1 //===-LTO.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 functions and classes used to support LTO.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/LTO/LTO.h"
14 #include "llvm/ADT/ScopeExit.h"
15 #include "llvm/ADT/SmallSet.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/ADT/StringExtras.h"
18 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
19 #include "llvm/Analysis/StackSafetyAnalysis.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/CodeGen/Analysis.h"
25 #include "llvm/Config/llvm-config.h"
26 #include "llvm/IR/AutoUpgrade.h"
27 #include "llvm/IR/DiagnosticPrinter.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/LLVMRemarkStreamer.h"
30 #include "llvm/IR/LegacyPassManager.h"
31 #include "llvm/IR/Mangler.h"
32 #include "llvm/IR/Metadata.h"
33 #include "llvm/LTO/LTOBackend.h"
34 #include "llvm/LTO/SummaryBasedOptimizations.h"
35 #include "llvm/Linker/IRMover.h"
36 #include "llvm/MC/TargetRegistry.h"
37 #include "llvm/Object/IRObjectFile.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Error.h"
40 #include "llvm/Support/FileSystem.h"
41 #include "llvm/Support/ManagedStatic.h"
42 #include "llvm/Support/MemoryBuffer.h"
43 #include "llvm/Support/Path.h"
44 #include "llvm/Support/SHA1.h"
45 #include "llvm/Support/SourceMgr.h"
46 #include "llvm/Support/ThreadPool.h"
47 #include "llvm/Support/Threading.h"
48 #include "llvm/Support/TimeProfiler.h"
49 #include "llvm/Support/ToolOutputFile.h"
50 #include "llvm/Support/VCSRevision.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Transforms/IPO.h"
54 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
55 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
56 #include "llvm/Transforms/Utils/SplitModule.h"
57
58 #include <optional>
59 #include <set>
60
61 using namespace llvm;
62 using namespace lto;
63 using namespace object;
64
65 #define DEBUG_TYPE "lto"
66
67 static cl::opt<bool>
68 DumpThinCGSCCs("dump-thin-cg-sccs", cl::init(false), cl::Hidden,
69 cl::desc("Dump the SCCs in the ThinLTO index's callgraph"));
70
71 namespace llvm {
72 /// Enable global value internalization in LTO.
73 cl::opt<bool> EnableLTOInternalization(
74 "enable-lto-internalization", cl::init(true), cl::Hidden,
75 cl::desc("Enable global value internalization in LTO"));
76 }
77
78 // Computes a unique hash for the Module considering the current list of
79 // export/import and other global analysis results.
80 // The hash is produced in \p Key.
computeLTOCacheKey(SmallString<40> & Key,const Config & Conf,const ModuleSummaryIndex & Index,StringRef ModuleID,const FunctionImporter::ImportMapTy & ImportList,const FunctionImporter::ExportSetTy & ExportList,const std::map<GlobalValue::GUID,GlobalValue::LinkageTypes> & ResolvedODR,const GVSummaryMapTy & DefinedGlobals,const std::set<GlobalValue::GUID> & CfiFunctionDefs,const std::set<GlobalValue::GUID> & CfiFunctionDecls)81 void llvm::computeLTOCacheKey(
82 SmallString<40> &Key, const Config &Conf, const ModuleSummaryIndex &Index,
83 StringRef ModuleID, const FunctionImporter::ImportMapTy &ImportList,
84 const FunctionImporter::ExportSetTy &ExportList,
85 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
86 const GVSummaryMapTy &DefinedGlobals,
87 const std::set<GlobalValue::GUID> &CfiFunctionDefs,
88 const std::set<GlobalValue::GUID> &CfiFunctionDecls) {
89 // Compute the unique hash for this entry.
90 // This is based on the current compiler version, the module itself, the
91 // export list, the hash for every single module in the import list, the
92 // list of ResolvedODR for the module, and the list of preserved symbols.
93 SHA1 Hasher;
94
95 // Start with the compiler revision
96 Hasher.update(LLVM_VERSION_STRING);
97 #ifdef LLVM_REVISION
98 Hasher.update(LLVM_REVISION);
99 #endif
100
101 // Include the parts of the LTO configuration that affect code generation.
102 auto AddString = [&](StringRef Str) {
103 Hasher.update(Str);
104 Hasher.update(ArrayRef<uint8_t>{0});
105 };
106 auto AddUnsigned = [&](unsigned I) {
107 uint8_t Data[4];
108 support::endian::write32le(Data, I);
109 Hasher.update(ArrayRef<uint8_t>{Data, 4});
110 };
111 auto AddUint64 = [&](uint64_t I) {
112 uint8_t Data[8];
113 support::endian::write64le(Data, I);
114 Hasher.update(ArrayRef<uint8_t>{Data, 8});
115 };
116 AddString(Conf.CPU);
117 // FIXME: Hash more of Options. For now all clients initialize Options from
118 // command-line flags (which is unsupported in production), but may set
119 // RelaxELFRelocations. The clang driver can also pass FunctionSections,
120 // DataSections and DebuggerTuning via command line flags.
121 AddUnsigned(Conf.Options.RelaxELFRelocations);
122 AddUnsigned(Conf.Options.FunctionSections);
123 AddUnsigned(Conf.Options.DataSections);
124 AddUnsigned((unsigned)Conf.Options.DebuggerTuning);
125 for (auto &A : Conf.MAttrs)
126 AddString(A);
127 if (Conf.RelocModel)
128 AddUnsigned(*Conf.RelocModel);
129 else
130 AddUnsigned(-1);
131 if (Conf.CodeModel)
132 AddUnsigned(*Conf.CodeModel);
133 else
134 AddUnsigned(-1);
135 for (const auto &S : Conf.MllvmArgs)
136 AddString(S);
137 AddUnsigned(Conf.CGOptLevel);
138 AddUnsigned(Conf.CGFileType);
139 AddUnsigned(Conf.OptLevel);
140 AddUnsigned(Conf.Freestanding);
141 AddString(Conf.OptPipeline);
142 AddString(Conf.AAPipeline);
143 AddString(Conf.OverrideTriple);
144 AddString(Conf.DefaultTriple);
145 AddString(Conf.DwoDir);
146
147 // Include the hash for the current module
148 auto ModHash = Index.getModuleHash(ModuleID);
149 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
150
151 std::vector<uint64_t> ExportsGUID;
152 ExportsGUID.reserve(ExportList.size());
153 for (const auto &VI : ExportList) {
154 auto GUID = VI.getGUID();
155 ExportsGUID.push_back(GUID);
156 }
157
158 // Sort the export list elements GUIDs.
159 llvm::sort(ExportsGUID);
160 for (uint64_t GUID : ExportsGUID) {
161 // The export list can impact the internalization, be conservative here
162 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&GUID, sizeof(GUID)));
163 }
164
165 // Include the hash for every module we import functions from. The set of
166 // imported symbols for each module may affect code generation and is
167 // sensitive to link order, so include that as well.
168 using ImportMapIteratorTy = FunctionImporter::ImportMapTy::const_iterator;
169 std::vector<ImportMapIteratorTy> ImportModulesVector;
170 ImportModulesVector.reserve(ImportList.size());
171
172 for (ImportMapIteratorTy It = ImportList.begin(); It != ImportList.end();
173 ++It) {
174 ImportModulesVector.push_back(It);
175 }
176 llvm::sort(ImportModulesVector,
177 [](const ImportMapIteratorTy &Lhs, const ImportMapIteratorTy &Rhs)
178 -> bool { return Lhs->getKey() < Rhs->getKey(); });
179 for (const ImportMapIteratorTy &EntryIt : ImportModulesVector) {
180 auto ModHash = Index.getModuleHash(EntryIt->first());
181 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
182
183 AddUint64(EntryIt->second.size());
184 for (auto &Fn : EntryIt->second)
185 AddUint64(Fn);
186 }
187
188 // Include the hash for the resolved ODR.
189 for (auto &Entry : ResolvedODR) {
190 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
191 sizeof(GlobalValue::GUID)));
192 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
193 sizeof(GlobalValue::LinkageTypes)));
194 }
195
196 // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or
197 // defined in this module.
198 std::set<GlobalValue::GUID> UsedCfiDefs;
199 std::set<GlobalValue::GUID> UsedCfiDecls;
200
201 // Typeids used in this module.
202 std::set<GlobalValue::GUID> UsedTypeIds;
203
204 auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) {
205 if (CfiFunctionDefs.count(ValueGUID))
206 UsedCfiDefs.insert(ValueGUID);
207 if (CfiFunctionDecls.count(ValueGUID))
208 UsedCfiDecls.insert(ValueGUID);
209 };
210
211 auto AddUsedThings = [&](GlobalValueSummary *GS) {
212 if (!GS) return;
213 AddUnsigned(GS->getVisibility());
214 AddUnsigned(GS->isLive());
215 AddUnsigned(GS->canAutoHide());
216 for (const ValueInfo &VI : GS->refs()) {
217 AddUnsigned(VI.isDSOLocal(Index.withDSOLocalPropagation()));
218 AddUsedCfiGlobal(VI.getGUID());
219 }
220 if (auto *GVS = dyn_cast<GlobalVarSummary>(GS)) {
221 AddUnsigned(GVS->maybeReadOnly());
222 AddUnsigned(GVS->maybeWriteOnly());
223 }
224 if (auto *FS = dyn_cast<FunctionSummary>(GS)) {
225 for (auto &TT : FS->type_tests())
226 UsedTypeIds.insert(TT);
227 for (auto &TT : FS->type_test_assume_vcalls())
228 UsedTypeIds.insert(TT.GUID);
229 for (auto &TT : FS->type_checked_load_vcalls())
230 UsedTypeIds.insert(TT.GUID);
231 for (auto &TT : FS->type_test_assume_const_vcalls())
232 UsedTypeIds.insert(TT.VFunc.GUID);
233 for (auto &TT : FS->type_checked_load_const_vcalls())
234 UsedTypeIds.insert(TT.VFunc.GUID);
235 for (auto &ET : FS->calls()) {
236 AddUnsigned(ET.first.isDSOLocal(Index.withDSOLocalPropagation()));
237 AddUsedCfiGlobal(ET.first.getGUID());
238 }
239 }
240 };
241
242 // Include the hash for the linkage type to reflect internalization and weak
243 // resolution, and collect any used type identifier resolutions.
