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/LoopPassManager.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Bitcode/ReaderWriter.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/Utils/FunctionImportUtils.h"
39 #include "llvm/Transforms/Utils/SplitModule.h"
40 
41 using namespace llvm;
42 using namespace lto;
43 
44 LLVM_ATTRIBUTE_NORETURN void reportOpenError(StringRef Path, Twine Msg) {
45   errs() << "failed to open " << Path << ": " << Msg << '\n';
46   errs().flush();
47   exit(1);
48 }
49 
50 Error Config::addSaveTemps(std::string OutputFileName,
51                            bool UseInputModulePath) {
52   ShouldDiscardValueNames = false;
53 
54   std::error_code EC;
55   ResolutionFile = llvm::make_unique<raw_fd_ostream>(
56       OutputFileName + "resolution.txt", EC, sys::fs::OpenFlags::F_Text);
57   if (EC)
58     return errorCodeToError(EC);
59 
60   auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
61     // Keep track of the hook provided by the linker, which also needs to run.
62     ModuleHookFn LinkerHook = Hook;
63     Hook = [=](unsigned Task, const Module &M) {
64       // If the linker's hook returned false, we need to pass that result
65       // through.
66       if (LinkerHook && !LinkerHook(Task, M))
67         return false;
68 
69       std::string PathPrefix;
70       // If this is the combined module (not a ThinLTO backend compile) or the
71       // user hasn't requested using the input module's path, emit to a file
72       // named from the provided OutputFileName with the Task ID appended.
73       if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
74         PathPrefix = OutputFileName + utostr(Task);
75       } else
76         PathPrefix = M.getModuleIdentifier();
77       std::string Path = PathPrefix + "." + PathSuffix + ".bc";
78       std::error_code EC;
79       raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
80       // Because -save-temps is a debugging feature, we report the error
81       // directly and exit.
82       if (EC)
83         reportOpenError(Path, EC.message());
84       WriteBitcodeToFile(&M, OS, /*ShouldPreserveUseListOrder=*/false);
85       return true;
86     };
87   };
88 
89   setHook("0.preopt", PreOptModuleHook);
90   setHook("1.promote", PostPromoteModuleHook);
91   setHook("2.internalize", PostInternalizeModuleHook);
92   setHook("3.import", PostImportModuleHook);
93   setHook("4.opt", PostOptModuleHook);
94   setHook("5.precodegen", PreCodeGenModuleHook);
95 
96   CombinedIndexHook = [=](const ModuleSummaryIndex &Index) {
97     std::string Path = OutputFileName + "index.bc";
98     std::error_code EC;
99     raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::F_None);
100     // Because -save-temps is a debugging feature, we report the error
101     // directly and exit.
102     if (EC)
103       reportOpenError(Path, EC.message());
104     WriteIndexToFile(Index, OS);
105     return true;
106   };
107 
108   return Error();
109 }
110 
111 namespace {
112 
113 std::unique_ptr<TargetMachine>
114 createTargetMachine(Config &Conf, StringRef TheTriple,
115                     const Target *TheTarget) {
116   SubtargetFeatures Features;
117   Features.getDefaultSubtargetFeatures(Triple(TheTriple));
118   for (const std::string &A : Conf.MAttrs)
119     Features.AddFeature(A);
120 
121   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
122       TheTriple, Conf.CPU, Features.getString(), Conf.Options, Conf.RelocModel,
123       Conf.CodeModel, Conf.CGOptLevel));
124 }
125 
126 static void runNewPMCustomPasses(Module &Mod, TargetMachine *TM,
127                                  std::string PipelineDesc,
128                                  std::string AAPipelineDesc,
129                                  bool DisableVerify) {
130   PassBuilder PB(TM);
131   AAManager AA;
132 
133   // Parse a custom AA pipeline if asked to.
134   if (!AAPipelineDesc.empty())
135     if (!PB.parseAAPipeline(AA, AAPipelineDesc))
136       report_fatal_error("unable to parse AA pipeline description: " +
137                          AAPipelineDesc);
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 
156   // Always verify the input.
157   MPM.addPass(VerifierPass());
158 
159   // Now, add all the passes we've been requested to.
160   if (!PB.parsePassPipeline(MPM, PipelineDesc))
161     report_fatal_error("unable to parse pass pipeline description: " +
162                        PipelineDesc);
163 
164   if (!DisableVerify)
165     MPM.addPass(VerifierPass());
166   MPM.run(Mod, MAM);
167 }
168 
169 static void runOldPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,
170                            bool IsThinLto) {
171   legacy::PassManager passes;
172   passes.add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
173 
174   PassManagerBuilder PMB;
175   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
176   PMB.Inliner = createFunctionInliningPass();
177   // Unconditionally verify input since it is not verified before this
178   // point and has unknown origin.
