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 + utostr(Task);
80       } else
81         PathPrefix = M.getModuleIdentifier();
82       std::string Path = PathPrefix + "." + PathSuffix + ".bc";
83       std::error_code EC;
84       raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
85       // Because -save-temps is a debugging feature, we report the error
86       // directly and exit.
87       if (EC)
88         reportOpenError(Path, EC.message());
89       WriteBitcodeToFile(M, OS, /*ShouldPreserveUseListOrder=*/false);
90       return true;
91     };
92   };
93 
94   setHook("0.preopt", PreOptModuleHook);
95   setHook("1.promote", PostPromoteModuleHook);
96   setHook("2.internalize", PostInternalizeModuleHook);
97   setHook("3.import", PostImportModuleHook);
98   setHook("4.opt", PostOptModuleHook);
99   setHook("5.precodegen", PreCodeGenModuleHook);
100 
101   CombinedIndexHook = [=](const ModuleSummaryIndex &Index) {
102     std::string Path = OutputFileName + "index.bc";
103     std::error_code EC;
104     raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
105     // Because -save-temps is a debugging feature, we report the error
106     // directly and exit.
107     if (EC)
108       reportOpenError(Path, EC.message());
109     WriteIndexToFile(Index, OS);
110 
111     Path = OutputFileName + "index.dot";
112     raw_fd_ostream OSDot(Path, EC, sys::fs::OpenFlags::F_None);
113     if (EC)
114       reportOpenError(Path, EC.message());
115     Index.exportToDot(OSDot);
116     return true;
117   };
118 
119   return Error::success();
120 }
121 
122 namespace {
123 
124 std::unique_ptr<TargetMachine>
125 createTargetMachine(Config &Conf, const Target *TheTarget, Module &M) {
126   StringRef TheTriple = M.getTargetTriple();
127   SubtargetFeatures Features;
128   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
129   for (const std::string &A : Conf.MAttrs)
130     Features.AddFeature(A);
131 
132   Reloc::Model RelocModel;
133   if (Conf.RelocModel)
134     RelocModel = *Conf.RelocModel;
135   else
136     RelocModel =
137         M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_;
138 
139   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
140       TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel,
141       Conf.CodeModel, Conf.CGOptLevel));
142 }
143 
144 static void runNewPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
145                            unsigned OptLevel, bool IsThinLTO) {
146   Optional<PGOOptions> PGOOpt;
147   if (!Conf.SampleProfile.empty())
148     PGOOpt = PGOOptions("", "", Conf.SampleProfile, false, true);
149 
150   PassBuilder PB(TM, PGOOpt);
151   AAManager AA;
152 
153   // Parse a custom AA pipeline if asked to.
154   if (!PB.parseAAPipeline(AA, "default"))
155     report_fatal_error("Error parsing default AA pipeline");
156 
157   LoopAnalysisManager LAM(Conf.DebugPassManager);
158   FunctionAnalysisManager FAM(Conf.DebugPassManager);
159   CGSCCAnalysisManager CGAM(Conf.DebugPassManager);
160   ModuleAnalysisManager MAM(Conf.DebugPassManager);
161 
162   // Register the AA manager first so that our version is the one used.
163   FAM.registerPass([&] { return std::move(AA); });
164 
165   // Register all the basic analyses with the managers.
166   PB.registerModuleAnalyses(MAM);
167   PB.registerCGSCCAnalyses(CGAM);
168   PB.registerFunctionAnalyses(FAM);
169   PB.registerLoopAnalyses(LAM);
170   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
171 
172   ModulePassManager MPM(Conf.DebugPassManager);
173   // FIXME (davide): verify the input.
174 
175   PassBuilder::OptimizationLevel OL;
176 
177   switch (OptLevel) {
178   default:
179     llvm_unreachable("Invalid optimization level");
180   case 0:
181     OL = PassBuilder::O0;
182     break;
183   case 1:
184     OL = PassBuilder::O1;
185     break;
186   case 2:
187     OL = PassBuilder::O2;
188     break;
189   case 3:
190     OL = PassBuilder::O3;
191     break;
192   }
193 
194   if (IsThinLTO)
195     MPM = PB.buildThinLTODefaultPipeline(OL, Conf.DebugPassManager);
196   else
197     MPM = PB.buildLTODefaultPipeline(OL, Conf.DebugPassManager);
198   MPM.run(Mod, MAM);
199 
200   // FIXME (davide): verify the output.
201 }
202 
203 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
204                                  std::string PipelineDesc,
205                                  std::string AAPipelineDesc,
206                                  bool DisableVerify) {
207   PassBuilder PB(TM);
208   AAManager AA;
209 
210   // Parse a custom AA pipeline if asked to.
