1 //===-LTO.cpp - LLVM Link Time Optimizer ----------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements functions and classes used to support LTO. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/LTO/LTO.h" 14 #include "llvm/ADT/Statistic.h" 15 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 16 #include "llvm/Analysis/TargetLibraryInfo.h" 17 #include "llvm/Analysis/TargetTransformInfo.h" 18 #include "llvm/Bitcode/BitcodeReader.h" 19 #include "llvm/Bitcode/BitcodeWriter.h" 20 #include "llvm/CodeGen/Analysis.h" 21 #include "llvm/Config/llvm-config.h" 22 #include "llvm/IR/AutoUpgrade.h" 23 #include "llvm/IR/DiagnosticPrinter.h" 24 #include "llvm/IR/Intrinsics.h" 25 #include "llvm/IR/LLVMRemarkStreamer.h" 26 #include "llvm/IR/LegacyPassManager.h" 27 #include "llvm/IR/Mangler.h" 28 #include "llvm/IR/Metadata.h" 29 #include "llvm/LTO/LTOBackend.h" 30 #include "llvm/LTO/SummaryBasedOptimizations.h" 31 #include "llvm/Linker/IRMover.h" 32 #include "llvm/Object/IRObjectFile.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Error.h" 35 #include "llvm/Support/ManagedStatic.h" 36 #include "llvm/Support/MemoryBuffer.h" 37 #include "llvm/Support/Path.h" 38 #include "llvm/Support/SHA1.h" 39 #include "llvm/Support/SourceMgr.h" 40 #include "llvm/Support/TargetRegistry.h" 41 #include "llvm/Support/ThreadPool.h" 42 #include "llvm/Support/Threading.h" 43 #include "llvm/Support/TimeProfiler.h" 44 #include "llvm/Support/VCSRevision.h" 45 #include "llvm/Support/raw_ostream.h" 46 #include "llvm/Target/TargetMachine.h" 47 #include "llvm/Target/TargetOptions.h" 48 #include "llvm/Transforms/IPO.h" 49 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 50 #include "llvm/Transforms/IPO/WholeProgramDevirt.h" 51 #include "llvm/Transforms/Utils/FunctionImportUtils.h" 52 #include "llvm/Transforms/Utils/SplitModule.h" 53 54 #include <set> 55 56 using namespace llvm; 57 using namespace lto; 58 using namespace object; 59 60 #define DEBUG_TYPE "lto" 61 62 static cl::opt<bool> 63 DumpThinCGSCCs("dump-thin-cg-sccs", cl::init(false), cl::Hidden, 64 cl::desc("Dump the SCCs in the ThinLTO index's callgraph")); 65 66 /// Enable global value internalization in LTO. 67 cl::opt<bool> EnableLTOInternalization( 68 "enable-lto-internalization", cl::init(true), cl::Hidden, 69 cl::desc("Enable global value internalization in LTO")); 70 71 // Computes a unique hash for the Module considering the current list of 72 // export/import and other global analysis results. 73 // The hash is produced in \p Key. 74 void llvm::computeLTOCacheKey( 75 SmallString<40> &Key, const Config &Conf, const ModuleSummaryIndex &Index, 76 StringRef ModuleID, const FunctionImporter::ImportMapTy &ImportList, 77 const FunctionImporter::ExportSetTy &ExportList, 78 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 79 const GVSummaryMapTy &DefinedGlobals, 80 const std::set<GlobalValue::GUID> &CfiFunctionDefs, 81 const std::set<GlobalValue::GUID> &CfiFunctionDecls) { 82 // Compute the unique hash for this entry. 83 // This is based on the current compiler version, the module itself, the 84 // export list, the hash for every single module in the import list, the 85 // list of ResolvedODR for the module, and the list of preserved symbols. 86 SHA1 Hasher; 87 88 // Start with the compiler revision 89 Hasher.update(LLVM_VERSION_STRING); 90 #ifdef LLVM_REVISION 91 Hasher.update(LLVM_REVISION); 92 #endif 93 94 // Include the parts of the LTO configuration that affect code generation. 95 auto AddString = [&](StringRef Str) { 96 Hasher.update(Str); 97 Hasher.update(ArrayRef<uint8_t>{0}); 98 }; 99 auto AddUnsigned = [&](unsigned I) { 100 uint8_t Data[4]; 101 support::endian::write32le(Data, I); 102 Hasher.update(ArrayRef<uint8_t>{Data, 4}); 103 }; 104 auto AddUint64 = [&](uint64_t I) { 105 uint8_t Data[8]; 106 support::endian::write64le(Data, I); 107 Hasher.update(ArrayRef<uint8_t>{Data, 8}); 108 }; 109 AddString(Conf.CPU); 110 // FIXME: Hash more of Options. For now all clients initialize Options from 111 // command-line flags (which is unsupported in production), but may set 112 // RelaxELFRelocations. The clang driver can also pass FunctionSections, 113 // DataSections and DebuggerTuning via command line flags. 114 AddUnsigned(Conf.Options.RelaxELFRelocations); 115 AddUnsigned(Conf.Options.FunctionSections); 116 AddUnsigned(Conf.Options.DataSections); 117 AddUnsigned((unsigned)Conf.Options.DebuggerTuning); 118 for (auto &A : Conf.MAttrs) 119 AddString(A); 120 if (Conf.RelocModel) 121 AddUnsigned(*Conf.RelocModel); 122 else 123 AddUnsigned(-1); 124 if (Conf.CodeModel) 125 AddUnsigned(*Conf.CodeModel); 126 else 127 AddUnsigned(-1); 128 AddUnsigned(Conf.CGOptLevel); 129 AddUnsigned(Conf.CGFileType); 130 AddUnsigned(Conf.OptLevel); 131 AddUnsigned(Conf.UseNewPM); 132 AddUnsigned(Conf.Freestanding); 133 AddString(Conf.OptPipeline); 134 AddString(Conf.AAPipeline); 135 AddString(Conf.OverrideTriple); 136 AddString(Conf.DefaultTriple); 137 AddString(Conf.DwoDir); 138 139 // Include the hash for the current module 140 auto ModHash = Index.getModuleHash(ModuleID); 141 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 142 143 std::vector<uint64_t> ExportsGUID; 144 ExportsGUID.reserve(ExportList.size()); 145 for (const auto &VI : ExportList) { 146 auto GUID = VI.getGUID(); 147 ExportsGUID.push_back(GUID); 148 } 149 150 // Sort the export list elements GUIDs. 151 llvm::sort(ExportsGUID); 152 for (uint64_t GUID : ExportsGUID) { 153 // The export list can impact the internalization, be conservative here 154 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&GUID, sizeof(GUID))); 155 } 156 157 // Include the hash for every module we import functions from. The set of 158 // imported symbols for each module may affect code generation and is 159 // sensitive to link order, so include that as well. 160 using ImportMapIteratorTy = FunctionImporter::ImportMapTy::const_iterator; 161 std::vector<ImportMapIteratorTy> ImportModulesVector; 162 ImportModulesVector.reserve(ImportList.size()); 163 164 for (ImportMapIteratorTy It = ImportList.begin(); It != ImportList.end(); 165 ++It) { 166 ImportModulesVector.push_back(It); 167 } 168 llvm::sort(ImportModulesVector, 169 [](const ImportMapIteratorTy &Lhs, const ImportMapIteratorTy &Rhs) 170 -> bool { return Lhs->getKey() < Rhs->getKey(); }); 171 for (const ImportMapIteratorTy &EntryIt : ImportModulesVector) { 172 auto ModHash = Index.getModuleHash(EntryIt->first()); 173 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 174 175 AddUint64(EntryIt->second.size()); 176 for (auto &Fn : EntryIt->second) 177 AddUint64(Fn); 178 } 179 180 // Include the hash for the resolved ODR. 181 for (auto &Entry : ResolvedODR) { 182 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first, 183 sizeof(GlobalValue::GUID))); 184 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second, 185 sizeof(GlobalValue::LinkageTypes))); 186 } 187 188 // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or 189 // defined in this module. 190 std::set<GlobalValue::GUID> UsedCfiDefs; 191 std::set<GlobalValue::GUID> UsedCfiDecls; 192 193 // Typeids used in this module. 