1 //===-ThinLTOCodeGenerator.cpp - LLVM Link Time Optimizer -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Thin Link Time Optimization library. This library is
11 // intended to be used by linker to optimize code at link time.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/LTO/legacy/ThinLTOCodeGenerator.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringExtras.h"
19 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
20 #include "llvm/Analysis/ProfileSummaryInfo.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Bitcode/BitcodeReader.h"
24 #include "llvm/Bitcode/BitcodeWriter.h"
25 #include "llvm/Bitcode/BitcodeWriterPass.h"
26 #include "llvm/ExecutionEngine/ObjectMemoryBuffer.h"
27 #include "llvm/IR/DebugInfo.h"
28 #include "llvm/IR/DiagnosticPrinter.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/LegacyPassManager.h"
31 #include "llvm/IR/Mangler.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/IRReader/IRReader.h"
34 #include "llvm/LTO/LTO.h"
35 #include "llvm/MC/SubtargetFeature.h"
36 #include "llvm/Object/IRObjectFile.h"
37 #include "llvm/Support/CachePruning.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/Error.h"
40 #include "llvm/Support/Path.h"
41 #include "llvm/Support/SHA1.h"
42 #include "llvm/Support/TargetRegistry.h"
43 #include "llvm/Support/ThreadPool.h"
44 #include "llvm/Support/Threading.h"
45 #include "llvm/Support/ToolOutputFile.h"
46 #include "llvm/Support/VCSRevision.h"
47 #include "llvm/Target/TargetMachine.h"
48 #include "llvm/Transforms/IPO.h"
49 #include "llvm/Transforms/IPO/FunctionImport.h"
50 #include "llvm/Transforms/IPO/Internalize.h"
51 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
52 #include "llvm/Transforms/ObjCARC.h"
53 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
59 #define DEBUG_TYPE "thinlto"
62 // Flags -discard-value-names, defined in LTOCodeGenerator.cpp
63 extern cl::opt<bool> LTODiscardValueNames;
64 extern cl::opt<std::string> LTORemarksFilename;
65 extern cl::opt<bool> LTOPassRemarksWithHotness;
66 extern cl::opt<bool> LTOStripInvalidDebugInfo;
72 ThreadCount("threads", cl::init(llvm::heavyweight_hardware_concurrency()));
74 // Simple helper to save temporary files for debug.
75 static void saveTempBitcode(const Module &TheModule, StringRef TempDir,
76 unsigned count, StringRef Suffix) {
79 // User asked to save temps, let dump the bitcode file after import.
80 std::string SaveTempPath = (TempDir + llvm::utostr(count) + Suffix).str();
82 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
84 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
85 " to save optimized bitcode\n");
86 WriteBitcodeToFile(&TheModule, OS, /* ShouldPreserveUseListOrder */ true);
89 static const GlobalValueSummary *
90 getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) {
91 // If there is any strong definition anywhere, get it.
92 auto StrongDefForLinker = llvm::find_if(
93 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
94 auto Linkage = Summary->linkage();
95 return !GlobalValue::isAvailableExternallyLinkage(Linkage) &&
96 !GlobalValue::isWeakForLinker(Linkage);
98 if (StrongDefForLinker != GVSummaryList.end())
99 return StrongDefForLinker->get();
100 // Get the first *linker visible* definition for this global in the summary
102 auto FirstDefForLinker = llvm::find_if(
103 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
104 auto Linkage = Summary->linkage();
105 return !GlobalValue::isAvailableExternallyLinkage(Linkage);
107 // Extern templates can be emitted as available_externally.
108 if (FirstDefForLinker == GVSummaryList.end())
110 return FirstDefForLinker->get();
113 // Populate map of GUID to the prevailing copy for any multiply defined
114 // symbols. Currently assume first copy is prevailing, or any strong
115 // definition. Can be refined with Linker information in the future.
116 static void computePrevailingCopies(
117 const ModuleSummaryIndex &Index,
118 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy) {
119 auto HasMultipleCopies = [&](const GlobalValueSummaryList &GVSummaryList) {
120 return GVSummaryList.size() > 1;
123 for (auto &I : Index) {
124 if (HasMultipleCopies(I.second.SummaryList))
125 PrevailingCopy[I.first] =
126 getFirstDefinitionForLinker(I.second.SummaryList);
130 static StringMap<MemoryBufferRef>
131 generateModuleMap(const std::vector<ThinLTOBuffer> &Modules) {
132 StringMap<MemoryBufferRef> ModuleMap;
133 for (auto &ModuleBuffer : Modules) {
134 assert(ModuleMap.find(ModuleBuffer.getBufferIdentifier()) ==
136 "Expect unique Buffer Identifier");
137 ModuleMap[ModuleBuffer.getBufferIdentifier()] = ModuleBuffer.getMemBuffer();
142 static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index) {
143 if (renameModuleForThinLTO(TheModule, Index))
144 report_fatal_error("renameModuleForThinLTO failed");
148 class ThinLTODiagnosticInfo : public DiagnosticInfo {
151 ThinLTODiagnosticInfo(const Twine &DiagMsg,
152 DiagnosticSeverity Severity = DS_Error)
153 : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {}
154 void print(DiagnosticPrinter &DP) const override { DP << Msg; }
158 /// Verify the module and strip broken debug info.
