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 #ifdef HAVE_LLVM_REVISION
18 #include "LLVMLTORevision.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/ADT/StringExtras.h"
23 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
24 #include "llvm/Analysis/ProfileSummaryInfo.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Bitcode/BitcodeReader.h"
28 #include "llvm/Bitcode/BitcodeWriter.h"
29 #include "llvm/Bitcode/BitcodeWriterPass.h"
30 #include "llvm/ExecutionEngine/ObjectMemoryBuffer.h"
31 #include "llvm/IR/DiagnosticPrinter.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/LegacyPassManager.h"
34 #include "llvm/IR/Mangler.h"
35 #include "llvm/IRReader/IRReader.h"
36 #include "llvm/LTO/LTO.h"
37 #include "llvm/Linker/Linker.h"
38 #include "llvm/MC/SubtargetFeature.h"
39 #include "llvm/Object/IRObjectFile.h"
40 #include "llvm/Object/ModuleSummaryIndexObjectFile.h"
41 #include "llvm/Support/CachePruning.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/Error.h"
44 #include "llvm/Support/Path.h"
45 #include "llvm/Support/SHA1.h"
46 #include "llvm/Support/TargetRegistry.h"
47 #include "llvm/Support/ThreadPool.h"
48 #include "llvm/Support/Threading.h"
49 #include "llvm/Support/ToolOutputFile.h"
50 #include "llvm/Target/TargetMachine.h"
51 #include "llvm/Transforms/IPO.h"
52 #include "llvm/Transforms/IPO/FunctionImport.h"
53 #include "llvm/Transforms/IPO/Internalize.h"
54 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
55 #include "llvm/Transforms/ObjCARC.h"
56 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
62 #define DEBUG_TYPE "thinlto"
65 // Flags -discard-value-names, defined in LTOCodeGenerator.cpp
66 extern cl::opt<bool> LTODiscardValueNames;
67 extern cl::opt<std::string> LTORemarksFilename;
68 extern cl::opt<bool> LTOPassRemarksWithHotness;
74 ThreadCount("threads", cl::init(llvm::heavyweight_hardware_concurrency()));
76 Expected<std::unique_ptr<tool_output_file>>
77 setupOptimizationRemarks(LLVMContext &Ctx, int Count) {
78 if (LTOPassRemarksWithHotness)
79 Ctx.setDiagnosticHotnessRequested(true);
81 if (LTORemarksFilename.empty())
84 std::string FileName =
85 LTORemarksFilename + ".thin." + llvm::utostr(Count) + ".yaml";
87 auto DiagnosticOutputFile =
88 llvm::make_unique<tool_output_file>(FileName, EC, sys::fs::F_None);
90 return errorCodeToError(EC);
91 Ctx.setDiagnosticsOutputFile(
92 llvm::make_unique<yaml::Output>(DiagnosticOutputFile->os()));
93 DiagnosticOutputFile->keep();
94 return std::move(DiagnosticOutputFile);
97 // Simple helper to save temporary files for debug.
98 static void saveTempBitcode(const Module &TheModule, StringRef TempDir,
99 unsigned count, StringRef Suffix) {
102 // User asked to save temps, let dump the bitcode file after import.
103 std::string SaveTempPath = (TempDir + llvm::utostr(count) + Suffix).str();
105 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
107 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
108 " to save optimized bitcode\n");
109 WriteBitcodeToFile(&TheModule, OS, /* ShouldPreserveUseListOrder */ true);
112 static const GlobalValueSummary *
113 getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) {
114 // If there is any strong definition anywhere, get it.
115 auto StrongDefForLinker = llvm::find_if(
116 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
117 auto Linkage = Summary->linkage();
118 return !GlobalValue::isAvailableExternallyLinkage(Linkage) &&
119 !GlobalValue::isWeakForLinker(Linkage);
121 if (StrongDefForLinker != GVSummaryList.end())
122 return StrongDefForLinker->get();
123 // Get the first *linker visible* definition for this global in the summary
125 auto FirstDefForLinker = llvm::find_if(
126 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
127 auto Linkage = Summary->linkage();
128 return !GlobalValue::isAvailableExternallyLinkage(Linkage);
130 // Extern templates can be emitted as available_externally.
