1 //===-LTO.cpp - LLVM Link Time Optimizer ----------------------------------===// 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 functions and classes used to support LTO. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/LTO/LTO.h" 15 #include "llvm/Analysis/TargetLibraryInfo.h" 16 #include "llvm/Analysis/TargetTransformInfo.h" 17 #include "llvm/Bitcode/BitcodeReader.h" 18 #include "llvm/Bitcode/BitcodeWriter.h" 19 #include "llvm/CodeGen/Analysis.h" 20 #include "llvm/IR/AutoUpgrade.h" 21 #include "llvm/IR/DiagnosticPrinter.h" 22 #include "llvm/IR/LegacyPassManager.h" 23 #include "llvm/LTO/LTOBackend.h" 24 #include "llvm/Linker/IRMover.h" 25 #include "llvm/Object/ModuleSummaryIndexObjectFile.h" 26 #include "llvm/Support/ManagedStatic.h" 27 #include "llvm/Support/MemoryBuffer.h" 28 #include "llvm/Support/Path.h" 29 #include "llvm/Support/SHA1.h" 30 #include "llvm/Support/SourceMgr.h" 31 #include "llvm/Support/TargetRegistry.h" 32 #include "llvm/Support/ThreadPool.h" 33 #include "llvm/Support/Threading.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Target/TargetMachine.h" 36 #include "llvm/Target/TargetOptions.h" 37 #include "llvm/Transforms/IPO.h" 38 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 39 #include "llvm/Transforms/Utils/SplitModule.h" 40 41 #include <set> 42 43 using namespace llvm; 44 using namespace lto; 45 using namespace object; 46 47 #define DEBUG_TYPE "lto" 48 49 // Returns a unique hash for the Module considering the current list of 50 // export/import and other global analysis results. 51 // The hash is produced in \p Key. 52 static void computeCacheKey( 53 SmallString<40> &Key, const Config &Conf, const ModuleSummaryIndex &Index, 54 StringRef ModuleID, const FunctionImporter::ImportMapTy &ImportList, 55 const FunctionImporter::ExportSetTy &ExportList, 56 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 57 const GVSummaryMapTy &DefinedGlobals) { 58 // Compute the unique hash for this entry. 59 // This is based on the current compiler version, the module itself, the 60 // export list, the hash for every single module in the import list, the 61 // list of ResolvedODR for the module, and the list of preserved symbols. 62 SHA1 Hasher; 63 64 // Start with the compiler revision 65 Hasher.update(LLVM_VERSION_STRING); 66 #ifdef HAVE_LLVM_REVISION 67 Hasher.update(LLVM_REVISION); 68 #endif 69 70 // Include the parts of the LTO configuration that affect code generation. 71 auto AddString = [&](StringRef Str) { 72 Hasher.update(Str); 73 Hasher.update(ArrayRef<uint8_t>{0}); 74 }; 75 auto AddUnsigned = [&](unsigned I) { 76 uint8_t Data[4]; 77 Data[0] = I; 78 Data[1] = I >> 8; 79 Data[2] = I >> 16; 80 Data[3] = I >> 24; 81 Hasher.update(ArrayRef<uint8_t>{Data, 4}); 82 }; 83 AddString(Conf.CPU); 84 // FIXME: Hash more of Options. For now all clients initialize Options from 85 // command-line flags (which is unsupported in production), but may set 86 // RelaxELFRelocations. The clang driver can also pass FunctionSections, 87 // DataSections and DebuggerTuning via command line flags. 88 AddUnsigned(Conf.Options.RelaxELFRelocations); 89 AddUnsigned(Conf.Options.FunctionSections); 90 AddUnsigned(Conf.Options.DataSections); 91 AddUnsigned((unsigned)Conf.Options.DebuggerTuning); 92 for (auto &A : Conf.MAttrs) 93 AddString(A); 94 AddUnsigned(Conf.RelocModel); 95 AddUnsigned(Conf.CodeModel); 96 AddUnsigned(Conf.CGOptLevel); 97 AddUnsigned(Conf.OptLevel); 98 AddString(Conf.OptPipeline); 99 AddString(Conf.AAPipeline); 100 AddString(Conf.OverrideTriple); 101 AddString(Conf.DefaultTriple); 102 103 // Include the hash for the current module 104 auto ModHash = Index.getModuleHash(ModuleID); 105 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 106 for (auto F : ExportList) 107 // The export list can impact the internalization, be conservative here 108 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&F, sizeof(F))); 109 110 // Include the hash for every module we import functions from 111 for (auto &Entry : ImportList) { 112 auto ModHash = Index.getModuleHash(Entry.first()); 113 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 114 } 115 116 // Include the hash for the resolved ODR. 117 for (auto &Entry : ResolvedODR) { 118 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first, 119 sizeof(GlobalValue::GUID))); 120 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second, 121 sizeof(GlobalValue::LinkageTypes))); 122 } 123 124 // Include the hash for the linkage type to reflect internalization and weak 125 // resolution. 