| 1 | //===-ThinLTOCodeGenerator.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 the Thin Link Time Optimization library. This library is |
| 10 | // intended to be used by linker to optimize code at link time. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/LTO/legacy/ThinLTOCodeGenerator.h" |
| 15 | #include "llvm/Support/CommandLine.h" |
| 16 | |
| 17 | #include "llvm/ADT/ScopeExit.h" |
| 18 | #include "llvm/ADT/Statistic.h" |
| 19 | #include "llvm/ADT/StringExtras.h" |
| 20 | #include "llvm/Analysis/ModuleSummaryAnalysis.h" |
| 21 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
| 22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 23 | #include "llvm/Bitcode/BitcodeReader.h" |
| 24 | #include "llvm/Bitcode/BitcodeWriter.h" |
| 25 | #include "llvm/Bitcode/BitcodeWriterPass.h" |
| 26 | #include "llvm/Config/llvm-config.h" |
| 27 | #include "llvm/IR/DebugInfo.h" |
| 28 | #include "llvm/IR/DiagnosticPrinter.h" |
| 29 | #include "llvm/IR/LLVMContext.h" |
| 30 | #include "llvm/IR/LLVMRemarkStreamer.h" |
| 31 | #include "llvm/IR/LegacyPassManager.h" |
| 32 | #include "llvm/IR/Mangler.h" |
| 33 | #include "llvm/IR/PassTimingInfo.h" |
| 34 | #include "llvm/IR/Verifier.h" |
| 35 | #include "llvm/IRReader/IRReader.h" |
| 36 | #include "llvm/LTO/LTO.h" |
| 37 | #include "llvm/MC/TargetRegistry.h" |
| 38 | #include "llvm/Object/IRObjectFile.h" |
| 39 | #include "llvm/Passes/PassBuilder.h" |
| 40 | #include "llvm/Passes/StandardInstrumentations.h" |
| 41 | #include "llvm/Remarks/HotnessThresholdParser.h" |
| 42 | #include "llvm/Support/CachePruning.h" |
| 43 | #include "llvm/Support/Debug.h" |
| 44 | #include "llvm/Support/Error.h" |
| 45 | #include "llvm/Support/FileSystem.h" |
| 46 | #include "llvm/Support/FormatVariadic.h" |
| 47 | #include "llvm/Support/Path.h" |
| 48 | #include "llvm/Support/SHA1.h" |
| 49 | #include "llvm/Support/SmallVectorMemoryBuffer.h" |
| 50 | #include "llvm/Support/ThreadPool.h" |
| 51 | #include "llvm/Support/Threading.h" |
| 52 | #include "llvm/Support/ToolOutputFile.h" |
| 53 | #include "llvm/Support/raw_ostream.h" |
| 54 | #include "llvm/Target/TargetMachine.h" |
| 55 | #include "llvm/TargetParser/SubtargetFeature.h" |
| 56 | #include "llvm/Transforms/IPO/FunctionAttrs.h" |
| 57 | #include "llvm/Transforms/IPO/FunctionImport.h" |
| 58 | #include "llvm/Transforms/IPO/Internalize.h" |
| 59 | #include "llvm/Transforms/IPO/WholeProgramDevirt.h" |
| 60 | #include "llvm/Transforms/Utils/FunctionImportUtils.h" |
| 61 | |
| 62 | #if !defined(_MSC_VER) && !defined(__MINGW32__) |
| 63 | #include <unistd.h> |
| 64 | #else |
| 65 | #include <io.h> |
| 66 | #endif |
| 67 | |
| 68 | using namespace llvm; |
| 69 | using namespace ThinLTOCodeGeneratorImpl; |
| 70 | |
| 71 | #define DEBUG_TYPE "thinlto" |
| 72 | |
| 73 | namespace llvm { |
| 74 | // Flags -discard-value-names, defined in LTOCodeGenerator.cpp |
| 75 | extern cl::opt<bool> LTODiscardValueNames; |
| 76 | extern cl::opt<std::string> ; |
| 77 | extern cl::opt<std::string> ; |
| 78 | extern cl::opt<bool> ; |
| 79 | extern cl::opt<std::optional<uint64_t>, false, remarks::HotnessThresholdParser> |
| 80 | ; |
| 81 | extern cl::opt<std::string> ; |
| 82 | } |
| 83 | |
| 84 | // Default to using all available threads in the system, but using only one |
| 85 | // thred per core, as indicated by the usage of |
| 86 | // heavyweight_hardware_concurrency() below. |
| 87 | static cl::opt<int> ThreadCount("threads" , cl::init(Val: 0)); |
| 88 | |
| 89 | // Simple helper to save temporary files for debug. |
| 90 | static void saveTempBitcode(const Module &TheModule, StringRef TempDir, |
| 91 | unsigned count, StringRef Suffix) { |
| 92 | if (TempDir.empty()) |
| 93 | return; |
| 94 | // User asked to save temps, let dump the bitcode file after import. |
| 95 | std::string SaveTempPath = (TempDir + llvm::Twine(count) + Suffix).str(); |
| 96 | std::error_code EC; |
| 97 | raw_fd_ostream OS(SaveTempPath, EC, sys::fs::OF_None); |
| 98 | if (EC) |
| 99 | report_fatal_error(reason: Twine("Failed to open " ) + SaveTempPath + |
| 100 | " to save optimized bitcode\n" ); |
| 101 | WriteBitcodeToFile(M: TheModule, Out&: OS, /* ShouldPreserveUseListOrder */ true); |
| 102 | } |
| 103 | |
| 104 | static const GlobalValueSummary * |
| 105 | getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) { |
| 106 | // If there is any strong definition anywhere, get it. |
| 107 | auto StrongDefForLinker = llvm::find_if( |
| 108 | Range: GVSummaryList, P: [](const std::unique_ptr<GlobalValueSummary> &Summary) { |
| 109 | auto Linkage = Summary->linkage(); |
| 110 | return !GlobalValue::isAvailableExternallyLinkage(Linkage) && |
| 111 | !GlobalValue::isWeakForLinker(Linkage); |
| 112 | }); |
| 113 | if (StrongDefForLinker != GVSummaryList.end()) |
| 114 | return StrongDefForLinker->get(); |
| 115 | // Get the first *linker visible* definition for this global in the summary |
| 116 | // list. |
| 117 | auto FirstDefForLinker = llvm::find_if( |
| 118 | Range: GVSummaryList, P: [](const std::unique_ptr<GlobalValueSummary> &Summary) { |
| 119 | auto Linkage = Summary->linkage(); |
| 120 | return !GlobalValue::isAvailableExternallyLinkage(Linkage); |
| 121 | }); |
| 122 | // Extern templates can be emitted as available_externally. |
| 123 | if (FirstDefForLinker == GVSummaryList.end()) |
| 124 | return nullptr; |
| 125 | return FirstDefForLinker->get(); |
| 126 | } |
| 127 | |
| 128 | // Populate map of GUID to the prevailing copy for any multiply defined |
| 129 | // symbols. Currently assume first copy is prevailing, or any strong |
| 130 | // definition. Can be refined with Linker information in the future. |
| 131 | static void computePrevailingCopies( |
| 132 | const ModuleSummaryIndex &Index, |
| 133 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy) { |
| 134 | auto HasMultipleCopies = [&](const GlobalValueSummaryList &GVSummaryList) { |
| 135 | return GVSummaryList.size() > 1; |
| 136 | }; |
| 137 | |
| 138 | for (auto &I : Index) { |
| 139 | if (HasMultipleCopies(I.second.SummaryList)) |
| 140 | PrevailingCopy[I.first] = |
| 141 | getFirstDefinitionForLinker(GVSummaryList: I.second.SummaryList); |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | static StringMap<lto::InputFile *> |
| 146 | generateModuleMap(std::vector<std::unique_ptr<lto::InputFile>> &Modules) { |
| 147 | StringMap<lto::InputFile *> ModuleMap; |
| 148 | for (auto &M : Modules) { |
| 149 | LLVM_DEBUG(dbgs() << "Adding module " << M->getName() << " to ModuleMap\n" ); |
| 150 | assert(!ModuleMap.contains(M->getName()) && |
| 151 | "Expect unique Buffer Identifier" ); |
| 152 | ModuleMap[M->getName()] = M.get(); |
| 153 | } |
| 154 | return ModuleMap; |
| 155 | } |
| 156 | |
| 157 | static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index, |
| 158 | bool ClearDSOLocalOnDeclarations) { |
| 159 | renameModuleForThinLTO(M&: TheModule, Index, ClearDSOLocalOnDeclarations); |
| 160 | } |
| 161 | |