211   if (!AAPipelineDesc.empty())
212     if (!PB.parseAAPipeline(AA, AAPipelineDesc))
213       report_fatal_error("unable to parse AA pipeline description: " +
214                          AAPipelineDesc);
215 
216   LoopAnalysisManager LAM;
217   FunctionAnalysisManager FAM;
218   CGSCCAnalysisManager CGAM;
219   ModuleAnalysisManager MAM;
220 
221   // Register the AA manager first so that our version is the one used.
222   FAM.registerPass([&] { return std::move(AA); });
223 
224   // Register all the basic analyses with the managers.
225   PB.registerModuleAnalyses(MAM);
226   PB.registerCGSCCAnalyses(CGAM);
227   PB.registerFunctionAnalyses(FAM);
228   PB.registerLoopAnalyses(LAM);
229   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
230 
231   ModulePassManager MPM;
232 
233   // Always verify the input.
234   MPM.addPass(VerifierPass());
235 
236   // Now, add all the passes we've been requested to.
237   if (!PB.parsePassPipeline(MPM, PipelineDesc))
238     report_fatal_error("unable to parse pass pipeline description: " +
239                        PipelineDesc);
240 
241   if (!DisableVerify)
242     MPM.addPass(VerifierPass());
243   MPM.run(Mod, MAM);
244 }
245 
246 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
247                            bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
248                            const ModuleSummaryIndex *ImportSummary) {
249   legacy::PassManager passes;
250   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
251 
252   PassManagerBuilder PMB;
253   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
254   PMB.Inliner = createFunctionInliningPass();
255   PMB.ExportSummary = ExportSummary;
256   PMB.ImportSummary = ImportSummary;
257   // Unconditionally verify input since it is not verified before this
258   // point and has unknown origin.
259   PMB.VerifyInput = true;
260   PMB.VerifyOutput = !Conf.DisableVerify;
261   PMB.LoopVectorize = true;
262   PMB.SLPVectorize = true;
263   PMB.OptLevel = Conf.OptLevel;
264   PMB.PGOSampleUse = Conf.SampleProfile;
265   if (IsThinLTO)
266     PMB.populateThinLTOPassManager(passes);
267   else
268     PMB.populateLTOPassManager(passes);
269   passes.run(Mod);
270 }
271 
272 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
273          bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
274          const ModuleSummaryIndex *ImportSummary) {
275   // FIXME: Plumb the combined index into the new pass manager.
276   if (!Conf.OptPipeline.empty())
277     runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline,
278                          Conf.DisableVerify);
279   else if (Conf.UseNewPM)
280     runNewPMPasses(Conf, Mod, TM, Conf.OptLevel, IsThinLTO);
281   else
282     runOldPMPasses(Conf, Mod, TM, IsThinLTO, ExportSummary, ImportSummary);
283   return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
284 }
285 
286 void codegenWithSplitDwarf(Config &Conf, TargetMachine *TM,
287                            AddStreamFn AddStream, unsigned Task, Module &Mod) {
288   SmallString<128> TempFile;
289   int FD = -1;
290   if (auto EC =
291       sys::fs::createTemporaryFile("lto-llvm-fission", "o", FD, TempFile))
292     report_fatal_error("Could not create temporary file " +
293         TempFile.str() + ": " + EC.message());
294   llvm::raw_fd_ostream OS(FD, true);
295   SmallString<1024> DwarfFile(Conf.DwoDir);
296   std::string DwoName = sys::path::filename(Mod.getModuleIdentifier()).str() +
297       "-" + std::to_string(Task) + "-";
298   size_t index = TempFile.str().rfind("lto-llvm-fission");
299   StringRef TempID = TempFile.str().substr(index + 17, 6);
300   DwoName += TempID.str() + ".dwo";
301   sys::path::append(DwarfFile, DwoName);
302   TM->Options.MCOptions.SplitDwarfFile = DwarfFile.str().str();
303 
304   legacy::PassManager CodeGenPasses;
305   if (TM->addPassesToEmitFile(CodeGenPasses, OS, Conf.CGFileType))
306     report_fatal_error("Failed to setup codegen");
307   CodeGenPasses.run(Mod);
308 
309   if (auto EC = llvm::sys::fs::create_directories(Conf.DwoDir))
310     report_fatal_error("Failed to create directory " +
311 		       Conf.DwoDir + ": " + EC.message());
312 
313   SmallVector<const char*, 5> ExtractArgs, StripArgs;
314   ExtractArgs.push_back(Conf.Objcopy.c_str());
315   ExtractArgs.push_back("--extract-dwo");
316   ExtractArgs.push_back(TempFile.c_str());
317   ExtractArgs.push_back(TM->Options.MCOptions.SplitDwarfFile.c_str());
318   ExtractArgs.push_back(nullptr);
319   StripArgs.push_back(Conf.Objcopy.c_str());