126 for (auto &GS : DefinedGlobals) { 127 GlobalValue::LinkageTypes Linkage = GS.second->linkage(); 128 Hasher.update( 129 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage))); 130 } 131 132 if (!Conf.SampleProfile.empty()) { 133 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile); 134 if (FileOrErr) 135 Hasher.update(FileOrErr.get()->getBuffer()); 136 } 137 138 Key = toHex(Hasher.result()); 139 } 140 141 static void thinLTOResolveWeakForLinkerGUID( 142 GlobalValueSummaryList &GVSummaryList, GlobalValue::GUID GUID, 143 DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias, 144 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 145 isPrevailing, 146 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 147 recordNewLinkage) { 148 for (auto &S : GVSummaryList) { 149 GlobalValue::LinkageTypes OriginalLinkage = S->linkage(); 150 if (!GlobalValue::isWeakForLinker(OriginalLinkage)) 151 continue; 152 // We need to emit only one of these. The prevailing module will keep it, 153 // but turned into a weak, while the others will drop it when possible. 154 // This is both a compile-time optimization and a correctness 155 // transformation. This is necessary for correctness when we have exported 156 // a reference - we need to convert the linkonce to weak to 157 // ensure a copy is kept to satisfy the exported reference. 158 // FIXME: We may want to split the compile time and correctness 159 // aspects into separate routines. 160 if (isPrevailing(GUID, S.get())) { 161 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) 162 S->setLinkage(GlobalValue::getWeakLinkage( 163 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage))); 164 } 165 // Alias and aliasee can't be turned into available_externally. 166 else if (!isa<AliasSummary>(S.get()) && 167 !GlobalInvolvedWithAlias.count(S.get())) 168 S->setLinkage(GlobalValue::AvailableExternallyLinkage); 169 if (S->linkage() != OriginalLinkage) 170 recordNewLinkage(S->modulePath(), GUID, S->linkage()); 171 } 172 } 173 174 // Resolve Weak and LinkOnce values in the \p Index. 175 // 176 // We'd like to drop these functions if they are no longer referenced in the 177 // current module. However there is a chance that another module is still 178 // referencing them because of the import. We make sure we always emit at least 179 // one copy. 180 void llvm::thinLTOResolveWeakForLinkerInIndex( 181 ModuleSummaryIndex &Index, 182 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 183 isPrevailing, 184 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 185 recordNewLinkage) { 186 // We won't optimize the globals that are referenced by an alias for now 187 // Ideally we should turn the alias into a global and duplicate the definition 188 // when needed. 189 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias; 190 for (auto &I : Index) 191 for (auto &S : I.second) 192 if (auto AS = dyn_cast<AliasSummary>(S.get())) 193 GlobalInvolvedWithAlias.insert(&AS->getAliasee()); 194 195 for (auto &I : Index) 196 thinLTOResolveWeakForLinkerGUID(I.second, I.first, GlobalInvolvedWithAlias, 197 isPrevailing, recordNewLinkage); 198 } 199 200 static void thinLTOInternalizeAndPromoteGUID( 201 GlobalValueSummaryList &GVSummaryList, GlobalValue::GUID GUID, 202 function_ref<bool(StringRef, GlobalValue::GUID)> isExported) { 203 for (auto &S : GVSummaryList) { 204 if (isExported(S->modulePath(), GUID)) { 205 if (GlobalValue::isLocalLinkage(S->linkage())) 206 S->setLinkage(GlobalValue::ExternalLinkage); 207 } else if (!GlobalValue::isLocalLinkage(S->linkage())) 208 S->setLinkage(GlobalValue::InternalLinkage); 209 } 210 } 211 212 // Update the linkages in the given \p Index to mark exported values 213 // as external and non-exported values as internal. 214 void llvm::thinLTOInternalizeAndPromoteInIndex( 215 ModuleSummaryIndex &Index, 216 function_ref<bool(StringRef, GlobalValue::GUID)> isExported) { 217 for (auto &I : Index) 218 thinLTOInternalizeAndPromoteGUID(I.second, I.first, isExported); 219 } 220 221 struct InputFile::InputModule { 222 BitcodeModule BM; 223 std::unique_ptr<Module> Mod; 224 225 // The range of ModuleSymbolTable entries for this input module. 226 size_t SymBegin, SymEnd; 227 }; 228 229 // Requires a destructor for std::vector<InputModule>. 