| 162 | namespace { |
| 163 | class ThinLTODiagnosticInfo : public DiagnosticInfo { |
| 164 | const Twine &Msg; |
| 165 | public: |
| 166 | ThinLTODiagnosticInfo(const Twine &DiagMsg LLVM_LIFETIME_BOUND, |
| 167 | DiagnosticSeverity Severity = DS_Error) |
| 168 | : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {} |
| 169 | void print(DiagnosticPrinter &DP) const override { DP << Msg; } |
| 170 | }; |
| 171 | } |
| 172 | |
| 173 | /// Verify the module and strip broken debug info. |
| 174 | static void verifyLoadedModule(Module &TheModule) { |
| 175 | bool BrokenDebugInfo = false; |
| 176 | if (verifyModule(M: TheModule, OS: &dbgs(), BrokenDebugInfo: &BrokenDebugInfo)) |
| 177 | report_fatal_error(reason: "Broken module found, compilation aborted!" ); |
| 178 | if (BrokenDebugInfo) { |
| 179 | TheModule.getContext().diagnose(DI: ThinLTODiagnosticInfo( |
| 180 | "Invalid debug info found, debug info will be stripped" , DS_Warning)); |
| 181 | StripDebugInfo(M&: TheModule); |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | static std::unique_ptr<Module> loadModuleFromInput(lto::InputFile *Input, |
| 186 | LLVMContext &Context, |
| 187 | bool Lazy, |
| 188 | bool IsImporting) { |
| 189 | auto &Mod = Input->getSingleBitcodeModule(); |
| 190 | SMDiagnostic Err; |
| 191 | Expected<std::unique_ptr<Module>> ModuleOrErr = |
| 192 | Lazy ? Mod.getLazyModule(Context, |
| 193 | /* ShouldLazyLoadMetadata */ true, IsImporting) |
| 194 | : Mod.parseModule(Context); |
| 195 | if (!ModuleOrErr) { |
| 196 | handleAllErrors(E: ModuleOrErr.takeError(), Handlers: [&](ErrorInfoBase &EIB) { |
| 197 | SMDiagnostic Err = SMDiagnostic(Mod.getModuleIdentifier(), |
| 198 | SourceMgr::DK_Error, EIB.message()); |
| 199 | Err.print(ProgName: "ThinLTO" , S&: errs()); |
| 200 | }); |
| 201 | report_fatal_error(reason: "Can't load module, abort." ); |
| 202 | } |
| 203 | if (!Lazy) |
| 204 | verifyLoadedModule(TheModule&: *ModuleOrErr.get()); |
| 205 | return std::move(*ModuleOrErr); |
| 206 | } |
| 207 | |
| 208 | static void |
| 209 | crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index, |
| 210 | StringMap<lto::InputFile *> &ModuleMap, |
| 211 | const FunctionImporter::ImportMapTy &ImportList, |
| 212 | bool ClearDSOLocalOnDeclarations) { |
| 213 | auto Loader = [&](StringRef Identifier) { |
| 214 | auto &Input = ModuleMap[Identifier]; |
| 215 | return loadModuleFromInput(Input, Context&: TheModule.getContext(), |
| 216 | /*Lazy=*/true, /*IsImporting*/ true); |
| 217 | }; |
| 218 | |
| 219 | FunctionImporter Importer(Index, Loader, ClearDSOLocalOnDeclarations); |
| 220 | Expected<bool> Result = Importer.importFunctions(M&: TheModule, ImportList); |
| 221 | if (!Result) { |
| 222 | handleAllErrors(E: Result.takeError(), Handlers: [&](ErrorInfoBase &EIB) { |
| 223 | SMDiagnostic Err = SMDiagnostic(TheModule.getModuleIdentifier(), |
| 224 | SourceMgr::DK_Error, EIB.message()); |
| 225 | Err.print(ProgName: "ThinLTO" , S&: errs()); |
| 226 | }); |
| 227 | report_fatal_error(reason: "importFunctions failed" ); |
| 228 | } |
| 229 | // Verify again after cross-importing. |
| 230 | verifyLoadedModule(TheModule); |
| 231 | } |
| 232 | |
| 233 | static void optimizeModule(Module &TheModule, TargetMachine &TM, |
| 234 | unsigned OptLevel, bool Freestanding, |
| 235 | bool DebugPassManager, ModuleSummaryIndex *Index) { |
| 236 | std::optional<PGOOptions> PGOOpt; |
| 237 | LoopAnalysisManager LAM; |
| 238 | FunctionAnalysisManager FAM; |
| 239 | CGSCCAnalysisManager CGAM; |
| 240 | ModuleAnalysisManager MAM; |
| 241 | |
| 242 | PassInstrumentationCallbacks PIC; |
| 243 | StandardInstrumentations SI(TheModule.getContext(), DebugPassManager); |
| 244 | SI.registerCallbacks(PIC, MAM: &MAM); |
| 245 | PipelineTuningOptions PTO; |
| 246 | PTO.LoopVectorization = true; |
| 247 | PTO.SLPVectorization = true; |
| 248 | PassBuilder PB(&TM, PTO, PGOOpt, &PIC); |
| 249 | |
| 250 | std::unique_ptr<TargetLibraryInfoImpl> TLII( |
| 251 | new TargetLibraryInfoImpl(TM.getTargetTriple())); |
| 252 | if (Freestanding) |
| 253 | TLII->disableAllFunctions(); |
| 254 | FAM.registerPass(PassBuilder: [&] { return TargetLibraryAnalysis(*TLII); }); |
| 255 | |
| 256 | // Register all the basic analyses with the managers. |
| 257 | PB.registerModuleAnalyses(MAM); |
| 258 | PB.registerCGSCCAnalyses(CGAM); |
| 259 | PB.registerFunctionAnalyses(FAM); |
| 260 | PB.registerLoopAnalyses(LAM); |
| 261 | PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); |
| 262 | |
| 263 | ModulePassManager MPM; |
| 264 | |
| 265 | OptimizationLevel OL; |
| 266 | |
| 267 | switch (OptLevel) { |
| 268 | default: |
| 269 | llvm_unreachable("Invalid optimization level" ); |
| 270 | case 0: |
| 271 | OL = OptimizationLevel::O0; |
| 272 | break; |
| 273 | case 1: |
| 274 | OL = OptimizationLevel::O1; |
| 275 | break; |
| 276 | case 2: |
| 277 | OL = OptimizationLevel::O2; |
| 278 | break; |
| 279 | case 3: |
| 280 | OL = OptimizationLevel::O3; |
| 281 | break; |
| 282 | } |
| 283 | |
| 284 | MPM.addPass(Pass: PB.buildThinLTODefaultPipeline(Level: OL, ImportSummary: Index)); |
| 285 | |
| 286 | MPM.run(IR&: TheModule, AM&: MAM); |
| 287 | } |
| 288 | |
| 289 | static void |
| 290 | addUsedSymbolToPreservedGUID(const lto::InputFile &File, |
| 291 | DenseSet<GlobalValue::GUID> &PreservedGUID) { |
| 292 | for (const auto &Sym : File.symbols()) { |
| 293 | if (Sym.isUsed()) |
| 294 | PreservedGUID.insert( |
| 295 | V: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: Sym.getIRName())); |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | // Convert the PreservedSymbols map from "Name" based to "GUID" based. |
| 300 | static void computeGUIDPreservedSymbols(const lto::InputFile &File, |
| 301 | const StringSet<> &PreservedSymbols, |
| 302 | const Triple &TheTriple, |
| 303 | DenseSet<GlobalValue::GUID> &GUIDs) { |
| 304 | // Iterate the symbols in the input file and if the input has preserved symbol |
| 305 | // compute the GUID for the symbol. |
| 306 | for (const auto &Sym : File.symbols()) { |
| 307 | if (PreservedSymbols.count(Key: Sym.getName()) && !Sym.getIRName().empty()) |
| 308 | GUIDs.insert(V: GlobalValue::getGUIDAssumingExternalLinkage( |
| 309 | GlobalName: GlobalValue::getGlobalIdentifier(Name: Sym.getIRName(), |
| 310 | Linkage: GlobalValue::ExternalLinkage, FileName: "" ))); |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | static DenseSet<GlobalValue::GUID> |
| 315 | computeGUIDPreservedSymbols(const lto::InputFile &File, |
| 316 | const StringSet<> &PreservedSymbols, |
| 317 | const Triple &TheTriple) { |
| 318 | DenseSet<GlobalValue::GUID> GUIDPreservedSymbols(PreservedSymbols.size()); |
| 319 | computeGUIDPreservedSymbols(File, PreservedSymbols, TheTriple, |
| 320 | GUIDs&: GUIDPreservedSymbols); |
| 321 | return GUIDPreservedSymbols; |
| 322 | } |
| 323 | |
| 324 | static std::unique_ptr<MemoryBuffer> codegenModule(Module &TheModule, |
| 325 | TargetMachine &TM) { |