194 std::set<GlobalValue::GUID> UsedTypeIds; 195 196 auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) { 197 if (CfiFunctionDefs.count(ValueGUID)) 198 UsedCfiDefs.insert(ValueGUID); 199 if (CfiFunctionDecls.count(ValueGUID)) 200 UsedCfiDecls.insert(ValueGUID); 201 }; 202 203 auto AddUsedThings = [&](GlobalValueSummary *GS) { 204 if (!GS) return; 205 AddUnsigned(GS->isLive()); 206 AddUnsigned(GS->canAutoHide()); 207 for (const ValueInfo &VI : GS->refs()) { 208 AddUnsigned(VI.isDSOLocal()); 209 AddUsedCfiGlobal(VI.getGUID()); 210 } 211 if (auto *GVS = dyn_cast<GlobalVarSummary>(GS)) { 212 AddUnsigned(GVS->maybeReadOnly()); 213 AddUnsigned(GVS->maybeWriteOnly()); 214 } 215 if (auto *FS = dyn_cast<FunctionSummary>(GS)) { 216 for (auto &TT : FS->type_tests()) 217 UsedTypeIds.insert(TT); 218 for (auto &TT : FS->type_test_assume_vcalls()) 219 UsedTypeIds.insert(TT.GUID); 220 for (auto &TT : FS->type_checked_load_vcalls()) 221 UsedTypeIds.insert(TT.GUID); 222 for (auto &TT : FS->type_test_assume_const_vcalls()) 223 UsedTypeIds.insert(TT.VFunc.GUID); 224 for (auto &TT : FS->type_checked_load_const_vcalls()) 225 UsedTypeIds.insert(TT.VFunc.GUID); 226 for (auto &ET : FS->calls()) { 227 AddUnsigned(ET.first.isDSOLocal()); 228 AddUsedCfiGlobal(ET.first.getGUID()); 229 } 230 } 231 }; 232 233 // Include the hash for the linkage type to reflect internalization and weak 234 // resolution, and collect any used type identifier resolutions. 235 for (auto &GS : DefinedGlobals) { 236 GlobalValue::LinkageTypes Linkage = GS.second->linkage(); 237 Hasher.update( 238 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage))); 239 AddUsedCfiGlobal(GS.first); 240 AddUsedThings(GS.second); 241 } 242 243 // Imported functions may introduce new uses of type identifier resolutions, 244 // so we need to collect their used resolutions as well. 245 for (auto &ImpM : ImportList) 246 for (auto &ImpF : ImpM.second) { 247 GlobalValueSummary *S = Index.findSummaryInModule(ImpF, ImpM.first()); 248 AddUsedThings(S); 249 // If this is an alias, we also care about any types/etc. that the aliasee 250 // may reference. 251 if (auto *AS = dyn_cast_or_null<AliasSummary>(S)) 252 AddUsedThings(AS->getBaseObject()); 253 } 254 255 auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) { 256 AddString(TId); 257 258 AddUnsigned(S.TTRes.TheKind); 259 AddUnsigned(S.TTRes.SizeM1BitWidth); 260 261 AddUint64(S.TTRes.AlignLog2); 262 AddUint64(S.TTRes.SizeM1); 263 AddUint64(S.TTRes.BitMask); 264 AddUint64(S.TTRes.InlineBits); 265 266 AddUint64(S.WPDRes.size()); 267 for (auto &WPD : S.WPDRes) { 268 AddUnsigned(WPD.first); 269 AddUnsigned(WPD.second.TheKind); 270 AddString(WPD.second.SingleImplName); 271 272 AddUint64(WPD.second.ResByArg.size()); 273 for (auto &ByArg : WPD.second.ResByArg) { 274 AddUint64(ByArg.first.size()); 275 for (uint64_t Arg : ByArg.first) 276 AddUint64(Arg); 277 AddUnsigned(ByArg.second.TheKind); 278 AddUint64(ByArg.second.Info); 279 AddUnsigned(ByArg.second.Byte); 280 AddUnsigned(ByArg.second.Bit); 281 } 282 } 283 }; 284 285 // Include the hash for all type identifiers used by this module. 286 for (GlobalValue::GUID TId : UsedTypeIds) { 287 auto TidIter = Index.typeIds().equal_range(TId); 288 for (auto It = TidIter.first; It != TidIter.second; ++It) 289 AddTypeIdSummary(It->second.first, It->second.second); 290 } 291 292 AddUnsigned(UsedCfiDefs.size()); 293 for (auto &V : UsedCfiDefs) 294 AddUint64(V); 295 296 AddUnsigned(UsedCfiDecls.size()); 297 for (auto &V : UsedCfiDecls) 298 AddUint64(V); 299 300 if (!Conf.SampleProfile.empty()) { 301 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile); 302 if (FileOrErr) { 303 Hasher.update(FileOrErr.get()->getBuffer()); 304 305 if (!Conf.ProfileRemapping.empty()) { 306 FileOrErr = MemoryBuffer::getFile(Conf.ProfileRemapping); 307 if (FileOrErr) 308 Hasher.update(FileOrErr.get()->getBuffer()); 309 } 310 } 311 } 312 313 Key = toHex(Hasher.result()); 314 } 315 316 static void thinLTOResolvePrevailingGUID( 317 ValueInfo VI, DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias, 318 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 319 isPrevailing, 320 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 321 recordNewLinkage, 322 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 323 for (auto &S : VI.getSummaryList()) { 324 GlobalValue::LinkageTypes OriginalLinkage = S->linkage(); 325 // Ignore local and appending linkage values since the linker 326 // doesn't resolve them. 327 if (GlobalValue::isLocalLinkage(OriginalLinkage) || 328 GlobalValue::isAppendingLinkage(S->linkage())) 329 continue; 330 // We need to emit only one of these. The prevailing module will keep it, 331 // but turned into a weak, while the others will drop it when possible. 332 // This is both a compile-time optimization and a correctness 333 // transformation. This is necessary for correctness when we have exported 334 // a reference - we need to convert the linkonce to weak to 335 // ensure a copy is kept to satisfy the exported reference. 336 // FIXME: We may want to split the compile time and correctness 337 // aspects into separate routines. 338 if (isPrevailing(VI.getGUID(), S.get())) { 339 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) { 340 S->setLinkage(GlobalValue::getWeakLinkage( 341 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage))); 342 // The kept copy is eligible for auto-hiding (hidden visibility) if all 343 // copies were (i.e. they were all linkonce_odr global unnamed addr). 344 // If any copy is not (e.g. it was originally weak_odr), then the symbol 345 // must remain externally available (e.g. a weak_odr from an explicitly 346 // instantiated template). Additionally, if it is in the 347 // GUIDPreservedSymbols set, that means that it is visibile outside 348 // the summary (e.g. in a native object or a bitcode file without 349 // summary), and in that case we cannot hide it as it isn't possible to 350 // check all copies. 351 S->setCanAutoHide(VI.canAutoHide() && 352 !GUIDPreservedSymbols.count(VI.getGUID())); 353 } 354 } 355 // Alias and aliasee can't be turned into available_externally. 356 else if (!isa<AliasSummary>(S.get()) && 357 !GlobalInvolvedWithAlias.count(S.get())) 358 S->setLinkage(GlobalValue::AvailableExternallyLinkage); 359 if (S->linkage() != OriginalLinkage) 360 recordNewLinkage(S->modulePath(), VI.getGUID(), S->linkage()); 361 } 362 } 363 364 /// Resolve linkage for prevailing symbols in the \p Index. 365 // 366 // We'd like to drop these functions if they are no longer referenced in the 367 // current module. However there is a chance that another module is still 368 // referencing them because of the import. We make sure we always emit at least 369 // one copy. 370 void llvm::thinLTOResolvePrevailingInIndex( 371 ModuleSummaryIndex &Index, 372 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 373 isPrevailing, 374 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 375 recordNewLinkage, 376 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 377 // We won't optimize the globals that are referenced by an alias for now 378 // Ideally we should turn the alias into a global and duplicate the definition 379 // when needed. 