230 InputFile::~InputFile() = default; 231 232 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) { 233 std::unique_ptr<InputFile> File(new InputFile); 234 235 ErrorOr<MemoryBufferRef> BCOrErr = 236 IRObjectFile::findBitcodeInMemBuffer(Object); 237 if (!BCOrErr) 238 return errorCodeToError(BCOrErr.getError()); 239 240 Expected<std::vector<BitcodeModule>> BMsOrErr = 241 getBitcodeModuleList(*BCOrErr); 242 if (!BMsOrErr) 243 return BMsOrErr.takeError(); 244 245 if (BMsOrErr->empty()) 246 return make_error<StringError>("Bitcode file does not contain any modules", 247 inconvertibleErrorCode()); 248 249 // Create an InputModule for each module in the InputFile, and add it to the 250 // ModuleSymbolTable. 251 for (auto BM : *BMsOrErr) { 252 Expected<std::unique_ptr<Module>> MOrErr = 253 BM.getLazyModule(File->Ctx, /*ShouldLazyLoadMetadata*/ true, 254 /*IsImporting*/ false); 255 if (!MOrErr) 256 return MOrErr.takeError(); 257 258 size_t SymBegin = File->SymTab.symbols().size(); 259 File->SymTab.addModule(MOrErr->get()); 260 size_t SymEnd = File->SymTab.symbols().size(); 261 262 for (const auto &C : (*MOrErr)->getComdatSymbolTable()) { 263 auto P = File->ComdatMap.insert( 264 std::make_pair(&C.second, File->Comdats.size())); 265 assert(P.second); 266 (void)P; 267 File->Comdats.push_back(C.first()); 268 } 269 270 File->Mods.push_back({BM, std::move(*MOrErr), SymBegin, SymEnd}); 271 } 272 273 return std::move(File); 274 } 275 276 Expected<int> InputFile::Symbol::getComdatIndex() const { 277 if (!isGV()) 278 return -1; 279 const GlobalObject *GO = getGV()->getBaseObject(); 280 if (!GO) 281 return make_error<StringError>("Unable to determine comdat of alias!", 282 inconvertibleErrorCode()); 283 if (const Comdat *C = GO->getComdat()) { 284 auto I = File->ComdatMap.find(C); 285 assert(I != File->ComdatMap.end()); 286 return I->second; 287 } 288 return -1; 289 } 290 291 StringRef InputFile::getName() const { 292 return Mods[0].BM.getModuleIdentifier(); 293 } 294 295 StringRef InputFile::getSourceFileName() const { 296 return Mods[0].Mod->getSourceFileName(); 297 } 298 299 iterator_range<InputFile::symbol_iterator> 300 InputFile::module_symbols(InputModule &IM) { 301 return llvm::make_range( 302 symbol_iterator(SymTab.symbols().data() + IM.SymBegin, SymTab, this), 303 symbol_iterator(SymTab.symbols().data() + IM.SymEnd, SymTab, this)); 304 } 305 306 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel, 307 Config &Conf) 308 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel), 309 Ctx(Conf) {} 310 311 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend) : Backend(Backend) { 312 if (!Backend) 313 this->Backend = 314 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency()); 315 } 316 317 LTO::LTO(Config Conf, ThinBackend Backend, 318 unsigned ParallelCodeGenParallelismLevel) 319 : Conf(std::move(Conf)), 320 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf), 321 ThinLTO(std::move(Backend)) {} 322 323 // Requires a destructor for MapVector<BitcodeModule>. 324 LTO::~LTO() = default; 325 326 // Add the given symbol to the GlobalResolutions map, and resolve its partition. 327 void LTO::addSymbolToGlobalRes(SmallPtrSet<GlobalValue *, 8> &Used, 328 const InputFile::Symbol &Sym, 329 SymbolResolution Res, unsigned Partition) { 330 GlobalValue *GV = Sym.isGV() ? Sym.getGV() : nullptr; 331 332 auto &GlobalRes = GlobalResolutions[Sym.getName()]; 333 if (GV) { 334 GlobalRes.UnnamedAddr &= GV->hasGlobalUnnamedAddr(); 335 if (Res.Prevailing) 336 GlobalRes.IRName = GV->getName(); 337 } 338 // Set the partition to external if we know it is used elsewhere, e.g. 339 // it is visible to a regular object, is referenced from llvm.compiler_used, 340 // or was already recorded as being referenced from a different partition. 341 if (Res.VisibleToRegularObj || (GV && Used.count(GV)) || 342 (GlobalRes.Partition != GlobalResolution::Unknown && 343 GlobalRes.Partition != Partition)) { 344 GlobalRes.Partition = GlobalResolution::External; 345 } else 346 // First recorded reference, save the current partition. 347 GlobalRes.Partition = Partition; 348 349 // Flag as visible outside of ThinLTO if visible from a regular object or 350 // if this is a reference in the regular LTO partition. 