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