131 if (FirstDefForLinker == GVSummaryList.end())
133 return FirstDefForLinker->get();
136 // Populate map of GUID to the prevailing copy for any multiply defined
137 // symbols. Currently assume first copy is prevailing, or any strong
138 // definition. Can be refined with Linker information in the future.
139 static void computePrevailingCopies(
140 const ModuleSummaryIndex &Index,
141 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy) {
142 auto HasMultipleCopies = [&](const GlobalValueSummaryList &GVSummaryList) {
143 return GVSummaryList.size() > 1;
146 for (auto &I : Index) {
147 if (HasMultipleCopies(I.second))
148 PrevailingCopy[I.first] = getFirstDefinitionForLinker(I.second);
152 static StringMap<MemoryBufferRef>
153 generateModuleMap(const std::vector<MemoryBufferRef> &Modules) {
154 StringMap<MemoryBufferRef> ModuleMap;
155 for (auto &ModuleBuffer : Modules) {
156 assert(ModuleMap.find(ModuleBuffer.getBufferIdentifier()) ==
158 "Expect unique Buffer Identifier");
159 ModuleMap[ModuleBuffer.getBufferIdentifier()] = ModuleBuffer;
164 static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index) {
165 if (renameModuleForThinLTO(TheModule, Index))
166 report_fatal_error("renameModuleForThinLTO failed");
169 static std::unique_ptr<Module>
170 loadModuleFromBuffer(const MemoryBufferRef &Buffer, LLVMContext &Context,
171 bool Lazy, bool IsImporting) {
173 Expected<std::unique_ptr<Module>> ModuleOrErr =
175 ? getLazyBitcodeModule(Buffer, Context,
176 /* ShouldLazyLoadMetadata */ true, IsImporting)
177 : parseBitcodeFile(Buffer, Context);
179 handleAllErrors(ModuleOrErr.takeError(), [&](ErrorInfoBase &EIB) {
180 SMDiagnostic Err = SMDiagnostic(Buffer.getBufferIdentifier(),
181 SourceMgr::DK_Error, EIB.message());
182 Err.print("ThinLTO", errs());
184 report_fatal_error("Can't load module, abort.");
186 return std::move(ModuleOrErr.get());
190 crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index,
191 StringMap<MemoryBufferRef> &ModuleMap,
192 const FunctionImporter::ImportMapTy &ImportList) {
193 auto Loader = [&](StringRef Identifier) {
194 return loadModuleFromBuffer(ModuleMap[Identifier], TheModule.getContext(),
195 /*Lazy=*/true, /*IsImporting*/ true);
198 FunctionImporter Importer(Index, Loader);
199 if (!Importer.importFunctions(TheModule, ImportList))
200 report_fatal_error("importFunctions failed");
203 static void optimizeModule(Module &TheModule, TargetMachine &TM,
205 // Populate the PassManager
206 PassManagerBuilder PMB;
207 PMB.LibraryInfo = new TargetLibraryInfoImpl(TM.getTargetTriple());
208 PMB.Inliner = createFunctionInliningPass();
209 // FIXME: should get it from the bitcode?
210 PMB.OptLevel = OptLevel;
211 PMB.LoopVectorize = true;
212 PMB.SLPVectorize = true;
213 PMB.VerifyInput = true;
214 PMB.VerifyOutput = false;
216 legacy::PassManager PM;
218 // Add the TTI (required to inform the vectorizer about register size for
220 PM.add(createTargetTransformInfoWrapperPass(TM.getTargetIRAnalysis()));
223 PMB.populateThinLTOPassManager(PM);
228 // Convert the PreservedSymbols map from "Name" based to "GUID" based.
229 static DenseSet<GlobalValue::GUID>
230 computeGUIDPreservedSymbols(const StringSet<> &PreservedSymbols,
231 const Triple &TheTriple) {
232 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols(PreservedSymbols.size());
233 for (auto &Entry : PreservedSymbols) {
234 StringRef Name = Entry.first();
235 if (TheTriple.isOSBinFormatMachO() && Name.size() > 0 && Name[0] == '_')
236 Name = Name.drop_front();
237 GUIDPreservedSymbols.insert(GlobalValue::getGUID(Name));
239 return GUIDPreservedSymbols;
242 std::unique_ptr<MemoryBuffer> codegenModule(Module &TheModule,
244 SmallVector<char, 128> OutputBuffer;
248 raw_svector_ostream OS(OutputBuffer);
249 legacy::PassManager PM;
251 // If the bitcode files contain ARC code and were compiled with optimization,
252 // the ObjCARCContractPass must be run, so do it unconditionally here.