244 for (auto &GS : DefinedGlobals) {
245 GlobalValue::LinkageTypes Linkage = GS.second->linkage();
246 Hasher.update(
247 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage)));
248 AddUsedCfiGlobal(GS.first);
249 AddUsedThings(GS.second);
250 }
251
252 // Imported functions may introduce new uses of type identifier resolutions,
253 // so we need to collect their used resolutions as well.
254 for (auto &ImpM : ImportList)
255 for (auto &ImpF : ImpM.second) {
256 GlobalValueSummary *S = Index.findSummaryInModule(ImpF, ImpM.first());
257 AddUsedThings(S);
258 // If this is an alias, we also care about any types/etc. that the aliasee
259 // may reference.
260 if (auto *AS = dyn_cast_or_null<AliasSummary>(S))
261 AddUsedThings(AS->getBaseObject());
262 }
263
264 auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) {
265 AddString(TId);
266
267 AddUnsigned(S.TTRes.TheKind);
268 AddUnsigned(S.TTRes.SizeM1BitWidth);
269
270 AddUint64(S.TTRes.AlignLog2);
271 AddUint64(S.TTRes.SizeM1);
272 AddUint64(S.TTRes.BitMask);
273 AddUint64(S.TTRes.InlineBits);
274
275 AddUint64(S.WPDRes.size());
276 for (auto &WPD : S.WPDRes) {
277 AddUnsigned(WPD.first);
278 AddUnsigned(WPD.second.TheKind);
279 AddString(WPD.second.SingleImplName);
280
281 AddUint64(WPD.second.ResByArg.size());
282 for (auto &ByArg : WPD.second.ResByArg) {
283 AddUint64(ByArg.first.size());
284 for (uint64_t Arg : ByArg.first)
285 AddUint64(Arg);
286 AddUnsigned(ByArg.second.TheKind);
287 AddUint64(ByArg.second.Info);
288 AddUnsigned(ByArg.second.Byte);
289 AddUnsigned(ByArg.second.Bit);
290 }
291 }
292 };
293
294 // Include the hash for all type identifiers used by this module.
295 for (GlobalValue::GUID TId : UsedTypeIds) {
296 auto TidIter = Index.typeIds().equal_range(TId);
297 for (auto It = TidIter.first; It != TidIter.second; ++It)
298 AddTypeIdSummary(It->second.first, It->second.second);
299 }
300
301 AddUnsigned(UsedCfiDefs.size());
302 for (auto &V : UsedCfiDefs)
303 AddUint64(V);
304
305 AddUnsigned(UsedCfiDecls.size());
306 for (auto &V : UsedCfiDecls)
307 AddUint64(V);
308
309 if (!Conf.SampleProfile.empty()) {
310 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile);
311 if (FileOrErr) {
312 Hasher.update(FileOrErr.get()->getBuffer());
313
314 if (!Conf.ProfileRemapping.empty()) {
315 FileOrErr = MemoryBuffer::getFile(Conf.ProfileRemapping);
316 if (FileOrErr)
317 Hasher.update(FileOrErr.get()->getBuffer());
318 }
319 }
320 }
321
322 Key = toHex(Hasher.result());
323 }
324
thinLTOResolvePrevailingGUID(const Config & C,ValueInfo VI,DenseSet<GlobalValueSummary * > & GlobalInvolvedWithAlias,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing,function_ref<void (StringRef,GlobalValue::GUID,GlobalValue::LinkageTypes)> recordNewLinkage,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols)325 static void thinLTOResolvePrevailingGUID(
326 const Config &C, ValueInfo VI,
327 DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias,
328 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
329 isPrevailing,
330 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
331 recordNewLinkage,
332 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
333 GlobalValue::VisibilityTypes Visibility =
334 C.VisibilityScheme == Config::ELF ? VI.getELFVisibility()
335 : GlobalValue::DefaultVisibility;
336 for (auto &S : VI.getSummaryList()) {
337 GlobalValue::LinkageTypes OriginalLinkage = S->linkage();
338 // Ignore local and appending linkage values since the linker
339 // doesn't resolve them.
340 if (GlobalValue::isLocalLinkage(OriginalLinkage) ||
341 GlobalValue::isAppendingLinkage(S->linkage()))
342 continue;
343 // We need to emit only one of these. The prevailing module will keep it,
344 // but turned into a weak, while the others will drop it when possible.
345 // This is both a compile-time optimization and a correctness
346 // transformation. This is necessary for correctness when we have exported
347 // a reference - we need to convert the linkonce to weak to
348 // ensure a copy is kept to satisfy the exported reference.
349 // FIXME: We may want to split the compile time and correctness
350 // aspects into separate routines.
351 if (isPrevailing(VI.getGUID(), S.get())) {
352 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) {
353 S->setLinkage(GlobalValue::getWeakLinkage(
354 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
355 // The kept copy is eligible for auto-hiding (hidden visibility) if all
356 // copies were (i.e. they were all linkonce_odr global unnamed addr).
357 // If any copy is not (e.g. it was originally weak_odr), then the symbol
358 // must remain externally available (e.g. a weak_odr from an explicitly
359 // instantiated template). Additionally, if it is in the
360 // GUIDPreservedSymbols set, that means that it is visibile outside
361 // the summary (e.g. in a native object or a bitcode file without
362 // summary), and in that case we cannot hide it as it isn't possible to
363 // check all copies.
364 S->setCanAutoHide(VI.canAutoHide() &&
365 !GUIDPreservedSymbols.count(VI.getGUID()));
366 }
367 if (C.VisibilityScheme == Config::FromPrevailing)
368 Visibility = S->getVisibility();
369 }
370 // Alias and aliasee can't be turned into available_externally.
371 else if (!isa<AliasSummary>(S.get()) &&
372 !GlobalInvolvedWithAlias.count(S.get()))
373 S->setLinkage(GlobalValue::AvailableExternallyLinkage);
374
375 // For ELF, set visibility to the computed visibility from summaries. We
376 // don't track visibility from declarations so this may be more relaxed than
377 // the most constraining one.
378 if (C.VisibilityScheme == Config::ELF)
379 S->setVisibility(Visibility);
380
381 if (S->linkage() != OriginalLinkage)
382 recordNewLinkage(S->modulePath(), VI.getGUID(), S->linkage());
383 }
384
385 if (C.VisibilityScheme == Config::FromPrevailing) {
386 for (auto &S : VI.getSummaryList()) {
387 GlobalValue::LinkageTypes OriginalLinkage = S->linkage();
388 if (GlobalValue::isLocalLinkage(OriginalLinkage) ||
389 GlobalValue::isAppendingLinkage(S->linkage()))
390 continue;
391 S->setVisibility(Visibility);
392 }
393 }
394 }
395
396 /// Resolve linkage for prevailing symbols in the \p Index.
397 //
398 // We'd like to drop these functions if they are no longer referenced in the
399 // current module. However there is a chance that another module is still
400 // referencing them because of the import. We make sure we always emit at least
401 // one copy.
thinLTOResolvePrevailingInIndex(const Config & C,ModuleSummaryIndex & Index,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing,function_ref<void (StringRef,GlobalValue::GUID,GlobalValue::LinkageTypes)> recordNewLinkage,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols)402 void llvm::thinLTOResolvePrevailingInIndex(
403 const Config &C, ModuleSummaryIndex &Index,
404 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
405 isPrevailing,
406 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
407 recordNewLinkage,
408 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
409 // We won't optimize the globals that are referenced by an alias for now
410 // Ideally we should turn the alias into a global and duplicate the definition
411 // when needed.
412 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias;
413 for (auto &I : Index)
414 for (auto &S : I.second.SummaryList)
415 if (auto AS = dyn_cast<AliasSummary>(S.get()))
416 GlobalInvolvedWithAlias.insert(&AS->getAliasee());
417
418 for (auto &I : Index)
419 thinLTOResolvePrevailingGUID(C, Index.getValueInfo(I),
420 GlobalInvolvedWithAlias, isPrevailing,
421 recordNewLinkage, GUIDPreservedSymbols);
422 }
423
isWeakObjectWithRWAccess(GlobalValueSummary * GVS)424 static bool isWeakObjectWithRWAccess(GlobalValueSummary *GVS) {
425 if (auto *VarSummary = dyn_cast<GlobalVarSummary>(GVS->getBaseObject()))
426 return !VarSummary->maybeReadOnly() && !VarSummary->maybeWriteOnly() &&
427 (VarSummary->linkage() == GlobalValue::WeakODRLinkage ||
428 VarSummary->linkage() == GlobalValue::LinkOnceODRLinkage);
429 return false;
430 }
431
thinLTOInternalizeAndPromoteGUID(ValueInfo VI,function_ref<bool (StringRef,ValueInfo)> isExported,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing)432 static void thinLTOInternalizeAndPromoteGUID(
433 ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported,
434 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
435 isPrevailing) {
436 for (auto &S : VI.getSummaryList()) {
437 if (isExported(S->modulePath(), VI)) {
438 if (GlobalValue::isLocalLinkage(S->linkage()))
439 S->setLinkage(GlobalValue::ExternalLinkage);
440 } else if (EnableLTOInternalization &&
441 // Ignore local and appending linkage values since the linker
442 // doesn't resolve them.