351 GlobalRes.VisibleOutsideThinLTO |= 352 (Res.VisibleToRegularObj || (Partition == GlobalResolution::RegularLTO)); 353 } 354 355 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input, 356 ArrayRef<SymbolResolution> Res) { 357 StringRef Path = Input->getName(); 358 OS << Path << '\n'; 359 auto ResI = Res.begin(); 360 for (const InputFile::Symbol &Sym : Input->symbols()) { 361 assert(ResI != Res.end()); 362 SymbolResolution Res = *ResI++; 363 364 OS << "-r=" << Path << ',' << Sym.getName() << ','; 365 if (Res.Prevailing) 366 OS << 'p'; 367 if (Res.FinalDefinitionInLinkageUnit) 368 OS << 'l'; 369 if (Res.VisibleToRegularObj) 370 OS << 'x'; 371 OS << '\n'; 372 } 373 OS.flush(); 374 assert(ResI == Res.end()); 375 } 376 377 Error LTO::add(std::unique_ptr<InputFile> Input, 378 ArrayRef<SymbolResolution> Res) { 379 assert(!CalledGetMaxTasks); 380 381 if (Conf.ResolutionFile) 382 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res); 383 384 const SymbolResolution *ResI = Res.begin(); 385 for (InputFile::InputModule &IM : Input->Mods) 386 if (Error Err = addModule(*Input, IM, ResI, Res.end())) 387 return Err; 388 389 assert(ResI == Res.end()); 390 return Error::success(); 391 } 392 393 Error LTO::addModule(InputFile &Input, InputFile::InputModule &IM, 394 const SymbolResolution *&ResI, 395 const SymbolResolution *ResE) { 396 // FIXME: move to backend 397 Module &M = *IM.Mod; 398 399 if (M.getDataLayoutStr().empty()) 400 return make_error<StringError>("input module has no datalayout", 401 inconvertibleErrorCode()); 402 403 if (!Conf.OverrideTriple.empty()) 404 M.setTargetTriple(Conf.OverrideTriple); 405 else if (M.getTargetTriple().empty()) 406 M.setTargetTriple(Conf.DefaultTriple); 407 408 Expected<bool> HasThinLTOSummary = IM.BM.hasSummary(); 409 if (!HasThinLTOSummary) 410 return HasThinLTOSummary.takeError(); 411 412 if (*HasThinLTOSummary) 413 return addThinLTO(IM.BM, M, Input.module_symbols(IM), ResI, ResE); 414 else 415 return addRegularLTO(IM.BM, ResI, ResE); 416 } 417 418 // Add a regular LTO object to the link. 419 Error LTO::addRegularLTO(BitcodeModule BM, const SymbolResolution *&ResI, 420 const SymbolResolution *ResE) { 421 if (!RegularLTO.CombinedModule) { 422 RegularLTO.CombinedModule = 423 llvm::make_unique<Module>("ld-temp.o", RegularLTO.Ctx); 424 RegularLTO.Mover = llvm::make_unique<IRMover>(*RegularLTO.CombinedModule); 425 } 426 Expected<std::unique_ptr<Module>> MOrErr = 427 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true, 428 /*IsImporting*/ false); 429 if (!MOrErr) 430 return MOrErr.takeError(); 431 432 Module &M = **MOrErr; 433 if (Error Err = M.materializeMetadata()) 434 return Err; 435 UpgradeDebugInfo(M); 436 437 ModuleSymbolTable SymTab; 438 SymTab.addModule(&M); 439 440 SmallPtrSet<GlobalValue *, 8> Used; 441 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ false); 442 443 std::vector<GlobalValue *> Keep; 444 445 for (GlobalVariable &GV : M.globals()) 446 if (GV.hasAppendingLinkage()) 447 Keep.push_back(&GV); 448 449 DenseSet<GlobalObject *> AliasedGlobals; 450 for (auto &GA : M.aliases()) 451 if (GlobalObject *GO = GA.getBaseObject()) 452 AliasedGlobals.insert(GO); 453 454 for (const InputFile::Symbol &Sym : 455 make_range(InputFile::symbol_iterator(SymTab.symbols().begin(), SymTab, 456 nullptr), 457 InputFile::symbol_iterator(SymTab.symbols().end(), SymTab, 458 nullptr))) { 459 assert(ResI != ResE); 460 SymbolResolution Res = *ResI++; 461 addSymbolToGlobalRes(Used, Sym, Res, 0); 462 463 if (Sym.isGV()) { 464 GlobalValue *GV = Sym.getGV(); 465 if (Res.Prevailing) { 466 if (Sym.getFlags() & object::BasicSymbolRef::SF_Undefined) 467 continue; 468 Keep.push_back(GV); 469 switch (GV->getLinkage()) { 470 default: 471 break; 472 case GlobalValue::LinkOnceAnyLinkage: 473 GV->setLinkage(GlobalValue::WeakAnyLinkage); 474 break; 475 case GlobalValue::LinkOnceODRLinkage: 476 GV->setLinkage(GlobalValue::WeakODRLinkage); 477 break; 478 } 479 } else if (isa<GlobalObject>(GV) && 480 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() || 481 GV->hasAvailableExternallyLinkage()) && 482 !AliasedGlobals.count(cast<GlobalObject>(GV))) { 483 // Either of the above three types of linkage indicates that the 484 // chosen prevailing symbol will have the same semantics as this copy of 485 // the symbol, so we can link it with available_externally linkage. We 486 // only need to do this if the symbol is undefined. 487 GlobalValue *CombinedGV = 488 RegularLTO.CombinedModule->getNamedValue(GV->getName()); 489 if (!CombinedGV || CombinedGV->isDeclaration()) { 490 Keep.push_back(GV); 491 GV->setLinkage(GlobalValue::AvailableExternallyLinkage); 492 cast<GlobalObject>(GV)->setComdat(nullptr); 493 } 494 } 495 } 496 // Common resolution: collect the maximum size/alignment over all commons. 