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