443 !GlobalValue::isLocalLinkage(S->linkage()) &&
444 (!GlobalValue::isInterposableLinkage(S->linkage()) ||
445 isPrevailing(VI.getGUID(), S.get())) &&
446 S->linkage() != GlobalValue::AppendingLinkage &&
447 // We can't internalize available_externally globals because this
448 // can break function pointer equality.
449 S->linkage() != GlobalValue::AvailableExternallyLinkage &&
450 // Functions and read-only variables with linkonce_odr and
451 // weak_odr linkage can be internalized. We can't internalize
452 // linkonce_odr and weak_odr variables which are both modified
453 // and read somewhere in the program because reads and writes
454 // will become inconsistent.
455 !isWeakObjectWithRWAccess(S.get()))
456 S->setLinkage(GlobalValue::InternalLinkage);
457 }
458 }
459
460 // Update the linkages in the given \p Index to mark exported values
461 // as external and non-exported values as internal.
thinLTOInternalizeAndPromoteInIndex(ModuleSummaryIndex & Index,function_ref<bool (StringRef,ValueInfo)> isExported,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing)462 void llvm::thinLTOInternalizeAndPromoteInIndex(
463 ModuleSummaryIndex &Index,
464 function_ref<bool(StringRef, ValueInfo)> isExported,
465 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
466 isPrevailing) {
467 for (auto &I : Index)
468 thinLTOInternalizeAndPromoteGUID(Index.getValueInfo(I), isExported,
469 isPrevailing);
470 }
471
472 // Requires a destructor for std::vector<InputModule>.
473 InputFile::~InputFile() = default;
474
create(MemoryBufferRef Object)475 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) {
476 std::unique_ptr<InputFile> File(new InputFile);
477
478 Expected<IRSymtabFile> FOrErr = readIRSymtab(Object);
479 if (!FOrErr)
480 return FOrErr.takeError();
481
482 File->TargetTriple = FOrErr->TheReader.getTargetTriple();
483 File->SourceFileName = FOrErr->TheReader.getSourceFileName();
484 File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts();
485 File->DependentLibraries = FOrErr->TheReader.getDependentLibraries();
486 File->ComdatTable = FOrErr->TheReader.getComdatTable();
487
488 for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) {
489 size_t Begin = File->Symbols.size();
490 for (const irsymtab::Reader::SymbolRef &Sym :
491 FOrErr->TheReader.module_symbols(I))
492 // Skip symbols that are irrelevant to LTO. Note that this condition needs
493 // to match the one in Skip() in LTO::addRegularLTO().
494 if (Sym.isGlobal() && !Sym.isFormatSpecific())
495 File->Symbols.push_back(Sym);
496 File->ModuleSymIndices.push_back({Begin, File->Symbols.size()});
497 }
498
499 File->Mods = FOrErr->Mods;
500 File->Strtab = std::move(FOrErr->Strtab);
501 return std::move(File);
502 }
503
getName() const504 StringRef InputFile::getName() const {
505 return Mods[0].getModuleIdentifier();
506 }
507
getSingleBitcodeModule()508 BitcodeModule &InputFile::getSingleBitcodeModule() {
509 assert(Mods.size() == 1 && "Expect only one bitcode module");
510 return Mods[0];
511 }
512
RegularLTOState(unsigned ParallelCodeGenParallelismLevel,const Config & Conf)513 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel,
514 const Config &Conf)
515 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel),
516 Ctx(Conf), CombinedModule(std::make_unique<Module>("ld-temp.o", Ctx)),
517 Mover(std::make_unique<IRMover>(*CombinedModule)) {}
518
ThinLTOState(ThinBackend Backend)519 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend)
520 : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) {
521 if (!Backend)
522 this->Backend =
523 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency());
524 }
525
LTO(Config Conf,ThinBackend Backend,unsigned ParallelCodeGenParallelismLevel)526 LTO::LTO(Config Conf, ThinBackend Backend,
527 unsigned ParallelCodeGenParallelismLevel)
528 : Conf(std::move(Conf)),
529 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf),
530 ThinLTO(std::move(Backend)) {}
531
532 // Requires a destructor for MapVector<BitcodeModule>.
533 LTO::~LTO() = default;
534
535 // Add the symbols in the given module to the GlobalResolutions map, and resolve
536 // their partitions.
addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,ArrayRef<SymbolResolution> Res,unsigned Partition,bool InSummary)537 void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
538 ArrayRef<SymbolResolution> Res,
539 unsigned Partition, bool InSummary) {
540 auto *ResI = Res.begin();
541 auto *ResE = Res.end();
542 (void)ResE;
543 const Triple TT(RegularLTO.CombinedModule->getTargetTriple());
544 for (const InputFile::Symbol &Sym : Syms) {
545 assert(ResI != ResE);
546 SymbolResolution Res = *ResI++;
547
548 StringRef Name = Sym.getName();
549 // Strip the __imp_ prefix from COFF dllimport symbols (similar to the
550 // way they are handled by lld), otherwise we can end up with two
551 // global resolutions (one with and one for a copy of the symbol without).
552 if (TT.isOSBinFormatCOFF() && Name.startswith("__imp_"))
553 Name = Name.substr(strlen("__imp_"));
554 auto &GlobalRes = GlobalResolutions[Name];
555 GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr();
556 if (Res.Prevailing) {
557 assert(!GlobalRes.Prevailing &&
558 "Multiple prevailing defs are not allowed");
559 GlobalRes.Prevailing = true;
560 GlobalRes.IRName = std::string(Sym.getIRName());
561 } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) {
562 // Sometimes it can be two copies of symbol in a module and prevailing
563 // symbol can have no IR name. That might happen if symbol is defined in
564 // module level inline asm block. In case we have multiple modules with
565 // the same symbol we want to use IR name of the prevailing symbol.
566 // Otherwise, if we haven't seen a prevailing symbol, set the name so that
567 // we can later use it to check if there is any prevailing copy in IR.
568 GlobalRes.IRName = std::string(Sym.getIRName());
569 }
570
571 // In rare occasion, the symbol used to initialize GlobalRes has a different
572 // IRName from the inspected Symbol. This can happen on macOS + iOS, when a
573 // symbol is referenced through its mangled name, say @"\01_symbol" while
574 // the IRName is @symbol (the prefix underscore comes from MachO mangling).
575 // In that case, we have the same actual Symbol that can get two different
576 // GUID, leading to some invalid internalization. Workaround this by marking
577 // the GlobalRes external.
578
579 // FIXME: instead of this check, it would be desirable to compute GUIDs
580 // based on mangled name, but this requires an access to the Target Triple
581 // and would be relatively invasive on the codebase.
582 if (GlobalRes.IRName != Sym.getIRName()) {
583 GlobalRes.Partition = GlobalResolution::External;
584 GlobalRes.VisibleOutsideSummary = true;
585 }
586
587 // Set the partition to external if we know it is re-defined by the linker
588 // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a
589 // regular object, is referenced from llvm.compiler.used/llvm.used, or was
590 // already recorded as being referenced from a different partition.
591 if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() ||
592 (GlobalRes.Partition != GlobalResolution::Unknown &&
593 GlobalRes.Partition != Partition)) {
594 GlobalRes.Partition = GlobalResolution::External;
595 } else
596 // First recorded reference, save the current partition.
597 GlobalRes.Partition = Partition;
598
599 // Flag as visible outside of summary if visible from a regular object or
600 // from a module that does not have a summary.
601 GlobalRes.VisibleOutsideSummary |=
602 (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary);
603
604 GlobalRes.ExportDynamic |= Res.ExportDynamic;
605 }
606 }
607
writeToResolutionFile(raw_ostream & OS,InputFile * Input,ArrayRef<SymbolResolution> Res)608 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input,
609 ArrayRef<SymbolResolution> Res) {
610 StringRef Path = Input->getName();
611 OS << Path << '\n';
612 auto ResI = Res.begin();
613 for (const InputFile::Symbol &Sym : Input->symbols()) {
614 assert(ResI != Res.end());
615 SymbolResolution Res = *ResI++;
616
617 OS << "-r=" << Path << ',' << Sym.getName() << ',';
618 if (Res.Prevailing)
619 OS << 'p';
620 if (Res.FinalDefinitionInLinkageUnit)
621 OS << 'l';
622 if (Res.VisibleToRegularObj)
623 OS << 'x';
624 if (Res.LinkerRedefined)
625 OS << 'r';
626 OS << '\n';
627 }
628 OS.flush();
629 assert(ResI == Res.end());
630 }
631
add(std::unique_ptr<InputFile> Input,ArrayRef<SymbolResolution> Res)632 Error LTO::add(std::unique_ptr<InputFile> Input,
633 ArrayRef<SymbolResolution> Res) {
634 assert(!CalledGetMaxTasks);
635
636 if (Conf.ResolutionFile)
637 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res);
638
639 if (RegularLTO.CombinedModule->getTargetTriple().empty()) {
640 RegularLTO.CombinedModule->setTargetTriple(Input->getTargetTriple());
641 if (Triple(Input->getTargetTriple()).isOSBinFormatELF())
642 Conf.VisibilityScheme = Config::ELF;
643 }
644
645 const SymbolResolution *ResI = Res.begin();
646 for (unsigned I = 0; I != Input->Mods.size(); ++I)
647 if (Error Err = addModule(*Input, I, ResI, Res.end()))
648 return Err;
649
650 assert(ResI == Res.end());
651 return Error::success();
652 }
653
addModule(InputFile & Input,unsigned ModI,const SymbolResolution * & ResI,const SymbolResolution * ResE)654 Error LTO::addModule(InputFile &Input, unsigned ModI,
655 const SymbolResolution *&ResI,
656 const SymbolResolution *ResE) {
657 Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo();
658 if (!LTOInfo)
659 return LTOInfo.takeError();
660
661 if (EnableSplitLTOUnit) {
662 // If only some modules were split, flag this in the index so that
663 // we can skip or error on optimizations that need consistently split
664 // modules (whole program devirt and lower type tests).
