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 support::endian::write32le(Data, I); 102 Hasher.update(ArrayRef<uint8_t>{Data, 4}); 103 }; 104 auto AddUint64 = [&](uint64_t I) { 105 uint8_t Data[8]; 106 support::endian::write64le(Data, I); 107 Hasher.update(ArrayRef<uint8_t>{Data, 8}); 108 }; 109 AddString(Conf.CPU); 110 // FIXME: Hash more of Options. For now all clients initialize Options from 111 // command-line flags (which is unsupported in production), but may set 112 // RelaxELFRelocations. The clang driver can also pass FunctionSections, 113 // DataSections and DebuggerTuning via command line flags. 114 AddUnsigned(Conf.Options.RelaxELFRelocations); 115 AddUnsigned(Conf.Options.FunctionSections); 116 AddUnsigned(Conf.Options.DataSections); 117 AddUnsigned((unsigned)Conf.Options.DebuggerTuning); 118 for (auto &A : Conf.MAttrs) 119 AddString(A); 120 if (Conf.RelocModel) 121 AddUnsigned(*Conf.RelocModel); 122 else 123 AddUnsigned(-1); 124 if (Conf.CodeModel) 125 AddUnsigned(*Conf.CodeModel); 126 else 127 AddUnsigned(-1); 128 AddUnsigned(Conf.CGOptLevel); 129 AddUnsigned(Conf.CGFileType); 130 AddUnsigned(Conf.OptLevel); 131 AddUnsigned(Conf.UseNewPM); 132 AddUnsigned(Conf.Freestanding); 133 AddString(Conf.OptPipeline); 134 AddString(Conf.AAPipeline); 135 AddString(Conf.OverrideTriple); 136 AddString(Conf.DefaultTriple); 137 AddString(Conf.DwoDir); 138 139 // Include the hash for the current module 140 auto ModHash = Index.getModuleHash(ModuleID); 141 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 142 for (const auto &VI : ExportList) { 143 auto GUID = VI.getGUID(); 144 // The export list can impact the internalization, be conservative here 145 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&GUID, sizeof(GUID))); 146 } 147 148 // Include the hash for every module we import functions from. The set of 149 // imported symbols for each module may affect code generation and is 150 // sensitive to link order, so include that as well. 151 for (auto &Entry : ImportList) { 152 auto ModHash = Index.getModuleHash(Entry.first()); 153 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); 154 155 AddUint64(Entry.second.size()); 156 for (auto &Fn : Entry.second) 157 AddUint64(Fn); 158 } 159 160 // Include the hash for the resolved ODR. 161 for (auto &Entry : ResolvedODR) { 162 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first, 163 sizeof(GlobalValue::GUID))); 164 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second, 165 sizeof(GlobalValue::LinkageTypes))); 166 } 167 168 // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or 169 // defined in this module. 170 std::set<GlobalValue::GUID> UsedCfiDefs; 171 std::set<GlobalValue::GUID> UsedCfiDecls; 172 173 // Typeids used in this module. 174 std::set<GlobalValue::GUID> UsedTypeIds; 175 176 auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) { 177 if (CfiFunctionDefs.count(ValueGUID)) 178 UsedCfiDefs.insert(ValueGUID); 179 if (CfiFunctionDecls.count(ValueGUID)) 180 UsedCfiDecls.insert(ValueGUID); 181 }; 182 183 auto AddUsedThings = [&](GlobalValueSummary *GS) { 184 if (!GS) return; 185 AddUnsigned(GS->isLive()); 186 AddUnsigned(GS->canAutoHide()); 187 for (const ValueInfo &VI : GS->refs()) { 188 AddUnsigned(VI.isDSOLocal()); 189 AddUsedCfiGlobal(VI.getGUID()); 190 } 191 if (auto *GVS = dyn_cast<GlobalVarSummary>(GS)) { 192 AddUnsigned(GVS->maybeReadOnly()); 193 AddUnsigned(GVS->maybeWriteOnly()); 194 } 195 if (auto *FS = dyn_cast<FunctionSummary>(GS)) { 196 for (auto &TT : FS->type_tests()) 197 UsedTypeIds.insert(TT); 198 for (auto &TT : FS->type_test_assume_vcalls()) 199 UsedTypeIds.insert(TT.GUID); 200 for (auto &TT : FS->type_checked_load_vcalls()) 201 UsedTypeIds.insert(TT.GUID); 202 for (auto &TT : FS->type_test_assume_const_vcalls()) 203 UsedTypeIds.insert(TT.VFunc.GUID); 204 for (auto &TT : FS->type_checked_load_const_vcalls()) 205 UsedTypeIds.insert(TT.VFunc.GUID); 206 for (auto &ET : FS->calls()) { 207 AddUnsigned(ET.first.isDSOLocal()); 208 AddUsedCfiGlobal(ET.first.getGUID()); 209 } 210 } 211 }; 212 213 // Include the hash for the linkage type to reflect internalization and weak 214 // resolution, and collect any used type identifier resolutions. 215 for (auto &GS : DefinedGlobals) { 216 GlobalValue::LinkageTypes Linkage = GS.second->linkage(); 217 Hasher.update( 218 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage))); 219 AddUsedCfiGlobal(GS.first); 220 AddUsedThings(GS.second); 221 } 222 223 // Imported functions may introduce new uses of type identifier resolutions, 224 // so we need to collect their used resolutions as well. 225 for (auto &ImpM : ImportList) 226 for (auto &ImpF : ImpM.second) { 227 GlobalValueSummary *S = Index.findSummaryInModule(ImpF, ImpM.first()); 228 AddUsedThings(S); 229 // If this is an alias, we also care about any types/etc. that the aliasee 230 // may reference. 231 if (auto *AS = dyn_cast_or_null<AliasSummary>(S)) 232 AddUsedThings(AS->getBaseObject()); 233 } 234 235 auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) { 236 AddString(TId); 237 238 AddUnsigned(S.TTRes.TheKind); 239 AddUnsigned(S.TTRes.SizeM1BitWidth); 240 241 AddUint64(S.TTRes.AlignLog2); 242 AddUint64(S.TTRes.SizeM1); 243 AddUint64(S.TTRes.BitMask); 244 AddUint64(S.TTRes.InlineBits); 245 246 AddUint64(S.WPDRes.size()); 247 for (auto &WPD : S.WPDRes) { 248 AddUnsigned(WPD.first); 249 AddUnsigned(WPD.second.TheKind); 250 AddString(WPD.second.SingleImplName); 251 252 AddUint64(WPD.second.ResByArg.size()); 253 for (auto &ByArg : WPD.second.ResByArg) { 254 AddUint64(ByArg.first.size()); 255 for (uint64_t Arg : ByArg.first) 256 AddUint64(Arg); 257 AddUnsigned(ByArg.second.TheKind); 258 AddUint64(ByArg.second.Info); 259 AddUnsigned(ByArg.second.Byte); 260 AddUnsigned(ByArg.second.Bit); 261 } 262 } 263 }; 264 265 // Include the hash for all type identifiers used by this module. 