[Alignment][NFC] Use Align with TargetLowering::setMinFunctionAlignment
[llvm-core.git] / lib / LTO / ThinLTOCodeGenerator.cpp
blob6c8f827d59aeac4232d3fcc493a9e5294c2351b6
1 //===-ThinLTOCodeGenerator.cpp - LLVM Link Time Optimizer -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the Thin Link Time Optimization library. This library is
10 // intended to be used by linker to optimize code at link time.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/LTO/legacy/ThinLTOCodeGenerator.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/ADT/StringExtras.h"
18 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
19 #include "llvm/Analysis/ProfileSummaryInfo.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/Bitcode/BitcodeWriterPass.h"
25 #include "llvm/Config/llvm-config.h"
26 #include "llvm/IR/DebugInfo.h"
27 #include "llvm/IR/DiagnosticPrinter.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/LegacyPassManager.h"
30 #include "llvm/IR/Mangler.h"
31 #include "llvm/IR/PassTimingInfo.h"
32 #include "llvm/IR/RemarkStreamer.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/IRReader/IRReader.h"
35 #include "llvm/LTO/LTO.h"
36 #include "llvm/LTO/SummaryBasedOptimizations.h"
37 #include "llvm/MC/SubtargetFeature.h"
38 #include "llvm/Object/IRObjectFile.h"
39 #include "llvm/Support/CachePruning.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/Error.h"
42 #include "llvm/Support/Path.h"
43 #include "llvm/Support/SHA1.h"
44 #include "llvm/Support/SmallVectorMemoryBuffer.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Support/ThreadPool.h"
47 #include "llvm/Support/Threading.h"
48 #include "llvm/Support/ToolOutputFile.h"
49 #include "llvm/Support/VCSRevision.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"
58 #include <numeric>
60 #if !defined(_MSC_VER) && !defined(__MINGW32__)
61 #include <unistd.h>
62 #else
63 #include <io.h>
64 #endif
66 using namespace llvm;
68 #define DEBUG_TYPE "thinlto"
70 namespace llvm {
71 // Flags -discard-value-names, defined in LTOCodeGenerator.cpp
72 extern cl::opt<bool> LTODiscardValueNames;
73 extern cl::opt<std::string> RemarksFilename;
74 extern cl::opt<std::string> RemarksPasses;
75 extern cl::opt<bool> RemarksWithHotness;
76 extern cl::opt<std::string> RemarksFormat;
79 namespace {
81 static cl::opt<int>
82 ThreadCount("threads", cl::init(llvm::heavyweight_hardware_concurrency()));
84 // Simple helper to save temporary files for debug.
85 static void saveTempBitcode(const Module &TheModule, StringRef TempDir,
86 unsigned count, StringRef Suffix) {
87 if (TempDir.empty())
88 return;
89 // User asked to save temps, let dump the bitcode file after import.
90 std::string SaveTempPath = (TempDir + llvm::Twine(count) + Suffix).str();
91 std::error_code EC;
92 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::OF_None);
93 if (EC)
94 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
95 " to save optimized bitcode\n");
96 WriteBitcodeToFile(TheModule, OS, /* ShouldPreserveUseListOrder */ true);
99 static const GlobalValueSummary *
100 getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) {
101 // If there is any strong definition anywhere, get it.
102 auto StrongDefForLinker = llvm::find_if(
103 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
104 auto Linkage = Summary->linkage();
105 return !GlobalValue::isAvailableExternallyLinkage(Linkage) &&
106 !GlobalValue::isWeakForLinker(Linkage);
108 if (StrongDefForLinker != GVSummaryList.end())
109 return StrongDefForLinker->get();
110 // Get the first *linker visible* definition for this global in the summary
111 // list.
112 auto FirstDefForLinker = llvm::find_if(
113 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
114 auto Linkage = Summary->linkage();
115 return !GlobalValue::isAvailableExternallyLinkage(Linkage);
117 // Extern templates can be emitted as available_externally.
118 if (FirstDefForLinker == GVSummaryList.end())
119 return nullptr;
120 return FirstDefForLinker->get();
123 // Populate map of GUID to the prevailing copy for any multiply defined
124 // symbols. Currently assume first copy is prevailing, or any strong
125 // definition. Can be refined with Linker information in the future.
126 static void computePrevailingCopies(
127 const ModuleSummaryIndex &Index,
128 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy) {
129 auto HasMultipleCopies = [&](const GlobalValueSummaryList &GVSummaryList) {
130 return GVSummaryList.size() > 1;
133 for (auto &I : Index) {
134 if (HasMultipleCopies(I.second.SummaryList))
135 PrevailingCopy[I.first] =
136 getFirstDefinitionForLinker(I.second.SummaryList);
140 static StringMap<lto::InputFile *>
141 generateModuleMap(std::vector<std::unique_ptr<lto::InputFile>> &Modules) {
142 StringMap<lto::InputFile *> ModuleMap;
143 for (auto &M : Modules) {
144 assert(ModuleMap.find(M->getName()) == ModuleMap.end() &&
145 "Expect unique Buffer Identifier");
146 ModuleMap[M->getName()] = M.get();
148 return ModuleMap;
151 static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index) {
152 if (renameModuleForThinLTO(TheModule, Index))
153 report_fatal_error("renameModuleForThinLTO failed");
156 namespace {
157 class ThinLTODiagnosticInfo : public DiagnosticInfo {
158 const Twine &Msg;
159 public:
160 ThinLTODiagnosticInfo(const Twine &DiagMsg,
161 DiagnosticSeverity Severity = DS_Error)
162 : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {}
163 void print(DiagnosticPrinter &DP) const override { DP << Msg; }
167 /// Verify the module and strip broken debug info.
