains of hot calls.
250 return Threshold * ImportHotInstrFactor;
251 return Threshold * ImportInstrFactor;
254 bool IsHotCallsite = Edge.second.Hotness == CalleeInfo::HotnessType::Hot;
255 const auto AdjThreshold = GetAdjustedThreshold(Threshold, IsHotCallsite);
257 auto ExportModulePath = ResolvedCalleeSummary->modulePath();
258 auto &ProcessedThreshold = ImportList[ExportModulePath][GUID];
259 /// Since the traversal of the call graph is DFS, we can revisit a function
260 /// a second time with a higher threshold. In this case, it is added back to
261 /// the worklist with the new threshold.
262 if (ProcessedThreshold && ProcessedThreshold >= AdjThreshold) {
263 DEBUG(dbgs() << "ignored! Target was already seen with Threshold "
264 << ProcessedThreshold << "\n");
267 bool PreviouslyImported = ProcessedThreshold != 0;
268 // Mark this function as imported in this module, with the current Threshold
269 ProcessedThreshold = AdjThreshold;
271 // Make exports in the source module.
273 auto &ExportList = (*ExportLists)[ExportModulePath];
274 ExportList.insert(GUID);
275 if (!PreviouslyImported) {
276 // This is the first time this function was exported from its source
277 // module, so mark all functions and globals it references as exported
278 // to the outside if they are defined in the same source module.
279 // For efficiency, we unconditionally add all the referenced GUIDs
280 // to the ExportList for this module, and will prune out any not
281 // defined in the module later in a single pass.
282 for (auto &Edge : ResolvedCalleeSummary->calls()) {
283 auto CalleeGUID = Edge.first.getGUID();
284 ExportList.insert(CalleeGUID);
286 for (auto &Ref : ResolvedCalleeSummary->refs()) {
287 auto GUID = Ref.getGUID();
288 ExportList.insert(GUID);
293 // Insert the newly imported function to the worklist.
294 Worklist.emplace_back(ResolvedCalleeSummary, AdjThreshold, GUID);
298 /// Given the list of globals defined in a module, compute the list of imports
299 /// as well as the list of "exports", i.e. the list of symbols referenced from
300 /// another module (that may require promotion).
301 static void ComputeImportForModule(
302 const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
303 FunctionImporter::ImportMapTy &ImportList,
304 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr,
305 const DenseSet<GlobalValue::GUID> *DeadSymbols = nullptr) {
306 // Worklist contains the list of function imported in this module, for which
307 // we will analyse the callees and may import further down the callgraph.
308 SmallVector<EdgeInfo, 128> Worklist;
310 // Populate the worklist with the import for the functions in the current
312 for (auto &GVSummary : DefinedGVSummaries) {
313 if (DeadSymbols && DeadSymbols->count(GVSummary.first)) {
314 DEBUG(dbgs() << "Ignores Dead GUID: " << GVSummary.first << "\n");
317 auto *Summary = GVSummary.second;
318 if (auto *AS = dyn_cast<AliasSummary>(Summary))
319 Summary = &AS->getAliasee();
320 auto *FuncSummary = dyn_cast<FunctionSummary>(Summary);
322 // Skip import for global variables
324 DEBUG(dbgs() << "Initalize import for " << GVSummary.first << "\n");
325 computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
326 DefinedGVSummaries, Worklist, ImportList,
330 // Process the newly imported functions and add callees to the worklist.
331 while (!Worklist.empty()) {
332 auto FuncInfo = Worklist.pop_back_val();
333 auto *Summary = std::get<0>(FuncInfo);
334 auto Threshold = std::get<1>(FuncInfo);
335 auto GUID = std::get<2>(FuncInfo);
337 // Check if we later added this summary with a higher threshold.
338 // If so, skip this entry.
339 auto ExportModulePath = Summary->modulePath();
340 auto &LatestProcessedThreshold = ImportList[ExportModulePath][GUID];
341 if (LatestProcessedThreshold > Threshold)
344 computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries,
345 Worklist, ImportList, ExportLists);
349 } // anonymous namespace
351 /// Compute all the import and export for every module using the Index.