159 static void verifyLoadedModule(Module &TheModule) {
160 bool BrokenDebugInfo = false;
161 if (verifyModule(TheModule, &dbgs(),
162 LTOStripInvalidDebugInfo ? &BrokenDebugInfo : nullptr))
163 report_fatal_error("Broken module found, compilation aborted!");
164 if (BrokenDebugInfo) {
165 TheModule.getContext().diagnose(ThinLTODiagnosticInfo(
166 "Invalid debug info found, debug info will be stripped", DS_Warning));
167 StripDebugInfo(TheModule);
171 static std::unique_ptr<Module>
172 loadModuleFromBuffer(const MemoryBufferRef &Buffer, LLVMContext &Context,
173 bool Lazy, bool IsImporting) {
175 Expected<std::unique_ptr<Module>> ModuleOrErr =
177 ? getLazyBitcodeModule(Buffer, Context,
178 /* ShouldLazyLoadMetadata */ true, IsImporting)
179 : parseBitcodeFile(Buffer, Context);
181 handleAllErrors(ModuleOrErr.takeError(), [&](ErrorInfoBase &EIB) {
182 SMDiagnostic Err = SMDiagnostic(Buffer.getBufferIdentifier(),
183 SourceMgr::DK_Error, EIB.message());
184 Err.print("ThinLTO", errs());
186 report_fatal_error("Can't load module, abort.");
189 verifyLoadedModule(*ModuleOrErr.get());
190 return std::move(ModuleOrErr.get());
194 crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index,
195 StringMap<MemoryBufferRef> &ModuleMap,
196 const FunctionImporter::ImportMapTy &ImportList) {
197 auto Loader = [&](StringRef Identifier) {
198 return loadModuleFromBuffer(ModuleMap[Identifier], TheModule.getContext(),
199 /*Lazy=*/true, /*IsImporting*/ true);
202 FunctionImporter Importer(Index, Loader);
203 Expected<bool> Result = Importer.importFunctions(TheModule, ImportList);
205 handleAllErrors(Result.takeError(), [&](ErrorInfoBase &EIB) {
206 SMDiagnostic Err = SMDiagnostic(TheModule.getModuleIdentifier(),
207 SourceMgr::DK_Error, EIB.message());
208 Err.print("ThinLTO", errs());
210 report_fatal_error("importFunctions failed");
212 // Verify again after cross-importing.
213 verifyLoadedModule(TheModule);
216 static void optimizeModule(Module &TheModule, TargetMachine &TM,
217 unsigned OptLevel, bool Freestanding) {
218 // Populate the PassManager
219 PassManagerBuilder PMB;
220 PMB.LibraryInfo = new TargetLibraryInfoImpl(TM.getTargetTriple());
222 PMB.LibraryInfo->disableAllFunctions();
223 PMB.Inliner = createFunctionInliningPass();
224 // FIXME: should get it from the bitcode?
225 PMB.OptLevel = OptLevel;
226 PMB.LoopVectorize = true;
227 PMB.SLPVectorize = true;
228 // Already did this in verifyLoadedModule().
229 PMB.VerifyInput = false;
230 PMB.VerifyOutput = false;
232 legacy::PassManager PM;
234 // Add the TTI (required to inform the vectorizer about register size for
236 PM.add(createTargetTransformInfoWrapperPass(TM.getTargetIRAnalysis()));
239 PMB.populateThinLTOPassManager(PM);
244 // Convert the PreservedSymbols map from "Name" based to "GUID" based.