380 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias; 381 for (auto &I : Index) 382 for (auto &S : I.second.SummaryList) 383 if (auto AS = dyn_cast<AliasSummary>(S.get())) 384 GlobalInvolvedWithAlias.insert(&AS->getAliasee()); 385 386 for (auto &I : Index) 387 thinLTOResolvePrevailingGUID(Index.getValueInfo(I), GlobalInvolvedWithAlias, 388 isPrevailing, recordNewLinkage, 389 GUIDPreservedSymbols); 390 } 391 392 static bool isWeakObjectWithRWAccess(GlobalValueSummary *GVS) { 393 if (auto *VarSummary = dyn_cast<GlobalVarSummary>(GVS->getBaseObject())) 394 return !VarSummary->maybeReadOnly() && !VarSummary->maybeWriteOnly() && 395 (VarSummary->linkage() == GlobalValue::WeakODRLinkage || 396 VarSummary->linkage() == GlobalValue::LinkOnceODRLinkage); 397 return false; 398 } 399 400 static void thinLTOInternalizeAndPromoteGUID( 401 ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported, 402 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 403 isPrevailing) { 404 for (auto &S : VI.getSummaryList()) { 405 if (isExported(S->modulePath(), VI)) { 406 if (GlobalValue::isLocalLinkage(S->linkage())) 407 S->setLinkage(GlobalValue::ExternalLinkage); 408 } else if (EnableLTOInternalization && 409 // Ignore local and appending linkage values since the linker 410 // doesn't resolve them. 411 !GlobalValue::isLocalLinkage(S->linkage()) && 412 (!GlobalValue::isInterposableLinkage(S->linkage()) || 413 isPrevailing(VI.getGUID(), S.get())) && 414 S->linkage() != GlobalValue::AppendingLinkage && 415 // We can't internalize available_externally globals because this 416 // can break function pointer equality. 417 S->linkage() != GlobalValue::AvailableExternallyLinkage && 418 // Functions and read-only variables with linkonce_odr and 419 // weak_odr linkage can be internalized. We can't internalize 420 // linkonce_odr and weak_odr variables which are both modified 421 // and read somewhere in the program because reads and writes 422 // will become inconsistent. 423 !isWeakObjectWithRWAccess(S.get())) 424 S->setLinkage(GlobalValue::InternalLinkage); 425 } 426 } 427 428 // Update the linkages in the given \p Index to mark exported values 429 // as external and non-exported values as internal. 430 void llvm::thinLTOInternalizeAndPromoteInIndex( 431 ModuleSummaryIndex &Index, 432 function_ref<bool(StringRef, ValueInfo)> isExported, 433 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 434 isPrevailing) { 435 for (auto &I : Index) 436 thinLTOInternalizeAndPromoteGUID(Index.getValueInfo(I), isExported, 437 isPrevailing); 438 } 439 440 // Requires a destructor for std::vector<InputModule>. 441 InputFile::~InputFile() = default; 442 443 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) { 444 std::unique_ptr<InputFile> File(new InputFile); 445 446 Expected<IRSymtabFile> FOrErr = readIRSymtab(Object); 447 if (!FOrErr) 448 return FOrErr.takeError(); 449 450 File->TargetTriple = FOrErr->TheReader.getTargetTriple(); 451 File->SourceFileName = FOrErr->TheReader.getSourceFileName(); 452 File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts(); 453 File->DependentLibraries = FOrErr->TheReader.getDependentLibraries(); 454 File->ComdatTable = FOrErr->TheReader.getComdatTable(); 455 456 for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) { 457 size_t Begin = File->Symbols.size(); 458 for (const irsymtab::Reader::SymbolRef &Sym : 459 FOrErr->TheReader.module_symbols(I)) 460 // Skip symbols that are irrelevant to LTO. Note that this condition needs 461 // to match the one in Skip() in LTO::addRegularLTO(). 462 if (Sym.isGlobal() && !Sym.isFormatSpecific()) 463 File->Symbols.push_back(Sym); 464 File->ModuleSymIndices.push_back({Begin, File->Symbols.size()}); 465 } 466 467 File->Mods = FOrErr->Mods; 468 File->Strtab = std::move(FOrErr->Strtab); 469 return std::move(File); 470 } 471 472 StringRef InputFile::getName() const { 473 return Mods[0].getModuleIdentifier(); 474 } 475 476 BitcodeModule &InputFile::getSingleBitcodeModule() { 477 assert(Mods.size() == 1 && "Expect only one bitcode module"); 478 return Mods[0]; 479 } 480 481 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel, 482 const Config &Conf) 483 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel), 484 Ctx(Conf), CombinedModule(std::make_unique<Module>("ld-temp.o", Ctx)), 485 Mover(std::make_unique<IRMover>(*CombinedModule)) {} 486 487 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend) 488 : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) { 489 if (!Backend) 490 this->Backend = 491 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency()); 492 } 493 494 LTO::LTO(Config Conf, ThinBackend Backend, 495 unsigned ParallelCodeGenParallelismLevel) 496 : Conf(std::move(Conf)), 497 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf), 498 ThinLTO(std::move(Backend)) {} 499 500 // Requires a destructor for MapVector<BitcodeModule>. 501 LTO::~LTO() = default; 502 503 // Add the symbols in the given module to the GlobalResolutions map, and resolve 504 // their partitions. 505 void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms, 506 ArrayRef<SymbolResolution> Res, 507 unsigned Partition, bool InSummary) { 508 auto *ResI = Res.begin(); 509 auto *ResE = Res.end(); 510 (void)ResE; 511 for (const InputFile::Symbol &Sym : Syms) { 512 assert(ResI != ResE); 513 SymbolResolution Res = *ResI++; 514 515 StringRef Name = Sym.getName(); 516 Triple TT(RegularLTO.CombinedModule->getTargetTriple()); 517 // Strip the __imp_ prefix from COFF dllimport symbols (similar to the 518 // way they are handled by lld), otherwise we can end up with two 519 // global resolutions (one with and one for a copy of the symbol without). 520 if (TT.isOSBinFormatCOFF() && Name.startswith("__imp_")) 521 Name = Name.substr(strlen("__imp_")); 522 auto &GlobalRes = GlobalResolutions[Name]; 523 GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr(); 524 if (Res.Prevailing) { 525 assert(!GlobalRes.Prevailing && 526 "Multiple prevailing defs are not allowed"); 527 GlobalRes.Prevailing = true; 528 GlobalRes.IRName = std::string(Sym.getIRName()); 529 } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) { 530 // Sometimes it can be two copies of symbol in a module and prevailing 531 // symbol can have no IR name. That might happen if symbol is defined in 532 // module level inline asm block. In case we have multiple modules with 533 // the same symbol we want to use IR name of the prevailing symbol. 534 // Otherwise, if we haven't seen a prevailing symbol, set the name so that 535 // we can later use it to check if there is any prevailing copy in IR. 536 GlobalRes.IRName = std::string(Sym.getIRName()); 537 } 538 539 // Set the partition to external if we know it is re-defined by the linker 540 // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a 541 // regular object, is referenced from llvm.compiler_used, or was already 542 // recorded as being referenced from a different partition. 543 if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() || 544 (GlobalRes.Partition != GlobalResolution::Unknown && 545 GlobalRes.Partition != Partition)) { 546 GlobalRes.Partition = GlobalResolution::External; 547 } else 548 // First recorded reference, save the current partition. 549 GlobalRes.Partition = Partition; 550 551 // Flag as visible outside of summary if visible from a regular object or 552 // from a module that does not have a summary. 