[ARM] MVE integer min and max
[llvm-complete.git] / lib / LTO / ThinLTOCodeGenerator.cpp
blob1c52218836caf14284d9e5e34cae5e2f9551c8ec
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::F_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 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 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) {
462 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
463 computePrevailingCopies(Index, PrevailingCopy);
465 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
466 const auto &Prevailing = PrevailingCopy.find(GUID);
467 // Not in map means that there was only one copy, which must be prevailing.
468 if (Prevailing == PrevailingCopy.end())
469 return true;
470 return Prevailing->second == S;
473 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
474 GlobalValue::GUID GUID,
475 GlobalValue::LinkageTypes NewLinkage) {
476 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
479 thinLTOResolvePrevailingInIndex(Index, isPrevailing, recordNewLinkage,
480 GUIDPreservedSymbols);
483 // Initialize the TargetMachine builder for a given Triple
484 static void initTMBuilder(TargetMachineBuilder &TMBuilder,
485 const Triple &TheTriple) {
486 // Set a default CPU for Darwin triples (copied from LTOCodeGenerator).
487 // FIXME this looks pretty terrible...
488 if (TMBuilder.MCpu.empty() && TheTriple.isOSDarwin()) {
489 if (TheTriple.getArch() == llvm::Triple::x86_64)
490 TMBuilder.MCpu = "core2";
491 else if (TheTriple.getArch() == llvm::Triple::x86)
492 TMBuilder.MCpu = "yonah";
493 else if (TheTriple.getArch() == llvm::Triple::aarch64)
494 TMBuilder.MCpu = "cyclone";
496 TMBuilder.TheTriple = std::move(TheTriple);
499 } // end anonymous namespace
501 void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) {
502 MemoryBufferRef Buffer(Data, Identifier);
504 auto InputOrError = lto::InputFile::create(Buffer);
505 if (!InputOrError)
506 report_fatal_error("ThinLTO cannot create input file: " +
507 toString(InputOrError.takeError()));
509 auto TripleStr = (*InputOrError)->getTargetTriple();
510 Triple TheTriple(TripleStr);
512 if (Modules.empty())
513 initTMBuilder(TMBuilder, Triple(TheTriple));
514 else if (TMBuilder.TheTriple != TheTriple) {
515 if (!TMBuilder.TheTriple.isCompatibleWith(TheTriple))
516 report_fatal_error("ThinLTO modules with incompatible triples not "
517 "supported");
518 initTMBuilder(TMBuilder, Triple(TMBuilder.TheTriple.merge(TheTriple)));
521 Modules.emplace_back(std::move(*InputOrError));
524 void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) {
525 PreservedSymbols.insert(Name);
528 void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) {
529 // FIXME: At the moment, we don't take advantage of this extra information,
530 // we're conservatively considering cross-references as preserved.
531 // CrossReferencedSymbols.insert(Name);
532 PreservedSymbols.insert(Name);
535 // TargetMachine factory
536 std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const {
537 std::string ErrMsg;
538 const Target *TheTarget =
539 TargetRegistry::lookupTarget(TheTriple.str(), ErrMsg);
540 if (!TheTarget) {
541 report_fatal_error("Can't load target for this Triple: " + ErrMsg);
544 // Use MAttr as the default set of features.
545 SubtargetFeatures Features(MAttr);
546 Features.getDefaultSubtargetFeatures(TheTriple);
547 std::string FeatureStr = Features.getString();
549 return std::unique_ptr<TargetMachine>(
550 TheTarget->createTargetMachine(TheTriple.str(), MCpu, FeatureStr, Options,
551 RelocModel, None, CGOptLevel));
555 * Produce the combined summary index from all the bitcode files:
556 * "thin-link".
558 std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() {
559 std::unique_ptr<ModuleSummaryIndex> CombinedIndex =
560 llvm::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false);
561 uint64_t NextModuleId = 0;
562 for (auto &Mod : Modules) {
563 auto &M = Mod->getSingleBitcodeModule();
564 if (Error Err =
565 M.readSummary(*CombinedIndex, Mod->getName(), NextModuleId++)) {
566 // FIXME diagnose
567 logAllUnhandledErrors(
568 std::move(Err), errs(),
569 "error: can't create module summary index for buffer: ");
570 return nullptr;
573 return CombinedIndex;
576 static void internalizeAndPromoteInIndex(
577 const StringMap<FunctionImporter::ExportSetTy> &ExportLists,
578 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
579 ModuleSummaryIndex &Index) {
580 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
581 const auto &ExportList = ExportLists.find(ModuleIdentifier);
582 return (ExportList != ExportLists.end() &&
583 ExportList->second.count(GUID)) ||
584 GUIDPreservedSymbols.count(GUID);
587 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
590 static void computeDeadSymbolsInIndex(
591 ModuleSummaryIndex &Index,
592 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
593 // We have no symbols resolution available. And can't do any better now in the
594 // case where the prevailing symbol is in a native object. It can be refined
595 // with linker information in the future.