497 // We also record if we see an instance of a common as prevailing, so that 498 // if none is prevailing we can ignore it later. 499 if (Sym.getFlags() & object::BasicSymbolRef::SF_Common) { 500 // FIXME: We should figure out what to do about commons defined by asm. 501 // For now they aren't reported correctly by ModuleSymbolTable. 502 auto &CommonRes = RegularLTO.Commons[Sym.getGV()->getName()]; 503 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize()); 504 CommonRes.Align = std::max(CommonRes.Align, Sym.getCommonAlignment()); 505 CommonRes.Prevailing |= Res.Prevailing; 506 } 507 508 // FIXME: use proposed local attribute for FinalDefinitionInLinkageUnit. 509 } 510 511 return RegularLTO.Mover->move(std::move(*MOrErr), Keep, 512 [](GlobalValue &, IRMover::ValueAdder) {}, 513 /* LinkModuleInlineAsm */ true, 514 /* IsPerformingImport */ false); 515 } 516 517 // Add a ThinLTO object to the link. 518 // FIXME: This function should not need to take as many parameters once we have 519 // a bitcode symbol table. 520 Error LTO::addThinLTO(BitcodeModule BM, Module &M, 521 iterator_range<InputFile::symbol_iterator> Syms, 522 const SymbolResolution *&ResI, 523 const SymbolResolution *ResE) { 524 SmallPtrSet<GlobalValue *, 8> Used; 525 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ false); 526 527 Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = BM.getSummary(); 528 if (!SummaryOrErr) 529 return SummaryOrErr.takeError(); 530 ThinLTO.CombinedIndex.mergeFrom(std::move(*SummaryOrErr), 531 ThinLTO.ModuleMap.size()); 532 533 for (const InputFile::Symbol &Sym : Syms) { 534 assert(ResI != ResE); 535 SymbolResolution Res = *ResI++; 536 addSymbolToGlobalRes(Used, Sym, Res, ThinLTO.ModuleMap.size() + 1); 537 538 if (Res.Prevailing && Sym.isGV()) 539 ThinLTO.PrevailingModuleForGUID[Sym.getGV()->getGUID()] = 540 BM.getModuleIdentifier(); 541 } 542 543 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second) 544 return make_error<StringError>( 545 "Expected at most one ThinLTO module per bitcode file", 546 inconvertibleErrorCode()); 547 548 return Error::success(); 549 } 550 551 unsigned LTO::getMaxTasks() const { 552 CalledGetMaxTasks = true; 553 return RegularLTO.ParallelCodeGenParallelismLevel + ThinLTO.ModuleMap.size(); 554 } 555 556 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) { 557 // Save the status of having a regularLTO combined module, as 558 // this is needed for generating the ThinLTO Task ID, and 559 // the CombinedModule will be moved at the end of runRegularLTO. 560 bool HasRegularLTO = RegularLTO.CombinedModule != nullptr; 561 // Invoke regular LTO if there was a regular LTO module to start with. 562 if (HasRegularLTO) 563 if (auto E = runRegularLTO(AddStream)) 564 return E; 565 return runThinLTO(AddStream, Cache, HasRegularLTO); 566 } 567 568 Error LTO::runRegularLTO(AddStreamFn AddStream) { 569 // Make sure commons have the right size/alignment: we kept the largest from 570 // all the prevailing when adding the inputs, and we apply it here. 571 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout(); 572 for (auto &I : RegularLTO.Commons) { 573 if (!I.second.Prevailing) 574 // Don't do anything if no instance of this common was prevailing. 575 continue; 576 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first); 577 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) { 578 // Don't create a new global if the type is already correct, just make 579 // sure the alignment is correct. 580 OldGV->setAlignment(I.second.Align); 581 continue; 582 } 583 ArrayType *Ty = 584 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size); 585 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false, 586 GlobalValue::CommonLinkage, 587 ConstantAggregateZero::get(Ty), ""); 588 GV->setAlignment(I.second.Align); 589 if (OldGV) { 590 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType())); 591 GV->takeName(OldGV); 592 OldGV->eraseFromParent(); 593 } else { 594 GV->setName(I.first); 595 } 596 } 597 598 if (Conf.PreOptModuleHook && 599 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule)) 600 return Error::success(); 601 602 if (!Conf.CodeGenOnly) { 603 for (const auto &R : GlobalResolutions) { 604 if (R.second.IRName.empty()) 605 continue; 606 if (R.second.Partition != 0 && 607 R.second.Partition != GlobalResolution::External) 608 continue; 609 610 GlobalValue *GV = 611 RegularLTO.CombinedModule->getNamedValue(R.second.IRName); 612 // Ignore symbols defined in other partitions. 