[PowerPC] Remove self-copies in pre-emit peephole
[llvm-core.git] / lib / LTO / ThinLTOCodeGenerator.cpp
blob9500b2ded70ea8de4d05c969ede91d1e1800ff65
1 //===-ThinLTOCodeGenerator.cpp - LLVM Link Time Optimizer -----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the Thin Link Time Optimization library. This library is
11 // intended to be used by linker to optimize code at link time.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/LTO/legacy/ThinLTOCodeGenerator.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringExtras.h"
19 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
20 #include "llvm/Analysis/ProfileSummaryInfo.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Bitcode/BitcodeReader.h"
24 #include "llvm/Bitcode/BitcodeWriter.h"
25 #include "llvm/Bitcode/BitcodeWriterPass.h"
26 #include "llvm/Config/llvm-config.h"
27 #include "llvm/IR/DebugInfo.h"
28 #include "llvm/IR/DiagnosticPrinter.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/LegacyPassManager.h"
31 #include "llvm/IR/Mangler.h"
32 #include "llvm/IR/PassTimingInfo.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/IRReader/IRReader.h"
35 #include "llvm/LTO/LTO.h"
36 #include "llvm/MC/SubtargetFeature.h"
37 #include "llvm/Object/IRObjectFile.h"
38 #include "llvm/Support/CachePruning.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/Error.h"
41 #include "llvm/Support/Path.h"
42 #include "llvm/Support/SHA1.h"
43 #include "llvm/Support/SmallVectorMemoryBuffer.h"
44 #include "llvm/Support/TargetRegistry.h"
45 #include "llvm/Support/ThreadPool.h"
46 #include "llvm/Support/Threading.h"
47 #include "llvm/Support/ToolOutputFile.h"
48 #include "llvm/Support/VCSRevision.h"
49 #include "llvm/Target/TargetMachine.h"
50 #include "llvm/Transforms/IPO.h"
51 #include "llvm/Transforms/IPO/FunctionImport.h"
52 #include "llvm/Transforms/IPO/Internalize.h"
53 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
54 #include "llvm/Transforms/ObjCARC.h"
55 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
57 #include <numeric>
59 #if !defined(_MSC_VER) && !defined(__MINGW32__)
60 #include <unistd.h>
61 #else
62 #include <io.h>
63 #endif
65 using namespace llvm;
67 #define DEBUG_TYPE "thinlto"
69 namespace llvm {
70 // Flags -discard-value-names, defined in LTOCodeGenerator.cpp
71 extern cl::opt<bool> LTODiscardValueNames;
72 extern cl::opt<std::string> LTORemarksFilename;
73 extern cl::opt<bool> LTOPassRemarksWithHotness;
76 namespace {
78 static cl::opt<int>
79 ThreadCount("threads", cl::init(llvm::heavyweight_hardware_concurrency()));
81 // Simple helper to save temporary files for debug.
82 static void saveTempBitcode(const Module &TheModule, StringRef TempDir,
83 unsigned count, StringRef Suffix) {
84 if (TempDir.empty())
85 return;
86 // User asked to save temps, let dump the bitcode file after import.
87 std::string SaveTempPath = (TempDir + llvm::Twine(count) + Suffix).str();
88 std::error_code EC;
89 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
90 if (EC)
91 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
92 " to save optimized bitcode\n");
93 WriteBitcodeToFile(TheModule, OS, /* ShouldPreserveUseListOrder */ true);
96 static const GlobalValueSummary *
97 getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) {
98 // If there is any strong definition anywhere, get it.
99 auto StrongDefForLinker = llvm::find_if(
100 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
101 auto Linkage = Summary->linkage();
102 return !GlobalValue::isAvailableExternallyLinkage(Linkage) &&
103 !GlobalValue::isWeakForLinker(Linkage);
105 if (StrongDefForLinker != GVSummaryList.end())
106 return StrongDefForLinker->get();
107 // Get the first *linker visible* definition for this global in the summary
108 // list.
109 auto FirstDefForLinker = llvm::find_if(
110 GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
111 auto Linkage = Summary->linkage();
112 return !GlobalValue::isAvailableExternallyLinkage(Linkage);
114 // Extern templates can be emitted as available_externally.
115 if (FirstDefForLinker == GVSummaryList.end())
116 return nullptr;
117 return FirstDefForLinker->get();
120 // Populate map of GUID to the prevailing copy for any multiply defined
121 // symbols. Currently assume first copy is prevailing, or any strong
122 // definition. Can be refined with Linker information in the future.
