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 static cl::opt<bool>
46     LTOUseNewPM("lto-use-new-pm",
47                 cl::desc("Run LTO passes using the new pass manager"),
48                 cl::init(false), cl::Hidden);
49 
50 LLVM_ATTRIBUTE_NORETURN static void reportOpenError(StringRef Path, Twine Msg) {
51   errs() << "failed to open " << Path << ": " << Msg << '\n';
52   errs().flush();
53   exit(1);
54 }
55 
56 Error Config::addSaveTemps(std::string OutputFileName,
57                            bool UseInputModulePath) {
58   ShouldDiscardValueNames = false;
59 
60   std::error_code EC;
61   ResolutionFile = llvm::make_unique<raw_fd_ostream>(
62       OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text);
63   if (EC)
64     return errorCodeToError(EC);
65 
66   auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
67     // Keep track of the hook provided by the linker, which also needs to run.
68     ModuleHookFn LinkerHook = Hook;
69     Hook = [=](unsigned Task, const Module &M) {
70       // If the linker's hook returned false, we need to pass that result
71       // through.
72       if (LinkerHook && !LinkerHook(Task, M))
73         return false;
74 
75       std::string PathPrefix;
76       // If this is the combined module (not a ThinLTO backend compile) or the
77       // user hasn't requested using the input module's path, emit to a file
78       // named from the provided OutputFileName with the Task ID appended.
79       if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
80         PathPrefix = OutputFileName + 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     return true;
112   };
113 
114   return Error::success();
115 }
116 
117 namespace {
118 
119 std::unique_ptr<TargetMachine>
120 createTargetMachine(Config &Conf, const Target *TheTarget, Module &M) {
121   StringRef TheTriple = M.getTargetTriple();
122   SubtargetFeatures Features;
123   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
124   for (const std::string &A : Conf.MAttrs)
125     Features.AddFeature(A);
126 
127   Reloc::Model RelocModel;
128   if (Conf.RelocModel)
129     RelocModel = *Conf.RelocModel;
130   else
131     RelocModel =
132         M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_;
133 
134   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
135       TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel,
136       Conf.CodeModel, Conf.CGOptLevel));
137 }
138 
139 static void runNewPMPasses(Module &Mod, TargetMachine *TM, unsigned OptLevel) {
140   PassBuilder PB(TM);
141   AAManager AA;
142 
143   // Parse a custom AA pipeline if asked to.
144   assert(PB.parseAAPipeline(AA, "default"));
145 
146   LoopAnalysisManager LAM;
147   FunctionAnalysisManager FAM;
148   CGSCCAnalysisManager CGAM;
149   ModuleAnalysisManager MAM;
150 
151   // Register the AA manager first so that our version is the one used.
152   FAM.registerPass([&] { return std::move(AA); });
153 
154   // Register all the basic analyses with the managers.
155   PB.registerModuleAnalyses(MAM);
156   PB.registerCGSCCAnalyses(CGAM);
157   PB.registerFunctionAnalyses(FAM);
158   PB.registerLoopAnalyses(LAM);
159   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
160 
161   ModulePassManager MPM;
162   // FIXME (davide): verify the input.
163 
164   PassBuilder::OptimizationLevel OL;
165 
166   switch (OptLevel) {
167   default:
168     llvm_unreachable("Invalid optimization level");
169   case 0:
170     OL = PassBuilder::O0;
171     break;
172   case 1:
173     OL = PassBuilder::O1;
174     break;
175   case 2:
176     OL = PassBuilder::O2;
177     break;
178   case 3:
179     OL = PassBuilder::O3;
180     break;
181   }
182 
183   MPM = PB.buildLTODefaultPipeline(OL, false /* DebugLogging */);
184   MPM.run(Mod, MAM);
185 
186   // FIXME (davide): verify the output.
187 }
188 
189 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
190                                  std::string PipelineDesc,
191                                  std::string AAPipelineDesc,
192                                  bool DisableVerify) {
193   PassBuilder PB(TM);
194   AAManager AA;
195 
196   // Parse a custom AA pipeline if asked to.
197   if (!AAPipelineDesc.empty())
198     if (!PB.parseAAPipeline(AA, AAPipelineDesc))
199       report_fatal_error("unable to parse AA pipeline description: " +
200                          AAPipelineDesc);
201 
202   LoopAnalysisManager LAM;
203   FunctionAnalysisManager FAM;
204   CGSCCAnalysisManager CGAM;
205   ModuleAnalysisManager MAM;
206 
207   // Register the AA manager first so that our version is the one used.
