1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements Function import based on summaries. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/IPO/FunctionImport.h" 15 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/IRReader/IRReader.h" 25 #include "llvm/Linker/Linker.h" 26 #include "llvm/Object/IRObjectFile.h" 27 #include "llvm/Object/ModuleSummaryIndexObjectFile.h" 28 #include "llvm/Support/CommandLine.h" 29 #include "llvm/Support/Debug.h" 30 #include "llvm/Support/SourceMgr.h" 31 #include "llvm/Transforms/IPO/Internalize.h" 32 #include "llvm/Transforms/Utils/FunctionImportUtils.h" 33 34 #define DEBUG_TYPE "function-import" 35 36 using namespace llvm; 37 38 STATISTIC(NumImported, "Number of functions imported"); 39 40 /// Limit on instruction count of imported functions. 41 static cl::opt<unsigned> ImportInstrLimit( 42 "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"), 43 cl::desc("Only import functions with less than N instructions")); 44 45 static cl::opt<float> 46 ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7), 47 cl::Hidden, cl::value_desc("x"), 48 cl::desc("As we import functions, multiply the " 49 "`import-instr-limit` threshold by this factor " 50 "before processing newly imported functions")); 51 52 static cl::opt<float> ImportHotInstrFactor( 53 "import-hot-evolution-factor", cl::init(1.0), cl::Hidden, 54 cl::value_desc("x"), 55 cl::desc("As we import functions called from hot callsite, multiply the " 56 "`import-instr-limit` threshold by this factor " 57 "before processing newly imported functions")); 58 59 static cl::opt<float> ImportHotMultiplier( 60 "import-hot-multiplier", cl::init(3.0), cl::Hidden, cl::value_desc("x"), 61 cl::desc("Multiply the `import-instr-limit` threshold for hot callsites")); 62 63 // FIXME: This multiplier was not really tuned up. 64 static cl::opt<float> ImportColdMultiplier( 65 "import-cold-multiplier", cl::init(0), cl::Hidden, cl::value_desc("N"), 66 cl::desc("Multiply the `import-instr-limit` threshold for cold callsites")); 67 68 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden, 69 cl::desc("Print imported functions")); 70 71 // Temporary allows the function import pass to disable always linking 72 // referenced discardable symbols. 73 static cl::opt<bool> 74 DontForceImportReferencedDiscardableSymbols("disable-force-link-odr", 75 cl::init(false), cl::Hidden); 76 77 static cl::opt<bool> EnableImportMetadata( 78 "enable-import-metadata", cl::init( 79 #if !defined(NDEBUG) 80 true /*Enabled with asserts.*/ 81 #else 82 false 83 #endif 84 ), 85 cl::Hidden, cl::desc("Enable import metadata like 'thinlto_src_module'")); 86 87 // Load lazily a module from \p FileName in \p Context. 88 static std::unique_ptr<Module> loadFile(const std::string &FileName, 89 LLVMContext &Context) { 90 SMDiagnostic Err; 91 DEBUG(dbgs() << "Loading '" << FileName << "'\n"); 92 // Metadata isn't loaded until functions are imported, to minimize 93 // the memory overhead. 94 std::unique_ptr<Module> Result = 95 getLazyIRFileModule(FileName, Err, Context, 96 /* ShouldLazyLoadMetadata = */ true); 97 if (!Result) { 98 Err.print("function-import", errs()); 99 report_fatal_error("Abort"); 100 } 101 102 return Result; 103 } 104 105 namespace { 106 107 // Return true if the Summary describes a GlobalValue that can be externally 108 // referenced, i.e. it does not need renaming (linkage is not local) or renaming 109 // is possible (does not have a section for instance). 110 static bool canBeExternallyReferenced(const GlobalValueSummary &Summary) { 111 if (!Summary.needsRenaming()) 112 return true; 113 114 if (Summary.noRename()) 115 // Can't externally reference a global that needs renaming if has a section 116 // or is referenced from inline assembly, for example. 117 return false; 118 119 return true; 120 } 121 122 // Return true if \p GUID describes a GlobalValue that can be externally 123 // referenced, i.e. it does not need renaming (linkage is not local) or 124 // renaming is possible (does not have a section for instance). 