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