| 326 | SmallVector<char, 128> OutputBuffer; |
| 327 | |
| 328 | // CodeGen |
| 329 | { |
| 330 | raw_svector_ostream OS(OutputBuffer); |
| 331 | legacy::PassManager PM; |
| 332 | |
| 333 | // Setup the codegen now. |
| 334 | if (TM.addPassesToEmitFile(PM, OS, nullptr, CodeGenFileType::ObjectFile, |
| 335 | /* DisableVerify */ true)) |
| 336 | report_fatal_error(reason: "Failed to setup codegen" ); |
| 337 | |
| 338 | // Run codegen now. resulting binary is in OutputBuffer. |
| 339 | PM.run(M&: TheModule); |
| 340 | } |
| 341 | return std::make_unique<SmallVectorMemoryBuffer>( |
| 342 | args: std::move(OutputBuffer), /*RequiresNullTerminator=*/args: false); |
| 343 | } |
| 344 | |
| 345 | namespace { |
| 346 | /// Manage caching for a single Module. |
| 347 | class ModuleCacheEntry { |
| 348 | SmallString<128> EntryPath; |
| 349 | |
| 350 | public: |
| 351 | // Create a cache entry. This compute a unique hash for the Module considering |
| 352 | // the current list of export/import, and offer an interface to query to |
| 353 | // access the content in the cache. |
| 354 | ModuleCacheEntry( |
| 355 | StringRef CachePath, const ModuleSummaryIndex &Index, StringRef ModuleID, |
| 356 | const FunctionImporter::ImportMapTy &ImportList, |
| 357 | const FunctionImporter::ExportSetTy &ExportList, |
| 358 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
| 359 | const GVSummaryMapTy &DefinedGVSummaries, unsigned OptLevel, |
| 360 | bool Freestanding, const TargetMachineBuilder &TMBuilder) { |
| 361 | if (CachePath.empty()) |
| 362 | return; |
| 363 | |
| 364 | if (!Index.modulePaths().count(Key: ModuleID)) |
| 365 | // The module does not have an entry, it can't have a hash at all |
| 366 | return; |
| 367 | |
| 368 | if (all_of(Range: Index.getModuleHash(ModPath: ModuleID), |
| 369 | P: [](uint32_t V) { return V == 0; })) |
| 370 | // No hash entry, no caching! |
| 371 | return; |
| 372 | |
| 373 | llvm::lto::Config Conf; |
| 374 | Conf.OptLevel = OptLevel; |
| 375 | Conf.Options = TMBuilder.Options; |
| 376 | Conf.CPU = TMBuilder.MCpu; |
| 377 | Conf.MAttrs.push_back(x: TMBuilder.MAttr); |
| 378 | Conf.RelocModel = TMBuilder.RelocModel; |
| 379 | Conf.CGOptLevel = TMBuilder.CGOptLevel; |
| 380 | Conf.Freestanding = Freestanding; |
| 381 | std::string Key = |
| 382 | computeLTOCacheKey(Conf, Index, ModuleID, ImportList, ExportList, |
| 383 | ResolvedODR, DefinedGlobals: DefinedGVSummaries); |
| 384 | |
| 385 | // This choice of file name allows the cache to be pruned (see pruneCache() |
| 386 | // in include/llvm/Support/CachePruning.h). |
| 387 | sys::path::append(path&: EntryPath, a: CachePath, b: Twine("llvmcache-" , Key)); |
| 388 | } |
| 389 | |
| 390 | // Access the path to this entry in the cache. |
| 391 | StringRef getEntryPath() { return EntryPath; } |
| 392 | |
| 393 | // Try loading the buffer for this cache entry. |
| 394 | ErrorOr<std::unique_ptr<MemoryBuffer>> tryLoadingBuffer() { |
| 395 | if (EntryPath.empty()) |
| 396 | return std::error_code(); |
| 397 | SmallString<64> ResultPath; |
| 398 | Expected<sys::fs::file_t> FDOrErr = sys::fs::openNativeFileForRead( |
| 399 | Name: Twine(EntryPath), Flags: sys::fs::OF_UpdateAtime, RealPath: &ResultPath); |
| 400 | if (!FDOrErr) |
| 401 | return errorToErrorCode(Err: FDOrErr.takeError()); |
| 402 | ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr = MemoryBuffer::getOpenFile( |
| 403 | FD: *FDOrErr, Filename: EntryPath, /*FileSize=*/-1, /*RequiresNullTerminator=*/false); |
| 404 | sys::fs::closeFile(F&: *FDOrErr); |
| 405 | return MBOrErr; |
| 406 | } |
| 407 | |
| 408 | // Cache the Produced object file |
| 409 | void write(const MemoryBuffer &OutputBuffer) { |
| 410 | if (EntryPath.empty()) |
| 411 | return; |
| 412 | |
| 413 | if (auto Err = llvm::writeToOutput( |
| 414 | OutputFileName: EntryPath, Write: [&OutputBuffer](llvm::raw_ostream &OS) -> llvm::Error { |
| 415 | OS << OutputBuffer.getBuffer(); |
| 416 | return llvm::Error::success(); |
| 417 | })) |
| 418 | report_fatal_error(reason: llvm::formatv(Fmt: "ThinLTO: Can't write file {0}: {1}" , |
| 419 | Vals&: EntryPath, |
| 420 | Vals: toString(E: std::move(Err)).c_str())); |
| 421 | } |
| 422 | }; |
| 423 | } // end anonymous namespace |
| 424 | |
| 425 | static std::unique_ptr<MemoryBuffer> |
| 426 | ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index, |
| 427 | StringMap<lto::InputFile *> &ModuleMap, TargetMachine &TM, |
| 428 | const FunctionImporter::ImportMapTy &ImportList, |
| 429 | const FunctionImporter::ExportSetTy &ExportList, |
| 430 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols, |
| 431 | const GVSummaryMapTy &DefinedGlobals, |
| 432 | const ThinLTOCodeGenerator::CachingOptions &CacheOptions, |
| 433 | bool DisableCodeGen, StringRef SaveTempsDir, |
| 434 | bool Freestanding, unsigned OptLevel, unsigned count, |
| 435 | bool DebugPassManager) { |
| 436 | // "Benchmark"-like optimization: single-source case |
| 437 | bool SingleModule = (ModuleMap.size() == 1); |
| 438 | |
| 439 | // When linking an ELF shared object, dso_local should be dropped. We |
| 440 | // conservatively do this for -fpic. |
| 441 | bool ClearDSOLocalOnDeclarations = |
| 442 | TM.getTargetTriple().isOSBinFormatELF() && |
| 443 | TM.getRelocationModel() != Reloc::Static && |
| 444 | TheModule.getPIELevel() == PIELevel::Default; |
| 445 | |
| 446 | if (!SingleModule) { |
| 447 | promoteModule(TheModule, Index, ClearDSOLocalOnDeclarations); |
| 448 | |
| 449 | // Apply summary-based prevailing-symbol resolution decisions. |
| 450 | thinLTOFinalizeInModule(TheModule, DefinedGlobals, /*PropagateAttrs=*/true); |
| 451 | |
| 452 | // Save temps: after promotion. |
| 453 | saveTempBitcode(TheModule, TempDir: SaveTempsDir, count, Suffix: ".1.promoted.bc" ); |
| 454 | } |
| 455 | |
| 456 | // Be friendly and don't nuke totally the module when the client didn't |
| 457 | // supply anything to preserve. |
| 458 | if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) { |
| 459 | // Apply summary-based internalization decisions. |
| 460 | thinLTOInternalizeModule(TheModule, DefinedGlobals); |
| 461 | } |
| 462 | |
| 463 | // Save internalized bitcode |
| 464 | saveTempBitcode(TheModule, TempDir: SaveTempsDir, count, Suffix: ".2.internalized.bc" ); |
| 465 | |
| 466 | if (!SingleModule) |
| 467 | crossImportIntoModule(TheModule, Index, ModuleMap, ImportList, |
| 468 | ClearDSOLocalOnDeclarations); |
| 469 | |
| 470 | // Do this after any importing so that imported code is updated. |
| 471 | // See comment at call to updateVCallVisibilityInIndex() for why |
| 472 | // WholeProgramVisibilityEnabledInLTO is false. |
| 473 | updatePublicTypeTestCalls(M&: TheModule, |
| 474 | /* WholeProgramVisibilityEnabledInLTO */ false); |
| 475 | |
| 476 | // Save temps: after cross-module import. |
| 477 | saveTempBitcode(TheModule, TempDir: SaveTempsDir, count, Suffix: ".3.imported.bc" ); |
| 478 | |