125 static bool canBeExternallyReferenced(const ModuleSummaryIndex &Index, 126 GlobalValue::GUID GUID) { 127 auto Summaries = Index.findGlobalValueSummaryList(GUID); 128 if (Summaries == Index.end()) 129 return true; 130 if (Summaries->second.size() != 1) 131 // If there are multiple globals with this GUID, then we know it is 132 // not a local symbol, and it is necessarily externally referenced. 133 return true; 134 135 // We don't need to check for the module path, because if it can't be 136 // externally referenced and we call it, it is necessarilly in the same 137 // module 138 return canBeExternallyReferenced(**Summaries->second.begin()); 139 } 140 141 // Return true if the global described by \p Summary can be imported in another 142 // module. 143 static bool eligibleForImport(const ModuleSummaryIndex &Index, 144 const GlobalValueSummary &Summary) { 145 if (!canBeExternallyReferenced(Summary)) 146 // Can't import a global that needs renaming if has a section for instance. 147 // FIXME: we may be able to import it by copying it without promotion. 148 return false; 149 150 // Don't import functions that are not viable to inline. 151 if (Summary.isNotViableToInline()) 152 return false; 153 154 // Check references (and potential calls) in the same module. If the current 155 // value references a global that can't be externally referenced it is not 156 // eligible for import. First check the flag set when we have possible 157 // opaque references (e.g. inline asm calls), then check the call and 158 // reference sets. 159 if (Summary.hasInlineAsmMaybeReferencingInternal()) 160 return false; 161 bool AllRefsCanBeExternallyReferenced = 162 llvm::all_of(Summary.refs(), [&](const ValueInfo &VI) { 163 return canBeExternallyReferenced(Index, VI.getGUID()); 164 }); 165 if (!AllRefsCanBeExternallyReferenced) 166 return false; 167 168 if (auto *FuncSummary = dyn_cast<FunctionSummary>(&Summary)) { 169 bool AllCallsCanBeExternallyReferenced = llvm::all_of( 170 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) { 171 return canBeExternallyReferenced(Index, Edge.first.getGUID()); 172 }); 173 if (!AllCallsCanBeExternallyReferenced) 174 return false; 175 } 176 return true; 177 } 178 179 /// Given a list of possible callee implementation for a call site, select one 180 /// that fits the \p Threshold. 181 /// 182 /// FIXME: select "best" instead of first that fits. But what is "best"? 183 /// - The smallest: more likely to be inlined. 184 /// - The one with the least outgoing edges (already well optimized). 185 /// - One from a module already being imported from in order to reduce the 186 /// number of source modules parsed/linked. 187 /// - One that has PGO data attached. 188 /// - [insert you fancy metric here] 189 static const GlobalValueSummary * 190 selectCallee(const ModuleSummaryIndex &Index, 191 const GlobalValueSummaryList &CalleeSummaryList, 192 unsigned Threshold) { 193 auto It = llvm::find_if( 194 CalleeSummaryList, 195 [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) { 196 auto *GVSummary = SummaryPtr.get(); 197 if (GlobalValue::isInterposableLinkage(GVSummary->linkage())) 198 // There is no point in importing these, we can't inline them 199 return false; 200 if (auto *AS = dyn_cast<AliasSummary>(GVSummary)) { 201 GVSummary = &AS->getAliasee(); 202 // Alias can't point to "available_externally". However when we import 203 // linkOnceODR the linkage does not change. So we import the alias 204 // and aliasee only in this case. 205 // FIXME: we should import alias as available_externally *function*, 206 // the destination module does need to know it is an alias. 