1 //===-LTO.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 functions and classes used to support LTO.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/LTO/LTO.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/TargetLibraryInfo.h"
17 #include "llvm/Analysis/TargetTransformInfo.h"
18 #include "llvm/Bitcode/BitcodeReader.h"
19 #include "llvm/Bitcode/BitcodeWriter.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/Config/llvm-config.h"
22 #include "llvm/IR/AutoUpgrade.h"
23 #include "llvm/IR/DiagnosticPrinter.h"
24 #include "llvm/IR/LegacyPassManager.h"
25 #include "llvm/IR/Mangler.h"
26 #include "llvm/IR/Metadata.h"
27 #include "llvm/LTO/LTOBackend.h"
28 #include "llvm/Linker/IRMover.h"
29 #include "llvm/Object/IRObjectFile.h"
30 #include "llvm/Support/Error.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Support/MemoryBuffer.h"
33 #include "llvm/Support/Path.h"
34 #include "llvm/Support/SHA1.h"
35 #include "llvm/Support/SourceMgr.h"
36 #include "llvm/Support/TargetRegistry.h"
37 #include "llvm/Support/ThreadPool.h"
38 #include "llvm/Support/Threading.h"
39 #include "llvm/Support/VCSRevision.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetMachine.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/Transforms/IPO.h"
44 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
45 #include "llvm/Transforms/Utils/SplitModule.h"
46
47 #include <set>
48
49 using namespace llvm;
50 using namespace lto;
51 using namespace object;
52
53 #define DEBUG_TYPE "lto"
54
55 static cl::opt<bool>
56 DumpThinCGSCCs("dump-thin-cg-sccs", cl::init(false), cl::Hidden,
57 cl::desc("Dump the SCCs in the ThinLTO index's callgraph"));
58
59 // The values are (type identifier, summary) pairs.
60 typedef DenseMap<
61 GlobalValue::GUID,
62 TinyPtrVector<const std::pair<const std::string, TypeIdSummary> *>>
63 TypeIdSummariesByGuidTy;
64
65 // Returns a unique hash for the Module considering the current list of
66 // export/import and other global analysis results.
67 // The hash is produced in \p Key.
computeCacheKey(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 TypeIdSummariesByGuidTy & TypeIdSummariesByGuid,const std::set<GlobalValue::GUID> & CfiFunctionDefs,const std::set<GlobalValue::GUID> & CfiFunctionDecls)68 static void computeCacheKey(
69 SmallString<40> &Key, const Config &Conf, const ModuleSummaryIndex &Index,
70 StringRef ModuleID, const FunctionImporter::ImportMapTy &ImportList,
71 const FunctionImporter::ExportSetTy &ExportList,
72 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
73 const GVSummaryMapTy &DefinedGlobals,
74 const TypeIdSummariesByGuidTy &TypeIdSummariesByGuid,
75 const std::set<GlobalValue::GUID> &CfiFunctionDefs,
76 const std::set<GlobalValue::GUID> &CfiFunctionDecls) {
77 // Compute the unique hash for this entry.
78 // This is based on the current compiler version, the module itself, the
79 // export list, the hash for every single module in the import list, the
80 // list of ResolvedODR for the module, and the list of preserved symbols.
81 SHA1 Hasher;
82
83 // Start with the compiler revision
84 Hasher.update(LLVM_VERSION_STRING);
85 #ifdef LLVM_REVISION
86 Hasher.update(LLVM_REVISION);
87 #endif
88
89 // Include the parts of the LTO configuration that affect code generation.
90 auto AddString = [&](StringRef Str) {
91 Hasher.update(Str);
92 Hasher.update(ArrayRef<uint8_t>{0});
93 };
94 auto AddUnsigned = [&](unsigned I) {
95 uint8_t Data[4];
96 Data[0] = I;
97 Data[1] = I >> 8;
98 Data[2] = I >> 16;
99 Data[3] = I >> 24;
100 Hasher.update(ArrayRef<uint8_t>{Data, 4});
101 };
102 auto AddUint64 = [&](uint64_t I) {
103 uint8_t Data[8];
104 Data[0] = I;
105 Data[1] = I >> 8;
106 Data[2] = I >> 16;
107 Data[3] = I >> 24;
108 Data[4] = I >> 32;
109 Data[5] = I >> 40;
110 Data[6] = I >> 48;
111 Data[7] = I >> 56;
112 Hasher.update(ArrayRef<uint8_t>{Data, 8});
113 };
114 AddString(Conf.CPU);
115 // FIXME: Hash more of Options. For now all clients initialize Options from
116 // command-line flags (which is unsupported in production), but may set
117 // RelaxELFRelocations. The clang driver can also pass FunctionSections,
118 // DataSections and DebuggerTuning via command line flags.
119 AddUnsigned(Conf.Options.RelaxELFRelocations);
120 AddUnsigned(Conf.Options.FunctionSections);
121 AddUnsigned(Conf.Options.DataSections);
122 AddUnsigned((unsigned)Conf.Options.DebuggerTuning);
123 for (auto &A : Conf.MAttrs)
124 AddString(A);
125 if (Conf.RelocModel)
126 AddUnsigned(*Conf.RelocModel);
127 else
128 AddUnsigned(-1);
129 if (Conf.CodeModel)
130 AddUnsigned(*Conf.CodeModel);
131 else
132 AddUnsigned(-1);
133 AddUnsigned(Conf.CGOptLevel);
134 AddUnsigned(Conf.CGFileType);
135 AddUnsigned(Conf.OptLevel);
136 AddUnsigned(Conf.UseNewPM);
137 AddString(Conf.OptPipeline);
138 AddString(Conf.AAPipeline);
139 AddString(Conf.OverrideTriple);
140 AddString(Conf.DefaultTriple);
141 AddString(Conf.DwoDir);
142
143 // Include the hash for the current module
144 auto ModHash = Index.getModuleHash(ModuleID);
145 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
146 for (auto F : ExportList)
147 // The export list can impact the internalization, be conservative here
148 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&F, sizeof(F)));
149
150 // Include the hash for every module we import functions from. The set of
151 // imported symbols for each module may affect code generation and is
152 // sensitive to link order, so include that as well.
153 for (auto &Entry : ImportList) {
154 auto ModHash = Index.getModuleHash(Entry.first());
155 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
156
157 AddUint64(Entry.second.size());
158 for (auto &Fn : Entry.second)
159 AddUint64(Fn);
160 }
161
162 // Include the hash for the resolved ODR.
163 for (auto &Entry : ResolvedODR) {
164 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
165 sizeof(GlobalValue::GUID)));
166 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
167 sizeof(GlobalValue::LinkageTypes)));
168 }
169
170 // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or
171 // defined in this module.