245 static DenseSet<GlobalValue::GUID>
246 computeGUIDPreservedSymbols(const StringSet<> &PreservedSymbols,
247 const Triple &TheTriple) {
248 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols(PreservedSymbols.size());
249 for (auto &Entry : PreservedSymbols) {
250 StringRef Name = Entry.first();
251 if (TheTriple.isOSBinFormatMachO() && Name.size() > 0 && Name[0] == '_')
252 Name = Name.drop_front();
253 GUIDPreservedSymbols.insert(GlobalValue::getGUID(Name));
255 return GUIDPreservedSymbols;
258 std::unique_ptr<MemoryBuffer> codegenModule(Module &TheModule,
260 SmallVector<char, 128> OutputBuffer;
264 raw_svector_ostream OS(OutputBuffer);
265 legacy::PassManager PM;
267 // If the bitcode files contain ARC code and were compiled with optimization,
268 // the ObjCARCContractPass must be run, so do it unconditionally here.
269 PM.add(createObjCARCContractPass());
271 // Setup the codegen now.
272 if (TM.addPassesToEmitFile(PM, OS, TargetMachine::CGFT_ObjectFile,
273 /* DisableVerify */ true))
274 report_fatal_error("Failed to setup codegen");
276 // Run codegen now. resulting binary is in OutputBuffer.
279 return make_unique<ObjectMemoryBuffer>(std::move(OutputBuffer));
282 /// Manage caching for a single Module.
283 class ModuleCacheEntry {
284 SmallString<128> EntryPath;
287 // Create a cache entry. This compute a unique hash for the Module considering
288 // the current list of export/import, and offer an interface to query to
289 // access the content in the cache.
291 StringRef CachePath, const ModuleSummaryIndex &Index, StringRef ModuleID,
292 const FunctionImporter::ImportMapTy &ImportList,
293 const FunctionImporter::ExportSetTy &ExportList,
294 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
295 const GVSummaryMapTy &DefinedFunctions,
296 const DenseSet<GlobalValue::GUID> &PreservedSymbols, unsigned OptLevel,
297 bool Freestanding, const TargetMachineBuilder &TMBuilder) {
298 if (CachePath.empty())
301 if (!Index.modulePaths().count(ModuleID))
302 // The module does not have an entry, it can't have a hash at all
305 // Compute the unique hash for this entry
306 // This is based on the current compiler version, the module itself, the
307 // export list, the hash for every single module in the import list, the
308 // list of ResolvedODR for the module, and the list of preserved symbols.
310 // Include the hash for the current module
311 auto ModHash = Index.getModuleHash(ModuleID);
313 if (all_of(ModHash, [](uint32_t V) { return V == 0; }))
314 // No hash entry, no caching!
319 // Include the parts of the LTO configuration that affect code generation.
320 auto AddString = [&](StringRef Str) {
322 Hasher.update(ArrayRef<uint8_t>{0});
324 auto AddUnsigned = [&](unsigned I) {
330 Hasher.update(ArrayRef<uint8_t>{Data, 4});
333 // Start with the compiler revision
334 Hasher.update(LLVM_VERSION_STRING);
336 Hasher.update(LLVM_REVISION);
339 // Hash the optimization level and the target machine settings.
340 AddString(TMBuilder.MCpu);
341 // FIXME: Hash more of Options. For now all clients initialize Options from
342 // command-line flags (which is unsupported in production), but may set
343 // RelaxELFRelocations. The clang driver can also pass FunctionSections,
344 // DataSections and DebuggerTuning via command line flags.
345 AddUnsigned(TMBuilder.Options.RelaxELFRelocations);
346 AddUnsigned(TMBuilder.Options.FunctionSections);
347 AddUnsigned(TMBuilder.Options.DataSections);
348 AddUnsigned((unsigned)TMBuilder.Options.DebuggerTuning);
349 AddString(TMBuilder.MAttr);
350 if (TMBuilder.RelocModel)
351 AddUnsigned(*TMBuilder.RelocModel);
352 AddUnsigned(TMBuilder.CGOptLevel);
353 AddUnsigned(OptLevel);
354 AddUnsigned(Freestanding);
356 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
357 for (auto F : ExportList)
358 // The export list can impact the internalization, be conservative here
359 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&F, sizeof(F)));
361 // Include the hash for every module we import functions from
362 for (auto &Entry : ImportList) {
363 auto ModHash = Index.getModuleHash(Entry.first());
364 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
367 // Include the hash for the resolved ODR.
368 for (auto &Entry : ResolvedODR) {
369 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
370 sizeof(GlobalValue::GUID)));
371 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
372 sizeof(GlobalValue::LinkageTypes)));
375 // Include the hash for the preserved symbols.
376 for (auto &Entry : PreservedSymbols) {
377 if (DefinedFunctions.count(Entry))
379 ArrayRef<uint8_t>((const uint8_t *)&Entry, sizeof(GlobalValue::GUID)));
382 // This choice of file name allows the cache to be pruned (see pruneCache()
383 // in include/llvm/Support/CachePruning.h).
