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/Statistic.h"
15 #include "llvm/Analysis/TargetLibraryInfo.h"
16 #include "llvm/Analysis/TargetTransformInfo.h"
17 #include "llvm/Bitcode/BitcodeReader.h"
18 #include "llvm/Bitcode/BitcodeWriter.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/Config/llvm-config.h"
21 #include "llvm/IR/AutoUpgrade.h"
22 #include "llvm/IR/DiagnosticPrinter.h"
23 #include "llvm/IR/Intrinsics.h"
24 #include "llvm/IR/LegacyPassManager.h"
25 #include "llvm/IR/Mangler.h"
26 #include "llvm/IR/Metadata.h"
27 #include "llvm/IR/RemarkStreamer.h"
28 #include "llvm/LTO/LTOBackend.h"
29 #include "llvm/LTO/SummaryBasedOptimizations.h"
30 #include "llvm/Linker/IRMover.h"
31 #include "llvm/Object/IRObjectFile.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Error.h"
34 #include "llvm/Support/ManagedStatic.h"
35 #include "llvm/Support/MemoryBuffer.h"
36 #include "llvm/Support/Path.h"
37 #include "llvm/Support/SHA1.h"
38 #include "llvm/Support/SourceMgr.h"
39 #include "llvm/Support/TargetRegistry.h"
40 #include "llvm/Support/ThreadPool.h"
41 #include "llvm/Support/Threading.h"
42 #include "llvm/Support/VCSRevision.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetMachine.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/Transforms/IPO.h"
47 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
48 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
49 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
50 #include "llvm/Transforms/Utils/SplitModule.h"
51
52 #include <set>
53
54 using namespace llvm;
55 using namespace lto;
56 using namespace object;
57
58 #define DEBUG_TYPE "lto"
59
60 static cl::opt<bool>
61 DumpThinCGSCCs("dump-thin-cg-sccs", cl::init(false), cl::Hidden,
62 cl::desc("Dump the SCCs in the ThinLTO index's callgraph"));
63
64 /// Enable global value internalization in LTO.
65 cl::opt<bool> EnableLTOInternalization(
66 "enable-lto-internalization", cl::init(true), cl::Hidden,
67 cl::desc("Enable global value internalization in LTO"));
68
69 // Computes a unique hash for the Module considering the current list of
70 // export/import and other global analysis results.
71 // 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)72 void llvm::computeLTOCacheKey(
73 SmallString<40> &Key, const Config &Conf, const ModuleSummaryIndex &Index,
74 StringRef ModuleID, const FunctionImporter::ImportMapTy &ImportList,
75 const FunctionImporter::ExportSetTy &ExportList,
76 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
77 const GVSummaryMapTy &DefinedGlobals,
78 const std::set<GlobalValue::GUID> &CfiFunctionDefs,
79 const std::set<GlobalValue::GUID> &CfiFunctionDecls) {
80 // Compute the unique hash for this entry.
81 // This is based on the current compiler version, the module itself, the
82 // export list, the hash for every single module in the import list, the
83 // list of ResolvedODR for the module, and the list of preserved symbols.
84 SHA1 Hasher;
85
86 // Start with the compiler revision
87 Hasher.update(LLVM_VERSION_STRING);
88 #ifdef LLVM_REVISION
89 Hasher.update(LLVM_REVISION);
90 #endif
91
92 // Include the parts of the LTO configuration that affect code generation.
93 auto AddString = [&](StringRef Str) {
94 Hasher.update(Str);
95 Hasher.update(ArrayRef<uint8_t>{0});
96 };
97 auto AddUnsigned = [&](unsigned I) {
98 uint8_t Data[4];
99 Data[0] = I;
100 Data[1] = I >> 8;
101 Data[2] = I >> 16;
102 Data[3] = I >> 24;
103 Hasher.update(ArrayRef<uint8_t>{Data, 4});
104 };
105 auto AddUint64 = [&](uint64_t I) {
106 uint8_t Data[8];
107 Data[0] = I;
108 Data[1] = I >> 8;
109 Data[2] = I >> 16;
110 Data[3] = I >> 24;
111 Data[4] = I >> 32;
112 Data[5] = I >> 40;
113 Data[6] = I >> 48;
114 Data[7] = I >> 56;
115 Hasher.update(ArrayRef<uint8_t>{Data, 8});
116 };
117 AddString(Conf.CPU);
118 // FIXME: Hash more of Options. For now all clients initialize Options from
119 // command-line flags (which is unsupported in production), but may set
120 // RelaxELFRelocations. The clang driver can also pass FunctionSections,
121 // DataSections and DebuggerTuning via command line flags.
122 AddUnsigned(Conf.Options.RelaxELFRelocations);
123 AddUnsigned(Conf.Options.FunctionSections);
124 AddUnsigned(Conf.Options.DataSections);
125 AddUnsigned((unsigned)Conf.Options.DebuggerTuning);
126 for (auto &A : Conf.MAttrs)
127 AddString(A);
128 if (Conf.RelocModel)
129 AddUnsigned(*Conf.RelocModel);
130 else
131 AddUnsigned(-1);
132 if (Conf.CodeModel)
133 AddUnsigned(*Conf.CodeModel);
134 else
135 AddUnsigned(-1);
136 AddUnsigned(Conf.CGOptLevel);
137 AddUnsigned(Conf.CGFileType);
138 AddUnsigned(Conf.OptLevel);
139 AddUnsigned(Conf.UseNewPM);
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 for (const auto &VI : ExportList) {
151 auto GUID = VI.getGUID();
152 // The export list can impact the internalization, be conservative here
153 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&GUID, sizeof(GUID)));
154 }
155
156 // Include the hash for every module we import functions from. The set of
157 // imported symbols for each module may affect code generation and is
158 // sensitive to link order, so include that as well.
159 for (auto &Entry : ImportList) {
160 auto ModHash = Index.getModuleHash(Entry.first());
161 Hasher.update(ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash)));
162
163 AddUint64(Entry.second.size());
164 for (auto &Fn : Entry.second)
165 AddUint64(Fn);
166 }
167
168 // Include the hash for the resolved ODR.
169 for (auto &Entry : ResolvedODR) {
170 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.first,
171 sizeof(GlobalValue::GUID)));
172 Hasher.update(ArrayRef<uint8_t>((const uint8_t *)&Entry.second,
173 sizeof(GlobalValue::LinkageTypes)));
174 }
175
176 // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or
177 // defined in this module.
178 std::set<GlobalValue::GUID> UsedCfiDefs;
179 std::set<GlobalValue::GUID> UsedCfiDecls;
180
181 // Typeids used in this module.
182 std::set<GlobalValue::GUID> UsedTypeIds;
183
184 auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) {
185 if (CfiFunctionDefs.count(ValueGUID))
186 UsedCfiDefs.insert(ValueGUID);
187 if (CfiFunctionDecls.count(ValueGUID))
188 UsedCfiDecls.insert(ValueGUID);
189 };
190
191 auto AddUsedThings = [&](GlobalValueSummary *GS) {
192 if (!GS) return;
193 AddUnsigned(GS->isLive());
194 AddUnsigned(GS->canAutoHide());
195 for (const ValueInfo &VI : GS->refs()) {
196 AddUnsigned(VI.isDSOLocal());
197 AddUsedCfiGlobal(VI.getGUID());
198 }
199 if (auto *GVS = dyn_cast<GlobalVarSummary>(GS)) {
200 AddUnsigned(GVS->maybeReadOnly());
201 AddUnsigned(GVS->maybeWriteOnly());
202 }
203 if (auto *FS = dyn_cast<FunctionSummary>(GS)) {
204 for (auto &TT : FS->type_tests())
205 UsedTypeIds.insert(TT);
206 for (auto &TT : FS->type_test_assume_vcalls())
207 UsedTypeIds.insert(TT.GUID);
208 for (auto &TT : FS->type_checked_load_vcalls())
209 UsedTypeIds.insert(TT.GUID);
210 for (auto &TT : FS->type_test_assume_const_vcalls())
211 UsedTypeIds.insert(TT.VFunc.GUID);
212 for (auto &TT : FS->type_checked_load_const_vcalls())
213 UsedTypeIds.insert(TT.VFunc.GUID);
214 for (auto &ET : FS->calls()) {
215 AddUnsigned(ET.first.isDSOLocal());
216 AddUsedCfiGlobal(ET.first.getGUID());
217 }
218 }
219 };
220
221 // Include the hash for the linkage type to reflect internalization and weak
222 // resolution, and collect any used type identifier resolutions.