| 479 | optimizeModule(TheModule, TM, OptLevel, Freestanding, DebugPassManager, |
| 480 | Index: &Index); |
| 481 | |
| 482 | saveTempBitcode(TheModule, TempDir: SaveTempsDir, count, Suffix: ".4.opt.bc" ); |
| 483 | |
| 484 | if (DisableCodeGen) { |
| 485 | // Configured to stop before CodeGen, serialize the bitcode and return. |
| 486 | SmallVector<char, 128> OutputBuffer; |
| 487 | { |
| 488 | raw_svector_ostream OS(OutputBuffer); |
| 489 | ProfileSummaryInfo PSI(TheModule); |
| 490 | auto Index = buildModuleSummaryIndex(M: TheModule, GetBFICallback: nullptr, PSI: &PSI); |
| 491 | WriteBitcodeToFile(M: TheModule, Out&: OS, ShouldPreserveUseListOrder: true, Index: &Index); |
| 492 | } |
| 493 | return std::make_unique<SmallVectorMemoryBuffer>( |
| 494 | args: std::move(OutputBuffer), /*RequiresNullTerminator=*/args: false); |
| 495 | } |
| 496 | |
| 497 | return codegenModule(TheModule, TM); |
| 498 | } |
| 499 | |
| 500 | /// Resolve prevailing symbols. Record resolutions in the \p ResolvedODR map |
| 501 | /// for caching, and in the \p Index for application during the ThinLTO |
| 502 | /// backends. This is needed for correctness for exported symbols (ensure |
| 503 | /// at least one copy kept) and a compile-time optimization (to drop duplicate |
| 504 | /// copies when possible). |
| 505 | static void resolvePrevailingInIndex( |
| 506 | ModuleSummaryIndex &Index, |
| 507 | StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> |
| 508 | &ResolvedODR, |
| 509 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols, |
| 510 | const DenseMap<GlobalValue::GUID, const GlobalValueSummary *> |
| 511 | &PrevailingCopy) { |
| 512 | |
| 513 | auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) { |
| 514 | const auto &Prevailing = PrevailingCopy.find(Val: GUID); |
| 515 | // Not in map means that there was only one copy, which must be prevailing. |
| 516 | if (Prevailing == PrevailingCopy.end()) |
| 517 | return true; |
| 518 | return Prevailing->second == S; |
| 519 | }; |
| 520 | |
| 521 | auto recordNewLinkage = [&](StringRef ModuleIdentifier, |
| 522 | GlobalValue::GUID GUID, |
| 523 | GlobalValue::LinkageTypes NewLinkage) { |
| 524 | ResolvedODR[ModuleIdentifier][GUID] = NewLinkage; |
| 525 | }; |
| 526 | |
| 527 | // TODO Conf.VisibilityScheme can be lto::Config::ELF for ELF. |
| 528 | lto::Config Conf; |
| 529 | thinLTOResolvePrevailingInIndex(C: Conf, Index, isPrevailing, recordNewLinkage, |
| 530 | GUIDPreservedSymbols); |
| 531 | } |
| 532 | |
| 533 | // Initialize the TargetMachine builder for a given Triple |
| 534 | static void initTMBuilder(TargetMachineBuilder &TMBuilder, |
| 535 | const Triple &TheTriple) { |
| 536 | if (TMBuilder.MCpu.empty()) |
| 537 | TMBuilder.MCpu = lto::getThinLTODefaultCPU(TheTriple); |
| 538 | TMBuilder.TheTriple = std::move(TheTriple); |
| 539 | } |
| 540 | |
| 541 | void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) { |
| 542 | MemoryBufferRef Buffer(Data, Identifier); |
| 543 | |
| 544 | auto InputOrError = lto::InputFile::create(Object: Buffer); |
| 545 | if (!InputOrError) |
| 546 | report_fatal_error(reason: Twine("ThinLTO cannot create input file: " ) + |
| 547 | toString(E: InputOrError.takeError())); |
| 548 | |
| 549 | auto TripleStr = (*InputOrError)->getTargetTriple(); |
| 550 | Triple TheTriple(TripleStr); |
| 551 | |
| 552 | if (Modules.empty()) |
| 553 | initTMBuilder(TMBuilder, TheTriple: Triple(TheTriple)); |
| 554 | else if (TMBuilder.TheTriple != TheTriple) { |
| 555 | if (!TMBuilder.TheTriple.isCompatibleWith(Other: TheTriple)) |
| 556 | report_fatal_error(reason: "ThinLTO modules with incompatible triples not " |
| 557 | "supported" ); |
| 558 | initTMBuilder(TMBuilder, TheTriple: Triple(TMBuilder.TheTriple.merge(Other: TheTriple))); |
| 559 | } |
| 560 | |
| 561 | Modules.emplace_back(args: std::move(*InputOrError)); |
| 562 | } |
| 563 | |
| 564 | void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) { |
| 565 | PreservedSymbols.insert(key: Name); |
| 566 | } |
| 567 | |
| 568 | void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) { |
| 569 | // FIXME: At the moment, we don't take advantage of this extra information, |
| 570 | // we're conservatively considering cross-references as preserved. |
| 571 | // CrossReferencedSymbols.insert(Name); |
| 572 | PreservedSymbols.insert(key: Name); |
| 573 | } |
| 574 | |
| 575 | // TargetMachine factory |
| 576 | std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const { |
| 577 | std::string ErrMsg; |
| 578 | const Target *TheTarget = TargetRegistry::lookupTarget(TheTriple, Error&: ErrMsg); |
| 579 | if (!TheTarget) { |
| 580 | report_fatal_error(reason: Twine("Can't load target for this Triple: " ) + ErrMsg); |
| 581 | } |
| 582 | |
| 583 | // Use MAttr as the default set of features. |
| 584 | SubtargetFeatures Features(MAttr); |
| 585 | Features.getDefaultSubtargetFeatures(Triple: TheTriple); |
| 586 | std::string FeatureStr = Features.getString(); |
| 587 | |
| 588 | std::unique_ptr<TargetMachine> TM( |
| 589 | TheTarget->createTargetMachine(TT: TheTriple, CPU: MCpu, Features: FeatureStr, Options, |
| 590 | RM: RelocModel, CM: std::nullopt, OL: CGOptLevel)); |
| 591 | assert(TM && "Cannot create target machine" ); |
| 592 | |
| 593 | return TM; |
| 594 | } |
| 595 | |
| 596 | /** |
| 597 | * Produce the combined summary index from all the bitcode files: |
| 598 | * "thin-link". |
| 599 | */ |
| 600 | std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() { |
| 601 | std::unique_ptr<ModuleSummaryIndex> CombinedIndex = |
| 602 | std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/args: false); |
| 603 | for (auto &Mod : Modules) { |
| 604 | auto &M = Mod->getSingleBitcodeModule(); |
| 605 | if (Error Err = M.readSummary(CombinedIndex&: *CombinedIndex, ModulePath: Mod->getName())) { |
| 606 | // FIXME diagnose |
| 607 | logAllUnhandledErrors( |
| 608 | E: std::move(Err), OS&: errs(), |
| 609 | ErrorBanner: "error: can't create module summary index for buffer: " ); |
| 610 | return nullptr; |
| 611 | } |
| 612 | } |
| 613 | return CombinedIndex; |
| 614 | } |
| 615 | |
| 616 | namespace { |
| 617 | struct IsExported { |
| 618 | const DenseMap<StringRef, FunctionImporter::ExportSetTy> &ExportLists; |
| 619 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols; |
| 620 | |
| 621 | IsExported( |
| 622 | const DenseMap<StringRef, FunctionImporter::ExportSetTy> &ExportLists, |
| 623 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) |
| 624 | : ExportLists(ExportLists), GUIDPreservedSymbols(GUIDPreservedSymbols) {} |
| 625 | |
| 626 | bool operator()(StringRef ModuleIdentifier, ValueInfo VI) const { |
| 627 | const auto &ExportList = ExportLists.find(Val: ModuleIdentifier); |
| 628 | return (ExportList != ExportLists.end() && ExportList->second.count(V: VI)) || |
| 629 | GUIDPreservedSymbols.count(V: VI.getGUID()); |
| 630 | } |
| 631 | }; |
| 632 | |
| 633 | struct IsPrevailing { |
| 634 | const DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy; |