320   StripArgs.push_back("--strip-dwo");
321   StripArgs.push_back(TempFile.c_str());
322   StripArgs.push_back(nullptr);
323 
324   if (auto Ret = sys::ExecuteAndWait(Conf.Objcopy, ExtractArgs.data())) {
325     report_fatal_error("Failed to extract dwo from " + TempFile.str() +
326         ". Exit code " + std::to_string(Ret));
327   }
328   if (auto Ret = sys::ExecuteAndWait(Conf.Objcopy, StripArgs.data())) {
329     report_fatal_error("Failed to strip dwo from " + TempFile.str() +
330         ". Exit code " + std::to_string(Ret));
331   }
332 
333   auto Stream = AddStream(Task);
334   auto Buffer = MemoryBuffer::getFile(TempFile);
335   if (auto EC = Buffer.getError())
336     report_fatal_error("Failed to load file " +
337                        TempFile.str() + ": " + EC.message());
338   *Stream->OS << Buffer.get()->getBuffer();
339   if (auto EC = sys::fs::remove(TempFile))
340     report_fatal_error("Failed to delete file " +
341                        TempFile.str() + ": " + EC.message());
342 }
343 
344 void codegen(Config &Conf, TargetMachine *TM, AddStreamFn AddStream,
345              unsigned Task, Module &Mod) {
346   if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
347     return;
348 
349   if (!Conf.DwoDir.empty()) {
350     codegenWithSplitDwarf(Conf, TM, AddStream, Task, Mod);
351     return;
352   }
353 
354   auto Stream = AddStream(Task);
355   legacy::PassManager CodeGenPasses;
356   if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS, Conf.CGFileType))
357     report_fatal_error("Failed to setup codegen");
358   CodeGenPasses.run(Mod);
359 }
360 
361 void splitCodeGen(Config &C, TargetMachine *TM, AddStreamFn AddStream,
362                   unsigned ParallelCodeGenParallelismLevel,
363                   std::unique_ptr<Module> Mod) {
364   ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel);
365   unsigned ThreadCount = 0;
366   const Target *T = &TM->getTarget();
367 
368   SplitModule(
369       std::move(Mod), ParallelCodeGenParallelismLevel,
370       [&](std::unique_ptr<Module> MPart) {
371         // We want to clone the module in a new context to multi-thread the
372         // codegen. We do it by serializing partition modules to bitcode
373         // (while still on the main thread, in order to avoid data races) and
374         // spinning up new threads which deserialize the partitions into
375         // separate contexts.
376         // FIXME: Provide a more direct way to do this in LLVM.
377         SmallString<0> BC;
378         raw_svector_ostream BCOS(BC);
379         WriteBitcodeToFile(*MPart, BCOS);
380 
381         // Enqueue the task
382         CodegenThreadPool.async(
383             [&](const SmallString<0> &BC, unsigned ThreadId) {
384               LTOLLVMContext Ctx(C);
385               Expected<std::unique_ptr<Module>> MOrErr = parseBitcodeFile(
386                   MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"),
387                   Ctx);
388               if (!MOrErr)
389                 report_fatal_error("Failed to read bitcode");
390               std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
391 
392               std::unique_ptr<TargetMachine> TM =
393                   createTargetMachine(C, T, *MPartInCtx);
394 
395               codegen(C, TM.get(), AddStream, ThreadId, *MPartInCtx);
396             },
397             // Pass BC using std::move to ensure that it get moved rather than
398             // copied into the thread's context.
399             std::move(BC), ThreadCount++);
400       },
401       false);
402 
403   // Because the inner lambda (which runs in a worker thread) captures our local
404   // variables, we need to wait for the worker threads to terminate before we
405   // can leave the function scope.
406   CodegenThreadPool.wait();
407 }
408 
409 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) {
410   if (!C.OverrideTriple.empty())
411     Mod.setTargetTriple(C.OverrideTriple);
412   else if (Mod.getTargetTriple().empty())
413     Mod.setTargetTriple(C.DefaultTriple);
414 
415   std::string Msg;
416   const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg);
417   if (!T)
418     return make_error<StringError>(Msg, inconvertibleErrorCode());
419   return T;
420 }
421 
422 }
423 
424 static void
425 finalizeOptimizationRemarks(std::unique_ptr<ToolOutputFile> DiagOutputFile) {
426   // Make sure we flush the diagnostic remarks file in case the linker doesn't
427   // call the global destructors before exiting.