266 for (GlobalValue::GUID TId : UsedTypeIds) { 267 auto TidIter = Index.typeIds().equal_range(TId); 268 for (auto It = TidIter.first; It != TidIter.second; ++It) 269 AddTypeIdSummary(It->second.first, It->second.second); 270 } 271 272 AddUnsigned(UsedCfiDefs.size()); 273 for (auto &V : UsedCfiDefs) 274 AddUint64(V); 275 276 AddUnsigned(UsedCfiDecls.size()); 277 for (auto &V : UsedCfiDecls) 278 AddUint64(V); 279 280 if (!Conf.SampleProfile.empty()) { 281 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile); 282 if (FileOrErr) { 283 Hasher.update(FileOrErr.get()->getBuffer()); 284 285 if (!Conf.ProfileRemapping.empty()) { 286 FileOrErr = MemoryBuffer::getFile(Conf.ProfileRemapping); 287 if (FileOrErr) 288 Hasher.update(FileOrErr.get()->getBuffer()); 289 } 290 } 291 } 292 293 Key = toHex(Hasher.result()); 294 } 295 296 static void thinLTOResolvePrevailingGUID( 297 ValueInfo VI, DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias, 298 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 299 isPrevailing, 300 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 301 recordNewLinkage, 302 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 303 for (auto &S : VI.getSummaryList()) { 304 GlobalValue::LinkageTypes OriginalLinkage = S->linkage(); 305 // Ignore local and appending linkage values since the linker 306 // doesn't resolve them. 307 if (GlobalValue::isLocalLinkage(OriginalLinkage) || 308 GlobalValue::isAppendingLinkage(S->linkage())) 309 continue; 310 // We need to emit only one of these. The prevailing module will keep it, 311 // but turned into a weak, while the others will drop it when possible. 312 // This is both a compile-time optimization and a correctness 313 // transformation. This is necessary for correctness when we have exported 314 // a reference - we need to convert the linkonce to weak to 315 // ensure a copy is kept to satisfy the exported reference. 316 // FIXME: We may want to split the compile time and correctness 317 // aspects into separate routines. 318 if (isPrevailing(VI.getGUID(), S.get())) { 319 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) { 320 S->setLinkage(GlobalValue::getWeakLinkage( 321 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage))); 322 // The kept copy is eligible for auto-hiding (hidden visibility) if all 323 // copies were (i.e. they were all linkonce_odr global unnamed addr). 324 // If any copy is not (e.g. it was originally weak_odr), then the symbol 325 // must remain externally available (e.g. a weak_odr from an explicitly 326 // instantiated template). Additionally, if it is in the 327 // GUIDPreservedSymbols set, that means that it is visibile outside 328 // the summary (e.g. in a native object or a bitcode file without 329 // summary), and in that case we cannot hide it as it isn't possible to 330 // check all copies. 331 S->setCanAutoHide(VI.canAutoHide() && 332 !GUIDPreservedSymbols.count(VI.getGUID())); 333 } 334 } 335 // Alias and aliasee can't be turned into available_externally. 336 else if (!isa<AliasSummary>(S.get()) && 337 !GlobalInvolvedWithAlias.count(S.get())) 338 S->setLinkage(GlobalValue::AvailableExternallyLinkage); 339 if (S->linkage() != OriginalLinkage) 340 recordNewLinkage(S->modulePath(), VI.getGUID(), S->linkage()); 341 } 342 } 343 344 /// Resolve linkage for prevailing symbols in the \p Index. 345 // 346 // We'd like to drop these functions if they are no longer referenced in the 347 // current module. However there is a chance that another module is still 348 // referencing them because of the import. We make sure we always emit at least 349 // one copy. 350 void llvm::thinLTOResolvePrevailingInIndex( 351 ModuleSummaryIndex &Index, 352 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 353 isPrevailing, 354 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> 355 recordNewLinkage, 356 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 357 // We won't optimize the globals that are referenced by an alias for now 358 // Ideally we should turn the alias into a global and duplicate the definition 359 // when needed. 360 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias; 361 for (auto &I : Index) 362 for (auto &S : I.second.SummaryList) 363 if (auto AS = dyn_cast<AliasSummary>(S.get())) 364 GlobalInvolvedWithAlias.insert(&AS->getAliasee()); 365 366 for (auto &I : Index) 367 thinLTOResolvePrevailingGUID(Index.getValueInfo(I), GlobalInvolvedWithAlias, 368 isPrevailing, recordNewLinkage, 369 GUIDPreservedSymbols); 370 } 371 372 static bool isWeakObjectWithRWAccess(GlobalValueSummary *GVS) { 373 if (auto *VarSummary = dyn_cast<GlobalVarSummary>(GVS->getBaseObject())) 374 return !VarSummary->maybeReadOnly() && !VarSummary->maybeWriteOnly() && 375 (VarSummary->linkage() == GlobalValue::WeakODRLinkage || 376 VarSummary->linkage() == GlobalValue::LinkOnceODRLinkage); 377 return false; 378 } 379 380 static void thinLTOInternalizeAndPromoteGUID( 381 ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported, 382 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 383 isPrevailing) { 384 for (auto &S : VI.getSummaryList()) { 385 if (isExported(S->modulePath(), VI)) { 386 if (GlobalValue::isLocalLinkage(S->linkage())) 387 S->setLinkage(GlobalValue::ExternalLinkage); 388 } else if (EnableLTOInternalization && 389 // Ignore local and appending linkage values since the linker 390 // doesn't resolve them. 391 !GlobalValue::isLocalLinkage(S->linkage()) && 392 (!GlobalValue::isInterposableLinkage(S->linkage()) || 393 isPrevailing(VI.getGUID(), S.get())) && 394 S->linkage() != GlobalValue::AppendingLinkage && 395 // We can't internalize available_externally globals because this 396 // can break function pointer equality. 397 S->linkage() != GlobalValue::AvailableExternallyLinkage && 398 // Functions and read-only variables with linkonce_odr and 399 // weak_odr linkage can be internalized. We can't internalize 400 // linkonce_odr and weak_odr variables which are both modified 401 // and read somewhere in the program because reads and writes 402 // will become inconsistent. 403 !isWeakObjectWithRWAccess(S.get())) 404 S->setLinkage(GlobalValue::InternalLinkage); 405 } 406 } 407 408 // Update the linkages in the given \p Index to mark exported values 409 // as external and non-exported values as internal. 410 void llvm::thinLTOInternalizeAndPromoteInIndex( 411 ModuleSummaryIndex &Index, 412 function_ref<bool(StringRef, ValueInfo)> isExported, 413 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> 414 isPrevailing) { 415 for (auto &I : Index) 416 thinLTOInternalizeAndPromoteGUID(Index.getValueInfo(I), isExported, 417 isPrevailing); 418 } 419 420 // Requires a destructor for std::vector<InputModule>. 421 InputFile::~InputFile() = default; 422 423 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) { 424 std::unique_ptr<InputFile> File(new InputFile); 425 426 Expected<IRSymtabFile> FOrErr = readIRSymtab(Object); 427 if (!FOrErr) 428 return FOrErr.takeError(); 429 430 File->TargetTriple = FOrErr->TheReader.getTargetTriple(); 431 File->SourceFileName = FOrErr->TheReader.getSourceFileName(); 432 File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts(); 433 File->DependentLibraries = FOrErr->TheReader.getDependentLibraries(); 434 File->ComdatTable = FOrErr->TheReader.getComdatTable(); 435 436 for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) { 437 size_t Begin = File->Symbols.size(); 438 for (const irsymtab::Reader::SymbolRef &Sym : 439 FOrErr->TheReader.module_symbols(I)) 440 // Skip symbols that are irrelevant to LTO. Note that this condition needs 441 // to match the one in Skip() in LTO::addRegularLTO(). 