384 sys::path::append(EntryPath, CachePath,
385 "llvmcache-" + toHex(Hasher.result()));
388 // Access the path to this entry in the cache.
389 StringRef getEntryPath() { return EntryPath; }
391 // Try loading the buffer for this cache entry.
392 ErrorOr<std::unique_ptr<MemoryBuffer>> tryLoadingBuffer() {
393 if (EntryPath.empty())
394 return std::error_code();
395 return MemoryBuffer::getFile(EntryPath);
398 // Cache the Produced object file
399 void write(const MemoryBuffer &OutputBuffer) {
400 if (EntryPath.empty())
403 // Write to a temporary to avoid race condition
404 SmallString<128> TempFilename;
407 sys::fs::createTemporaryFile("Thin", "tmp.o", TempFD, TempFilename);
409 errs() << "Error: " << EC.message() << "\n";
410 report_fatal_error("ThinLTO: Can't get a temporary file");
413 raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
414 OS << OutputBuffer.getBuffer();
416 // Rename to final destination (hopefully race condition won't matter here)
417 EC = sys::fs::rename(TempFilename, EntryPath);
419 sys::fs::remove(TempFilename);
420 raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
422 report_fatal_error(Twine("Failed to open ") + EntryPath +
423 " to save cached entry\n");
424 OS << OutputBuffer.getBuffer();
429 static std::unique_ptr<MemoryBuffer>
430 ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index,
431 StringMap<MemoryBufferRef> &ModuleMap, TargetMachine &TM,
432 const FunctionImporter::ImportMapTy &ImportList,
433 const FunctionImporter::ExportSetTy &ExportList,
434 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
435 const GVSummaryMapTy &DefinedGlobals,
436 const ThinLTOCodeGenerator::CachingOptions &CacheOptions,
437 bool DisableCodeGen, StringRef SaveTempsDir,
438 bool Freestanding, unsigned OptLevel, unsigned count) {
440 // "Benchmark"-like optimization: single-source case
441 bool SingleModule = (ModuleMap.size() == 1);
444 promoteModule(TheModule, Index);
446 // Apply summary-based LinkOnce/Weak resolution decisions.
447 thinLTOResolveWeakForLinkerModule(TheModule, DefinedGlobals);
449 // Save temps: after promotion.
450 saveTempBitcode(TheModule, SaveTempsDir, count, ".1.promoted.bc");
453 // Be friendly and don't nuke totally the module when the client didn't
454 // supply anything to preserve.
455 if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) {
456 // Apply summary-based internalization decisions.
457 thinLTOInternalizeModule(TheModule, DefinedGlobals);
460 // Save internalized bitcode
461 saveTempBitcode(TheModule, SaveTempsDir, count, ".2.internalized.bc");
464 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
466 // Save temps: after cross-module import.
467 saveTempBitcode(TheModule, SaveTempsDir, count, ".3.imported.bc");
470 optimizeModule(TheModule, TM, OptLevel, Freestanding);
472 saveTempBitcode(TheModule, SaveTempsDir, count, ".4.opt.bc");
474 if (DisableCodeGen) {
475 // Configured to stop before CodeGen, serialize the bitcode and return.
476 SmallVector<char, 128> OutputBuffer;
478 raw_svector_ostream OS(OutputBuffer);
479 ProfileSummaryInfo PSI(TheModule);
480 auto Index = buildModuleSummaryIndex(TheModule, nullptr, &PSI);
481 WriteBitcodeToFile(&TheModule, OS, true, &Index);
483 return make_unique<ObjectMemoryBuffer>(std::move(OutputBuffer));
486 return codegenModule(TheModule, TM);
489 /// Resolve LinkOnce/Weak symbols. Record resolutions in the \p ResolvedODR map
490 /// for caching, and in the \p Index for application during the ThinLTO
491 /// backends. This is needed for correctness for exported symbols (ensure
492 /// at least one copy kept) and a compile-time optimization (to drop duplicate
493 /// copies when possible).
494 static void resolveWeakForLinkerInIndex(
495 ModuleSummaryIndex &Index,
496 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>>
499 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
500 computePrevailingCopies(Index, PrevailingCopy);
502 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
503 const auto &Prevailing = PrevailingCopy.find(GUID);
504 // Not in map means that there was only one copy, which must be prevailing.