168 static void verifyLoadedModule(Module &TheModule) {
169 bool BrokenDebugInfo = false;
170 if (verifyModule(TheModule, &dbgs(), &BrokenDebugInfo))
171 report_fatal_error("Broken module found, compilation aborted!");
172 if (BrokenDebugInfo) {
173 TheModule.getContext().diagnose(ThinLTODiagnosticInfo(
174 "Invalid debug info found, debug info will be stripped", DS_Warning));
175 StripDebugInfo(TheModule);
179 static std::unique_ptr<Module> loadModuleFromInput(lto::InputFile *Input,
180 LLVMContext &Context,
181 bool Lazy,
182 bool IsImporting) {
183 auto &Mod = Input->getSingleBitcodeModule();
184 SMDiagnostic Err;
185 Expected<std::unique_ptr<Module>> ModuleOrErr =
186 Lazy ? Mod.getLazyModule(Context,
187 /* ShouldLazyLoadMetadata */ true, IsImporting)
188 : Mod.parseModule(Context);
189 if (!ModuleOrErr) {
190 handleAllErrors(ModuleOrErr.takeError(), [&](ErrorInfoBase &EIB) {
191 SMDiagnostic Err = SMDiagnostic(Mod.getModuleIdentifier(),
192 SourceMgr::DK_Error, EIB.message());
193 Err.print("ThinLTO", errs());
195 report_fatal_error("Can't load module, abort.");
197 if (!Lazy)
198 verifyLoadedModule(*ModuleOrErr.get());
199 return std::move(*ModuleOrErr);
202 static void
203 crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index,
204 StringMap<lto::InputFile*> &ModuleMap,
205 const FunctionImporter::ImportMapTy &ImportList) {
206 auto Loader = [&](StringRef Identifier) {
207 auto &Input = ModuleMap[Identifier];
208 return loadModuleFromInput(Input, TheModule.getContext(),
209 /*Lazy=*/true, /*IsImporting*/ true);
212 FunctionImporter Importer(Index, Loader);
213 Expected<bool> Result = Importer.importFunctions(TheModule, ImportList);
214 if (!Result) {
215 handleAllErrors(Result.takeError(), [&](ErrorInfoBase &EIB) {
216 SMDiagnostic Err = SMDiagnostic(TheModule.getModuleIdentifier(),
217 SourceMgr::DK_Error, EIB.message());
218 Err.print("ThinLTO", errs());
220 report_fatal_error("importFunctions failed");
222 // Verify again after cross-importing.
223 verifyLoadedModule(TheModule);
226 static void optimizeModule(Module &TheModule, TargetMachine &TM,
227 unsigned OptLevel, bool Freestanding) {
228 // Populate the PassManager
229 PassManagerBuilder PMB;
230 PMB.LibraryInfo = new TargetLibraryInfoImpl(TM.getTargetTriple());
231 if (Freestanding)
232 PMB.LibraryInfo->disableAllFunctions();
233 PMB.Inliner = createFunctionInliningPass();
234 // FIXME: should get it from the bitcode?
235 PMB.OptLevel = OptLevel;
236 PMB.LoopVectorize = true;
237 PMB.SLPVectorize = true;
238 // Already did this in verifyLoadedModule().
239 PMB.VerifyInput = false;
240 PMB.VerifyOutput = false;
242 legacy::PassManager PM;
244 // Add the TTI (required to inform the vectorizer about register size for
245 // instance)
246 PM.add(createTargetTransformInfoWrapperPass(TM.getTargetIRAnalysis()));
248 // Add optimizations
249 PMB.populateThinLTOPassManager(PM);
251 PM.run(TheModule);
254 static void
255 addUsedSymbolToPreservedGUID(const lto::InputFile &File,
256 DenseSet<GlobalValue::GUID> &PreservedGUID) {
257 for (const auto &Sym : File.symbols()) {
258 if (Sym.isUsed())
259 PreservedGUID.insert(GlobalValue::getGUID(Sym.getIRName()));
263 // Convert the PreservedSymbols map from "Name" based to "GUID" based.