179   PMB.VerifyInput = true;
180   PMB.VerifyOutput = !Conf.DisableVerify;
181   PMB.LoopVectorize = true;
182   PMB.SLPVectorize = true;
183   PMB.OptLevel = Conf.OptLevel;
184   if (IsThinLto)
185     PMB.populateThinLTOPassManager(passes);
186   else
187     PMB.populateLTOPassManager(passes);
188   passes.run(Mod);
189 }
190 
191 bool opt(Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
192          bool IsThinLto) {
193   Mod.setDataLayout(TM->createDataLayout());
194   if (Conf.OptPipeline.empty())
195     runOldPMPasses(Conf, Mod, TM, IsThinLto);
196   else
197     runNewPMCustomPasses(Mod, TM, Conf.OptPipeline, Conf.AAPipeline,
198                          Conf.DisableVerify);
199   return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
200 }
201 
202 /// Monolithic LTO does not support caching (yet), this is a convenient wrapper
203 /// around AddOutput to workaround this.
204 static AddOutputFn getUncachedOutputWrapper(AddOutputFn &AddOutput,
205                                             unsigned Task) {
206   return [Task, &AddOutput](unsigned TaskId) {
207     auto Output = AddOutput(Task);
208     if (Output->isCachingEnabled() && Output->tryLoadFromCache(""))
209       report_fatal_error("Cache hit without a valid key?");
210     assert(Task == TaskId && "Unexpexted TaskId mismatch");
211     return Output;
212   };
213 }
214 
215 void codegen(Config &Conf, TargetMachine *TM, AddOutputFn AddOutput,
216              unsigned Task, Module &Mod) {
217   if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
218     return;
219 
220   auto Output = AddOutput(Task);
221   std::unique_ptr<raw_pwrite_stream> OS = Output->getStream();
222   legacy::PassManager CodeGenPasses;
223   if (TM->addPassesToEmitFile(CodeGenPasses, *OS,
224                               TargetMachine::CGFT_ObjectFile))
225     report_fatal_error("Failed to setup codegen");
226   CodeGenPasses.run(Mod);
227 }
228 
229 void splitCodeGen(Config &C, TargetMachine *TM, AddOutputFn AddOutput,
230                   unsigned ParallelCodeGenParallelismLevel,
231                   std::unique_ptr<Module> Mod) {
232   ThreadPool CodegenThreadPool(ParallelCodeGenParallelismLevel);
233   unsigned ThreadCount = 0;
234   const Target *T = &TM->getTarget();
235 
236   SplitModule(
237       std::move(Mod), ParallelCodeGenParallelismLevel,
238       [&](std::unique_ptr<Module> MPart) {
239         // We want to clone the module in a new context to multi-thread the
240         // codegen. We do it by serializing partition modules to bitcode
241         // (while still on the main thread, in order to avoid data races) and
242         // spinning up new threads which deserialize the partitions into
243         // separate contexts.
244         // FIXME: Provide a more direct way to do this in LLVM.
245         SmallString<0> BC;
246         raw_svector_ostream BCOS(BC);
247         WriteBitcodeToFile(MPart.get(), BCOS);
248 
249         // Enqueue the task
250         CodegenThreadPool.async(
251             [&](const SmallString<0> &BC, unsigned ThreadId) {
252               LTOLLVMContext Ctx(C);
253               ErrorOr<std::unique_ptr<Module>> MOrErr = parseBitcodeFile(
254                   MemoryBufferRef(StringRef(BC.data(), BC.size()), "ld-temp.o"),
255                   Ctx);
256               if (!MOrErr)
257                 report_fatal_error("Failed to read bitcode");
258               std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
259 
260               std::unique_ptr<TargetMachine> TM =
261                   createTargetMachine(C, MPartInCtx->getTargetTriple(), T);
262 
263               codegen(C, TM.get(),
264                       getUncachedOutputWrapper(AddOutput, ThreadId), ThreadId,
265                       *MPartInCtx);
266             },
267             // Pass BC using std::move to ensure that it get moved rather than
268             // copied into the thread's context.