172 std::set<GlobalValue::GUID> UsedCfiDefs;
173 std::set<GlobalValue::GUID> UsedCfiDecls;
174
175 // Typeids used in this module.
176 std::set<GlobalValue::GUID> UsedTypeIds;
177
178 auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) {
179 if (CfiFunctionDefs.count(ValueGUID))
180 UsedCfiDefs.insert(ValueGUID);
181 if (CfiFunctionDecls.count(ValueGUID))
182 UsedCfiDecls.insert(ValueGUID);
183 };
184
185 auto AddUsedThings = [&](GlobalValueSummary *GS) {
186 if (!GS) return;
187 AddUnsigned(GS->isLive());
188 for (const ValueInfo &VI : GS->refs()) {
189 AddUnsigned(VI.isDSOLocal());
190 AddUsedCfiGlobal(VI.getGUID());
191 }
192 if (auto *FS = dyn_cast<FunctionSummary>(GS)) {
193 for (auto &TT : FS->type_tests())
194 UsedTypeIds.insert(TT);
195 for (auto &TT : FS->type_test_assume_vcalls())
196 UsedTypeIds.insert(TT.GUID);
197 for (auto &TT : FS->type_checked_load_vcalls())
198 UsedTypeIds.insert(TT.GUID);
199 for (auto &TT : FS->type_test_assume_const_vcalls())
200 UsedTypeIds.insert(TT.VFunc.GUID);
201 for (auto &TT : FS->type_checked_load_const_vcalls())
202 UsedTypeIds.insert(TT.VFunc.GUID);
203 for (auto &ET : FS->calls()) {
204 AddUnsigned(ET.first.isDSOLocal());
205 AddUsedCfiGlobal(ET.first.getGUID());
206 }
207 }
208 };
209
210 // Include the hash for the linkage type to reflect internalization and weak
211 // resolution, and collect any used type identifier resolutions.
212 for (auto &GS : DefinedGlobals) {
213 GlobalValue::LinkageTypes Linkage = GS.second->linkage();
214 Hasher.update(
215 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage)));
216 AddUsedCfiGlobal(GS.first);
217 AddUsedThings(GS.second);
218 }
219
220 // Imported functions may introduce new uses of type identifier resolutions,
221 // so we need to collect their used resolutions as well.
222 for (auto &ImpM : ImportList)
223 for (auto &ImpF : ImpM.second)
224 AddUsedThings(Index.findSummaryInModule(ImpF, ImpM.first()));
225
226 auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) {
227 AddString(TId);
228
229 AddUnsigned(S.TTRes.TheKind);
230 AddUnsigned(S.TTRes.SizeM1BitWidth);
231
232 AddUint64(S.TTRes.AlignLog2);
233 AddUint64(S.TTRes.SizeM1);
234 AddUint64(S.TTRes.BitMask);
235 AddUint64(S.TTRes.InlineBits);
236
237 AddUint64(S.WPDRes.size());
238 for (auto &WPD : S.WPDRes) {
239 AddUnsigned(WPD.first);
240 AddUnsigned(WPD.second.TheKind);
241 AddString(WPD.second.SingleImplName);
242
243 AddUint64(WPD.second.ResByArg.size());
244 for (auto &ByArg : WPD.second.ResByArg) {
245 AddUint64(ByArg.first.size());
246 for (uint64_t Arg : ByArg.first)
247 AddUint64(Arg);
248 AddUnsigned(ByArg.second.TheKind);
249 AddUint64(ByArg.second.Info);
250 AddUnsigned(ByArg.second.Byte);
251 AddUnsigned(ByArg.second.Bit);
252 }
253 }
254 };
255
256 // Include the hash for all type identifiers used by this module.
257 for (GlobalValue::GUID TId : UsedTypeIds) {
258 auto SummariesI = TypeIdSummariesByGuid.find(TId);
259 if (SummariesI != TypeIdSummariesByGuid.end())
260 for (auto *Summary : SummariesI->second)
261 AddTypeIdSummary(Summary->first, Summary->second);
262 }
263
264 AddUnsigned(UsedCfiDefs.size());
265 for (auto &V : UsedCfiDefs)
266 AddUint64(V);
267
268 AddUnsigned(UsedCfiDecls.size());
269 for (auto &V : UsedCfiDecls)
270 AddUint64(V);
271
272 if (!Conf.SampleProfile.empty()) {
273 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile);
274 if (FileOrErr)
275 Hasher.update(FileOrErr.get()->getBuffer());
276 }
277
278 Key = toHex(Hasher.result());
279 }
280
thinLTOResolveWeakForLinkerGUID(GlobalValueSummaryList & GVSummaryList,GlobalValue::GUID GUID,DenseSet<GlobalValueSummary * > & GlobalInvolvedWithAlias,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing,function_ref<void (StringRef,GlobalValue::GUID,GlobalValue::LinkageTypes)> recordNewLinkage)281 static void thinLTOResolveWeakForLinkerGUID(
282 GlobalValueSummaryList &GVSummaryList, GlobalValue::GUID GUID,
283 DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias,
284 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
285 isPrevailing,
286 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
287 recordNewLinkage) {
288 for (auto &S : GVSummaryList) {
289 GlobalValue::LinkageTypes OriginalLinkage = S->linkage();
290 if (!GlobalValue::isWeakForLinker(OriginalLinkage))
291 continue;
292 // We need to emit only one of these. The prevailing module will keep it,
293 // but turned into a weak, while the others will drop it when possible.
294 // This is both a compile-time optimization and a correctness
295 // transformation. This is necessary for correctness when we have exported
296 // a reference - we need to convert the linkonce to weak to
297 // ensure a copy is kept to satisfy the exported reference.
298 // FIXME: We may want to split the compile time and correctness
299 // aspects into separate routines.
300 if (isPrevailing(GUID, S.get())) {
301 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage))
302 S->setLinkage(GlobalValue::getWeakLinkage(
303 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
304 }
305 // Alias and aliasee can't be turned into available_externally.
306 else if (!isa<AliasSummary>(S.get()) &&
307 !GlobalInvolvedWithAlias.count(S.get()))
308 S->setLinkage(GlobalValue::AvailableExternallyLinkage);
309 if (S->linkage() != OriginalLinkage)
310 recordNewLinkage(S->modulePath(), GUID, S->linkage());
311 }
312 }
313
314 // Resolve Weak and LinkOnce values in the \p Index.