352 void llvm::ComputeCrossModuleImport(
353 const ModuleSummaryIndex &Index,
354 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
355 StringMap<FunctionImporter::ImportMapTy> &ImportLists,
356 StringMap<FunctionImporter::ExportSetTy> &ExportLists,
357 const DenseSet<GlobalValue::GUID> *DeadSymbols) {
358 // For each module that has function defined, compute the import/export lists.
359 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
360 auto &ImportList = ImportLists[DefinedGVSummaries.first()];
361 DEBUG(dbgs() << "Computing import for Module '"
362 << DefinedGVSummaries.first() << "'\n");
363 ComputeImportForModule(DefinedGVSummaries.second, Index, ImportList,
364 &ExportLists, DeadSymbols);
367 // When computing imports we added all GUIDs referenced by anything
368 // imported from the module to its ExportList. Now we prune each ExportList
369 // of any not defined in that module. This is more efficient than checking
370 // while computing imports because some of the summary lists may be long
371 // due to linkonce (comdat) copies.
372 for (auto &ELI : ExportLists) {
373 const auto &DefinedGVSummaries =
374 ModuleToDefinedGVSummaries.lookup(ELI.first());
375 for (auto EI = ELI.second.begin(); EI != ELI.second.end();) {
376 if (!DefinedGVSummaries.count(*EI))
377 EI = ELI.second.erase(EI);
384 DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
386 for (auto &ModuleImports : ImportLists) {
387 auto ModName = ModuleImports.first();
388 auto &Exports = ExportLists[ModName];
389 DEBUG(dbgs() << "* Module " << ModName << " exports " << Exports.size()
390 << " functions. Imports from " << ModuleImports.second.size()
392 for (auto &Src : ModuleImports.second) {
393 auto SrcModName = Src.first();
394 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
395 << SrcModName << "\n");
401 /// Compute all the imports for the given module in the Index.
402 void llvm::ComputeCrossModuleImportForModule(
403 StringRef ModulePath, const ModuleSummaryIndex &Index,
404 FunctionImporter::ImportMapTy &ImportList) {
406 // Collect the list of functions this module defines.
408 GVSummaryMapTy FunctionSummaryMap;
409 Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
411 // Compute the import list for this module.
412 DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
413 ComputeImportForModule(FunctionSummaryMap, Index, ImportList);
416 DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
417 << ImportList.size() << " modules.\n");
418 for (auto &Src : ImportList) {
419 auto SrcModName = Src.first();
420 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
421 << SrcModName << "\n");
426 DenseSet<GlobalValue::GUID> llvm::computeDeadSymbols(
427 const ModuleSummaryIndex &Index,
428 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
430 return DenseSet<GlobalValue::GUID>();
431 if (GUIDPreservedSymbols.empty())
432 // Don't do anything when nothing is live, this is friendly with tests.
433 return DenseSet<GlobalValue::GUID>();
434 DenseSet<GlobalValue::GUID> LiveSymbols = GUIDPreservedSymbols;
435 SmallVector<GlobalValue::GUID, 128> Worklist;
436 Worklist.reserve(LiveSymbols.size() * 2);
437 for (auto GUID : LiveSymbols) {
438 DEBUG(dbgs() << "Live root: " << GUID << "\n");
439 Worklist.push_back(GUID);
441 // Add values flagged in the index as live roots to the worklist.