553 GlobalRes.VisibleOutsideSummary |= 554 (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary); 555 } 556 } 557 558 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input, 559 ArrayRef<SymbolResolution> Res) { 560 StringRef Path = Input->getName(); 561 OS << Path << '\n'; 562 auto ResI = Res.begin(); 563 for (const InputFile::Symbol &Sym : Input->symbols()) { 564 assert(ResI != Res.end()); 565 SymbolResolution Res = *ResI++; 566 567 OS << "-r=" << Path << ',' << Sym.getName() << ','; 568 if (Res.Prevailing) 569 OS << 'p'; 570 if (Res.FinalDefinitionInLinkageUnit) 571 OS << 'l'; 572 if (Res.VisibleToRegularObj) 573 OS << 'x'; 574 if (Res.LinkerRedefined) 575 OS << 'r'; 576 OS << '\n'; 577 } 578 OS.flush(); 579 assert(ResI == Res.end()); 580 } 581 582 Error LTO::add(std::unique_ptr<InputFile> Input, 583 ArrayRef<SymbolResolution> Res) { 584 assert(!CalledGetMaxTasks); 585 586 if (Conf.ResolutionFile) 587 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res); 588 589 if (RegularLTO.CombinedModule->getTargetTriple().empty()) 590 RegularLTO.CombinedModule->setTargetTriple(Input->getTargetTriple()); 591 592 const SymbolResolution *ResI = Res.begin(); 593 for (unsigned I = 0; I != Input->Mods.size(); ++I) 594 if (Error Err = addModule(*Input, I, ResI, Res.end())) 595 return Err; 596 597 assert(ResI == Res.end()); 598 return Error::success(); 599 } 600 601 Error LTO::addModule(InputFile &Input, unsigned ModI, 602 const SymbolResolution *&ResI, 603 const SymbolResolution *ResE) { 604 Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo(); 605 if (!LTOInfo) 606 return LTOInfo.takeError(); 607 608 if (EnableSplitLTOUnit.hasValue()) { 609 // If only some modules were split, flag this in the index so that 610 // we can skip or error on optimizations that need consistently split 611 // modules (whole program devirt and lower type tests). 612 if (EnableSplitLTOUnit.getValue() != LTOInfo->EnableSplitLTOUnit) 613 ThinLTO.CombinedIndex.setPartiallySplitLTOUnits(); 614 } else 615 EnableSplitLTOUnit = LTOInfo->EnableSplitLTOUnit; 616 617 BitcodeModule BM = Input.Mods[ModI]; 618 auto ModSyms = Input.module_symbols(ModI); 619 addModuleToGlobalRes(ModSyms, {ResI, ResE}, 620 LTOInfo->IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0, 621 LTOInfo->HasSummary); 622 623 if (LTOInfo->IsThinLTO) 624 return addThinLTO(BM, ModSyms, ResI, ResE); 625 626 RegularLTO.EmptyCombinedModule = false; 627 Expected<RegularLTOState::AddedModule> ModOrErr = 628 addRegularLTO(BM, ModSyms, ResI, ResE); 629 if (!ModOrErr) 630 return ModOrErr.takeError(); 631 632 if (!LTOInfo->HasSummary) 633 return linkRegularLTO(std::move(*ModOrErr), /*LivenessFromIndex=*/false); 634 635 // Regular LTO module summaries are added to a dummy module that represents 636 // the combined regular LTO module. 637 if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, "", -1ull)) 638 return Err; 639 RegularLTO.ModsWithSummaries.push_back(std::move(*ModOrErr)); 640 return Error::success(); 641 } 642 643 // Checks whether the given global value is in a non-prevailing comdat 644 // (comdat containing values the linker indicated were not prevailing, 645 // which we then dropped to available_externally), and if so, removes 646 // it from the comdat. This is called for all global values to ensure the 647 // comdat is empty rather than leaving an incomplete comdat. It is needed for 648 // regular LTO modules, in case we are in a mixed-LTO mode (both regular 649 // and thin LTO modules) compilation. Since the regular LTO module will be 650 // linked first in the final native link, we want to make sure the linker 651 // doesn't select any of these incomplete comdats that would be left 652 // in the regular LTO module without this cleanup. 653 static void 654 handleNonPrevailingComdat(GlobalValue &GV, 655 std::set<const Comdat *> &NonPrevailingComdats) { 656 Comdat *C = GV.getComdat(); 657 if (!C) 658 return; 659 660 if (!NonPrevailingComdats.count(C)) 661 return; 662 663 // Additionally need to drop externally visible global values from the comdat 664 // to available_externally, so that there aren't multiply defined linker 665 // errors. 666 if (!GV.hasLocalLinkage()) 667 GV.setLinkage(GlobalValue::AvailableExternallyLinkage); 668 669 if (auto GO = dyn_cast<GlobalObject>(&GV)) 670 GO->setComdat(nullptr); 671 } 672 673 // Add a regular LTO object to the link. 674 // The resulting module needs to be linked into the combined LTO module with 675 // linkRegularLTO. 676 Expected<LTO::RegularLTOState::AddedModule> 677 LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, 678 const SymbolResolution *&ResI, 679 const SymbolResolution *ResE) { 680 RegularLTOState::AddedModule Mod; 681 Expected<std::unique_ptr<Module>> MOrErr = 682 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true, 683 /*IsImporting*/ false); 684 if (!MOrErr) 685 return MOrErr.takeError(); 686 Module &M = **MOrErr; 687 Mod.M = std::move(*MOrErr); 688 689 if (Error Err = M.materializeMetadata()) 690 return std::move(Err); 691 UpgradeDebugInfo(M); 692 693 ModuleSymbolTable SymTab; 694 SymTab.addModule(&M); 695 696 for (GlobalVariable &GV : M.globals()) 697 if (GV.hasAppendingLinkage()) 698 Mod.Keep.push_back(&GV); 699 700 DenseSet<GlobalObject *> AliasedGlobals; 701 for (auto &GA : M.aliases()) 702 if (GlobalObject *GO = GA.getBaseObject()) 703 AliasedGlobals.insert(GO); 704 705 // In this function we need IR GlobalValues matching the symbols in Syms 706 // (which is not backed by a module), so we need to enumerate them in the same 707 // order. The symbol enumeration order of a ModuleSymbolTable intentionally 708 // matches the order of an irsymtab, but when we read the irsymtab in 709 // InputFile::create we omit some symbols that are irrelevant to LTO. The 710 // Skip() function skips the same symbols from the module as InputFile does 711 // from the symbol table. 712 auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end(); 713 auto Skip = [&]() { 714 while (MsymI != MsymE) { 715 auto Flags = SymTab.getSymbolFlags(*MsymI); 716 if ((Flags & object::BasicSymbolRef::SF_Global) && 717 !(Flags & object::BasicSymbolRef::SF_FormatSpecific)) 718 return; 719 ++MsymI; 720 } 721 }; 722 Skip(); 723 724 std::set<const Comdat *> NonPrevailingComdats; 725 for (const InputFile::Symbol &Sym : Syms) { 726 assert(ResI != ResE); 727 SymbolResolution Res = *ResI++; 728 729 assert(MsymI != MsymE); 730 ModuleSymbolTable::Symbol Msym = *MsymI++; 731 Skip(); 732 733 if (GlobalValue *GV = Msym.dyn_cast<GlobalValue *>()) { 734 if (Res.Prevailing) { 735 if (Sym.isUndefined()) 736 continue; 737 Mod.Keep.push_back(GV); 738 // For symbols re-defined with linker -wrap and -defsym options, 739 // set the linkage to weak to inhibit IPO. The linkage will be 740 // restored by the linker. 741 if (Res.LinkerRedefined) 742 GV->setLinkage(GlobalValue::WeakAnyLinkage); 743 744 GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage(); 745 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) 746 GV->setLinkage(GlobalValue::getWeakLinkage( 747 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage))); 748 } else if (isa<GlobalObject>(GV) && 749 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() || 750 GV->hasAvailableExternallyLinkage()) && 751 !AliasedGlobals.count(cast<GlobalObject>(GV))) { 752 // Any of the above three types of linkage indicates that the 753 // chosen prevailing symbol will have the same semantics as this copy of 754 // the symbol, so we may be able to link it with available_externally 755 // linkage. We will decide later whether to do that when we link this 756 // module (in linkRegularLTO), based on whether it is undefined. 