315 //
316 // We'd like to drop these functions if they are no longer referenced in the
317 // current module. However there is a chance that another module is still
318 // referencing them because of the import. We make sure we always emit at least
319 // one copy.
thinLTOResolveWeakForLinkerInIndex(ModuleSummaryIndex & Index,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing,function_ref<void (StringRef,GlobalValue::GUID,GlobalValue::LinkageTypes)> recordNewLinkage)320 void llvm::thinLTOResolveWeakForLinkerInIndex(
321 ModuleSummaryIndex &Index,
322 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
323 isPrevailing,
324 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
325 recordNewLinkage) {
326 // We won't optimize the globals that are referenced by an alias for now
327 // Ideally we should turn the alias into a global and duplicate the definition
328 // when needed.
329 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias;
330 for (auto &I : Index)
331 for (auto &S : I.second.SummaryList)
332 if (auto AS = dyn_cast<AliasSummary>(S.get()))
333 GlobalInvolvedWithAlias.insert(&AS->getAliasee());
334
335 for (auto &I : Index)
336 thinLTOResolveWeakForLinkerGUID(I.second.SummaryList, I.first,
337 GlobalInvolvedWithAlias, isPrevailing,
338 recordNewLinkage);
339 }
340
thinLTOInternalizeAndPromoteGUID(GlobalValueSummaryList & GVSummaryList,GlobalValue::GUID GUID,function_ref<bool (StringRef,GlobalValue::GUID)> isExported)341 static void thinLTOInternalizeAndPromoteGUID(
342 GlobalValueSummaryList &GVSummaryList, GlobalValue::GUID GUID,
343 function_ref<bool(StringRef, GlobalValue::GUID)> isExported) {
344 for (auto &S : GVSummaryList) {
345 if (isExported(S->modulePath(), GUID)) {
346 if (GlobalValue::isLocalLinkage(S->linkage()))
347 S->setLinkage(GlobalValue::ExternalLinkage);
348 } else if (!GlobalValue::isLocalLinkage(S->linkage()))
349 S->setLinkage(GlobalValue::InternalLinkage);
350 }
351 }
352
353 // Update the linkages in the given \p Index to mark exported values
354 // as external and non-exported values as internal.
thinLTOInternalizeAndPromoteInIndex(ModuleSummaryIndex & Index,function_ref<bool (StringRef,GlobalValue::GUID)> isExported)355 void llvm::thinLTOInternalizeAndPromoteInIndex(
356 ModuleSummaryIndex &Index,
357 function_ref<bool(StringRef, GlobalValue::GUID)> isExported) {
358 for (auto &I : Index)
359 thinLTOInternalizeAndPromoteGUID(I.second.SummaryList, I.first, isExported);
360 }
361
362 // Requires a destructor for std::vector<InputModule>.
363 InputFile::~InputFile() = default;
364
create(MemoryBufferRef Object)365 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) {
366 std::unique_ptr<InputFile> File(new InputFile);
367
368 Expected<IRSymtabFile> FOrErr = readIRSymtab(Object);
369 if (!FOrErr)
370 return FOrErr.takeError();
371
372 File->TargetTriple = FOrErr->TheReader.getTargetTriple();
373 File->SourceFileName = FOrErr->TheReader.getSourceFileName();
374 File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts();
375 File->ComdatTable = FOrErr->TheReader.getComdatTable();
376
377 for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) {
378 size_t Begin = File->Symbols.size();
379 for (const irsymtab::Reader::SymbolRef &Sym :
380 FOrErr->TheReader.module_symbols(I))
381 // Skip symbols that are irrelevant to LTO. Note that this condition needs
382 // to match the one in Skip() in LTO::addRegularLTO().
383 if (Sym.isGlobal() && !Sym.isFormatSpecific())
384 File->Symbols.push_back(Sym);
385 File->ModuleSymIndices.push_back({Begin, File->Symbols.size()});
386 }
387
388 File->Mods = FOrErr->Mods;
389 File->Strtab = std::move(FOrErr->Strtab);
390 return std::move(File);
391 }
392
getName() const393 StringRef InputFile::getName() const {
394 return Mods[0].getModuleIdentifier();
395 }
396
RegularLTOState(unsigned ParallelCodeGenParallelismLevel,Config & Conf)397 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel,
398 Config &Conf)
399 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel),
400 Ctx(Conf), CombinedModule(llvm::make_unique<Module>("ld-temp.o", Ctx)),
401 Mover(llvm::make_unique<IRMover>(*CombinedModule)) {}
402
ThinLTOState(ThinBackend Backend)403 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend)
404 : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) {
405 if (!Backend)
406 this->Backend =
407 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency());
408 }
409
LTO(Config Conf,ThinBackend Backend,unsigned ParallelCodeGenParallelismLevel)410 LTO::LTO(Config Conf, ThinBackend Backend,
411 unsigned ParallelCodeGenParallelismLevel)
412 : Conf(std::move(Conf)),
413 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf),
414 ThinLTO(std::move(Backend)) {}
415
416 // Requires a destructor for MapVector<BitcodeModule>.
417 LTO::~LTO() = default;
418
419 // Add the symbols in the given module to the GlobalResolutions map, and resolve
420 // their partitions.
addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,ArrayRef<SymbolResolution> Res,unsigned Partition,bool InSummary)421 void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
422 ArrayRef<SymbolResolution> Res,
423 unsigned Partition, bool InSummary) {
424 auto *ResI = Res.begin();
425 auto *ResE = Res.end();
426 (void)ResE;
427 for (const InputFile::Symbol &Sym : Syms) {
428 assert(ResI != ResE);
429 SymbolResolution Res = *ResI++;
430
431 StringRef Name = Sym.getName();
432 Triple TT(RegularLTO.CombinedModule->getTargetTriple());
433 // Strip the __imp_ prefix from COFF dllimport symbols (similar to the
434 // way they are handled by lld), otherwise we can end up with two
435 // global resolutions (one with and one for a copy of the symbol without).
436 if (TT.isOSBinFormatCOFF() && Name.startswith("__imp_"))
437 Name = Name.substr(strlen("__imp_"));
438 auto &GlobalRes = GlobalResolutions[Name];
439 GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr();
440 if (Res.Prevailing) {
441 assert(!GlobalRes.Prevailing &&
442 "Multiple prevailing defs are not allowed");
443 GlobalRes.Prevailing = true;
444 GlobalRes.IRName = Sym.getIRName();
445 } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) {
446 // Sometimes it can be two copies of symbol in a module and prevailing
447 // symbol can have no IR name. That might happen if symbol is defined in
448 // module level inline asm block. In case we have multiple modules with
449 // the same symbol we want to use IR name of the prevailing symbol.
450 // Otherwise, if we haven't seen a prevailing symbol, set the name so that
451 // we can later use it to check if there is any prevailing copy in IR.
452 GlobalRes.IRName = Sym.getIRName();
453 }
454
455 // Set the partition to external if we know it is re-defined by the linker
456 // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a
457 // regular object, is referenced from llvm.compiler_used, or was already
458 // recorded as being referenced from a different partition.