123 static void computePrevailingCopies(
124 const ModuleSummaryIndex &Index,
125 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> &PrevailingCopy) {
126 auto HasMultipleCopies = [&](const GlobalValueSummaryList &GVSummaryList) {
127 return GVSummaryList.size() > 1;
130 for (auto &I : Index) {
131 if (HasMultipleCopies(I.second.SummaryList))
132 PrevailingCopy[I.first] =
133 getFirstDefinitionForLinker(I.second.SummaryList);
137 static StringMap<MemoryBufferRef>
138 generateModuleMap(const std::vector<ThinLTOBuffer> &Modules) {
139 StringMap<MemoryBufferRef> ModuleMap;
140 for (auto &ModuleBuffer : Modules) {
141 assert(ModuleMap.find(ModuleBuffer.getBufferIdentifier()) ==
142 ModuleMap.end() &&
143 "Expect unique Buffer Identifier");
144 ModuleMap[ModuleBuffer.getBufferIdentifier()] = ModuleBuffer.getMemBuffer();
146 return ModuleMap;
149 static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index) {
150 if (renameModuleForThinLTO(TheModule, Index))
151 report_fatal_error("renameModuleForThinLTO failed");
154 namespace {
155 class ThinLTODiagnosticInfo : public DiagnosticInfo {
156 const Twine &Msg;
157 public:
158 ThinLTODiagnosticInfo(const Twine &DiagMsg,
159 DiagnosticSeverity Severity = DS_Error)
160 : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {}
161 void print(DiagnosticPrinter &DP) const override { DP << Msg; }
165 /// Verify the module and strip broken debug info.
166 static void verifyLoadedModule(Module &TheModule) {
167 bool BrokenDebugInfo = false;
168 if (verifyModule(TheModule, &dbgs(), &BrokenDebugInfo))
169 report_fatal_error("Broken module found, compilation aborted!");
170 if (BrokenDebugInfo) {
171 TheModule.getContext().diagnose(ThinLTODiagnosticInfo(
172 "Invalid debug info found, debug info will be stripped", DS_Warning));
173 StripDebugInfo(TheModule);
177 static std::unique_ptr<Module>
178 loadModuleFromBuffer(const MemoryBufferRef &Buffer, LLVMContext &Context,
179 bool Lazy, bool IsImporting) {
180 SMDiagnostic Err;
181 Expected<std::unique_ptr<Module>> ModuleOrErr =
182 Lazy
183 ? getLazyBitcodeModule(Buffer, Context,
184 /* ShouldLazyLoadMetadata */ true, IsImporting)
185 : parseBitcodeFile(Buffer, Context);
186 if (!ModuleOrErr) {
187 handleAllErrors(ModuleOrErr.takeError(), [&](ErrorInfoBase &EIB) {
188 SMDiagnostic Err = SMDiagnostic(Buffer.getBufferIdentifier(),
189 SourceMgr::DK_Error, EIB.message());
190 Err.print("ThinLTO", errs());
192 report_fatal_error("Can't load module, abort.");
194 if (!Lazy)
195 verifyLoadedModule(*ModuleOrErr.get());
196 return std::move(ModuleOrErr.get());
199 static void
200 crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index,
201 StringMap<MemoryBufferRef> &ModuleMap,
202 const FunctionImporter::ImportMapTy &ImportList) {
203 auto Loader = [&](StringRef Identifier) {
204 return loadModuleFromBuffer(ModuleMap[Identifier], TheModule.getContext(),
205 /*Lazy=*/true, /*IsImporting*/ true);
208 FunctionImporter Importer(Index, Loader);
209 Expected<bool> Result = Importer.importFunctions(TheModule, ImportList);
210 if (!Result) {
211 handleAllErrors(Result.takeError(), [&](ErrorInfoBase &EIB) {
212 SMDiagnostic Err = SMDiagnostic(TheModule.getModuleIdentifier(),
213 SourceMgr::DK_Error, EIB.message());
214 Err.print("ThinLTO", errs());
216 report_fatal_error("importFunctions failed");
218 // Verify again after cross-importing.
219 verifyLoadedModule(TheModule);
222 static void optimizeModule(Module &TheModule, TargetMachine &TM,
223 unsigned OptLevel, bool Freestanding) {
224 // Populate the PassManager
225 PassManagerBuilder PMB;
226 PMB.LibraryInfo = new TargetLibraryInfoImpl(TM.getTargetTriple());
227 if (Freestanding)
228 PMB.LibraryInfo->disableAllFunctions();
229 PMB.Inliner = createFunctionInliningPass();
230 // FIXME: should get it from the bitcode?
231 PMB.OptLevel = OptLevel;
232 PMB.LoopVectorize = true;
233 PMB.SLPVectorize = true;
234 // Already did this in verifyLoadedModule().
235 PMB.VerifyInput = false;
236 PMB.VerifyOutput = false;
238 legacy::PassManager PM;
240 // Add the TTI (required to inform the vectorizer about register size for
241 // instance)
242 PM.add(createTargetTransformInfoWrapperPass(TM.getTargetIRAnalysis()));
244 // Add optimizations
245 PMB.populateThinLTOPassManager(PM);
247 PM.run(TheModule);
250 // Convert the PreservedSymbols map from "Name" based to "GUID" based.
251 static DenseSet<GlobalValue::GUID>
252 computeGUIDPreservedSymbols(const StringSet<> &PreservedSymbols,
253 const Triple &TheTriple) {
254 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols(PreservedSymbols.size());
255 for (auto &Entry : PreservedSymbols) {
256 StringRef Name = Entry.first();
257 if (TheTriple.isOSBinFormatMachO() && Name.size() > 0 && Name[0] == '_')
258 Name = Name.drop_front();
259 GUIDPreservedSymbols.insert(GlobalValue::getGUID(Name));
261 return GUIDPreservedSymbols;
264 std::unique_ptr<MemoryBuffer> codegenModule(Module &TheModule,
265 TargetMachine &TM) {
266 SmallVector<char, 128> OutputBuffer;
268 // CodeGen
270 raw_svector_ostream OS(OutputBuffer);
271 legacy::PassManager PM;
273 // If the bitcode files contain ARC code and were compiled with optimization,
274 // the ObjCARCContractPass must be run, so do it unconditionally here.