223 for (auto &GS : DefinedGlobals) {
224 GlobalValue::LinkageTypes Linkage = GS.second->linkage();
225 Hasher.update(
226 ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage)));
227 AddUsedCfiGlobal(GS.first);
228 AddUsedThings(GS.second);
229 }
230
231 // Imported functions may introduce new uses of type identifier resolutions,
232 // so we need to collect their used resolutions as well.
233 for (auto &ImpM : ImportList)
234 for (auto &ImpF : ImpM.second) {
235 GlobalValueSummary *S = Index.findSummaryInModule(ImpF, ImpM.first());
236 AddUsedThings(S);
237 // If this is an alias, we also care about any types/etc. that the aliasee
238 // may reference.
239 if (auto *AS = dyn_cast_or_null<AliasSummary>(S))
240 AddUsedThings(AS->getBaseObject());
241 }
242
243 auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) {
244 AddString(TId);
245
246 AddUnsigned(S.TTRes.TheKind);
247 AddUnsigned(S.TTRes.SizeM1BitWidth);
248
249 AddUint64(S.TTRes.AlignLog2);
250 AddUint64(S.TTRes.SizeM1);
251 AddUint64(S.TTRes.BitMask);
252 AddUint64(S.TTRes.InlineBits);
253
254 AddUint64(S.WPDRes.size());
255 for (auto &WPD : S.WPDRes) {
256 AddUnsigned(WPD.first);
257 AddUnsigned(WPD.second.TheKind);
258 AddString(WPD.second.SingleImplName);
259
260 AddUint64(WPD.second.ResByArg.size());
261 for (auto &ByArg : WPD.second.ResByArg) {
262 AddUint64(ByArg.first.size());
263 for (uint64_t Arg : ByArg.first)
264 AddUint64(Arg);
265 AddUnsigned(ByArg.second.TheKind);
266 AddUint64(ByArg.second.Info);
267 AddUnsigned(ByArg.second.Byte);
268 AddUnsigned(ByArg.second.Bit);
269 }
270 }
271 };
272
273 // Include the hash for all type identifiers used by this module.
274 for (GlobalValue::GUID TId : UsedTypeIds) {
275 auto TidIter = Index.typeIds().equal_range(TId);
276 for (auto It = TidIter.first; It != TidIter.second; ++It)
277 AddTypeIdSummary(It->second.first, It->second.second);
278 }
279
280 AddUnsigned(UsedCfiDefs.size());
281 for (auto &V : UsedCfiDefs)
282 AddUint64(V);
283
284 AddUnsigned(UsedCfiDecls.size());
285 for (auto &V : UsedCfiDecls)
286 AddUint64(V);
287
288 if (!Conf.SampleProfile.empty()) {
289 auto FileOrErr = MemoryBuffer::getFile(Conf.SampleProfile);
290 if (FileOrErr) {
291 Hasher.update(FileOrErr.get()->getBuffer());
292
293 if (!Conf.ProfileRemapping.empty()) {
294 FileOrErr = MemoryBuffer::getFile(Conf.ProfileRemapping);
295 if (FileOrErr)
296 Hasher.update(FileOrErr.get()->getBuffer());
297 }
298 }
299 }
300
301 Key = toHex(Hasher.result());
302 }
303
thinLTOResolvePrevailingGUID(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)304 static void thinLTOResolvePrevailingGUID(
305 ValueInfo VI, DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias,
306 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
307 isPrevailing,
308 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
309 recordNewLinkage,
310 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
311 for (auto &S : VI.getSummaryList()) {
312 GlobalValue::LinkageTypes OriginalLinkage = S->linkage();
313 // Ignore local and appending linkage values since the linker
314 // doesn't resolve them.
315 if (GlobalValue::isLocalLinkage(OriginalLinkage) ||
316 GlobalValue::isAppendingLinkage(S->linkage()))
317 continue;
318 // We need to emit only one of these. The prevailing module will keep it,
319 // but turned into a weak, while the others will drop it when possible.
320 // This is both a compile-time optimization and a correctness
321 // transformation. This is necessary for correctness when we have exported
322 // a reference - we need to convert the linkonce to weak to
323 // ensure a copy is kept to satisfy the exported reference.
324 // FIXME: We may want to split the compile time and correctness
325 // aspects into separate routines.
326 if (isPrevailing(VI.getGUID(), S.get())) {
327 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage)) {
328 S->setLinkage(GlobalValue::getWeakLinkage(
329 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
330 // The kept copy is eligible for auto-hiding (hidden visibility) if all
331 // copies were (i.e. they were all linkonce_odr global unnamed addr).
332 // If any copy is not (e.g. it was originally weak_odr), then the symbol
333 // must remain externally available (e.g. a weak_odr from an explicitly
334 // instantiated template). Additionally, if it is in the
335 // GUIDPreservedSymbols set, that means that it is visibile outside
336 // the summary (e.g. in a native object or a bitcode file without
337 // summary), and in that case we cannot hide it as it isn't possible to
338 // check all copies.
339 S->setCanAutoHide(VI.canAutoHide() &&
340 !GUIDPreservedSymbols.count(VI.getGUID()));
341 }
342 }
343 // Alias and aliasee can't be turned into available_externally.
344 else if (!isa<AliasSummary>(S.get()) &&
345 !GlobalInvolvedWithAlias.count(S.get()))
346 S->setLinkage(GlobalValue::AvailableExternallyLinkage);
347 if (S->linkage() != OriginalLinkage)
348 recordNewLinkage(S->modulePath(), VI.getGUID(), S->linkage());
349 }
350 }
351
352 /// Resolve linkage for prevailing symbols in the \p Index.
353 //
354 // We'd like to drop these functions if they are no longer referenced in the
355 // current module. However there is a chance that another module is still
356 // referencing them because of the import. We make sure we always emit at least
357 // one copy.
thinLTOResolvePrevailingInIndex(ModuleSummaryIndex & Index,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing,function_ref<void (StringRef,GlobalValue::GUID,GlobalValue::LinkageTypes)> recordNewLinkage,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols)358 void llvm::thinLTOResolvePrevailingInIndex(
359 ModuleSummaryIndex &Index,
360 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
361 isPrevailing,
362 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
363 recordNewLinkage,
364 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
365 // We won't optimize the globals that are referenced by an alias for now
366 // Ideally we should turn the alias into a global and duplicate the definition
367 // when needed.