459 if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() ||
460 (GlobalRes.Partition != GlobalResolution::Unknown &&
461 GlobalRes.Partition != Partition)) {
462 GlobalRes.Partition = GlobalResolution::External;
463 } else
464 // First recorded reference, save the current partition.
465 GlobalRes.Partition = Partition;
466
467 // Flag as visible outside of summary if visible from a regular object or
468 // from a module that does not have a summary.
469 GlobalRes.VisibleOutsideSummary |=
470 (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary);
471 }
472 }
473
writeToResolutionFile(raw_ostream & OS,InputFile * Input,ArrayRef<SymbolResolution> Res)474 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input,
475 ArrayRef<SymbolResolution> Res) {
476 StringRef Path = Input->getName();
477 OS << Path << '\n';
478 auto ResI = Res.begin();
479 for (const InputFile::Symbol &Sym : Input->symbols()) {
480 assert(ResI != Res.end());
481 SymbolResolution Res = *ResI++;
482
483 OS << "-r=" << Path << ',' << Sym.getName() << ',';
484 if (Res.Prevailing)
485 OS << 'p';
486 if (Res.FinalDefinitionInLinkageUnit)
487 OS << 'l';
488 if (Res.VisibleToRegularObj)
489 OS << 'x';
490 if (Res.LinkerRedefined)
491 OS << 'r';
492 OS << '\n';
493 }
494 OS.flush();
495 assert(ResI == Res.end());
496 }
497
add(std::unique_ptr<InputFile> Input,ArrayRef<SymbolResolution> Res)498 Error LTO::add(std::unique_ptr<InputFile> Input,
499 ArrayRef<SymbolResolution> Res) {
500 assert(!CalledGetMaxTasks);
501
502 if (Conf.ResolutionFile)
503 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res);
504
505 if (RegularLTO.CombinedModule->getTargetTriple().empty())
506 RegularLTO.CombinedModule->setTargetTriple(Input->getTargetTriple());
507
508 const SymbolResolution *ResI = Res.begin();
509 for (unsigned I = 0; I != Input->Mods.size(); ++I)
510 if (Error Err = addModule(*Input, I, ResI, Res.end()))
511 return Err;
512
513 assert(ResI == Res.end());
514 return Error::success();
515 }
516
addModule(InputFile & Input,unsigned ModI,const SymbolResolution * & ResI,const SymbolResolution * ResE)517 Error LTO::addModule(InputFile &Input, unsigned ModI,
518 const SymbolResolution *&ResI,
519 const SymbolResolution *ResE) {
520 Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo();
521 if (!LTOInfo)
522 return LTOInfo.takeError();
523
524 BitcodeModule BM = Input.Mods[ModI];
525 auto ModSyms = Input.module_symbols(ModI);
526 addModuleToGlobalRes(ModSyms, {ResI, ResE},
527 LTOInfo->IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0,
528 LTOInfo->HasSummary);
529
530 if (LTOInfo->IsThinLTO)
531 return addThinLTO(BM, ModSyms, ResI, ResE);
532
533 Expected<RegularLTOState::AddedModule> ModOrErr =
534 addRegularLTO(BM, ModSyms, ResI, ResE);
535 if (!ModOrErr)
536 return ModOrErr.takeError();
537
538 if (!LTOInfo->HasSummary)
539 return linkRegularLTO(std::move(*ModOrErr), /*LivenessFromIndex=*/false);
540
541 // Regular LTO module summaries are added to a dummy module that represents
542 // the combined regular LTO module.
543 if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, "", -1ull))
544 return Err;
545 RegularLTO.ModsWithSummaries.push_back(std::move(*ModOrErr));
546 return Error::success();
547 }
548
549 // Checks whether the given global value is in a non-prevailing comdat
550 // (comdat containing values the linker indicated were not prevailing,
551 // which we then dropped to available_externally), and if so, removes
552 // it from the comdat. This is called for all global values to ensure the
553 // comdat is empty rather than leaving an incomplete comdat. It is needed for
554 // regular LTO modules, in case we are in a mixed-LTO mode (both regular
555 // and thin LTO modules) compilation. Since the regular LTO module will be
556 // linked first in the final native link, we want to make sure the linker
557 // doesn't select any of these incomplete comdats that would be left
558 // in the regular LTO module without this cleanup.
559 static void
handleNonPrevailingComdat(GlobalValue & GV,std::set<const Comdat * > & NonPrevailingComdats)560 handleNonPrevailingComdat(GlobalValue &GV,
561 std::set<const Comdat *> &NonPrevailingComdats) {
562 Comdat *C = GV.getComdat();
563 if (!C)
564 return;
565
566 if (!NonPrevailingComdats.count(C))
567 return;
568
569 // Additionally need to drop externally visible global values from the comdat
570 // to available_externally, so that there aren't multiply defined linker
571 // errors.
572 if (!GV.hasLocalLinkage())
573 GV.setLinkage(GlobalValue::AvailableExternallyLinkage);
574
575 if (auto GO = dyn_cast<GlobalObject>(&GV))
576 GO->setComdat(nullptr);
577 }
578
579 // Add a regular LTO object to the link.
580 // The resulting module needs to be linked into the combined LTO module with
581 // linkRegularLTO.
582 Expected<LTO::RegularLTOState::AddedModule>
addRegularLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)583 LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
584 const SymbolResolution *&ResI,
585 const SymbolResolution *ResE) {
586 RegularLTOState::AddedModule Mod;
587 Expected<std::unique_ptr<Module>> MOrErr =
588 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true,
589 /*IsImporting*/ false);
590 if (!MOrErr)
591 return MOrErr.takeError();
592 Module &M = **MOrErr;
593 Mod.M = std::move(*MOrErr);
594
595 if (Error Err = M.materializeMetadata())
596 return std::move(Err);
597 UpgradeDebugInfo(M);
598
599 ModuleSymbolTable SymTab;
600 SymTab.addModule(&M);
601
602 for (GlobalVariable &GV : M.globals())
603 if (GV.hasAppendingLinkage())
604 Mod.Keep.push_back(&GV);
605
606 DenseSet<GlobalObject *> AliasedGlobals;
607 for (auto &GA : M.aliases())
608 if (GlobalObject *GO = GA.getBaseObject())
609 AliasedGlobals.insert(GO);
610
611 // In this function we need IR GlobalValues matching the symbols in Syms
612 // (which is not backed by a module), so we need to enumerate them in the same
613 // order. The symbol enumeration order of a ModuleSymbolTable intentionally
614 // matches the order of an irsymtab, but when we read the irsymtab in
615 // InputFile::create we omit some symbols that are irrelevant to LTO. The
616 // Skip() function skips the same symbols from the module as InputFile does
617 // from the symbol table.
