1 //===-LTOBackend.cpp - LLVM Link Time Optimizer Backend -------------------===//
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 "backend" phase of LTO, i.e. it performs
10 // optimization and code generation on a loaded module. It is generally used
11 // internally by the LTO class but can also be used independently, for example
12 // to implement a standalone ThinLTO backend.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/LTO/LTOBackend.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/CGSCCPassManager.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Bitcode/BitcodeReader.h"
22 #include "llvm/Bitcode/BitcodeWriter.h"
23 #include "llvm/IR/LegacyPassManager.h"
24 #include "llvm/IR/PassManager.h"
25 #include "llvm/IR/Verifier.h"
26 #include "llvm/LTO/LTO.h"
27 #include "llvm/MC/SubtargetFeature.h"
28 #include "llvm/Object/ModuleSymbolTable.h"
29 #include "llvm/Passes/PassBuilder.h"
30 #include "llvm/Support/Error.h"
31 #include "llvm/Support/FileSystem.h"
32 #include "llvm/Support/MemoryBuffer.h"
33 #include "llvm/Support/Path.h"
34 #include "llvm/Support/Program.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Support/TargetRegistry.h"
37 #include "llvm/Support/ThreadPool.h"
38 #include "llvm/Target/TargetMachine.h"
39 #include "llvm/Transforms/IPO.h"
40 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
41 #include "llvm/Transforms/Scalar/LoopPassManager.h"
42 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
43 #include "llvm/Transforms/Utils/SplitModule.h"
44 
45 using namespace llvm;
46 using namespace lto;
47 
48 LLVM_ATTRIBUTE_NORETURN static void reportOpenError(StringRef Path, Twine Msg) {
49   errs() << "failed to open " << Path << ": " << Msg << '\n';
50   errs().flush();
51   exit(1);
52 }
53 
54 Error Config::addSaveTemps(std::string OutputFileName,
55                            bool UseInputModulePath) {
56   ShouldDiscardValueNames = false;
57 
58   std::error_code EC;
59   ResolutionFile = llvm::make_unique<raw_fd_ostream>(
60       OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text);
61   if (EC)
62     return errorCodeToError(EC);
63 
64   auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
65     // Keep track of the hook provided by the linker, which also needs to run.
66     ModuleHookFn LinkerHook = Hook;
67     Hook = [=](unsigned Task, const Module &M) {
68       // If the linker's hook returned false, we need to pass that result
69       // through.
70       if (LinkerHook && !LinkerHook(Task, M))
71         return false;
72 
73       std::string PathPrefix;
74       // If this is the combined module (not a ThinLTO backend compile) or the
75       // user hasn't requested using the input module's path, emit to a file
76       // named from the provided OutputFileName with the Task ID appended.
77       if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
78         PathPrefix = OutputFileName;
79         if (Task != (unsigned)-1)
80           PathPrefix += utostr(Task) + ".";
81       } else
82         PathPrefix = M.getModuleIdentifier() + ".";
83       std::string Path = PathPrefix + PathSuffix + ".bc";
84       std::error_code EC;
85       raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
86       // Because -save-temps is a debugging feature, we report the error
87       // directly and exit.
88       if (EC)
89         reportOpenError(Path, EC.message());
90       WriteBitcodeToFile(M, OS, /*ShouldPreserveUseListOrder=*/false);
91       return true;
92     };
93   };
94 
95   setHook("0.preopt", PreOptModuleHook);
96   setHook("1.promote", PostPromoteModuleHook);
97   setHook("2.internalize", PostInternalizeModuleHook);
98   setHook("3.import", PostImportModuleHook);
99   setHook("4.opt", PostOptModuleHook);
100   setHook("5.precodegen", PreCodeGenModuleHook);
101 
102   CombinedIndexHook = [=](const ModuleSummaryIndex &Index) {
103     std::string Path = OutputFileName + "index.bc";
104     std::error_code EC;
105     raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
106     // Because -save-temps is a debugging feature, we report the error
107     // directly and exit.