368 DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias;
369 for (auto &I : Index)
370 for (auto &S : I.second.SummaryList)
371 if (auto AS = dyn_cast<AliasSummary>(S.get()))
372 GlobalInvolvedWithAlias.insert(&AS->getAliasee());
373
374 for (auto &I : Index)
375 thinLTOResolvePrevailingGUID(Index.getValueInfo(I), GlobalInvolvedWithAlias,
376 isPrevailing, recordNewLinkage,
377 GUIDPreservedSymbols);
378 }
379
isWeakObjectWithRWAccess(GlobalValueSummary * GVS)380 static bool isWeakObjectWithRWAccess(GlobalValueSummary *GVS) {
381 if (auto *VarSummary = dyn_cast<GlobalVarSummary>(GVS->getBaseObject()))
382 return !VarSummary->maybeReadOnly() && !VarSummary->maybeWriteOnly() &&
383 (VarSummary->linkage() == GlobalValue::WeakODRLinkage ||
384 VarSummary->linkage() == GlobalValue::LinkOnceODRLinkage);
385 return false;
386 }
387
thinLTOInternalizeAndPromoteGUID(ValueInfo VI,function_ref<bool (StringRef,ValueInfo)> isExported,function_ref<bool (GlobalValue::GUID,const GlobalValueSummary *)> isPrevailing)388 static void thinLTOInternalizeAndPromoteGUID(
389 ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported,
390 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
391 isPrevailing) {
392 for (auto &S : VI.getSummaryList()) {
393 if (isExported(S->modulePath(), VI)) {
394 if (GlobalValue::isLocalLinkage(S->linkage()))
395 S->setLinkage(GlobalValue::ExternalLinkage);
396 } else if (EnableLTOInternalization &&
397 // Ignore local and appending linkage values since the linker
398 // doesn't resolve them.
399 !GlobalValue::isLocalLinkage(S->linkage()) &&
400 (!GlobalValue::isInterposableLinkage(S->linkage()) ||
401 isPrevailing(VI.getGUID(), S.get())) &&
402 S->linkage() != GlobalValue::AppendingLinkage &&
403 // We can't internalize available_externally globals because this
404 // can break function pointer equality.
405 S->linkage() != GlobalValue::AvailableExternallyLinkage &&
406 // Functions and read-only variables with linkonce_odr and
407 // weak_odr linkage can be internalized. We can't internalize
408 // linkonce_odr and weak_odr variables which are both modified
409 // and read somewhere in the program because reads and writes
410 // will become inconsistent.
411 !isWeakObjectWithRWAccess(S.get()))
412 S->setLinkage(GlobalValue::InternalLinkage);
413 }
414 }
415
416 // Update the linkages in the given \p Index to mark exported values
417 // 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)418 void llvm::thinLTOInternalizeAndPromoteInIndex(
419 ModuleSummaryIndex &Index,
420 function_ref<bool(StringRef, ValueInfo)> isExported,
421 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
422 isPrevailing) {
423 for (auto &I : Index)
424 thinLTOInternalizeAndPromoteGUID(Index.getValueInfo(I), isExported,
425 isPrevailing);
426 }
427
428 // Requires a destructor for std::vector<InputModule>.
429 InputFile::~InputFile() = default;
430
create(MemoryBufferRef Object)431 Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) {
432 std::unique_ptr<InputFile> File(new InputFile);
433
434 Expected<IRSymtabFile> FOrErr = readIRSymtab(Object);
435 if (!FOrErr)
436 return FOrErr.takeError();
437
438 File->TargetTriple = FOrErr->TheReader.getTargetTriple();
439 File->SourceFileName = FOrErr->TheReader.getSourceFileName();
440 File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts();
441 File->DependentLibraries = FOrErr->TheReader.getDependentLibraries();
442 File->ComdatTable = FOrErr->TheReader.getComdatTable();
443
444 for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) {
445 size_t Begin = File->Symbols.size();
446 for (const irsymtab::Reader::SymbolRef &Sym :
447 FOrErr->TheReader.module_symbols(I))
448 // Skip symbols that are irrelevant to LTO. Note that this condition needs
449 // to match the one in Skip() in LTO::addRegularLTO().
450 if (Sym.isGlobal() && !Sym.isFormatSpecific())
451 File->Symbols.push_back(Sym);
452 File->ModuleSymIndices.push_back({Begin, File->Symbols.size()});
453 }
454
455 File->Mods = FOrErr->Mods;
456 File->Strtab = std::move(FOrErr->Strtab);
457 return std::move(File);
458 }
459
getName() const460 StringRef InputFile::getName() const {
461 return Mods[0].getModuleIdentifier();
462 }
463
getSingleBitcodeModule()464 BitcodeModule &InputFile::getSingleBitcodeModule() {
465 assert(Mods.size() == 1 && "Expect only one bitcode module");
466 return Mods[0];
467 }
468
RegularLTOState(unsigned ParallelCodeGenParallelismLevel,const Config & Conf)469 LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel,
470 const Config &Conf)
471 : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel),
472 Ctx(Conf), CombinedModule(std::make_unique<Module>("ld-temp.o", Ctx)),
473 Mover(std::make_unique<IRMover>(*CombinedModule)) {}
474
ThinLTOState(ThinBackend Backend)475 LTO::ThinLTOState::ThinLTOState(ThinBackend Backend)
476 : Backend(Backend), CombinedIndex(/*HaveGVs*/ false) {
477 if (!Backend)
478 this->Backend =
479 createInProcessThinBackend(llvm::heavyweight_hardware_concurrency());
480 }
481
LTO(Config Conf,ThinBackend Backend,unsigned ParallelCodeGenParallelismLevel)482 LTO::LTO(Config Conf, ThinBackend Backend,
483 unsigned ParallelCodeGenParallelismLevel)
484 : Conf(std::move(Conf)),
485 RegularLTO(ParallelCodeGenParallelismLevel, this->Conf),
486 ThinLTO(std::move(Backend)) {}
487
488 // Requires a destructor for MapVector<BitcodeModule>.
489 LTO::~LTO() = default;
490
491 // Add the symbols in the given module to the GlobalResolutions map, and resolve
492 // their partitions.
addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,ArrayRef<SymbolResolution> Res,unsigned Partition,bool InSummary)493 void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
494 ArrayRef<SymbolResolution> Res,
495 unsigned Partition, bool InSummary) {
496 auto *ResI = Res.begin();
497 auto *ResE = Res.end();
498 (void)ResE;
499 for (const InputFile::Symbol &Sym : Syms) {
500 assert(ResI != ResE);
501 SymbolResolution Res = *ResI++;
502
503 StringRef Name = Sym.getName();
504 Triple TT(RegularLTO.CombinedModule->getTargetTriple());
505 // Strip the __imp_ prefix from COFF dllimport symbols (similar to the
506 // way they are handled by lld), otherwise we can end up with two
507 // global resolutions (one with and one for a copy of the symbol without).
508 if (TT.isOSBinFormatCOFF() && Name.startswith("__imp_"))
509 Name = Name.substr(strlen("__imp_"));
510 auto &GlobalRes = GlobalResolutions[Name];
511 GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr();
512 if (Res.Prevailing) {
513 assert(!GlobalRes.Prevailing &&
514 "Multiple prevailing defs are not allowed");
515 GlobalRes.Prevailing = true;
516 GlobalRes.IRName = Sym.getIRName();
517 } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) {
518 // Sometimes it can be two copies of symbol in a module and prevailing
519 // symbol can have no IR name. That might happen if symbol is defined in
520 // module level inline asm block. In case we have multiple modules with
521 // the same symbol we want to use IR name of the prevailing symbol.
522 // Otherwise, if we haven't seen a prevailing symbol, set the name so that
523 // we can later use it to check if there is any prevailing copy in IR.
524 GlobalRes.IRName = Sym.getIRName();
525 }
526
527 // Set the partition to external if we know it is re-defined by the linker
528 // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a
529 // regular object, is referenced from llvm.compiler_used, or was already
530 // recorded as being referenced from a different partition.
531 if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() ||
532 (GlobalRes.Partition != GlobalResolution::Unknown &&
533 GlobalRes.Partition != Partition)) {
534 GlobalRes.Partition = GlobalResolution::External;
535 } else
536 // First recorded reference, save the current partition.
