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/LTO/legacy/UpdateCompilerUsed.h"
29 #include "llvm/MC/SubtargetFeature.h"
30 #include "llvm/Object/ModuleSymbolTable.h"
31 #include "llvm/Passes/PassBuilder.h"
32 #include "llvm/Support/Error.h"
33 #include "llvm/Support/FileSystem.h"
34 #include "llvm/Support/TargetRegistry.h"
35 #include "llvm/Support/ThreadPool.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Transforms/IPO.h"
38 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
39 #include "llvm/Transforms/Scalar/LoopPassManager.h"
40 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
41 #include "llvm/Transforms/Utils/SplitModule.h"
42 
43 using namespace llvm;
44 using namespace lto;
45 
46 static cl::opt<bool>
47     LTOUseNewPM("lto-use-new-pm",
48                 cl::desc("Run LTO passes using the new pass manager"),
49                 cl::init(false), cl::Hidden);
50 
51 LLVM_ATTRIBUTE_NORETURN static void reportOpenError(StringRef Path, Twine Msg) {
52   errs() << "failed to open " << Path << ": " << Msg << '\n';
53   errs().flush();
54   exit(1);
55 }
56 
57 Error Config::addSaveTemps(std::string OutputFileName,
58                            bool UseInputModulePath) {
59   ShouldDiscardValueNames = false;
60 
61   std::error_code EC;
62   ResolutionFile = llvm::make_unique<raw_fd_ostream>(
63       OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text);
64   if (EC)
65     return errorCodeToError(EC);
66 
67   auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
68     // Keep track of the hook provided by the linker, which also needs to run.
69     ModuleHookFn LinkerHook = Hook;
70     Hook = [=](unsigned Task, const Module &M) {
71       // If the linker's hook returned false, we need to pass that result
72       // through.
73       if (LinkerHook && !LinkerHook(Task, M))
74         return false;
75 
76       std::string PathPrefix;
77       // If this is the combined module (not a ThinLTO backend compile) or the
78       // user hasn't requested using the input module's path, emit to a file
79       // named from the provided OutputFileName with the Task ID appended.
80       if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
81         PathPrefix = OutputFileName + 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     return true;
113   };
114 
115   return Error::success();
116 }
117 
118 namespace {
119 
120 std::unique_ptr<TargetMachine>
121 createTargetMachine(Config &Conf, StringRef TheTriple,
122                     const Target *TheTarget) {
123   SubtargetFeatures Features;
124   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
125   for (const std::string &A : Conf.MAttrs)
126     Features.AddFeature(A);
127 
128   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
129       TheTriple, Conf.CPU, Features.getString(), Conf.Options, Conf.RelocModel,
130       Conf.CodeModel, Conf.CGOptLevel));
131 }
132 
133 static void runNewPMPasses(Module &Mod, TargetMachine *TM, unsigned OptLevel) {
134   PassBuilder PB(TM);
135   AAManager AA;
136 
137   // Parse a custom AA pipeline if asked to.
138   assert(PB.parseAAPipeline(AA, "default"));
139 
140   LoopAnalysisManager LAM;
141   FunctionAnalysisManager FAM;
142   CGSCCAnalysisManager CGAM;
143   ModuleAnalysisManager MAM;
144 
145   // Register the AA manager first so that our version is the one used.
146   FAM.registerPass([&] { return std::move(AA); });
147 
148   // Register all the basic analyses with the managers.
149   PB.registerModuleAnalyses(MAM);
150   PB.registerCGSCCAnalyses(CGAM);
151   PB.registerFunctionAnalyses(FAM);
152   PB.registerLoopAnalyses(LAM);
153   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
154 
155   ModulePassManager MPM;
156   // FIXME (davide): verify the input.
157 
158   PassBuilder::OptimizationLevel OL;
159 
160   switch (OptLevel) {
161   default:
162     llvm_unreachable("Invalid optimization level");
163   case 0:
164     OL = PassBuilder::O0;
165     break;
166   case 1:
167     OL = PassBuilder::O1;
168     break;
169   case 2:
170     OL = PassBuilder::O2;
171     break;
172   case 3:
173     OL = PassBuilder::O3;
174     break;
175   }
176 
177   MPM = PB.buildLTODefaultPipeline(OL, false /* DebugLogging */);
178   MPM.run(Mod, MAM);
179 
180   // FIXME (davide): verify the output.
181 }
182 
183 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
184                                  std::string PipelineDesc,
185                                  std::string AAPipelineDesc,
186                                  bool DisableVerify) {
187   PassBuilder PB(TM);
188   AAManager AA;
189 
190   // Parse a custom AA pipeline if asked to.
191   if (!AAPipelineDesc.empty())
192     if (!PB.parseAAPipeline(AA, AAPipelineDesc))
193       report_fatal_error("unable to parse AA pipeline description: " +
194                          AAPipelineDesc);
195 
196   LoopAnalysisManager LAM;
197   FunctionAnalysisManager FAM;
198   CGSCCAnalysisManager CGAM;
199   ModuleAnalysisManager MAM;
200 
201   // Register the AA manager first so that our version is the one used.