264 static DenseSet<GlobalValue::GUID>
265 computeGUIDPreservedSymbols(const StringSet<> &PreservedSymbols,
266 const Triple &TheTriple) {
267 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols(PreservedSymbols.size());
268 for (auto &Entry : PreservedSymbols) {
269 StringRef Name = Entry.first();
270 if (TheTriple.isOSBinFormatMachO() && Name.size() > 0 && Name[0] == '_')
271 Name = Name.drop_front();
272 GUIDPreservedSymbols.insert(GlobalValue::getGUID(Name));
274 return GUIDPreservedSymbols;
277 std::unique_ptr<MemoryBuffer> codegenModule(Module &TheModule,
278 TargetMachine &TM) {
279 SmallVector<char, 128> OutputBuffer;
281 // CodeGen
283 raw_svector_ostream OS(OutputBuffer);
284 legacy::PassManager PM;
286 // If the bitcode files contain ARC code and were compiled with optimization,
287 // the ObjCARCContractPass must be run, so do it unconditionally here.
288 PM.add(createObjCARCContractPass());
290 // Setup the codegen now.
291 if (TM.addPassesToEmitFile(PM, OS, nullptr, TargetMachine::CGFT_ObjectFile,
292 /* DisableVerify */ true))
293 report_fatal_error("Failed to setup codegen");
295 // Run codegen now. resulting binary is in OutputBuffer.
296 PM.run(TheModule);
298 return std::make_unique<SmallVectorMemoryBuffer>(std::move(OutputBuffer));
301 /// Manage caching for a single Module.
302 class ModuleCacheEntry {
303 SmallString<128> EntryPath;
305 public:
306 // Create a cache entry. This compute a unique hash for the Module considering
307 // the current list of export/import, and offer an interface to query to
308 // access the content in the cache.
309 ModuleCacheEntry(
310 StringRef CachePath, const ModuleSummaryIndex &Index, StringRef ModuleID,
311 const FunctionImporter::ImportMapTy &ImportList,
312 const FunctionImporter::ExportSetTy &ExportList,
313 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
314 const GVSummaryMapTy &DefinedGVSummaries, unsigned OptLevel,
315 bool Freestanding, const TargetMachineBuilder &TMBuilder) {
316 if (CachePath.empty())
317 return;
319 if (!Index.modulePaths().count(ModuleID))
320 // The module does not have an entry, it can't have a hash at all
321 return;
323 if (all_of(Index.getModuleHash(ModuleID),
324 [](uint32_t V) { return V == 0; }))
325 // No hash entry, no caching!
326 return;
328 llvm::lto::Config Conf;
329 Conf.OptLevel = OptLevel;
330 Conf.Options = TMBuilder.Options;
331 Conf.CPU = TMBuilder.MCpu;
332 Conf.MAttrs.push_back(TMBuilder.MAttr);
333 Conf.RelocModel = TMBuilder.RelocModel;
334 Conf.CGOptLevel = TMBuilder.CGOptLevel;
335 Conf.Freestanding = Freestanding;
336 SmallString<40> Key;
337 computeLTOCacheKey(Key, Conf, Index, ModuleID, ImportList, ExportList,
338 ResolvedODR, DefinedGVSummaries);
340 // This choice of file name allows the cache to be pruned (see pruneCache()
341 // in include/llvm/Support/CachePruning.h).
342 sys::path::append(EntryPath, CachePath, "llvmcache-" + Key);
345 // Access the path to this entry in the cache.
346 StringRef getEntryPath() { return EntryPath; }
348 // Try loading the buffer for this cache entry.
349 ErrorOr<std::unique_ptr<MemoryBuffer>> tryLoadingBuffer() {
350 if (EntryPath.empty())
351 return std::error_code();
352 SmallString<64> ResultPath;
353 Expected<sys::fs::file_t> FDOrErr = sys::fs::openNativeFileForRead(
354 Twine(EntryPath), sys::fs::OF_UpdateAtime, &ResultPath);
355 if (!FDOrErr)
356 return errorToErrorCode(FDOrErr.takeError());
357 ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr = MemoryBuffer::getOpenFile(
358 *FDOrErr, EntryPath, /*FileSize=*/-1, /*RequiresNullTerminator=*/false);
359 sys::fs::closeFile(*FDOrErr);
360 return MBOrErr;
363 // Cache the Produced object file
364 void write(const MemoryBuffer &OutputBuffer) {
365 if (EntryPath.empty())
366 return;
368 // Write to a temporary to avoid race condition
369 SmallString<128> TempFilename;
370 SmallString<128> CachePath(EntryPath);
371 int TempFD;
372 llvm::sys::path::remove_filename(CachePath);
373 sys::path::append(TempFilename, CachePath, "Thin-%%%%%%.tmp.o");
374 std::error_code EC =
375 sys::fs::createUniqueFile(TempFilename, TempFD, TempFilename);
376 if (EC) {
377 errs() << "Error: " << EC.message() << "\n";
378 report_fatal_error("ThinLTO: Can't get a temporary file");
381 raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
382 OS << OutputBuffer.getBuffer();
384 // Rename temp file to final destination; rename is atomic
385 EC = sys::fs::rename(TempFilename, EntryPath);
386 if (EC)
387 sys::fs::remove(TempFilename);
391 static std::unique_ptr<MemoryBuffer>
392 ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index,
393 StringMap<lto::InputFile *> &ModuleMap, TargetMachine &TM,
394 const FunctionImporter::ImportMapTy &ImportList,
395 const FunctionImporter::ExportSetTy &ExportList,
396 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
397 const GVSummaryMapTy &DefinedGlobals,
398 const ThinLTOCodeGenerator::CachingOptions &CacheOptions,
399 bool DisableCodeGen, StringRef SaveTempsDir,
400 bool Freestanding, unsigned OptLevel, unsigned count) {
402 // "Benchmark"-like optimization: single-source case
403 bool SingleModule = (ModuleMap.size() == 1);
405 if (!SingleModule) {
406 promoteModule(TheModule, Index);
408 // Apply summary-based prevailing-symbol resolution decisions.