| 635 | IsPrevailing(const DenseMap<GlobalValue::GUID, const GlobalValueSummary *> |
| 636 | &PrevailingCopy) |
| 637 | : PrevailingCopy(PrevailingCopy) {} |
| 638 | |
| 639 | bool operator()(GlobalValue::GUID GUID, const GlobalValueSummary *S) const { |
| 640 | const auto &Prevailing = PrevailingCopy.find(Val: GUID); |
| 641 | // Not in map means that there was only one copy, which must be prevailing. |
| 642 | if (Prevailing == PrevailingCopy.end()) |
| 643 | return true; |
| 644 | return Prevailing->second == S; |
| 645 | }; |
| 646 | }; |
| 647 | } // namespace |
| 648 | |
| 649 | static void computeDeadSymbolsInIndex( |
| 650 | ModuleSummaryIndex &Index, |
| 651 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { |
| 652 | // We have no symbols resolution available. And can't do any better now in the |
| 653 | // case where the prevailing symbol is in a native object. It can be refined |
| 654 | // with linker information in the future. |
| 655 | auto isPrevailing = [&](GlobalValue::GUID G) { |
| 656 | return PrevailingType::Unknown; |
| 657 | }; |
| 658 | computeDeadSymbolsWithConstProp(Index, GUIDPreservedSymbols, isPrevailing, |
| 659 | /* ImportEnabled = */ true); |
| 660 | } |
| 661 | |
| 662 | /** |
| 663 | * Perform promotion and renaming of exported internal functions. |
| 664 | * Index is updated to reflect linkage changes from weak resolution. |
| 665 | */ |
| 666 | void ThinLTOCodeGenerator::promote(Module &TheModule, ModuleSummaryIndex &Index, |
| 667 | const lto::InputFile &File) { |
| 668 | auto ModuleCount = Index.modulePaths().size(); |
| 669 | auto ModuleIdentifier = TheModule.getModuleIdentifier(); |
| 670 | |
| 671 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 672 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries; |
| 673 | Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 674 | |
| 675 | // Convert the preserved symbols set from string to GUID |
| 676 | auto GUIDPreservedSymbols = computeGUIDPreservedSymbols( |
| 677 | File, PreservedSymbols, TheTriple: TheModule.getTargetTriple()); |
| 678 | |
| 679 | // Add used symbol to the preserved symbols. |
| 680 | addUsedSymbolToPreservedGUID(File, PreservedGUID&: GUIDPreservedSymbols); |
| 681 | |
| 682 | // Compute "dead" symbols, we don't want to import/export these! |
| 683 | computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols); |
| 684 | |
| 685 | // Compute prevailing symbols |
| 686 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 687 | computePrevailingCopies(Index, PrevailingCopy); |
| 688 | |
| 689 | // Generate import/export list |
| 690 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 691 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 692 | ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, |
| 693 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 694 | ExportLists); |
| 695 | |
| 696 | // Resolve prevailing symbols |
| 697 | StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; |
| 698 | resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols, |
| 699 | PrevailingCopy); |
| 700 | |
| 701 | thinLTOFinalizeInModule(TheModule, |
| 702 | DefinedGlobals: ModuleToDefinedGVSummaries[ModuleIdentifier], |
| 703 | /*PropagateAttrs=*/false); |
| 704 | |
| 705 | // Promote the exported values in the index, so that they are promoted |
| 706 | // in the module. |
| 707 | thinLTOInternalizeAndPromoteInIndex( |
| 708 | Index, isExported: IsExported(ExportLists, GUIDPreservedSymbols), |
| 709 | isPrevailing: IsPrevailing(PrevailingCopy)); |
| 710 | |
| 711 | // FIXME Set ClearDSOLocalOnDeclarations. |
| 712 | promoteModule(TheModule, Index, /*ClearDSOLocalOnDeclarations=*/false); |
| 713 | } |
| 714 | |
| 715 | /** |
| 716 | * Perform cross-module importing for the module identified by ModuleIdentifier. |
| 717 | */ |
| 718 | void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule, |
| 719 | ModuleSummaryIndex &Index, |
| 720 | const lto::InputFile &File) { |
| 721 | auto ModuleMap = generateModuleMap(Modules); |
| 722 | auto ModuleCount = Index.modulePaths().size(); |
| 723 | |
| 724 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 725 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount); |
| 726 | Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 727 | |
| 728 | // Convert the preserved symbols set from string to GUID |
| 729 | auto GUIDPreservedSymbols = computeGUIDPreservedSymbols( |
| 730 | File, PreservedSymbols, TheTriple: TheModule.getTargetTriple()); |
| 731 | |
| 732 | addUsedSymbolToPreservedGUID(File, PreservedGUID&: GUIDPreservedSymbols); |
| 733 | |
| 734 | // Compute "dead" symbols, we don't want to import/export these! |
| 735 | computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols); |
| 736 | |
| 737 | // Compute prevailing symbols |
| 738 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 739 | computePrevailingCopies(Index, PrevailingCopy); |
| 740 | |
| 741 | // Generate import/export list |
| 742 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 743 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 744 | ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, |
| 745 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 746 | ExportLists); |
| 747 | auto &ImportList = ImportLists[TheModule.getModuleIdentifier()]; |
| 748 | |
| 749 | // FIXME Set ClearDSOLocalOnDeclarations. |
| 750 | crossImportIntoModule(TheModule, Index, ModuleMap, ImportList, |
| 751 | /*ClearDSOLocalOnDeclarations=*/false); |
| 752 | } |
| 753 | |
| 754 | /** |
| 755 | * Compute the list of summaries needed for importing into module. |
| 756 | */ |
| 757 | void ThinLTOCodeGenerator::gatherImportedSummariesForModule( |
| 758 | Module &TheModule, ModuleSummaryIndex &Index, |
| 759 | ModuleToSummariesForIndexTy &ModuleToSummariesForIndex, |
| 760 | GVSummaryPtrSet &DecSummaries, const lto::InputFile &File) { |
| 761 | auto ModuleCount = Index.modulePaths().size(); |
| 762 | auto ModuleIdentifier = TheModule.getModuleIdentifier(); |
| 763 | |
| 764 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 765 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount); |
| 766 | Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 767 | |
| 768 | // Convert the preserved symbols set from string to GUID |
| 769 | auto GUIDPreservedSymbols = computeGUIDPreservedSymbols( |
| 770 | File, PreservedSymbols, TheTriple: TheModule.getTargetTriple()); |
| 771 | |
| 772 | addUsedSymbolToPreservedGUID(File, PreservedGUID&: GUIDPreservedSymbols); |
| 773 | |
| 774 | // Compute "dead" symbols, we don't want to import/export these! |
| 775 | computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols); |
| 776 | |
| 777 | // Compute prevailing symbols |
| 778 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 779 | computePrevailingCopies(Index, PrevailingCopy); |
| 780 | |
| 781 | // Generate import/export list |