207 if (!GlobalValue::isLinkOnceODRLinkage(GVSummary->linkage())) 208 return false; 209 } 210 211 auto *Summary = cast<FunctionSummary>(GVSummary); 212 213 if (Summary->instCount() > Threshold) 214 return false; 215 216 if (!eligibleForImport(Index, *Summary)) 217 return false; 218 219 return true; 220 }); 221 if (It == CalleeSummaryList.end()) 222 return nullptr; 223 224 return cast<GlobalValueSummary>(It->get()); 225 } 226 227 /// Return the summary for the function \p GUID that fits the \p Threshold, or 228 /// null if there's no match. 229 static const GlobalValueSummary *selectCallee(GlobalValue::GUID GUID, 230 unsigned Threshold, 231 const ModuleSummaryIndex &Index) { 232 auto CalleeSummaryList = Index.findGlobalValueSummaryList(GUID); 233 if (CalleeSummaryList == Index.end()) 234 return nullptr; // This function does not have a summary 235 return selectCallee(Index, CalleeSummaryList->second, Threshold); 236 } 237 238 using EdgeInfo = std::tuple<const FunctionSummary *, unsigned /* Threshold */, 239 GlobalValue::GUID>; 240 241 /// Compute the list of functions to import for a given caller. Mark these 242 /// imported functions and the symbols they reference in their source module as 243 /// exported from their source module. 244 static void computeImportForFunction( 245 const FunctionSummary &Summary, const ModuleSummaryIndex &Index, 246 const unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries, 247 SmallVectorImpl<EdgeInfo> &Worklist, 248 FunctionImporter::ImportMapTy &ImportList, 249 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) { 250 for (auto &Edge : Summary.calls()) { 251 auto GUID = Edge.first.getGUID(); 252 DEBUG(dbgs() << " edge -> " << GUID << " Threshold:" << Threshold << "\n"); 253 254 if (DefinedGVSummaries.count(GUID)) { 255 DEBUG(dbgs() << "ignored! Target already in destination module.\n"); 256 continue; 257 } 258 259 auto GetBonusMultiplier = [](CalleeInfo::HotnessType Hotness) -> float { 260 if (Hotness == CalleeInfo::HotnessType::Hot) 261 return ImportHotMultiplier; 262 if (Hotness == CalleeInfo::HotnessType::Cold) 263 return ImportColdMultiplier; 264 return 1.0; 265 }; 266 267 const auto NewThreshold = 268 Threshold * GetBonusMultiplier(Edge.second.Hotness); 269 270 auto *CalleeSummary = selectCallee(GUID, NewThreshold, Index); 271 if (!CalleeSummary) { 272 DEBUG(dbgs() << "ignored! No qualifying callee with summary found.\n"); 273 continue; 274 } 275 // "Resolve" the summary, traversing alias, 276 const FunctionSummary *ResolvedCalleeSummary; 277 if (isa<AliasSummary>(CalleeSummary)) { 278 ResolvedCalleeSummary = cast<FunctionSummary>( 279 &cast<AliasSummary>(CalleeSummary)->getAliasee()); 280 assert( 281 GlobalValue::isLinkOnceODRLinkage(ResolvedCalleeSummary->linkage()) && 282 "Unexpected alias to a non-linkonceODR in import list"); 283 } else 284 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary); 285 286 assert(ResolvedCalleeSummary->instCount() <= NewThreshold && 287 "selectCallee() didn't honor the threshold"); 288 289 auto GetAdjustedThreshold = [](unsigned Threshold, bool IsHotCallsite) { 290 // Adjust the threshold for next level of imported functions. 291 // The threshold is different for hot callsites because we can then 292 // inline chains of hot calls. 293 if (IsHotCallsite) 294 return Threshold * ImportHotInstrFactor; 295 return Threshold * ImportInstrFactor; 296 }; 297 298 bool IsHotCallsite = Edge.second.Hotness == CalleeInfo::HotnessType::Hot; 299 const auto AdjThreshold = GetAdjustedThreshold(Threshold, IsHotCallsite); 300 301 auto ExportModulePath = ResolvedCalleeSummary->modulePath(); 302 auto &ProcessedThreshold = ImportList[ExportModulePath][GUID]; 303 /// Since the traversal of the call graph is DFS, we can revisit a function 304 /// a second time with a higher threshold. In this case, it is added back to 305 /// the worklist with the new threshold. 306 if (ProcessedThreshold && ProcessedThreshold >= AdjThreshold) { 307 DEBUG(dbgs() << "ignored! Target was already seen with Threshold " 308 << ProcessedThreshold << "\n"); 309 continue; 310 } 311 bool PreviouslyImported = ProcessedThreshold != 0; 312 // Mark this function as imported in this module, with the current Threshold 313 ProcessedThreshold = AdjThreshold; 314 315 // Make exports in the source module. 