275 PM.add(createObjCARCContractPass());
277 // Setup the codegen now.
278 if (TM.addPassesToEmitFile(PM, OS, nullptr, TargetMachine::CGFT_ObjectFile,
279 /* DisableVerify */ true))
280 report_fatal_error("Failed to setup codegen");
282 // Run codegen now. resulting binary is in OutputBuffer.
283 PM.run(TheModule);
285 return make_unique<SmallVectorMemoryBuffer>(std::move(OutputBuffer));
288 /// Manage caching for a single Module.
289 class ModuleCacheEntry {
290 SmallString<128> EntryPath;
292 public:
293 // Create a cache entry. This compute a unique hash for the Module considering
294 // the current list of export/import, and offer an interface to query to
295 // access the content in the cache.
296 ModuleCacheEntry(
297 StringRef CachePath, const ModuleSummaryIndex &Index, StringRef ModuleID,
298 const FunctionImporter::ImportMapTy &ImportList,
299 const FunctionImporter::ExportSetTy &ExportList,
300 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
301 const GVSummaryMapTy &DefinedFunctions,
302 const DenseSet<GlobalValue::GUID> &PreservedSymbols, unsigned OptLevel,
303 bool Freestanding, const TargetMachineBuilder &TMBuilder) {
304 if (CachePath.empty())
305 return;
307 if (!Index.modulePaths().count(ModuleID))
308 // The module does not have an entry, it can't have a hash at all
309 return;
311 // Compute the unique hash for this entry
312 // This is based on the current compiler version, the module itself, the
313 // export list, the hash for every single module in the import list, the
314 // list of ResolvedODR for the module, and the list of preserved symbols.
316 // Include the hash for the current module
317 auto ModHash = Index.getModuleHash(ModuleID);
319 if (all_of(ModHash, [](uint32_t V) { return V == 0; }))
320 // No hash entry, no caching!
321 return;
323 SHA1 Hasher;
325 // Include the parts of the LTO configuration that affect code generation.
326 auto AddString = [&](StringRef Str) {
327 Hasher.update(Str);
328 Hasher.update(ArrayRef<uint8_t>{0});
330 auto AddUnsigned = [&](unsigned I) {
331 uint8_t Data[4];
332 Data[0] = I;
333 Data[1] = I >> 8;
334 Data[2] = I >> 16;
335 Data[3] = I >> 24;
336 Hasher.update(ArrayRef<uint8_t>{Data, 4});
339 // Start with the compiler revision
340 Hasher.update(LLVM_VERSION_STRING);
341 #ifdef LLVM_REVISION
342 Hasher.update(LLVM_REVISION);
343 #endif
345 // Hash the optimization level and the target machine settings.
346 AddString(TMBuilder.MCpu);
347 // FIXME: Hash more of Options. For now all clients initialize Options from
348 // command-line flags (which is unsupported in production), but may set
349 // RelaxELFRelocations. The clang driver can also pass FunctionSections,
350 // DataSections and DebuggerTuning via command line flags.
351 AddUnsigned(TMBuilder.Options.RelaxELFRelocations);
352 AddUnsigned(TMBuilder.Options.FunctionSections);
353 AddUnsigned(TMBuilder.Options.DataSections);
354 AddUnsigned((unsigned)TMBuilder.Options.DebuggerTuning);
355 AddString(TMBuilder.MAttr);
356 if (TMBuilder.RelocModel)
357 AddUnsigned(*TMBuilder.RelocModel);
358 AddUnsigned(TMBuilder.CGOptLevel);
359 AddUnsigned(OptLevel);
360 AddUnsigned(Freestanding);
362 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
363 for (auto F : ExportList)
364 // The export list can impact the internalization, be conservative here
365 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&F, sizeof(F)));
367 // Include the hash for every module we import functions from
368 for (auto &Entry : ImportList) {
369 auto ModHash = Index.getModuleHash(Entry.first());
370 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
373 // Include the hash for the resolved ODR.
374 for (auto &Entry : ResolvedODR) {
375 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
376 sizeof(GlobalValue::GUID)));
377 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
378 sizeof(GlobalValue::LinkageTypes)));
381 // Include the hash for the preserved symbols.
382 for (auto &Entry : PreservedSymbols) {
383 if (DefinedFunctions.count(Entry))
384 Hasher.update(
385 ArrayRef<uint8_t>((const uint8_t *)&Entry, sizeof(GlobalValue::GUID)));
388 // This choice of file name allows the cache to be pruned (see pruneCache()
389 // in include/llvm/Support/CachePruning.h).
390 sys::path::append(EntryPath, CachePath,
391 "llvmcache-" + toHex(Hasher.result()));
394 // Access the path to this entry in the cache.
395 StringRef getEntryPath() { return EntryPath; }
397 // Try loading the buffer for this cache entry.