| 782 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 783 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 784 | ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, |
| 785 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 786 | ExportLists); |
| 787 | |
| 788 | llvm::gatherImportedSummariesForModule( |
| 789 | ModulePath: ModuleIdentifier, ModuleToDefinedGVSummaries, |
| 790 | ImportList: ImportLists[ModuleIdentifier], ModuleToSummariesForIndex, DecSummaries); |
| 791 | } |
| 792 | |
| 793 | /** |
| 794 | * Emit the list of files needed for importing into module. |
| 795 | */ |
| 796 | void ThinLTOCodeGenerator::emitImports(Module &TheModule, StringRef OutputName, |
| 797 | ModuleSummaryIndex &Index, |
| 798 | const lto::InputFile &File) { |
| 799 | auto ModuleCount = Index.modulePaths().size(); |
| 800 | auto ModuleIdentifier = TheModule.getModuleIdentifier(); |
| 801 | |
| 802 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 803 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount); |
| 804 | Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 805 | |
| 806 | // Convert the preserved symbols set from string to GUID |
| 807 | auto GUIDPreservedSymbols = computeGUIDPreservedSymbols( |
| 808 | File, PreservedSymbols, TheTriple: TheModule.getTargetTriple()); |
| 809 | |
| 810 | addUsedSymbolToPreservedGUID(File, PreservedGUID&: GUIDPreservedSymbols); |
| 811 | |
| 812 | // Compute "dead" symbols, we don't want to import/export these! |
| 813 | computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols); |
| 814 | |
| 815 | // Compute prevailing symbols |
| 816 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 817 | computePrevailingCopies(Index, PrevailingCopy); |
| 818 | |
| 819 | // Generate import/export list |
| 820 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 821 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 822 | ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, |
| 823 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 824 | ExportLists); |
| 825 | |
| 826 | // 'EmitImportsFiles' emits the list of modules from which to import from, and |
| 827 | // the set of keys in `ModuleToSummariesForIndex` should be a superset of keys |
| 828 | // in `DecSummaries`, so no need to use `DecSummaries` in `EmitImportsFiles`. |
| 829 | GVSummaryPtrSet DecSummaries; |
| 830 | ModuleToSummariesForIndexTy ModuleToSummariesForIndex; |
| 831 | llvm::gatherImportedSummariesForModule( |
| 832 | ModulePath: ModuleIdentifier, ModuleToDefinedGVSummaries, |
| 833 | ImportList: ImportLists[ModuleIdentifier], ModuleToSummariesForIndex, DecSummaries); |
| 834 | |
| 835 | if (Error EC = EmitImportsFiles(ModulePath: ModuleIdentifier, OutputFilename: OutputName, |
| 836 | ModuleToSummariesForIndex)) |
| 837 | report_fatal_error(reason: Twine("Failed to open " ) + OutputName + |
| 838 | " to save imports lists\n" ); |
| 839 | } |
| 840 | |
| 841 | /** |
| 842 | * Perform internalization. Runs promote and internalization together. |
| 843 | * Index is updated to reflect linkage changes. |
| 844 | */ |
| 845 | void ThinLTOCodeGenerator::internalize(Module &TheModule, |
| 846 | ModuleSummaryIndex &Index, |
| 847 | const lto::InputFile &File) { |
| 848 | initTMBuilder(TMBuilder, TheTriple: TheModule.getTargetTriple()); |
| 849 | auto ModuleCount = Index.modulePaths().size(); |
| 850 | auto ModuleIdentifier = TheModule.getModuleIdentifier(); |
| 851 | |
| 852 | // Convert the preserved symbols set from string to GUID |
| 853 | auto GUIDPreservedSymbols = |
| 854 | computeGUIDPreservedSymbols(File, PreservedSymbols, TheTriple: TMBuilder.TheTriple); |
| 855 | |
| 856 | addUsedSymbolToPreservedGUID(File, PreservedGUID&: GUIDPreservedSymbols); |
| 857 | |
| 858 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 859 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount); |
| 860 | Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 861 | |
| 862 | // Compute "dead" symbols, we don't want to import/export these! |
| 863 | computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols); |
| 864 | |
| 865 | // Compute prevailing symbols |
| 866 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 867 | computePrevailingCopies(Index, PrevailingCopy); |
| 868 | |
| 869 | // Generate import/export list |
| 870 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 871 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 872 | ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, |
| 873 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 874 | ExportLists); |
| 875 | auto &ExportList = ExportLists[ModuleIdentifier]; |
| 876 | |
| 877 | // Be friendly and don't nuke totally the module when the client didn't |
| 878 | // supply anything to preserve. |
| 879 | if (ExportList.empty() && GUIDPreservedSymbols.empty()) |
| 880 | return; |
| 881 | |
| 882 | // Resolve prevailing symbols |
| 883 | StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; |
| 884 | resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols, |
| 885 | PrevailingCopy); |
| 886 | |
| 887 | // Promote the exported values in the index, so that they are promoted |
| 888 | // in the module. |
| 889 | thinLTOInternalizeAndPromoteInIndex( |
| 890 | Index, isExported: IsExported(ExportLists, GUIDPreservedSymbols), |
| 891 | isPrevailing: IsPrevailing(PrevailingCopy)); |
| 892 | |
| 893 | // FIXME Set ClearDSOLocalOnDeclarations. |
| 894 | promoteModule(TheModule, Index, /*ClearDSOLocalOnDeclarations=*/false); |
| 895 | |
| 896 | // Internalization |
| 897 | thinLTOFinalizeInModule(TheModule, |
| 898 | DefinedGlobals: ModuleToDefinedGVSummaries[ModuleIdentifier], |
| 899 | /*PropagateAttrs=*/false); |
| 900 | |
| 901 | thinLTOInternalizeModule(TheModule, |
| 902 | DefinedGlobals: ModuleToDefinedGVSummaries[ModuleIdentifier]); |
| 903 | } |
| 904 | |
| 905 | /** |
| 906 | * Perform post-importing ThinLTO optimizations. |
| 907 | */ |
| 908 | void ThinLTOCodeGenerator::optimize(Module &TheModule) { |
| 909 | initTMBuilder(TMBuilder, TheTriple: TheModule.getTargetTriple()); |
| 910 | |
| 911 | // Optimize now |
| 912 | optimizeModule(TheModule, TM&: *TMBuilder.create(), OptLevel, Freestanding, |
| 913 | DebugPassManager, Index: nullptr); |
| 914 | } |
| 915 | |
| 916 | /// Write out the generated object file, either from CacheEntryPath or from |
| 917 | /// OutputBuffer, preferring hard-link when possible. |
| 918 | /// Returns the path to the generated file in SavedObjectsDirectoryPath. |
| 919 | std::string |
| 920 | ThinLTOCodeGenerator::writeGeneratedObject(int count, StringRef CacheEntryPath, |
| 921 | const MemoryBuffer &OutputBuffer) { |
| 922 | auto ArchName = TMBuilder.TheTriple.getArchName(); |
| 923 | SmallString<128> OutputPath(SavedObjectsDirectoryPath); |
| 924 | llvm::sys::path::append(path&: OutputPath, |
| 925 | a: Twine(count) + "." + ArchName + ".thinlto.o" ); |
| 926 | OutputPath.c_str(); // Ensure the string is null terminated. |
| 927 | if (sys::fs::exists(Path: OutputPath)) |