253 PM.add(createObjCARCContractPass());
255 // Setup the codegen now.
256 if (TM.addPassesToEmitFile(PM, OS, TargetMachine::CGFT_ObjectFile,
257 /* DisableVerify */ true))
258 report_fatal_error("Failed to setup codegen");
260 // Run codegen now. resulting binary is in OutputBuffer.
263 return make_unique<ObjectMemoryBuffer>(std::move(OutputBuffer));
266 /// Manage caching for a single Module.
267 class ModuleCacheEntry {
268 SmallString<128> EntryPath;
271 // Create a cache entry. This compute a unique hash for the Module considering
272 // the current list of export/import, and offer an interface to query to
273 // access the content in the cache.
275 StringRef CachePath, const ModuleSummaryIndex &Index, StringRef ModuleID,
276 const FunctionImporter::ImportMapTy &ImportList,
277 const FunctionImporter::ExportSetTy &ExportList,
278 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
279 const GVSummaryMapTy &DefinedFunctions,
280 const DenseSet<GlobalValue::GUID> &PreservedSymbols) {
281 if (CachePath.empty())
284 if (!Index.modulePaths().count(ModuleID))
285 // The module does not have an entry, it can't have a hash at all
288 // Compute the unique hash for this entry
289 // This is based on the current compiler version, the module itself, the
290 // export list, the hash for every single module in the import list, the
291 // list of ResolvedODR for the module, and the list of preserved symbols.
293 // Include the hash for the current module
294 auto ModHash = Index.getModuleHash(ModuleID);
296 if (all_of(ModHash, [](uint32_t V) { return V == 0; }))
297 // No hash entry, no caching!
302 // Start with the compiler revision
303 Hasher.update(LLVM_VERSION_STRING);
304 #ifdef HAVE_LLVM_REVISION
305 Hasher.update(LLVM_REVISION);
308 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
309 for (auto F : ExportList)
310 // The export list can impact the internalization, be conservative here
311 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&F, sizeof(F)));
313 // Include the hash for every module we import functions from
314 for (auto &Entry : ImportList) {
315 auto ModHash = Index.getModuleHash(Entry.first());
316 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
319 // Include the hash for the resolved ODR.
320 for (auto &Entry : ResolvedODR) {
321 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
322 sizeof(GlobalValue::GUID)));
323 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
324 sizeof(GlobalValue::LinkageTypes)));
327 // Include the hash for the preserved symbols.
328 for (auto &Entry : PreservedSymbols) {
329 if (DefinedFunctions.count(Entry))
331 ArrayRef<uint8_t>((const uint8_t *)&Entry, sizeof(GlobalValue::GUID)));
334 sys::path::append(EntryPath, CachePath, toHex(Hasher.result()));
337 // Access the path to this entry in the cache.
338 StringRef getEntryPath() { return EntryPath; }
340 // Try loading the buffer for this cache entry.
341 ErrorOr<std::unique_ptr<MemoryBuffer>> tryLoadingBuffer() {
342 if (EntryPath.empty())
343 return std::error_code();
344 return MemoryBuffer::getFile(EntryPath);
347 // Cache the Produced object file
348 void write(const MemoryBuffer &OutputBuffer) {
349 if (EntryPath.empty())
352 // Write to a temporary to avoid race condition
353 SmallString<128> TempFilename;
356 sys::fs::createTemporaryFile("Thin", "tmp.o", TempFD, TempFilename);
358 errs() << "Error: " << EC.message() << "\n";
359 report_fatal_error("ThinLTO: Can't get a temporary file");
362 raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
363 OS << OutputBuffer.getBuffer();
365 // Rename to final destination (hopefully race condition won't matter here)
366 EC = sys::fs::rename(TempFilename, EntryPath);
368 sys::fs::remove(TempFilename);
369 raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
371 report_fatal_error(Twine("Failed to open ") + EntryPath +
372 " to save cached entry\n");
373 OS << OutputBuffer.getBuffer();
378 static std::unique_ptr<MemoryBuffer>
379 ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index,
380 StringMap<MemoryBufferRef> &ModuleMap, TargetMachine &TM,
381 const FunctionImporter::ImportMapTy &ImportList,
382 const FunctionImporter::ExportSetTy &ExportList,
383 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
384 const GVSummaryMapTy &DefinedGlobals,
385 const ThinLTOCodeGenerator::CachingOptions &CacheOptions,
386 bool DisableCodeGen, StringRef SaveTempsDir,
387 unsigned OptLevel, unsigned count) {
389 // "Benchmark"-like optimization: single-source case
390 bool SingleModule = (ModuleMap.size() == 1);
393 promoteModule(TheModule, Index);
395 // Apply summary-based LinkOnce/Weak resolution decisions.