757 Mod.Keep.push_back(GV); 758 GV->setLinkage(GlobalValue::AvailableExternallyLinkage); 759 if (GV->hasComdat()) 760 NonPrevailingComdats.insert(GV->getComdat()); 761 cast<GlobalObject>(GV)->setComdat(nullptr); 762 } 763 764 // Set the 'local' flag based on the linker resolution for this symbol. 765 if (Res.FinalDefinitionInLinkageUnit) { 766 GV->setDSOLocal(true); 767 if (GV->hasDLLImportStorageClass()) 768 GV->setDLLStorageClass(GlobalValue::DLLStorageClassTypes:: 769 DefaultStorageClass); 770 } 771 } 772 // Common resolution: collect the maximum size/alignment over all commons. 773 // We also record if we see an instance of a common as prevailing, so that 774 // if none is prevailing we can ignore it later. 775 if (Sym.isCommon()) { 776 // FIXME: We should figure out what to do about commons defined by asm. 777 // For now they aren't reported correctly by ModuleSymbolTable. 778 auto &CommonRes = RegularLTO.Commons[std::string(Sym.getIRName())]; 779 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize()); 780 MaybeAlign SymAlign(Sym.getCommonAlignment()); 781 if (SymAlign) 782 CommonRes.Align = max(*SymAlign, CommonRes.Align); 783 CommonRes.Prevailing |= Res.Prevailing; 784 } 785 786 } 787 if (!M.getComdatSymbolTable().empty()) 788 for (GlobalValue &GV : M.global_values()) 789 handleNonPrevailingComdat(GV, NonPrevailingComdats); 790 assert(MsymI == MsymE); 791 return std::move(Mod); 792 } 793 794 Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod, 795 bool LivenessFromIndex) { 796 std::vector<GlobalValue *> Keep; 797 for (GlobalValue *GV : Mod.Keep) { 798 if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GV->getGUID())) { 799 if (Function *F = dyn_cast<Function>(GV)) { 800 OptimizationRemarkEmitter ORE(F); 801 ORE.emit(OptimizationRemark(DEBUG_TYPE, "deadfunction", F) 802 << ore::NV("Function", F) 803 << " not added to the combined module "); 804 } 805 continue; 806 } 807 808 if (!GV->hasAvailableExternallyLinkage()) { 809 Keep.push_back(GV); 810 continue; 811 } 812 813 // Only link available_externally definitions if we don't already have a 814 // definition. 815 GlobalValue *CombinedGV = 816 RegularLTO.CombinedModule->getNamedValue(GV->getName()); 817 if (CombinedGV && !CombinedGV->isDeclaration()) 818 continue; 819 820 Keep.push_back(GV); 821 } 822 823 return RegularLTO.Mover->move(std::move(Mod.M), Keep, 824 [](GlobalValue &, IRMover::ValueAdder) {}, 825 /* IsPerformingImport */ false); 826 } 827 828 // Add a ThinLTO module to the link. 829 Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, 830 const SymbolResolution *&ResI, 831 const SymbolResolution *ResE) { 832 if (Error Err = 833 BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(), 834 ThinLTO.ModuleMap.size())) 835 return Err; 836 837 for (const InputFile::Symbol &Sym : Syms) { 838 assert(ResI != ResE); 839 SymbolResolution Res = *ResI++; 840 841 if (!Sym.getIRName().empty()) { 842 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier( 843 Sym.getIRName(), GlobalValue::ExternalLinkage, "")); 844 if (Res.Prevailing) { 845 ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier(); 846 847 // For linker redefined symbols (via --wrap or --defsym) we want to 848 // switch the linkage to `weak` to prevent IPOs from happening. 849 // Find the summary in the module for this very GV and record the new 850 // linkage so that we can switch it when we import the GV. 851 if (Res.LinkerRedefined) 852 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( 853 GUID, BM.getModuleIdentifier())) 854 S->setLinkage(GlobalValue::WeakAnyLinkage); 855 } 856 857 // If the linker resolved the symbol to a local definition then mark it 858 // as local in the summary for the module we are adding. 859 if (Res.FinalDefinitionInLinkageUnit) { 860 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( 861 GUID, BM.getModuleIdentifier())) { 862 S->setDSOLocal(true); 863 } 864 } 865 } 866 } 867 868 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second) 869 return make_error<StringError>( 870 "Expected at most one ThinLTO module per bitcode file", 871 inconvertibleErrorCode()); 872 873 if (!Conf.ThinLTOModulesToCompile.empty()) { 874 if (!ThinLTO.ModulesToCompile) 875 ThinLTO.ModulesToCompile = ModuleMapType(); 876 // This is a fuzzy name matching where only modules with name containing the 877 // specified switch values are going to be compiled. 878 for (const std::string &Name : Conf.ThinLTOModulesToCompile) { 879 if (BM.getModuleIdentifier().contains(Name)) { 880 ThinLTO.ModulesToCompile->insert({BM.getModuleIdentifier(), BM}); 881 llvm::errs() << "[ThinLTO] Selecting " << BM.getModuleIdentifier() 882 << " to compile\n"; 883 } 884 } 885 } 886 887 return Error::success(); 888 } 889 890 unsigned LTO::getMaxTasks() const { 891 CalledGetMaxTasks = true; 892 auto ModuleCount = ThinLTO.ModulesToCompile ? ThinLTO.ModulesToCompile->size() 893 : ThinLTO.ModuleMap.size(); 894 return RegularLTO.ParallelCodeGenParallelismLevel + ModuleCount; 895 } 896 897 // If only some of the modules were split, we cannot correctly handle 898 // code that contains type tests or type checked loads. 899 Error LTO::checkPartiallySplit() { 900 if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits()) 901 return Error::success(); 902 903 Function *TypeTestFunc = RegularLTO.CombinedModule->getFunction( 904 Intrinsic::getName(Intrinsic::type_test)); 905 Function *TypeCheckedLoadFunc = RegularLTO.CombinedModule->getFunction( 906 Intrinsic::getName(Intrinsic::type_checked_load)); 907 908 // First check if there are type tests / type checked loads in the 909 // merged regular LTO module IR. 910 if ((TypeTestFunc && !TypeTestFunc->use_empty()) || 911 (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty())) 912 return make_error<StringError>( 913 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)", 914 inconvertibleErrorCode()); 915 916 // Otherwise check if there are any recorded in the combined summary from the 917 // ThinLTO modules. 918 for (auto &P : ThinLTO.CombinedIndex) { 919 for (auto &S : P.second.SummaryList) { 920 auto *FS = dyn_cast<FunctionSummary>(S.get()); 921 if (!FS) 922 continue; 923 if (!FS->type_test_assume_vcalls().empty() || 924 !FS->type_checked_load_vcalls().empty() || 925 !FS->type_test_assume_const_vcalls().empty() || 926 !FS->type_checked_load_const_vcalls().empty() || 927 !FS->type_tests().empty()) 928 return make_error<StringError>( 929 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)", 930 inconvertibleErrorCode()); 931 } 932 } 933 return Error::success(); 934 } 935 936 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) { 937 // Compute "dead" symbols, we don't want to import/export these! 