398 ErrorOr<std::unique_ptr<MemoryBuffer>> tryLoadingBuffer() {
399 if (EntryPath.empty())
400 return std::error_code();
401 int FD;
402 SmallString<64> ResultPath;
403 std::error_code EC = sys::fs::openFileForRead(
404 Twine(EntryPath), FD, sys::fs::OF_UpdateAtime, &ResultPath);
405 if (EC)
406 return EC;
407 ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr =
408 MemoryBuffer::getOpenFile(FD, EntryPath,
409 /*FileSize*/ -1,
410 /*RequiresNullTerminator*/ false);
411 close(FD);
412 return MBOrErr;
415 // Cache the Produced object file
416 void write(const MemoryBuffer &OutputBuffer) {
417 if (EntryPath.empty())
418 return;
420 // Write to a temporary to avoid race condition
421 SmallString<128> TempFilename;
422 SmallString<128> CachePath(EntryPath);
423 int TempFD;
424 llvm::sys::path::remove_filename(CachePath);
425 sys::path::append(TempFilename, CachePath, "Thin-%%%%%%.tmp.o");
426 std::error_code EC =
427 sys::fs::createUniqueFile(TempFilename, TempFD, TempFilename);
428 if (EC) {
429 errs() << "Error: " << EC.message() << "\n";
430 report_fatal_error("ThinLTO: Can't get a temporary file");
433 raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
434 OS << OutputBuffer.getBuffer();
436 // Rename temp file to final destination; rename is atomic
437 EC = sys::fs::rename(TempFilename, EntryPath);
438 if (EC)
439 sys::fs::remove(TempFilename);
443 static std::unique_ptr<MemoryBuffer>
444 ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index,
445 StringMap<MemoryBufferRef> &ModuleMap, TargetMachine &TM,
446 const FunctionImporter::ImportMapTy &ImportList,
447 const FunctionImporter::ExportSetTy &ExportList,
448 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
449 const GVSummaryMapTy &DefinedGlobals,
450 const ThinLTOCodeGenerator::CachingOptions &CacheOptions,
451 bool DisableCodeGen, StringRef SaveTempsDir,
452 bool Freestanding, unsigned OptLevel, unsigned count) {
454 // "Benchmark"-like optimization: single-source case
455 bool SingleModule = (ModuleMap.size() == 1);
457 if (!SingleModule) {
458 promoteModule(TheModule, Index);
460 // Apply summary-based LinkOnce/Weak resolution decisions.
461 thinLTOResolveWeakForLinkerModule(TheModule, DefinedGlobals);
463 // Save temps: after promotion.
464 saveTempBitcode(TheModule, SaveTempsDir, count, ".1.promoted.bc");
467 // Be friendly and don't nuke totally the module when the client didn't
468 // supply anything to preserve.
469 if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) {
470 // Apply summary-based internalization decisions.
471 thinLTOInternalizeModule(TheModule, DefinedGlobals);
474 // Save internalized bitcode
475 saveTempBitcode(TheModule, SaveTempsDir, count, ".2.internalized.bc");
477 if (!SingleModule) {
478 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
480 // Save temps: after cross-module import.
481 saveTempBitcode(TheModule, SaveTempsDir, count, ".3.imported.bc");
484 optimizeModule(TheModule, TM, OptLevel, Freestanding);
486 saveTempBitcode(TheModule, SaveTempsDir, count, ".4.opt.bc");
488 if (DisableCodeGen) {
489 // Configured to stop before CodeGen, serialize the bitcode and return.
490 SmallVector<char, 128> OutputBuffer;
492 raw_svector_ostream OS(OutputBuffer);
493 ProfileSummaryInfo PSI(TheModule);
494 auto Index = buildModuleSummaryIndex(TheModule, nullptr, &PSI);
495 WriteBitcodeToFile(TheModule, OS, true, &Index);
497 return make_unique<SmallVectorMemoryBuffer>(std::move(OutputBuffer));
500 return codegenModule(TheModule, TM);
503 /// Resolve LinkOnce/Weak symbols. Record resolutions in the \p ResolvedODR map
504 /// for caching, and in the \p Index for application during the ThinLTO
505 /// backends. This is needed for correctness for exported symbols (ensure
506 /// at least one copy kept) and a compile-time optimization (to drop duplicate
507 /// copies when possible).
508 static void resolveWeakForLinkerInIndex(
509 ModuleSummaryIndex &Index,
510 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>>
511 &ResolvedODR) {
513 DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
514 computePrevailingCopies(Index, PrevailingCopy);
516 auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
517 const auto &Prevailing = PrevailingCopy.find(GUID);
518 // Not in map means that there was only one copy, which must be prevailing.
519 if (Prevailing == PrevailingCopy.end())
520 return true;
521 return Prevailing->second == S;
524 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
525 GlobalValue::GUID GUID,
526 GlobalValue::LinkageTypes NewLinkage) {
527 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
530 thinLTOResolveWeakForLinkerInIndex(Index, isPrevailing, recordNewLinkage);
533 // Initialize the TargetMachine builder for a given Triple
534 static void initTMBuilder(TargetMachineBuilder &TMBuilder,
535 const Triple &TheTriple) {
536 // Set a default CPU for Darwin triples (copied from LTOCodeGenerator).
537 // FIXME this looks pretty terrible...