208   FAM.registerPass([&] { return std::move(AA); });
209 
210   // Register all the basic analyses with the managers.
211   PB.registerModuleAnalyses(MAM);
212   PB.registerCGSCCAnalyses(CGAM);
213   PB.registerFunctionAnalyses(FAM);
214   PB.registerLoopAnalyses(LAM);
215   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
216 
217   ModulePassManager MPM;
218 
219   // Always verify the input.
220   MPM.addPass(VerifierPass());
221 
222   // Now, add all the passes we've been requested to.
223   if (!PB.parsePassPipeline(MPM, PipelineDesc))
224     report_fatal_error("unable to parse pass pipeline description: " +
225                        PipelineDesc);
226 
227   if (!DisableVerify)
228     MPM.addPass(VerifierPass());
229   MPM.run(Mod, MAM);
230 }
231 
232 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
233                            bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
234                            const ModuleSummaryIndex *ImportSummary) {
235   legacy::PassManager passes;
236   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
237 
238   PassManagerBuilder PMB;
239   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
240   PMB.Inliner = createFunctionInliningPass();
241   PMB.ExportSummary = ExportSummary;
242   PMB.ImportSummary = ImportSummary;
243   // Unconditionally verify input since it is not verified before this
244   // point and has unknown origin.
245   PMB.VerifyInput = true;
246   PMB.VerifyOutput = !Conf.DisableVerify;
247   PMB.LoopVectorize = true;
248   PMB.SLPVectorize = true;
249   PMB.OptLevel = Conf.OptLevel;
250   PMB.PGOSampleUse = Conf.SampleProfile;
251   if (IsThinLTO)
252     PMB.populateThinLTOPassManager(passes);
253   else
254     PMB.populateLTOPassManager(passes);
255   passes.run(Mod);
256 }
257 
258 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
259          bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
260          const ModuleSummaryIndex *ImportSummary) {
261   // There's still no ThinLTO pipeline hooked up in the new pass manager,
262   // once there is one, we can just remove this.
263   if (LTOUseNewPM && IsThinLTO)
264     report_fatal_error("ThinLTO not supported with the new PM yet!");
265 
266   // FIXME: Plumb the combined index into the new pass manager.
267   if (!Conf.OptPipeline.empty())
268     runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline,
269                          Conf.DisableVerify);
270   else if (LTOUseNewPM)
271     runNewPMPasses(Mod, TM, Conf.OptLevel);
272   else
273     runOldPMPasses(Conf, Mod, TM, IsThinLTO, ExportSummary, ImportSummary);
274   return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
275 }
276 
277 void codegen(Config &Conf, TargetMachine *TM, AddStreamFn AddStream,
278              unsigned Task, Module &Mod) {
279   if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
280     return;
281 
282   auto Stream = AddStream(Task);
283   legacy::PassManager CodeGenPasses;
284   if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS, Conf.CGFileType))
285     report_fatal_error("Failed to setup codegen");
286   CodeGenPasses.run(Mod);
287 }
288 
289 void splitCodeGen(Config &C, TargetMachine *TM, AddStreamFn AddStream,
290                   unsigned ParallelCodeGenParallelismLevel,
291                   std::unique_ptr<Module> Mod) {
292   ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel);
293   unsigned ThreadCount = 0;
294   const Target *T = &TM->getTarget();
295 
296   SplitModule(
297       std::move(Mod), ParallelCodeGenParallelismLevel,
298       [&](std::unique_ptr<Module> MPart) {
299         // We want to clone the module in a new context to multi-thread the
300         // codegen. We do it by serializing partition modules to bitcode
301         // (while still on the main thread, in order to avoid data races) and
302         // spinning up new threads which deserialize the partitions into
303         // separate contexts.
304         // FIXME: Provide a more direct way to do this in LLVM.
305         SmallString<0> BC;
306         raw_svector_ostream BCOS(BC);
307         WriteBitcodeToFile(MPart.get(), BCOS);
308 
309         // Enqueue the task
310         CodegenThreadPool.async(
311             [&](const SmallString<0> &BC, unsigned ThreadId) {
312               LTOLLVMContext Ctx(C);
313               Expected<std::unique_ptr<Module>> MOrErr = parseBitcodeFile(
314                   MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"),
315                   Ctx);
316               if (!MOrErr)
317                 report_fatal_error("Failed to read bitcode");
318               std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
319 
320               std::unique_ptr<TargetMachine> TM =
321                   createTargetMachine(C, T, *MPartInCtx);
322 
323               codegen(C, TM.get(), AddStream, ThreadId, *MPartInCtx);
324             },
325             // Pass BC using std::move to ensure that it get moved rather than
326             // copied into the thread's context.