442 if (Sym.isGlobal() && !Sym.isFormatSpecific()) 443 File->Symbols.push_back(Sym); 444 File->ModuleSymIndices.push_back({Begin, File->Symbols.size()}); 445 } 446 447 File->Mods = FOrErr->Mods; 448 File->Strtab = std::move(FOrErr->Strtab); 449 return std::move(File); 450 } 451 452 StringRef InputFile::getName() const { 453 return Mods[0].getModuleIdentifier(); 454 } 455 456 BitcodeModule &InputFile::getSingleBitcodeModule() { 457 assert(Mods.size() == 1 && "Expect only one bitcode module"); 458 return Mods[0]; 459 } 460 461 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel, 462 const Config &Conf) 463 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel), 464 Ctx(Conf), CombinedModule(std::make_unique<Module>("ld-temp.o", Ctx)), 465 Mover(std::make_unique<IRMover>(*CombinedModule)) {} 466 467 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend) 468 : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) { 469 if (!Backend) 470 this->Backend = 471 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency()); 472 } 473 474 LTO::LTO(Config Conf, ThinBackend Backend, 475 unsigned ParallelCodeGenParallelismLevel) 476 : Conf(std::move(Conf)), 477 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf), 478 ThinLTO(std::move(Backend)) {} 479 480 // Requires a destructor for MapVector<BitcodeModule>. 481 LTO::~LTO() = default; 482 483 // Add the symbols in the given module to the GlobalResolutions map, and resolve 484 // their partitions. 485 void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms, 486 ArrayRef<SymbolResolution> Res, 487 unsigned Partition, bool InSummary) { 488 auto *ResI = Res.begin(); 489 auto *ResE = Res.end(); 490 (void)ResE; 491 for (const InputFile::Symbol &Sym : Syms) { 492 assert(ResI != ResE); 493 SymbolResolution Res = *ResI++; 494 495 StringRef Name = Sym.getName(); 496 Triple TT(RegularLTO.CombinedModule->getTargetTriple()); 497 // Strip the __imp_ prefix from COFF dllimport symbols (similar to the 498 // way they are handled by lld), otherwise we can end up with two 499 // global resolutions (one with and one for a copy of the symbol without). 500 if (TT.isOSBinFormatCOFF() && Name.startswith("__imp_")) 501 Name = Name.substr(strlen("__imp_")); 502 auto &GlobalRes = GlobalResolutions[Name]; 503 GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr(); 504 if (Res.Prevailing) { 505 assert(!GlobalRes.Prevailing && 506 "Multiple prevailing defs are not allowed"); 507 GlobalRes.Prevailing = true; 508 GlobalRes.IRName = std::string(Sym.getIRName()); 509 } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) { 510 // Sometimes it can be two copies of symbol in a module and prevailing 511 // symbol can have no IR name. That might happen if symbol is defined in 512 // module level inline asm block. In case we have multiple modules with 513 // the same symbol we want to use IR name of the prevailing symbol. 514 // Otherwise, if we haven't seen a prevailing symbol, set the name so that 515 // we can later use it to check if there is any prevailing copy in IR. 516 GlobalRes.IRName = std::string(Sym.getIRName()); 517 } 518 519 // Set the partition to external if we know it is re-defined by the linker 520 // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a 521 // regular object, is referenced from llvm.compiler_used, or was already 522 // recorded as being referenced from a different partition. 523 if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() || 524 (GlobalRes.Partition != GlobalResolution::Unknown && 525 GlobalRes.Partition != Partition)) { 526 GlobalRes.Partition = GlobalResolution::External; 527 } else 528 // First recorded reference, save the current partition. 529 GlobalRes.Partition = Partition; 530 531 // Flag as visible outside of summary if visible from a regular object or 532 // from a module that does not have a summary. 533 GlobalRes.VisibleOutsideSummary |= 534 (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary); 535 } 536 } 537 538 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input, 539 ArrayRef<SymbolResolution> Res) { 540 StringRef Path = Input->getName(); 541 OS << Path << '\n'; 542 auto ResI = Res.begin(); 543 for (const InputFile::Symbol &Sym : Input->symbols()) { 544 assert(ResI != Res.end()); 545 SymbolResolution Res = *ResI++; 546 547 OS << "-r=" << Path << ',' << Sym.getName() << ','; 548 if (Res.Prevailing) 549 OS << 'p'; 550 if (Res.FinalDefinitionInLinkageUnit) 551 OS << 'l'; 552 if (Res.VisibleToRegularObj) 553 OS << 'x'; 554 if (Res.LinkerRedefined) 555 OS << 'r'; 556 OS << '\n'; 557 } 558 OS.flush(); 559 assert(ResI == Res.end()); 560 } 561 562 Error LTO::add(std::unique_ptr<InputFile> Input, 563 ArrayRef<SymbolResolution> Res) { 564 assert(!CalledGetMaxTasks); 565 566 if (Conf.ResolutionFile) 567 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res); 568 569 if (RegularLTO.CombinedModule->getTargetTriple().empty()) 570 RegularLTO.CombinedModule->setTargetTriple(Input->getTargetTriple()); 571 572 const SymbolResolution *ResI = Res.begin(); 573 for (unsigned I = 0; I != Input->Mods.size(); ++I) 574 if (Error Err = addModule(*Input, I, ResI, Res.end())) 575 return Err; 576 577 assert(ResI == Res.end()); 578 return Error::success(); 579 } 580 581 Error LTO::addModule(InputFile &Input, unsigned ModI, 582 const SymbolResolution *&ResI, 583 const SymbolResolution *ResE) { 584 Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo(); 585 if (!LTOInfo) 586 return LTOInfo.takeError(); 587 588 if (EnableSplitLTOUnit.hasValue()) { 589 // If only some modules were split, flag this in the index so that 590 // we can skip or error on optimizations that need consistently split 591 // modules (whole program devirt and lower type tests). 592 if (EnableSplitLTOUnit.getValue() != LTOInfo->EnableSplitLTOUnit) 593 ThinLTO.CombinedIndex.setPartiallySplitLTOUnits(); 594 } else 595 EnableSplitLTOUnit = LTOInfo->EnableSplitLTOUnit; 596 597 BitcodeModule BM = Input.Mods[ModI]; 598 auto ModSyms = Input.module_symbols(ModI); 599 addModuleToGlobalRes(ModSyms, {ResI, ResE}, 600 LTOInfo->IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0, 601 LTOInfo->HasSummary); 602 603 if (LTOInfo->IsThinLTO) 604 return addThinLTO(BM, ModSyms, ResI, ResE); 605 606 RegularLTO.EmptyCombinedModule = false; 607 Expected<RegularLTOState::AddedModule> ModOrErr = 608 addRegularLTO(BM, ModSyms, ResI, ResE); 609 if (!ModOrErr) 610 return ModOrErr.takeError(); 611 612 if (!LTOInfo->HasSummary) 613 return linkRegularLTO(std::move(*ModOrErr), /*LivenessFromIndex=*/false); 614 615 // Regular LTO module summaries are added to a dummy module that represents 616 // the combined regular LTO module. 