409 thinLTOResolvePrevailingInModule(TheModule, DefinedGlobals);
411 // Save temps: after promotion.
412 saveTempBitcode(TheModule, SaveTempsDir, count, ".1.promoted.bc");
415 // Be friendly and don't nuke totally the module when the client didn't
416 // supply anything to preserve.
417 if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) {
418 // Apply summary-based internalization decisions.
419 thinLTOInternalizeModule(TheModule, DefinedGlobals);
422 // Save internalized bitcode
423 saveTempBitcode(TheModule, SaveTempsDir, count, ".2.internalized.bc");
425 if (!SingleModule) {
426 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
428 // Save temps: after cross-module import.
429 saveTempBitcode(TheModule, SaveTempsDir, count, ".3.imported.bc");
432 optimizeModule(TheModule, TM, OptLevel, Freestanding);
434 saveTempBitcode(TheModule, SaveTempsDir, count, ".4.opt.bc");
436 if (DisableCodeGen) {
437 // Configured to stop before CodeGen, serialize the bitcode and return.
438 SmallVector<char, 128> OutputBuffer;
440 raw_svector_ostream OS(OutputBuffer);
441 ProfileSummaryInfo PSI(TheModule);
442 auto Index = buildModuleSummaryIndex(TheModule, nullptr, &PSI);
443 WriteBitcodeToFile(TheModule, OS, true, &Index);
445 return std::make_unique<SmallVectorMemoryBuffer>(std::move(OutputBuffer));
448 return codegenModule(TheModule, TM);
451 /// Resolve prevailing symbols. Record resolutions in the \p ResolvedODR map
452 /// for caching, and in the \p Index for application during the ThinLTO
453 /// backends. This is needed for correctness for exported symbols (ensure
454 /// at least one copy kept) and a compile-time optimization (to drop duplicate
455 /// copies when possible).
456 static void resolvePrevailingInIndex(
457 ModuleSummaryIndex &Index,
458 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>>
459 &ResolvedODR,
460 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
461 const DenseMap<GlobalValue::GUID, const GlobalValueSummary *>
462 &PrevailingCopy) {
464 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
465 const auto &Prevailing = PrevailingCopy.find(GUID);
466 // Not in map means that there was only one copy, which must be prevailing.
467 if (Prevailing == PrevailingCopy.end())
468 return true;
469 return Prevailing->second == S;
472 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
473 GlobalValue::GUID GUID,
474 GlobalValue::LinkageTypes NewLinkage) {
475 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
478 thinLTOResolvePrevailingInIndex(Index, isPrevailing, recordNewLinkage,
479 GUIDPreservedSymbols);
482 // Initialize the TargetMachine builder for a given Triple
483 static void initTMBuilder(TargetMachineBuilder &TMBuilder,
484 const Triple &TheTriple) {
485 // Set a default CPU for Darwin triples (copied from LTOCodeGenerator).
486 // FIXME this looks pretty terrible...
487 if (TMBuilder.MCpu.empty() && TheTriple.isOSDarwin()) {
488 if (TheTriple.getArch() == llvm::Triple::x86_64)
489 TMBuilder.MCpu = "core2";
490 else if (TheTriple.getArch() == llvm::Triple::x86)
491 TMBuilder.MCpu = "yonah";
492 else if (TheTriple.getArch() == llvm::Triple::aarch64)
493 TMBuilder.MCpu = "cyclone";
495 TMBuilder.TheTriple = std::move(TheTriple);
498 } // end anonymous namespace
500 void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) {
501 MemoryBufferRef Buffer(Data, Identifier);
503 auto InputOrError = lto::InputFile::create(Buffer);
504 if (!InputOrError)
505 report_fatal_error("ThinLTO cannot create input file: " +
506 toString(InputOrError.takeError()));
508 auto TripleStr = (*InputOrError)->getTargetTriple();
509 Triple TheTriple(TripleStr);
511 if (Modules.empty())
512 initTMBuilder(TMBuilder, Triple(TheTriple));
513 else if (TMBuilder.TheTriple != TheTriple) {
514 if (!TMBuilder.TheTriple.isCompatibleWith(TheTriple))
515 report_fatal_error("ThinLTO modules with incompatible triples not "
516 "supported");
517 initTMBuilder(TMBuilder, Triple(TMBuilder.TheTriple.merge(TheTriple)));
520 Modules.emplace_back(std::move(*InputOrError));
523 void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) {
524 PreservedSymbols.insert(Name);
527 void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) {
528 // FIXME: At the moment, we don't take advantage of this extra information,
529 // we're conservatively considering cross-references as preserved.