442 for (const auto &Entry : Index) {
443 bool IsLiveRoot = llvm::any_of(
445 [&](const std::unique_ptr<llvm::GlobalValueSummary> &Summary) {
446 return Summary->liveRoot();
450 DEBUG(dbgs() << "Live root (summary): " << Entry.first << "\n");
451 Worklist.push_back(Entry.first);
454 while (!Worklist.empty()) {
455 auto GUID = Worklist.pop_back_val();
456 auto It = Index.findGlobalValueSummaryList(GUID);
457 if (It == Index.end()) {
458 DEBUG(dbgs() << "Not in index: " << GUID << "\n");
462 // FIXME: we should only make the prevailing copy live here
463 for (auto &Summary : It->second) {
464 for (auto Ref : Summary->refs()) {
465 auto RefGUID = Ref.getGUID();
466 if (LiveSymbols.insert(RefGUID).second) {
467 DEBUG(dbgs() << "Marking live (ref): " << RefGUID << "\n");
468 Worklist.push_back(RefGUID);
471 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) {
472 for (auto Call : FS->calls()) {
473 auto CallGUID = Call.first.getGUID();
474 if (LiveSymbols.insert(CallGUID).second) {
475 DEBUG(dbgs() << "Marking live (call): " << CallGUID << "\n");
476 Worklist.push_back(CallGUID);
480 if (auto *AS = dyn_cast<AliasSummary>(Summary.get())) {
481 auto AliaseeGUID = AS->getAliasee().getOriginalName();
482 if (LiveSymbols.insert(AliaseeGUID).second) {
483 DEBUG(dbgs() << "Marking live (alias): " << AliaseeGUID << "\n");
484 Worklist.push_back(AliaseeGUID);
489 DenseSet<GlobalValue::GUID> DeadSymbols;
491 std::min(Index.size(), Index.size() - LiveSymbols.size()));
492 for (auto &Entry : Index) {
493 auto GUID = Entry.first;
494 if (!LiveSymbols.count(GUID)) {
495 DEBUG(dbgs() << "Marking dead: " << GUID << "\n");
496 DeadSymbols.insert(GUID);
499 DEBUG(dbgs() << LiveSymbols.size() << " symbols Live, and "
500 << DeadSymbols.size() << " symbols Dead \n");
501 NumDeadSymbols += DeadSymbols.size();
502 NumLiveSymbols += LiveSymbols.size();
506 /// Compute the set of summaries needed for a ThinLTO backend compilation of
508 void llvm::gatherImportedSummariesForModule(
509 StringRef ModulePath,
510 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
511 const FunctionImporter::ImportMapTy &ImportList,
512 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
513 // Include all summaries from the importing module.
514 ModuleToSummariesForIndex[ModulePath] =
515 ModuleToDefinedGVSummaries.lookup(ModulePath);
516 // Include summaries for imports.
517 for (auto &ILI : ImportList) {
518 auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()];
519 const auto &DefinedGVSummaries =
520 ModuleToDefinedGVSummaries.lookup(ILI.first());
521 for (auto &GI : ILI.second) {
522 const auto &DS = DefinedGVSummaries.find(GI.first);
523 assert(DS != DefinedGVSummaries.end() &&
524 "Expected a defined summary for imported global value");
525 SummariesForIndex[GI.first] = DS->second;
530 /// Emit the files \p ModulePath will import from into \p OutputFilename.
532 llvm::EmitImportsFiles(StringRef ModulePath, StringRef OutputFilename,
533 const FunctionImporter::ImportMapTy &ModuleImports) {
535 raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::F_None);
538 for (auto &ILI : ModuleImports)
539 ImportsOS << ILI.first() << "\n";
540 return std::error_code();
543 /// Fixup WeakForLinker linkages in \p TheModule based on summary analysis.
544 void llvm::thinLTOResolveWeakForLinkerModule(
545 Module &TheModule, const GVSummaryMapTy &DefinedGlobals) {
546 auto ConvertToDeclaration = [](GlobalValue &GV) {
547 DEBUG(dbgs() << "Converting to a declaration: `" << GV.getName() << "\n");
548 if (Function *F = dyn_cast<Function>(&GV)) {
551 } else if (GlobalVariable *V = dyn_cast<GlobalVariable>(&GV)) {
552 V->setInitializer(nullptr);
553 V->setLinkage(GlobalValue::ExternalLinkage);
556 // For now we don't resolve or drop aliases. Once we do we'll
557 // need to add support here for creating either a function or
558 // variable declaration, and return the new GlobalValue* for
559 // the caller to use.