938 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols; 939 DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions; 940 for (auto &Res : GlobalResolutions) { 941 // Normally resolution have IR name of symbol. We can do nothing here 942 // otherwise. See comments in GlobalResolution struct for more details. 943 if (Res.second.IRName.empty()) 944 continue; 945 946 GlobalValue::GUID GUID = GlobalValue::getGUID( 947 GlobalValue::dropLLVMManglingEscape(Res.second.IRName)); 948 949 if (Res.second.VisibleOutsideSummary && Res.second.Prevailing) 950 GUIDPreservedSymbols.insert(GUID); 951 952 GUIDPrevailingResolutions[GUID] = 953 Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No; 954 } 955 956 auto isPrevailing = [&](GlobalValue::GUID G) { 957 auto It = GUIDPrevailingResolutions.find(G); 958 if (It == GUIDPrevailingResolutions.end()) 959 return PrevailingType::Unknown; 960 return It->second; 961 }; 962 computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols, 963 isPrevailing, Conf.OptLevel > 0); 964 965 // Setup output file to emit statistics. 966 auto StatsFileOrErr = setupStatsFile(Conf.StatsFile); 967 if (!StatsFileOrErr) 968 return StatsFileOrErr.takeError(); 969 std::unique_ptr<ToolOutputFile> StatsFile = std::move(StatsFileOrErr.get()); 970 971 Error Result = runRegularLTO(AddStream); 972 if (!Result) 973 Result = runThinLTO(AddStream, Cache, GUIDPreservedSymbols); 974 975 if (StatsFile) 976 PrintStatisticsJSON(StatsFile->os()); 977 978 return Result; 979 } 980 981 Error LTO::runRegularLTO(AddStreamFn AddStream) { 982 // Setup optimization remarks. 983 auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks( 984 RegularLTO.CombinedModule->getContext(), Conf.RemarksFilename, 985 Conf.RemarksPasses, Conf.RemarksFormat, Conf.RemarksWithHotness); 986 if (!DiagFileOrErr) 987 return DiagFileOrErr.takeError(); 988 989 // Finalize linking of regular LTO modules containing summaries now that 990 // we have computed liveness information. 991 for (auto &M : RegularLTO.ModsWithSummaries) 992 if (Error Err = linkRegularLTO(std::move(M), 993 /*LivenessFromIndex=*/true)) 994 return Err; 995 996 // Ensure we don't have inconsistently split LTO units with type tests. 997 // FIXME: this checks both LTO and ThinLTO. It happens to work as we take 998 // this path both cases but eventually this should be split into two and 999 // do the ThinLTO checks in `runThinLTO`. 1000 if (Error Err = checkPartiallySplit()) 1001 return Err; 1002 1003 // Make sure commons have the right size/alignment: we kept the largest from 1004 // all the prevailing when adding the inputs, and we apply it here. 1005 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout(); 1006 for (auto &I : RegularLTO.Commons) { 1007 if (!I.second.Prevailing) 1008 // Don't do anything if no instance of this common was prevailing. 1009 continue; 1010 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first); 1011 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) { 1012 // Don't create a new global if the type is already correct, just make 1013 // sure the alignment is correct. 1014 OldGV->setAlignment(I.second.Align); 1015 continue; 1016 } 1017 ArrayType *Ty = 1018 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size); 1019 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false, 1020 GlobalValue::CommonLinkage, 1021 ConstantAggregateZero::get(Ty), ""); 1022 GV->setAlignment(I.second.Align); 1023 if (OldGV) { 1024 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType())); 1025 GV->takeName(OldGV); 1026 OldGV->eraseFromParent(); 1027 } else { 1028 GV->setName(I.first); 1029 } 1030 } 1031 1032 // If allowed, upgrade public vcall visibility metadata to linkage unit 1033 // visibility before whole program devirtualization in the optimizer. 1034 updateVCallVisibilityInModule(*RegularLTO.CombinedModule, 1035 Conf.HasWholeProgramVisibility); 1036 1037 if (Conf.PreOptModuleHook && 1038 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule)) 1039 return Error::success(); 1040 1041 if (!Conf.CodeGenOnly) { 1042 for (const auto &R : GlobalResolutions) { 1043 if (!R.second.isPrevailingIRSymbol()) 1044 continue; 1045 if (R.second.Partition != 0 && 1046 R.second.Partition != GlobalResolution::External) 1047 continue; 1048 1049 GlobalValue *GV = 1050 RegularLTO.CombinedModule->getNamedValue(R.second.IRName); 1051 // Ignore symbols defined in other partitions. 1052 // Also skip declarations, which are not allowed to have internal linkage. 1053 if (!GV || GV->hasLocalLinkage() || GV->isDeclaration()) 1054 continue; 1055 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global 1056 : GlobalValue::UnnamedAddr::None); 1057 if (EnableLTOInternalization && R.second.Partition == 0) 1058 GV->setLinkage(GlobalValue::InternalLinkage); 1059 } 1060 1061 RegularLTO.CombinedModule->addModuleFlag(Module::Error, "LTOPostLink", 1); 1062 1063 if (Conf.PostInternalizeModuleHook && 1064 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule)) 1065 return Error::success(); 1066 } 1067 1068 if (!RegularLTO.EmptyCombinedModule || Conf.AlwaysEmitRegularLTOObj) { 1069 if (Error Err = backend( 1070 Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel, 1071 std::move(RegularLTO.CombinedModule), ThinLTO.CombinedIndex)) 1072 return Err; 1073 } 1074 1075 return finalizeOptimizationRemarks(std::move(*DiagFileOrErr)); 1076 } 1077 1078 static const char *libcallRoutineNames[] = { 1079 #define HANDLE_LIBCALL(code, name) name, 1080 #include "llvm/IR/RuntimeLibcalls.def" 1081 #undef HANDLE_LIBCALL 1082 }; 1083 1084 ArrayRef<const char*> LTO::getRuntimeLibcallSymbols() { 1085 return makeArrayRef(libcallRoutineNames); 1086 } 1087 1088 /// This class defines the interface to the ThinLTO backend. 1089 class lto::ThinBackendProc { 1090 protected: 1091 const Config &Conf; 1092 ModuleSummaryIndex &CombinedIndex; 1093 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries; 1094 1095 public: 1096 ThinBackendProc(const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1097 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries) 1098 : Conf(Conf), CombinedIndex(CombinedIndex), 1099 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries) {} 1100 1101 virtual ~ThinBackendProc() {} 1102 virtual Error start( 1103 unsigned Task, BitcodeModule BM, 1104 const FunctionImporter::ImportMapTy &ImportList, 1105 const FunctionImporter::ExportSetTy &ExportList, 1106 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1107 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0; 1108 virtual Error wait() = 0; 1109 }; 1110 1111 namespace { 1112 class InProcessThinBackend : public ThinBackendProc { 1113 ThreadPool BackendThreadPool; 1114 AddStreamFn AddStream; 1115 NativeObjectCache Cache; 1116 std::set<GlobalValue::GUID> CfiFunctionDefs; 1117 std::set<GlobalValue::GUID> CfiFunctionDecls; 1118 1119 Optional<Error> Err; 1120 std::mutex ErrMu; 1121 1122 public: 1123 InProcessThinBackend( 1124 const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1125 ThreadPoolStrategy ThinLTOParallelism, 1126 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1127 AddStreamFn AddStream, NativeObjectCache Cache) 1128 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 1129 BackendThreadPool(ThinLTOParallelism), AddStream(std::move(AddStream)), 1130 Cache(std::move(Cache)) { 1131 for (auto &Name : CombinedIndex.cfiFunctionDefs()) 1132 CfiFunctionDefs.insert( 