108     if (EC)
109       reportOpenError(Path, EC.message());
110     WriteIndexToFile(Index, OS);
111 
112     Path = OutputFileName + "index.dot";
113     raw_fd_ostream OSDot(Path, EC, sys::fs::OpenFlags::F_None);
114     if (EC)
115       reportOpenError(Path, EC.message());
116     Index.exportToDot(OSDot);
117     return true;
118   };
119 
120   return Error::success();
121 }
122 
123 namespace {
124 
125 std::unique_ptr<TargetMachine>
126 createTargetMachine(Config &Conf, const Target *TheTarget, Module &M) {
127   StringRef TheTriple = M.getTargetTriple();
128   SubtargetFeatures Features;
129   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
130   for (const std::string &A : Conf.MAttrs)
131     Features.AddFeature(A);
132 
133   Reloc::Model RelocModel;
134   if (Conf.RelocModel)
135     RelocModel = *Conf.RelocModel;
136   else
137     RelocModel =
138         M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_;
139 
140   Optional<CodeModel::Model> CodeModel;
141   if (Conf.CodeModel)
142     CodeModel = *Conf.CodeModel;
143   else
144     CodeModel = M.getCodeModel();
145 
146   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
147       TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel,
148       CodeModel, Conf.CGOptLevel));
149 }
150 
151 static void runNewPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
152                            unsigned OptLevel, bool IsThinLTO,
153                            ModuleSummaryIndex *ExportSummary,
154                            const ModuleSummaryIndex *ImportSummary) {
155   Optional<PGOOptions> PGOOpt;
156   if (!Conf.SampleProfile.empty())
157     PGOOpt = PGOOptions(Conf.SampleProfile, "", Conf.ProfileRemapping,
158                         PGOOptions::SampleUse, PGOOptions::NoCSAction, true);
159   else if (Conf.RunCSIRInstr) {
160     PGOOpt = PGOOptions("", Conf.CSIRProfile, Conf.ProfileRemapping,
161                         PGOOptions::IRUse, PGOOptions::CSIRInstr);
162   } else if (!Conf.CSIRProfile.empty()) {
163     PGOOpt = PGOOptions(Conf.CSIRProfile, "", Conf.ProfileRemapping,
164                         PGOOptions::IRUse, PGOOptions::CSIRUse);
165   }
166 
167   PassBuilder PB(TM, PipelineTuningOptions(), PGOOpt);
168   AAManager AA;
169 
170   // Parse a custom AA pipeline if asked to.
171   if (auto Err = PB.parseAAPipeline(AA, "default"))
172     report_fatal_error("Error parsing default AA pipeline");
173 
174   LoopAnalysisManager LAM(Conf.DebugPassManager);
175   FunctionAnalysisManager FAM(Conf.DebugPassManager);
176   CGSCCAnalysisManager CGAM(Conf.DebugPassManager);
177   ModuleAnalysisManager MAM(Conf.DebugPassManager);
178 
179   // Register the AA manager first so that our version is the one used.
180   FAM.registerPass([&] { return std::move(AA); });
181 
182   // Register all the basic analyses with the managers.
183   PB.registerModuleAnalyses(MAM);
184   PB.registerCGSCCAnalyses(CGAM);
185   PB.registerFunctionAnalyses(FAM);
186   PB.registerLoopAnalyses(LAM);
187   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
188 
189   ModulePassManager MPM(Conf.DebugPassManager);
190   // FIXME (davide): verify the input.
191 
192   PassBuilder::OptimizationLevel OL;
193 
194   switch (OptLevel) {
195   default:
196     llvm_unreachable("Invalid optimization level");
197   case 0:
198     OL = PassBuilder::O0;
199     break;
200   case 1:
201     OL = PassBuilder::O1;
202     break;
203   case 2:
204     OL = PassBuilder::O2;
205     break;
206   case 3:
207     OL = PassBuilder::O3;
208     break;
209   }
210 
211   if (IsThinLTO)
212     MPM = PB.buildThinLTODefaultPipeline(OL, Conf.DebugPassManager,
213                                          ImportSummary);
214   else
215     MPM = PB.buildLTODefaultPipeline(OL, Conf.DebugPassManager, ExportSummary);
216   MPM.run(Mod, MAM);
217 
218   // FIXME (davide): verify the output.