530 // CrossReferencedSymbols.insert(Name);
531 PreservedSymbols.insert(Name);
534 // TargetMachine factory
535 std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const {
536 std::string ErrMsg;
537 const Target *TheTarget =
538 TargetRegistry::lookupTarget(TheTriple.str(), ErrMsg);
539 if (!TheTarget) {
540 report_fatal_error("Can't load target for this Triple: " + ErrMsg);
543 // Use MAttr as the default set of features.
544 SubtargetFeatures Features(MAttr);
545 Features.getDefaultSubtargetFeatures(TheTriple);
546 std::string FeatureStr = Features.getString();
548 return std::unique_ptr<TargetMachine>(
549 TheTarget->createTargetMachine(TheTriple.str(), MCpu, FeatureStr, Options,
550 RelocModel, None, CGOptLevel));
554 * Produce the combined summary index from all the bitcode files:
555 * "thin-link".
557 std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() {
558 std::unique_ptr<ModuleSummaryIndex> CombinedIndex =
559 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false);
560 uint64_t NextModuleId = 0;
561 for (auto &Mod : Modules) {
562 auto &M = Mod->getSingleBitcodeModule();
563 if (Error Err =
564 M.readSummary(*CombinedIndex, Mod->getName(), NextModuleId++)) {
565 // FIXME diagnose
566 logAllUnhandledErrors(
567 std::move(Err), errs(),
568 "error: can't create module summary index for buffer: ");
569 return nullptr;
572 return CombinedIndex;
575 static void internalizeAndPromoteInIndex(
576 const StringMap<FunctionImporter::ExportSetTy> &ExportLists,
577 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
578 const DenseMap<GlobalValue::GUID, const GlobalValueSummary *>
579 &PrevailingCopy,
580 ModuleSummaryIndex &Index) {
581 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
582 const auto &ExportList = ExportLists.find(ModuleIdentifier);
583 return (ExportList != ExportLists.end() &&
584 ExportList->second.count(GUID)) ||
585 GUIDPreservedSymbols.count(GUID);
588 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
589 const auto &Prevailing = PrevailingCopy.find(GUID);
590 // Not in map means that there was only one copy, which must be prevailing.
591 if (Prevailing == PrevailingCopy.end())
592 return true;
593 return Prevailing->second == S;
596 thinLTOInternalizeAndPromoteInIndex(Index, isExported, isPrevailing);
599 static void computeDeadSymbolsInIndex(
600 ModuleSummaryIndex &Index,
601 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
602 // We have no symbols resolution available. And can't do any better now in the
603 // case where the prevailing symbol is in a native object. It can be refined
604 // with linker information in the future.
605 auto isPrevailing = [&](GlobalValue::GUID G) {
606 return PrevailingType::Unknown;
608 computeDeadSymbolsWithConstProp(Index, GUIDPreservedSymbols, isPrevailing,
609 /* ImportEnabled = */ true);
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, ModuleSummaryIndex &Index,
617 const lto::InputFile &File) {
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 // Add used symbol to the preserved symbols.
630 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
632 // Compute "dead" symbols, we don't want to import/export these!
633 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
635 // Generate import/export list
636 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
637 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
638 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
639 ExportLists);
641 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
642 computePrevailingCopies(Index, PrevailingCopy);
644 // Resolve prevailing symbols
645 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
646 resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols,
647 PrevailingCopy);
649 thinLTOResolvePrevailingInModule(
650 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
652 // Promote the exported values in the index, so that they are promoted
653 // in the module.
654 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols,
655 PrevailingCopy, Index);
657 promoteModule(TheModule, Index);
661 * Perform cross-module importing for the module identified by ModuleIdentifier.
663 void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule,
664 ModuleSummaryIndex &Index,
665 const lto::InputFile &File) {
666 auto ModuleMap = generateModuleMap(Modules);
667 auto ModuleCount = Index.modulePaths().size();
669 // Collect for each module the list of function it defines (GUID -> Summary).
670 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
671 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
673 // Convert the preserved symbols set from string to GUID
674 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
675 PreservedSymbols, Triple(TheModule.getTargetTriple()));
677 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
679 // Compute "dead" symbols, we don't want to import/export these!
680 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
682 // Generate import/export list
683 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
684 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
685 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
686 ExportLists);
687 auto &ImportList = ImportLists[TheModule.getModuleIdentifier()];
689 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
693 * Compute the list of summaries needed for importing into module.