505 if (Prevailing == PrevailingCopy.end())
507 return Prevailing->second == S;
510 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
511 GlobalValue::GUID GUID,
512 GlobalValue::LinkageTypes NewLinkage) {
513 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
516 thinLTOResolveWeakForLinkerInIndex(Index, isPrevailing, recordNewLinkage);
519 // Initialize the TargetMachine builder for a given Triple
520 static void initTMBuilder(TargetMachineBuilder &TMBuilder,
521 const Triple &TheTriple) {
522 // Set a default CPU for Darwin triples (copied from LTOCodeGenerator).
523 // FIXME this looks pretty terrible...
524 if (TMBuilder.MCpu.empty() && TheTriple.isOSDarwin()) {
525 if (TheTriple.getArch() == llvm::Triple::x86_64)
526 TMBuilder.MCpu = "core2";
527 else if (TheTriple.getArch() == llvm::Triple::x86)
528 TMBuilder.MCpu = "yonah";
529 else if (TheTriple.getArch() == llvm::Triple::aarch64)
530 TMBuilder.MCpu = "cyclone";
532 TMBuilder.TheTriple = std::move(TheTriple);
535 } // end anonymous namespace
537 void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) {
538 ThinLTOBuffer Buffer(Data, Identifier);
541 ErrorOr<std::string> TripleOrErr = expectedToErrorOrAndEmitErrors(
542 Context, getBitcodeTargetTriple(Buffer.getMemBuffer()));
545 TripleStr = *TripleOrErr;
547 Triple TheTriple(TripleStr);
550 initTMBuilder(TMBuilder, Triple(TheTriple));
551 else if (TMBuilder.TheTriple != TheTriple) {
552 if (!TMBuilder.TheTriple.isCompatibleWith(TheTriple))
553 report_fatal_error("ThinLTO modules with incompatible triples not "
555 initTMBuilder(TMBuilder, Triple(TMBuilder.TheTriple.merge(TheTriple)));
558 Modules.push_back(Buffer);
561 void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) {
562 PreservedSymbols.insert(Name);
565 void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) {
566 // FIXME: At the moment, we don't take advantage of this extra information,
567 // we're conservatively considering cross-references as preserved.
568 // CrossReferencedSymbols.insert(Name);
569 PreservedSymbols.insert(Name);
572 // TargetMachine factory
573 std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const {
575 const Target *TheTarget =
576 TargetRegistry::lookupTarget(TheTriple.str(), ErrMsg);
578 report_fatal_error("Can't load target for this Triple: " + ErrMsg);
581 // Use MAttr as the default set of features.
582 SubtargetFeatures Features(MAttr);
583 Features.getDefaultSubtargetFeatures(TheTriple);
584 std::string FeatureStr = Features.getString();
586 return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
587 TheTriple.str(), MCpu, FeatureStr, Options, RelocModel,
588 CodeModel::Default, CGOptLevel));
592 * Produce the combined summary index from all the bitcode files:
595 std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() {
596 std::unique_ptr<ModuleSummaryIndex> CombinedIndex =
597 llvm::make_unique<ModuleSummaryIndex>();
598 uint64_t NextModuleId = 0;
599 for (auto &ModuleBuffer : Modules) {
600 if (Error Err = readModuleSummaryIndex(ModuleBuffer.getMemBuffer(),
601 *CombinedIndex, NextModuleId++)) {
603 logAllUnhandledErrors(
604 std::move(Err), errs(),
605 "error: can't create module summary index for buffer: ");
609 return CombinedIndex;
613 * Perform promotion and renaming of exported internal functions.
614 * Index is updated to reflect linkage changes from weak resolution.
616 void ThinLTOCodeGenerator::promote(Module &TheModule,
617 ModuleSummaryIndex &Index) {
618 auto ModuleCount = Index.modulePaths().size();
619 auto ModuleIdentifier = TheModule.getModuleIdentifier();
621 // Collect for each module the list of function it defines (GUID -> Summary).
622 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries;
623 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
625 // Convert the preserved symbols set from string to GUID
626 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
627 PreservedSymbols, Triple(TheModule.getTargetTriple()));
629 // Compute "dead" symbols, we don't want to import/export these!
630 computeDeadSymbols(Index, GUIDPreservedSymbols);
632 // Generate import/export list
633 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
634 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
635 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
638 // Resolve LinkOnce/Weak symbols.
639 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
640 resolveWeakForLinkerInIndex(Index, ResolvedODR);
642 thinLTOResolveWeakForLinkerModule(
643 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
645 // Promote the exported values in the index, so that they are promoted
647 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
648 const auto &ExportList = ExportLists.find(ModuleIdentifier);
649 return (ExportList != ExportLists.end() &&
650 ExportList->second.count(GUID)) ||
651 GUIDPreservedSymbols.count(GUID);
653 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
655 promoteModule(TheModule, Index);
659 * Perform cross-module importing for the module identified by ModuleIdentifier.