396 thinLTOResolveWeakForLinkerModule(TheModule, DefinedGlobals);
398 // Save temps: after promotion.
399 saveTempBitcode(TheModule, SaveTempsDir, count, ".1.promoted.bc");
402 // Be friendly and don't nuke totally the module when the client didn't
403 // supply anything to preserve.
404 if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) {
405 // Apply summary-based internalization decisions.
406 thinLTOInternalizeModule(TheModule, DefinedGlobals);
409 // Save internalized bitcode
410 saveTempBitcode(TheModule, SaveTempsDir, count, ".2.internalized.bc");
413 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
415 // Save temps: after cross-module import.
416 saveTempBitcode(TheModule, SaveTempsDir, count, ".3.imported.bc");
419 optimizeModule(TheModule, TM, OptLevel);
421 saveTempBitcode(TheModule, SaveTempsDir, count, ".4.opt.bc");
423 if (DisableCodeGen) {
424 // Configured to stop before CodeGen, serialize the bitcode and return.
425 SmallVector<char, 128> OutputBuffer;
427 raw_svector_ostream OS(OutputBuffer);
428 ProfileSummaryInfo PSI(TheModule);
429 auto Index = buildModuleSummaryIndex(TheModule, nullptr, nullptr);
430 WriteBitcodeToFile(&TheModule, OS, true, &Index);
432 return make_unique<ObjectMemoryBuffer>(std::move(OutputBuffer));
435 return codegenModule(TheModule, TM);
438 /// Resolve LinkOnce/Weak symbols. Record resolutions in the \p ResolvedODR map
439 /// for caching, and in the \p Index for application during the ThinLTO
440 /// backends. This is needed for correctness for exported symbols (ensure
441 /// at least one copy kept) and a compile-time optimization (to drop duplicate
442 /// copies when possible).
443 static void resolveWeakForLinkerInIndex(
444 ModuleSummaryIndex &Index,
445 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>>
448 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
449 computePrevailingCopies(Index, PrevailingCopy);
451 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
452 const auto &Prevailing = PrevailingCopy.find(GUID);
453 // Not in map means that there was only one copy, which must be prevailing.
454 if (Prevailing == PrevailingCopy.end())
456 return Prevailing->second == S;
459 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
460 GlobalValue::GUID GUID,
461 GlobalValue::LinkageTypes NewLinkage) {
462 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
465 thinLTOResolveWeakForLinkerInIndex(Index, isPrevailing, recordNewLinkage);
468 // Initialize the TargetMachine builder for a given Triple
469 static void initTMBuilder(TargetMachineBuilder &TMBuilder,
470 const Triple &TheTriple) {
471 // Set a default CPU for Darwin triples (copied from LTOCodeGenerator).
472 // FIXME this looks pretty terrible...
473 if (TMBuilder.MCpu.empty() && TheTriple.isOSDarwin()) {
474 if (TheTriple.getArch() == llvm::Triple::x86_64)
475 TMBuilder.MCpu = "core2";
476 else if (TheTriple.getArch() == llvm::Triple::x86)
477 TMBuilder.MCpu = "yonah";
478 else if (TheTriple.getArch() == llvm::Triple::aarch64)
479 TMBuilder.MCpu = "cyclone";
481 TMBuilder.TheTriple = std::move(TheTriple);
484 } // end anonymous namespace
486 void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) {
487 MemoryBufferRef Buffer(Data, Identifier);
488 if (Modules.empty()) {
489 // First module added, so initialize the triple and some options
492 ErrorOr<std::string> TripleOrErr =
493 expectedToErrorOrAndEmitErrors(Context, getBitcodeTargetTriple(Buffer));
495 TripleStr = *TripleOrErr;
496 Triple TheTriple(TripleStr);
497 initTMBuilder(TMBuilder, Triple(TheTriple));
503 ErrorOr<std::string> TripleOrErr =
504 expectedToErrorOrAndEmitErrors(Context, getBitcodeTargetTriple(Buffer));
506 TripleStr = *TripleOrErr;
507 assert(TMBuilder.TheTriple.str() == TripleStr &&
508 "ThinLTO modules with different triple not supported");
511 Modules.push_back(Buffer);
514 void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) {
515 PreservedSymbols.insert(Name);
518 void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) {
519 // FIXME: At the moment, we don't take advantage of this extra information,
520 // we're conservatively considering cross-references as preserved.
