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