327             std::move(BC), ThreadCount++);
328       },
329       false);
330 
331   // Because the inner lambda (which runs in a worker thread) captures our local
332   // variables, we need to wait for the worker threads to terminate before we
333   // can leave the function scope.
334   CodegenThreadPool.wait();
335 }
336 
337 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) {
338   if (!C.OverrideTriple.empty())
339     Mod.setTargetTriple(C.OverrideTriple);
340   else if (Mod.getTargetTriple().empty())
341     Mod.setTargetTriple(C.DefaultTriple);
342 
343   std::string Msg;
344   const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg);
345   if (!T)
346     return make_error<StringError>(Msg, inconvertibleErrorCode());
347   return T;
348 }
349 
350 }
351 
352 static void
353 finalizeOptimizationRemarks(std::unique_ptr<tool_output_file> DiagOutputFile) {
354   // Make sure we flush the diagnostic remarks file in case the linker doesn't
355   // call the global destructors before exiting.
356   if (!DiagOutputFile)
357     return;
358   DiagOutputFile->keep();
359   DiagOutputFile->os().flush();
360 }
361 
362 Error lto::backend(Config &C, AddStreamFn AddStream,
363                    unsigned ParallelCodeGenParallelismLevel,
364                    std::unique_ptr<Module> Mod,
365                    ModuleSummaryIndex &CombinedIndex) {
366   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
367   if (!TOrErr)
368     return TOrErr.takeError();
369 
370   std::unique_ptr<TargetMachine> TM = createTargetMachine(C, *TOrErr, *Mod);
371 
372   // Setup optimization remarks.
373   auto DiagFileOrErr = lto::setupOptimizationRemarks(
374       Mod->getContext(), C.RemarksFilename, C.RemarksWithHotness);
375   if (!DiagFileOrErr)
376     return DiagFileOrErr.takeError();
377   auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
378 
379   if (!C.CodeGenOnly) {
380     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLTO=*/false,
381              /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr)) {
382       finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
383       return Error::success();
384     }
385   }
386 
387   if (ParallelCodeGenParallelismLevel == 1) {
388     codegen(C, TM.get(), AddStream, 0, *Mod);
389   } else {
390     splitCodeGen(C, TM.get(), AddStream, ParallelCodeGenParallelismLevel,
391                  std::move(Mod));
392   }
393   finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
394   return Error::success();
395 }
396 
397 Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
398                        Module &Mod, const ModuleSummaryIndex &CombinedIndex,
399                        const FunctionImporter::ImportMapTy &ImportList,
400                        const GVSummaryMapTy &DefinedGlobals,
401                        MapVector<StringRef, BitcodeModule> &ModuleMap) {
402   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
403   if (!TOrErr)
404     return TOrErr.takeError();
405 
406   std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, *TOrErr, Mod);
407 
408   if (Conf.CodeGenOnly) {
409     codegen(Conf, TM.get(), AddStream, Task, Mod);
410     return Error::success();
411   }
412 
413   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
414     return Error::success();
415 
416   renameModuleForThinLTO(Mod, CombinedIndex);
417 
418   thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals);
419 
420   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
421     return Error::success();
422 
423   if (!DefinedGlobals.empty())
424     thinLTOInternalizeModule(Mod, DefinedGlobals);
425 
426   if (Conf.PostInternalizeModuleHook &&
427       !Conf.PostInternalizeModuleHook(Task, Mod))
428     return Error::success();
429 
430   auto ModuleLoader = [&](StringRef Identifier) {
431     assert(Mod.getContext().isODRUniquingDebugTypes() &&
432            "ODR Type uniquing should be enabled on the context");
433     auto I = ModuleMap.find(Identifier);
434     assert(I != ModuleMap.end());
435     return I->second.getLazyModule(Mod.getContext(),
436                                    /*ShouldLazyLoadMetadata=*/true,
437                                    /*IsImporting*/ true);
438   };
439 
440   FunctionImporter Importer(CombinedIndex, ModuleLoader);
441   if (Error Err = Importer.importFunctions(Mod, ImportList).takeError())
442     return Err;
443 
444   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
445     return Error::success();
446 
447   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLTO=*/true,
448            /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex))
449     return Error::success();
450 
451   codegen(Conf, TM.get(), AddStream, Task, Mod);
452   return Error::success();
453 }
454