316 if (ExportLists) { 317 auto &ExportList = (*ExportLists)[ExportModulePath]; 318 ExportList.insert(GUID); 319 if (!PreviouslyImported) { 320 // This is the first time this function was exported from its source 321 // module, so mark all functions and globals it references as exported 322 // to the outside if they are defined in the same source module. 323 // For efficiency, we unconditionally add all the referenced GUIDs 324 // to the ExportList for this module, and will prune out any not 325 // defined in the module later in a single pass. 326 for (auto &Edge : ResolvedCalleeSummary->calls()) { 327 auto CalleeGUID = Edge.first.getGUID(); 328 ExportList.insert(CalleeGUID); 329 } 330 for (auto &Ref : ResolvedCalleeSummary->refs()) { 331 auto GUID = Ref.getGUID(); 332 ExportList.insert(GUID); 333 } 334 } 335 } 336 337 // Insert the newly imported function to the worklist. 338 Worklist.emplace_back(ResolvedCalleeSummary, AdjThreshold, GUID); 339 } 340 } 341 342 /// Given the list of globals defined in a module, compute the list of imports 343 /// as well as the list of "exports", i.e. the list of symbols referenced from 344 /// another module (that may require promotion). 345 static void ComputeImportForModule( 346 const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index, 347 FunctionImporter::ImportMapTy &ImportList, 348 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) { 349 // Worklist contains the list of function imported in this module, for which 350 // we will analyse the callees and may import further down the callgraph. 351 SmallVector<EdgeInfo, 128> Worklist; 352 353 // Populate the worklist with the import for the functions in the current 354 // module 355 for (auto &GVSummary : DefinedGVSummaries) { 356 auto *Summary = GVSummary.second; 357 if (auto *AS = dyn_cast<AliasSummary>(Summary)) 358 Summary = &AS->getAliasee(); 359 auto *FuncSummary = dyn_cast<FunctionSummary>(Summary); 360 if (!FuncSummary) 361 // Skip import for global variables 362 continue; 363 DEBUG(dbgs() << "Initalize import for " << GVSummary.first << "\n"); 364 computeImportForFunction(*FuncSummary, Index, ImportInstrLimit, 365 DefinedGVSummaries, Worklist, ImportList, 366 ExportLists); 367 } 368 369 // Process the newly imported functions and add callees to the worklist. 370 while (!Worklist.empty()) { 371 auto FuncInfo = Worklist.pop_back_val(); 372 auto *Summary = std::get<0>(FuncInfo); 373 auto Threshold = std::get<1>(FuncInfo); 374 auto GUID = std::get<2>(FuncInfo); 375 376 // Check if we later added this summary with a higher threshold. 377 // If so, skip this entry. 378 auto ExportModulePath = Summary->modulePath(); 379 auto &LatestProcessedThreshold = ImportList[ExportModulePath][GUID]; 380 if (LatestProcessedThreshold > Threshold) 381 continue; 382 383 computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries, 384 Worklist, ImportList, ExportLists); 385 } 386 } 387 388 } // anonymous namespace 389 390 /// Compute all the import and export for every module using the Index. 391 void llvm::ComputeCrossModuleImport( 392 const ModuleSummaryIndex &Index, 393 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 394 StringMap<FunctionImporter::ImportMapTy> &ImportLists, 395 StringMap<FunctionImporter::ExportSetTy> &ExportLists) { 396 // For each module that has function defined, compute the import/export lists. 397 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) { 398 auto &ImportList = ImportLists[DefinedGVSummaries.first()]; 399 DEBUG(dbgs() << "Computing import for Module '" 400 << DefinedGVSummaries.first() << "'\n"); 401 ComputeImportForModule(DefinedGVSummaries.second, Index, ImportList, 402 &ExportLists); 403 } 404 405 // When computing imports we added all GUIDs referenced by anything 406 // imported from the module to its ExportList. Now we prune each ExportList 407 // of any not defined in that module. This is more efficient than checking 408 // while computing imports because some of the summary lists may be long 409 // due to linkonce (comdat) copies. 