1 //===-LTOBackend.cpp - LLVM Link Time Optimizer Backend -------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the "backend" phase of LTO, i.e. it performs 11 // optimization and code generation on a loaded module. It is generally used 12 // internally by the LTO class but can also be used independently, for example 13 // to implement a standalone ThinLTO backend. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "llvm/LTO/LTOBackend.h" 18 #include "llvm/Analysis/AliasAnalysis.h" 19 #include "llvm/Analysis/CGSCCPassManager.h" 20 #include "llvm/Analysis/TargetLibraryInfo.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/Bitcode/BitcodeReader.h" 23 #include "llvm/Bitcode/BitcodeWriter.h" 24 #include "llvm/IR/LegacyPassManager.h" 25 #include "llvm/IR/PassManager.h" 26 #include "llvm/IR/Verifier.h" 27 #include "llvm/LTO/LTO.h" 28 #include "llvm/MC/SubtargetFeature.h" 29 #include "llvm/Object/ModuleSymbolTable.h" 30 #include "llvm/Passes/PassBuilder.h" 31 #include "llvm/Support/Error.h" 32 #include "llvm/Support/FileSystem.h" 33 #include "llvm/Support/MemoryBuffer.h" 34 #include "llvm/Support/Path.h" 35 #include "llvm/Support/Program.h" 36 #include "llvm/Support/raw_ostream.h" 37 #include "llvm/Support/TargetRegistry.h" 38 #include "llvm/Support/ThreadPool.h" 39 #include "llvm/Target/TargetMachine.h" 40 #include "llvm/Transforms/IPO.h" 41 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 42 #include "llvm/Transforms/Scalar/LoopPassManager.h" 43 #include "llvm/Transforms/Utils/FunctionImportUtils.h" 44 #include "llvm/Transforms/Utils/SplitModule.h" 45 46 using namespace llvm; 47 using namespace lto; 48 49 LLVM_ATTRIBUTE_NORETURN static void reportOpenError(StringRef Path, Twine Msg) { 50 errs() << "failed to open " << Path << ": " << Msg << '\n'; 51 errs().flush(); 52 exit(1); 53 } 54 55 Error Config::addSaveTemps(std::string OutputFileName, 56 bool UseInputModulePath) { 57 ShouldDiscardValueNames = false; 58 59 std::error_code EC; 60 ResolutionFile = llvm::make_unique<raw_fd_ostream>( 61 OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text); 62 if (EC) 63 return errorCodeToError(EC); 64 65 auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) { 66 // Keep track of the hook provided by the linker, which also needs to run. 67 ModuleHookFn LinkerHook = Hook; 68 Hook = [=](unsigned Task, const Module &M) { 69 // If the linker's hook returned false, we need to pass that result 70 // through. 71 if (LinkerHook && !LinkerHook(Task, M)) 72 return false; 73 74 std::string PathPrefix; 75 // If this is the combined module (not a ThinLTO backend compile) or the 76 // user hasn't requested using the input module's path, emit to a file 77 // named from the provided OutputFileName with the Task ID appended. 78 if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) { 79 PathPrefix = OutputFileName; 80 if (Task != (unsigned)-1) 81 PathPrefix += utostr(Task) + "."; 82 } else 83 PathPrefix = M.getModuleIdentifier() + "."; 84 std::string Path = PathPrefix + PathSuffix + ".bc"; 85 std::error_code EC; 86 raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None); 87 // Because -save-temps is a debugging feature, we report the error 88 // directly and exit. 89 if (EC) 90 reportOpenError(Path, EC.message()); 91 WriteBitcodeToFile(M, OS, /*ShouldPreserveUseListOrder=*/false); 92 return true; 93 }; 94 }; 95 96 setHook("0.preopt", PreOptModuleHook); 97 setHook("1.promote", PostPromoteModuleHook); 98 setHook("2.internalize", PostInternalizeModuleHook); 99 setHook("3.import", PostImportModuleHook); 100 setHook("4.opt", PostOptModuleHook); 101 setHook("5.precodegen", PreCodeGenModuleHook); 