1133 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name))); 1134 for (auto &Name : CombinedIndex.cfiFunctionDecls()) 1135 CfiFunctionDecls.insert( 1136 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name))); 1137 } 1138 1139 Error runThinLTOBackendThread( 1140 AddStreamFn AddStream, NativeObjectCache Cache, unsigned Task, 1141 BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 1142 const FunctionImporter::ImportMapTy &ImportList, 1143 const FunctionImporter::ExportSetTy &ExportList, 1144 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1145 const GVSummaryMapTy &DefinedGlobals, 1146 MapVector<StringRef, BitcodeModule> &ModuleMap) { 1147 auto RunThinBackend = [&](AddStreamFn AddStream) { 1148 LTOLLVMContext BackendContext(Conf); 1149 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext); 1150 if (!MOrErr) 1151 return MOrErr.takeError(); 1152 1153 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex, 1154 ImportList, DefinedGlobals, ModuleMap); 1155 }; 1156 1157 auto ModuleID = BM.getModuleIdentifier(); 1158 1159 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) || 1160 all_of(CombinedIndex.getModuleHash(ModuleID), 1161 [](uint32_t V) { return V == 0; })) 1162 // Cache disabled or no entry for this module in the combined index or 1163 // no module hash. 1164 return RunThinBackend(AddStream); 1165 1166 SmallString<40> Key; 1167 // The module may be cached, this helps handling it. 1168 computeLTOCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList, 1169 ExportList, ResolvedODR, DefinedGlobals, CfiFunctionDefs, 1170 CfiFunctionDecls); 1171 if (AddStreamFn CacheAddStream = Cache(Task, Key)) 1172 return RunThinBackend(CacheAddStream); 1173 1174 return Error::success(); 1175 } 1176 1177 Error start( 1178 unsigned Task, BitcodeModule BM, 1179 const FunctionImporter::ImportMapTy &ImportList, 1180 const FunctionImporter::ExportSetTy &ExportList, 1181 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1182 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 1183 StringRef ModulePath = BM.getModuleIdentifier(); 1184 assert(ModuleToDefinedGVSummaries.count(ModulePath)); 1185 const GVSummaryMapTy &DefinedGlobals = 1186 ModuleToDefinedGVSummaries.find(ModulePath)->second; 1187 BackendThreadPool.async( 1188 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 1189 const FunctionImporter::ImportMapTy &ImportList, 1190 const FunctionImporter::ExportSetTy &ExportList, 1191 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> 1192 &ResolvedODR, 1193 const GVSummaryMapTy &DefinedGlobals, 1194 MapVector<StringRef, BitcodeModule> &ModuleMap) { 1195 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) 1196 timeTraceProfilerInitialize(Conf.TimeTraceGranularity, 1197 "thin backend"); 1198 Error E = runThinLTOBackendThread( 1199 AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList, 1200 ResolvedODR, DefinedGlobals, ModuleMap); 1201 if (E) { 1202 std::unique_lock<std::mutex> L(ErrMu); 1203 if (Err) 1204 Err = joinErrors(std::move(*Err), std::move(E)); 1205 else 1206 Err = std::move(E); 1207 } 1208 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) 1209 timeTraceProfilerFinishThread(); 1210 }, 1211 BM, std::ref(CombinedIndex), std::ref(ImportList), std::ref(ExportList), 1212 std::ref(ResolvedODR), std::ref(DefinedGlobals), std::ref(ModuleMap)); 1213 return Error::success(); 1214 } 1215 1216 Error wait() override { 1217 BackendThreadPool.wait(); 1218 if (Err) 1219 return std::move(*Err); 1220 else 1221 return Error::success(); 1222 } 1223 }; 1224 } // end anonymous namespace 1225 1226 ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism) { 1227 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1228 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1229 AddStreamFn AddStream, NativeObjectCache Cache) { 1230 return std::make_unique<InProcessThinBackend>( 1231 Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries, AddStream, 1232 Cache); 1233 }; 1234 } 1235 1236 // Given the original \p Path to an output file, replace any path 1237 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the 1238 // resulting directory if it does not yet exist. 1239 std::string lto::getThinLTOOutputFile(const std::string &Path, 1240 const std::string &OldPrefix, 1241 const std::string &NewPrefix) { 1242 if (OldPrefix.empty() && NewPrefix.empty()) 1243 return Path; 1244 SmallString<128> NewPath(Path); 1245 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix); 1246 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str()); 1247 if (!ParentPath.empty()) { 1248 // Make sure the new directory exists, creating it if necessary. 1249 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath)) 1250 llvm::errs() << "warning: could not create directory '" << ParentPath 1251 << "': " << EC.message() << '\n'; 1252 } 1253 return std::string(NewPath.str()); 1254 } 1255 1256 namespace { 1257 class WriteIndexesThinBackend : public ThinBackendProc { 1258 std::string OldPrefix, NewPrefix; 1259 bool ShouldEmitImportsFiles; 1260 raw_fd_ostream *LinkedObjectsFile; 1261 lto::IndexWriteCallback OnWrite; 1262 1263 public: 1264 WriteIndexesThinBackend( 1265 const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1266 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1267 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles, 1268 raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite) 1269 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 1270 OldPrefix(OldPrefix), NewPrefix(NewPrefix), 1271 ShouldEmitImportsFiles(ShouldEmitImportsFiles), 1272 LinkedObjectsFile(LinkedObjectsFile), OnWrite(OnWrite) {} 1273 1274 Error start( 1275 unsigned Task, BitcodeModule BM, 1276 const FunctionImporter::ImportMapTy &ImportList, 1277 const FunctionImporter::ExportSetTy &ExportList, 1278 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1279 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 1280 StringRef ModulePath = BM.getModuleIdentifier(); 1281 std::string NewModulePath = 1282 getThinLTOOutputFile(std::string(ModulePath), OldPrefix, NewPrefix); 1283 1284 if (LinkedObjectsFile) 1285 *LinkedObjectsFile << NewModulePath << '\n'; 1286 1287 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex; 1288 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries, 1289 ImportList, ModuleToSummariesForIndex); 1290 1291 std::error_code EC; 1292 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC, 1293 sys::fs::OpenFlags::OF_None); 1294 if (EC) 1295 return errorCodeToError(EC); 1296 WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex); 1297 1298 if (ShouldEmitImportsFiles) { 1299 EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports", 1300 ModuleToSummariesForIndex); 1301 if (EC) 1302 return errorCodeToError(EC); 1303 } 1304 1305 if (OnWrite) 1306 OnWrite(std::string(ModulePath)); 1307 return Error::success(); 1308 } 1309 1310 Error wait() override { return Error::success(); } 1311 }; 1312 } // end anonymous namespace 1313 1314 ThinBackend lto::createWriteIndexesThinBackend( 1315 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles, 1316 raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) { 1317 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1318 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1319 AddStreamFn AddStream, NativeObjectCache Cache) { 1320 return std::make_unique<WriteIndexesThinBackend>( 1321 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix, 1322 ShouldEmitImportsFiles, LinkedObjectsFile, OnWrite); 1323 }; 1324 } 1325 1326 Error LTO::runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache, 1327 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 1328 if (ThinLTO.ModuleMap.empty()) 1329 return Error::success(); 1330 1331 if (ThinLTO.ModulesToCompile && ThinLTO.ModulesToCompile->empty()) { 1332 llvm::errs() << "warning: [ThinLTO] No module compiled\n"; 1333 return Error::success(); 1334 } 1335 1336 if (Conf.CombinedIndexHook && 1337 !Conf.CombinedIndexHook(ThinLTO.CombinedIndex, GUIDPreservedSymbols)) 1338 return Error::success(); 1339 1340 // Collect for each module the list of function it defines (GUID -> 1341 // Summary). 