560 llvm_unreachable("Expected function or variable");
563 auto updateLinkage = [&](GlobalValue &GV) {
564 if (!GlobalValue::isWeakForLinker(GV.getLinkage()))
566 // See if the global summary analysis computed a new resolved linkage.
567 const auto &GS = DefinedGlobals.find(GV.getGUID());
568 if (GS == DefinedGlobals.end())
570 auto NewLinkage = GS->second->linkage();
571 if (NewLinkage == GV.getLinkage())
573 // Check for a non-prevailing def that has interposable linkage
574 // (e.g. non-odr weak or linkonce). In that case we can't simply
575 // convert to available_externally, since it would lose the
576 // interposable property and possibly get inlined. Simply drop
577 // the definition in that case.
578 if (GlobalValue::isAvailableExternallyLinkage(NewLinkage) &&
579 GlobalValue::isInterposableLinkage(GV.getLinkage()))
580 ConvertToDeclaration(GV);
582 DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName() << "` from "
583 << GV.getLinkage() << " to " << NewLinkage << "\n");
584 GV.setLinkage(NewLinkage);
586 // Remove declarations from comdats, including available_externally
587 // as this is a declaration for the linker, and will be dropped eventually.
588 // It is illegal for comdats to contain declarations.
589 auto *GO = dyn_cast_or_null<GlobalObject>(&GV);
590 if (GO && GO->isDeclarationForLinker() && GO->hasComdat())
591 GO->setComdat(nullptr);
594 // Process functions and global now
595 for (auto &GV : TheModule)
597 for (auto &GV : TheModule.globals())
599 for (auto &GV : TheModule.aliases())
603 /// Run internalization on \p TheModule based on symmary analysis.
604 void llvm::thinLTOInternalizeModule(Module &TheModule,
605 const GVSummaryMapTy &DefinedGlobals) {
606 // Parse inline ASM and collect the list of symbols that are not defined in
607 // the current module.
608 StringSet<> AsmUndefinedRefs;
609 ModuleSymbolTable::CollectAsmSymbols(
611 [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
612 if (Flags & object::BasicSymbolRef::SF_Undefined)
613 AsmUndefinedRefs.insert(Name);
616 // Declare a callback for the internalize pass that will ask for every
617 // candidate GlobalValue if it can be internalized or not.
618 auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
619 // Can't be internalized if referenced in inline asm.
620 if (AsmUndefinedRefs.count(GV.getName()))
623 // Lookup the linkage recorded in the summaries during global analysis.
624 const auto &GS = DefinedGlobals.find(GV.getGUID());
625 GlobalValue::LinkageTypes Linkage;
626 if (GS == DefinedGlobals.end()) {
627 // Must have been promoted (possibly conservatively). Find original
628 // name so that we can access the correct summary and see if it can
629 // be internalized again.
630 // FIXME: Eventually we should control promotion instead of promoting
631 // and internalizing again.
633 ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
634 std::string OrigId = GlobalValue::getGlobalIdentifier(
635 OrigName, GlobalValue::InternalLinkage,
636 TheModule.getSourceFileName());
637 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
638 if (GS == DefinedGlobals.end()) {
639 // Also check the original non-promoted non-globalized name. In some
640 // cases a preempted weak value is linked in as a local copy because
641 // it is referenced by an alias (IRLinker::linkGlobalValueProto).
642 // In that case, since it was originally not a local value, it was
643 // recorded in the index using the original name.
644 // FIXME: This may not be needed once PR27866 is fixed.
645 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
646 assert(GS != DefinedGlobals.end());
647 Linkage = GS->second->linkage();
649 Linkage = GS->second->linkage();
652 Linkage = GS->second->linkage();
653 return !GlobalValue::isLocalLinkage(Linkage);
656 // FIXME: See if we can just internalize directly here via linkage changes
657 // based on the index, rather than invoking internalizeModule.