661 void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule,
662 ModuleSummaryIndex &Index) {
663 auto ModuleMap = generateModuleMap(Modules);
664 auto ModuleCount = Index.modulePaths().size();
666 // Collect for each module the list of function it defines (GUID -> Summary).
667 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
668 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
670 // Convert the preserved symbols set from string to GUID
671 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
672 PreservedSymbols, Triple(TheModule.getTargetTriple()));
674 // Compute "dead" symbols, we don't want to import/export these!
675 computeDeadSymbols(Index, GUIDPreservedSymbols);
677 // Generate import/export list
678 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
679 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
680 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
682 auto &ImportList = ImportLists[TheModule.getModuleIdentifier()];
684 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
688 * Compute the list of summaries needed for importing into module.
690 void ThinLTOCodeGenerator::gatherImportedSummariesForModule(
691 StringRef ModulePath, ModuleSummaryIndex &Index,
692 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
693 auto ModuleCount = Index.modulePaths().size();
695 // Collect for each module the list of function it defines (GUID -> Summary).
696 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
697 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
699 // Generate import/export list
700 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
701 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
702 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
705 llvm::gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
706 ImportLists[ModulePath],
707 ModuleToSummariesForIndex);
711 * Emit the list of files needed for importing into module.
713 void ThinLTOCodeGenerator::emitImports(StringRef ModulePath,
714 StringRef OutputName,
715 ModuleSummaryIndex &Index) {
716 auto ModuleCount = Index.modulePaths().size();
718 // Collect for each module the list of function it defines (GUID -> Summary).
719 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
720 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
722 // Generate import/export list
723 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
724 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
725 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
729 if ((EC = EmitImportsFiles(ModulePath, OutputName, ImportLists[ModulePath])))
730 report_fatal_error(Twine("Failed to open ") + OutputName +
731 " to save imports lists\n");
735 * Perform internalization. Index is updated to reflect linkage changes.
737 void ThinLTOCodeGenerator::internalize(Module &TheModule,
738 ModuleSummaryIndex &Index) {
739 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
740 auto ModuleCount = Index.modulePaths().size();
741 auto ModuleIdentifier = TheModule.getModuleIdentifier();
743 // Convert the preserved symbols set from string to GUID
744 auto GUIDPreservedSymbols =
745 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
747 // Collect for each module the list of function it defines (GUID -> Summary).
748 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
749 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
751 // Compute "dead" symbols, we don't want to import/export these!
752 computeDeadSymbols(Index, GUIDPreservedSymbols);
754 // Generate import/export list
755 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
756 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
757 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
759 auto &ExportList = ExportLists[ModuleIdentifier];
761 // Be friendly and don't nuke totally the module when the client didn't
762 // supply anything to preserve.
763 if (ExportList.empty() && GUIDPreservedSymbols.empty())
767 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
768 const auto &ExportList = ExportLists.find(ModuleIdentifier);
769 return (ExportList != ExportLists.end() &&
770 ExportList->second.count(GUID)) ||
771 GUIDPreservedSymbols.count(GUID);
773 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
774 thinLTOInternalizeModule(TheModule,
775 ModuleToDefinedGVSummaries[ModuleIdentifier]);
779 * Perform post-importing ThinLTO optimizations.
781 void ThinLTOCodeGenerator::optimize(Module &TheModule) {
782 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
785 optimizeModule(TheModule, *TMBuilder.create(), OptLevel, Freestanding);
789 * Perform ThinLTO CodeGen.
791 std::unique_ptr<MemoryBuffer> ThinLTOCodeGenerator::codegen(Module &TheModule) {
792 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
793 return codegenModule(TheModule, *TMBuilder.create());
796 /// Write out the generated object file, either from CacheEntryPath or from
797 /// OutputBuffer, preferring hard-link when possible.
798 /// Returns the path to the generated file in SavedObjectsDirectoryPath.
799 static std::string writeGeneratedObject(int count, StringRef CacheEntryPath,
800 StringRef SavedObjectsDirectoryPath,
801 const MemoryBuffer &OutputBuffer) {
802 SmallString<128> OutputPath(SavedObjectsDirectoryPath);
803 llvm::sys::path::append(OutputPath, Twine(count) + ".thinlto.o");
804 OutputPath.c_str(); // Ensure the string is null terminated.
805 if (sys::fs::exists(OutputPath))
806 sys::fs::remove(OutputPath);
808 // We don't return a memory buffer to the linker, just a list of files.
