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/TargetRegistry.h"
34 #include "llvm/Support/ThreadPool.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Transforms/IPO.h"
37 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
38 #include "llvm/Transforms/Scalar/LoopPassManager.h"
39 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
40 #include "llvm/Transforms/Utils/SplitModule.h"
41 
42 using namespace llvm;
43 using namespace lto;
44 
45 LLVM_ATTRIBUTE_NORETURN static void reportOpenError(StringRef Path, Twine Msg) {
46   errs() << "failed to open " << Path << ": " << Msg << '\n';
47   errs().flush();
48   exit(1);
49 }
50 
51 Error Config::addSaveTemps(std::string OutputFileName,
52                            bool UseInputModulePath) {
53   ShouldDiscardValueNames = false;
54 
55   std::error_code EC;
56   ResolutionFile = llvm::make_unique<raw_fd_ostream>(
57       OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text);
58   if (EC)
59     return errorCodeToError(EC);
60 
61   auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
62     // Keep track of the hook provided by the linker, which also needs to run.
63     ModuleHookFn LinkerHook = Hook;
64     Hook = [=](unsigned Task, const Module &M) {
65       // If the linker's hook returned false, we need to pass that result
66       // through.
67       if (LinkerHook && !LinkerHook(Task, M))
68         return false;
69 
70       std::string PathPrefix;
71       // If this is the combined module (not a ThinLTO backend compile) or the
72       // user hasn't requested using the input module's path, emit to a file
73       // named from the provided OutputFileName with the Task ID appended.
74       if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
75         PathPrefix = OutputFileName + utostr(Task);
76       } else
77         PathPrefix = M.getModuleIdentifier();
78       std::string Path = PathPrefix + "." + PathSuffix + ".bc";
79       std::error_code EC;
80       raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
81       // Because -save-temps is a debugging feature, we report the error
82       // directly and exit.
83       if (EC)
84         reportOpenError(Path, EC.message());
85       WriteBitcodeToFile(&M, OS, /*ShouldPreserveUseListOrder=*/false);
86       return true;
87     };
88   };
89 
90   setHook("0.preopt", PreOptModuleHook);
91   setHook("1.promote", PostPromoteModuleHook);
92   setHook("2.internalize", PostInternalizeModuleHook);
93   setHook("3.import", PostImportModuleHook);
94   setHook("4.opt", PostOptModuleHook);
95   setHook("5.precodegen", PreCodeGenModuleHook);
96 
97   CombinedIndexHook = [=](const ModuleSummaryIndex &Index) {
98     std::string Path = OutputFileName + "index.bc";
99     std::error_code EC;
100     raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
101     // Because -save-temps is a debugging feature, we report the error
102     // directly and exit.
103     if (EC)
104       reportOpenError(Path, EC.message());
105     WriteIndexToFile(Index, OS);
106     return true;
107   };
108 
109   return Error::success();
110 }
111 
112 namespace {
113 
114 std::unique_ptr<TargetMachine>
115 createTargetMachine(Config &Conf, const Target *TheTarget, Module &M) {
116   StringRef TheTriple = M.getTargetTriple();
117   SubtargetFeatures Features;
118   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
119   for (const std::string &A : Conf.MAttrs)
120     Features.AddFeature(A);
121 
122   Reloc::Model RelocModel;
123   if (Conf.RelocModel)
124     RelocModel = *Conf.RelocModel;
125   else
126     RelocModel =
127         M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_;
128 
129   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
130       TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel,
131       Conf.CodeModel, Conf.CGOptLevel));
132 }
133 
134 static void runNewPMPasses(Module &Mod, TargetMachine *TM, unsigned OptLevel,
135                            bool IsThinLTO) {
136   PassBuilder PB(TM);
137   AAManager AA;
138 
139   // Parse a custom AA pipeline if asked to.
140   assert(PB.parseAAPipeline(AA, "default"));
141 
142   LoopAnalysisManager LAM;
143   FunctionAnalysisManager FAM;
144   CGSCCAnalysisManager CGAM;
145   ModuleAnalysisManager MAM;
146 
147   // Register the AA manager first so that our version is the one used.
148   FAM.registerPass([&] { return std::move(AA); });
149 
150   // Register all the basic analyses with the managers.
151   PB.registerModuleAnalyses(MAM);
152   PB.registerCGSCCAnalyses(CGAM);
153   PB.registerFunctionAnalyses(FAM);
154   PB.registerLoopAnalyses(LAM);
155   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
156 
157   ModulePassManager MPM;
158   // FIXME (davide): verify the input.
159 
160   PassBuilder::OptimizationLevel OL;
161 
162   switch (OptLevel) {
163   default:
164     llvm_unreachable("Invalid optimization level");
165   case 0:
166     OL = PassBuilder::O0;
167     break;
168   case 1:
169     OL = PassBuilder::O1;
170     break;
171   case 2:
172     OL = PassBuilder::O2;
173     break;
174   case 3:
175     OL = PassBuilder::O3;
176     break;
177   }
178 
179   if (IsThinLTO)
180     MPM = PB.buildThinLTODefaultPipeline(OL, false /* DebugLogging */);
181   else
182     MPM = PB.buildLTODefaultPipeline(OL, false /* DebugLogging */);
183   MPM.run(Mod, MAM);
184 
185   // FIXME (davide): verify the output.
186 }
187 
188 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
189                                  std::string PipelineDesc,
190                                  std::string AAPipelineDesc,
191                                  bool DisableVerify) {
192   PassBuilder PB(TM);
193   AAManager AA;
194 
195   // Parse a custom AA pipeline if asked to.