538 if (TMBuilder.MCpu.empty() && TheTriple.isOSDarwin()) {
539 if (TheTriple.getArch() == llvm::Triple::x86_64)
540 TMBuilder.MCpu = "core2";
541 else if (TheTriple.getArch() == llvm::Triple::x86)
542 TMBuilder.MCpu = "yonah";
543 else if (TheTriple.getArch() == llvm::Triple::aarch64)
544 TMBuilder.MCpu = "cyclone";
546 TMBuilder.TheTriple = std::move(TheTriple);
549 } // end anonymous namespace
551 void ThinLTOCodeGenerator::addModule(StringRef Identifier, StringRef Data) {
552 ThinLTOBuffer Buffer(Data, Identifier);
553 LLVMContext Context;
554 StringRef TripleStr;
555 ErrorOr<std::string> TripleOrErr = expectedToErrorOrAndEmitErrors(
556 Context, getBitcodeTargetTriple(Buffer.getMemBuffer()));
558 if (TripleOrErr)
559 TripleStr = *TripleOrErr;
561 Triple TheTriple(TripleStr);
563 if (Modules.empty())
564 initTMBuilder(TMBuilder, Triple(TheTriple));
565 else if (TMBuilder.TheTriple != TheTriple) {
566 if (!TMBuilder.TheTriple.isCompatibleWith(TheTriple))
567 report_fatal_error("ThinLTO modules with incompatible triples not "
568 "supported");
569 initTMBuilder(TMBuilder, Triple(TMBuilder.TheTriple.merge(TheTriple)));
572 Modules.push_back(Buffer);
575 void ThinLTOCodeGenerator::preserveSymbol(StringRef Name) {
576 PreservedSymbols.insert(Name);
579 void ThinLTOCodeGenerator::crossReferenceSymbol(StringRef Name) {
580 // FIXME: At the moment, we don't take advantage of this extra information,
581 // we're conservatively considering cross-references as preserved.
582 // CrossReferencedSymbols.insert(Name);
583 PreservedSymbols.insert(Name);
586 // TargetMachine factory
587 std::unique_ptr<TargetMachine> TargetMachineBuilder::create() const {
588 std::string ErrMsg;
589 const Target *TheTarget =
590 TargetRegistry::lookupTarget(TheTriple.str(), ErrMsg);
591 if (!TheTarget) {
592 report_fatal_error("Can't load target for this Triple: " + ErrMsg);
595 // Use MAttr as the default set of features.
596 SubtargetFeatures Features(MAttr);
597 Features.getDefaultSubtargetFeatures(TheTriple);
598 std::string FeatureStr = Features.getString();
600 return std::unique_ptr<TargetMachine>(
601 TheTarget->createTargetMachine(TheTriple.str(), MCpu, FeatureStr, Options,
602 RelocModel, None, CGOptLevel));
606 * Produce the combined summary index from all the bitcode files:
607 * "thin-link".
609 std::unique_ptr<ModuleSummaryIndex> ThinLTOCodeGenerator::linkCombinedIndex() {
610 std::unique_ptr<ModuleSummaryIndex> CombinedIndex =
611 llvm::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false);
612 uint64_t NextModuleId = 0;
613 for (auto &ModuleBuffer : Modules) {
614 if (Error Err = readModuleSummaryIndex(ModuleBuffer.getMemBuffer(),
615 *CombinedIndex, NextModuleId++)) {
616 // FIXME diagnose
617 logAllUnhandledErrors(
618 std::move(Err), errs(),
619 "error: can't create module summary index for buffer: ");
620 return nullptr;
623 return CombinedIndex;
626 static void internalizeAndPromoteInIndex(
627 const StringMap<FunctionImporter::ExportSetTy> &ExportLists,
628 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
629 ModuleSummaryIndex &Index) {
630 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
631 const auto &ExportList = ExportLists.find(ModuleIdentifier);
632 return (ExportList != ExportLists.end() &&
633 ExportList->second.count(GUID)) ||
634 GUIDPreservedSymbols.count(GUID);
637 thinLTOInternalizeAndPromoteInIndex(Index, isExported);
640 static void computeDeadSymbolsInIndex(
641 ModuleSummaryIndex &Index,
642 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
643 // We have no symbols resolution available. And can't do any better now in the
644 // case where the prevailing symbol is in a native object. It can be refined
645 // with linker information in the future.
646 auto isPrevailing = [&](GlobalValue::GUID G) {
647 return PrevailingType::Unknown;
649 computeDeadSymbols(Index, GUIDPreservedSymbols, isPrevailing);
653 * Perform promotion and renaming of exported internal functions.
654 * Index is updated to reflect linkage changes from weak resolution.
656 void ThinLTOCodeGenerator::promote(Module &TheModule,
657 ModuleSummaryIndex &Index) {
658 auto ModuleCount = Index.modulePaths().size();
659 auto ModuleIdentifier = TheModule.getModuleIdentifier();
661 // Collect for each module the list of function it defines (GUID -> Summary).
662 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries;
663 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
665 // Convert the preserved symbols set from string to GUID
666 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
667 PreservedSymbols, Triple(TheModule.getTargetTriple()));
669 // Compute "dead" symbols, we don't want to import/export these!
670 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
672 // Generate import/export list
673 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
674 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
675 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
676 ExportLists);
678 // Resolve LinkOnce/Weak symbols.