| 928 | sys::fs::remove(path: OutputPath); |
| 929 | |
| 930 | // We don't return a memory buffer to the linker, just a list of files. |
| 931 | if (!CacheEntryPath.empty()) { |
| 932 | // Cache is enabled, hard-link the entry (or copy if hard-link fails). |
| 933 | auto Err = sys::fs::create_hard_link(to: CacheEntryPath, from: OutputPath); |
| 934 | if (!Err) |
| 935 | return std::string(OutputPath); |
| 936 | // Hard linking failed, try to copy. |
| 937 | Err = sys::fs::copy_file(From: CacheEntryPath, To: OutputPath); |
| 938 | if (!Err) |
| 939 | return std::string(OutputPath); |
| 940 | // Copy failed (could be because the CacheEntry was removed from the cache |
| 941 | // in the meantime by another process), fall back and try to write down the |
| 942 | // buffer to the output. |
| 943 | errs() << "remark: can't link or copy from cached entry '" << CacheEntryPath |
| 944 | << "' to '" << OutputPath << "'\n" ; |
| 945 | } |
| 946 | // No cache entry, just write out the buffer. |
| 947 | std::error_code Err; |
| 948 | raw_fd_ostream OS(OutputPath, Err, sys::fs::OF_None); |
| 949 | if (Err) |
| 950 | report_fatal_error(reason: Twine("Can't open output '" ) + OutputPath + "'\n" ); |
| 951 | OS << OutputBuffer.getBuffer(); |
| 952 | return std::string(OutputPath); |
| 953 | } |
| 954 | |
| 955 | // Main entry point for the ThinLTO processing |
| 956 | void ThinLTOCodeGenerator::run() { |
| 957 | timeTraceProfilerBegin(Name: "ThinLink" , Detail: StringRef("" )); |
| 958 | auto TimeTraceScopeExit = llvm::make_scope_exit(F: []() { |
| 959 | if (llvm::timeTraceProfilerEnabled()) |
| 960 | llvm::timeTraceProfilerEnd(); |
| 961 | }); |
| 962 | // Prepare the resulting object vector |
| 963 | assert(ProducedBinaries.empty() && "The generator should not be reused" ); |
| 964 | if (SavedObjectsDirectoryPath.empty()) |
| 965 | ProducedBinaries.resize(new_size: Modules.size()); |
| 966 | else { |
| 967 | sys::fs::create_directories(path: SavedObjectsDirectoryPath); |
| 968 | bool IsDir; |
| 969 | sys::fs::is_directory(path: SavedObjectsDirectoryPath, result&: IsDir); |
| 970 | if (!IsDir) |
| 971 | report_fatal_error(reason: Twine("Unexistent dir: '" ) + SavedObjectsDirectoryPath + "'" ); |
| 972 | ProducedBinaryFiles.resize(new_size: Modules.size()); |
| 973 | } |
| 974 | |
| 975 | if (CodeGenOnly) { |
| 976 | // Perform only parallel codegen and return. |
| 977 | DefaultThreadPool Pool; |
| 978 | int count = 0; |
| 979 | for (auto &Mod : Modules) { |
| 980 | Pool.async(F: [&](int count) { |
| 981 | LLVMContext Context; |
| 982 | Context.setDiscardValueNames(LTODiscardValueNames); |
| 983 | |
| 984 | // Parse module now |
| 985 | auto TheModule = loadModuleFromInput(Input: Mod.get(), Context, Lazy: false, |
| 986 | /*IsImporting*/ false); |
| 987 | |
| 988 | // CodeGen |
| 989 | auto OutputBuffer = codegenModule(TheModule&: *TheModule, TM&: *TMBuilder.create()); |
| 990 | if (SavedObjectsDirectoryPath.empty()) |
| 991 | ProducedBinaries[count] = std::move(OutputBuffer); |
| 992 | else |
| 993 | ProducedBinaryFiles[count] = |
| 994 | writeGeneratedObject(count, CacheEntryPath: "" , OutputBuffer: *OutputBuffer); |
| 995 | }, ArgList: count++); |
| 996 | } |
| 997 | |
| 998 | return; |
| 999 | } |
| 1000 | |
| 1001 | // Sequential linking phase |
| 1002 | auto Index = linkCombinedIndex(); |
| 1003 | |
| 1004 | // Save temps: index. |
| 1005 | if (!SaveTempsDir.empty()) { |
| 1006 | auto SaveTempPath = SaveTempsDir + "index.bc" ; |
| 1007 | std::error_code EC; |
| 1008 | raw_fd_ostream OS(SaveTempPath, EC, sys::fs::OF_None); |
| 1009 | if (EC) |
| 1010 | report_fatal_error(reason: Twine("Failed to open " ) + SaveTempPath + |
| 1011 | " to save optimized bitcode\n" ); |
| 1012 | writeIndexToFile(Index: *Index, Out&: OS); |
| 1013 | } |
| 1014 | |
| 1015 | |
| 1016 | // Prepare the module map. |
| 1017 | auto ModuleMap = generateModuleMap(Modules); |
| 1018 | auto ModuleCount = Modules.size(); |
| 1019 | |
| 1020 | // Collect for each module the list of function it defines (GUID -> Summary). |
| 1021 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount); |
| 1022 | Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries); |
| 1023 | |
| 1024 | // Convert the preserved symbols set from string to GUID, this is needed for |
| 1025 | // computing the caching hash and the internalization. |
| 1026 | DenseSet<GlobalValue::GUID> GUIDPreservedSymbols; |
| 1027 | for (const auto &M : Modules) |
| 1028 | computeGUIDPreservedSymbols(File: *M, PreservedSymbols, TheTriple: TMBuilder.TheTriple, |
| 1029 | GUIDs&: GUIDPreservedSymbols); |
| 1030 | |
| 1031 | // Add used symbol from inputs to the preserved symbols. |
| 1032 | for (const auto &M : Modules) |
| 1033 | addUsedSymbolToPreservedGUID(File: *M, PreservedGUID&: GUIDPreservedSymbols); |
| 1034 | |
| 1035 | // Compute "dead" symbols, we don't want to import/export these! |
| 1036 | computeDeadSymbolsInIndex(Index&: *Index, GUIDPreservedSymbols); |
| 1037 | |
| 1038 | // Currently there is no support for enabling whole program visibility via a |
| 1039 | // linker option in the old LTO API, but this call allows it to be specified |
| 1040 | // via the internal option. Must be done before WPD below. |
| 1041 | if (hasWholeProgramVisibility(/* WholeProgramVisibilityEnabledInLTO */ false)) |
| 1042 | Index->setWithWholeProgramVisibility(); |
| 1043 | |
| 1044 | // FIXME: This needs linker information via a TBD new interface |
| 1045 | updateVCallVisibilityInIndex(Index&: *Index, |
| 1046 | /*WholeProgramVisibilityEnabledInLTO=*/false, |
| 1047 | // FIXME: These need linker information via a |
| 1048 | // TBD new interface. |
| 1049 | /*DynamicExportSymbols=*/{}, |
| 1050 | /*VisibleToRegularObjSymbols=*/{}); |
| 1051 | |
| 1052 | // Perform index-based WPD. This will return immediately if there are |
| 1053 | // no index entries in the typeIdMetadata map (e.g. if we are instead |
| 1054 | // performing IR-based WPD in hybrid regular/thin LTO mode). |
| 1055 | std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap; |
| 1056 | std::set<GlobalValue::GUID> ExportedGUIDs; |
| 1057 | runWholeProgramDevirtOnIndex(Summary&: *Index, ExportedGUIDs, LocalWPDTargetsMap); |
| 1058 | GUIDPreservedSymbols.insert_range(R&: ExportedGUIDs); |
| 1059 | |
| 1060 | // Compute prevailing symbols |
| 1061 | DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy; |
| 1062 | computePrevailingCopies(Index: *Index, PrevailingCopy); |
| 1063 | |
| 1064 | // Collect the import/export lists for all modules from the call-graph in the |
| 1065 | // combined index. |
| 1066 | FunctionImporter::ImportListsTy ImportLists(ModuleCount); |
| 1067 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists(ModuleCount); |
| 1068 | ComputeCrossModuleImport(Index: *Index, ModuleToDefinedGVSummaries, |
| 1069 | isPrevailing: IsPrevailing(PrevailingCopy), ImportLists, |
| 1070 | ExportLists); |
| 1071 | |