665 if (*EnableSplitLTOUnit != LTOInfo->EnableSplitLTOUnit)
666 ThinLTO.CombinedIndex.setPartiallySplitLTOUnits();
667 } else
668 EnableSplitLTOUnit = LTOInfo->EnableSplitLTOUnit;
669
670 BitcodeModule BM = Input.Mods[ModI];
671 auto ModSyms = Input.module_symbols(ModI);
672 addModuleToGlobalRes(ModSyms, {ResI, ResE},
673 LTOInfo->IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0,
674 LTOInfo->HasSummary);
675
676 if (LTOInfo->IsThinLTO)
677 return addThinLTO(BM, ModSyms, ResI, ResE);
678
679 RegularLTO.EmptyCombinedModule = false;
680 Expected<RegularLTOState::AddedModule> ModOrErr =
681 addRegularLTO(BM, ModSyms, ResI, ResE);
682 if (!ModOrErr)
683 return ModOrErr.takeError();
684
685 if (!LTOInfo->HasSummary)
686 return linkRegularLTO(std::move(*ModOrErr), /*LivenessFromIndex=*/false);
687
688 // Regular LTO module summaries are added to a dummy module that represents
689 // the combined regular LTO module.
690 if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, "", -1ull))
691 return Err;
692 RegularLTO.ModsWithSummaries.push_back(std::move(*ModOrErr));
693 return Error::success();
694 }
695
696 // Checks whether the given global value is in a non-prevailing comdat
697 // (comdat containing values the linker indicated were not prevailing,
698 // which we then dropped to available_externally), and if so, removes
699 // it from the comdat. This is called for all global values to ensure the
700 // comdat is empty rather than leaving an incomplete comdat. It is needed for
701 // regular LTO modules, in case we are in a mixed-LTO mode (both regular
702 // and thin LTO modules) compilation. Since the regular LTO module will be
703 // linked first in the final native link, we want to make sure the linker
704 // doesn't select any of these incomplete comdats that would be left
705 // in the regular LTO module without this cleanup.
706 static void
handleNonPrevailingComdat(GlobalValue & GV,std::set<const Comdat * > & NonPrevailingComdats)707 handleNonPrevailingComdat(GlobalValue &GV,
708 std::set<const Comdat *> &NonPrevailingComdats) {
709 Comdat *C = GV.getComdat();
710 if (!C)
711 return;
712
713 if (!NonPrevailingComdats.count(C))
714 return;
715
716 // Additionally need to drop all global values from the comdat to
717 // available_externally, to satisfy the COMDAT requirement that all members
718 // are discarded as a unit. The non-local linkage global values avoid
719 // duplicate definition linker errors.
720 GV.setLinkage(GlobalValue::AvailableExternallyLinkage);
721
722 if (auto GO = dyn_cast<GlobalObject>(&GV))
723 GO->setComdat(nullptr);
724 }
725
726 // Add a regular LTO object to the link.
727 // The resulting module needs to be linked into the combined LTO module with
728 // linkRegularLTO.
729 Expected<LTO::RegularLTOState::AddedModule>
addRegularLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)730 LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
731 const SymbolResolution *&ResI,
732 const SymbolResolution *ResE) {
733 RegularLTOState::AddedModule Mod;
734 Expected<std::unique_ptr<Module>> MOrErr =
735 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true,
736 /*IsImporting*/ false);
737 if (!MOrErr)
738 return MOrErr.takeError();
739 Module &M = **MOrErr;
740 Mod.M = std::move(*MOrErr);
741
742 if (Error Err = M.materializeMetadata())
743 return std::move(Err);
744 UpgradeDebugInfo(M);
745
746 ModuleSymbolTable SymTab;
747 SymTab.addModule(&M);
748
749 for (GlobalVariable &GV : M.globals())
750 if (GV.hasAppendingLinkage())
751 Mod.Keep.push_back(&GV);
752
753 DenseSet<GlobalObject *> AliasedGlobals;
754 for (auto &GA : M.aliases())
755 if (GlobalObject *GO = GA.getAliaseeObject())
756 AliasedGlobals.insert(GO);
757
758 // In this function we need IR GlobalValues matching the symbols in Syms
759 // (which is not backed by a module), so we need to enumerate them in the same
760 // order. The symbol enumeration order of a ModuleSymbolTable intentionally
761 // matches the order of an irsymtab, but when we read the irsymtab in
762 // InputFile::create we omit some symbols that are irrelevant to LTO. The
763 // Skip() function skips the same symbols from the module as InputFile does
764 // from the symbol table.
765 auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end();
766 auto Skip = [&]() {
767 while (MsymI != MsymE) {
768 auto Flags = SymTab.getSymbolFlags(*MsymI);
769 if ((Flags & object::BasicSymbolRef::SF_Global) &&
770 !(Flags & object::BasicSymbolRef::SF_FormatSpecific))
771 return;
772 ++MsymI;
773 }
774 };
775 Skip();
776
777 std::set<const Comdat *> NonPrevailingComdats;
778 SmallSet<StringRef, 2> NonPrevailingAsmSymbols;
779 for (const InputFile::Symbol &Sym : Syms) {
780 assert(ResI != ResE);
781 SymbolResolution Res = *ResI++;
782
783 assert(MsymI != MsymE);
784 ModuleSymbolTable::Symbol Msym = *MsymI++;
785 Skip();
786
787 if (GlobalValue *GV = Msym.dyn_cast<GlobalValue *>()) {
788 if (Res.Prevailing) {
789 if (Sym.isUndefined())
790 continue;
791 Mod.Keep.push_back(GV);
792 // For symbols re-defined with linker -wrap and -defsym options,
793 // set the linkage to weak to inhibit IPO. The linkage will be
794 // restored by the linker.
795 if (Res.LinkerRedefined)
796 GV->setLinkage(GlobalValue::WeakAnyLinkage);
797
798 GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage();
799 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage))
800 GV->setLinkage(GlobalValue::getWeakLinkage(
801 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
802 } else if (isa<GlobalObject>(GV) &&
803 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() ||
804 GV->hasAvailableExternallyLinkage()) &&
805 !AliasedGlobals.count(cast<GlobalObject>(GV))) {
806 // Any of the above three types of linkage indicates that the
807 // chosen prevailing symbol will have the same semantics as this copy of
808 // the symbol, so we may be able to link it with available_externally
809 // linkage. We will decide later whether to do that when we link this
810 // module (in linkRegularLTO), based on whether it is undefined.
811 Mod.Keep.push_back(GV);
812 GV->setLinkage(GlobalValue::AvailableExternallyLinkage);
813 if (GV->hasComdat())
814 NonPrevailingComdats.insert(GV->getComdat());
815 cast<GlobalObject>(GV)->setComdat(nullptr);
816 }
817
818 // Set the 'local' flag based on the linker resolution for this symbol.
819 if (Res.FinalDefinitionInLinkageUnit) {
820 GV->setDSOLocal(true);
821 if (GV->hasDLLImportStorageClass())
822 GV->setDLLStorageClass(GlobalValue::DLLStorageClassTypes::
823 DefaultStorageClass);
824 }
825 } else if (auto *AS = Msym.dyn_cast<ModuleSymbolTable::AsmSymbol *>()) {
826 // Collect non-prevailing symbols.
827 if (!Res.Prevailing)
828 NonPrevailingAsmSymbols.insert(AS->first);
829 } else {
830 llvm_unreachable("unknown symbol type");
831 }
832
833 // Common resolution: collect the maximum size/alignment over all commons.
834 // We also record if we see an instance of a common as prevailing, so that
835 // if none is prevailing we can ignore it later.
836 if (Sym.isCommon()) {
837 // FIXME: We should figure out what to do about commons defined by asm.
838 // For now they aren't reported correctly by ModuleSymbolTable.
839 auto &CommonRes = RegularLTO.Commons[std::string(Sym.getIRName())];
840 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize());
841 if (uint32_t SymAlignValue = Sym.getCommonAlignment()) {
842 const Align SymAlign(SymAlignValue);
843 CommonRes.Align = std::max(SymAlign, CommonRes.Align.valueOrOne());
844 }
845 CommonRes.Prevailing |= Res.Prevailing;
846 }
847 }
848
849 if (!M.getComdatSymbolTable().empty())
850 for (GlobalValue &GV : M.global_values())
851 handleNonPrevailingComdat(GV, NonPrevailingComdats);
852
853 // Prepend ".lto_discard <sym>, <sym>*" directive to each module inline asm
854 // block.