537 GlobalRes.Partition = Partition;
538
539 // Flag as visible outside of summary if visible from a regular object or
540 // from a module that does not have a summary.
541 GlobalRes.VisibleOutsideSummary |=
542 (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary);
543 }
544 }
545
writeToResolutionFile(raw_ostream & OS,InputFile * Input,ArrayRef<SymbolResolution> Res)546 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input,
547 ArrayRef<SymbolResolution> Res) {
548 StringRef Path = Input->getName();
549 OS << Path << '\n';
550 auto ResI = Res.begin();
551 for (const InputFile::Symbol &Sym : Input->symbols()) {
552 assert(ResI != Res.end());
553 SymbolResolution Res = *ResI++;
554
555 OS << "-r=" << Path << ',' << Sym.getName() << ',';
556 if (Res.Prevailing)
557 OS << 'p';
558 if (Res.FinalDefinitionInLinkageUnit)
559 OS << 'l';
560 if (Res.VisibleToRegularObj)
561 OS << 'x';
562 if (Res.LinkerRedefined)
563 OS << 'r';
564 OS << '\n';
565 }
566 OS.flush();
567 assert(ResI == Res.end());
568 }
569
add(std::unique_ptr<InputFile> Input,ArrayRef<SymbolResolution> Res)570 Error LTO::add(std::unique_ptr<InputFile> Input,
571 ArrayRef<SymbolResolution> Res) {
572 assert(!CalledGetMaxTasks);
573
574 if (Conf.ResolutionFile)
575 writeToResolutionFile(*Conf.ResolutionFile, Input.get(), Res);
576
577 if (RegularLTO.CombinedModule->getTargetTriple().empty())
578 RegularLTO.CombinedModule->setTargetTriple(Input->getTargetTriple());
579
580 const SymbolResolution *ResI = Res.begin();
581 for (unsigned I = 0; I != Input->Mods.size(); ++I)
582 if (Error Err = addModule(*Input, I, ResI, Res.end()))
583 return Err;
584
585 assert(ResI == Res.end());
586 return Error::success();
587 }
588
addModule(InputFile & Input,unsigned ModI,const SymbolResolution * & ResI,const SymbolResolution * ResE)589 Error LTO::addModule(InputFile &Input, unsigned ModI,
590 const SymbolResolution *&ResI,
591 const SymbolResolution *ResE) {
592 Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo();
593 if (!LTOInfo)
594 return LTOInfo.takeError();
595
596 if (EnableSplitLTOUnit.hasValue()) {
597 // If only some modules were split, flag this in the index so that
598 // we can skip or error on optimizations that need consistently split
599 // modules (whole program devirt and lower type tests).
600 if (EnableSplitLTOUnit.getValue() != LTOInfo->EnableSplitLTOUnit)
601 ThinLTO.CombinedIndex.setPartiallySplitLTOUnits();
602 } else
603 EnableSplitLTOUnit = LTOInfo->EnableSplitLTOUnit;
604
605 BitcodeModule BM = Input.Mods[ModI];
606 auto ModSyms = Input.module_symbols(ModI);
607 addModuleToGlobalRes(ModSyms, {ResI, ResE},
608 LTOInfo->IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0,
609 LTOInfo->HasSummary);
610
611 if (LTOInfo->IsThinLTO)
612 return addThinLTO(BM, ModSyms, ResI, ResE);
613
614 Expected<RegularLTOState::AddedModule> ModOrErr =
615 addRegularLTO(BM, ModSyms, ResI, ResE);
616 if (!ModOrErr)
617 return ModOrErr.takeError();
618
619 if (!LTOInfo->HasSummary)
620 return linkRegularLTO(std::move(*ModOrErr), /*LivenessFromIndex=*/false);
621
622 // Regular LTO module summaries are added to a dummy module that represents
623 // the combined regular LTO module.
624 if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, "", -1ull))
625 return Err;
626 RegularLTO.ModsWithSummaries.push_back(std::move(*ModOrErr));
627 return Error::success();
628 }
629
630 // Checks whether the given global value is in a non-prevailing comdat
631 // (comdat containing values the linker indicated were not prevailing,
632 // which we then dropped to available_externally), and if so, removes
633 // it from the comdat. This is called for all global values to ensure the
634 // comdat is empty rather than leaving an incomplete comdat. It is needed for
635 // regular LTO modules, in case we are in a mixed-LTO mode (both regular
636 // and thin LTO modules) compilation. Since the regular LTO module will be
637 // linked first in the final native link, we want to make sure the linker
638 // doesn't select any of these incomplete comdats that would be left
639 // in the regular LTO module without this cleanup.
640 static void
handleNonPrevailingComdat(GlobalValue & GV,std::set<const Comdat * > & NonPrevailingComdats)641 handleNonPrevailingComdat(GlobalValue &GV,
642 std::set<const Comdat *> &NonPrevailingComdats) {
643 Comdat *C = GV.getComdat();
644 if (!C)
645 return;
646
647 if (!NonPrevailingComdats.count(C))
648 return;
649
650 // Additionally need to drop externally visible global values from the comdat
651 // to available_externally, so that there aren't multiply defined linker
652 // errors.
653 if (!GV.hasLocalLinkage())
654 GV.setLinkage(GlobalValue::AvailableExternallyLinkage);
655
656 if (auto GO = dyn_cast<GlobalObject>(&GV))
657 GO->setComdat(nullptr);
658 }
659
660 // Add a regular LTO object to the link.
661 // The resulting module needs to be linked into the combined LTO module with
662 // linkRegularLTO.
663 Expected<LTO::RegularLTOState::AddedModule>
addRegularLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)664 LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
665 const SymbolResolution *&ResI,
666 const SymbolResolution *ResE) {
667 RegularLTOState::AddedModule Mod;
668 Expected<std::unique_ptr<Module>> MOrErr =
669 BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true,
670 /*IsImporting*/ false);
671 if (!MOrErr)
672 return MOrErr.takeError();
673 Module &M = **MOrErr;
674 Mod.M = std::move(*MOrErr);
675
676 if (Error Err = M.materializeMetadata())
677 return std::move(Err);
678 UpgradeDebugInfo(M);
679
680 ModuleSymbolTable SymTab;
681 SymTab.addModule(&M);
682
683 for (GlobalVariable &GV : M.globals())
684 if (GV.hasAppendingLinkage())
685 Mod.Keep.push_back(&GV);
686
687 DenseSet<GlobalObject *> AliasedGlobals;
688 for (auto &GA : M.aliases())
689 if (GlobalObject *GO = GA.getBaseObject())
690 AliasedGlobals.insert(GO);
691
692 // In this function we need IR GlobalValues matching the symbols in Syms
693 // (which is not backed by a module), so we need to enumerate them in the same
694 // order. The symbol enumeration order of a ModuleSymbolTable intentionally
695 // matches the order of an irsymtab, but when we read the irsymtab in
696 // InputFile::create we omit some symbols that are irrelevant to LTO. The
697 // Skip() function skips the same symbols from the module as InputFile does
698 // from the symbol table.
699 auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end();
700 auto Skip = [&]() {
701 while (MsymI != MsymE) {
702 auto Flags = SymTab.getSymbolFlags(*MsymI);
703 if ((Flags & object::BasicSymbolRef::SF_Global) &&
704 !(Flags & object::BasicSymbolRef::SF_FormatSpecific))
705 return;
706 ++MsymI;
707 }
708 };
709 Skip();
710
711 std::set<const Comdat *> NonPrevailingComdats;
712 for (const InputFile::Symbol &Sym : Syms) {
713 assert(ResI != ResE);
714 SymbolResolution Res = *ResI++;
715
716 assert(MsymI != MsymE);
717 ModuleSymbolTable::Symbol Msym = *MsymI++;
718 Skip();
719
720 if (GlobalValue *GV = Msym.dyn_cast<GlobalValue *>()) {
721 if (Res.Prevailing) {
722 if (Sym.isUndefined())
723 continue;
724 Mod.Keep.push_back(GV);
725 // For symbols re-defined with linker -wrap and -defsym options,
726 // set the linkage to weak to inhibit IPO. The linkage will be
727 // restored by the linker.