| 1072 | // We use a std::map here to be able to have a defined ordering when |
| 1073 | // producing a hash for the cache entry. |
| 1074 | // FIXME: we should be able to compute the caching hash for the entry based |
| 1075 | // on the index, and nuke this map. |
| 1076 | StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; |
| 1077 | |
| 1078 | // Resolve prevailing symbols, this has to be computed early because it |
| 1079 | // impacts the caching. |
| 1080 | resolvePrevailingInIndex(Index&: *Index, ResolvedODR, GUIDPreservedSymbols, |
| 1081 | PrevailingCopy); |
| 1082 | |
| 1083 | // Use global summary-based analysis to identify symbols that can be |
| 1084 | // internalized (because they aren't exported or preserved as per callback). |
| 1085 | // Changes are made in the index, consumed in the ThinLTO backends. |
| 1086 | updateIndexWPDForExports(Summary&: *Index, |
| 1087 | isExported: IsExported(ExportLists, GUIDPreservedSymbols), |
| 1088 | LocalWPDTargetsMap); |
| 1089 | thinLTOInternalizeAndPromoteInIndex( |
| 1090 | Index&: *Index, isExported: IsExported(ExportLists, GUIDPreservedSymbols), |
| 1091 | isPrevailing: IsPrevailing(PrevailingCopy)); |
| 1092 | |
| 1093 | thinLTOPropagateFunctionAttrs(Index&: *Index, isPrevailing: IsPrevailing(PrevailingCopy)); |
| 1094 | |
| 1095 | // Make sure that every module has an entry in the ExportLists, ImportList, |
| 1096 | // GVSummary and ResolvedODR maps to enable threaded access to these maps |
| 1097 | // below. |
| 1098 | for (auto &Module : Modules) { |
| 1099 | auto ModuleIdentifier = Module->getName(); |
| 1100 | ExportLists[ModuleIdentifier]; |
| 1101 | ImportLists[ModuleIdentifier]; |
| 1102 | ResolvedODR[ModuleIdentifier]; |
| 1103 | ModuleToDefinedGVSummaries[ModuleIdentifier]; |
| 1104 | } |
| 1105 | |
| 1106 | std::vector<BitcodeModule *> ModulesVec; |
| 1107 | ModulesVec.reserve(n: Modules.size()); |
| 1108 | for (auto &Mod : Modules) |
| 1109 | ModulesVec.push_back(x: &Mod->getSingleBitcodeModule()); |
| 1110 | std::vector<int> ModulesOrdering = lto::generateModulesOrdering(R: ModulesVec); |
| 1111 | |
| 1112 | if (llvm::timeTraceProfilerEnabled()) |
| 1113 | llvm::timeTraceProfilerEnd(); |
| 1114 | |
| 1115 | TimeTraceScopeExit.release(); |
| 1116 | |
| 1117 | // Parallel optimizer + codegen |
| 1118 | { |
| 1119 | DefaultThreadPool Pool(heavyweight_hardware_concurrency(ThreadCount)); |
| 1120 | for (auto IndexCount : ModulesOrdering) { |
| 1121 | auto &Mod = Modules[IndexCount]; |
| 1122 | Pool.async(F: [&](int count) { |
| 1123 | auto ModuleIdentifier = Mod->getName(); |
| 1124 | auto &ExportList = ExportLists[ModuleIdentifier]; |
| 1125 | |
| 1126 | auto &DefinedGVSummaries = ModuleToDefinedGVSummaries[ModuleIdentifier]; |
| 1127 | |
| 1128 | // The module may be cached, this helps handling it. |
| 1129 | ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier, |
| 1130 | ImportLists[ModuleIdentifier], ExportList, |
| 1131 | ResolvedODR[ModuleIdentifier], |
| 1132 | DefinedGVSummaries, OptLevel, Freestanding, |
| 1133 | TMBuilder); |
| 1134 | auto CacheEntryPath = CacheEntry.getEntryPath(); |
| 1135 | |
| 1136 | { |
| 1137 | auto ErrOrBuffer = CacheEntry.tryLoadingBuffer(); |
| 1138 | LLVM_DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss" ) |
| 1139 | << " '" << CacheEntryPath << "' for buffer " |
| 1140 | << count << " " << ModuleIdentifier << "\n" ); |
| 1141 | |
| 1142 | if (ErrOrBuffer) { |
| 1143 | // Cache Hit! |
| 1144 | if (SavedObjectsDirectoryPath.empty()) |
| 1145 | ProducedBinaries[count] = std::move(ErrOrBuffer.get()); |
| 1146 | else |
| 1147 | ProducedBinaryFiles[count] = writeGeneratedObject( |
| 1148 | count, CacheEntryPath, OutputBuffer: *ErrOrBuffer.get()); |
| 1149 | return; |
| 1150 | } |
| 1151 | } |
| 1152 | |
| 1153 | LLVMContext Context; |
| 1154 | Context.setDiscardValueNames(LTODiscardValueNames); |
| 1155 | Context.enableDebugTypeODRUniquing(); |
| 1156 | auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks( |
| 1157 | Context, RemarksFilename, RemarksPasses, RemarksFormat, |
| 1158 | RemarksWithHotness, RemarksHotnessThreshold, Count: count); |
| 1159 | if (!DiagFileOrErr) { |
| 1160 | errs() << "Error: " << toString(E: DiagFileOrErr.takeError()) << "\n" ; |
| 1161 | report_fatal_error(reason: "ThinLTO: Can't get an output file for the " |
| 1162 | "remarks" ); |
| 1163 | } |
| 1164 | |
| 1165 | // Parse module now |
| 1166 | auto TheModule = loadModuleFromInput(Input: Mod.get(), Context, Lazy: false, |
| 1167 | /*IsImporting*/ false); |
| 1168 | |
| 1169 | // Save temps: original file. |
| 1170 | saveTempBitcode(TheModule: *TheModule, TempDir: SaveTempsDir, count, Suffix: ".0.original.bc" ); |
| 1171 | |
| 1172 | auto &ImportList = ImportLists[ModuleIdentifier]; |
| 1173 | // Run the main process now, and generates a binary |
| 1174 | auto OutputBuffer = ProcessThinLTOModule( |
| 1175 | TheModule&: *TheModule, Index&: *Index, ModuleMap, TM&: *TMBuilder.create(), ImportList, |
| 1176 | ExportList, GUIDPreservedSymbols, |
| 1177 | DefinedGlobals: ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions, |
| 1178 | DisableCodeGen, SaveTempsDir, Freestanding, OptLevel, count, |
| 1179 | DebugPassManager); |
| 1180 | |
| 1181 | // Commit to the cache (if enabled) |
| 1182 | CacheEntry.write(OutputBuffer: *OutputBuffer); |
| 1183 | |
| 1184 | if (SavedObjectsDirectoryPath.empty()) { |
| 1185 | // We need to generated a memory buffer for the linker. |
| 1186 | if (!CacheEntryPath.empty()) { |
| 1187 | // When cache is enabled, reload from the cache if possible. |
| 1188 | // Releasing the buffer from the heap and reloading it from the |
| 1189 | // cache file with mmap helps us to lower memory pressure. |
| 1190 | // The freed memory can be used for the next input file. |
| 1191 | // The final binary link will read from the VFS cache (hopefully!) |
| 1192 | // or from disk (if the memory pressure was too high). |
| 1193 | auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer(); |
| 1194 | if (auto EC = ReloadedBufferOrErr.getError()) { |
| 1195 | // On error, keep the preexisting buffer and print a diagnostic. |
| 1196 | errs() << "remark: can't reload cached file '" << CacheEntryPath |
| 1197 | << "': " << EC.message() << "\n" ; |
| 1198 | } else { |
| 1199 | OutputBuffer = std::move(*ReloadedBufferOrErr); |
| 1200 | } |
| 1201 | } |
| 1202 | ProducedBinaries[count] = std::move(OutputBuffer); |
| 1203 | return; |
| 1204 | } |
| 1205 | ProducedBinaryFiles[count] = writeGeneratedObject( |
| 1206 | count, CacheEntryPath, OutputBuffer: *OutputBuffer); |
| 1207 | }, ArgList&: IndexCount); |
| 1208 | } |
| 1209 | } |
| 1210 | |
| 1211 | pruneCache(Path: CacheOptions.Path, Policy: CacheOptions.Policy, Files: ProducedBinaries); |
| 1212 | |
| 1213 | // If statistics were requested, print them out now. |
| 1214 | if (llvm::AreStatisticsEnabled()) |
| 1215 | llvm::PrintStatistics(); |
| 1216 | reportAndResetTimings(); |
| 1217 | } |
| 1218 | |