617 if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, "", -1ull)) 618 return Err; 619 RegularLTO.ModsWithSummaries.push_back(std::move(*ModOrErr)); 620 return Error::success(); 621 } 622 623 // Checks whether the given global value is in a non-prevailing comdat 624 // (comdat containing values the linker indicated were not prevailing, 625 // which we then dropped to available_externally), and if so, removes 626 // it from the comdat. This is called for all global values to ensure the 627 // comdat is empty rather than leaving an incomplete comdat. It is needed for 628 // regular LTO modules, in case we are in a mixed-LTO mode (both regular 629 // and thin LTO modules) compilation. Since the regular LTO module will be 630 // linked first in the final native link, we want to make sure the linker 631 // doesn't select any of these incomplete comdats that would be left 632 // in the regular LTO module without this cleanup. 633 static void 634 handleNonPrevailingComdat(GlobalValue &GV, 635 std::set<const Comdat *> &NonPrevailingComdats) { 636 Comdat *C = GV.getComdat(); 637 if (!C) 638 return; 639 640 if (!NonPrevailingComdats.count(C)) 641 return; 642 643 // Additionally need to drop externally visible global values from the comdat 644 // to available_externally, so that there aren't multiply defined linker 645 // errors. 646 if (!GV.hasLocalLinkage()) 647 GV.setLinkage(GlobalValue::AvailableExternallyLinkage); 648 649 if (auto GO = dyn_cast<GlobalObject>(&GV)) 650 GO->setComdat(nullptr); 651 } 652 653 // Add a regular LTO object to the link. 654 // The resulting module needs to be linked into the combined LTO module with 655 // linkRegularLTO. 656 Expected<LTO::RegularLTOState::AddedModule> 657 LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, 658 const SymbolResolution *&ResI, 659 const SymbolResolution *ResE) { 660 RegularLTOState::AddedModule Mod; 661 Expected<std::unique_ptr<Module>> MOrErr = 662 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true, 663 /*IsImporting*/ false); 664 if (!MOrErr) 665 return MOrErr.takeError(); 666 Module &M = **MOrErr; 667 Mod.M = std::move(*MOrErr); 668 669 if (Error Err = M.materializeMetadata()) 670 return std::move(Err); 671 UpgradeDebugInfo(M); 672 673 ModuleSymbolTable SymTab; 674 SymTab.addModule(&M); 675 676 for (GlobalVariable &GV : M.globals()) 677 if (GV.hasAppendingLinkage()) 678 Mod.Keep.push_back(&GV); 679 680 DenseSet<GlobalObject *> AliasedGlobals; 681 for (auto &GA : M.aliases()) 682 if (GlobalObject *GO = GA.getBaseObject()) 683 AliasedGlobals.insert(GO); 684 685 // In this function we need IR GlobalValues matching the symbols in Syms 686 // (which is not backed by a module), so we need to enumerate them in the same 687 // order. The symbol enumeration order of a ModuleSymbolTable intentionally 688 // matches the order of an irsymtab, but when we read the irsymtab in 689 // InputFile::create we omit some symbols that are irrelevant to LTO. The 690 // Skip() function skips the same symbols from the module as InputFile does 691 // from the symbol table. 692 auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end(); 693 auto Skip = [&]() { 694 while (MsymI != MsymE) { 695 auto Flags = SymTab.getSymbolFlags(*MsymI); 696 if ((Flags & object::BasicSymbolRef::SF_Global) && 697 !(Flags & object::BasicSymbolRef::SF_FormatSpecific)) 698 return; 699 ++MsymI; 700 } 701 }; 702 Skip(); 703 704 std::set<const Comdat *> NonPrevailingComdats; 705 for (const InputFile::Symbol &Sym : Syms) { 706 assert(ResI != ResE); 707 SymbolResolution Res = *ResI++; 708 709 assert(MsymI != MsymE); 710 ModuleSymbolTable::Symbol Msym = *MsymI++; 711 Skip(); 712 713 if (GlobalValue *GV = Msym.dyn_cast<GlobalValue *>()) { 714 if (Res.Prevailing) { 715 if (Sym.isUndefined()) 716 continue; 717 Mod.Keep.push_back(GV); 718 // For symbols re-defined with linker -wrap and -defsym options, 719 // set the linkage to weak to inhibit IPO. The linkage will be 720 // restored by the linker. 721 if (Res.LinkerRedefined) 722 GV->setLinkage(GlobalValue::WeakAnyLinkage); 723 724 GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage(); 725 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) 726 GV->setLinkage(GlobalValue::getWeakLinkage( 727 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage))); 728 } else if (isa<GlobalObject>(GV) && 729 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() || 730 GV->hasAvailableExternallyLinkage()) && 731 !AliasedGlobals.count(cast<GlobalObject>(GV))) { 732 // Any of the above three types of linkage indicates that the 733 // chosen prevailing symbol will have the same semantics as this copy of 734 // the symbol, so we may be able to link it with available_externally 735 // linkage. We will decide later whether to do that when we link this 736 // module (in linkRegularLTO), based on whether it is undefined. 737 Mod.Keep.push_back(GV); 738 GV->setLinkage(GlobalValue::AvailableExternallyLinkage); 739 if (GV->hasComdat()) 740 NonPrevailingComdats.insert(GV->getComdat()); 741 cast<GlobalObject>(GV)->setComdat(nullptr); 742 } 743 744 // Set the 'local' flag based on the linker resolution for this symbol. 745 if (Res.FinalDefinitionInLinkageUnit) { 746 GV->setDSOLocal(true); 747 if (GV->hasDLLImportStorageClass()) 748 GV->setDLLStorageClass(GlobalValue::DLLStorageClassTypes:: 749 DefaultStorageClass); 750 } 751 } 752 // Common resolution: collect the maximum size/alignment over all commons. 753 // We also record if we see an instance of a common as prevailing, so that 754 // if none is prevailing we can ignore it later. 755 if (Sym.isCommon()) { 756 // FIXME: We should figure out what to do about commons defined by asm. 757 // For now they aren't reported correctly by ModuleSymbolTable. 