855 if (!M.getModuleInlineAsm().empty()) {
856 std::string NewIA = ".lto_discard";
857 if (!NonPrevailingAsmSymbols.empty()) {
858 // Don't dicard a symbol if there is a live .symver for it.
859 ModuleSymbolTable::CollectAsmSymvers(
860 M, [&](StringRef Name, StringRef Alias) {
861 if (!NonPrevailingAsmSymbols.count(Alias))
862 NonPrevailingAsmSymbols.erase(Name);
863 });
864 NewIA += " " + llvm::join(NonPrevailingAsmSymbols, ", ");
865 }
866 NewIA += "\n";
867 M.setModuleInlineAsm(NewIA + M.getModuleInlineAsm());
868 }
869
870 assert(MsymI == MsymE);
871 return std::move(Mod);
872 }
873
linkRegularLTO(RegularLTOState::AddedModule Mod,bool LivenessFromIndex)874 Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod,
875 bool LivenessFromIndex) {
876 std::vector<GlobalValue *> Keep;
877 for (GlobalValue *GV : Mod.Keep) {
878 if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GV->getGUID())) {
879 if (Function *F = dyn_cast<Function>(GV)) {
880 if (DiagnosticOutputFile) {
881 if (Error Err = F->materialize())
882 return Err;
883 OptimizationRemarkEmitter ORE(F, nullptr);
884 ORE.emit(OptimizationRemark(DEBUG_TYPE, "deadfunction", F)
885 << ore::NV("Function", F)
886 << " not added to the combined module ");
887 }
888 }
889 continue;
890 }
891
892 if (!GV->hasAvailableExternallyLinkage()) {
893 Keep.push_back(GV);
894 continue;
895 }
896
897 // Only link available_externally definitions if we don't already have a
898 // definition.
899 GlobalValue *CombinedGV =
900 RegularLTO.CombinedModule->getNamedValue(GV->getName());
901 if (CombinedGV && !CombinedGV->isDeclaration())
902 continue;
903
904 Keep.push_back(GV);
905 }
906
907 return RegularLTO.Mover->move(std::move(Mod.M), Keep, nullptr,
908 /* IsPerformingImport */ false);
909 }
910
911 // Add a ThinLTO module to the link.
addThinLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)912 Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
913 const SymbolResolution *&ResI,
914 const SymbolResolution *ResE) {
915 const SymbolResolution *ResITmp = ResI;
916 for (const InputFile::Symbol &Sym : Syms) {
917 assert(ResITmp != ResE);
918 SymbolResolution Res = *ResITmp++;
919
920 if (!Sym.getIRName().empty()) {
921 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier(
922 Sym.getIRName(), GlobalValue::ExternalLinkage, ""));
923 if (Res.Prevailing)
924 ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier();
925 }
926 }
927
928 if (Error Err =
929 BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(),
930 ThinLTO.ModuleMap.size(), [&](GlobalValue::GUID GUID) {
931 return ThinLTO.PrevailingModuleForGUID[GUID] ==
932 BM.getModuleIdentifier();
933 }))
934 return Err;
935
936 for (const InputFile::Symbol &Sym : Syms) {
937 assert(ResI != ResE);
938 SymbolResolution Res = *ResI++;
939
940 if (!Sym.getIRName().empty()) {
941 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier(
942 Sym.getIRName(), GlobalValue::ExternalLinkage, ""));
943 if (Res.Prevailing) {
944 assert(ThinLTO.PrevailingModuleForGUID[GUID] ==
945 BM.getModuleIdentifier());
946
947 // For linker redefined symbols (via --wrap or --defsym) we want to
948 // switch the linkage to `weak` to prevent IPOs from happening.
949 // Find the summary in the module for this very GV and record the new
950 // linkage so that we can switch it when we import the GV.
951 if (Res.LinkerRedefined)
952 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
953 GUID, BM.getModuleIdentifier()))
954 S->setLinkage(GlobalValue::WeakAnyLinkage);
955 }
956
957 // If the linker resolved the symbol to a local definition then mark it
958 // as local in the summary for the module we are adding.
959 if (Res.FinalDefinitionInLinkageUnit) {
960 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
961 GUID, BM.getModuleIdentifier())) {
962 S->setDSOLocal(true);
963 }
964 }
965 }
966 }
967
968 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second)
969 return make_error<StringError>(
970 "Expected at most one ThinLTO module per bitcode file",
971 inconvertibleErrorCode());
972
973 if (!Conf.ThinLTOModulesToCompile.empty()) {
974 if (!ThinLTO.ModulesToCompile)
975 ThinLTO.ModulesToCompile = ModuleMapType();
976 // This is a fuzzy name matching where only modules with name containing the
977 // specified switch values are going to be compiled.
978 for (const std::string &Name : Conf.ThinLTOModulesToCompile) {
979 if (BM.getModuleIdentifier().contains(Name)) {
980 ThinLTO.ModulesToCompile->insert({BM.getModuleIdentifier(), BM});
981 llvm::errs() << "[ThinLTO] Selecting " << BM.getModuleIdentifier()
982 << " to compile\n";
983 }
984 }
985 }
986
987 return Error::success();
988 }
989
getMaxTasks() const990 unsigned LTO::getMaxTasks() const {
991 CalledGetMaxTasks = true;
992 auto ModuleCount = ThinLTO.ModulesToCompile ? ThinLTO.ModulesToCompile->size()
993 : ThinLTO.ModuleMap.size();
994 return RegularLTO.ParallelCodeGenParallelismLevel + ModuleCount;
995 }
996
997 // If only some of the modules were split, we cannot correctly handle
998 // code that contains type tests or type checked loads.
checkPartiallySplit()999 Error LTO::checkPartiallySplit() {
1000 if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits())
1001 return Error::success();
1002
1003 Function *TypeTestFunc = RegularLTO.CombinedModule->getFunction(
1004 Intrinsic::getName(Intrinsic::type_test));
1005 Function *TypeCheckedLoadFunc = RegularLTO.CombinedModule->getFunction(
1006 Intrinsic::getName(Intrinsic::type_checked_load));
1007
1008 // First check if there are type tests / type checked loads in the
1009 // merged regular LTO module IR.
1010 if ((TypeTestFunc && !TypeTestFunc->use_empty()) ||
1011 (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()))
1012 return make_error<StringError>(
1013 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
1014 inconvertibleErrorCode());
1015
1016 // Otherwise check if there are any recorded in the combined summary from the
1017 // ThinLTO modules.
1018 for (auto &P : ThinLTO.CombinedIndex) {
1019 for (auto &S : P.second.SummaryList) {
1020 auto *FS = dyn_cast<FunctionSummary>(S.get());
1021 if (!FS)
1022 continue;
1023 if (!FS->type_test_assume_vcalls().empty() ||
1024 !FS->type_checked_load_vcalls().empty() ||
1025 !FS->type_test_assume_const_vcalls().empty() ||
1026 !FS->type_checked_load_const_vcalls().empty() ||
1027 !FS->type_tests().empty())
1028 return make_error<StringError>(
1029 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
1030 inconvertibleErrorCode());
1031 }
1032 }
1033 return Error::success();
1034 }
1035
run(AddStreamFn AddStream,FileCache Cache)1036 Error LTO::run(AddStreamFn AddStream, FileCache Cache) {
1037 // Compute "dead" symbols, we don't want to import/export these!
1038 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols;
1039 DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions;
1040 for (auto &Res : GlobalResolutions) {
1041 // Normally resolution have IR name of symbol. We can do nothing here
1042 // otherwise. See comments in GlobalResolution struct for more details.
1043 if (Res.second.IRName.empty())
1044 continue;
1045
1046 GlobalValue::GUID GUID = GlobalValue::getGUID(
1047 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
1048
1049 if (Res.second.VisibleOutsideSummary && Res.second.Prevailing)
1050 GUIDPreservedSymbols.insert(GUID);
1051
1052 if (Res.second.ExportDynamic)
1053 DynamicExportSymbols.insert(GUID);
1054
1055 GUIDPrevailingResolutions[GUID] =
1056 Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No;
1057 }
1058
1059 auto isPrevailing = [&](GlobalValue::GUID G) {
1060 auto It = GUIDPrevailingResolutions.find(G);
1061 if (It == GUIDPrevailingResolutions.end())
1062 return PrevailingType::Unknown;
1063 return It->second;
1064 };
1065 computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols,
1066 isPrevailing, Conf.OptLevel > 0);
1067
1068 // Setup output file to emit statistics.
1069 auto StatsFileOrErr = setupStatsFile(Conf.StatsFile);
1070 if (!StatsFileOrErr)
1071 return StatsFileOrErr.takeError();
1072 std::unique_ptr<ToolOutputFile> StatsFile = std::move(StatsFileOrErr.get());
1073
1074 Error Result = runRegularLTO(AddStream);
1075 if (!Result)
1076 Result = runThinLTO(AddStream, Cache, GUIDPreservedSymbols);
1077
1078 if (StatsFile)
1079 PrintStatisticsJSON(StatsFile->os());
1080
1081 return Result;
1082 }
1083
runRegularLTO(AddStreamFn AddStream)1084 Error LTO::runRegularLTO(AddStreamFn AddStream) {
1085 // Setup optimization remarks.