613 if (!GV || GV->hasLocalLinkage()) 614 continue; 615 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global 616 : GlobalValue::UnnamedAddr::None); 617 if (R.second.Partition == 0) 618 GV->setLinkage(GlobalValue::InternalLinkage); 619 } 620 621 if (Conf.PostInternalizeModuleHook && 622 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule)) 623 return Error::success(); 624 } 625 return backend(Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel, 626 std::move(RegularLTO.CombinedModule), ThinLTO.CombinedIndex); 627 } 628 629 /// This class defines the interface to the ThinLTO backend. 630 class lto::ThinBackendProc { 631 protected: 632 Config &Conf; 633 ModuleSummaryIndex &CombinedIndex; 634 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries; 635 636 public: 637 ThinBackendProc(Config &Conf, ModuleSummaryIndex &CombinedIndex, 638 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries) 639 : Conf(Conf), CombinedIndex(CombinedIndex), 640 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries) {} 641 642 virtual ~ThinBackendProc() {} 643 virtual Error start( 644 unsigned Task, BitcodeModule BM, 645 const FunctionImporter::ImportMapTy &ImportList, 646 const FunctionImporter::ExportSetTy &ExportList, 647 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 648 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0; 649 virtual Error wait() = 0; 650 }; 651 652 namespace { 653 class InProcessThinBackend : public ThinBackendProc { 654 ThreadPool BackendThreadPool; 655 AddStreamFn AddStream; 656 NativeObjectCache Cache; 657 658 Optional<Error> Err; 659 std::mutex ErrMu; 660 661 public: 662 InProcessThinBackend( 663 Config &Conf, ModuleSummaryIndex &CombinedIndex, 664 unsigned ThinLTOParallelismLevel, 665 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 666 AddStreamFn AddStream, NativeObjectCache Cache) 667 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 668 BackendThreadPool(ThinLTOParallelismLevel), 669 AddStream(std::move(AddStream)), Cache(std::move(Cache)) {} 670 671 Error runThinLTOBackendThread( 672 AddStreamFn AddStream, NativeObjectCache Cache, unsigned Task, 673 BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 674 const FunctionImporter::ImportMapTy &ImportList, 675 const FunctionImporter::ExportSetTy &ExportList, 676 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 677 const GVSummaryMapTy &DefinedGlobals, 678 MapVector<StringRef, BitcodeModule> &ModuleMap) { 679 auto RunThinBackend = [&](AddStreamFn AddStream) { 680 LTOLLVMContext BackendContext(Conf); 681 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext); 682 if (!MOrErr) 683 return MOrErr.takeError(); 684 685 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex, 686 ImportList, DefinedGlobals, ModuleMap); 687 }; 688 689 auto ModuleID = BM.getModuleIdentifier(); 690 691 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) || 692 all_of(CombinedIndex.getModuleHash(ModuleID), 693 [](uint32_t V) { return V == 0; })) 694 // Cache disabled or no entry for this module in the combined index or 695 // no module hash. 696 return RunThinBackend(AddStream); 697 698 SmallString<40> Key; 699 // The module may be cached, this helps handling it. 700 computeCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList, ExportList, 701 ResolvedODR, DefinedGlobals); 702 if (AddStreamFn CacheAddStream = Cache(Task, Key)) 703 return RunThinBackend(CacheAddStream); 704 705 return Error::success(); 706 } 707 708 Error start( 709 unsigned Task, BitcodeModule BM, 710 const FunctionImporter::ImportMapTy &ImportList, 711 const FunctionImporter::ExportSetTy &ExportList, 712 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 713 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 714 StringRef ModulePath = BM.getModuleIdentifier(); 715 assert(ModuleToDefinedGVSummaries.count(ModulePath)); 716 const GVSummaryMapTy &DefinedGlobals = 717 ModuleToDefinedGVSummaries.find(ModulePath)->second; 718 BackendThreadPool.async( 719 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 720 const FunctionImporter::ImportMapTy &ImportList, 721 const FunctionImporter::ExportSetTy &ExportList, 722 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> 723 &ResolvedODR, 724 const GVSummaryMapTy &DefinedGlobals, 725 MapVector<StringRef, BitcodeModule> &ModuleMap) { 726 Error E = runThinLTOBackendThread( 727 AddStream, Cache, Task, BM, CombinedIndex, ImportList, 728 