219 }
220 
221 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
222                                  std::string PipelineDesc,
223                                  std::string AAPipelineDesc,
224                                  bool DisableVerify) {
225   PassBuilder PB(TM);
226   AAManager AA;
227 
228   // Parse a custom AA pipeline if asked to.
229   if (!AAPipelineDesc.empty())
230     if (auto Err = PB.parseAAPipeline(AA, AAPipelineDesc))
231       report_fatal_error("unable to parse AA pipeline description '" +
232                          AAPipelineDesc + "': " + toString(std::move(Err)));
233 
234   LoopAnalysisManager LAM;
235   FunctionAnalysisManager FAM;
236   CGSCCAnalysisManager CGAM;
237   ModuleAnalysisManager MAM;
238 
239   // Register the AA manager first so that our version is the one used.
240   FAM.registerPass([&] { return std::move(AA); });
241 
242   // Register all the basic analyses with the managers.
243   PB.registerModuleAnalyses(MAM);
244   PB.registerCGSCCAnalyses(CGAM);
245   PB.registerFunctionAnalyses(FAM);
246   PB.registerLoopAnalyses(LAM);
247   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
248 
249   ModulePassManager MPM;
250 
251   // Always verify the input.
252   MPM.addPass(VerifierPass());
253 
254   // Now, add all the passes we've been requested to.
255   if (auto Err = PB.parsePassPipeline(MPM, PipelineDesc))
256     report_fatal_error("unable to parse pass pipeline description '" +
257                        PipelineDesc + "': " + toString(std::move(Err)));
258 
259   if (!DisableVerify)
260     MPM.addPass(VerifierPass());
261   MPM.run(Mod, MAM);
262 }
263 
264 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
265                            bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
266                            const ModuleSummaryIndex *ImportSummary) {
267   legacy::PassManager passes;
268   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
269 
270   PassManagerBuilder PMB;
271   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
272   PMB.Inliner = createFunctionInliningPass();
273   PMB.ExportSummary = ExportSummary;
274   PMB.ImportSummary = ImportSummary;
275   // Unconditionally verify input since it is not verified before this
276   // point and has unknown origin.
277   PMB.VerifyInput = true;
278   PMB.VerifyOutput = !Conf.DisableVerify;
279   PMB.LoopVectorize = true;
280   PMB.SLPVectorize = true;
281   PMB.OptLevel = Conf.OptLevel;
282   PMB.PGOSampleUse = Conf.SampleProfile;
283   PMB.EnablePGOCSInstrGen = Conf.RunCSIRInstr;
284   if (!Conf.RunCSIRInstr && !Conf.CSIRProfile.empty()) {
285     PMB.EnablePGOCSInstrUse = true;
286     PMB.PGOInstrUse = Conf.CSIRProfile;
287   }
288   if (IsThinLTO)
289     PMB.populateThinLTOPassManager(passes);
290   else
291     PMB.populateLTOPassManager(passes);
292   passes.run(Mod);
293 }
294 
295 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
296          bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
297          const ModuleSummaryIndex *ImportSummary) {
298   // FIXME: Plumb the combined index into the new pass manager.