728 if (Res.LinkerRedefined)
729 GV->setLinkage(GlobalValue::WeakAnyLinkage);
730
731 GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage();
732 if (GlobalValue::isLinkOnceLinkage(OriginalLinkage))
733 GV->setLinkage(GlobalValue::getWeakLinkage(
734 GlobalValue::isLinkOnceODRLinkage(OriginalLinkage)));
735 } else if (isa<GlobalObject>(GV) &&
736 (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() ||
737 GV->hasAvailableExternallyLinkage()) &&
738 !AliasedGlobals.count(cast<GlobalObject>(GV))) {
739 // Any of the above three types of linkage indicates that the
740 // chosen prevailing symbol will have the same semantics as this copy of
741 // the symbol, so we may be able to link it with available_externally
742 // linkage. We will decide later whether to do that when we link this
743 // module (in linkRegularLTO), based on whether it is undefined.
744 Mod.Keep.push_back(GV);
745 GV->setLinkage(GlobalValue::AvailableExternallyLinkage);
746 if (GV->hasComdat())
747 NonPrevailingComdats.insert(GV->getComdat());
748 cast<GlobalObject>(GV)->setComdat(nullptr);
749 }
750
751 // Set the 'local' flag based on the linker resolution for this symbol.
752 if (Res.FinalDefinitionInLinkageUnit) {
753 GV->setDSOLocal(true);
754 if (GV->hasDLLImportStorageClass())
755 GV->setDLLStorageClass(GlobalValue::DLLStorageClassTypes::
756 DefaultStorageClass);
757 }
758 }
759 // Common resolution: collect the maximum size/alignment over all commons.
760 // We also record if we see an instance of a common as prevailing, so that
761 // if none is prevailing we can ignore it later.
762 if (Sym.isCommon()) {
763 // FIXME: We should figure out what to do about commons defined by asm.
764 // For now they aren't reported correctly by ModuleSymbolTable.
765 auto &CommonRes = RegularLTO.Commons[Sym.getIRName()];
766 CommonRes.Size = std::max(CommonRes.Size, Sym.getCommonSize());
767 CommonRes.Align =
768 std::max(CommonRes.Align, MaybeAlign(Sym.getCommonAlignment()));
769 CommonRes.Prevailing |= Res.Prevailing;
770 }
771
772 }
773 if (!M.getComdatSymbolTable().empty())
774 for (GlobalValue &GV : M.global_values())
775 handleNonPrevailingComdat(GV, NonPrevailingComdats);
776 assert(MsymI == MsymE);
777 return std::move(Mod);
778 }
779
linkRegularLTO(RegularLTOState::AddedModule Mod,bool LivenessFromIndex)780 Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod,
781 bool LivenessFromIndex) {
782 std::vector<GlobalValue *> Keep;
783 for (GlobalValue *GV : Mod.Keep) {
784 if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GV->getGUID()))
785 continue;
786
787 if (!GV->hasAvailableExternallyLinkage()) {
788 Keep.push_back(GV);
789 continue;
790 }
791
792 // Only link available_externally definitions if we don't already have a
793 // definition.
794 GlobalValue *CombinedGV =
795 RegularLTO.CombinedModule->getNamedValue(GV->getName());
796 if (CombinedGV && !CombinedGV->isDeclaration())
797 continue;
798
799 Keep.push_back(GV);
800 }
801
802 return RegularLTO.Mover->move(std::move(Mod.M), Keep,
803 [](GlobalValue &, IRMover::ValueAdder) {},
804 /* IsPerformingImport */ false);
805 }
806
807 // Add a ThinLTO module to the link.
addThinLTO(BitcodeModule BM,ArrayRef<InputFile::Symbol> Syms,const SymbolResolution * & ResI,const SymbolResolution * ResE)808 Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
809 const SymbolResolution *&ResI,
810 const SymbolResolution *ResE) {
811 if (Error Err =
812 BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(),
813 ThinLTO.ModuleMap.size()))
814 return Err;
815
816 for (const InputFile::Symbol &Sym : Syms) {
817 assert(ResI != ResE);
818 SymbolResolution Res = *ResI++;
819
820 if (!Sym.getIRName().empty()) {
821 auto GUID = GlobalValue::getGUID(GlobalValue::getGlobalIdentifier(
822 Sym.getIRName(), GlobalValue::ExternalLinkage, ""));
823 if (Res.Prevailing) {
824 ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier();
825
826 // For linker redefined symbols (via --wrap or --defsym) we want to
827 // switch the linkage to `weak` to prevent IPOs from happening.
828 // Find the summary in the module for this very GV and record the new
829 // linkage so that we can switch it when we import the GV.
830 if (Res.LinkerRedefined)
831 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
832 GUID, BM.getModuleIdentifier()))
833 S->setLinkage(GlobalValue::WeakAnyLinkage);
834 }
835
836 // If the linker resolved the symbol to a local definition then mark it
837 // as local in the summary for the module we are adding.
838 if (Res.FinalDefinitionInLinkageUnit) {
839 if (auto S = ThinLTO.CombinedIndex.findSummaryInModule(
840 GUID, BM.getModuleIdentifier())) {
841 S->setDSOLocal(true);
842 }
843 }
844 }
845 }
846
847 if (!ThinLTO.ModuleMap.insert({BM.getModuleIdentifier(), BM}).second)
848 return make_error<StringError>(
849 "Expected at most one ThinLTO module per bitcode file",
850 inconvertibleErrorCode());
851
852 return Error::success();
853 }
854
getMaxTasks() const855 unsigned LTO::getMaxTasks() const {
856 CalledGetMaxTasks = true;
857 return RegularLTO.ParallelCodeGenParallelismLevel + ThinLTO.ModuleMap.size();
858 }
859
860 // If only some of the modules were split, we cannot correctly handle
861 // code that contains type tests or type checked loads.
checkPartiallySplit()862 Error LTO::checkPartiallySplit() {
863 if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits())
864 return Error::success();
865
866 Function *TypeTestFunc = RegularLTO.CombinedModule->getFunction(
867 Intrinsic::getName(Intrinsic::type_test));
868 Function *TypeCheckedLoadFunc = RegularLTO.CombinedModule->getFunction(
869 Intrinsic::getName(Intrinsic::type_checked_load));
870
871 // First check if there are type tests / type checked loads in the
872 // merged regular LTO module IR.
873 if ((TypeTestFunc && !TypeTestFunc->use_empty()) ||
874 (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()))
875 return make_error<StringError>(
876 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
877 inconvertibleErrorCode());
878
879 // Otherwise check if there are any recorded in the combined summary from the
880 // ThinLTO modules.
881 for (auto &P : ThinLTO.CombinedIndex) {
882 for (auto &S : P.second.SummaryList) {
883 auto *FS = dyn_cast<FunctionSummary>(S.get());
884 if (!FS)
885 continue;
886 if (!FS->type_test_assume_vcalls().empty() ||
887 !FS->type_checked_load_vcalls().empty() ||
888 !FS->type_test_assume_const_vcalls().empty() ||
889 !FS->type_checked_load_const_vcalls().empty() ||
890 !FS->type_tests().empty())
891 return make_error<StringError>(
892 "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
893 inconvertibleErrorCode());
894 }
895 }
896 return Error::success();
897 }
898
run(AddStreamFn AddStream,NativeObjectCache Cache)899 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) {
900 // Compute "dead" symbols, we don't want to import/export these!