521 // CrossReferencedSymbols.insert(Name);
522 PreservedSymbols.insert(Name);
525 // TargetMachine factory
526 std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const {
528 const Target *TheTarget =
529 TargetRegistry::lookupTarget(TheTriple.str(), ErrMsg);
531 report_fatal_error("Can't load target for this Triple: " + ErrMsg);
534 // Use MAttr as the default set of features.
535 SubtargetFeatures Features(MAttr);
536 Features.getDefaultSubtargetFeatures(TheTriple);
537 std::string FeatureStr = Features.getString();
539 return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
540 TheTriple.str(), MCpu, FeatureStr, Options, RelocModel,
541 CodeModel::Default, CGOptLevel));
545 * Produce the combined summary index from all the bitcode files:
548 std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() {
549 std::unique_ptr<ModuleSummaryIndex> CombinedIndex;
550 uint64_t NextModuleId = 0;
551 for (auto &ModuleBuffer : Modules) {
552 Expected<std::unique_ptr<object::ModuleSummaryIndexObjectFile>> ObjOrErr =
553 object::ModuleSummaryIndexObjectFile::create(ModuleBuffer);
556 logAllUnhandledErrors(
557 ObjOrErr.takeError(), errs(),
558 "error: can't create ModuleSummaryIndexObjectFile for buffer: ");
561 auto Index = (*ObjOrErr)->takeIndex();
563 CombinedIndex->mergeFrom(std::move(Index), ++NextModuleId);
565 CombinedIndex = std::move(Index);
568 return CombinedIndex;
572 * Perform promotion and renaming of exported internal functions.
573 * Index is updated to reflect linkage changes from weak resolution.
575 void ThinLTOCodeGenerator::promote(Module &TheModule,
576 ModuleSummaryIndex &Index) {
577 auto ModuleCount = Index.modulePaths().size();
578 auto ModuleIdentifier = TheModule.getModuleIdentifier();
580 // Collect for each module the list of function it defines (GUID -> Summary).
581 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries;
582 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
584 // Convert the preserved symbols set from string to GUID
585 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
586 PreservedSymbols, Triple(TheModule.getTargetTriple()));
588 // Compute "dead" symbols, we don't want to import/export these!
589 auto DeadSymbols = computeDeadSymbols(Index, GUIDPreservedSymbols);
591 // Generate import/export list
592 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
593 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
594 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
595 ExportLists, &DeadSymbols);
597 // Resolve LinkOnce/Weak symbols.
598 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
599 resolveWeakForLinkerInIndex(Index, ResolvedODR);
601 thinLTOResolveWeakForLinkerModule(
602 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
604 // Promote the exported values in the index, so that they are promoted
606 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
607 const auto &ExportList = ExportLists.find(ModuleIdentifier);
608 return (ExportList != ExportLists.end() &&
609 ExportList->second.count(GUID)) ||
610 GUIDPreservedSymbols.count(GUID);
612 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
614 promoteModule(TheModule, Index);
618 * Perform cross-module importing for the module identified by ModuleIdentifier.
620 void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule,
621 ModuleSummaryIndex &Index) {
622 auto ModuleMap = generateModuleMap(Modules);
623 auto ModuleCount = Index.modulePaths().size();
625 // Collect for each module the list of function it defines (GUID -> Summary).
626 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
627 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
629 // Convert the preserved symbols set from string to GUID
630 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
631 PreservedSymbols, Triple(TheModule.getTargetTriple()));
633 // Compute "dead" symbols, we don't want to import/export these!