679 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
680 resolveWeakForLinkerInIndex(Index, ResolvedODR);
682 thinLTOResolveWeakForLinkerModule(
683 TheModule, ModuleToDefinedGVSummaries[ModuleIdentifier]);
685 // Promote the exported values in the index, so that they are promoted
686 // in the module.
687 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, Index);
689 promoteModule(TheModule, Index);
693 * Perform cross-module importing for the module identified by ModuleIdentifier.
695 void ThinLTOCodeGenerator::crossModuleImport(Module &TheModule,
696 ModuleSummaryIndex &Index) {
697 auto ModuleMap = generateModuleMap(Modules);
698 auto ModuleCount = Index.modulePaths().size();
700 // Collect for each module the list of function it defines (GUID -> Summary).
701 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
702 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
704 // Convert the preserved symbols set from string to GUID
705 auto GUIDPreservedSymbols = computeGUIDPreservedSymbols(
706 PreservedSymbols, Triple(TheModule.getTargetTriple()));
708 // Compute "dead" symbols, we don't want to import/export these!
709 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
711 // Generate import/export list
712 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
713 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
714 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
715 ExportLists);
716 auto &ImportList = ImportLists[TheModule.getModuleIdentifier()];
718 crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
722 * Compute the list of summaries needed for importing into module.
724 void ThinLTOCodeGenerator::gatherImportedSummariesForModule(
725 StringRef ModulePath, ModuleSummaryIndex &Index,
726 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
727 auto ModuleCount = Index.modulePaths().size();
729 // Collect for each module the list of function it defines (GUID -> Summary).
730 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
731 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
733 // Generate import/export list
734 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
735 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
736 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
737 ExportLists);
739 llvm::gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
740 ImportLists[ModulePath],
741 ModuleToSummariesForIndex);
745 * Emit the list of files needed for importing into module.
747 void ThinLTOCodeGenerator::emitImports(StringRef ModulePath,
748 StringRef OutputName,
749 ModuleSummaryIndex &Index) {
750 auto ModuleCount = Index.modulePaths().size();
752 // Collect for each module the list of function it defines (GUID -> Summary).
753 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
754 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
756 // Generate import/export list
757 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
758 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
759 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
760 ExportLists);
762 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
763 llvm::gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
764 ImportLists[ModulePath],
765 ModuleToSummariesForIndex);
767 std::error_code EC;
768 if ((EC =
769 EmitImportsFiles(ModulePath, OutputName, ModuleToSummariesForIndex)))
770 report_fatal_error(Twine("Failed to open ") + OutputName +
771 " to save imports lists\n");
775 * Perform internalization. Index is updated to reflect linkage changes.
777 void ThinLTOCodeGenerator::internalize(Module &TheModule,
778 ModuleSummaryIndex &Index) {
779 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
780 auto ModuleCount = Index.modulePaths().size();
781 auto ModuleIdentifier = TheModule.getModuleIdentifier();
783 // Convert the preserved symbols set from string to GUID
784 auto GUIDPreservedSymbols =
785 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
787 // Collect for each module the list of function it defines (GUID -> Summary).
788 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
789 Index.collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
791 // Compute "dead" symbols, we don't want to import/export these!
792 computeDeadSymbolsInIndex(Index, GUIDPreservedSymbols);
794 // Generate import/export list
795 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
796 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
797 ComputeCrossModuleImport(Index, ModuleToDefinedGVSummaries, ImportLists,
798 ExportLists);
799 auto &ExportList = ExportLists[ModuleIdentifier];
801 // Be friendly and don't nuke totally the module when the client didn't
802 // supply anything to preserve.
803 if (ExportList.empty() && GUIDPreservedSymbols.empty())
804 return;
806 // Internalization
807 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, Index);
808 thinLTOInternalizeModule(TheModule,
809 ModuleToDefinedGVSummaries[ModuleIdentifier]);
813 * Perform post-importing ThinLTO optimizations.
815 void ThinLTOCodeGenerator::optimize(Module &TheModule) {
816 initTMBuilder(TMBuilder, Triple(TheModule.getTargetTriple()));
818 // Optimize now
819 optimizeModule(TheModule, *TMBuilder.create(), OptLevel, Freestanding);
822 /// Write out the generated object file, either from CacheEntryPath or from
823 /// OutputBuffer, preferring hard-link when possible.
824 /// Returns the path to the generated file in SavedObjectsDirectoryPath.
825 static std::string writeGeneratedObject(int count, StringRef CacheEntryPath,
826 StringRef SavedObjectsDirectoryPath,
827 const MemoryBuffer &OutputBuffer) {
828 SmallString<128> OutputPath(SavedObjectsDirectoryPath);
829 llvm::sys::path::append(OutputPath, Twine(count) + ".thinlto.o");
830 OutputPath.c_str(); // Ensure the string is null terminated.
831 if (sys::fs::exists(OutputPath))
832 sys::fs::remove(OutputPath);
834 // We don't return a memory buffer to the linker, just a list of files.
835 if (!CacheEntryPath.empty()) {
836 // Cache is enabled, hard-link the entry (or copy if hard-link fails).
837 auto Err = sys::fs::create_hard_link(CacheEntryPath, OutputPath);
838 if (!Err)
839 return OutputPath.str();
840 // Hard linking failed, try to copy.