618 auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end();
619 auto Skip = [&]() {
620 while (MsymI != MsymE) {
621 auto Flags = SymTab.getSymbolFlags(*MsymI);
622 if ((Flags & object::BasicSymbolRef::SF_Global) &&
623 !(Flags & object::BasicSymbolRef::SF_FormatSpecific))
624 return;
625 ++MsymI;
626 }
627 };
628 Skip();
629
630 std::set<const Comdat *> NonPrevailingComdats;
631 for (const InputFile::Symbol &Sym : Syms) {
632 assert(ResI != ResE);
633 SymbolResolution Res = *ResI++;
634
635 assert(MsymI != MsymE);
636 ModuleSymbolTable::Symbol Msym = *MsymI++;
637 Skip();
638
639 if (GlobalValue *GV = Msym.dyn_cast<GlobalValue *>()) {
640 if (Res.Prevailing) {
641 if (Sym.isUndefined())
642 continue;
643 Mod.Keep.push_back(GV);
644 // For symbols re-defined with linker -wrap and -defsym options,
645 // set the linkage to weak to inhibit IPO. The linkage will be
646 // restored by the linker.
647 if (Res.LinkerRedefined)
648 GV->setLinkage(GlobalValue::WeakAnyLinkage);
649
650 GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage();
651 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage))
652 GV->setLinkage(GlobalValue::getWeakLinkage(
653 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
654 } else if (isa<GlobalObject>(GV) &&
655 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() ||
656 GV->hasAvailableExternallyLinkage()) &&
657 !AliasedGlobals.count(cast<GlobalObject>(GV))) {
658 // Any of the above three types of linkage indicates that the
659 // chosen prevailing symbol will have the same semantics as this copy of
660 // the symbol, so we may be able to link it with available_externally
661 // linkage. We will decide later whether to do that when we link this
662 // module (in linkRegularLTO), based on whether it is undefined.
663 Mod.Keep.push_back(GV);
664 GV->setLinkage(GlobalValue::AvailableExternallyLinkage);
665 if (GV->hasComdat())
666 NonPrevailingComdats.insert(GV->getComdat());
667 cast<GlobalObject>(GV)->setComdat(nullptr);
668 }
669
670 // Set the 'local' flag based on the linker resolution for this symbol.
671 if (Res.FinalDefinitionInLinkageUnit)
672 GV->setDSOLocal(true);
673 }
674 // Common resolution: collect the maximum size/alignment over all commons.
675 // We also record if we see an instance of a common as prevailing, so that
676 // if none is prevailing we can ignore it later.
677 if (Sym.isCommon()) {
678 // FIXME: We should figure out what to do about commons defined by asm.
679 // For now they aren't reported correctly by ModuleSymbolTable.
680 auto &CommonRes = RegularLTO.Commons[Sym.getIRName()];
681 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize());
682 CommonRes.Align = std::max(CommonRes.Align, Sym.getCommonAlignment());
683 CommonRes.Prevailing |= Res.Prevailing;
684 }
685
686 }
687 if (!M.getComdatSymbolTable().empty())
688 for (GlobalValue &GV : M.global_values())
689 handleNonPrevailingComdat(GV, NonPrevailingComdats);
690 assert(MsymI == MsymE);
691 return std::move(Mod);
692 }
693
linkRegularLTO(RegularLTOState::AddedModule Mod,bool LivenessFromIndex)694 Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod,
695 bool LivenessFromIndex) {
696 std::vector<GlobalValue *> Keep;
697 for (GlobalValue *GV : Mod.Keep) {
698 if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GV->getGUID()))
699 continue;
700
701 if (!GV->hasAvailableExternallyLinkage()) {
702 Keep.push_back(GV);
703 continue;
704 }
705
706 // Only link available_externally definitions if we don't already have a
707 // definition.
708 GlobalValue *CombinedGV =
709 RegularLTO.CombinedModule->getNamedValue(GV->getName());
710 if (CombinedGV && !CombinedGV->isDeclaration())
711 continue;
712
713 Keep.push_back(GV);
714 }
715
716 return RegularLTO.Mover->move(std::move(Mod.M), Keep,
717 [](GlobalValue &, IRMover::ValueAdder) {},
718 /* IsPerformingImport */ false);
719 }
720
721 // Add a ThinLTO module to the link.
addThinLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)722 Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
723 const SymbolResolution *&ResI,
724 const SymbolResolution *ResE) {
725 if (Error Err =
726 BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(),
727 ThinLTO.ModuleMap.size()))
728 return Err;
729
730 for (const InputFile::Symbol &Sym : Syms) {
731 assert(ResI != ResE);
732 SymbolResolution Res = *ResI++;
733
734 if (!Sym.getIRName().empty()) {
735 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier(
736 Sym.getIRName(), GlobalValue::ExternalLinkage, ""));
737 if (Res.Prevailing) {
738 ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier();
739
740 // For linker redefined symbols (via --wrap or --defsym) we want to
741 // switch the linkage to `weak` to prevent IPOs from happening.
742 // Find the summary in the module for this very GV and record the new
743 // linkage so that we can switch it when we import the GV.
744 if (Res.LinkerRedefined)
745 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
746 GUID, BM.getModuleIdentifier()))
747 S->setLinkage(GlobalValue::WeakAnyLinkage);
748 }
749
750 // If the linker resolved the symbol to a local definition then mark it
751 // as local in the summary for the module we are adding.
752 if (Res.FinalDefinitionInLinkageUnit) {
753 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
754 GUID, BM.getModuleIdentifier())) {
755 S->setDSOLocal(true);
756 }
757 }
758 }
759 }
760
761 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second)
762 return make_error<StringError>(
763 "Expected at most one ThinLTO module per bitcode file",
764 inconvertibleErrorCode());
765
766 return Error::success();
767 }
768
getMaxTasks() const769 unsigned LTO::getMaxTasks() const {
770 CalledGetMaxTasks = true;
771 return RegularLTO.ParallelCodeGenParallelismLevel + ThinLTO.ModuleMap.size();
772 }
773
run(AddStreamFn AddStream,NativeObjectCache Cache)774 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) {
775 // Compute "dead" symbols, we don't want to import/export these!
776 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols;
777 DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions;
778 for (auto &Res : GlobalResolutions) {
779 // Normally resolution have IR name of symbol. We can do nothing here
780 // otherwise. See comments in GlobalResolution struct for more details.