634 auto DeadSymbols = computeDeadSymbols(Index, GUIDPreservedSymbols);
636 // Generate import/export list
637 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
638 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
639 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
640 ExportLists, &DeadSymbols);
641 auto &ImportList = ImportLists[TheModule.getModuleIdentifier()];
643 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
647 * Compute the list of summaries needed for importing into module.
649 void ThinLTOCodeGenerator::gatherImportedSummariesForModule(
650 StringRef ModulePath, ModuleSummaryIndex &Index,
651 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
652 auto ModuleCount = Index.modulePaths().size();
654 // Collect for each module the list of function it defines (GUID -> Summary).
655 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
656 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
658 // Generate import/export list
659 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
660 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
661 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
664 llvm::gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
665 ImportLists[ModulePath],
666 ModuleToSummariesForIndex);
670 * Emit the list of files needed for importing into module.
672 void ThinLTOCodeGenerator::emitImports(StringRef ModulePath,
673 StringRef OutputName,
674 ModuleSummaryIndex &Index) {
675 auto ModuleCount = Index.modulePaths().size();
677 // Collect for each module the list of function it defines (GUID -> Summary).
678 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
679 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
681 // Generate import/export list
682 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
683 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
684 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
688 if ((EC = EmitImportsFiles(ModulePath, OutputName, ImportLists[ModulePath])))
689 report_fatal_error(Twine("Failed to open ") + OutputName +
690 " to save imports lists\n");
694 * Perform internalization. Index is updated to reflect linkage changes.
696 void ThinLTOCodeGenerator::internalize(Module &TheModule,
697 ModuleSummaryIndex &Index) {
698 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
699 auto ModuleCount = Index.modulePaths().size();
700 auto ModuleIdentifier = TheModule.getModuleIdentifier();
702 // Convert the preserved symbols set from string to GUID
703 auto GUIDPreservedSymbols =
704 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
706 // Collect for each module the list of function it defines (GUID -> Summary).
707 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
708 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
710 // Compute "dead" symbols, we don't want to import/export these!
711 auto DeadSymbols = computeDeadSymbols(Index, GUIDPreservedSymbols);
713 // Generate import/export list
714 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
715 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
716 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
717 ExportLists, &DeadSymbols);
718 auto &ExportList = ExportLists[ModuleIdentifier];
720 // Be friendly and don't nuke totally the module when the client didn't
721 // supply anything to preserve.
722 if (ExportList.empty() && GUIDPreservedSymbols.empty())
726 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
727 const auto &ExportList = ExportLists.find(ModuleIdentifier);
728 return (ExportList != ExportLists.end() &&
729 ExportList->second.count(GUID)) ||
730 GUIDPreservedSymbols.count(GUID);
732 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
733 thinLTOInternalizeModule(TheModule,
734 ModuleToDefinedGVSummaries[ModuleIdentifier]);
738 * Perform post-importing ThinLTO optimizations.
740 void ThinLTOCodeGenerator::optimize(Module &TheModule) {
741 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
744 optimizeModule(TheModule, *TMBuilder.create(), OptLevel);
748 * Perform ThinLTO CodeGen.
750 std::unique_ptr<MemoryBuffer> ThinLTOCodeGenerator::codegen(Module &TheModule) {
751 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
752 return codegenModule(TheModule, *TMBuilder.create());
755 /// Write out the generated object file, either from CacheEntryPath or from
756 /// OutputBuffer, preferring hard-link when possible.
757 /// Returns the path to the generated file in SavedObjectsDirectoryPath.
758 static std::string writeGeneratedObject(int count, StringRef CacheEntryPath,
759 StringRef SavedObjectsDirectoryPath,
760 const MemoryBuffer &OutputBuffer) {
761 SmallString<128> OutputPath(SavedObjectsDirectoryPath);
762 llvm::sys::path::append(OutputPath, Twine(count) + ".thinlto.o");
763 OutputPath.c_str(); // Ensure the string is null terminated.
764 if (sys::fs::exists(OutputPath))
765 sys::fs::remove(OutputPath);
767 // We don't return a memory buffer to the linker, just a list of files.
768 if (!CacheEntryPath.empty()) {
769 // Cache is enabled, hard-link the entry (or copy if hard-link fails).
770 auto Err = sys::fs::create_hard_link(CacheEntryPath, OutputPath);
772 return OutputPath.str();
773 // Hard linking failed, try to copy.