658 llvm::internalizeModule(TheModule, MustPreserveGV);
661 // Automatically import functions in Module \p DestModule based on the summaries
664 Expected<bool> FunctionImporter::importFunctions(
665 Module &DestModule, const FunctionImporter::ImportMapTy &ImportList) {
666 DEBUG(dbgs() << "Starting import for Module "
667 << DestModule.getModuleIdentifier() << "\n");
668 unsigned ImportedCount = 0;
670 IRMover Mover(DestModule);
671 // Do the actual import of functions now, one Module at a time
672 std::set<StringRef> ModuleNameOrderedList;
673 for (auto &FunctionsToImportPerModule : ImportList) {
674 ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
676 for (auto &Name : ModuleNameOrderedList) {
677 // Get the module for the import
678 const auto &FunctionsToImportPerModule = ImportList.find(Name);
679 assert(FunctionsToImportPerModule != ImportList.end());
680 Expected<std::unique_ptr<Module>> SrcModuleOrErr = ModuleLoader(Name);
682 return SrcModuleOrErr.takeError();
683 std::unique_ptr<Module> SrcModule = std::move(*SrcModuleOrErr);
684 assert(&DestModule.getContext() == &SrcModule->getContext() &&
687 // If modules were created with lazy metadata loading, materialize it
688 // now, before linking it (otherwise this will be a noop).
689 if (Error Err = SrcModule->materializeMetadata())
690 return std::move(Err);
692 auto &ImportGUIDs = FunctionsToImportPerModule->second;
693 // Find the globals to import
694 SetVector<GlobalValue *> GlobalsToImport;
695 for (Function &F : *SrcModule) {
698 auto GUID = F.getGUID();
699 auto Import = ImportGUIDs.count(GUID);
700 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function " << GUID
701 << " " << F.getName() << " from "
702 << SrcModule->getSourceFileName() << "\n");
704 if (Error Err = F.materialize())
705 return std::move(Err);
706 if (EnableImportMetadata) {
707 // Add 'thinlto_src_module' metadata for statistics and debugging.
709 "thinlto_src_module",
711 DestModule.getContext(),
712 {llvm::MDString::get(DestModule.getContext(),
713 SrcModule->getSourceFileName())}));
715 GlobalsToImport.insert(&F);
718 for (GlobalVariable &GV : SrcModule->globals()) {
721 auto GUID = GV.getGUID();
722 auto Import = ImportGUIDs.count(GUID);
723 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global " << GUID
724 << " " << GV.getName() << " from "
725 << SrcModule->getSourceFileName() << "\n");
727 if (Error Err = GV.materialize())
728 return std::move(Err);
729 GlobalsToImport.insert(&GV);
732 for (GlobalAlias &GA : SrcModule->aliases()) {
733 // FIXME: This should eventually be controlled entirely by the summary.
734 if (FunctionImportGlobalProcessing::doImportAsDefinition(
735 &GA, &GlobalsToImport)) {
736 GlobalsToImport.insert(&GA);
742 auto GUID = GA.getGUID();
743 auto Import = ImportGUIDs.count(GUID);
744 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias " << GUID
745 << " " << GA.getName() << " from "
746 << SrcModule->getSourceFileName() << "\n");
748 // Alias can't point to "available_externally". However when we import
749 // linkOnceODR the linkage does not change. So we import the alias
750 // and aliasee only in this case. This has been handled by
751 // computeImportForFunction()
752 GlobalObject *GO = GA.getBaseObject();
753 assert(GO->hasLinkOnceODRLinkage() &&
754 "Unexpected alias to a non-linkonceODR in import list");
756 if (!GlobalsToImport.count(GO))
757 DEBUG(dbgs() << " alias triggers importing aliasee " << GO->getGUID()
758 << " " << GO->getName() << " from "
759 << SrcModule->getSourceFileName() << "\n");
761 if (Error Err = GO->materialize())
762 return std::move(Err);
763 GlobalsToImport.insert(GO);
764 if (Error Err = GA.materialize())
765 return std::move(Err);
766 GlobalsToImport.insert(&GA);
770 // Upgrade debug info after we're done materializing all the globals and we
771 // have loaded all the required metadata!