695 void ThinLTOCodeGenerator::gatherImportedSummariesForModule(
696 Module &TheModule, ModuleSummaryIndex &Index,
697 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex,
698 const lto::InputFile &File) {
699 auto ModuleCount = Index.modulePaths().size();
700 auto ModuleIdentifier = TheModule.getModuleIdentifier();
702 // Collect for each module the list of function it defines (GUID -> Summary).
703 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
704 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
706 // Convert the preserved symbols set from string to GUID
707 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
708 PreservedSymbols, Triple(TheModule.getTargetTriple()));
710 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
712 // Compute "dead" symbols, we don't want to import/export these!
713 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
715 // Generate import/export list
716 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
717 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
718 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
719 ExportLists);
721 llvm::gatherImportedSummariesForModule(
722 ModuleIdentifier, ModuleToDefinedGVSummaries,
723 ImportLists[ModuleIdentifier], ModuleToSummariesForIndex);
727 * Emit the list of files needed for importing into module.
729 void ThinLTOCodeGenerator::emitImports(Module &TheModule, StringRef OutputName,
730 ModuleSummaryIndex &Index,
731 const lto::InputFile &File) {
732 auto ModuleCount = Index.modulePaths().size();
733 auto ModuleIdentifier = TheModule.getModuleIdentifier();
735 // Collect for each module the list of function it defines (GUID -> Summary).
736 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
737 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
739 // Convert the preserved symbols set from string to GUID
740 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
741 PreservedSymbols, Triple(TheModule.getTargetTriple()));
743 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
745 // Compute "dead" symbols, we don't want to import/export these!
746 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
748 // Generate import/export list
749 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
750 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
751 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
752 ExportLists);
754 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
755 llvm::gatherImportedSummariesForModule(
756 ModuleIdentifier, ModuleToDefinedGVSummaries,
757 ImportLists[ModuleIdentifier], ModuleToSummariesForIndex);
759 std::error_code EC;
760 if ((EC = EmitImportsFiles(ModuleIdentifier, OutputName,
761 ModuleToSummariesForIndex)))
762 report_fatal_error(Twine("Failed to open ") + OutputName +
763 " to save imports lists\n");
767 * Perform internalization. Runs promote and internalization together.
768 * Index is updated to reflect linkage changes.
770 void ThinLTOCodeGenerator::internalize(Module &TheModule,
771 ModuleSummaryIndex &Index,
772 const lto::InputFile &File) {
773 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
774 auto ModuleCount = Index.modulePaths().size();
775 auto ModuleIdentifier = TheModule.getModuleIdentifier();
777 // Convert the preserved symbols set from string to GUID
778 auto GUIDPreservedSymbols =
779 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
781 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
783 // Collect for each module the list of function it defines (GUID -> Summary).
784 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
785 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
787 // Compute "dead" symbols, we don't want to import/export these!
788 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
790 // Generate import/export list
791 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
792 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
793 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
794 ExportLists);
795 auto &ExportList = ExportLists[ModuleIdentifier];
797 // Be friendly and don't nuke totally the module when the client didn't
798 // supply anything to preserve.
799 if (ExportList.empty() && GUIDPreservedSymbols.empty())
800 return;
802 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
803 computePrevailingCopies(Index, PrevailingCopy);
805 // Resolve prevailing symbols
806 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
807 resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols,
808 PrevailingCopy);
810 // Promote the exported values in the index, so that they are promoted
811 // in the module.
812 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols,
813 PrevailingCopy, Index);
815 promoteModule(TheModule, Index);
817 // Internalization
818 thinLTOResolvePrevailingInModule(
819 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
821 thinLTOInternalizeModule(TheModule,
822 ModuleToDefinedGVSummaries[ModuleIdentifier]);
826 * Perform post-importing ThinLTO optimizations.
828 void ThinLTOCodeGenerator::optimize(Module &TheModule) {
829 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
831 // Optimize now
832 optimizeModule(TheModule, *TMBuilder.create(), OptLevel, Freestanding);
835 /// Write out the generated object file, either from CacheEntryPath or from
836 /// OutputBuffer, preferring hard-link when possible.
837 /// Returns the path to the generated file in SavedObjectsDirectoryPath.
838 std::string
839 ThinLTOCodeGenerator::writeGeneratedObject(int count, StringRef CacheEntryPath,
840 const MemoryBuffer &OutputBuffer) {
841 auto ArchName = TMBuilder.TheTriple.getArchName();
842 SmallString<128> OutputPath(SavedObjectsDirectoryPath);
843 llvm::sys::path::append(OutputPath,
844 Twine(count) + "." + ArchName + ".thinlto.o");
845 OutputPath.c_str(); // Ensure the string is null terminated.
846 if (sys::fs::exists(OutputPath))
847 sys::fs::remove(OutputPath);
849 // We don't return a memory buffer to the linker, just a list of files.
850 if (!CacheEntryPath.empty()) {
851 // Cache is enabled, hard-link the entry (or copy if hard-link fails).