ExportList, ResolvedODR, DefinedGlobals, ModuleMap); 729 if (E) { 730 std::unique_lock<std::mutex> L(ErrMu); 731 if (Err) 732 Err = joinErrors(std::move(*Err), std::move(E)); 733 else 734 Err = std::move(E); 735 } 736 }, 737 BM, std::ref(CombinedIndex), std::ref(ImportList), 738 std::ref(ExportList), std::ref(ResolvedODR), std::ref(DefinedGlobals), 739 std::ref(ModuleMap)); 740 return Error::success(); 741 } 742 743 Error wait() override { 744 BackendThreadPool.wait(); 745 if (Err) 746 return std::move(*Err); 747 else 748 return Error::success(); 749 } 750 }; 751 } // end anonymous namespace 752 753 ThinBackend lto::createInProcessThinBackend(unsigned ParallelismLevel) { 754 return [=](Config &Conf, ModuleSummaryIndex &CombinedIndex, 755 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 756 AddStreamFn AddStream, NativeObjectCache Cache) { 757 return llvm::make_unique<InProcessThinBackend>( 758 Conf, CombinedIndex, ParallelismLevel, ModuleToDefinedGVSummaries, 759 AddStream, Cache); 760 }; 761 } 762 763 // Given the original \p Path to an output file, replace any path 764 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the 765 // resulting directory if it does not yet exist. 766 std::string lto::getThinLTOOutputFile(const std::string &Path, 767 const std::string &OldPrefix, 768 const std::string &NewPrefix) { 769 if (OldPrefix.empty() && NewPrefix.empty()) 770 return Path; 771 SmallString<128> NewPath(Path); 772 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix); 773 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str()); 774 if (!ParentPath.empty()) { 775 // Make sure the new directory exists, creating it if necessary. 776 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath)) 777 llvm::errs() << "warning: could not create directory '" << ParentPath 778 << "': " << EC.message() << '\n'; 779 } 780 return NewPath.str(); 781 } 782 783 namespace { 784 class WriteIndexesThinBackend : public ThinBackendProc { 785 std::string OldPrefix, NewPrefix; 786 bool ShouldEmitImportsFiles; 787 788 std::string LinkedObjectsFileName; 789 std::unique_ptr<llvm::raw_fd_ostream> LinkedObjectsFile; 790 791 public: 792 WriteIndexesThinBackend( 793 Config &Conf, ModuleSummaryIndex &CombinedIndex, 794 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 795 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles, 796 std::string LinkedObjectsFileName) 797 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 798 OldPrefix(OldPrefix), NewPrefix(NewPrefix), 799 ShouldEmitImportsFiles(ShouldEmitImportsFiles), 800 LinkedObjectsFileName(LinkedObjectsFileName) {} 801 802 Error start( 803 unsigned Task, BitcodeModule BM, 804 const FunctionImporter::ImportMapTy &ImportList, 805 const FunctionImporter::ExportSetTy &ExportList, 806 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 807 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 808 StringRef ModulePath = BM.getModuleIdentifier(); 809 std::string NewModulePath = 810 getThinLTOOutputFile(ModulePath, OldPrefix, NewPrefix); 811 812 std::error_code EC; 813 if (!LinkedObjectsFileName.empty()) { 814 if (!LinkedObjectsFile) { 815 LinkedObjectsFile = llvm::make_unique<raw_fd_ostream>( 816 LinkedObjectsFileName, EC, sys::fs::OpenFlags::F_None); 817 if (EC) 818 return errorCodeToError(EC); 819 } 820 *LinkedObjectsFile << NewModulePath << '\n'; 821 } 822 823 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex; 824 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries, 825 ImportList, ModuleToSummariesForIndex); 826 827 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC, 828 sys::fs::OpenFlags::F_None); 829 if (EC) 830 return errorCodeToError(EC); 831 WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex); 832 833 if (ShouldEmitImportsFiles) 834 return errorCodeToError( 835 EmitImportsFiles(ModulePath, NewModulePath + ".imports", ImportList)); 836 return Error::success(); 837 } 838 839 Error wait() override { return Error::success(); } 840 }; 841 } // end anonymous namespace 842 843 ThinBackend lto::createWriteIndexesThinBackend(std::string OldPrefix, 844 std::string NewPrefix, 845 bool ShouldEmitImportsFiles, 846 std::string LinkedObjectsFile) { 847 return [=](Config &Conf, ModuleSummaryIndex &CombinedIndex, 848 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 849 AddStreamFn AddStream, NativeObjectCache Cache) { 850 return llvm::make_unique<WriteIndexesThinBackend>( 851 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix, 852 ShouldEmitImportsFiles, LinkedObjectsFile); 853 }; 854 } 855 856 Error LTO::runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache, 857 bool HasRegularLTO) { 858 if (ThinLTO.ModuleMap.empty()) 859 return Error::success(); 860 861 if (Conf.CombinedIndexHook && !Conf.CombinedIndexHook(ThinLTO.CombinedIndex)) 862 return Error::success(); 863 864 // Collect for each module the list of function it defines (GUID -> 865 // Summary). 