781 if (Res.second.IRName.empty())
782 continue;
783
784 GlobalValue::GUID GUID = GlobalValue::getGUID(
785 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
786
787 if (Res.second.VisibleOutsideSummary && Res.second.Prevailing)
788 GUIDPreservedSymbols.insert(GlobalValue::getGUID(
789 GlobalValue::dropLLVMManglingEscape(Res.second.IRName)));
790
791 GUIDPrevailingResolutions[GUID] =
792 Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No;
793 }
794
795 auto isPrevailing = [&](GlobalValue::GUID G) {
796 auto It = GUIDPrevailingResolutions.find(G);
797 if (It == GUIDPrevailingResolutions.end())
798 return PrevailingType::Unknown;
799 return It->second;
800 };
801 computeDeadSymbols(ThinLTO.CombinedIndex, GUIDPreservedSymbols, isPrevailing);
802
803 // Setup output file to emit statistics.
804 std::unique_ptr<ToolOutputFile> StatsFile = nullptr;
805 if (!Conf.StatsFile.empty()) {
806 EnableStatistics(false);
807 std::error_code EC;
808 StatsFile =
809 llvm::make_unique<ToolOutputFile>(Conf.StatsFile, EC, sys::fs::F_None);
810 if (EC)
811 return errorCodeToError(EC);
812 StatsFile->keep();
813 }
814
815 Error Result = runRegularLTO(AddStream);
816 if (!Result)
817 Result = runThinLTO(AddStream, Cache);
818
819 if (StatsFile)
820 PrintStatisticsJSON(StatsFile->os());
821
822 return Result;
823 }
824
runRegularLTO(AddStreamFn AddStream)825 Error LTO::runRegularLTO(AddStreamFn AddStream) {
826 for (auto &M : RegularLTO.ModsWithSummaries)
827 if (Error Err = linkRegularLTO(std::move(M),
828 /*LivenessFromIndex=*/true))
829 return Err;
830
831 // Make sure commons have the right size/alignment: we kept the largest from
832 // all the prevailing when adding the inputs, and we apply it here.
833 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout();
834 for (auto &I : RegularLTO.Commons) {
835 if (!I.second.Prevailing)
836 // Don't do anything if no instance of this common was prevailing.
837 continue;
838 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first);
839 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) {
840 // Don't create a new global if the type is already correct, just make
841 // sure the alignment is correct.
842 OldGV->setAlignment(I.second.Align);
843 continue;
844 }
845 ArrayType *Ty =
846 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size);
847 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false,
848 GlobalValue::CommonLinkage,
849 ConstantAggregateZero::get(Ty), "");
850 GV->setAlignment(I.second.Align);
851 if (OldGV) {
852 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType()));
853 GV->takeName(OldGV);
854 OldGV->eraseFromParent();
855 } else {
856 GV->setName(I.first);
857 }
858 }
859
860 if (Conf.PreOptModuleHook &&
861 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule))
862 return Error::success();
863
864 if (!Conf.CodeGenOnly) {
865 for (const auto &R : GlobalResolutions) {
866 if (!R.second.isPrevailingIRSymbol())
867 continue;
868 if (R.second.Partition != 0 &&
869 R.second.Partition != GlobalResolution::External)
870 continue;
871
872 GlobalValue *GV =
873 RegularLTO.CombinedModule->getNamedValue(R.second.IRName);
874 // Ignore symbols defined in other partitions.
875 // Also skip declarations, which are not allowed to have internal linkage.
876 if (!GV || GV->hasLocalLinkage() || GV->isDeclaration())
877 continue;
878 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global
879 : GlobalValue::UnnamedAddr::None);
880 if (R.second.Partition == 0)
881 GV->setLinkage(GlobalValue::InternalLinkage);
882 }
883
884 if (Conf.PostInternalizeModuleHook &&
885 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule))
886 return Error::success();
887 }
888 return backend(Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel,
889 std::move(RegularLTO.CombinedModule), ThinLTO.CombinedIndex);
890 }
891
892 /// This class defines the interface to the ThinLTO backend.
893 class lto::ThinBackendProc {
894 protected:
895 Config &Conf;
896 ModuleSummaryIndex &CombinedIndex;
897 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries;
898
899 public:
ThinBackendProc(Config & Conf,ModuleSummaryIndex & CombinedIndex,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries)900 ThinBackendProc(Config &Conf, ModuleSummaryIndex &CombinedIndex,
901 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries)
902 : Conf(Conf), CombinedIndex(CombinedIndex),
903 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries) {}
904
~ThinBackendProc()905 virtual ~ThinBackendProc() {}
906 virtual Error start(
907 unsigned Task, BitcodeModule BM,
908 const FunctionImporter::ImportMapTy &ImportList,
909 const FunctionImporter::ExportSetTy &ExportList,
910 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
911 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0;
912 virtual Error wait() = 0;
913 };
914
915 namespace {
916 class InProcessThinBackend : public ThinBackendProc {
917 ThreadPool BackendThreadPool;
918 AddStreamFn AddStream;
919 NativeObjectCache Cache;
920 TypeIdSummariesByGuidTy TypeIdSummariesByGuid;
921 std::set<GlobalValue::GUID> CfiFunctionDefs;
922 std::set<GlobalValue::GUID> CfiFunctionDecls;
923
924 Optional<Error> Err;
925 std::mutex ErrMu;
926
927 public:
InProcessThinBackend(Config & Conf,ModuleSummaryIndex & CombinedIndex,unsigned ThinLTOParallelismLevel,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,AddStreamFn AddStream,NativeObjectCache Cache)928 InProcessThinBackend(
929 Config &Conf, ModuleSummaryIndex &CombinedIndex,
930 unsigned ThinLTOParallelismLevel,
931 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
932 AddStreamFn AddStream, NativeObjectCache Cache)
933 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries),
934 BackendThreadPool(ThinLTOParallelismLevel),
935 AddStream(std::move(AddStream)), Cache(std::move(Cache)) {
936 // Create a mapping from type identifier GUIDs to type identifier summaries.
937 // This allows backends to use the type identifier GUIDs stored in the
938 // function summaries to determine which type identifier summaries affect
939 // each function without needing to compute GUIDs in each backend.