758 auto &CommonRes = RegularLTO.Commons[std::string(Sym.getIRName())]; 759 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize()); 760 CommonRes.Align = 761 std::max(CommonRes.Align, MaybeAlign(Sym.getCommonAlignment())); 762 CommonRes.Prevailing |= Res.Prevailing; 763 } 764 765 } 766 if (!M.getComdatSymbolTable().empty()) 767 for (GlobalValue &GV : M.global_values()) 768 handleNonPrevailingComdat(GV, NonPrevailingComdats); 769 assert(MsymI == MsymE); 770 return std::move(Mod); 771 } 772 773 Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod, 774 bool LivenessFromIndex) { 775 std::vector<GlobalValue *> Keep; 776 for (GlobalValue *GV : Mod.Keep) { 777 if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GV->getGUID())) { 778 if (Function *F = dyn_cast<Function>(GV)) { 779 OptimizationRemarkEmitter ORE(F); 780 ORE.emit(OptimizationRemark(DEBUG_TYPE, "deadfunction", F) 781 << ore::NV("Function", F) 782 << " not added to the combined module "); 783 } 784 continue; 785 } 786 787 if (!GV->hasAvailableExternallyLinkage()) { 788 Keep.push_back(GV); 789 continue; 790 } 791 792 // Only link available_externally definitions if we don't already have a 793 // definition. 794 GlobalValue *CombinedGV = 795 RegularLTO.CombinedModule->getNamedValue(GV->getName()); 796 if (CombinedGV && !CombinedGV->isDeclaration()) 797 continue; 798 799 Keep.push_back(GV); 800 } 801 802 return RegularLTO.Mover->move(std::move(Mod.M), Keep, 803 [](GlobalValue &, IRMover::ValueAdder) {}, 804 /* IsPerformingImport */ false); 805 } 806 807 // Add a ThinLTO module to the link. 808 Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, 809 const SymbolResolution *&ResI, 810 const SymbolResolution *ResE) { 811 if (Error Err = 812 BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(), 813 ThinLTO.ModuleMap.size())) 814 return Err; 815 816 for (const InputFile::Symbol &Sym : Syms) { 817 assert(ResI != ResE); 818 SymbolResolution Res = *ResI++; 819 820 if (!Sym.getIRName().empty()) { 821 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier( 822 Sym.getIRName(), GlobalValue::ExternalLinkage, "")); 823 if (Res.Prevailing) { 824 ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier(); 825 826 // For linker redefined symbols (via --wrap or --defsym) we want to 827 // switch the linkage to `weak` to prevent IPOs from happening. 828 // Find the summary in the module for this very GV and record the new 829 // linkage so that we can switch it when we import the GV. 830 if (Res.LinkerRedefined) 831 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( 832 GUID, BM.getModuleIdentifier())) 833 S->setLinkage(GlobalValue::WeakAnyLinkage); 834 } 835 836 // If the linker resolved the symbol to a local definition then mark it 837 // as local in the summary for the module we are adding. 838 if (Res.FinalDefinitionInLinkageUnit) { 839 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( 840 GUID, BM.getModuleIdentifier())) { 841 S->setDSOLocal(true); 842 } 843 } 844 } 845 } 846 847 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second) 848 return make_error<StringError>( 849 "Expected at most one ThinLTO module per bitcode file", 850 inconvertibleErrorCode()); 851 852 return Error::success(); 853 } 854 855 unsigned LTO::getMaxTasks() const { 856 CalledGetMaxTasks = true; 857 return RegularLTO.ParallelCodeGenParallelismLevel + ThinLTO.ModuleMap.size(); 858 } 859 860 // If only some of the modules were split, we cannot correctly handle 861 // code that contains type tests or type checked loads. 862 Error LTO::checkPartiallySplit() { 863 if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits()) 864 return Error::success(); 865 866 Function *TypeTestFunc = RegularLTO.CombinedModule->getFunction( 867 Intrinsic::getName(Intrinsic::type_test)); 868 Function *TypeCheckedLoadFunc = RegularLTO.CombinedModule->getFunction( 869 Intrinsic::getName(Intrinsic::type_checked_load)); 870 871 // First check if there are type tests / type checked loads in the 872 // merged regular LTO module IR. 873 if ((TypeTestFunc && !TypeTestFunc->use_empty()) || 874 (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty())) 875 return make_error<StringError>( 876 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)", 877 inconvertibleErrorCode()); 878 879 // Otherwise check if there are any recorded in the combined summary from the 880 // ThinLTO modules. 881 for (auto &P : ThinLTO.CombinedIndex) { 882 for (auto &S : P.second.SummaryList) { 883 auto *FS = dyn_cast<FunctionSummary>(S.get()); 884 if (!FS) 885 continue; 886 if (!FS->type_test_assume_vcalls().empty() || 887 !FS->type_checked_load_vcalls().empty() || 888 !FS->type_test_assume_const_vcalls().empty() || 889 !FS->type_checked_load_const_vcalls().empty() || 890 !FS->type_tests().empty()) 891 return make_error<StringError>( 892 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)", 893 inconvertibleErrorCode()); 894 } 895 } 896 return Error::success(); 897 } 898 899 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) { 900 // Compute "dead" symbols, we don't want to import/export these! 901 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols; 902 DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions; 903 for (auto &Res : GlobalResolutions) { 904 // Normally resolution have IR name of symbol. We can do nothing here 905 // otherwise. See comments in GlobalResolution struct for more details. 906 if (Res.second.IRName.empty()) 907 continue; 908 909 GlobalValue::GUID GUID = GlobalValue::getGUID( 910 GlobalValue::dropLLVMManglingEscape(Res.second.IRName)); 911 912 if (Res.second.VisibleOutsideSummary && Res.second.Prevailing) 913 GUIDPreservedSymbols.insert(GUID); 914 915 GUIDPrevailingResolutions[GUID] = 916 Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No; 917 } 918 919 auto isPrevailing = [&](GlobalValue::GUID G) { 920 auto It = GUIDPrevailingResolutions.find(G); 921 if (It == GUIDPrevailingResolutions.end()) 922 return PrevailingType::Unknown; 923 return It->second; 924 }; 925 computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols, 926 isPrevailing, Conf.OptLevel > 0); 927 928 // Setup output file to emit statistics. 929 auto StatsFileOrErr = setupStatsFile(Conf.StatsFile); 930 if (!StatsFileOrErr) 931 return StatsFileOrErr.takeError(); 932 std::unique_ptr<ToolOutputFile> StatsFile = std::move(StatsFileOrErr.get()); 933 934 Error Result = runRegularLTO(AddStream); 935 if (!Result) 936 Result = runThinLTO(AddStream, Cache, GUIDPreservedSymbols); 937 938 if (StatsFile) 939 PrintStatisticsJSON(StatsFile->os()); 940 941 return Result; 942 } 943 944 Error LTO::runRegularLTO(AddStreamFn AddStream) { 945 // Setup optimization remarks. 946 auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks( 947 RegularLTO.CombinedModule->getContext(), Conf.RemarksFilename, 948 Conf.RemarksPasses, Conf.RemarksFormat, Conf.RemarksWithHotness); 949 if (!DiagFileOrErr) 950 return DiagFileOrErr.takeError(); 951 952 // Finalize linking of regular LTO modules containing summaries now that 953 // we have computed liveness information. 