410 for (auto &ELI : ExportLists) { 411 const auto &DefinedGVSummaries = 412 ModuleToDefinedGVSummaries.lookup(ELI.first()); 413 for (auto EI = ELI.second.begin(); EI != ELI.second.end();) { 414 if (!DefinedGVSummaries.count(*EI)) 415 EI = ELI.second.erase(EI); 416 else 417 ++EI; 418 } 419 } 420 421 #ifndef NDEBUG 422 DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size() 423 << " modules:\n"); 424 for (auto &ModuleImports : ImportLists) { 425 auto ModName = ModuleImports.first(); 426 auto &Exports = ExportLists[ModName]; 427 DEBUG(dbgs() << "* Module " << ModName << " exports " << Exports.size() 428 << " functions. Imports from " << ModuleImports.second.size() 429 << " modules.\n"); 430 for (auto &Src : ModuleImports.second) { 431 auto SrcModName = Src.first(); 432 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from " 433 << SrcModName << "\n"); 434 } 435 } 436 #endif 437 } 438 439 /// Compute all the imports for the given module in the Index. 440 void llvm::ComputeCrossModuleImportForModule( 441 StringRef ModulePath, const ModuleSummaryIndex &Index, 442 FunctionImporter::ImportMapTy &ImportList) { 443 444 // Collect the list of functions this module defines. 445 // GUID -> Summary 446 GVSummaryMapTy FunctionSummaryMap; 447 Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap); 448 449 // Compute the import list for this module. 450 DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n"); 451 ComputeImportForModule(FunctionSummaryMap, Index, ImportList); 452 453 #ifndef NDEBUG 454 DEBUG(dbgs() << "* Module " << ModulePath << " imports from " 455 << ImportList.size() << " modules.\n"); 456 for (auto &Src : ImportList) { 457 auto SrcModName = Src.first(); 458 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from " 459 << SrcModName << "\n"); 460 } 461 #endif 462 } 463 464 /// Compute the set of summaries needed for a ThinLTO backend compilation of 465 /// \p ModulePath. 466 void llvm::gatherImportedSummariesForModule( 467 StringRef ModulePath, 468 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 469 const FunctionImporter::ImportMapTy &ImportList, 470 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) { 471 // Include all summaries from the importing module. 472 ModuleToSummariesForIndex[ModulePath] = 473 ModuleToDefinedGVSummaries.lookup(ModulePath); 474 // Include summaries for imports. 475 for (auto &ILI : ImportList) { 476 auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()]; 477 const auto &DefinedGVSummaries = 478 ModuleToDefinedGVSummaries.lookup(ILI.first()); 479 for (auto &GI : ILI.second) { 480 const auto &DS = DefinedGVSummaries.find(GI.first); 481 assert(DS != DefinedGVSummaries.end() && 482 "Expected a defined summary for imported global value"); 483 SummariesForIndex[GI.first] = DS->second; 484 } 485 } 486 } 487 488 /// Emit the files \p ModulePath will import from into \p OutputFilename. 489 std::error_code 490 llvm::EmitImportsFiles(StringRef ModulePath, StringRef OutputFilename, 491 const FunctionImporter::ImportMapTy &ModuleImports) { 492 std::error_code EC; 493 raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::F_None); 494 if (EC) 495 return EC; 496 for (auto &ILI : ModuleImports) 497 ImportsOS << ILI.first() << "\n"; 498 return std::error_code(); 499 } 500 501 /// Fixup WeakForLinker linkages in \p TheModule based on summary analysis. 502 void llvm::thinLTOResolveWeakForLinkerModule( 503 Module &TheModule, const GVSummaryMapTy &DefinedGlobals) { 504 auto updateLinkage = [&](GlobalValue &GV) { 505 if (!GlobalValue::isWeakForLinker(GV.getLinkage())) 506 return; 507 // See if the global summary analysis computed a new resolved linkage. 