841 Err = sys::fs::copy_file(CacheEntryPath, OutputPath);
842 if (!Err)
843 return OutputPath.str();
844 // Copy failed (could be because the CacheEntry was removed from the cache
845 // in the meantime by another process), fall back and try to write down the
846 // buffer to the output.
847 errs() << "error: can't link or copy from cached entry '" << CacheEntryPath
848 << "' to '" << OutputPath << "'\n";
850 // No cache entry, just write out the buffer.
851 std::error_code Err;
852 raw_fd_ostream OS(OutputPath, Err, sys::fs::F_None);
853 if (Err)
854 report_fatal_error("Can't open output '" + OutputPath + "'\n");
855 OS << OutputBuffer.getBuffer();
856 return OutputPath.str();
859 // Main entry point for the ThinLTO processing
860 void ThinLTOCodeGenerator::run() {
861 // Prepare the resulting object vector
862 assert(ProducedBinaries.empty() && "The generator should not be reused");
863 if (SavedObjectsDirectoryPath.empty())
864 ProducedBinaries.resize(Modules.size());
865 else {
866 sys::fs::create_directories(SavedObjectsDirectoryPath);
867 bool IsDir;
868 sys::fs::is_directory(SavedObjectsDirectoryPath, IsDir);
869 if (!IsDir)
870 report_fatal_error("Unexistent dir: '" + SavedObjectsDirectoryPath + "'");
871 ProducedBinaryFiles.resize(Modules.size());
874 if (CodeGenOnly) {
875 // Perform only parallel codegen and return.
876 ThreadPool Pool;
877 int count = 0;
878 for (auto &ModuleBuffer : Modules) {
879 Pool.async([&](int count) {
880 LLVMContext Context;
881 Context.setDiscardValueNames(LTODiscardValueNames);
883 // Parse module now
884 auto TheModule =
885 loadModuleFromBuffer(ModuleBuffer.getMemBuffer(), Context, false,
886 /*IsImporting*/ false);
888 // CodeGen
889 auto OutputBuffer = codegenModule(*TheModule, *TMBuilder.create());
890 if (SavedObjectsDirectoryPath.empty())
891 ProducedBinaries[count] = std::move(OutputBuffer);
892 else
893 ProducedBinaryFiles[count] = writeGeneratedObject(
894 count, "", SavedObjectsDirectoryPath, *OutputBuffer);
895 }, count++);
898 return;
901 // Sequential linking phase
902 auto Index = linkCombinedIndex();
904 // Save temps: index.
905 if (!SaveTempsDir.empty()) {
906 auto SaveTempPath = SaveTempsDir + "index.bc";
907 std::error_code EC;
908 raw_fd_ostream OS(SaveTempPath, EC, sys::fs::F_None);
909 if (EC)
910 report_fatal_error(Twine("Failed to open ") + SaveTempPath +
911 " to save optimized bitcode\n");
912 WriteIndexToFile(*Index, OS);
916 // Prepare the module map.
917 auto ModuleMap = generateModuleMap(Modules);
918 auto ModuleCount = Modules.size();
920 // Collect for each module the list of function it defines (GUID -> Summary).
921 StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries(ModuleCount);
922 Index->collectDefinedGVSummariesPerModule(ModuleToDefinedGVSummaries);
924 // Convert the preserved symbols set from string to GUID, this is needed for
925 // computing the caching hash and the internalization.
926 auto GUIDPreservedSymbols =
927 computeGUIDPreservedSymbols(PreservedSymbols, TMBuilder.TheTriple);
929 // Compute "dead" symbols, we don't want to import/export these!
930 computeDeadSymbolsInIndex(*Index, GUIDPreservedSymbols);
932 // Collect the import/export lists for all modules from the call-graph in the
933 // combined index.
934 StringMap<FunctionImporter::ImportMapTy> ImportLists(ModuleCount);
935 StringMap<FunctionImporter::ExportSetTy> ExportLists(ModuleCount);
936 ComputeCrossModuleImport(*Index, ModuleToDefinedGVSummaries, ImportLists,
937 ExportLists);
939 // We use a std::map here to be able to have a defined ordering when
940 // producing a hash for the cache entry.
941 // FIXME: we should be able to compute the caching hash for the entry based
942 // on the index, and nuke this map.
943 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
945 // Resolve LinkOnce/Weak symbols, this has to be computed early because it
946 // impacts the caching.
947 resolveWeakForLinkerInIndex(*Index, ResolvedODR);
949 // Use global summary-based analysis to identify symbols that can be
950 // internalized (because they aren't exported or preserved as per callback).
951 // Changes are made in the index, consumed in the ThinLTO backends.
952 internalizeAndPromoteInIndex(ExportLists, GUIDPreservedSymbols, *Index);
954 // Make sure that every module has an entry in the ExportLists, ImportList,
955 // GVSummary and ResolvedODR maps to enable threaded access to these maps
956 // below.
957 for (auto &Module : Modules) {
958 auto ModuleIdentifier = Module.getBufferIdentifier();
959 ExportLists[ModuleIdentifier];
960 ImportLists[ModuleIdentifier];
961 ResolvedODR[ModuleIdentifier];
962 ModuleToDefinedGVSummaries[ModuleIdentifier];
965 // Compute the ordering we will process the inputs: the rough heuristic here
966 // is to sort them per size so that the largest module get schedule as soon as
967 // possible. This is purely a compile-time optimization.
