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