1086 auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks(
1087 RegularLTO.CombinedModule->getContext(), Conf.RemarksFilename,
1088 Conf.RemarksPasses, Conf.RemarksFormat, Conf.RemarksWithHotness,
1089 Conf.RemarksHotnessThreshold);
1090 if (!DiagFileOrErr)
1091 return DiagFileOrErr.takeError();
1092 DiagnosticOutputFile = std::move(*DiagFileOrErr);
1093
1094 // Finalize linking of regular LTO modules containing summaries now that
1095 // we have computed liveness information.
1096 for (auto &M : RegularLTO.ModsWithSummaries)
1097 if (Error Err = linkRegularLTO(std::move(M),
1098 /*LivenessFromIndex=*/true))
1099 return Err;
1100
1101 // Ensure we don't have inconsistently split LTO units with type tests.
1102 // FIXME: this checks both LTO and ThinLTO. It happens to work as we take
1103 // this path both cases but eventually this should be split into two and
1104 // do the ThinLTO checks in `runThinLTO`.
1105 if (Error Err = checkPartiallySplit())
1106 return Err;
1107
1108 // Make sure commons have the right size/alignment: we kept the largest from
1109 // all the prevailing when adding the inputs, and we apply it here.
1110 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout();
1111 for (auto &I : RegularLTO.Commons) {
1112 if (!I.second.Prevailing)
1113 // Don't do anything if no instance of this common was prevailing.
1114 continue;
1115 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first);
1116 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) {
1117 // Don't create a new global if the type is already correct, just make
1118 // sure the alignment is correct.
1119 OldGV->setAlignment(I.second.Align);
1120 continue;
1121 }
1122 ArrayType *Ty =
1123 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size);
1124 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false,
1125 GlobalValue::CommonLinkage,
1126 ConstantAggregateZero::get(Ty), "");
1127 GV->setAlignment(I.second.Align);
1128 if (OldGV) {
1129 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType()));
1130 GV->takeName(OldGV);
1131 OldGV->eraseFromParent();
1132 } else {
1133 GV->setName(I.first);
1134 }
1135 }
1136
1137 // If allowed, upgrade public vcall visibility metadata to linkage unit
1138 // visibility before whole program devirtualization in the optimizer.
1139 updateVCallVisibilityInModule(*RegularLTO.CombinedModule,
1140 Conf.HasWholeProgramVisibility,
1141 DynamicExportSymbols);
1142 updatePublicTypeTestCalls(*RegularLTO.CombinedModule,
1143 Conf.HasWholeProgramVisibility);
1144
1145 if (Conf.PreOptModuleHook &&
1146 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule))
1147 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
1148
1149 if (!Conf.CodeGenOnly) {
1150 for (const auto &R : GlobalResolutions) {
1151 if (!R.second.isPrevailingIRSymbol())
1152 continue;
1153 if (R.second.Partition != 0 &&
1154 R.second.Partition != GlobalResolution::External)
1155 continue;
1156
1157 GlobalValue *GV =
1158 RegularLTO.CombinedModule->getNamedValue(R.second.IRName);
1159 // Ignore symbols defined in other partitions.
1160 // Also skip declarations, which are not allowed to have internal linkage.
1161 if (!GV || GV->hasLocalLinkage() || GV->isDeclaration())
1162 continue;
1163 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global
1164 : GlobalValue::UnnamedAddr::None);
1165 if (EnableLTOInternalization && R.second.Partition == 0)
1166 GV->setLinkage(GlobalValue::InternalLinkage);
1167 }
1168
1169 RegularLTO.CombinedModule->addModuleFlag(Module::Error, "LTOPostLink", 1);
1170
1171 if (Conf.PostInternalizeModuleHook &&
1172 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule))
1173 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
1174 }
1175
1176 if (!RegularLTO.EmptyCombinedModule || Conf.AlwaysEmitRegularLTOObj) {
1177 if (Error Err =
1178 backend(Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel,
1179 *RegularLTO.CombinedModule, ThinLTO.CombinedIndex))
1180 return Err;
1181 }
1182
1183 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
1184 }
1185
1186 static const char *libcallRoutineNames[] = {
1187 #define HANDLE_LIBCALL(code, name) name,
1188 #include "llvm/IR/RuntimeLibcalls.def"
1189 #undef HANDLE_LIBCALL
1190 };
1191
getRuntimeLibcallSymbols()1192 ArrayRef<const char*> LTO::getRuntimeLibcallSymbols() {
1193 return ArrayRef(libcallRoutineNames);
1194 }
1195
1196 /// This class defines the interface to the ThinLTO backend.
1197 class lto::ThinBackendProc {
1198 protected:
1199 const Config &Conf;
1200 ModuleSummaryIndex &CombinedIndex;
1201 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries;
1202 lto::IndexWriteCallback OnWrite;
1203 bool ShouldEmitImportsFiles;
1204
1205 public:
ThinBackendProc(const Config & Conf,ModuleSummaryIndex & CombinedIndex,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,lto::IndexWriteCallback OnWrite,bool ShouldEmitImportsFiles)1206 ThinBackendProc(const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1207 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1208 lto::IndexWriteCallback OnWrite, bool ShouldEmitImportsFiles)
1209 : Conf(Conf), CombinedIndex(CombinedIndex),
1210 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries),
1211 OnWrite(OnWrite), ShouldEmitImportsFiles(ShouldEmitImportsFiles) {}
1212
1213 virtual ~ThinBackendProc() = default;
1214 virtual Error start(
1215 unsigned Task, BitcodeModule BM,
1216 const FunctionImporter::ImportMapTy &ImportList,
1217 const FunctionImporter::ExportSetTy &ExportList,
1218 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1219 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0;
1220 virtual Error wait() = 0;
1221 virtual unsigned getThreadCount() = 0;
1222
1223 // Write sharded indices and (optionally) imports to disk
emitFiles(const FunctionImporter::ImportMapTy & ImportList,llvm::StringRef ModulePath,const std::string & NewModulePath)1224 Error emitFiles(const FunctionImporter::ImportMapTy &ImportList,
1225 llvm::StringRef ModulePath,
1226 const std::string &NewModulePath) {
1227 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
1228 std::error_code EC;
1229 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
1230 ImportList, ModuleToSummariesForIndex);
1231
1232 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC,
1233 sys::fs::OpenFlags::OF_None);
1234 if (EC)
1235 return errorCodeToError(EC);
1236 writeIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex);
1237
1238 if (ShouldEmitImportsFiles) {
1239 EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports",
1240 ModuleToSummariesForIndex);
1241 if (EC)
1242 return errorCodeToError(EC);
1243 }
1244 return Error::success();
1245 }
1246 };
1247
1248 namespace {
1249 class InProcessThinBackend : public ThinBackendProc {
1250 ThreadPool BackendThreadPool;
1251 AddStreamFn AddStream;
1252 FileCache Cache;
1253 std::set<GlobalValue::GUID> CfiFunctionDefs;
1254 std::set<GlobalValue::GUID> CfiFunctionDecls;
1255
1256 std::optional<Error> Err;
1257 std::mutex ErrMu;
1258
1259 bool ShouldEmitIndexFiles;
1260
1261 public:
InProcessThinBackend(const Config & Conf,ModuleSummaryIndex & CombinedIndex,ThreadPoolStrategy ThinLTOParallelism,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,AddStreamFn AddStream,FileCache Cache,lto::IndexWriteCallback OnWrite,bool ShouldEmitIndexFiles,bool ShouldEmitImportsFiles)1262 InProcessThinBackend(
1263 const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1264 ThreadPoolStrategy ThinLTOParallelism,
1265 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1266 AddStreamFn AddStream, FileCache Cache, lto::IndexWriteCallback OnWrite,
1267 bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles)
1268 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
1269 OnWrite, ShouldEmitImportsFiles),
1270 BackendThreadPool(ThinLTOParallelism), AddStream(std::move(AddStream)),
1271 Cache(std::move(Cache)), ShouldEmitIndexFiles(ShouldEmitIndexFiles) {
1272 for (auto &Name : CombinedIndex.cfiFunctionDefs())
1273 CfiFunctionDefs.insert(
1274 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
1275 for (auto &Name : CombinedIndex.cfiFunctionDecls())
1276 CfiFunctionDecls.insert(
1277 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
1278 }
1279
runThinLTOBackendThread(AddStreamFn AddStream,FileCache Cache,unsigned Task,BitcodeModule BM,ModuleSummaryIndex & CombinedIndex,const FunctionImporter::ImportMapTy & ImportList,const FunctionImporter::ExportSetTy & ExportList,const std::map<GlobalValue::GUID,GlobalValue::LinkageTypes> & ResolvedODR,const GVSummaryMapTy & DefinedGlobals,MapVector<StringRef,BitcodeModule> & ModuleMap)1280 Error runThinLTOBackendThread(
1281 AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM,
1282 ModuleSummaryIndex &CombinedIndex,
1283 const FunctionImporter::ImportMapTy &ImportList,
1284 const FunctionImporter::ExportSetTy &ExportList,
1285 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1286 const GVSummaryMapTy &DefinedGlobals,
1287 MapVector<StringRef, BitcodeModule> &ModuleMap) {
1288 auto RunThinBackend = [&](AddStreamFn AddStream) {
1289 LTOLLVMContext BackendContext(Conf);
1290 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext);
1291 if (!MOrErr)
1292 return MOrErr.takeError();
1293
1294 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex,
1295 ImportList, DefinedGlobals, &ModuleMap);
1296 };
1297
1298 auto ModuleID = BM.getModuleIdentifier();
1299
1300 if (ShouldEmitIndexFiles) {
1301 if (auto E = emitFiles(ImportList, ModuleID, ModuleID.str()))
1302 return E;
1303 }
1304
1305 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) ||
1306 all_of(CombinedIndex.getModuleHash(ModuleID),
1307 [](uint32_t V) { return V == 0; }))
1308 // Cache disabled or no entry for this module in the combined index or
1309 // no module hash.