1342 StringMap<GVSummaryMapTy> 1343 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size()); 1344 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule( 1345 ModuleToDefinedGVSummaries); 1346 // Create entries for any modules that didn't have any GV summaries 1347 // (either they didn't have any GVs to start with, or we suppressed 1348 // generation of the summaries because they e.g. had inline assembly 1349 // uses that couldn't be promoted/renamed on export). This is so 1350 // InProcessThinBackend::start can still launch a backend thread, which 1351 // is passed the map of summaries for the module, without any special 1352 // handling for this case. 1353 for (auto &Mod : ThinLTO.ModuleMap) 1354 if (!ModuleToDefinedGVSummaries.count(Mod.first)) 1355 ModuleToDefinedGVSummaries.try_emplace(Mod.first); 1356 1357 // Synthesize entry counts for functions in the CombinedIndex. 1358 computeSyntheticCounts(ThinLTO.CombinedIndex); 1359 1360 StringMap<FunctionImporter::ImportMapTy> ImportLists( 1361 ThinLTO.ModuleMap.size()); 1362 StringMap<FunctionImporter::ExportSetTy> ExportLists( 1363 ThinLTO.ModuleMap.size()); 1364 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; 1365 1366 if (DumpThinCGSCCs) 1367 ThinLTO.CombinedIndex.dumpSCCs(outs()); 1368 1369 std::set<GlobalValue::GUID> ExportedGUIDs; 1370 1371 // If allowed, upgrade public vcall visibility to linkage unit visibility in 1372 // the summaries before whole program devirtualization below. 1373 updateVCallVisibilityInIndex(ThinLTO.CombinedIndex, 1374 Conf.HasWholeProgramVisibility); 1375 1376 // Perform index-based WPD. This will return immediately if there are 1377 // no index entries in the typeIdMetadata map (e.g. if we are instead 1378 // performing IR-based WPD in hybrid regular/thin LTO mode). 1379 std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap; 1380 runWholeProgramDevirtOnIndex(ThinLTO.CombinedIndex, ExportedGUIDs, 1381 LocalWPDTargetsMap); 1382 1383 if (Conf.OptLevel > 0) 1384 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 1385 ImportLists, ExportLists); 1386 1387 // Figure out which symbols need to be internalized. This also needs to happen 1388 // at -O0 because summary-based DCE is implemented using internalization, and 1389 // we must apply DCE consistently with the full LTO module in order to avoid 1390 // undefined references during the final link. 1391 for (auto &Res : GlobalResolutions) { 1392 // If the symbol does not have external references or it is not prevailing, 1393 // then not need to mark it as exported from a ThinLTO partition. 1394 if (Res.second.Partition != GlobalResolution::External || 1395 !Res.second.isPrevailingIRSymbol()) 1396 continue; 1397 auto GUID = GlobalValue::getGUID( 1398 GlobalValue::dropLLVMManglingEscape(Res.second.IRName)); 1399 // Mark exported unless index-based analysis determined it to be dead. 1400 if (ThinLTO.CombinedIndex.isGUIDLive(GUID)) 1401 ExportedGUIDs.insert(GUID); 1402 } 1403 1404 // Any functions referenced by the jump table in the regular LTO object must 1405 // be exported. 1406 for (auto &Def : ThinLTO.CombinedIndex.cfiFunctionDefs()) 1407 ExportedGUIDs.insert( 1408 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Def))); 1409 1410 auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) { 1411 const auto &ExportList = ExportLists.find(ModuleIdentifier); 1412 return (ExportList != ExportLists.end() && ExportList->second.count(VI)) || 1413 ExportedGUIDs.count(VI.getGUID()); 1414 }; 1415 1416 // Update local devirtualized targets that were exported by cross-module 1417 // importing or by other devirtualizations marked in the ExportedGUIDs set. 1418 updateIndexWPDForExports(ThinLTO.CombinedIndex, isExported, 1419 LocalWPDTargetsMap); 1420 1421 auto isPrevailing = [&](GlobalValue::GUID GUID, 1422 const GlobalValueSummary *S) { 1423 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath(); 1424 }; 1425 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported, 1426 isPrevailing); 1427 1428 auto recordNewLinkage = [&](StringRef ModuleIdentifier, 1429 GlobalValue::GUID GUID, 1430 GlobalValue::LinkageTypes NewLinkage) { 1431 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage; 1432 }; 1433 thinLTOResolvePrevailingInIndex(ThinLTO.CombinedIndex, isPrevailing, 1434 recordNewLinkage, GUIDPreservedSymbols); 1435 1436 std::unique_ptr<ThinBackendProc> BackendProc = 1437 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 1438 AddStream, Cache); 1439 1440 auto &ModuleMap = 1441 ThinLTO.ModulesToCompile ? *ThinLTO.ModulesToCompile : ThinLTO.ModuleMap; 1442 1443 // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for combined 1444 // module and parallel code generation partitions. 1445 unsigned Task = RegularLTO.ParallelCodeGenParallelismLevel; 1446 for (auto &Mod : ModuleMap) { 1447 if (Error E = BackendProc->start(Task, Mod.second, ImportLists[Mod.first], 1448 ExportLists[Mod.first], 1449 ResolvedODR[Mod.first], ThinLTO.ModuleMap)) 1450 return E; 1451 ++Task; 1452 } 1453 1454 return BackendProc->wait(); 1455 } 1456 1457 Expected<std::unique_ptr<ToolOutputFile>> lto::setupLLVMOptimizationRemarks( 1458 LLVMContext &Context, StringRef RemarksFilename, StringRef RemarksPasses, 1459 StringRef RemarksFormat, bool RemarksWithHotness, int Count) { 1460 std::string Filename = std::string(RemarksFilename); 1461 // For ThinLTO, file.opt.<format> becomes 1462 // file.opt.<format>.thin.<num>.<format>. 1463 if (!Filename.empty() && Count != -1) 1464 Filename = 1465 (Twine(Filename) + ".thin." + llvm::utostr(Count) + "." + RemarksFormat) 1466 .str(); 1467 1468 auto ResultOrErr = llvm::setupLLVMOptimizationRemarks( 1469 Context, Filename, RemarksPasses, RemarksFormat, RemarksWithHotness); 1470 if (Error E = ResultOrErr.takeError()) 1471 return std::move(E); 1472 1473 if (*ResultOrErr) 1474 (*ResultOrErr)->keep(); 1475 1476 return ResultOrErr; 1477 } 1478 1479 Expected<std::unique_ptr<ToolOutputFile>> 1480 lto::setupStatsFile(StringRef StatsFilename) { 1481 // Setup output file to emit statistics. 1482 if (StatsFilename.empty()) 1483 return nullptr; 1484 1485 llvm::EnableStatistics(false); 1486 std::error_code EC; 1487 auto StatsFile = 1488 std::make_unique<ToolOutputFile>(StatsFilename, EC, sys::fs::OF_None); 1489 if (EC) 1490 return errorCodeToError(EC); 1491 1492 StatsFile->keep(); 1493 return std::move(StatsFile); 1494 } 1495