596 auto isPrevailing = [&](GlobalValue::GUID G) {
597 return PrevailingType::Unknown;
599 computeDeadSymbolsWithConstProp(Index, GUIDPreservedSymbols, isPrevailing,
600 /* ImportEnabled = */ true);
604 * Perform promotion and renaming of exported internal functions.
605 * Index is updated to reflect linkage changes from weak resolution.
607 void ThinLTOCodeGenerator::promote(Module &TheModule, ModuleSummaryIndex &Index,
608 const lto::InputFile &File) {
609 auto ModuleCount = Index.modulePaths().size();
610 auto ModuleIdentifier = TheModule.getModuleIdentifier();
612 // Collect for each module the list of function it defines (GUID -> Summary).
613 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries;
614 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
616 // Convert the preserved symbols set from string to GUID
617 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
618 PreservedSymbols, Triple(TheModule.getTargetTriple()));
620 // Add used symbol to the preserved symbols.
621 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
623 // Compute "dead" symbols, we don't want to import/export these!
624 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
626 // Generate import/export list
627 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
628 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
629 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
630 ExportLists);
632 // Resolve prevailing symbols
633 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
634 resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols);
636 thinLTOResolvePrevailingInModule(
637 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
639 // Promote the exported values in the index, so that they are promoted
640 // in the module.
641 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, Index);
643 promoteModule(TheModule, Index);
647 * Perform cross-module importing for the module identified by ModuleIdentifier.
649 void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule,
650 ModuleSummaryIndex &Index,
651 const lto::InputFile &File) {
652 auto ModuleMap = generateModuleMap(Modules);
653 auto ModuleCount = Index.modulePaths().size();
655 // Collect for each module the list of function it defines (GUID -> Summary).
656 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
657 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
659 // Convert the preserved symbols set from string to GUID
660 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
661 PreservedSymbols, Triple(TheModule.getTargetTriple()));
663 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
665 // Compute "dead" symbols, we don't want to import/export these!
666 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
668 // Generate import/export list
669 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
670 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
671 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
672 ExportLists);
673 auto &ImportList = ImportLists[TheModule.getModuleIdentifier()];
675 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
679 * Compute the list of summaries needed for importing into module.
681 void ThinLTOCodeGenerator::gatherImportedSummariesForModule(
682 Module &TheModule, ModuleSummaryIndex &Index,
683 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex,
684 const lto::InputFile &File) {
685 auto ModuleCount = Index.modulePaths().size();
686 auto ModuleIdentifier = TheModule.getModuleIdentifier();
688 // Collect for each module the list of function it defines (GUID -> Summary).
689 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
690 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
692 // Convert the preserved symbols set from string to GUID
693 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
694 PreservedSymbols, Triple(TheModule.getTargetTriple()));
696 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
698 // Compute "dead" symbols, we don't want to import/export these!
699 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
701 // Generate import/export list
702 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
703 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
704 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
705 ExportLists);
707 llvm::gatherImportedSummariesForModule(
708 ModuleIdentifier, ModuleToDefinedGVSummaries,
709 ImportLists[ModuleIdentifier], ModuleToSummariesForIndex);
713 * Emit the list of files needed for importing into module.
715 void ThinLTOCodeGenerator::emitImports(Module &TheModule, StringRef OutputName,
716 ModuleSummaryIndex &Index,
717 const lto::InputFile &File) {
718 auto ModuleCount = Index.modulePaths().size();
719 auto ModuleIdentifier = TheModule.getModuleIdentifier();
721 // Collect for each module the list of function it defines (GUID -> Summary).
722 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
723 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
725 // Convert the preserved symbols set from string to GUID
726 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
727 PreservedSymbols, Triple(TheModule.getTargetTriple()));
729 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
731 // Compute "dead" symbols, we don't want to import/export these!