1310 return RunThinBackend(AddStream);
1311
1312 SmallString<40> Key;
1313 // The module may be cached, this helps handling it.
1314 computeLTOCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList,
1315 ExportList, ResolvedODR, DefinedGlobals, CfiFunctionDefs,
1316 CfiFunctionDecls);
1317 Expected<AddStreamFn> CacheAddStreamOrErr = Cache(Task, Key, ModuleID);
1318 if (Error Err = CacheAddStreamOrErr.takeError())
1319 return Err;
1320 AddStreamFn &CacheAddStream = *CacheAddStreamOrErr;
1321 if (CacheAddStream)
1322 return RunThinBackend(CacheAddStream);
1323
1324 return Error::success();
1325 }
1326
start(unsigned Task,BitcodeModule BM,const FunctionImporter::ImportMapTy & ImportList,const FunctionImporter::ExportSetTy & ExportList,const std::map<GlobalValue::GUID,GlobalValue::LinkageTypes> & ResolvedODR,MapVector<StringRef,BitcodeModule> & ModuleMap)1327 Error start(
1328 unsigned Task, BitcodeModule BM,
1329 const FunctionImporter::ImportMapTy &ImportList,
1330 const FunctionImporter::ExportSetTy &ExportList,
1331 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1332 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
1333 StringRef ModulePath = BM.getModuleIdentifier();
1334 assert(ModuleToDefinedGVSummaries.count(ModulePath));
1335 const GVSummaryMapTy &DefinedGlobals =
1336 ModuleToDefinedGVSummaries.find(ModulePath)->second;
1337 BackendThreadPool.async(
1338 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex,
1339 const FunctionImporter::ImportMapTy &ImportList,
1340 const FunctionImporter::ExportSetTy &ExportList,
1341 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>
1342 &ResolvedODR,
1343 const GVSummaryMapTy &DefinedGlobals,
1344 MapVector<StringRef, BitcodeModule> &ModuleMap) {
1345 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled)
1346 timeTraceProfilerInitialize(Conf.TimeTraceGranularity,
1347 "thin backend");
1348 Error E = runThinLTOBackendThread(
1349 AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList,
1350 ResolvedODR, DefinedGlobals, ModuleMap);
1351 if (E) {
1352 std::unique_lock<std::mutex> L(ErrMu);
1353 if (Err)
1354 Err = joinErrors(std::move(*Err), std::move(E));
1355 else
1356 Err = std::move(E);
1357 }
1358 if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled)
1359 timeTraceProfilerFinishThread();
1360 },
1361 BM, std::ref(CombinedIndex), std::ref(ImportList), std::ref(ExportList),
1362 std::ref(ResolvedODR), std::ref(DefinedGlobals), std::ref(ModuleMap));
1363
1364 if (OnWrite)
1365 OnWrite(std::string(ModulePath));
1366 return Error::success();
1367 }
1368
wait()1369 Error wait() override {
1370 BackendThreadPool.wait();
1371 if (Err)
1372 return std::move(*Err);
1373 else
1374 return Error::success();
1375 }
1376
getThreadCount()1377 unsigned getThreadCount() override {
1378 return BackendThreadPool.getThreadCount();
1379 }
1380 };
1381 } // end anonymous namespace
1382
createInProcessThinBackend(ThreadPoolStrategy Parallelism,lto::IndexWriteCallback OnWrite,bool ShouldEmitIndexFiles,bool ShouldEmitImportsFiles)1383 ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism,
1384 lto::IndexWriteCallback OnWrite,
1385 bool ShouldEmitIndexFiles,
1386 bool ShouldEmitImportsFiles) {
1387 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1388 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1389 AddStreamFn AddStream, FileCache Cache) {
1390 return std::make_unique<InProcessThinBackend>(
1391 Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries, AddStream,
1392 Cache, OnWrite, ShouldEmitIndexFiles, ShouldEmitImportsFiles);
1393 };
1394 }
1395
1396 // Given the original \p Path to an output file, replace any path
1397 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the
1398 // resulting directory if it does not yet exist.
getThinLTOOutputFile(const std::string & Path,const std::string & OldPrefix,const std::string & NewPrefix)1399 std::string lto::getThinLTOOutputFile(const std::string &Path,
1400 const std::string &OldPrefix,
1401 const std::string &NewPrefix) {
1402 if (OldPrefix.empty() && NewPrefix.empty())
1403 return Path;
1404 SmallString<128> NewPath(Path);
1405 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix);
1406 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str());
1407 if (!ParentPath.empty()) {
1408 // Make sure the new directory exists, creating it if necessary.
1409 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath))
1410 llvm::errs() << "warning: could not create directory '" << ParentPath
1411 << "': " << EC.message() << '\n';
1412 }
1413 return std::string(NewPath.str());
1414 }
1415
1416 namespace {
1417 class WriteIndexesThinBackend : public ThinBackendProc {
1418 std::string OldPrefix, NewPrefix;
1419 raw_fd_ostream *LinkedObjectsFile;
1420
1421 public:
WriteIndexesThinBackend(const Config & Conf,ModuleSummaryIndex & CombinedIndex,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,std::string OldPrefix,std::string NewPrefix,bool ShouldEmitImportsFiles,raw_fd_ostream * LinkedObjectsFile,lto::IndexWriteCallback OnWrite)1422 WriteIndexesThinBackend(
1423 const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1424 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1425 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1426 raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite)
1427 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
1428 OnWrite, ShouldEmitImportsFiles),
1429 OldPrefix(OldPrefix), NewPrefix(NewPrefix),
1430 LinkedObjectsFile(LinkedObjectsFile) {}
1431
start(unsigned Task,BitcodeModule BM,const FunctionImporter::ImportMapTy & ImportList,const FunctionImporter::ExportSetTy & ExportList,const std::map<GlobalValue::GUID,GlobalValue::LinkageTypes> & ResolvedODR,MapVector<StringRef,BitcodeModule> & ModuleMap)1432 Error start(
1433 unsigned Task, BitcodeModule BM,
1434 const FunctionImporter::ImportMapTy &ImportList,
1435 const FunctionImporter::ExportSetTy &ExportList,
1436 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1437 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
1438 StringRef ModulePath = BM.getModuleIdentifier();
1439 std::string NewModulePath =
1440 getThinLTOOutputFile(std::string(ModulePath), OldPrefix, NewPrefix);
1441
1442 if (LinkedObjectsFile)
1443 *LinkedObjectsFile << NewModulePath << '\n';
1444
1445 if (auto E = emitFiles(ImportList, ModulePath, NewModulePath))
1446 return E;
1447
1448 if (OnWrite)
1449 OnWrite(std::string(ModulePath));
1450 return Error::success();
1451 }
1452
wait()1453 Error wait() override { return Error::success(); }
1454
1455 // WriteIndexesThinBackend should always return 1 to prevent module
1456 // re-ordering and avoid non-determinism in the final link.
getThreadCount()1457 unsigned getThreadCount() override { return 1; }
1458 };
1459 } // end anonymous namespace
1460
createWriteIndexesThinBackend(std::string OldPrefix,std::string NewPrefix,bool ShouldEmitImportsFiles,raw_fd_ostream * LinkedObjectsFile,IndexWriteCallback OnWrite)1461 ThinBackend lto::createWriteIndexesThinBackend(
1462 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1463 raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) {
1464 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1465 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1466 AddStreamFn AddStream, FileCache Cache) {
1467 return std::make_unique<WriteIndexesThinBackend>(
1468 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix,
1469 ShouldEmitImportsFiles, LinkedObjectsFile, OnWrite);
1470 };
1471 }
1472
runThinLTO(AddStreamFn AddStream,FileCache Cache,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols)1473 Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
1474 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
1475 ThinLTO.CombinedIndex.releaseTemporaryMemory();
1476 timeTraceProfilerBegin("ThinLink", StringRef(""));
1477 auto TimeTraceScopeExit = llvm::make_scope_exit([]() {
1478 if (llvm::timeTraceProfilerEnabled())
1479 llvm::timeTraceProfilerEnd();
1480 });
1481 if (ThinLTO.ModuleMap.empty())
1482 return Error::success();
1483
1484 if (ThinLTO.ModulesToCompile && ThinLTO.ModulesToCompile->empty()) {
1485 llvm::errs() << "warning: [ThinLTO] No module compiled\n";
1486 return Error::success();
1487 }
1488
1489 if (Conf.CombinedIndexHook &&
1490 !Conf.CombinedIndexHook(ThinLTO.CombinedIndex, GUIDPreservedSymbols))
1491 return Error::success();
1492
1493 // Collect for each module the list of function it defines (GUID ->
1494 // Summary).
1495 StringMap<GVSummaryMapTy>
1496 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size());
1497 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule(
1498 ModuleToDefinedGVSummaries);
1499 // Create entries for any modules that didn't have any GV summaries
1500 // (either they didn't have any GVs to start with, or we suppressed
1501 // generation of the summaries because they e.g. had inline assembly
1502 // uses that couldn't be promoted/renamed on export). This is so
1503 // InProcessThinBackend::start can still launch a backend thread, which
1504 // is passed the map of summaries for the module, without any special
1505 // handling for this case.