102 103 CombinedIndexHook = [=](const ModuleSummaryIndex &Index) { 104 std::string Path = OutputFileName + "index.bc"; 105 std::error_code EC; 106 raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None); 107 // Because -save-temps is a debugging feature, we report the error 108 // directly and exit. 109 if (EC) 110 reportOpenError(Path, EC.message()); 111 WriteIndexToFile(Index, OS); 112 113 Path = OutputFileName + "index.dot"; 114 raw_fd_ostream OSDot(Path, EC, sys::fs::OpenFlags::F_None); 115 if (EC) 116 reportOpenError(Path, EC.message()); 117 Index.exportToDot(OSDot); 118 return true; 119 }; 120 121 return Error::success(); 122 } 123 124 namespace { 125 126 std::unique_ptr<TargetMachine> 127 createTargetMachine(Config &Conf, const Target *TheTarget, Module &M) { 128 StringRef TheTriple = M.getTargetTriple(); 129 SubtargetFeatures Features; 130 Features.getDefaultSubtargetFeatures(Triple(TheTriple)); 131 for (const std::string &A : Conf.MAttrs) 132 Features.AddFeature(A); 133 134 Reloc::Model RelocModel; 135 if (Conf.RelocModel) 136 RelocModel = *Conf.RelocModel; 137 else 138 RelocModel = 139 M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_; 140 141 return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine( 142 TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel, 143 Conf.CodeModel, Conf.CGOptLevel)); 144 } 145 146 static void runNewPMPasses(Config &Conf, Module &Mod, TargetMachine *TM, 147 unsigned OptLevel, bool IsThinLTO, 148 ModuleSummaryIndex *ExportSummary, 149 const ModuleSummaryIndex *ImportSummary) { 150 Optional<PGOOptions> PGOOpt; 151 if (!Conf.SampleProfile.empty()) 152 PGOOpt = PGOOptions("", "", Conf.SampleProfile, false, true); 153 154 PassBuilder PB(TM, PGOOpt); 155 AAManager AA; 156 157 // Parse a custom AA pipeline if asked to. 158 if (!PB.parseAAPipeline(AA, "default")) 159 report_fatal_error("Error parsing default AA pipeline"); 160 161 LoopAnalysisManager LAM(Conf.DebugPassManager); 162 FunctionAnalysisManager FAM(Conf.DebugPassManager); 163 CGSCCAnalysisManager CGAM(Conf.DebugPassManager); 164 ModuleAnalysisManager MAM(Conf.DebugPassManager); 165 166 // Register the AA manager first so that our version is the one used. 167 FAM.registerPass([&] { return std::move(AA); }); 168 169 // Register all the basic analyses with the managers. 170 PB.registerModuleAnalyses(MAM); 171 PB.registerCGSCCAnalyses(CGAM); 172 PB.registerFunctionAnalyses(FAM); 173 PB.registerLoopAnalyses(LAM); 174 PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); 175 176 ModulePassManager MPM(Conf.DebugPassManager); 177 // FIXME (davide): verify the input. 178 179 PassBuilder::OptimizationLevel OL; 180 181 switch (OptLevel) { 182 default: 183 llvm_unreachable("Invalid optimization level"); 184 case 0: 185 OL = PassBuilder::O0; 186 break; 187 case 1: 188 OL = PassBuilder::O1; 189 break; 190 case 2: 191 OL = PassBuilder::O2; 192 break; 193 case 3: 194 OL = PassBuilder::O3; 195 break; 196 } 197 198 if (IsThinLTO) 199 MPM = PB.buildThinLTODefaultPipeline(OL, Conf.DebugPassManager, 200 ImportSummary); 201 else 202 MPM = PB.buildLTODefaultPipeline(OL, Conf.DebugPassManager, ExportSummary); 203 MPM.run(Mod, MAM); 204 205 // FIXME (davide): verify the output. 206 } 207 208 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM, 209 std::string PipelineDesc, 210 std::string AAPipelineDesc, 211 bool DisableVerify) { 212 PassBuilder PB(TM); 213 AAManager AA; 214 215 // Parse a custom AA pipeline if asked to. 216 if (!AAPipelineDesc.empty()) 217 if (!PB.parseAAPipeline(AA, AAPipelineDesc)) 218 report_fatal_error("unable to parse AA pipeline description: " + 219 AAPipelineDesc); 220 221 LoopAnalysisManager LAM; 222 FunctionAnalysisManager FAM; 223 CGSCCAnalysisManager CGAM; 224 ModuleAnalysisManager MAM; 225 226 // Register the AA manager first so that our version is the one used. 227 FAM.registerPass([&] { return std::move(AA); }); 228 229 // Register all the basic analyses with the managers. 