202   FAM.registerPass([&] { return std::move(AA); });
203 
204   // Register all the basic analyses with the managers.
205   PB.registerModuleAnalyses(MAM);
206   PB.registerCGSCCAnalyses(CGAM);
207   PB.registerFunctionAnalyses(FAM);
208   PB.registerLoopAnalyses(LAM);
209   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
210 
211   ModulePassManager MPM;
212 
213   // Always verify the input.
214   MPM.addPass(VerifierPass());
215 
216   // Now, add all the passes we've been requested to.
217   if (!PB.parsePassPipeline(MPM, PipelineDesc))
218     report_fatal_error("unable to parse pass pipeline description: " +
219                        PipelineDesc);
220 
221   if (!DisableVerify)
222     MPM.addPass(VerifierPass());
223   MPM.run(Mod, MAM);
224 }
225 
226 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
227                            bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
228                            const ModuleSummaryIndex *ImportSummary) {
229   legacy::PassManager passes;
230   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
231 
232   PassManagerBuilder PMB;
233   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
234   PMB.Inliner = createFunctionInliningPass();
235   PMB.ExportSummary = ExportSummary;
236   PMB.ImportSummary = ImportSummary;
237   // Unconditionally verify input since it is not verified before this
238   // point and has unknown origin.
239   PMB.VerifyInput = true;
240   PMB.VerifyOutput = !Conf.DisableVerify;
241   PMB.LoopVectorize = true;
242   PMB.SLPVectorize = true;
243   PMB.OptLevel = Conf.OptLevel;
244   PMB.PGOSampleUse = Conf.SampleProfile;
245   if (IsThinLTO)
246     PMB.populateThinLTOPassManager(passes);
247   else
248     PMB.populateLTOPassManager(passes);
249   passes.run(Mod);
250 }
251 
252 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
253          bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
254          const ModuleSummaryIndex *ImportSummary) {
255   // There's still no ThinLTO pipeline hooked up in the new pass manager,
256   // once there is one, we can just remove this.
257   if (LTOUseNewPM && IsThinLTO)
258     report_fatal_error("ThinLTO not supported with the new PM yet!");
259 
260   // FIXME: Plumb the combined index into the new pass manager.
261   if (!Conf.OptPipeline.empty())
262     runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline,
263                          Conf.DisableVerify);
264   else if (LTOUseNewPM)
265     runNewPMPasses(Mod, TM, Conf.OptLevel);
266   else
267     runOldPMPasses(Conf, Mod, TM, IsThinLTO, ExportSummary, ImportSummary);
268   return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
269 }
270 
271 void codegen(Config &Conf, TargetMachine *TM, AddStreamFn AddStream,
272              unsigned Task, Module &Mod) {
273   if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
274     return;
275 
276   auto Stream = AddStream(Task);
277   legacy::PassManager CodeGenPasses;
278   if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS, Conf.CGFileType))
279     report_fatal_error("Failed to setup codegen");
280   CodeGenPasses.run(Mod);
281 }
282 
283 void splitCodeGen(Config &C, TargetMachine *TM, AddStreamFn AddStream,
284                   unsigned ParallelCodeGenParallelismLevel,
285                   std::unique_ptr<Module> Mod) {
286   ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel);
287   unsigned ThreadCount = 0;
288   const Target *T = &TM->getTarget();
289 
290   SplitModule(
291       std::move(Mod), ParallelCodeGenParallelismLevel,
292       [&](std::unique_ptr<Module> MPart) {
293         // We want to clone the module in a new context to multi-thread the
294         // codegen. We do it by serializing partition modules to bitcode
295         // (while still on the main thread, in order to avoid data races) and
296         // spinning up new threads which deserialize the partitions into
297         // separate contexts.
298         // FIXME: Provide a more direct way to do this in LLVM.
299         SmallString<0> BC;
300         raw_svector_ostream BCOS(BC);
301         WriteBitcodeToFile(MPart.get(), BCOS);
302 
303         // Enqueue the task
304         CodegenThreadPool.async(
305             [&](const SmallString<0> &BC, unsigned ThreadId) {
306               LTOLLVMContext Ctx(C);
307               Expected<std::unique_ptr<Module>> MOrErr = parseBitcodeFile(
308                   MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"),
309                   Ctx);
310               if (!MOrErr)
311                 report_fatal_error("Failed to read bitcode");
312               std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
313 
314               std::unique_ptr<TargetMachine> TM =
315                   createTargetMachine(C, MPartInCtx->getTargetTriple(), T);
316 
317               codegen(C, TM.get(), AddStream, ThreadId, *MPartInCtx);
318             },
319             // Pass BC using std::move to ensure that it get moved rather than
320             // copied into the thread's context.