428   if (!DiagOutputFile)
429     return;
430   DiagOutputFile->keep();
431   DiagOutputFile->os().flush();
432 }
433 
434 Error lto::backend(Config &C, AddStreamFn AddStream,
435                    unsigned ParallelCodeGenParallelismLevel,
436                    std::unique_ptr<Module> Mod,
437                    ModuleSummaryIndex &CombinedIndex) {
438   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
439   if (!TOrErr)
440     return TOrErr.takeError();
441 
442   std::unique_ptr<TargetMachine> TM = createTargetMachine(C, *TOrErr, *Mod);
443 
444   // Setup optimization remarks.
445   auto DiagFileOrErr = lto::setupOptimizationRemarks(
446       Mod->getContext(), C.RemarksFilename, C.RemarksWithHotness);
447   if (!DiagFileOrErr)
448     return DiagFileOrErr.takeError();
449   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
450 
451   if (!C.CodeGenOnly) {
452     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false,
453              /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr)) {
454       finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
455       return Error::success();
456     }
457   }
458 
459   if (ParallelCodeGenParallelismLevel == 1) {
460     codegen(C, TM.get(), AddStream, 0, *Mod);
461   } else {
462     splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel,
463                  std::move(Mod));
464   }
465   finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
466   return Error::success();
467 }
468 
469 static void dropDeadSymbols(Module &Mod, const GVSummaryMapTy &DefinedGlobals,
470                             const ModuleSummaryIndex &Index) {
471   std::vector<GlobalValue*> DeadGVs;
472   for (auto &GV : Mod.global_values())
473     if (GlobalValueSummary *GVS = DefinedGlobals.lookup(GV.getGUID()))
474       if (!Index.isGlobalValueLive(GVS)) {
475         DeadGVs.push_back(&GV);
476         convertToDeclaration(GV);
477       }
478 
479   // Now that all dead bodies have been dropped, delete the actual objects
480   // themselves when possible.
481   for (GlobalValue *GV : DeadGVs) {
482     GV->removeDeadConstantUsers();
483     // Might reference something defined in native object (i.e. dropped a
484     // non-prevailing IR def, but we need to keep the declaration).
485     if (GV->use_empty())
486       GV->eraseFromParent();
487   }
488 }
489 
490 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
491                        Module &Mod, const ModuleSummaryIndex &CombinedIndex,
492                        const FunctionImporter::ImportMapTy &ImportList,
493                        const GVSummaryMapTy &DefinedGlobals,
494                        MapVector<StringRef, BitcodeModule> &ModuleMap) {
495   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
496   if (!TOrErr)
497     return TOrErr.takeError();
498 
499   std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, *TOrErr, Mod);
500 
501   if (Conf.CodeGenOnly) {
502     codegen(Conf, TM.get(), AddStream, Task, Mod);
503     return Error::success();
504   }
505 
506   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
507     return Error::success();
508 
509   renameModuleForThinLTO(Mod, CombinedIndex);
510 
511   dropDeadSymbols(Mod, DefinedGlobals, CombinedIndex);
512 
513   thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals);
514 
515   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
516     return Error::success();
517 
518   if (!DefinedGlobals.empty())
519     thinLTOInternalizeModule(Mod, DefinedGlobals);
520 
521   if (Conf.PostInternalizeModuleHook &&
522       !Conf.PostInternalizeModuleHook(Task, Mod))
523     return Error::success();
524 
525   auto ModuleLoader = [&](StringRef Identifier) {
526     assert(Mod.getContext().isODRUniquingDebugTypes() &&
527            "ODR Type uniquing should be enabled on the context");
528     auto I = ModuleMap.find(Identifier);
529     assert(I != ModuleMap.end());
530     return I->second.getLazyModule(Mod.getContext(),
531                                    /*ShouldLazyLoadMetadata=*/true,
532                                    /*IsImporting*/ true);
533   };
534 
535   FunctionImporter Importer(CombinedIndex, ModuleLoader);
536   if (Error Err = Importer.importFunctions(Mod, ImportList).takeError())
537     return Err;
538 
539   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
540     return Error::success();
541 
542   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true,
543            /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex))
544     return Error::success();
545 
546   codegen(Conf, TM.get(), AddStream, Task, Mod);
547   return Error::success();
548 }
549