852 auto Err = sys::fs::create_hard_link(CacheEntryPath, OutputPath);
853 if (!Err)
854 return OutputPath.str();
855 // Hard linking failed, try to copy.
856 Err = sys::fs::copy_file(CacheEntryPath, OutputPath);
857 if (!Err)
858 return OutputPath.str();
859 // Copy failed (could be because the CacheEntry was removed from the cache
860 // in the meantime by another process), fall back and try to write down the
861 // buffer to the output.
862 errs() << "error: can't link or copy from cached entry '" << CacheEntryPath
863 << "' to '" << OutputPath << "'\n";
865 // No cache entry, just write out the buffer.
866 std::error_code Err;
867 raw_fd_ostream OS(OutputPath, Err, sys::fs::OF_None);
868 if (Err)
869 report_fatal_error("Can't open output '" + OutputPath + "'\n");
870 OS << OutputBuffer.getBuffer();
871 return OutputPath.str();
874 // Main entry point for the ThinLTO processing
875 void ThinLTOCodeGenerator::run() {
876 // Prepare the resulting object vector
877 assert(ProducedBinaries.empty() && "The generator should not be reused");
878 if (SavedObjectsDirectoryPath.empty())
879 ProducedBinaries.resize(Modules.size());
880 else {
881 sys::fs::create_directories(SavedObjectsDirectoryPath);
882 bool IsDir;
883 sys::fs::is_directory(SavedObjectsDirectoryPath, IsDir);
884 if (!IsDir)
885 report_fatal_error("Unexistent dir: '" + SavedObjectsDirectoryPath + "'");
886 ProducedBinaryFiles.resize(Modules.size());
889 if (CodeGenOnly) {
890 // Perform only parallel codegen and return.
891 ThreadPool Pool;
892 int count = 0;
893 for (auto &Mod : Modules) {
894 Pool.async([&](int count) {
895 LLVMContext Context;
896 Context.setDiscardValueNames(LTODiscardValueNames);
898 // Parse module now
899 auto TheModule = loadModuleFromInput(Mod.get(), Context, false,
900 /*IsImporting*/ false);
902 // CodeGen
903 auto OutputBuffer = codegenModule(*TheModule, *TMBuilder.create());
904 if (SavedObjectsDirectoryPath.empty())
905 ProducedBinaries[count] = std::move(OutputBuffer);
906 else
907 ProducedBinaryFiles[count] =
908 writeGeneratedObject(count, "", *OutputBuffer);
909 }, count++);
912 return;
915 // Sequential linking phase
916 auto Index = linkCombinedIndex();
918 // Save temps: index.
919 if (!SaveTempsDir.empty()) {
920 auto SaveTempPath = SaveTempsDir + "index.bc";
921 std::error_code EC;
922 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::OF_None);
923 if (EC)
924 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
925 " to save optimized bitcode\n");
926 WriteIndexToFile(*Index, OS);
930 // Prepare the module map.
931 auto ModuleMap = generateModuleMap(Modules);
932 auto ModuleCount = Modules.size();
934 // Collect for each module the list of function it defines (GUID -> Summary).
935 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
936 Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
938 // Convert the preserved symbols set from string to GUID, this is needed for
939 // computing the caching hash and the internalization.
940 auto GUIDPreservedSymbols =
941 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
943 // Add used symbol from inputs to the preserved symbols.
944 for (const auto &M : Modules)
945 addUsedSymbolToPreservedGUID(*M, GUIDPreservedSymbols);
947 // Compute "dead" symbols, we don't want to import/export these!
948 computeDeadSymbolsInIndex(*Index, GUIDPreservedSymbols);
950 // Synthesize entry counts for functions in the combined index.
951 computeSyntheticCounts(*Index);
953 // Collect the import/export lists for all modules from the call-graph in the
954 // combined index.
955 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
956 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
957 ComputeCrossModuleImport(*Index, ModuleToDefinedGVSummaries, ImportLists,
958 ExportLists);
960 // We use a std::map here to be able to have a defined ordering when
961 // producing a hash for the cache entry.
962 // FIXME: we should be able to compute the caching hash for the entry based
963 // on the index, and nuke this map.
964 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
966 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
967 computePrevailingCopies(*Index, PrevailingCopy);
969 // Resolve prevailing symbols, this has to be computed early because it
970 // impacts the caching.
971 resolvePrevailingInIndex(*Index, ResolvedODR, GUIDPreservedSymbols,
972 PrevailingCopy);
974 // Use global summary-based analysis to identify symbols that can be
975 // internalized (because they aren't exported or preserved as per callback).
976 // Changes are made in the index, consumed in the ThinLTO backends.
977 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols,
978 PrevailingCopy, *Index);
980 // Make sure that every module has an entry in the ExportLists, ImportList,
981 // GVSummary and ResolvedODR maps to enable threaded access to these maps
982 // below.