901 DenseSet<GlobalValue::GUID> GUIDPreservedSymbols;
902 DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions;
903 for (auto &Res : GlobalResolutions) {
904 // Normally resolution have IR name of symbol. We can do nothing here
905 // otherwise. See comments in GlobalResolution struct for more details.
906 if (Res.second.IRName.empty())
907 continue;
908
909 GlobalValue::GUID GUID = GlobalValue::getGUID(
910 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
911
912 if (Res.second.VisibleOutsideSummary && Res.second.Prevailing)
913 GUIDPreservedSymbols.insert(GUID);
914
915 GUIDPrevailingResolutions[GUID] =
916 Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No;
917 }
918
919 auto isPrevailing = [&](GlobalValue::GUID G) {
920 auto It = GUIDPrevailingResolutions.find(G);
921 if (It == GUIDPrevailingResolutions.end())
922 return PrevailingType::Unknown;
923 return It->second;
924 };
925 computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols,
926 isPrevailing, Conf.OptLevel > 0);
927
928 // Setup output file to emit statistics.
929 auto StatsFileOrErr = setupStatsFile(Conf.StatsFile);
930 if (!StatsFileOrErr)
931 return StatsFileOrErr.takeError();
932 std::unique_ptr<ToolOutputFile> StatsFile = std::move(StatsFileOrErr.get());
933
934 // Finalize linking of regular LTO modules containing summaries now that
935 // we have computed liveness information.
936 for (auto &M : RegularLTO.ModsWithSummaries)
937 if (Error Err = linkRegularLTO(std::move(M),
938 /*LivenessFromIndex=*/true))
939 return Err;
940
941 // Ensure we don't have inconsistently split LTO units with type tests.
942 if (Error Err = checkPartiallySplit())
943 return Err;
944
945 Error Result = runRegularLTO(AddStream);
946 if (!Result)
947 Result = runThinLTO(AddStream, Cache, GUIDPreservedSymbols);
948
949 if (StatsFile)
950 PrintStatisticsJSON(StatsFile->os());
951
952 return Result;
953 }
954
runRegularLTO(AddStreamFn AddStream)955 Error LTO::runRegularLTO(AddStreamFn AddStream) {
956 // Make sure commons have the right size/alignment: we kept the largest from
957 // all the prevailing when adding the inputs, and we apply it here.
958 const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout();
959 for (auto &I : RegularLTO.Commons) {
960 if (!I.second.Prevailing)
961 // Don't do anything if no instance of this common was prevailing.
962 continue;
963 GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first);
964 if (OldGV && DL.getTypeAllocSize(OldGV->getValueType()) == I.second.Size) {
965 // Don't create a new global if the type is already correct, just make
966 // sure the alignment is correct.
967 OldGV->setAlignment(I.second.Align);
968 continue;
969 }
970 ArrayType *Ty =
971 ArrayType::get(Type::getInt8Ty(RegularLTO.Ctx), I.second.Size);
972 auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false,
973 GlobalValue::CommonLinkage,
974 ConstantAggregateZero::get(Ty), "");
975 GV->setAlignment(I.second.Align);
976 if (OldGV) {
977 OldGV->replaceAllUsesWith(ConstantExpr::getBitCast(GV, OldGV->getType()));
978 GV->takeName(OldGV);
979 OldGV->eraseFromParent();
980 } else {
981 GV->setName(I.first);
982 }
983 }
984
985 if (Conf.PreOptModuleHook &&
986 !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule))
987 return Error::success();
988
989 if (!Conf.CodeGenOnly) {
990 for (const auto &R : GlobalResolutions) {
991 if (!R.second.isPrevailingIRSymbol())
992 continue;
993 if (R.second.Partition != 0 &&
994 R.second.Partition != GlobalResolution::External)
995 continue;
996
997 GlobalValue *GV =
998 RegularLTO.CombinedModule->getNamedValue(R.second.IRName);
999 // Ignore symbols defined in other partitions.
1000 // Also skip declarations, which are not allowed to have internal linkage.
1001 if (!GV || GV->hasLocalLinkage() || GV->isDeclaration())
1002 continue;
1003 GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global
1004 : GlobalValue::UnnamedAddr::None);
1005 if (EnableLTOInternalization && R.second.Partition == 0)
1006 GV->setLinkage(GlobalValue::InternalLinkage);
1007 }
1008
1009 RegularLTO.CombinedModule->addModuleFlag(Module::Error, "LTOPostLink", 1);
1010
1011 if (Conf.PostInternalizeModuleHook &&
1012 !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule))
1013 return Error::success();
1014 }
1015 return backend(Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel,
1016 std::move(RegularLTO.CombinedModule), ThinLTO.CombinedIndex);
1017 }
1018
1019 static const char *libcallRoutineNames[] = {
1020 #define HANDLE_LIBCALL(code, name) name,
1021 #include "llvm/IR/RuntimeLibcalls.def"
1022 #undef HANDLE_LIBCALL
1023 };
1024
getRuntimeLibcallSymbols()1025 ArrayRef<const char*> LTO::getRuntimeLibcallSymbols() {
1026 return makeArrayRef(libcallRoutineNames);
1027 }
1028
1029 /// This class defines the interface to the ThinLTO backend.
1030 class lto::ThinBackendProc {
1031 protected:
1032 const Config &Conf;
1033 ModuleSummaryIndex &CombinedIndex;
1034 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries;
1035
1036 public:
ThinBackendProc(const Config & Conf,ModuleSummaryIndex & CombinedIndex,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries)1037 ThinBackendProc(const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1038 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries)
1039 : Conf(Conf), CombinedIndex(CombinedIndex),
1040 ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries) {}
1041
~ThinBackendProc()1042 virtual ~ThinBackendProc() {}
1043 virtual Error start(
1044 unsigned Task, BitcodeModule BM,
1045 const FunctionImporter::ImportMapTy &ImportList,
1046 const FunctionImporter::ExportSetTy &ExportList,
1047 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1048 MapVector<StringRef, BitcodeModule> &ModuleMap) = 0;
1049 virtual Error wait() = 0;
1050 };
1051
1052 namespace {
1053 class InProcessThinBackend : public ThinBackendProc {
1054 ThreadPool BackendThreadPool;
1055 AddStreamFn AddStream;
1056 NativeObjectCache Cache;
1057 std::set<GlobalValue::GUID> CfiFunctionDefs;
1058 std::set<GlobalValue::GUID> CfiFunctionDecls;
1059
1060 Optional<Error> Err;
1061 std::mutex ErrMu;
1062
1063 public:
InProcessThinBackend(const Config & Conf,ModuleSummaryIndex & CombinedIndex,unsigned ThinLTOParallelismLevel,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,AddStreamFn AddStream,NativeObjectCache Cache)1064 InProcessThinBackend(
1065 const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1066 unsigned ThinLTOParallelismLevel,
1067 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1068 AddStreamFn AddStream, NativeObjectCache Cache)
1069 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries),
1070 BackendThreadPool(ThinLTOParallelismLevel),
1071 AddStream(std::move(AddStream)), Cache(std::move(Cache)) {
1072 for (auto &Name : CombinedIndex.cfiFunctionDefs())
1073 CfiFunctionDefs.insert(
1074 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
1075 for (auto &Name : CombinedIndex.cfiFunctionDecls())
1076 CfiFunctionDecls.insert(
1077 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
1078 }
1079
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)1080 Error runThinLTOBackendThread(
1081 AddStreamFn AddStream, NativeObjectCache Cache, unsigned Task,
1082 BitcodeModule BM, ModuleSummaryIndex &CombinedIndex,
1083 const FunctionImporter::ImportMapTy &ImportList,
1084 const FunctionImporter::ExportSetTy &ExportList,
1085 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1086 const GVSummaryMapTy &DefinedGlobals,
1087 MapVector<StringRef, BitcodeModule> &ModuleMap) {
1088 auto RunThinBackend = [&](AddStreamFn AddStream) {
1089 LTOLLVMContext BackendContext(Conf);
1090 Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(BackendContext);
1091 if (!MOrErr)
1092 return MOrErr.takeError();
1093
1094 return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex,
1095 ImportList, DefinedGlobals, ModuleMap);
1096 };
1097
1098 auto ModuleID = BM.getModuleIdentifier();
1099
1100 if (!Cache || !CombinedIndex.modulePaths().count(ModuleID) ||
1101 all_of(CombinedIndex.getModuleHash(ModuleID),
1102 [](uint32_t V) { return V == 0; }))
1103 // Cache disabled or no entry for this module in the combined index or
1104 // no module hash.