774 Err = sys::fs::copy_file(CacheEntryPath, OutputPath);
776 return OutputPath.str();
777 // Copy failed (could be because the CacheEntry was removed from the cache
778 // in the meantime by another process), fall back and try to write down the
779 // buffer to the output.
780 errs() << "error: can't link or copy from cached entry '" << CacheEntryPath
781 << "' to '" << OutputPath << "'\n";
783 // No cache entry, just write out the buffer.
785 raw_fd_ostream OS(OutputPath, Err, sys::fs::F_None);
787 report_fatal_error("Can't open output '" + OutputPath + "'\n");
788 OS << OutputBuffer.getBuffer();
789 return OutputPath.str();
792 // Main entry point for the ThinLTO processing
793 void ThinLTOCodeGenerator::run() {
795 // Perform only parallel codegen and return.
797 assert(ProducedBinaries.empty() && "The generator should not be reused");
798 ProducedBinaries.resize(Modules.size());
800 for (auto &ModuleBuffer : Modules) {
801 Pool.async([&](int count) {
803 Context.setDiscardValueNames(LTODiscardValueNames);
806 auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false,
807 /*IsImporting*/ false);
810 ProducedBinaries[count] = codegen(*TheModule);
817 // Sequential linking phase
818 auto Index = linkCombinedIndex();
820 // Save temps: index.
821 if (!SaveTempsDir.empty()) {
822 auto SaveTempPath = SaveTempsDir + "index.bc";
824 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
826 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
827 " to save optimized bitcode\n");
828 WriteIndexToFile(*Index, OS);
831 // Prepare the resulting object vector
832 assert(ProducedBinaries.empty() && "The generator should not be reused");
833 if (SavedObjectsDirectoryPath.empty())
834 ProducedBinaries.resize(Modules.size());
836 sys::fs::create_directories(SavedObjectsDirectoryPath);
838 sys::fs::is_directory(SavedObjectsDirectoryPath, IsDir);
840 report_fatal_error("Unexistent dir: '" + SavedObjectsDirectoryPath + "'");
841 ProducedBinaryFiles.resize(Modules.size());
844 // Prepare the module map.
845 auto ModuleMap = generateModuleMap(Modules);
846 auto ModuleCount = Modules.size();
848 // Collect for each module the list of function it defines (GUID -> Summary).
849 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
850 Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
852 // Convert the preserved symbols set from string to GUID, this is needed for
853 // computing the caching hash and the internalization.
854 auto GUIDPreservedSymbols =
855 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
857 // Compute "dead" symbols, we don't want to import/export these!
858 auto DeadSymbols = computeDeadSymbols(*Index, GUIDPreservedSymbols);
860 // Collect the import/export lists for all modules from the call-graph in the
862 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
863 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
864 ComputeCrossModuleImport(*Index, ModuleToDefinedGVSummaries, ImportLists,
865 ExportLists, &DeadSymbols);
867 // We use a std::map here to be able to have a defined ordering when
868 // producing a hash for the cache entry.
869 // FIXME: we should be able to compute the caching hash for the entry based
870 // on the index, and nuke this map.
871 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
873 // Resolve LinkOnce/Weak symbols, this has to be computed early because it
874 // impacts the caching.
875 resolveWeakForLinkerInIndex(*Index, ResolvedODR);
877 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
878 const auto &ExportList = ExportLists.find(ModuleIdentifier);
879 return (ExportList != ExportLists.end() &&
880 ExportList->second.count(GUID)) ||
881 GUIDPreservedSymbols.count(GUID);
884 // Use global summary-based analysis to identify symbols that can be
885 // internalized (because they aren't exported or preserved as per callback).
886 // Changes are made in the index, consumed in the ThinLTO backends.
887 thinLTOInternalizeAndPromoteInIndex(*Index, isExported);
889 // Make sure that every module has an entry in the ExportLists and
890 // ResolvedODR maps to enable threaded access to these maps below.
891 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
892 ExportLists[DefinedGVSummaries.first()];
893 ResolvedODR[DefinedGVSummaries.first()];
896 // Compute the ordering we will process the inputs: the rough heuristic here
897 // is to sort them per size so that the largest module get schedule as soon as
898 // possible. This is purely a compile-time optimization.