983 for (auto &Module : Modules) {
984 auto ModuleIdentifier = Module->getName();
985 ExportLists[ModuleIdentifier];
986 ImportLists[ModuleIdentifier];
987 ResolvedODR[ModuleIdentifier];
988 ModuleToDefinedGVSummaries[ModuleIdentifier];
991 // Compute the ordering we will process the inputs: the rough heuristic here
992 // is to sort them per size so that the largest module get schedule as soon as
993 // possible. This is purely a compile-time optimization.
994 std::vector<int> ModulesOrdering;
995 ModulesOrdering.resize(Modules.size());
996 std::iota(ModulesOrdering.begin(), ModulesOrdering.end(), 0);
997 llvm::sort(ModulesOrdering, [&](int LeftIndex, int RightIndex) {
998 auto LSize =
999 Modules[LeftIndex]->getSingleBitcodeModule().getBuffer().size();
1000 auto RSize =
1001 Modules[RightIndex]->getSingleBitcodeModule().getBuffer().size();
1002 return LSize > RSize;
1005 // Parallel optimizer + codegen
1007 ThreadPool Pool(ThreadCount);
1008 for (auto IndexCount : ModulesOrdering) {
1009 auto &Mod = Modules[IndexCount];
1010 Pool.async([&](int count) {
1011 auto ModuleIdentifier = Mod->getName();
1012 auto &ExportList = ExportLists[ModuleIdentifier];
1014 auto &DefinedGVSummaries = ModuleToDefinedGVSummaries[ModuleIdentifier];
1016 // The module may be cached, this helps handling it.
1017 ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier,
1018 ImportLists[ModuleIdentifier], ExportList,
1019 ResolvedODR[ModuleIdentifier],
1020 DefinedGVSummaries, OptLevel, Freestanding,
1021 TMBuilder);
1022 auto CacheEntryPath = CacheEntry.getEntryPath();
1025 auto ErrOrBuffer = CacheEntry.tryLoadingBuffer();
1026 LLVM_DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss")
1027 << " '" << CacheEntryPath << "' for buffer "
1028 << count << " " << ModuleIdentifier << "\n");
1030 if (ErrOrBuffer) {
1031 // Cache Hit!
1032 if (SavedObjectsDirectoryPath.empty())
1033 ProducedBinaries[count] = std::move(ErrOrBuffer.get());
1034 else
1035 ProducedBinaryFiles[count] = writeGeneratedObject(
1036 count, CacheEntryPath, *ErrOrBuffer.get());
1037 return;
1041 LLVMContext Context;
1042 Context.setDiscardValueNames(LTODiscardValueNames);
1043 Context.enableDebugTypeODRUniquing();
1044 auto DiagFileOrErr = lto::setupOptimizationRemarks(
1045 Context, RemarksFilename, RemarksPasses, RemarksFormat,
1046 RemarksWithHotness, count);
1047 if (!DiagFileOrErr) {
1048 errs() << "Error: " << toString(DiagFileOrErr.takeError()) << "\n";
1049 report_fatal_error("ThinLTO: Can't get an output file for the "
1050 "remarks");
1053 // Parse module now
1054 auto TheModule = loadModuleFromInput(Mod.get(), Context, false,
1055 /*IsImporting*/ false);
1057 // Save temps: original file.
1058 saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");
1060 auto &ImportList = ImportLists[ModuleIdentifier];
1061 // Run the main process now, and generates a binary
1062 auto OutputBuffer = ProcessThinLTOModule(
1063 *TheModule, *Index, ModuleMap, *TMBuilder.create(), ImportList,
1064 ExportList, GUIDPreservedSymbols,
1065 ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions,
1066 DisableCodeGen, SaveTempsDir, Freestanding, OptLevel, count);
1068 // Commit to the cache (if enabled)
1069 CacheEntry.write(*OutputBuffer);
1071 if (SavedObjectsDirectoryPath.empty()) {
1072 // We need to generated a memory buffer for the linker.
1073 if (!CacheEntryPath.empty()) {
1074 // When cache is enabled, reload from the cache if possible.
1075 // Releasing the buffer from the heap and reloading it from the
1076 // cache file with mmap helps us to lower memory pressure.
1077 // The freed memory can be used for the next input file.
1078 // The final binary link will read from the VFS cache (hopefully!)
1079 // or from disk (if the memory pressure was too high).
1080 auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
1081 if (auto EC = ReloadedBufferOrErr.getError()) {
1082 // On error, keep the preexisting buffer and print a diagnostic.
1083 errs() << "error: can't reload cached file '" << CacheEntryPath
1084 << "': " << EC.message() << "\n";
1085 } else {
1086 OutputBuffer = std::move(*ReloadedBufferOrErr);
1089 ProducedBinaries[count] = std::move(OutputBuffer);
1090 return;
1092 ProducedBinaryFiles[count] = writeGeneratedObject(
1093 count, CacheEntryPath, *OutputBuffer);
1094 }, IndexCount);
1098 pruneCache(CacheOptions.Path, CacheOptions.Policy);
1100 // If statistics were requested, print them out now.
1101 if (llvm::AreStatisticsEnabled())
1102 llvm::PrintStatistics();
1103 reportAndResetTimings();