954 for (auto &M : RegularLTO.ModsWithSummaries) 955 if (Error Err = linkRegularLTO(std::move(M), 956 /*LivenessFromIndex=*/true)) 957 return Err; 958 959 // Ensure we don't have inconsistently split LTO units with type tests. 960 // FIXME: this checks both LTO and ThinLTO. It happens to work as we take 961 // this path both cases but eventually this should be split into two and 962 // do the ThinLTO checks in `runThinLTO`. 963 if (Error Err = checkPartiallySplit()) 964 return Err; 965 966 // Make sure commons have the right size/alignment: we kept the largest from 967 // all the prevailing when adding the inputs, and we apply it here. 968 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout(); 969 for (auto &I : RegularLTO.Commons) { 970 if (!I.second.Prevailing) 971 // Don't do anything if no instance of this common was prevailing. 972 continue; 973 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first); 974 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) { 975 // Don't create a new global if the type is already correct, just make 976 // sure the alignment is correct. 977 OldGV->setAlignment(I.second.Align); 978 continue; 979 } 980 ArrayType *Ty = 981 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size); 982 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false, 983 GlobalValue::CommonLinkage, 984 ConstantAggregateZero::get(Ty), ""); 985 GV->setAlignment(I.second.Align); 986 if (OldGV) { 987 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType())); 988 GV->takeName(OldGV); 989 OldGV->eraseFromParent(); 990 } else { 991 GV->setName(I.first); 992 } 993 } 994 995 // If allowed, upgrade public vcall visibility metadata to linkage unit 996 // visibility before whole program devirtualization in the optimizer. 997 updateVCallVisibilityInModule(*RegularLTO.CombinedModule, 998 Conf.HasWholeProgramVisibility); 999 1000 if (Conf.PreOptModuleHook && 1001 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule)) 1002 return Error::success(); 1003 1004 if (!Conf.CodeGenOnly) { 1005 for (const auto &R : GlobalResolutions) { 1006 if (!R.second.isPrevailingIRSymbol()) 1007 continue; 1008 if (R.second.Partition != 0 && 1009 R.second.Partition != GlobalResolution::External) 1010 continue; 1011 1012 GlobalValue *GV = 1013 RegularLTO.CombinedModule->getNamedValue(R.second.IRName); 1014 // Ignore symbols defined in other partitions. 1015 // Also skip declarations, which are not allowed to have internal linkage. 1016 if (!GV || GV->hasLocalLinkage() || GV->isDeclaration()) 1017 continue; 1018 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global 1019 : GlobalValue::UnnamedAddr::None); 1020 if (EnableLTOInternalization && R.second.Partition == 0) 1021 GV->setLinkage(GlobalValue::InternalLinkage); 1022 } 1023 1024 RegularLTO.CombinedModule->addModuleFlag(Module::Error, "LTOPostLink", 1); 1025 1026 if (Conf.PostInternalizeModuleHook && 1027 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule)) 1028 return Error::success(); 1029 } 1030 1031 if (!RegularLTO.EmptyCombinedModule || Conf.AlwaysEmitRegularLTOObj) { 1032 if (Error Err = backend( 1033 Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel, 1034 std::move(RegularLTO.CombinedModule), ThinLTO.CombinedIndex)) 1035 return Err; 1036 } 1037 1038 return finalizeOptimizationRemarks(std::move(*DiagFileOrErr)); 1039 } 1040 1041 static const char *libcallRoutineNames[] = { 1042 #define HANDLE_LIBCALL(code, name) name, 1043 #include "llvm/IR/RuntimeLibcalls.def" 1044 #undef HANDLE_LIBCALL 1045 }; 1046 1047 ArrayRef<const char*> LTO::getRuntimeLibcallSymbols() { 1048 return makeArrayRef(libcallRoutineNames); 1049 } 1050 1051 /// This class defines the interface to the ThinLTO backend. 1052 class lto::ThinBackendProc { 1053 protected: 1054 const Config &Conf; 1055 ModuleSummaryIndex &CombinedIndex; 1056 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries; 1057 1058 public: 1059 ThinBackendProc(const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1060 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries) 1061 : Conf(Conf), CombinedIndex(CombinedIndex), 1062 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries) {} 1063 1064 virtual ~ThinBackendProc() {} 1065 virtual Error start( 1066 unsigned Task, BitcodeModule BM, 1067 const FunctionImporter::ImportMapTy &ImportList, 1068 const FunctionImporter::ExportSetTy &ExportList, 1069 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1070 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0; 1071 virtual Error wait() = 0; 1072 }; 1073 1074 namespace { 1075 class InProcessThinBackend : public ThinBackendProc { 1076 ThreadPool BackendThreadPool; 1077 AddStreamFn AddStream; 1078 NativeObjectCache Cache; 1079 std::set<GlobalValue::GUID> CfiFunctionDefs; 1080 std::set<GlobalValue::GUID> CfiFunctionDecls; 1081 1082 Optional<Error> Err; 1083 std::mutex ErrMu; 1084 1085 public: 1086 InProcessThinBackend( 1087 const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1088 ThreadPoolStrategy ThinLTOParallelism, 1089 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1090 AddStreamFn AddStream, NativeObjectCache Cache) 1091 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 1092 BackendThreadPool(ThinLTOParallelism), AddStream(std::move(AddStream)), 1093 Cache(std::move(Cache)) { 1094 for (auto &Name : CombinedIndex.cfiFunctionDefs()) 1095 CfiFunctionDefs.insert( 1096 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name))); 1097 for (auto &Name : CombinedIndex.cfiFunctionDecls()) 1098 CfiFunctionDecls.insert( 1099 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name))); 1100 } 1101 1102 Error runThinLTOBackendThread( 1103 AddStreamFn AddStream, NativeObjectCache Cache, unsigned Task, 1104 BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 1105 const FunctionImporter::ImportMapTy &ImportList, 1106 const FunctionImporter::ExportSetTy &ExportList, 1107 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1108 const GVSummaryMapTy &DefinedGlobals, 1109 MapVector<StringRef, BitcodeModule> &ModuleMap) { 1110 auto RunThinBackend = [&](AddStreamFn AddStream) { 1111 LTOLLVMContext BackendContext(Conf); 1112 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext); 1113 if (!MOrErr) 1114 return MOrErr.takeError(); 1115 1116 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex, 1117 ImportList, DefinedGlobals, ModuleMap); 1118 }; 1119 1120 auto ModuleID = BM.getModuleIdentifier(); 1121 1122 