968 std::vector<int> ModulesOrdering;
969 ModulesOrdering.resize(Modules.size());
970 std::iota(ModulesOrdering.begin(), ModulesOrdering.end(), 0);
971 llvm::sort(ModulesOrdering, [&](int LeftIndex, int RightIndex) {
972 auto LSize = Modules[LeftIndex].getBuffer().size();
973 auto RSize = Modules[RightIndex].getBuffer().size();
974 return LSize > RSize;
977 // Parallel optimizer + codegen
979 ThreadPool Pool(ThreadCount);
980 for (auto IndexCount : ModulesOrdering) {
981 auto &ModuleBuffer = Modules[IndexCount];
982 Pool.async([&](int count) {
983 auto ModuleIdentifier = ModuleBuffer.getBufferIdentifier();
984 auto &ExportList = ExportLists[ModuleIdentifier];
986 auto &DefinedFunctions = ModuleToDefinedGVSummaries[ModuleIdentifier];
988 // The module may be cached, this helps handling it.
989 ModuleCacheEntry CacheEntry(CacheOptions.Path, *Index, ModuleIdentifier,
990 ImportLists[ModuleIdentifier], ExportList,
991 ResolvedODR[ModuleIdentifier],
992 DefinedFunctions, GUIDPreservedSymbols,
993 OptLevel, Freestanding, TMBuilder);
994 auto CacheEntryPath = CacheEntry.getEntryPath();
997 auto ErrOrBuffer = CacheEntry.tryLoadingBuffer();
998 LLVM_DEBUG(dbgs() << "Cache " << (ErrOrBuffer ? "hit" : "miss")
999 << " '" << CacheEntryPath << "' for buffer "
1000 << count << " " << ModuleIdentifier << "\n");
1002 if (ErrOrBuffer) {
1003 // Cache Hit!
1004 if (SavedObjectsDirectoryPath.empty())
1005 ProducedBinaries[count] = std::move(ErrOrBuffer.get());
1006 else
1007 ProducedBinaryFiles[count] = writeGeneratedObject(
1008 count, CacheEntryPath, SavedObjectsDirectoryPath,
1009 *ErrOrBuffer.get());
1010 return;
1014 LLVMContext Context;
1015 Context.setDiscardValueNames(LTODiscardValueNames);
1016 Context.enableDebugTypeODRUniquing();
1017 auto DiagFileOrErr = lto::setupOptimizationRemarks(
1018 Context, LTORemarksFilename, LTOPassRemarksWithHotness, count);
1019 if (!DiagFileOrErr) {
1020 errs() << "Error: " << toString(DiagFileOrErr.takeError()) << "\n";
1021 report_fatal_error("ThinLTO: Can't get an output file for the "
1022 "remarks");
1025 // Parse module now
1026 auto TheModule =
1027 loadModuleFromBuffer(ModuleBuffer.getMemBuffer(), Context, false,
1028 /*IsImporting*/ false);
1030 // Save temps: original file.
1031 saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");
1033 auto &ImportList = ImportLists[ModuleIdentifier];
1034 // Run the main process now, and generates a binary
1035 auto OutputBuffer = ProcessThinLTOModule(
1036 *TheModule, *Index, ModuleMap, *TMBuilder.create(), ImportList,
1037 ExportList, GUIDPreservedSymbols,
1038 ModuleToDefinedGVSummaries[ModuleIdentifier], CacheOptions,
1039 DisableCodeGen, SaveTempsDir, Freestanding, OptLevel, count);
1041 // Commit to the cache (if enabled)
1042 CacheEntry.write(*OutputBuffer);
1044 if (SavedObjectsDirectoryPath.empty()) {
1045 // We need to generated a memory buffer for the linker.
1046 if (!CacheEntryPath.empty()) {
1047 // When cache is enabled, reload from the cache if possible.
1048 // Releasing the buffer from the heap and reloading it from the
1049 // cache file with mmap helps us to lower memory pressure.
1050 // The freed memory can be used for the next input file.
1051 // The final binary link will read from the VFS cache (hopefully!)
1052 // or from disk (if the memory pressure was too high).
1053 auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
1054 if (auto EC = ReloadedBufferOrErr.getError()) {
1055 // On error, keep the preexisting buffer and print a diagnostic.
1056 errs() << "error: can't reload cached file '" << CacheEntryPath
1057 << "': " << EC.message() << "\n";
1058 } else {
1059 OutputBuffer = std::move(*ReloadedBufferOrErr);
1062 ProducedBinaries[count] = std::move(OutputBuffer);
1063 return;
1065 ProducedBinaryFiles[count] = writeGeneratedObject(
1066 count, CacheEntryPath, SavedObjectsDirectoryPath, *OutputBuffer);
1067 }, IndexCount);
1071 pruneCache(CacheOptions.Path, CacheOptions.Policy);
1073 // If statistics were requested, print them out now.
1074 if (llvm::AreStatisticsEnabled())
1075 llvm::PrintStatistics();
1076 reportAndResetTimings();