230 PB.registerModuleAnalyses(MAM); 231 PB.registerCGSCCAnalyses(CGAM); 232 PB.registerFunctionAnalyses(FAM); 233 PB.registerLoopAnalyses(LAM); 234 PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); 235 236 ModulePassManager MPM; 237 238 // Always verify the input. 239 MPM.addPass(VerifierPass()); 240 241 // Now, add all the passes we've been requested to. 242 if (!PB.parsePassPipeline(MPM, PipelineDesc)) 243 report_fatal_error("unable to parse pass pipeline description: " + 244 PipelineDesc); 245 246 if (!DisableVerify) 247 MPM.addPass(VerifierPass()); 248 MPM.run(Mod, MAM); 249 } 250 251 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM, 252 bool IsThinLTO, ModuleSummaryIndex *ExportSummary, 253 const ModuleSummaryIndex *ImportSummary) { 254 legacy::PassManager passes; 255 passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis())); 256 257 PassManagerBuilder PMB; 258 PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple())); 259 PMB.Inliner = createFunctionInliningPass(); 260 PMB.ExportSummary = ExportSummary; 261 PMB.ImportSummary = ImportSummary; 262 // Unconditionally verify input since it is not verified before this 263 // point and has unknown origin. 264 PMB.VerifyInput = true; 265 PMB.VerifyOutput = !Conf.DisableVerify; 266 PMB.LoopVectorize = true; 267 PMB.SLPVectorize = true; 268 PMB.OptLevel = Conf.OptLevel; 269 PMB.PGOSampleUse = Conf.SampleProfile; 270 if (IsThinLTO) 271 PMB.populateThinLTOPassManager(passes); 272 else 273 PMB.populateLTOPassManager(passes); 274 passes.run(Mod); 275 } 276 277 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod, 278 bool IsThinLTO, ModuleSummaryIndex *ExportSummary, 279 const ModuleSummaryIndex *ImportSummary) { 280 // FIXME: Plumb the combined index into the new pass manager. 281 if (!Conf.OptPipeline.empty()) 282 runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline, 283 Conf.DisableVerify); 284 else if (Conf.UseNewPM) 285 runNewPMPasses(Conf, Mod, TM, Conf.OptLevel, IsThinLTO, ExportSummary, 286 ImportSummary); 287 else 288 runOldPMPasses(Conf, Mod, TM, IsThinLTO, ExportSummary, ImportSummary); 289 return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod); 290 } 291 292 void codegen(Config &Conf, TargetMachine *TM, AddStreamFn AddStream, 293 unsigned Task, Module &Mod) { 294 if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod)) 295 return; 296 297 std::unique_ptr<ToolOutputFile> DwoOut; 298 SmallString<1024> DwoFile(Conf.DwoPath); 299 if (!Conf.DwoDir.empty()) { 300 std::error_code EC; 301 if (auto EC = llvm::sys::fs::create_directories(Conf.DwoDir)) 302 report_fatal_error("Failed to create directory " + Conf.DwoDir + ": " + 303 EC.message()); 304 305 DwoFile = Conf.DwoDir; 306 sys::path::append(DwoFile, std::to_string(Task) + ".dwo"); 307 } 308 309 if (!DwoFile.empty()) { 310 std::error_code EC; 311 TM->Options.MCOptions.SplitDwarfFile = DwoFile.str().str(); 312 DwoOut = llvm::make_unique<ToolOutputFile>(DwoFile, EC, sys::fs::F_None); 313 if (EC) 314 report_fatal_error("Failed to open " + DwoFile + ": " + EC.message()); 315 } 316 317 auto Stream = AddStream(Task); 318 legacy::PassManager CodeGenPasses; 319 if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS, 320 DwoOut ? &DwoOut->os() : nullptr, 321 Conf.CGFileType)) 