299   if (!Conf.OptPipeline.empty())
300     runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline,
301                          Conf.DisableVerify);
302   else if (Conf.UseNewPM)
303     runNewPMPasses(Conf, Mod, TM, Conf.OptLevel, IsThinLTO, ExportSummary,
304                    ImportSummary);
305   else
306     runOldPMPasses(Conf, Mod, TM, IsThinLTO, ExportSummary, ImportSummary);
307   return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
308 }
309 
310 void codegen(Config &Conf, TargetMachine *TM, AddStreamFn AddStream,
311              unsigned Task, Module &Mod) {
312   if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
313     return;
314 
315   std::unique_ptr<ToolOutputFile> DwoOut;
316   SmallString<1024> DwoFile(Conf.DwoPath);
317   if (!Conf.DwoDir.empty()) {
318     std::error_code EC;
319     if (auto EC = llvm::sys::fs::create_directories(Conf.DwoDir))
320       report_fatal_error("Failed to create directory " + Conf.DwoDir + ": " +
321                          EC.message());
322 
323     DwoFile = Conf.DwoDir;
324     sys::path::append(DwoFile, std::to_string(Task) + ".dwo");
325   }
326 
327   if (!DwoFile.empty()) {
328     std::error_code EC;
329     TM->Options.MCOptions.SplitDwarfFile = DwoFile.str().str();
330     DwoOut = llvm::make_unique<ToolOutputFile>(DwoFile, EC, sys::fs::F_None);
331     if (EC)
332       report_fatal_error("Failed to open " + DwoFile + ": " + EC.message());
333   }
334 
335   auto Stream = AddStream(Task);
336   legacy::PassManager CodeGenPasses;
337   if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS,
338                               DwoOut ? &DwoOut->os() : nullptr,
339                               Conf.CGFileType))
340     report_fatal_error("Failed to setup codegen");
341   CodeGenPasses.run(Mod);
342 
343   if (DwoOut)
344     DwoOut->keep();
345 }
346 
347 void splitCodeGen(Config &C, TargetMachine *TM, AddStreamFn AddStream,
348                   unsigned ParallelCodeGenParallelismLevel,
349                   std::unique_ptr<Module> Mod) {
350   ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel);
351   unsigned ThreadCount = 0;
352   const Target *T = &TM->getTarget();
353 
354   SplitModule(
355       std::move(Mod), ParallelCodeGenParallelismLevel,
356       [&](std::unique_ptr<Module> MPart) {
357         // We want to clone the module in a new context to multi-thread the
358         // codegen. We do it by serializing partition modules to bitcode
359         // (while still on the main thread, in order to avoid data races) and
360         // spinning up new threads which deserialize the partitions into
361         // separate contexts.
362         // FIXME: Provide a more direct way to do this in LLVM.
363         SmallString<0> BC;
364         raw_svector_ostream BCOS(BC);
365         WriteBitcodeToFile(*MPart, BCOS);
366 
367         // Enqueue the task
368         CodegenThreadPool.async(
369             [&](const SmallString<0> &BC, unsigned ThreadId) {
370               LTOLLVMContext Ctx(C);
371               Expected<std::unique_ptr<Module>> MOrErr = parseBitcodeFile(
372                   MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"),
373                   Ctx);
374               if (!MOrErr)
375                 report_fatal_error("Failed to read bitcode");
376               std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
377 
378               std::unique_ptr<TargetMachine> TM =
379                   createTargetMachine(C, T, *MPartInCtx);
380 
381               codegen(C, TM.get(), AddStream, ThreadId, *MPartInCtx);
382             },
383             // Pass BC using std::move to ensure that it get moved rather than
384             // copied into the thread's context.
385             std::move(BC), ThreadCount++);
386       },
387       false);
388 
389   // Because the inner lambda (which runs in a worker thread) captures our local
390   // variables, we need to wait for the worker threads to terminate before we
391   // can leave the function scope.
392   CodegenThreadPool.wait();
393 }
394 
395 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) {
396   if (!C.OverrideTriple.empty())
397     Mod.setTargetTriple(C.OverrideTriple);
398   else if (Mod.getTargetTriple().empty())
399     Mod.setTargetTriple(C.DefaultTriple);
400 
401   std::string Msg;
402   const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg);
403   if (!T)
404     return make_error<StringError>(Msg, inconvertibleErrorCode());
405   return T;
406 }
407 
408 }
409 
410 static Error
411 finalizeOptimizationRemarks(std::unique_ptr<ToolOutputFile> DiagOutputFile) {
412   // Make sure we flush the diagnostic remarks file in case the linker doesn't
413   // call the global destructors before exiting.
414   if (!DiagOutputFile)
415     return Error::success();
416   DiagOutputFile->keep();
417   DiagOutputFile->os().flush();
418   return Error::success();
419 }
420 
421 Error lto::backend(Config &C, AddStreamFn AddStream,
422                    unsigned ParallelCodeGenParallelismLevel,
423                    std::unique_ptr<Module> Mod,
424                    ModuleSummaryIndex &CombinedIndex) {
425   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
426   if (!TOrErr)
427     return TOrErr.takeError();
428 
429   std::unique_ptr<TargetMachine> TM = createTargetMachine(C, *TOrErr, *Mod);
430 
431   // Setup optimization remarks.