732 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
734 // Generate import/export list
735 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
736 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
737 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
738 ExportLists);
740 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
741 llvm::gatherImportedSummariesForModule(
742 ModuleIdentifier, ModuleToDefinedGVSummaries,
743 ImportLists[ModuleIdentifier], ModuleToSummariesForIndex);
745 std::error_code EC;
746 if ((EC = EmitImportsFiles(ModuleIdentifier, OutputName,
747 ModuleToSummariesForIndex)))
748 report_fatal_error(Twine("Failed to open ") + OutputName +
749 " to save imports lists\n");
753 * Perform internalization. Runs promote and internalization together.
754 * Index is updated to reflect linkage changes.
756 void ThinLTOCodeGenerator::internalize(Module &TheModule,
757 ModuleSummaryIndex &Index,
758 const lto::InputFile &File) {
759 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
760 auto ModuleCount = Index.modulePaths().size();
761 auto ModuleIdentifier = TheModule.getModuleIdentifier();
763 // Convert the preserved symbols set from string to GUID
764 auto GUIDPreservedSymbols =
765 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
767 addUsedSymbolToPreservedGUID(File, GUIDPreservedSymbols);
769 // Collect for each module the list of function it defines (GUID -> Summary).
770 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
771 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
773 // Compute "dead" symbols, we don't want to import/export these!
774 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
776 // Generate import/export list
777 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
778 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
779 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
780 ExportLists);
781 auto &ExportList = ExportLists[ModuleIdentifier];
783 // Be friendly and don't nuke totally the module when the client didn't
784 // supply anything to preserve.
785 if (ExportList.empty() && GUIDPreservedSymbols.empty())
786 return;
788 // Resolve prevailing symbols
789 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
790 resolvePrevailingInIndex(Index, ResolvedODR, GUIDPreservedSymbols);
792 // Promote the exported values in the index, so that they are promoted
793 // in the module.
794 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, Index);
796 promoteModule(TheModule, Index);
798 // Internalization
799 thinLTOResolvePrevailingInModule(
800 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
802 thinLTOInternalizeModule(TheModule,
803 ModuleToDefinedGVSummaries[ModuleIdentifier]);
807 * Perform post-importing ThinLTO optimizations.
809 void ThinLTOCodeGenerator::optimize(Module &TheModule) {
810 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
812 // Optimize now
813 optimizeModule(TheModule, *TMBuilder.create(), OptLevel, Freestanding);
816 /// Write out the generated object file, either from CacheEntryPath or from
817 /// OutputBuffer, preferring hard-link when possible.
818 /// Returns the path to the generated file in SavedObjectsDirectoryPath.
819 std::string
820 ThinLTOCodeGenerator::writeGeneratedObject(int count, StringRef CacheEntryPath,
821 const MemoryBuffer &OutputBuffer) {
822 auto ArchName = TMBuilder.TheTriple.getArchName();
823 SmallString<128> OutputPath(SavedObjectsDirectoryPath);
824 llvm::sys::path::append(OutputPath,
825 Twine(count) + "." + ArchName + ".thinlto.o");
826 OutputPath.c_str(); // Ensure the string is null terminated.
827 if (sys::fs::exists(OutputPath))
828 sys::fs::remove(OutputPath);
830 // We don't return a memory buffer to the linker, just a list of files.
831 if (!CacheEntryPath.empty()) {
832 // Cache is enabled, hard-link the entry (or copy if hard-link fails).
833 auto Err = sys::fs::create_hard_link(CacheEntryPath, OutputPath);
834 if (!Err)
835 return OutputPath.str();
836 // Hard linking failed, try to copy.
837 Err = sys::fs::copy_file(CacheEntryPath, OutputPath);
838 if (!Err)
839 return OutputPath.str();
840 // Copy failed (could be because the CacheEntry was removed from the cache
841 // in the meantime by another process), fall back and try to write down the
842 // buffer to the output.
843 errs() << "error: can't link or copy from cached entry '" << CacheEntryPath
844 << "' to '" << OutputPath << "'\n";
846 // No cache entry, just write out the buffer.