321             std::move(BC), ThreadCount++);
322       },
323       false);
324 
325   // Because the inner lambda (which runs in a worker thread) captures our local
326   // variables, we need to wait for the worker threads to terminate before we
327   // can leave the function scope.
328   CodegenThreadPool.wait();
329 }
330 
331 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) {
332   if (!C.OverrideTriple.empty())
333     Mod.setTargetTriple(C.OverrideTriple);
334   else if (Mod.getTargetTriple().empty())
335     Mod.setTargetTriple(C.DefaultTriple);
336 
337   std::string Msg;
338   const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg);
339   if (!T)
340     return make_error<StringError>(Msg, inconvertibleErrorCode());
341   return T;
342 }
343 
344 }
345 
346 static void
347 finalizeOptimizationRemarks(std::unique_ptr<tool_output_file> DiagOutputFile) {
348   // Make sure we flush the diagnostic remarks file in case the linker doesn't
349   // call the global destructors before exiting.
350   if (!DiagOutputFile)
351     return;
352   DiagOutputFile->keep();
353   DiagOutputFile->os().flush();
354 }
355 
356 static void handleAsmUndefinedRefs(Module &Mod, TargetMachine &TM) {
357   // Collect the list of undefined symbols used in asm and update
358   // llvm.compiler.used to prevent optimization to drop these from the output.
359   StringSet<> AsmUndefinedRefs;
360   ModuleSymbolTable::CollectAsmSymbols(
361       Mod,
362       [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
363         if (Flags & object::BasicSymbolRef::SF_Undefined)
364           AsmUndefinedRefs.insert(Name);
365       });
366   updateCompilerUsed(Mod, TM, AsmUndefinedRefs);
367 }
368 
369 Error lto::backend(Config &C, AddStreamFn AddStream,
370                    unsigned ParallelCodeGenParallelismLevel,
371                    std::unique_ptr<Module> Mod,
372                    ModuleSummaryIndex &CombinedIndex) {
373   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
374   if (!TOrErr)
375     return TOrErr.takeError();
376 
377   std::unique_ptr<TargetMachine> TM =
378       createTargetMachine(C, Mod->getTargetTriple(), *TOrErr);
379 
380   handleAsmUndefinedRefs(*Mod, *TM);
381 
382   // Setup optimization remarks.
383   auto DiagFileOrErr = lto::setupOptimizationRemarks(
384       Mod->getContext(), C.RemarksFilename, C.RemarksWithHotness);
385   if (!DiagFileOrErr)
386     return DiagFileOrErr.takeError();
387   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
388 
389   if (!C.CodeGenOnly) {
390     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false,
391              /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr)) {
392       finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
393       return Error::success();
394     }
395   }
396 
397   if (ParallelCodeGenParallelismLevel == 1) {
398     codegen(C, TM.get(), AddStream, 0, *Mod);
399   } else {
400     splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel,
401                  std::move(Mod));
402   }
403   finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
404   return Error::success();
405 }
406 
407 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
408                        Module &Mod, const ModuleSummaryIndex &CombinedIndex,
409                        const FunctionImporter::ImportMapTy &ImportList,
410                        const GVSummaryMapTy &DefinedGlobals,
411                        MapVector<StringRef, BitcodeModule> &ModuleMap) {
412   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
413   if (!TOrErr)
414     return TOrErr.takeError();
415 
416   std::unique_ptr<TargetMachine> TM =
417       createTargetMachine(Conf, Mod.getTargetTriple(), *TOrErr);
418 
419   handleAsmUndefinedRefs(Mod, *TM);
420 
421   if (Conf.CodeGenOnly) {
422     codegen(Conf, TM.get(), AddStream, Task, Mod);
423     return Error::success();
424   }
425 
426   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
427     return Error::success();
428 
429   renameModuleForThinLTO(Mod, CombinedIndex);
430 
431   thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals);
432 
433   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
434     return Error::success();
435 
436   if (!DefinedGlobals.empty())
437     thinLTOInternalizeModule(Mod, DefinedGlobals);
438 
439   if (Conf.PostInternalizeModuleHook &&
440       !Conf.PostInternalizeModuleHook(Task, Mod))
441     return Error::success();
442 
443   auto ModuleLoader = [&](StringRef Identifier) {
444     assert(Mod.getContext().isODRUniquingDebugTypes() &&
445            "ODR Type uniquing should be enabled on the context");
446     auto I = ModuleMap.find(Identifier);
447     assert(I != ModuleMap.end());
448     return I->second.getLazyModule(Mod.getContext(),
449                                    /*ShouldLazyLoadMetadata=*/true,
450                                    /*IsImporting*/ true);
451   };
452 
453   FunctionImporter Importer(CombinedIndex, ModuleLoader);
454   if (Error Err = Importer.importFunctions(Mod, ImportList).takeError())
455     return Err;
456 
457   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
458     return Error::success();
459 
460   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true,
461            /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex))
462     return Error::success();
463 
464   codegen(Conf, TM.get(), AddStream, Task, Mod);
465   return Error::success();
466 }
467