809 if (!CacheEntryPath.empty()) {
810 // Cache is enabled, hard-link the entry (or copy if hard-link fails).
811 auto Err = sys::fs::create_hard_link(CacheEntryPath, OutputPath);
813 return OutputPath.str();
814 // Hard linking failed, try to copy.
815 Err = sys::fs::copy_file(CacheEntryPath, OutputPath);
817 return OutputPath.str();
818 // Copy failed (could be because the CacheEntry was removed from the cache
819 // in the meantime by another process), fall back and try to write down the
820 // buffer to the output.
821 errs() << "error: can't link or copy from cached entry '" << CacheEntryPath
822 << "' to '" << OutputPath << "'\n";
824 // No cache entry, just write out the buffer.
826 raw_fd_ostream OS(OutputPath, Err, sys::fs::F_None);
828 report_fatal_error("Can't open output '" + OutputPath + "'\n");
829 OS << OutputBuffer.getBuffer();
830 return OutputPath.str();
833 // Main entry point for the ThinLTO processing
834 void ThinLTOCodeGenerator::run() {
835 // Prepare the resulting object vector
836 assert(ProducedBinaries.empty() && "The generator should not be reused");
837 if (SavedObjectsDirectoryPath.empty())
838 ProducedBinaries.resize(Modules.size());
840 sys::fs::create_directories(SavedObjectsDirectoryPath);
842 sys::fs::is_directory(SavedObjectsDirectoryPath, IsDir);
844 report_fatal_error("Unexistent dir: '" + SavedObjectsDirectoryPath + "'");
845 ProducedBinaryFiles.resize(Modules.size());
849 // Perform only parallel codegen and return.
852 for (auto &ModuleBuffer : Modules) {
853 Pool.async([&](int count) {
855 Context.setDiscardValueNames(LTODiscardValueNames);
859 loadModuleFromBuffer(ModuleBuffer.getMemBuffer(), Context, false,
860 /*IsImporting*/ false);
863 auto OutputBuffer = codegen(*TheModule);
864 if (SavedObjectsDirectoryPath.empty())
865 ProducedBinaries[count] = std::move(OutputBuffer);
867 ProducedBinaryFiles[count] = writeGeneratedObject(
868 count, "", SavedObjectsDirectoryPath, *OutputBuffer);
875 // Sequential linking phase
876 auto Index = linkCombinedIndex();
878 // Save temps: index.
879 if (!SaveTempsDir.empty()) {
880 auto SaveTempPath = SaveTempsDir + "index.bc";
882 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
884 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
885 " to save optimized bitcode\n");
886 WriteIndexToFile(*Index, OS);
890 // Prepare the module map.
891 auto ModuleMap = generateModuleMap(Modules);
892 auto ModuleCount = Modules.size();
894 // Collect for each module the list of function it defines (GUID -> Summary).
895 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
896 Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
898 // Convert the preserved symbols set from string to GUID, this is needed for
899 // computing the caching hash and the internalization.
900 auto GUIDPreservedSymbols =
901 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
903 // Compute "dead" symbols, we don't want to import/export these!
904 computeDeadSymbols(*Index, GUIDPreservedSymbols);
906 // Collect the import/export lists for all modules from the call-graph in the
908 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
909 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
910 ComputeCrossModuleImport(*Index, ModuleToDefinedGVSummaries, ImportLists,
913 // We use a std::map here to be able to have a defined ordering when
914 // producing a hash for the cache entry.
915 // FIXME: we should be able to compute the caching hash for the entry based
916 // on the index, and nuke this map.
917 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
919 // Resolve LinkOnce/Weak symbols, this has to be computed early because it
920 // impacts the caching.
921 resolveWeakForLinkerInIndex(*Index, ResolvedODR);
923 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
924 const auto &ExportList = ExportLists.find(ModuleIdentifier);
925 return (ExportList != ExportLists.end() &&
926 ExportList->second.count(GUID)) ||
927 GUIDPreservedSymbols.count(GUID);
930 // Use global summary-based analysis to identify symbols that can be
931 // internalized (because they aren't exported or preserved as per callback).
932 // Changes are made in the index, consumed in the ThinLTO backends.
933 thinLTOInternalizeAndPromoteInIndex(*Index, isExported);
935 // Make sure that every module has an entry in the ExportLists and
936 // ResolvedODR maps to enable threaded access to these maps below.
937 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
938 ExportLists[DefinedGVSummaries.first()];
939 ResolvedODR[DefinedGVSummaries.first()];
942 // Compute the ordering we will process the inputs: the rough heuristic here
943 // is to sort them per size so that the largest module get schedule as soon as
944 // possible. This is purely a compile-time optimization.