508 const auto &GS = DefinedGlobals.find(GV.getGUID()); 509 if (GS == DefinedGlobals.end()) 510 return; 511 auto NewLinkage = GS->second->linkage(); 512 if (NewLinkage == GV.getLinkage()) 513 return; 514 DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName() << "` from " 515 << GV.getLinkage() << " to " << NewLinkage << "\n"); 516 GV.setLinkage(NewLinkage); 517 // Remove functions converted to available_externally from comdats, 518 // as this is a declaration for the linker, and will be dropped eventually. 519 // It is illegal for comdats to contain declarations. 520 auto *GO = dyn_cast_or_null<GlobalObject>(&GV); 521 if (GO && GO->isDeclarationForLinker() && GO->hasComdat()) { 522 assert(GO->hasAvailableExternallyLinkage() && 523 "Expected comdat on definition (possibly available external)"); 524 GO->setComdat(nullptr); 525 } 526 }; 527 528 // Process functions and global now 529 for (auto &GV : TheModule) 530 updateLinkage(GV); 531 for (auto &GV : TheModule.globals()) 532 updateLinkage(GV); 533 for (auto &GV : TheModule.aliases()) 534 updateLinkage(GV); 535 } 536 537 /// Run internalization on \p TheModule based on symmary analysis. 538 void llvm::thinLTOInternalizeModule(Module &TheModule, 539 const GVSummaryMapTy &DefinedGlobals) { 540 // Parse inline ASM and collect the list of symbols that are not defined in 541 // the current module. 542 StringSet<> AsmUndefinedRefs; 543 ModuleSymbolTable::CollectAsmSymbols( 544 Triple(TheModule.getTargetTriple()), TheModule.getModuleInlineAsm(), 545 [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) { 546 if (Flags & object::BasicSymbolRef::SF_Undefined) 547 AsmUndefinedRefs.insert(Name); 548 }); 549 550 // Declare a callback for the internalize pass that will ask for every 551 // candidate GlobalValue if it can be internalized or not. 552 auto MustPreserveGV = [&](const GlobalValue &GV) -> bool { 553 // Can't be internalized if referenced in inline asm. 554 if (AsmUndefinedRefs.count(GV.getName())) 555 return true; 556 557 // Lookup the linkage recorded in the summaries during global analysis. 558 const auto &GS = DefinedGlobals.find(GV.getGUID()); 559 GlobalValue::LinkageTypes Linkage; 560 if (GS == DefinedGlobals.end()) { 561 // Must have been promoted (possibly conservatively). Find original 562 // name so that we can access the correct summary and see if it can 563 // be internalized again. 564 // FIXME: Eventually we should control promotion instead of promoting 565 // and internalizing again. 566 StringRef OrigName = 567 ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName()); 568 std::string OrigId = GlobalValue::getGlobalIdentifier( 569 OrigName, GlobalValue::InternalLinkage, 570 TheModule.getSourceFileName()); 571 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId)); 572 if (GS == DefinedGlobals.end()) { 573 // Also check the original non-promoted non-globalized name. In some 574 // cases a preempted weak value is linked in as a local copy because 575 // it is referenced by an alias (IRLinker::linkGlobalValueProto). 576 // In that case, since it was originally not a local value, it was 577 // recorded in the index using the original name. 578 // FIXME: This may not be needed once PR27866 is fixed. 579 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName)); 580 assert(GS != DefinedGlobals.end()); 581 Linkage = GS->second->linkage(); 582 } else { 583 Linkage = GS->second->linkage(); 584 } 585 } else 586 Linkage = GS->second->linkage(); 587 return !GlobalValue::isLocalLinkage(Linkage); 588 }; 589 590 // FIXME: See if we can just internalize directly here via linkage changes 591 // based on the index, rather than invoking internalizeModule. 592 llvm::internalizeModule(TheModule, MustPreserveGV); 593 } 594 595 // Automatically import functions in Module \p DestModule based on the summaries 596 // index. 