772 UpgradeDebugInfo(*SrcModule);
774 // Link in the specified functions.
775 if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport))
779 for (const auto *GV : GlobalsToImport)
780 dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
781 << " from " << SrcModule->getSourceFileName() << "\n";
784 if (Mover.move(std::move(SrcModule), GlobalsToImport.getArrayRef(),
785 [](GlobalValue &, IRMover::ValueAdder) {},
786 /*IsPerformingImport=*/true))
787 report_fatal_error("Function Import: link error");
789 ImportedCount += GlobalsToImport.size();
790 NumImportedModules++;
793 NumImportedFunctions += ImportedCount;
795 DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
796 << DestModule.getModuleIdentifier() << "\n");
797 return ImportedCount;
800 /// Summary file to use for function importing when using -function-import from
801 /// the command line.
802 static cl::opt<std::string>
803 SummaryFile("summary-file",
804 cl::desc("The summary file to use for function importing."));
806 static bool doImportingForModule(Module &M) {
807 if (SummaryFile.empty())
808 report_fatal_error("error: -function-import requires -summary-file\n");
809 Expected<std::unique_ptr<ModuleSummaryIndex>> IndexPtrOrErr =
810 getModuleSummaryIndexForFile(SummaryFile);
811 if (!IndexPtrOrErr) {
812 logAllUnhandledErrors(IndexPtrOrErr.takeError(), errs(),
813 "Error loading file '" + SummaryFile + "': ");
816 std::unique_ptr<ModuleSummaryIndex> Index = std::move(*IndexPtrOrErr);
818 // First step is collecting the import list.
819 FunctionImporter::ImportMapTy ImportList;
820 ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
823 // Conservatively mark all internal values as promoted. This interface is
824 // only used when doing importing via the function importing pass. The pass
825 // is only enabled when testing importing via the 'opt' tool, which does
826 // not do the ThinLink that would normally determine what values to promote.
827 for (auto &I : *Index) {
828 for (auto &S : I.second) {
829 if (GlobalValue::isLocalLinkage(S->linkage()))
830 S->setLinkage(GlobalValue::ExternalLinkage);
834 // Next we need to promote to global scope and rename any local values that
835 // are potentially exported to other modules.
836 if (renameModuleForThinLTO(M, *Index, nullptr)) {
837 errs() << "Error renaming module\n";
841 // Perform the import now.
842 auto ModuleLoader = [&M](StringRef Identifier) {
843 return loadFile(Identifier, M.getContext());
845 FunctionImporter Importer(*Index, ModuleLoader);
846 Expected<bool> Result = Importer.importFunctions(M, ImportList);
848 // FIXME: Probably need to propagate Errors through the pass manager.
850 logAllUnhandledErrors(Result.takeError(), errs(),
851 "Error importing module: ");
859 /// Pass that performs cross-module function import provided a summary file.
860 class FunctionImportLegacyPass : public ModulePass {
862 /// Pass identification, replacement for typeid
865 /// Specify pass name for debug output
866 StringRef getPassName() const override { return "Function Importing"; }
868 explicit FunctionImportLegacyPass() : ModulePass(ID) {}
870 bool runOnModule(Module &M) override {
874 return doImportingForModule(M);
877 } // anonymous namespace
879 PreservedAnalyses FunctionImportPass::run(Module &M,
880 ModuleAnalysisManager &AM) {
881 if (!doImportingForModule(M))
882 return PreservedAnalyses::all();
884 return PreservedAnalyses::none();
887 char FunctionImportLegacyPass::ID = 0;
888 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import",
889 "Summary Based Function Import", false, false)
892 Pass *createFunctionImportPass() {
893 return new FunctionImportLegacyPass();