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