322 report_fatal_error("Failed to setup codegen"); 323 CodeGenPasses.run(Mod); 324 325 if (DwoOut) 326 DwoOut->keep(); 327 } 328 329 void splitCodeGen(Config &C, TargetMachine *TM, AddStreamFn AddStream, 330 unsigned ParallelCodeGenParallelismLevel, 331 std::unique_ptr<Module> Mod) { 332 ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel); 333 unsigned ThreadCount = 0; 334 const Target *T = &TM->getTarget(); 335 336 SplitModule( 337 std::move(Mod), ParallelCodeGenParallelismLevel, 338 [&](std::unique_ptr<Module> MPart) { 339 // We want to clone the module in a new context to multi-thread the 340 // codegen. We do it by serializing partition modules to bitcode 341 // (while still on the main thread, in order to avoid data races) and 342 // spinning up new threads which deserialize the partitions into 343 // separate contexts. 344 // FIXME: Provide a more direct way to do this in LLVM. 345 SmallString<0> BC; 346 raw_svector_ostream BCOS(BC); 347 WriteBitcodeToFile(*MPart, BCOS); 348 349 // Enqueue the task 350 CodegenThreadPool.async( 351 [&](const SmallString<0> &BC, unsigned ThreadId) { 352 LTOLLVMContext Ctx(C); 353 Expected<std::unique_ptr<Module>> MOrErr = parseBitcodeFile( 354 MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"), 355 Ctx); 356 if (!MOrErr) 357 report_fatal_error("Failed to read bitcode"); 358 std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get()); 359 360 std::unique_ptr<TargetMachine> TM = 361 createTargetMachine(C, T, *MPartInCtx); 362 363 codegen(C, TM.get(), AddStream, ThreadId, *MPartInCtx); 364 }, 365 // Pass BC using std::move to ensure that it get moved rather than 366 // copied into the thread's context. 367 std::move(BC), ThreadCount++); 368 }, 369 false); 370 371 // Because the inner lambda (which runs in a worker thread) captures our local 372 // variables, we need to wait for the worker threads to terminate before we 373 // can leave the function scope. 374 CodegenThreadPool.wait(); 375 } 376 377 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) { 378 if (!C.OverrideTriple.empty()) 379 Mod.setTargetTriple(C.OverrideTriple); 380 else if (Mod.getTargetTriple().empty()) 381 Mod.setTargetTriple(C.DefaultTriple); 382 383 std::string Msg; 384 const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg); 385 if (!T) 386 return make_error<StringError>(Msg, inconvertibleErrorCode()); 387 return T; 388 } 389 390 } 391 392 static Error 393 finalizeOptimizationRemarks(std::unique_ptr<ToolOutputFile> DiagOutputFile) { 394 // Make sure we flush the diagnostic remarks file in case the linker doesn't 395 // call the global destructors before exiting. 396 if (!DiagOutputFile) 397 return Error::success(); 398 DiagOutputFile->keep(); 399 DiagOutputFile->os().flush(); 400 return Error::success(); 401 } 402 403 Error lto::backend(Config &C, AddStreamFn AddStream, 404 unsigned ParallelCodeGenParallelismLevel, 405 std::unique_ptr<Module> Mod, 406 ModuleSummaryIndex &CombinedIndex) { 407 Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod); 408 if (!TOrErr) 409 return TOrErr.takeError(); 410 411 std::unique_ptr<TargetMachine> TM = createTargetMachine(C, *TOrErr, *Mod); 412 413 // Setup optimization remarks. 