866 StringMap<std::map<GlobalValue::GUID, GlobalValueSummary *>> 867 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size()); 868 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule( 869 ModuleToDefinedGVSummaries); 870 // Create entries for any modules that didn't have any GV summaries 871 // (either they didn't have any GVs to start with, or we suppressed 872 // generation of the summaries because they e.g. had inline assembly 873 // uses that couldn't be promoted/renamed on export). This is so 874 // InProcessThinBackend::start can still launch a backend thread, which 875 // is passed the map of summaries for the module, without any special 876 // handling for this case. 877 for (auto &Mod : ThinLTO.ModuleMap) 878 if (!ModuleToDefinedGVSummaries.count(Mod.first)) 879 ModuleToDefinedGVSummaries.try_emplace(Mod.first); 880 881 StringMap<FunctionImporter::ImportMapTy> ImportLists( 882 ThinLTO.ModuleMap.size()); 883 StringMap<FunctionImporter::ExportSetTy> ExportLists( 884 ThinLTO.ModuleMap.size()); 885 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; 886 887 if (Conf.OptLevel > 0) { 888 // Compute "dead" symbols, we don't want to import/export these! 889 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols; 890 for (auto &Res : GlobalResolutions) { 891 if (Res.second.VisibleOutsideThinLTO && 892 // IRName will be defined if we have seen the prevailing copy of 893 // this value. If not, no need to preserve any ThinLTO copies. 894 !Res.second.IRName.empty()) 895 GUIDPreservedSymbols.insert(GlobalValue::getGUID(Res.second.IRName)); 896 } 897 898 auto DeadSymbols = 899 computeDeadSymbols(ThinLTO.CombinedIndex, GUIDPreservedSymbols); 900 901 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 902 ImportLists, ExportLists, &DeadSymbols); 903 904 std::set<GlobalValue::GUID> ExportedGUIDs; 905 for (auto &Res : GlobalResolutions) { 906 // First check if the symbol was flagged as having external references. 907 if (Res.second.Partition != GlobalResolution::External) 908 continue; 909 // IRName will be defined if we have seen the prevailing copy of 910 // this value. If not, no need to mark as exported from a ThinLTO 911 // partition (and we can't get the GUID). 912 if (Res.second.IRName.empty()) 913 continue; 914 auto GUID = GlobalValue::getGUID(Res.second.IRName); 915 // Mark exported unless index-based analysis determined it to be dead. 916 if (!DeadSymbols.count(GUID)) 917 ExportedGUIDs.insert(GlobalValue::getGUID(Res.second.IRName)); 918 } 919 920 auto isPrevailing = [&](GlobalValue::GUID GUID, 921 const GlobalValueSummary *S) { 922 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath(); 923 }; 924 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) { 925 const auto &ExportList = ExportLists.find(ModuleIdentifier); 926 return (ExportList != ExportLists.end() && 927 ExportList->second.count(GUID)) || 928 ExportedGUIDs.count(GUID); 929 }; 930 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported); 931 932 auto recordNewLinkage = [&](StringRef ModuleIdentifier, 933 GlobalValue::GUID GUID, 934 GlobalValue::LinkageTypes NewLinkage) { 935 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage; 936 }; 937 938 thinLTOResolveWeakForLinkerInIndex(ThinLTO.CombinedIndex, isPrevailing, 939 recordNewLinkage); 940 } 941 942 std::unique_ptr<ThinBackendProc> BackendProc = 943 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 944 AddStream, Cache); 945 946 // Task numbers start at ParallelCodeGenParallelismLevel if an LTO 947 // module is present, as tasks 0 through ParallelCodeGenParallelismLevel-1 948 // are reserved for parallel code generation partitions. 949 unsigned Task = 950 HasRegularLTO ? RegularLTO.ParallelCodeGenParallelismLevel : 0; 951 for (auto &Mod : ThinLTO.ModuleMap) { 952 if (Error E = BackendProc->start(Task, Mod.second, ImportLists[Mod.first], 953 ExportLists[Mod.first], 954 ResolvedODR[Mod.first], ThinLTO.ModuleMap)) 955 return E; 956 ++Task; 957 } 958 959 return BackendProc->wait(); 960 } 961