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) || 1123 all_of(CombinedIndex.getModuleHash(ModuleID), 1124 [](uint32_t V) { return V == 0; })) 1125 // Cache disabled or no entry for this module in the combined index or 1126 // no module hash. 1127 return RunThinBackend(AddStream); 1128 1129 SmallString<40> Key; 1130 // The module may be cached, this helps handling it. 1131 computeLTOCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList, 1132 ExportList, ResolvedODR, DefinedGlobals, CfiFunctionDefs, 1133 CfiFunctionDecls); 1134 if (AddStreamFn CacheAddStream = Cache(Task, Key)) 1135 return RunThinBackend(CacheAddStream); 1136 1137 return Error::success(); 1138 } 1139 1140 Error start( 1141 unsigned Task, BitcodeModule BM, 1142 const FunctionImporter::ImportMapTy &ImportList, 1143 const FunctionImporter::ExportSetTy &ExportList, 1144 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1145 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 1146 StringRef ModulePath = BM.getModuleIdentifier(); 1147 assert(ModuleToDefinedGVSummaries.count(ModulePath)); 1148 const GVSummaryMapTy &DefinedGlobals = 1149 ModuleToDefinedGVSummaries.find(ModulePath)->second; 1150 BackendThreadPool.async( 1151 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, 1152 const FunctionImporter::ImportMapTy &ImportList, 1153 const FunctionImporter::ExportSetTy &ExportList, 1154 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> 1155 &ResolvedODR, 1156 const GVSummaryMapTy &DefinedGlobals, 1157 MapVector<StringRef, BitcodeModule> &ModuleMap) { 1158 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) 1159 timeTraceProfilerInitialize(Conf.TimeTraceGranularity, 1160 "thin backend"); 1161 Error E = runThinLTOBackendThread( 1162 AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList, 1163 ResolvedODR, DefinedGlobals, ModuleMap); 1164 if (E) { 1165 std::unique_lock<std::mutex> L(ErrMu); 1166 if (Err) 1167 Err = joinErrors(std::move(*Err), std::move(E)); 1168 else 1169 Err = std::move(E); 1170 } 1171 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) 1172 timeTraceProfilerFinishThread(); 1173 }, 1174 BM, std::ref(CombinedIndex), std::ref(ImportList), std::ref(ExportList), 1175 std::ref(ResolvedODR), std::ref(DefinedGlobals), std::ref(ModuleMap)); 1176 return Error::success(); 1177 } 1178 1179 Error wait() override { 1180 BackendThreadPool.wait(); 1181 if (Err) 1182 return std::move(*Err); 1183 else 1184 return Error::success(); 1185 } 1186 }; 1187 } // end anonymous namespace 1188 1189 ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism) { 1190 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1191 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1192 AddStreamFn AddStream, NativeObjectCache Cache) { 1193 return std::make_unique<InProcessThinBackend>( 1194 Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries, AddStream, 1195 Cache); 1196 }; 1197 } 1198 1199 // Given the original \p Path to an output file, replace any path 1200 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the 1201 // resulting directory if it does not yet exist. 1202 std::string lto::getThinLTOOutputFile(const std::string &Path, 1203 const std::string &OldPrefix, 1204 const std::string &NewPrefix) { 1205 if (OldPrefix.empty() && NewPrefix.empty()) 1206 return Path; 1207 SmallString<128> NewPath(Path); 1208 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix); 1209 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str()); 1210 if (!ParentPath.empty()) { 1211 // Make sure the new directory exists, creating it if necessary. 1212 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath)) 1213 llvm::errs() << "warning: could not create directory '" << ParentPath 1214 << "': " << EC.message() << '\n'; 1215 } 1216 return std::string(NewPath.str()); 1217 } 1218 1219 namespace { 1220 class WriteIndexesThinBackend : public ThinBackendProc { 1221 std::string OldPrefix, NewPrefix; 1222 bool ShouldEmitImportsFiles; 1223 raw_fd_ostream *LinkedObjectsFile; 1224 lto::IndexWriteCallback OnWrite; 1225 1226 public: 1227 WriteIndexesThinBackend( 1228 const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1229 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1230 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles, 1231 raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite) 1232 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries), 1233 OldPrefix(OldPrefix), NewPrefix(NewPrefix), 1234 ShouldEmitImportsFiles(ShouldEmitImportsFiles), 1235 LinkedObjectsFile(LinkedObjectsFile), OnWrite(OnWrite) {} 1236 1237 Error start( 1238 unsigned Task, BitcodeModule BM, 1239 const FunctionImporter::ImportMapTy &ImportList, 1240 const FunctionImporter::ExportSetTy &ExportList, 1241 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, 1242 MapVector<StringRef, BitcodeModule> &ModuleMap) override { 1243 StringRef ModulePath = BM.getModuleIdentifier(); 1244 std::string NewModulePath = 1245 getThinLTOOutputFile(std::string(ModulePath), OldPrefix, NewPrefix); 1246 1247 if (LinkedObjectsFile) 1248 *LinkedObjectsFile << NewModulePath << '\n'; 1249 1250 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex; 1251 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries, 1252 ImportList, ModuleToSummariesForIndex); 1253 1254 std::error_code EC; 1255 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC, 1256 sys::fs::OpenFlags::OF_None); 1257 if (EC) 1258 return errorCodeToError(EC); 1259 WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex); 1260 1261 if (ShouldEmitImportsFiles) { 1262 EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports", 1263 ModuleToSummariesForIndex); 1264 if (EC) 1265 return errorCodeToError(EC); 1266 } 1267 1268 if (OnWrite) 1269 OnWrite(std::string(ModulePath)); 1270 return Error::success(); 1271 } 1272 1273 Error wait() override { return Error::success(); } 1274 }; 1275 } // end anonymous namespace 1276 1277 ThinBackend lto::createWriteIndexesThinBackend( 1278 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles, 1279 raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) { 1280 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, 1281 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, 1282 AddStreamFn AddStream, NativeObjectCache Cache) { 1283 return std::make_unique<WriteIndexesThinBackend>( 1284 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix, 1285 ShouldEmitImportsFiles, LinkedObjectsFile, OnWrite); 1286 }; 1287 } 1288 1289 Error LTO::runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache, 1290 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { 1291 if (ThinLTO.ModuleMap.empty()) 1292 return Error::success(); 1293 1294 if (Conf.CombinedIndexHook && 1295 !Conf.CombinedIndexHook(ThinLTO.CombinedIndex, GUIDPreservedSymbols)) 1296 return Error::success(); 1297 1298 // Collect for each module the list of function it defines (GUID -> 1299 // Summary). 