597 // 598 Expected<bool> FunctionImporter::importFunctions( 599 Module &DestModule, const FunctionImporter::ImportMapTy &ImportList, 600 bool ForceImportReferencedDiscardableSymbols) { 601 DEBUG(dbgs() << "Starting import for Module " 602 << DestModule.getModuleIdentifier() << "\n"); 603 unsigned ImportedCount = 0; 604 605 // Linker that will be used for importing function 606 Linker TheLinker(DestModule); 607 // Do the actual import of functions now, one Module at a time 608 std::set<StringRef> ModuleNameOrderedList; 609 for (auto &FunctionsToImportPerModule : ImportList) { 610 ModuleNameOrderedList.insert(FunctionsToImportPerModule.first()); 611 } 612 for (auto &Name : ModuleNameOrderedList) { 613 // Get the module for the import 614 const auto &FunctionsToImportPerModule = ImportList.find(Name); 615 assert(FunctionsToImportPerModule != ImportList.end()); 616 Expected<std::unique_ptr<Module>> SrcModuleOrErr = ModuleLoader(Name); 617 if (!SrcModuleOrErr) 618 return SrcModuleOrErr.takeError(); 619 std::unique_ptr<Module> SrcModule = std::move(*SrcModuleOrErr); 620 assert(&DestModule.getContext() == &SrcModule->getContext() && 621 "Context mismatch"); 622 623 // If modules were created with lazy metadata loading, materialize it 624 // now, before linking it (otherwise this will be a noop). 625 if (Error Err = SrcModule->materializeMetadata()) 626 return std::move(Err); 627 UpgradeDebugInfo(*SrcModule); 628 629 auto &ImportGUIDs = FunctionsToImportPerModule->second; 630 // Find the globals to import 631 DenseSet<const GlobalValue *> GlobalsToImport; 632 for (Function &F : *SrcModule) { 633 if (!F.hasName()) 634 continue; 635 auto GUID = F.getGUID(); 636 auto Import = ImportGUIDs.count(GUID); 637 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function " << GUID 638 << " " << F.getName() << " from " 639 << SrcModule->getSourceFileName() << "\n"); 640 if (Import) { 641 if (Error Err = F.materialize()) 642 return std::move(Err); 643 if (EnableImportMetadata) { 644 // Add 'thinlto_src_module' metadata for statistics and debugging. 645 F.setMetadata( 646 "thinlto_src_module", 647 llvm::MDNode::get( 648 DestModule.getContext(), 649 {llvm::MDString::get(DestModule.getContext(), 650 SrcModule->getSourceFileName())})); 651 } 652 GlobalsToImport.insert(&F); 653 } 654 } 655 for (GlobalVariable &GV : SrcModule->globals()) { 656 if (!GV.hasName()) 657 continue; 658 auto GUID = GV.getGUID(); 659 auto Import = ImportGUIDs.count(GUID); 660 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global " << GUID 661 << " " << GV.getName() << " from " 662 << SrcModule->getSourceFileName() << "\n"); 663 if (Import) { 664 if (Error Err = GV.materialize()) 665 return std::move(Err); 666 GlobalsToImport.insert(&GV); 667 } 668 } 669 for (GlobalAlias &GA : SrcModule->aliases()) { 670 if (!GA.hasName()) 671 continue; 672 auto GUID = GA.getGUID(); 673 auto Import = ImportGUIDs.count(GUID); 674 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias " << GUID 675 << " " << GA.getName() << " from " 676 << SrcModule->getSourceFileName() << "\n"); 677 if (Import) { 678 // Alias can't point to "available_externally". However when we import 679 // linkOnceODR the linkage does not change. So we import the alias 680 // and aliasee only in this case. This has been handled by 681 // computeImportForFunction() 682 GlobalObject *GO = GA.getBaseObject(); 683 assert(GO->hasLinkOnceODRLinkage() && 684 "Unexpected alias to a non-linkonceODR in import list"); 685 #ifndef NDEBUG 686 if (!GlobalsToImport.count(GO)) 687 DEBUG(dbgs() << " alias triggers importing aliasee " << GO->getGUID() 688 << " " << GO->getName() << " from " 689 << SrcModule->getSourceFileName() << "\n"); 690 #endif 691 if (Error Err = GO->materialize()) 692 return std::move(Err); 693 GlobalsToImport.insert(GO); 694 if (Error Err = GA.materialize()) 695 return std::move(Err); 696 GlobalsToImport.insert(&GA); 697 } 698 } 699 700 // Link in the specified functions. 