1105 return RunThinBackend(AddStream);
1106
1107 SmallString<40> Key;
1108 // The module may be cached, this helps handling it.
1109 computeLTOCacheKey(Key, Conf, CombinedIndex, ModuleID, ImportList,
1110 ExportList, ResolvedODR, DefinedGlobals, CfiFunctionDefs,
1111 CfiFunctionDecls);
1112 if (AddStreamFn CacheAddStream = Cache(Task, Key))
1113 return RunThinBackend(CacheAddStream);
1114
1115 return Error::success();
1116 }
1117
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)1118 Error start(
1119 unsigned Task, BitcodeModule BM,
1120 const FunctionImporter::ImportMapTy &ImportList,
1121 const FunctionImporter::ExportSetTy &ExportList,
1122 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1123 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
1124 StringRef ModulePath = BM.getModuleIdentifier();
1125 assert(ModuleToDefinedGVSummaries.count(ModulePath));
1126 const GVSummaryMapTy &DefinedGlobals =
1127 ModuleToDefinedGVSummaries.find(ModulePath)->second;
1128 BackendThreadPool.async(
1129 [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex,
1130 const FunctionImporter::ImportMapTy &ImportList,
1131 const FunctionImporter::ExportSetTy &ExportList,
1132 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>
1133 &ResolvedODR,
1134 const GVSummaryMapTy &DefinedGlobals,
1135 MapVector<StringRef, BitcodeModule> &ModuleMap) {
1136 Error E = runThinLTOBackendThread(
1137 AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList,
1138 ResolvedODR, DefinedGlobals, ModuleMap);
1139 if (E) {
1140 std::unique_lock<std::mutex> L(ErrMu);
1141 if (Err)
1142 Err = joinErrors(std::move(*Err), std::move(E));
1143 else
1144 Err = std::move(E);
1145 }
1146 },
1147 BM, std::ref(CombinedIndex), std::ref(ImportList), std::ref(ExportList),
1148 std::ref(ResolvedODR), std::ref(DefinedGlobals), std::ref(ModuleMap));
1149 return Error::success();
1150 }
1151
wait()1152 Error wait() override {
1153 BackendThreadPool.wait();
1154 if (Err)
1155 return std::move(*Err);
1156 else
1157 return Error::success();
1158 }
1159 };
1160 } // end anonymous namespace
1161
createInProcessThinBackend(unsigned ParallelismLevel)1162 ThinBackend lto::createInProcessThinBackend(unsigned ParallelismLevel) {
1163 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1164 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1165 AddStreamFn AddStream, NativeObjectCache Cache) {
1166 return std::make_unique<InProcessThinBackend>(
1167 Conf, CombinedIndex, ParallelismLevel, ModuleToDefinedGVSummaries,
1168 AddStream, Cache);
1169 };
1170 }
1171
1172 // Given the original \p Path to an output file, replace any path
1173 // prefix matching \p OldPrefix with \p NewPrefix. Also, create the
1174 // resulting directory if it does not yet exist.
getThinLTOOutputFile(const std::string & Path,const std::string & OldPrefix,const std::string & NewPrefix)1175 std::string lto::getThinLTOOutputFile(const std::string &Path,
1176 const std::string &OldPrefix,
1177 const std::string &NewPrefix) {
1178 if (OldPrefix.empty() && NewPrefix.empty())
1179 return Path;
1180 SmallString<128> NewPath(Path);
1181 llvm::sys::path::replace_path_prefix(NewPath, OldPrefix, NewPrefix);
1182 StringRef ParentPath = llvm::sys::path::parent_path(NewPath.str());
1183 if (!ParentPath.empty()) {
1184 // Make sure the new directory exists, creating it if necessary.
1185 if (std::error_code EC = llvm::sys::fs::create_directories(ParentPath))
1186 llvm::errs() << "warning: could not create directory '" << ParentPath
1187 << "': " << EC.message() << '\n';
1188 }
1189 return NewPath.str();
1190 }
1191
1192 namespace {
1193 class WriteIndexesThinBackend : public ThinBackendProc {
1194 std::string OldPrefix, NewPrefix;
1195 bool ShouldEmitImportsFiles;
1196 raw_fd_ostream *LinkedObjectsFile;
1197 lto::IndexWriteCallback OnWrite;
1198
1199 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)1200 WriteIndexesThinBackend(
1201 const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1202 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1203 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1204 raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite)
1205 : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries),
1206 OldPrefix(OldPrefix), NewPrefix(NewPrefix),
1207 ShouldEmitImportsFiles(ShouldEmitImportsFiles),
1208 LinkedObjectsFile(LinkedObjectsFile), OnWrite(OnWrite) {}
1209
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)1210 Error start(
1211 unsigned Task, BitcodeModule BM,
1212 const FunctionImporter::ImportMapTy &ImportList,
1213 const FunctionImporter::ExportSetTy &ExportList,
1214 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
1215 MapVector<StringRef, BitcodeModule> &ModuleMap) override {
1216 StringRef ModulePath = BM.getModuleIdentifier();
1217 std::string NewModulePath =
1218 getThinLTOOutputFile(ModulePath, OldPrefix, NewPrefix);
1219
1220 if (LinkedObjectsFile)
1221 *LinkedObjectsFile << NewModulePath << '\n';
1222
1223 std::map<std::string, GVSummaryMapTy> ModuleToSummariesForIndex;
1224 gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries,
1225 ImportList, ModuleToSummariesForIndex);
1226
1227 std::error_code EC;
1228 raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC,
1229 sys::fs::OpenFlags::OF_None);
1230 if (EC)
1231 return errorCodeToError(EC);
1232 WriteIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex);
1233
1234 if (ShouldEmitImportsFiles) {
1235 EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports",
1236 ModuleToSummariesForIndex);
1237 if (EC)
1238 return errorCodeToError(EC);
1239 }
1240
1241 if (OnWrite)
1242 OnWrite(ModulePath);
1243 return Error::success();
1244 }
1245
wait()1246 Error wait() override { return Error::success(); }
1247 };
1248 } // end anonymous namespace
1249
createWriteIndexesThinBackend(std::string OldPrefix,std::string NewPrefix,bool ShouldEmitImportsFiles,raw_fd_ostream * LinkedObjectsFile,IndexWriteCallback OnWrite)1250 ThinBackend lto::createWriteIndexesThinBackend(
1251 std::string OldPrefix, std::string NewPrefix, bool ShouldEmitImportsFiles,
1252 raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) {
1253 return [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
1254 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
1255 AddStreamFn AddStream, NativeObjectCache Cache) {
1256 return std::make_unique<WriteIndexesThinBackend>(
1257 Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix, NewPrefix,
1258 ShouldEmitImportsFiles, LinkedObjectsFile, OnWrite);
1259 };
1260 }
1261
runThinLTO(AddStreamFn AddStream,NativeObjectCache Cache,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols)1262 Error LTO::runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache,
1263 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
1264 if (ThinLTO.ModuleMap.empty())
1265 return Error::success();
1266
1267 if (Conf.CombinedIndexHook &&
1268 !Conf.CombinedIndexHook(ThinLTO.CombinedIndex, GUIDPreservedSymbols))
1269 return Error::success();
1270
1271 // Collect for each module the list of function it defines (GUID ->
1272 // Summary).