940 for (auto &TId : CombinedIndex.typeIds())
941 TypeIdSummariesByGuid[GlobalValue::getGUID(TId.first)].push_back(&TId);
942 for (auto &Name : CombinedIndex.cfiFunctionDefs())
943 CfiFunctionDefs.insert(
944 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
945 for (auto &Name : CombinedIndex.cfiFunctionDecls())
946 CfiFunctionDecls.insert(
947 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
948 }
949
runThinLTOBackendThread(AddStreamFn AddStream,NativeObjectCache 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,const TypeIdSummariesByGuidTy & TypeIdSummariesByGuid)950 Error runThinLTOBackendThread(
951 AddStreamFn AddStream, NativeObjectCache Cache, unsigned Task,
952 BitcodeModule BM, ModuleSummaryIndex &CombinedIndex,
953 const FunctionImporter::ImportMapTy &ImportList,
954 const FunctionImporter::ExportSetTy &ExportList,
955 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
956 const GVSummaryMapTy &DefinedGlobals,
957 MapVector<StringRef, BitcodeModule> &ModuleMap,
958 const TypeIdSummariesByGuidTy &TypeIdSummariesByGuid) {
959 auto RunThinBackend = [&](AddStreamFn AddStream) {
960 LTOLLVMContext BackendContext(Conf);
961 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext);
962 if (!MOrErr)
963 return MOrErr.takeError();
964
965 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex,
966 ImportList, DefinedGlobals, ModuleMap);
967 };
968
969 auto ModuleID = BM.getModuleIdentifier();
970
971 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) ||
972 all_of(CombinedIndex.getModuleHash(ModuleID),
973 [](uint32_t V) { return V == 0; }))
974 // Cache disabled or no entry for this module in the combined index or
975 // no module hash.
976 return RunThinBackend(AddStream);
977
978 SmallString<40> Key;
979 // The module may be cached, this helps handling it.
980 computeCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList, ExportList,
981 ResolvedODR, DefinedGlobals, TypeIdSummariesByGuid,
982 CfiFunctionDefs, CfiFunctionDecls);
983 if (AddStreamFn CacheAddStream = Cache(Task, Key))
984 return RunThinBackend(CacheAddStream);
985
986 return Error::success();
987 }
988
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)989 Error start(
990 unsigned Task, BitcodeModule BM,
991 const FunctionImporter::ImportMapTy &ImportList,
992 const FunctionImporter::ExportSetTy &ExportList,
993 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
994 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
995 StringRef ModulePath = BM.getModuleIdentifier();
996 assert(ModuleToDefinedGVSummaries.count(ModulePath));
997 const GVSummaryMapTy &DefinedGlobals =
998 ModuleToDefinedGVSummaries.find(ModulePath)->second;
999 BackendThreadPool.async(
1000 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex,
1001 const FunctionImporter::ImportMapTy &ImportList,
1002 const FunctionImporter::ExportSetTy &ExportList,
1003 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>
1004 &ResolvedODR,
1005 const GVSummaryMapTy &DefinedGlobals,
1006 MapVector<StringRef, BitcodeModule> &ModuleMap,
1007 const TypeIdSummariesByGuidTy &TypeIdSummariesByGuid) {
1008 Error E = runThinLTOBackendThread(
1009 AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList,
1010 ResolvedODR, DefinedGlobals, ModuleMap, TypeIdSummariesByGuid);
1011 if (E) {
1012 std::unique_lock<std::mutex> L(ErrMu);
1013 if (Err)
1014 Err = joinErrors(std::move(*Err), std::move(E));
1015 else
1016 Err = std::move(E);
1017 }
1018 },
1019 BM, std::ref(CombinedIndex), std::ref(ImportList), std::ref(ExportList),
1020 std::ref(ResolvedODR), std::ref(DefinedGlobals), std::ref(ModuleMap),
1021 std::ref(TypeIdSummariesByGuid));
1022 return Error::success();
1023 }
1024
wait()1025 Error wait() override {
1026 BackendThreadPool.wait();
1027 if (Err)
1028 return std::move(*Err);
1029 else
1030 return Error::success();
1031 }
1032 };
1033 } // end anonymous namespace
1034
createInProcessThinBackend(unsigned ParallelismLevel)1035 ThinBackend lto::createInProcessThinBackend(unsigned ParallelismLevel) {
1036 return [=](Config &Conf, ModuleSummaryIndex &CombinedIndex,
1037 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1038 AddStreamFn AddStream, NativeObjectCache Cache) {
1039 return llvm::make_unique<InProcessThinBackend>(
1040 Conf, CombinedIndex, ParallelismLevel, ModuleToDefinedGVSummaries,
1041 AddStream, Cache);
1042 };
1043 }
1044
1045 // Given the original \p Path to an output file, replace any path
1046 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the
1047 // resulting directory if it does not yet exist.
getThinLTOOutputFile(const std::string & Path,const std::string & OldPrefix,const std::string & NewPrefix)1048 std::string lto::getThinLTOOutputFile(const std::string &Path,
1049 const std::string &OldPrefix,
1050 const std::string &NewPrefix) {
1051 if (OldPrefix.empty() && NewPrefix.empty())
1052 return Path;
1053 SmallString<128> NewPath(Path);
1054 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix);
1055 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str());
1056 if (!ParentPath.empty()) {
1057 // Make sure the new directory exists, creating it if necessary.
1058 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath))
1059 llvm::errs() << "warning: could not create directory '" << ParentPath
1060 << "': " << EC.message() << '\n';
1061 }
1062 return NewPath.str();
1063 }
1064
1065 namespace {
1066 class WriteIndexesThinBackend : public ThinBackendProc {
1067 std::string OldPrefix, NewPrefix;
1068 bool ShouldEmitImportsFiles;
1069 raw_fd_ostream *LinkedObjectsFile;
1070 lto::IndexWriteCallback OnWrite;
1071
1072 public:
WriteIndexesThinBackend(Config & Conf,ModuleSummaryIndex & CombinedIndex,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,std::string OldPrefix,std::string NewPrefix,bool ShouldEmitImportsFiles,raw_fd_ostream * LinkedObjectsFile,lto::IndexWriteCallback OnWrite)1073 WriteIndexesThinBackend(
1074 Config &Conf, ModuleSummaryIndex &CombinedIndex,
1075 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1076 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1077 raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite)
1078 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries),
1079 OldPrefix(OldPrefix), NewPrefix(NewPrefix),
1080 ShouldEmitImportsFiles(ShouldEmitImportsFiles),
1081 LinkedObjectsFile(LinkedObjectsFile), OnWrite(OnWrite) {}
1082
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)1083 Error start(
1084 unsigned Task, BitcodeModule BM,
1085 const FunctionImporter::ImportMapTy &ImportList,
1086 const FunctionImporter::ExportSetTy &ExportList,
1087 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1088 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
1089 StringRef ModulePath = BM.getModuleIdentifier();
1090 std::string NewModulePath =
1091 getThinLTOOutputFile(ModulePath, OldPrefix, NewPrefix);
1092
1093 if (LinkedObjectsFile)
1094 *LinkedObjectsFile << NewModulePath << '\n';
1095
1096 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
1097 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
1098 ImportList, ModuleToSummariesForIndex);
1099
1100 std::error_code EC;
1101 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC,
1102 sys::fs::OpenFlags::F_None);
1103 if (EC)
1104 return errorCodeToError(EC);
1105 WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex);
1106
1107 if (ShouldEmitImportsFiles) {
1108 EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports",
1109 ModuleToSummariesForIndex);
1110 if (EC)
1111 return errorCodeToError(EC);
1112 }
1113
1114 if (OnWrite)
1115 OnWrite(ModulePath);
1116 return Error::success();
1117 }
1118
wait()1119 Error wait() override { return Error::success(); }
1120 };
1121 } // end anonymous namespace
1122
createWriteIndexesThinBackend(std::string OldPrefix,std::string NewPrefix,bool ShouldEmitImportsFiles,raw_fd_ostream * LinkedObjectsFile,IndexWriteCallback OnWrite)1123 ThinBackend lto::createWriteIndexesThinBackend(
1124 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1125 raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) {
1126 return [=](Config &Conf, ModuleSummaryIndex &CombinedIndex,
1127 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1128 AddStreamFn AddStream, NativeObjectCache Cache) {
1129 return llvm::make_unique<WriteIndexesThinBackend>(
1130 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix,
1131 ShouldEmitImportsFiles, LinkedObjectsFile, OnWrite);
1132 };
1133 }
1134
runThinLTO(AddStreamFn AddStream,NativeObjectCache Cache)1135 Error LTO::runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache) {
1136 if (ThinLTO.ModuleMap.empty())
1137 return Error::success();
1138
1139 if (Conf.CombinedIndexHook && !Conf.CombinedIndexHook(ThinLTO.CombinedIndex))
1140 return Error::success();
1141
1142 // Collect for each module the list of function it defines (GUID ->
1143 // Summary).