269             std::move(BC), ThreadCount++);
270       },
271       false);
272 }
273 
274 Expected<const Target *> initAndLookupTarget(Config &C, Module &Mod) {
275   if (!C.OverrideTriple.empty())
276     Mod.setTargetTriple(C.OverrideTriple);
277   else if (Mod.getTargetTriple().empty())
278     Mod.setTargetTriple(C.DefaultTriple);
279 
280   std::string Msg;
281   const Target *T = TargetRegistry::lookupTarget(Mod.getTargetTriple(), Msg);
282   if (!T)
283     return make_error<StringError>(Msg, inconvertibleErrorCode());
284   return T;
285 }
286 
287 }
288 
289 static void handleAsmUndefinedRefs(Module &Mod, TargetMachine &TM) {
290   // Collect the list of undefined symbols used in asm and update
291   // llvm.compiler.used to prevent optimization to drop these from the output.
292   StringSet<> AsmUndefinedRefs;
293   object::IRObjectFile::CollectAsmUndefinedRefs(
294       Triple(Mod.getTargetTriple()), Mod.getModuleInlineAsm(),
295       [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
296         if (Flags & object::BasicSymbolRef::SF_Undefined)
297           AsmUndefinedRefs.insert(Name);
298       });
299   updateCompilerUsed(Mod, TM, AsmUndefinedRefs);
300 }
301 
302 Error lto::backend(Config &C, AddOutputFn AddOutput,
303                    unsigned ParallelCodeGenParallelismLevel,
304                    std::unique_ptr<Module> Mod) {
305   Expected<const Target *> TOrErr = initAndLookupTarget(C, *Mod);
306   if (!TOrErr)
307     return TOrErr.takeError();
308 
309   std::unique_ptr<TargetMachine> TM =
310       createTargetMachine(C, Mod->getTargetTriple(), *TOrErr);
311 
312   handleAsmUndefinedRefs(*Mod, *TM);
313 
314   if (!C.CodeGenOnly)
315     if (!opt(C, TM.get(), 0, *Mod, /*IsThinLto=*/false))
316       return Error();
317 
318   if (ParallelCodeGenParallelismLevel == 1) {
319     codegen(C, TM.get(), getUncachedOutputWrapper(AddOutput, 0), 0, *Mod);
320   } else {
321     splitCodeGen(C, TM.get(), AddOutput, ParallelCodeGenParallelismLevel,
322                  std::move(Mod));
323   }
324   return Error();
325 }
326 
327 Error lto::thinBackend(Config &Conf, unsigned Task, AddOutputFn AddOutput,
328                        Module &Mod, ModuleSummaryIndex &CombinedIndex,
329                        const FunctionImporter::ImportMapTy &ImportList,
330                        const GVSummaryMapTy &DefinedGlobals,
331                        MapVector<StringRef, MemoryBufferRef> &ModuleMap) {
332   Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
333   if (!TOrErr)
334     return TOrErr.takeError();
335 
336   std::unique_ptr<TargetMachine> TM =
337       createTargetMachine(Conf, Mod.getTargetTriple(), *TOrErr);
338 
339   handleAsmUndefinedRefs(Mod, *TM);
340 
341   if (Conf.CodeGenOnly) {
342     codegen(Conf, TM.get(), AddOutput, Task, Mod);
343     return Error();
344   }
345 
346   if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
347     return Error();
348 
349   renameModuleForThinLTO(Mod, CombinedIndex);
350 
351   thinLTOResolveWeakForLinkerModule(Mod, DefinedGlobals);
352 
353   if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
354     return Error();
355 
356   if (!DefinedGlobals.empty())
357     thinLTOInternalizeModule(Mod, DefinedGlobals);
358 
359   if (Conf.PostInternalizeModuleHook &&
360       !Conf.PostInternalizeModuleHook(Task, Mod))
361     return Error();
362 
363   auto ModuleLoader = [&](StringRef Identifier) {
364     assert(Mod.getContext().isODRUniquingDebugTypes() &&
365            "ODR Type uniquing should be enabled on the context");
366     return std::move(getLazyBitcodeModule(MemoryBuffer::getMemBuffer(
367                                               ModuleMap[Identifier], false),
368                                           Mod.getContext(),
369                                           /*ShouldLazyLoadMetadata=*/true)
370                          .get());
371   };
372 
373   FunctionImporter Importer(CombinedIndex, ModuleLoader);
374   Importer.importFunctions(Mod, ImportList);
375 
376   if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
377     return Error();
378 
379   if (!opt(Conf, TM.get(), Task, Mod, /*IsThinLto=*/true))
380     return Error();
381 
382   codegen(Conf, TM.get(), AddOutput, Task, Mod);
383   return Error();
384 }
385