1506 for (auto &Mod : ThinLTO.ModuleMap)
1507 if (!ModuleToDefinedGVSummaries.count(Mod.first))
1508 ModuleToDefinedGVSummaries.try_emplace(Mod.first);
1509
1510 // Synthesize entry counts for functions in the CombinedIndex.
1511 computeSyntheticCounts(ThinLTO.CombinedIndex);
1512
1513 StringMap<FunctionImporter::ImportMapTy> ImportLists(
1514 ThinLTO.ModuleMap.size());
1515 StringMap<FunctionImporter::ExportSetTy> ExportLists(
1516 ThinLTO.ModuleMap.size());
1517 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
1518
1519 if (DumpThinCGSCCs)
1520 ThinLTO.CombinedIndex.dumpSCCs(outs());
1521
1522 std::set<GlobalValue::GUID> ExportedGUIDs;
1523
1524 if (hasWholeProgramVisibility(Conf.HasWholeProgramVisibility))
1525 ThinLTO.CombinedIndex.setWithWholeProgramVisibility();
1526 // If allowed, upgrade public vcall visibility to linkage unit visibility in
1527 // the summaries before whole program devirtualization below.
1528 updateVCallVisibilityInIndex(ThinLTO.CombinedIndex,
1529 Conf.HasWholeProgramVisibility,
1530 DynamicExportSymbols);
1531
1532 // Perform index-based WPD. This will return immediately if there are
1533 // no index entries in the typeIdMetadata map (e.g. if we are instead
1534 // performing IR-based WPD in hybrid regular/thin LTO mode).
1535 std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap;
1536 runWholeProgramDevirtOnIndex(ThinLTO.CombinedIndex, ExportedGUIDs,
1537 LocalWPDTargetsMap);
1538
1539 if (Conf.OptLevel > 0)
1540 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1541 ImportLists, ExportLists);
1542
1543 // Figure out which symbols need to be internalized. This also needs to happen
1544 // at -O0 because summary-based DCE is implemented using internalization, and
1545 // we must apply DCE consistently with the full LTO module in order to avoid
1546 // undefined references during the final link.
1547 for (auto &Res : GlobalResolutions) {
1548 // If the symbol does not have external references or it is not prevailing,
1549 // then not need to mark it as exported from a ThinLTO partition.
1550 if (Res.second.Partition != GlobalResolution::External ||
1551 !Res.second.isPrevailingIRSymbol())
1552 continue;
1553 auto GUID = GlobalValue::getGUID(
1554 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
1555 // Mark exported unless index-based analysis determined it to be dead.
1556 if (ThinLTO.CombinedIndex.isGUIDLive(GUID))
1557 ExportedGUIDs.insert(GUID);
1558 }
1559
1560 // Any functions referenced by the jump table in the regular LTO object must
1561 // be exported.
1562 for (auto &Def : ThinLTO.CombinedIndex.cfiFunctionDefs())
1563 ExportedGUIDs.insert(
1564 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Def)));
1565 for (auto &Decl : ThinLTO.CombinedIndex.cfiFunctionDecls())
1566 ExportedGUIDs.insert(
1567 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Decl)));
1568
1569 auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) {
1570 const auto &ExportList = ExportLists.find(ModuleIdentifier);
1571 return (ExportList != ExportLists.end() && ExportList->second.count(VI)) ||
1572 ExportedGUIDs.count(VI.getGUID());
1573 };
1574
1575 // Update local devirtualized targets that were exported by cross-module
1576 // importing or by other devirtualizations marked in the ExportedGUIDs set.
1577 updateIndexWPDForExports(ThinLTO.CombinedIndex, isExported,
1578 LocalWPDTargetsMap);
1579
1580 auto isPrevailing = [&](GlobalValue::GUID GUID,
1581 const GlobalValueSummary *S) {
1582 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath();
1583 };
1584 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported,
1585 isPrevailing);
1586
1587 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
1588 GlobalValue::GUID GUID,
1589 GlobalValue::LinkageTypes NewLinkage) {
1590 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
1591 };
1592 thinLTOResolvePrevailingInIndex(Conf, ThinLTO.CombinedIndex, isPrevailing,
1593 recordNewLinkage, GUIDPreservedSymbols);
1594
1595 thinLTOPropagateFunctionAttrs(ThinLTO.CombinedIndex, isPrevailing);
1596
1597 generateParamAccessSummary(ThinLTO.CombinedIndex);
1598
1599 if (llvm::timeTraceProfilerEnabled())
1600 llvm::timeTraceProfilerEnd();
1601
1602 TimeTraceScopeExit.release();
1603
1604 std::unique_ptr<ThinBackendProc> BackendProc =
1605 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1606 AddStream, Cache);
1607
1608 auto &ModuleMap =
1609 ThinLTO.ModulesToCompile ? *ThinLTO.ModulesToCompile : ThinLTO.ModuleMap;
1610
1611 auto ProcessOneModule = [&](int I) -> Error {
1612 auto &Mod = *(ModuleMap.begin() + I);
1613 // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for
1614 // combined module and parallel code generation partitions.
1615 return BackendProc->start(RegularLTO.ParallelCodeGenParallelismLevel + I,
1616 Mod.second, ImportLists[Mod.first],
1617 ExportLists[Mod.first], ResolvedODR[Mod.first],
1618 ThinLTO.ModuleMap);
1619 };
1620
1621 if (BackendProc->getThreadCount() == 1) {
1622 // Process the modules in the order they were provided on the command-line.
1623 // It is important for this codepath to be used for WriteIndexesThinBackend,
1624 // to ensure the emitted LinkedObjectsFile lists ThinLTO objects in the same
1625 // order as the inputs, which otherwise would affect the final link order.
1626 for (int I = 0, E = ModuleMap.size(); I != E; ++I)
1627 if (Error E = ProcessOneModule(I))
1628 return E;
1629 } else {
1630 // When executing in parallel, process largest bitsize modules first to
1631 // improve parallelism, and avoid starving the thread pool near the end.
1632 // This saves about 15 sec on a 36-core machine while link `clang.exe` (out
1633 // of 100 sec).
1634 std::vector<BitcodeModule *> ModulesVec;
1635 ModulesVec.reserve(ModuleMap.size());
1636 for (auto &Mod : ModuleMap)
1637 ModulesVec.push_back(&Mod.second);
1638 for (int I : generateModulesOrdering(ModulesVec))
1639 if (Error E = ProcessOneModule(I))
1640 return E;
1641 }
1642 return BackendProc->wait();
1643 }
1644
setupLLVMOptimizationRemarks(LLVMContext & Context,StringRef RemarksFilename,StringRef RemarksPasses,StringRef RemarksFormat,bool RemarksWithHotness,std::optional<uint64_t> RemarksHotnessThreshold,int Count)1645 Expected<std::unique_ptr<ToolOutputFile>> lto::setupLLVMOptimizationRemarks(
1646 LLVMContext &Context, StringRef RemarksFilename, StringRef RemarksPasses,
1647 StringRef RemarksFormat, bool RemarksWithHotness,
1648 std::optional<uint64_t> RemarksHotnessThreshold, int Count) {
1649 std::string Filename = std::string(RemarksFilename);
1650 // For ThinLTO, file.opt.<format> becomes
1651 // file.opt.<format>.thin.<num>.<format>.
1652 if (!Filename.empty() && Count != -1)
1653 Filename =
1654 (Twine(Filename) + ".thin." + llvm::utostr(Count) + "." + RemarksFormat)
1655 .str();
1656
1657 auto ResultOrErr = llvm::setupLLVMOptimizationRemarks(
1658 Context, Filename, RemarksPasses, RemarksFormat, RemarksWithHotness,
1659 RemarksHotnessThreshold);
1660 if (Error E = ResultOrErr.takeError())
1661 return std::move(E);
1662
1663 if (*ResultOrErr)
1664 (*ResultOrErr)->keep();
1665
1666 return ResultOrErr;
1667 }
1668
1669 Expected<std::unique_ptr<ToolOutputFile>>
setupStatsFile(StringRef StatsFilename)1670 lto::setupStatsFile(StringRef StatsFilename) {
1671 // Setup output file to emit statistics.
1672 if (StatsFilename.empty())
1673 return nullptr;
1674
1675 llvm::EnableStatistics(false);
1676 std::error_code EC;
1677 auto StatsFile =
1678 std::make_unique<ToolOutputFile>(StatsFilename, EC, sys::fs::OF_None);
1679 if (EC)
1680 return errorCodeToError(EC);
1681
1682 StatsFile->keep();
1683 return std::move(StatsFile);
1684 }
1685
1686 // Compute the ordering we will process the inputs: the rough heuristic here
1687 // is to sort them per size so that the largest module get schedule as soon as
1688 // possible. This is purely a compile-time optimization.
generateModulesOrdering(ArrayRef<BitcodeModule * > R)1689 std::vector<int> lto::generateModulesOrdering(ArrayRef<BitcodeModule *> R) {
1690 auto Seq = llvm::seq<int>(0, R.size());
1691 std::vector<int> ModulesOrdering(Seq.begin(), Seq.end());
1692 llvm::sort(ModulesOrdering, [&](int LeftIndex, int RightIndex) {
1693 auto LSize = R[LeftIndex]->getBuffer().size();
1694 auto RSize = R[RightIndex]->getBuffer().size();
1695 return LSize > RSize;
1696 });
1697 return ModulesOrdering;
1698 }
1699