1273 StringMap<GVSummaryMapTy>
1274 ModuleToDefinedGVSummaries(ThinLTO.ModuleMap.size());
1275 ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule(
1276 ModuleToDefinedGVSummaries);
1277 // Create entries for any modules that didn't have any GV summaries
1278 // (either they didn't have any GVs to start with, or we suppressed
1279 // generation of the summaries because they e.g. had inline assembly
1280 // uses that couldn't be promoted/renamed on export). This is so
1281 // InProcessThinBackend::start can still launch a backend thread, which
1282 // is passed the map of summaries for the module, without any special
1283 // handling for this case.
1284 for (auto &Mod : ThinLTO.ModuleMap)
1285 if (!ModuleToDefinedGVSummaries.count(Mod.first))
1286 ModuleToDefinedGVSummaries.try_emplace(Mod.first);
1287
1288 // Synthesize entry counts for functions in the CombinedIndex.
1289 computeSyntheticCounts(ThinLTO.CombinedIndex);
1290
1291 StringMap<FunctionImporter::ImportMapTy> ImportLists(
1292 ThinLTO.ModuleMap.size());
1293 StringMap<FunctionImporter::ExportSetTy> ExportLists(
1294 ThinLTO.ModuleMap.size());
1295 StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
1296
1297 if (DumpThinCGSCCs)
1298 ThinLTO.CombinedIndex.dumpSCCs(outs());
1299
1300 std::set<GlobalValue::GUID> ExportedGUIDs;
1301
1302 // Perform index-based WPD. This will return immediately if there are
1303 // no index entries in the typeIdMetadata map (e.g. if we are instead
1304 // performing IR-based WPD in hybrid regular/thin LTO mode).
1305 std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap;
1306 runWholeProgramDevirtOnIndex(ThinLTO.CombinedIndex, ExportedGUIDs,
1307 LocalWPDTargetsMap);
1308
1309 if (Conf.OptLevel > 0)
1310 ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1311 ImportLists, ExportLists);
1312
1313 // Figure out which symbols need to be internalized. This also needs to happen
1314 // at -O0 because summary-based DCE is implemented using internalization, and
1315 // we must apply DCE consistently with the full LTO module in order to avoid
1316 // undefined references during the final link.
1317 for (auto &Res : GlobalResolutions) {
1318 // If the symbol does not have external references or it is not prevailing,
1319 // then not need to mark it as exported from a ThinLTO partition.
1320 if (Res.second.Partition != GlobalResolution::External ||
1321 !Res.second.isPrevailingIRSymbol())
1322 continue;
1323 auto GUID = GlobalValue::getGUID(
1324 GlobalValue::dropLLVMManglingEscape(Res.second.IRName));
1325 // Mark exported unless index-based analysis determined it to be dead.
1326 if (ThinLTO.CombinedIndex.isGUIDLive(GUID))
1327 ExportedGUIDs.insert(GUID);
1328 }
1329
1330 // Any functions referenced by the jump table in the regular LTO object must
1331 // be exported.
1332 for (auto &Def : ThinLTO.CombinedIndex.cfiFunctionDefs())
1333 ExportedGUIDs.insert(
1334 GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Def)));
1335
1336 auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) {
1337 const auto &ExportList = ExportLists.find(ModuleIdentifier);
1338 return (ExportList != ExportLists.end() && ExportList->second.count(VI)) ||
1339 ExportedGUIDs.count(VI.getGUID());
1340 };
1341
1342 // Update local devirtualized targets that were exported by cross-module
1343 // importing or by other devirtualizations marked in the ExportedGUIDs set.
1344 updateIndexWPDForExports(ThinLTO.CombinedIndex, isExported,
1345 LocalWPDTargetsMap);
1346
1347 auto isPrevailing = [&](GlobalValue::GUID GUID,
1348 const GlobalValueSummary *S) {
1349 return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath();
1350 };
1351 thinLTOInternalizeAndPromoteInIndex(ThinLTO.CombinedIndex, isExported,
1352 isPrevailing);
1353
1354 auto recordNewLinkage = [&](StringRef ModuleIdentifier,
1355 GlobalValue::GUID GUID,
1356 GlobalValue::LinkageTypes NewLinkage) {
1357 ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
1358 };
1359 thinLTOResolvePrevailingInIndex(ThinLTO.CombinedIndex, isPrevailing,
1360 recordNewLinkage, GUIDPreservedSymbols);
1361
1362 std::unique_ptr<ThinBackendProc> BackendProc =
1363 ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
1364 AddStream, Cache);
1365
1366 // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for combined
1367 // module and parallel code generation partitions.
1368 unsigned Task = RegularLTO.ParallelCodeGenParallelismLevel;
1369 for (auto &Mod : ThinLTO.ModuleMap) {
1370 if (Error E = BackendProc->start(Task, Mod.second, ImportLists[Mod.first],
1371 ExportLists[Mod.first],
1372 ResolvedODR[Mod.first], ThinLTO.ModuleMap))
1373 return E;
1374 ++Task;
1375 }
1376
1377 return BackendProc->wait();
1378 }
1379
1380 Expected<std::unique_ptr<ToolOutputFile>>
setupOptimizationRemarks(LLVMContext & Context,StringRef RemarksFilename,StringRef RemarksPasses,StringRef RemarksFormat,bool RemarksWithHotness,int Count)1381 lto::setupOptimizationRemarks(LLVMContext &Context, StringRef RemarksFilename,
1382 StringRef RemarksPasses, StringRef RemarksFormat,
1383 bool RemarksWithHotness, int Count) {
1384 std::string Filename = RemarksFilename;
1385 // For ThinLTO, file.opt.<format> becomes
1386 // file.opt.<format>.thin.<num>.<format>.
1387 if (!Filename.empty() && Count != -1)
1388 Filename =
1389 (Twine(Filename) + ".thin." + llvm::utostr(Count) + "." + RemarksFormat)
1390 .str();
1391
1392 auto ResultOrErr = llvm::setupOptimizationRemarks(
1393 Context, Filename, RemarksPasses, RemarksFormat, RemarksWithHotness);
1394 if (Error E = ResultOrErr.takeError())
1395 return std::move(E);
1396
1397 if (*ResultOrErr)
1398 (*ResultOrErr)->keep();
1399
1400 return ResultOrErr;
1401 }
1402
1403 Expected<std::unique_ptr<ToolOutputFile>>
setupStatsFile(StringRef StatsFilename)1404 lto::setupStatsFile(StringRef StatsFilename) {
1405 // Setup output file to emit statistics.
1406 if (StatsFilename.empty())
1407 return nullptr;
1408
1409 llvm::EnableStatistics(false);
1410 std::error_code EC;
1411 auto StatsFile =
1412 std::make_unique<ToolOutputFile>(StatsFilename, EC, sys::fs::OF_None);
1413 if (EC)
1414 return errorCodeToError(EC);
1415
1416 StatsFile->keep();
1417 return std::move(StatsFile);
1418 }
1419