432   auto DiagFileOrErr =
433       lto::setupOptimizationRemarks(Mod->getContext(), C.RemarksFilename,
434                                     C.RemarksPasses, C.RemarksWithHotness);
435   if (!DiagFileOrErr)
436     return DiagFileOrErr.takeError();
437   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
438 
439   if (!C.CodeGenOnly) {
440     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false,
441              /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr))
442       return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
443   }
444 
445   if (ParallelCodeGenParallelismLevel == 1) {
446     codegen(C, TM.get(), AddStream, 0, *Mod);
447   } else {
448     splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel,
449                  std::move(Mod));
450   }
451   return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
452 }
453 
454 static void dropDeadSymbols(Module &Mod, const GVSummaryMapTy &DefinedGlobals,
455                             const ModuleSummaryIndex &Index) {
456   std::vector<GlobalValue*> DeadGVs;
457   for (auto &GV : Mod.global_values())
458     if (GlobalValueSummary *GVS = DefinedGlobals.lookup(GV.getGUID()))
459       if (!Index.isGlobalValueLive(GVS)) {
460         DeadGVs.push_back(&GV);
461         convertToDeclaration(GV);
462       }
463 
464   // Now that all dead bodies have been dropped, delete the actual objects
465   // themselves when possible.
466   for (GlobalValue *GV : DeadGVs) {
467     GV->removeDeadConstantUsers();
468     // Might reference something defined in native object (i.e. dropped a
469     // non-prevailing IR def, but we need to keep the declaration).
470     if (GV->use_empty())
471       GV->eraseFromParent();
472   }
473 }
474 
475 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
476                        Module &Mod, const ModuleSummaryIndex &CombinedIndex,
477                        const FunctionImporter::ImportMapTy &ImportList,
478                        const GVSummaryMapTy &DefinedGlobals,
479                        MapVector<StringRef, BitcodeModule> &ModuleMap) {
480   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
481   if (!TOrErr)
482     return TOrErr.takeError();
483 
484   std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, *TOrErr, Mod);
485 
486   // Setup optimization remarks.
487   auto DiagFileOrErr = lto::setupOptimizationRemarks(
488       Mod.getContext(), Conf.RemarksFilename, Conf.RemarksPasses,
489       Conf.RemarksWithHotness, Task);
490   if (!DiagFileOrErr)
491     return DiagFileOrErr.takeError();
492   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
493 
494   if (Conf.CodeGenOnly) {
495     codegen(Conf, TM.get(), AddStream, Task, Mod);
496     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
497   }
498 
499   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
500     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
501 
502   renameModuleForThinLTO(Mod, CombinedIndex);
503 
504   dropDeadSymbols(Mod, DefinedGlobals, CombinedIndex);
505 
506   thinLTOResolvePrevailingInModule(Mod, DefinedGlobals);
507 
508   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
509     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
510 
511   if (!DefinedGlobals.empty())
512     thinLTOInternalizeModule(Mod, DefinedGlobals);
513 
514   if (Conf.PostInternalizeModuleHook &&
515       !Conf.PostInternalizeModuleHook(Task, Mod))
516     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
517 
518   auto ModuleLoader = [&](StringRef Identifier) {
519     assert(Mod.getContext().isODRUniquingDebugTypes() &&
520            "ODR Type uniquing should be enabled on the context");
521     auto I = ModuleMap.find(Identifier);
522     assert(I != ModuleMap.end());
523     return I->second.getLazyModule(Mod.getContext(),
524                                    /*ShouldLazyLoadMetadata=*/true,
525                                    /*IsImporting*/ true);
526   };
527 
528   FunctionImporter Importer(CombinedIndex, ModuleLoader);
529   if (Error Err = Importer.importFunctions(Mod, ImportList).takeError())
530     return Err;
531 
532   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
533     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
534 
535   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true,
536            /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex))
537     return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
538 
539   codegen(Conf, TM.get(), AddStream, Task, Mod);
540   return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
541 }
542