701 if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport)) 702 return true; 703 704 if (PrintImports) { 705 for (const auto *GV : GlobalsToImport) 706 dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName() 707 << " from " << SrcModule->getSourceFileName() << "\n"; 708 } 709 710 // Instruct the linker that the client will take care of linkonce resolution 711 unsigned Flags = Linker::Flags::None; 712 if (!ForceImportReferencedDiscardableSymbols) 713 Flags |= Linker::Flags::DontForceLinkLinkonceODR; 714 715 if (TheLinker.linkInModule(std::move(SrcModule), Flags, &GlobalsToImport)) 716 report_fatal_error("Function Import: link error"); 717 718 ImportedCount += GlobalsToImport.size(); 719 } 720 721 NumImported += ImportedCount; 722 723 DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module " 724 << DestModule.getModuleIdentifier() << "\n"); 725 return ImportedCount; 726 } 727 728 /// Summary file to use for function importing when using -function-import from 729 /// the command line. 730 static cl::opt<std::string> 731 SummaryFile("summary-file", 732 cl::desc("The summary file to use for function importing.")); 733 734 static bool doImportingForModule(Module &M) { 735 if (SummaryFile.empty()) 736 report_fatal_error("error: -function-import requires -summary-file\n"); 737 Expected<std::unique_ptr<ModuleSummaryIndex>> IndexPtrOrErr = 738 getModuleSummaryIndexForFile(SummaryFile); 739 if (!IndexPtrOrErr) { 740 logAllUnhandledErrors(IndexPtrOrErr.takeError(), errs(), 741 "Error loading file '" + SummaryFile + "': "); 742 return false; 743 } 744 std::unique_ptr<ModuleSummaryIndex> Index = std::move(*IndexPtrOrErr); 745 746 // First step is collecting the import list. 747 FunctionImporter::ImportMapTy ImportList; 748 ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index, 749 ImportList); 750 751 // Conservatively mark all internal values as promoted. This interface is 752 // only used when doing importing via the function importing pass. The pass 753 // is only enabled when testing importing via the 'opt' tool, which does 754 // not do the ThinLink that would normally determine what values to promote. 755 for (auto &I : *Index) { 756 for (auto &S : I.second) { 757 if (GlobalValue::isLocalLinkage(S->linkage())) 758 S->setLinkage(GlobalValue::ExternalLinkage); 759 } 760 } 761 762 // Next we need to promote to global scope and rename any local values that 763 // are potentially exported to other modules. 764 if (renameModuleForThinLTO(M, *Index, nullptr)) { 765 errs() << "Error renaming module\n"; 766 return false; 767 } 768 769 // Perform the import now. 770 auto ModuleLoader = [&M](StringRef Identifier) { 771 return loadFile(Identifier, M.getContext()); 772 }; 773 FunctionImporter Importer(*Index, ModuleLoader); 774 Expected<bool> Result = Importer.importFunctions( 775 M, ImportList, !DontForceImportReferencedDiscardableSymbols); 776 777 // FIXME: Probably need to propagate Errors through the pass manager. 778 if (!Result) { 779 logAllUnhandledErrors(Result.takeError(), errs(), 780 "Error importing module: "); 781 return false; 782 } 783 784 return *Result; 785 } 786 787 namespace { 788 /// Pass that performs cross-module function import provided a summary file. 789 class FunctionImportLegacyPass : public ModulePass { 790 public: 791 /// Pass identification, replacement for typeid 792 static char ID; 793 794 /// Specify pass name for debug output 795 StringRef getPassName() const override { return "Function Importing"; } 796 797 explicit FunctionImportLegacyPass() : ModulePass(ID) {} 798 799 bool runOnModule(Module &M) override { 800 if (skipModule(M)) 801 return false; 802 803 return doImportingForModule(M); 804 } 805 }; 806 } // anonymous namespace 807 808 PreservedAnalyses FunctionImportPass::run(Module &M, 809 ModuleAnalysisManager &AM) { 810 if (!doImportingForModule(M)) 811 return PreservedAnalyses::all(); 812 813 return PreservedAnalyses::none(); 814 } 815 816 char FunctionImportLegacyPass::ID = 0; 817 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import", 818 "Summary Based Function Import", false, false) 819 820 namespace llvm { 821 Pass *createFunctionImportPass() { 822 return new FunctionImportLegacyPass(); 823 } 824 } 825