196   if (!AAPipelineDesc.empty())
197     if (!PB.parseAAPipeline(AA, AAPipelineDesc))
198       report_fatal_error("unable to parse AA pipeline description: " +
199                          AAPipelineDesc);
200 
201   LoopAnalysisManager LAM;
202   FunctionAnalysisManager FAM;
203   CGSCCAnalysisManager CGAM;
204   ModuleAnalysisManager MAM;
205 
206   // Register the AA manager first so that our version is the one used.
207   FAM.registerPass([&] { return std::move(AA); });
208 
209   // Register all the basic analyses with the managers.
210   PB.registerModuleAnalyses(MAM);
211   PB.registerCGSCCAnalyses(CGAM);
212   PB.registerFunctionAnalyses(FAM);
213   PB.registerLoopAnalyses(LAM);
214   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
215 
216   ModulePassManager MPM;
217 
218   // Always verify the input.
219   MPM.addPass(VerifierPass());
220 
221   // Now, add all the passes we've been requested to.
222   if (!PB.parsePassPipeline(MPM, PipelineDesc))
223     report_fatal_error("unable to parse pass pipeline description: " +
224                        PipelineDesc);
225 
226   if (!DisableVerify)
227     MPM.addPass(VerifierPass());
228   MPM.run(Mod, MAM);
229 }
230 
231 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
232                            bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
233                            const ModuleSummaryIndex *ImportSummary) {
234   legacy::PassManager passes;
235   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
236 
237   PassManagerBuilder PMB;
238   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
239   PMB.Inliner = createFunctionInliningPass();
240   PMB.ExportSummary = ExportSummary;
241   PMB.ImportSummary = ImportSummary;
242   // Unconditionally verify input since it is not verified before this
243   // point and has unknown origin.
244   PMB.VerifyInput = true;
245   PMB.VerifyOutput = !Conf.DisableVerify;
246   PMB.LoopVectorize = true;
247   PMB.SLPVectorize = true;
248   PMB.OptLevel = Conf.OptLevel;
249   PMB.PGOSampleUse = Conf.SampleProfile;
250   if (IsThinLTO)
251     PMB.populateThinLTOPassManager(passes);
252   else
253     PMB.populateLTOPassManager(passes);
254   passes.run(Mod);
255 }
256 
257 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
258          bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
259          const ModuleSummaryIndex *ImportSummary) {
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 (Conf.UseNewPM)
265     runNewPMPasses(Mod, TM, Conf.OptLevel, IsThinLTO);
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, T, *MPartInCtx);
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 Error lto::backend(Config &C, AddStreamFn AddStream,
357                    unsigned ParallelCodeGenParallelismLevel,
358                    std::unique_ptr<Module> Mod,
359                    ModuleSummaryIndex &CombinedIndex) {
360   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
361   if (!TOrErr)
362     return TOrErr.takeError();
363 
364   std::unique_ptr<TargetMachine> TM = createTargetMachine(C, *TOrErr, *Mod);
365 
366   // Setup optimization remarks.
367   auto DiagFileOrErr = lto::setupOptimizationRemarks(
368       Mod->getContext(), C.RemarksFilename, C.RemarksWithHotness);
369   if (!DiagFileOrErr)
370     return DiagFileOrErr.takeError();
371   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
372 
373   if (!C.CodeGenOnly) {
374     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false,
375              /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr)) {
376       finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
377       return Error::success();
378     }
379   }
380 
381   if (ParallelCodeGenParallelismLevel == 1) {
382     codegen(C, TM.get(), AddStream, 0, *Mod);
383   } else {
384     splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel,
385                  std::move(Mod));
386   }
387   finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
388   return Error::success();
389 }
390 
391 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
392                        Module &Mod, const ModuleSummaryIndex &CombinedIndex,
393                        const FunctionImporter::ImportMapTy &ImportList,
394                        const GVSummaryMapTy &DefinedGlobals,
395                        MapVector<StringRef, BitcodeModule> &ModuleMap) {
396   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
397   if (!TOrErr)
398     return TOrErr.takeError();
399 
400   std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, *TOrErr, Mod);
401 
402   if (Conf.CodeGenOnly) {
403     codegen(Conf, TM.get(), AddStream, Task, Mod);
404     return Error::success();
405   }
406 
407   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
408     return Error::success();
409 
410   renameModuleForThinLTO(Mod, CombinedIndex);
411 
412   thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals);
413 
414   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
415     return Error::success();
416 
417   if (!DefinedGlobals.empty())
418     thinLTOInternalizeModule(Mod, DefinedGlobals);
419 
420   if (Conf.PostInternalizeModuleHook &&
421       !Conf.PostInternalizeModuleHook(Task, Mod))
422     return Error::success();
423 
424   auto ModuleLoader = [&](StringRef Identifier) {
425     assert(Mod.getContext().isODRUniquingDebugTypes() &&
426            "ODR Type uniquing should be enabled on the context");
427     auto I = ModuleMap.find(Identifier);
428     assert(I != ModuleMap.end());
429     return I->second.getLazyModule(Mod.getContext(),
430                                    /*ShouldLazyLoadMetadata=*/true,
431                                    /*IsImporting*/ true);
432   };
433 
434   FunctionImporter Importer(CombinedIndex, ModuleLoader);
435   if (Error Err = Importer.importFunctions(Mod, ImportList).takeError())
436     return Err;
437 
438   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
439     return Error::success();
440 
441   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true,
442            /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex))
443     return Error::success();
444 
445   codegen(Conf, TM.get(), AddStream, Task, Mod);
446   return Error::success();
447 }
448