1300 StringMap<GVSummaryMapTy> 1301 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size()); 1302 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule( 1303 ModuleToDefinedGVSummaries); 1304 // Create entries for any modules that didn't have any GV summaries 1305 // (either they didn't have any GVs to start with, or we suppressed 1306 // generation of the summaries because they e.g. had inline assembly 1307 // uses that couldn't be promoted/renamed on export). This is so 1308 // InProcessThinBackend::start can still launch a backend thread, which 1309 // is passed the map of summaries for the module, without any special 1310 // handling for this case. 1311 for (auto &Mod : ThinLTO.ModuleMap) 1312 if (!ModuleToDefinedGVSummaries.count(Mod.first)) 1313 ModuleToDefinedGVSummaries.try_emplace(Mod.first); 1314 1315 // Synthesize entry counts for functions in the CombinedIndex. 1316 computeSyntheticCounts(ThinLTO.CombinedIndex); 1317 1318 StringMap<FunctionImporter::ImportMapTy> ImportLists( 1319 ThinLTO.ModuleMap.size()); 1320 StringMap<FunctionImporter::ExportSetTy> ExportLists( 1321 ThinLTO.ModuleMap.size()); 1322 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; 1323 1324 if (DumpThinCGSCCs) 1325 ThinLTO.CombinedIndex.dumpSCCs(outs()); 1326 1327 std::set<GlobalValue::GUID> ExportedGUIDs; 1328 1329 // If allowed, upgrade public vcall visibility to linkage unit visibility in 1330 // the summaries before whole program devirtualization below. 1331 updateVCallVisibilityInIndex(ThinLTO.CombinedIndex, 1332 Conf.HasWholeProgramVisibility); 1333 1334 // Perform index-based WPD. This will return immediately if there are 1335 // no index entries in the typeIdMetadata map (e.g. if we are instead 1336 // performing IR-based WPD in hybrid regular/thin LTO mode). 1337 std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap; 1338 runWholeProgramDevirtOnIndex(ThinLTO.CombinedIndex, ExportedGUIDs, 1339 LocalWPDTargetsMap); 1340 1341 if (Conf.OptLevel > 0) 1342 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 1343 ImportLists, ExportLists); 1344 1345 // Figure out which symbols need to be internalized. This also needs to happen 1346 // at -O0 because summary-based DCE is implemented using internalization, and 1347 // we must apply DCE consistently with the full LTO module in order to avoid 1348 // undefined references during the final link. 1349 for (auto &Res : GlobalResolutions) { 1350 // If the symbol does not have external references or it is not prevailing, 1351 // then not need to mark it as exported from a ThinLTO partition. 1352 if (Res.second.Partition != GlobalResolution::External || 1353 !Res.second.isPrevailingIRSymbol()) 1354 continue; 1355 auto GUID = GlobalValue::getGUID( 1356 GlobalValue::dropLLVMManglingEscape(Res.second.IRName)); 1357 // Mark exported unless index-based analysis determined it to be dead. 1358 if (ThinLTO.CombinedIndex.isGUIDLive(GUID)) 1359 ExportedGUIDs.insert(GUID); 1360 } 1361 1362 // Any functions referenced by the jump table in the regular LTO object must 1363 // be exported. 1364 for (auto &Def : ThinLTO.CombinedIndex.cfiFunctionDefs()) 1365 ExportedGUIDs.insert( 1366 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Def))); 1367 1368 auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) { 1369 const auto &ExportList = ExportLists.find(ModuleIdentifier); 1370 return (ExportList != ExportLists.end() && ExportList->second.count(VI)) || 1371 ExportedGUIDs.count(VI.getGUID()); 1372 }; 1373 1374 // Update local devirtualized targets that were exported by cross-module 1375 // importing or by other devirtualizations marked in the ExportedGUIDs set. 1376 updateIndexWPDForExports(ThinLTO.CombinedIndex, isExported, 1377 LocalWPDTargetsMap); 1378 1379 auto isPrevailing = [&](GlobalValue::GUID GUID, 1380 const GlobalValueSummary *S) { 1381 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath(); 1382 }; 1383 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported, 1384 isPrevailing); 1385 1386 auto recordNewLinkage = [&](StringRef ModuleIdentifier, 1387 GlobalValue::GUID GUID, 1388 GlobalValue::LinkageTypes NewLinkage) { 1389 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage; 1390 }; 1391 thinLTOResolvePrevailingInIndex(ThinLTO.CombinedIndex, isPrevailing, 1392 recordNewLinkage, GUIDPreservedSymbols); 1393 1394 std::unique_ptr<ThinBackendProc> BackendProc = 1395 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, 1396 AddStream, Cache); 1397 1398 // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for combined 1399 // module and parallel code generation partitions. 1400 unsigned Task = RegularLTO.ParallelCodeGenParallelismLevel; 1401 for (auto &Mod : ThinLTO.ModuleMap) { 1402 if (Error E = BackendProc->start(Task, Mod.second, ImportLists[Mod.first], 1403 ExportLists[Mod.first], 1404 ResolvedODR[Mod.first], ThinLTO.ModuleMap)) 1405 return E; 1406 ++Task; 1407 } 1408 1409 return BackendProc->wait(); 1410 } 1411 1412 Expected<std::unique_ptr<ToolOutputFile>> lto::setupLLVMOptimizationRemarks( 1413 LLVMContext &Context, StringRef RemarksFilename, StringRef RemarksPasses, 1414 StringRef RemarksFormat, bool RemarksWithHotness, int Count) { 1415 std::string Filename = std::string(RemarksFilename); 1416 // For ThinLTO, file.opt.<format> becomes 1417 // file.opt.<format>.thin.<num>.<format>. 1418 if (!Filename.empty() && Count != -1) 1419 Filename = 1420 (Twine(Filename) + ".thin." + llvm::utostr(Count) + "." + RemarksFormat) 1421 .str(); 1422 1423 auto ResultOrErr = llvm::setupLLVMOptimizationRemarks( 1424 Context, Filename, RemarksPasses, RemarksFormat, RemarksWithHotness); 1425 if (Error E = ResultOrErr.takeError()) 1426 return std::move(E); 1427 1428 if (*ResultOrErr) 1429 (*ResultOrErr)->keep(); 1430 1431 return ResultOrErr; 1432 } 1433 1434 Expected<std::unique_ptr<ToolOutputFile>> 1435 lto::setupStatsFile(StringRef StatsFilename) { 1436 // Setup output file to emit statistics. 1437 if (StatsFilename.empty()) 1438 return nullptr; 1439 1440 llvm::EnableStatistics(false); 1441 std::error_code EC; 1442 auto StatsFile = 1443 std::make_unique<ToolOutputFile>(StatsFilename, EC, sys::fs::OF_None); 1444 if (EC) 1445 return errorCodeToError(EC); 1446 1447 StatsFile->keep(); 1448 return std::move(StatsFile); 1449 } 1450