414 auto DiagFileOrErr = lto::setupOptimizationRemarks( 415 Mod->getContext(), C.RemarksFilename, C.RemarksWithHotness); 416 if (!DiagFileOrErr) 417 return DiagFileOrErr.takeError(); 418 auto DiagnosticOutputFile = std::move(*DiagFileOrErr); 419 420 if (!C.CodeGenOnly) { 421 if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false, 422 /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr)) 423 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 424 } 425 426 if (ParallelCodeGenParallelismLevel == 1) { 427 codegen(C, TM.get(), AddStream, 0, *Mod); 428 } else { 429 splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel, 430 std::move(Mod)); 431 } 432 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 433 } 434 435 static void dropDeadSymbols(Module &Mod, const GVSummaryMapTy &DefinedGlobals, 436 const ModuleSummaryIndex &Index) { 437 std::vector<GlobalValue*> DeadGVs; 438 for (auto &GV : Mod.global_values()) 439 if (GlobalValueSummary *GVS = DefinedGlobals.lookup(GV.getGUID())) 440 if (!Index.isGlobalValueLive(GVS)) { 441 DeadGVs.push_back(&GV); 442 convertToDeclaration(GV); 443 } 444 445 // Now that all dead bodies have been dropped, delete the actual objects 446 // themselves when possible. 447 for (GlobalValue *GV : DeadGVs) { 448 GV->removeDeadConstantUsers(); 449 // Might reference something defined in native object (i.e. dropped a 450 // non-prevailing IR def, but we need to keep the declaration). 451 if (GV->use_empty()) 452 GV->eraseFromParent(); 453 } 454 } 455 456 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream, 457 Module &Mod, const ModuleSummaryIndex &CombinedIndex, 458 const FunctionImporter::ImportMapTy &ImportList, 459 const GVSummaryMapTy &DefinedGlobals, 460 MapVector<StringRef, BitcodeModule> &ModuleMap) { 461 Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod); 462 if (!TOrErr) 463 return TOrErr.takeError(); 464 465 std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, *TOrErr, Mod); 466 467 // Setup optimization remarks. 468 auto DiagFileOrErr = lto::setupOptimizationRemarks( 469 Mod.getContext(), Conf.RemarksFilename, Conf.RemarksWithHotness, Task); 470 if (!DiagFileOrErr) 471 return DiagFileOrErr.takeError(); 472 auto DiagnosticOutputFile = std::move(*DiagFileOrErr); 473 474 if (Conf.CodeGenOnly) { 475 codegen(Conf, TM.get(), AddStream, Task, Mod); 476 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 477 } 478 479 if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod)) 480 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 481 482 renameModuleForThinLTO(Mod, CombinedIndex); 483 484 dropDeadSymbols(Mod, DefinedGlobals, CombinedIndex); 485 486 thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals); 487 488 if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod)) 489 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 490 491 if (!DefinedGlobals.empty()) 492 thinLTOInternalizeModule(Mod, DefinedGlobals); 493 494 if (Conf.PostInternalizeModuleHook && 495 !Conf.PostInternalizeModuleHook(Task, Mod)) 496 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 497 498 auto ModuleLoader = [&](StringRef Identifier) { 499 assert(Mod.getContext().isODRUniquingDebugTypes() && 500 "ODR Type uniquing should be enabled on the context"); 501 auto I = ModuleMap.find(Identifier); 502 assert(I != ModuleMap.end()); 503 return I->second.getLazyModule(Mod.getContext(), 504 /*ShouldLazyLoadMetadata=*/true, 505 /*IsImporting*/ true); 506 }; 507 508 FunctionImporter Importer(CombinedIndex, ModuleLoader); 509 if (Error Err = Importer.importFunctions(Mod, ImportList).takeError()) 510 return Err; 511 512 if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod)) 513 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 514 515 if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true, 516 /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex)) 517 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 518 519 codegen(Conf, TM.get(), AddStream, Task, Mod); 520 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile)); 521 } 522