847 std::error_code Err;
848 raw_fd_ostream OS(OutputPath, Err, sys::fs::F_None);
849 if (Err)
850 report_fatal_error("Can't open output '" + OutputPath + "'\n");
851 OS << OutputBuffer.getBuffer();
852 return OutputPath.str();
855 // Main entry point for the ThinLTO processing
856 void ThinLTOCodeGenerator::run() {
857 // Prepare the resulting object vector
858 assert(ProducedBinaries.empty() && "The generator should not be reused");
859 if (SavedObjectsDirectoryPath.empty())
860 ProducedBinaries.resize(Modules.size());
861 else {
862 sys::fs::create_directories(SavedObjectsDirectoryPath);
863 bool IsDir;
864 sys::fs::is_directory(SavedObjectsDirectoryPath, IsDir);
865 if (!IsDir)
866 report_fatal_error("Unexistent dir: '" + SavedObjectsDirectoryPath + "'");
867 ProducedBinaryFiles.resize(Modules.size());
870 if (CodeGenOnly) {
871 // Perform only parallel codegen and return.
872 ThreadPool Pool;
873 int count = 0;
874 for (auto &Mod : Modules) {
875 Pool.async([&](int count) {
876 LLVMContext Context;
877 Context.setDiscardValueNames(LTODiscardValueNames);
879 // Parse module now
880 auto TheModule = loadModuleFromInput(Mod.get(), Context, false,
881 /*IsImporting*/ false);
883 // CodeGen
884 auto OutputBuffer = codegenModule(*TheModule, *TMBuilder.create());
885 if (SavedObjectsDirectoryPath.empty())
886 ProducedBinaries[count] = std::move(OutputBuffer);
887 else
888 ProducedBinaryFiles[count] =
889 writeGeneratedObject(count, "", *OutputBuffer);
890 }, count++);
893 return;
896 // Sequential linking phase
897 auto Index = linkCombinedIndex();
899 // Save temps: index.
900 if (!SaveTempsDir.empty()) {
901 auto SaveTempPath = SaveTempsDir + "index.bc";
902 std::error_code EC;
903 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
904 if (EC)
905 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
906 " to save optimized bitcode\n");
907 WriteIndexToFile(*Index, OS);
911 // Prepare the module map.
912 auto ModuleMap = generateModuleMap(Modules);
913 auto ModuleCount = Modules.size();
915 // Collect for each module the list of function it defines (GUID -> Summary).
916 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
917 Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
919 // Convert the preserved symbols set from string to GUID, this is needed for
920 // computing the caching hash and the internalization.
921 auto GUIDPreservedSymbols =
922 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
924 // Add used symbol from inputs to the preserved symbols.
925 for (const auto &M : Modules)
926 addUsedSymbolToPreservedGUID(*M, GUIDPreservedSymbols);
928 // Compute "dead" symbols, we don't want to import/export these!
929 computeDeadSymbolsInIndex(*Index, GUIDPreservedSymbols);
931 // Synthesize entry counts for functions in the combined index.
932 computeSyntheticCounts(*Index);
934 // Collect the import/export lists for all modules from the call-graph in the
935 // combined index.
936 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
937 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
938 ComputeCrossModuleImport(*Index, ModuleToDefinedGVSummaries, ImportLists,
939 ExportLists);
941 // We use a std::map here to be able to have a defined ordering when
942 // producing a hash for the cache entry.
943 // FIXME: we should be able to compute the caching hash for the entry based
944 // on the index, and nuke this map.
945 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
947 // Resolve prevailing symbols, this has to be computed early because it
948 // impacts the caching.
949 resolvePrevailingInIndex(*Index, ResolvedODR, GUIDPreservedSymbols);
951 // Use global summary-based analysis to identify symbols that can be
952 // internalized (because they aren't exported or preserved as per callback).
953 // Changes are made in the index, consumed in the ThinLTO backends.
954 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, *Index);
956 // Make sure that every module has an entry in the ExportLists, ImportList,
957 // GVSummary and ResolvedODR maps to enable threaded access to these maps
958 // below.
959 for (auto &Module : Modules) {
960 auto ModuleIdentifier = Module->getName();
961 ExportLists[ModuleIdentifier];
962 ImportLists[ModuleIdentifier];
963 ResolvedODR[ModuleIdentifier];
964 ModuleToDefinedGVSummaries[ModuleIdentifier];
967 // Compute the ordering we will process the inputs: the rough heuristic here
968 // is to sort them per size so that the largest module get schedule as soon as
969 // possible. This is purely a compile-time optimization.