1144 StringMap<GVSummaryMapTy>
1145 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size());
1146 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule(
1147 ModuleToDefinedGVSummaries);
1148 // Create entries for any modules that didn't have any GV summaries
1149 // (either they didn't have any GVs to start with, or we suppressed
1150 // generation of the summaries because they e.g. had inline assembly
1151 // uses that couldn't be promoted/renamed on export). This is so
1152 // InProcessThinBackend::start can still launch a backend thread, which
1153 // is passed the map of summaries for the module, without any special
1154 // handling for this case.
1155 for (auto &Mod : ThinLTO.ModuleMap)
1156 if (!ModuleToDefinedGVSummaries.count(Mod.first))
1157 ModuleToDefinedGVSummaries.try_emplace(Mod.first);
1158
1159 StringMap<FunctionImporter::ImportMapTy> ImportLists(
1160 ThinLTO.ModuleMap.size());
1161 StringMap<FunctionImporter::ExportSetTy> ExportLists(
1162 ThinLTO.ModuleMap.size());
1163 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
1164
1165 if (DumpThinCGSCCs)
1166 ThinLTO.CombinedIndex.dumpSCCs(outs());
1167
1168 if (Conf.OptLevel > 0)
1169 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1170 ImportLists, ExportLists);
1171
1172 // Figure out which symbols need to be internalized. This also needs to happen
1173 // at -O0 because summary-based DCE is implemented using internalization, and
1174 // we must apply DCE consistently with the full LTO module in order to avoid
1175 // undefined references during the final link.
1176 std::set<GlobalValue::GUID> ExportedGUIDs;
1177 for (auto &Res : GlobalResolutions) {
1178 // If the symbol does not have external references or it is not prevailing,
1179 // then not need to mark it as exported from a ThinLTO partition.
1180 if (Res.second.Partition != GlobalResolution::External ||
1181 !Res.second.isPrevailingIRSymbol())
1182 continue;
1183 auto GUID = GlobalValue::getGUID(
1184 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
1185 // Mark exported unless index-based analysis determined it to be dead.
1186 if (ThinLTO.CombinedIndex.isGUIDLive(GUID))
1187 ExportedGUIDs.insert(GUID);
1188 }
1189
1190 // Any functions referenced by the jump table in the regular LTO object must
1191 // be exported.
1192 for (auto &Def : ThinLTO.CombinedIndex.cfiFunctionDefs())
1193 ExportedGUIDs.insert(
1194 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Def)));
1195
1196 auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
1197 const auto &ExportList = ExportLists.find(ModuleIdentifier);
1198 return (ExportList != ExportLists.end() &&
1199 ExportList->second.count(GUID)) ||
1200 ExportedGUIDs.count(GUID);
1201 };
1202 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported);
1203
1204 auto isPrevailing = [&](GlobalValue::GUID GUID,
1205 const GlobalValueSummary *S) {
1206 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath();
1207 };
1208 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
1209 GlobalValue::GUID GUID,
1210 GlobalValue::LinkageTypes NewLinkage) {
1211 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
1212 };
1213 thinLTOResolveWeakForLinkerInIndex(ThinLTO.CombinedIndex, isPrevailing,
1214 recordNewLinkage);
1215
1216 std::unique_ptr<ThinBackendProc> BackendProc =
1217 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1218 AddStream, Cache);
1219
1220 // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for combined
1221 // module and parallel code generation partitions.
1222 unsigned Task = RegularLTO.ParallelCodeGenParallelismLevel;
1223 for (auto &Mod : ThinLTO.ModuleMap) {
1224 if (Error E = BackendProc->start(Task, Mod.second, ImportLists[Mod.first],
1225 ExportLists[Mod.first],
1226 ResolvedODR[Mod.first], ThinLTO.ModuleMap))
1227 return E;
1228 ++Task;
1229 }
1230
1231 return BackendProc->wait();
1232 }
1233
1234 Expected<std::unique_ptr<ToolOutputFile>>
setupOptimizationRemarks(LLVMContext & Context,StringRef LTORemarksFilename,bool LTOPassRemarksWithHotness,int Count)1235 lto::setupOptimizationRemarks(LLVMContext &Context,
1236 StringRef LTORemarksFilename,
1237 bool LTOPassRemarksWithHotness, int Count) {
1238 if (LTOPassRemarksWithHotness)
1239 Context.setDiagnosticsHotnessRequested(true);
1240 if (LTORemarksFilename.empty())
1241 return nullptr;
1242
1243 std::string Filename = LTORemarksFilename;
1244 if (Count != -1)
1245 Filename += ".thin." + llvm::utostr(Count) + ".yaml";
1246
1247 std::error_code EC;
1248 auto DiagnosticFile =
1249 llvm::make_unique<ToolOutputFile>(Filename, EC, sys::fs::F_None);
1250 if (EC)
1251 return errorCodeToError(EC);
1252 Context.setDiagnosticsOutputFile(
1253 llvm::make_unique<yaml::Output>(DiagnosticFile->os()));
1254 DiagnosticFile->keep();
1255 return std::move(DiagnosticFile);
1256 }
1257