945 std::vector<int> ModulesOrdering;
946 ModulesOrdering.resize(Modules.size());
947 std::iota(ModulesOrdering.begin(), ModulesOrdering.end(), 0);
948 std::sort(ModulesOrdering.begin(), ModulesOrdering.end(),
949 [&](int LeftIndex, int RightIndex) {
950 auto LSize = Modules[LeftIndex].getBuffer().size();
951 auto RSize = Modules[RightIndex].getBuffer().size();
952 return LSize > RSize;
955 // Parallel optimizer + codegen
957 ThreadPool Pool(ThreadCount);
958 for (auto IndexCount : ModulesOrdering) {
959 auto &ModuleBuffer = Modules[IndexCount];
960 Pool.async([&](int count) {
961 auto ModuleIdentifier = ModuleBuffer.getBufferIdentifier();
962 auto &ExportList = ExportLists[ModuleIdentifier];
964 auto &DefinedFunctions = ModuleToDefinedGVSummaries[ModuleIdentifier];
966 // The module may be cached, this helps handling it.
967 ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier,
968 ImportLists[ModuleIdentifier], ExportList,
969 ResolvedODR[ModuleIdentifier],
970 DefinedFunctions, GUIDPreservedSymbols,
971 OptLevel, Freestanding, TMBuilder);
972 auto CacheEntryPath = CacheEntry.getEntryPath();
975 auto ErrOrBuffer = CacheEntry.tryLoadingBuffer();
976 DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss") << " '"
977 << CacheEntryPath << "' for buffer " << count << " "
978 << ModuleIdentifier << "\n");
982 if (SavedObjectsDirectoryPath.empty())
983 ProducedBinaries[count] = std::move(ErrOrBuffer.get());
985 ProducedBinaryFiles[count] = writeGeneratedObject(
986 count, CacheEntryPath, SavedObjectsDirectoryPath,
993 Context.setDiscardValueNames(LTODiscardValueNames);
994 Context.enableDebugTypeODRUniquing();
995 auto DiagFileOrErr = lto::setupOptimizationRemarks(
996 Context, LTORemarksFilename, LTOPassRemarksWithHotness, count);
997 if (!DiagFileOrErr) {
998 errs() << "Error: " << toString(DiagFileOrErr.takeError()) << "\n";
999 report_fatal_error("ThinLTO: Can't get an output file for the "
1005 loadModuleFromBuffer(ModuleBuffer.getMemBuffer(), Context, false,
1006 /*IsImporting*/ false);
1008 // Save temps: original file.
1009 saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");
1011 auto &ImportList = ImportLists[ModuleIdentifier];
1012 // Run the main process now, and generates a binary
1013 auto OutputBuffer = ProcessThinLTOModule(
1014 *TheModule, *Index, ModuleMap, *TMBuilder.create(), ImportList,
1015 ExportList, GUIDPreservedSymbols,
1016 ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions,
1017 DisableCodeGen, SaveTempsDir, Freestanding, OptLevel, count);
1019 // Commit to the cache (if enabled)
1020 CacheEntry.write(*OutputBuffer);
1022 if (SavedObjectsDirectoryPath.empty()) {
1023 // We need to generated a memory buffer for the linker.
1024 if (!CacheEntryPath.empty()) {
1025 // Cache is enabled, reload from the cache
1026 // We do this to lower memory pressuree: the buffer is on the heap
1027 // and releasing it frees memory that can be used for the next input
1028 // file. The final binary link will read from the VFS cache
1029 // (hopefully!) or from disk if the memory pressure wasn't too high.
1030 auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
1031 if (auto EC = ReloadedBufferOrErr.getError()) {
1032 // On error, keeping the preexisting buffer and printing a
1033 // diagnostic is more friendly than just crashing.
1034 errs() << "error: can't reload cached file '" << CacheEntryPath
1035 << "': " << EC.message() << "\n";
1037 OutputBuffer = std::move(*ReloadedBufferOrErr);
1040 ProducedBinaries[count] = std::move(OutputBuffer);
1043 ProducedBinaryFiles[count] = writeGeneratedObject(
1044 count, CacheEntryPath, SavedObjectsDirectoryPath, *OutputBuffer);
1049 pruneCache(CacheOptions.Path, CacheOptions.Policy);
1051 // If statistics were requested, print them out now.
1052 if (llvm::AreStatisticsEnabled())
1053 llvm::PrintStatistics();
1054 reportAndResetTimings();