970 std::vector<int> ModulesOrdering;
971 ModulesOrdering.resize(Modules.size());
972 std::iota(ModulesOrdering.begin(), ModulesOrdering.end(), 0);
973 llvm::sort(ModulesOrdering, [&](int LeftIndex, int RightIndex) {
974 auto LSize =
975 Modules[LeftIndex]->getSingleBitcodeModule().getBuffer().size();
976 auto RSize =
977 Modules[RightIndex]->getSingleBitcodeModule().getBuffer().size();
978 return LSize > RSize;
981 // Parallel optimizer + codegen
983 ThreadPool Pool(ThreadCount);
984 for (auto IndexCount : ModulesOrdering) {
985 auto &Mod = Modules[IndexCount];
986 Pool.async([&](int count) {
987 auto ModuleIdentifier = Mod->getName();
988 auto &ExportList = ExportLists[ModuleIdentifier];
990 auto &DefinedGVSummaries = ModuleToDefinedGVSummaries[ModuleIdentifier];
992 // The module may be cached, this helps handling it.
993 ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier,
994 ImportLists[ModuleIdentifier], ExportList,
995 ResolvedODR[ModuleIdentifier],
996 DefinedGVSummaries, OptLevel, Freestanding,
997 TMBuilder);
998 auto CacheEntryPath = CacheEntry.getEntryPath();
1001 auto ErrOrBuffer = CacheEntry.tryLoadingBuffer();
1002 LLVM_DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss")
1003 << " '" << CacheEntryPath << "' for buffer "
1004 << count << " " << ModuleIdentifier << "\n");
1006 if (ErrOrBuffer) {
1007 // Cache Hit!
1008 if (SavedObjectsDirectoryPath.empty())
1009 ProducedBinaries[count] = std::move(ErrOrBuffer.get());
1010 else
1011 ProducedBinaryFiles[count] = writeGeneratedObject(
1012 count, CacheEntryPath, *ErrOrBuffer.get());
1013 return;
1017 LLVMContext Context;
1018 Context.setDiscardValueNames(LTODiscardValueNames);
1019 Context.enableDebugTypeODRUniquing();
1020 auto DiagFileOrErr = lto::setupOptimizationRemarks(
1021 Context, RemarksFilename, RemarksPasses, RemarksFormat,
1022 RemarksWithHotness, count);
1023 if (!DiagFileOrErr) {
1024 errs() << "Error: " << toString(DiagFileOrErr.takeError()) << "\n";
1025 report_fatal_error("ThinLTO: Can't get an output file for the "
1026 "remarks");
1029 // Parse module now
1030 auto TheModule = loadModuleFromInput(Mod.get(), Context, false,
1031 /*IsImporting*/ false);
1033 // Save temps: original file.
1034 saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");
1036 auto &ImportList = ImportLists[ModuleIdentifier];
1037 // Run the main process now, and generates a binary
1038 auto OutputBuffer = ProcessThinLTOModule(
1039 *TheModule, *Index, ModuleMap, *TMBuilder.create(), ImportList,
1040 ExportList, GUIDPreservedSymbols,
1041 ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions,
1042 DisableCodeGen, SaveTempsDir, Freestanding, OptLevel, count);
1044 // Commit to the cache (if enabled)
1045 CacheEntry.write(*OutputBuffer);
1047 if (SavedObjectsDirectoryPath.empty()) {
1048 // We need to generated a memory buffer for the linker.
1049 if (!CacheEntryPath.empty()) {
1050 // When cache is enabled, reload from the cache if possible.
1051 // Releasing the buffer from the heap and reloading it from the
1052 // cache file with mmap helps us to lower memory pressure.
1053 // The freed memory can be used for the next input file.
1054 // The final binary link will read from the VFS cache (hopefully!)
1055 // or from disk (if the memory pressure was too high).
1056 auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
1057 if (auto EC = ReloadedBufferOrErr.getError()) {
1058 // On error, keep the preexisting buffer and print a diagnostic.
1059 errs() << "error: can't reload cached file '" << CacheEntryPath
1060 << "': " << EC.message() << "\n";
1061 } else {
1062 OutputBuffer = std::move(*ReloadedBufferOrErr);
1065 ProducedBinaries[count] = std::move(OutputBuffer);
1066 return;
1068 ProducedBinaryFiles[count] = writeGeneratedObject(
1069 count, CacheEntryPath, *OutputBuffer);
1070 }, IndexCount);
1074 pruneCache(CacheOptions.Path, CacheOptions.Policy);
1076 // If statistics were requested, print them out now.
1077 if (llvm::AreStatisticsEnabled())
1078 llvm::PrintStatistics();
1079 reportAndResetTimings();