899 std::vector<int> ModulesOrdering;
900 ModulesOrdering.resize(Modules.size());
901 std::iota(ModulesOrdering.begin(), ModulesOrdering.end(), 0);
902 std::sort(ModulesOrdering.begin(), ModulesOrdering.end(),
903 [&](int LeftIndex, int RightIndex) {
904 auto LSize = Modules[LeftIndex].getBufferSize();
905 auto RSize = Modules[RightIndex].getBufferSize();
906 return LSize > RSize;
909 // Parallel optimizer + codegen
911 ThreadPool Pool(ThreadCount);
912 for (auto IndexCount : ModulesOrdering) {
913 auto &ModuleBuffer = Modules[IndexCount];
914 Pool.async([&](int count) {
915 auto ModuleIdentifier = ModuleBuffer.getBufferIdentifier();
916 auto &ExportList = ExportLists[ModuleIdentifier];
918 auto &DefinedFunctions = ModuleToDefinedGVSummaries[ModuleIdentifier];
920 // The module may be cached, this helps handling it.
921 ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier,
922 ImportLists[ModuleIdentifier], ExportList,
923 ResolvedODR[ModuleIdentifier],
924 DefinedFunctions, GUIDPreservedSymbols);
925 auto CacheEntryPath = CacheEntry.getEntryPath();
928 auto ErrOrBuffer = CacheEntry.tryLoadingBuffer();
929 DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss") << " '"
930 << CacheEntryPath << "' for buffer " << count << " "
931 << ModuleIdentifier << "\n");
935 if (SavedObjectsDirectoryPath.empty())
936 ProducedBinaries[count] = std::move(ErrOrBuffer.get());
938 ProducedBinaryFiles[count] = writeGeneratedObject(
939 count, CacheEntryPath, SavedObjectsDirectoryPath,
946 Context.setDiscardValueNames(LTODiscardValueNames);
947 Context.enableDebugTypeODRUniquing();
948 auto DiagFileOrErr = setupOptimizationRemarks(Context, count);
949 if (!DiagFileOrErr) {
950 errs() << "Error: " << toString(DiagFileOrErr.takeError()) << "\n";
951 report_fatal_error("ThinLTO: Can't get an output file for the "
956 auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false,
957 /*IsImporting*/ false);
959 // Save temps: original file.
960 saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");
962 auto &ImportList = ImportLists[ModuleIdentifier];
963 // Run the main process now, and generates a binary
964 auto OutputBuffer = ProcessThinLTOModule(
965 *TheModule, *Index, ModuleMap, *TMBuilder.create(), ImportList,
966 ExportList, GUIDPreservedSymbols,
967 ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions,
968 DisableCodeGen, SaveTempsDir, OptLevel, count);
970 // Commit to the cache (if enabled)
971 CacheEntry.write(*OutputBuffer);
973 if (SavedObjectsDirectoryPath.empty()) {
974 // We need to generated a memory buffer for the linker.
975 if (!CacheEntryPath.empty()) {
976 // Cache is enabled, reload from the cache
977 // We do this to lower memory pressuree: the buffer is on the heap
978 // and releasing it frees memory that can be used for the next input
979 // file. The final binary link will read from the VFS cache
980 // (hopefully!) or from disk if the memory pressure wasn't too high.
981 auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
982 if (auto EC = ReloadedBufferOrErr.getError()) {
983 // On error, keeping the preexisting buffer and printing a
984 // diagnostic is more friendly than just crashing.
985 errs() << "error: can't reload cached file '" << CacheEntryPath
986 << "': " << EC.message() << "\n";
988 OutputBuffer = std::move(*ReloadedBufferOrErr);
991 ProducedBinaries[count] = std::move(OutputBuffer);
994 ProducedBinaryFiles[count] = writeGeneratedObject(
995 count, CacheEntryPath, SavedObjectsDirectoryPath, *OutputBuffer);
1000 CachePruning(CacheOptions.Path)
1001 .setPruningInterval(std::chrono::seconds(CacheOptions.PruningInterval))
1002 .setEntryExpiration(std::chrono::seconds(CacheOptions.Expiration))
1003 .setMaxSize(CacheOptions.MaxPercentageOfAvailableSpace)
1006 // If statistics were requested, print them out now.
1007 if (llvm::AreStatisticsEnabled())
1008 llvm::PrintStatistics();