1 | //===-LTO.cpp - LLVM Link Time Optimizer ----------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements functions and classes used to support LTO. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/LTO/LTO.h" |
14 | #include "llvm/ADT/ScopeExit.h" |
15 | #include "llvm/ADT/SmallSet.h" |
16 | #include "llvm/ADT/StableHashing.h" |
17 | #include "llvm/ADT/Statistic.h" |
18 | #include "llvm/ADT/StringExtras.h" |
19 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
20 | #include "llvm/Analysis/StackSafetyAnalysis.h" |
21 | #include "llvm/Analysis/TargetLibraryInfo.h" |
22 | #include "llvm/Analysis/TargetTransformInfo.h" |
23 | #include "llvm/Bitcode/BitcodeReader.h" |
24 | #include "llvm/Bitcode/BitcodeWriter.h" |
25 | #include "llvm/CGData/CodeGenData.h" |
26 | #include "llvm/CodeGen/Analysis.h" |
27 | #include "llvm/Config/llvm-config.h" |
28 | #include "llvm/IR/AutoUpgrade.h" |
29 | #include "llvm/IR/DiagnosticPrinter.h" |
30 | #include "llvm/IR/Intrinsics.h" |
31 | #include "llvm/IR/LLVMRemarkStreamer.h" |
32 | #include "llvm/IR/LegacyPassManager.h" |
33 | #include "llvm/IR/Mangler.h" |
34 | #include "llvm/IR/Metadata.h" |
35 | #include "llvm/IR/RuntimeLibcalls.h" |
36 | #include "llvm/LTO/LTOBackend.h" |
37 | #include "llvm/Linker/IRMover.h" |
38 | #include "llvm/MC/TargetRegistry.h" |
39 | #include "llvm/Object/IRObjectFile.h" |
40 | #include "llvm/Support/Caching.h" |
41 | #include "llvm/Support/CommandLine.h" |
42 | #include "llvm/Support/Compiler.h" |
43 | #include "llvm/Support/Error.h" |
44 | #include "llvm/Support/FileSystem.h" |
45 | #include "llvm/Support/JSON.h" |
46 | #include "llvm/Support/MemoryBuffer.h" |
47 | #include "llvm/Support/Path.h" |
48 | #include "llvm/Support/Process.h" |
49 | #include "llvm/Support/SHA1.h" |
50 | #include "llvm/Support/SourceMgr.h" |
51 | #include "llvm/Support/ThreadPool.h" |
52 | #include "llvm/Support/Threading.h" |
53 | #include "llvm/Support/TimeProfiler.h" |
54 | #include "llvm/Support/ToolOutputFile.h" |
55 | #include "llvm/Support/VCSRevision.h" |
56 | #include "llvm/Support/raw_ostream.h" |
57 | #include "llvm/Target/TargetOptions.h" |
58 | #include "llvm/Transforms/IPO.h" |
59 | #include "llvm/Transforms/IPO/MemProfContextDisambiguation.h" |
60 | #include "llvm/Transforms/IPO/WholeProgramDevirt.h" |
61 | #include "llvm/Transforms/Utils/FunctionImportUtils.h" |
62 | #include "llvm/Transforms/Utils/SplitModule.h" |
63 | |
64 | #include <optional> |
65 | #include <set> |
66 | |
67 | using namespace llvm; |
68 | using namespace lto; |
69 | using namespace object; |
70 | |
71 | #define DEBUG_TYPE "lto" |
72 | |
73 | static cl::opt<bool> |
74 | DumpThinCGSCCs("dump-thin-cg-sccs" , cl::init(Val: false), cl::Hidden, |
75 | cl::desc("Dump the SCCs in the ThinLTO index's callgraph" )); |
76 | |
77 | extern cl::opt<bool> CodeGenDataThinLTOTwoRounds; |
78 | |
79 | extern cl::opt<bool> ForceImportAll; |
80 | |
81 | namespace llvm { |
82 | /// Enable global value internalization in LTO. |
83 | cl::opt<bool> EnableLTOInternalization( |
84 | "enable-lto-internalization" , cl::init(Val: true), cl::Hidden, |
85 | cl::desc("Enable global value internalization in LTO" )); |
86 | |
87 | static cl::opt<bool> |
88 | LTOKeepSymbolCopies("lto-keep-symbol-copies" , cl::init(Val: false), cl::Hidden, |
89 | cl::desc("Keep copies of symbols in LTO indexing" )); |
90 | |
91 | /// Indicate we are linking with an allocator that supports hot/cold operator |
92 | /// new interfaces. |
93 | extern cl::opt<bool> SupportsHotColdNew; |
94 | |
95 | /// Enable MemProf context disambiguation for thin link. |
96 | extern cl::opt<bool> EnableMemProfContextDisambiguation; |
97 | } // namespace llvm |
98 | |
99 | // Computes a unique hash for the Module considering the current list of |
100 | // export/import and other global analysis results. |
101 | // Returns the hash in its hexadecimal representation. |
102 | std::string llvm::computeLTOCacheKey( |
103 | const Config &Conf, const ModuleSummaryIndex &Index, StringRef ModuleID, |
104 | const FunctionImporter::ImportMapTy &ImportList, |
105 | const FunctionImporter::ExportSetTy &ExportList, |
106 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
107 | const GVSummaryMapTy &DefinedGlobals, |
108 | const DenseSet<GlobalValue::GUID> &CfiFunctionDefs, |
109 | const DenseSet<GlobalValue::GUID> &CfiFunctionDecls) { |
110 | // Compute the unique hash for this entry. |
111 | // This is based on the current compiler version, the module itself, the |
112 | // export list, the hash for every single module in the import list, the |
113 | // list of ResolvedODR for the module, and the list of preserved symbols. |
114 | SHA1 Hasher; |
115 | |
116 | // Start with the compiler revision |
117 | Hasher.update(LLVM_VERSION_STRING); |
118 | #ifdef LLVM_REVISION |
119 | Hasher.update(LLVM_REVISION); |
120 | #endif |
121 | |
122 | // Include the parts of the LTO configuration that affect code generation. |
123 | auto AddString = [&](StringRef Str) { |
124 | Hasher.update(Str); |
125 | Hasher.update(Data: ArrayRef<uint8_t>{0}); |
126 | }; |
127 | auto AddUnsigned = [&](unsigned I) { |
128 | uint8_t Data[4]; |
129 | support::endian::write32le(P: Data, V: I); |
130 | Hasher.update(Data); |
131 | }; |
132 | auto AddUint64 = [&](uint64_t I) { |
133 | uint8_t Data[8]; |
134 | support::endian::write64le(P: Data, V: I); |
135 | Hasher.update(Data); |
136 | }; |
137 | auto AddUint8 = [&](const uint8_t I) { |
138 | Hasher.update(Data: ArrayRef<uint8_t>(&I, 1)); |
139 | }; |
140 | AddString(Conf.CPU); |
141 | // FIXME: Hash more of Options. For now all clients initialize Options from |
142 | // command-line flags (which is unsupported in production), but may set |
143 | // X86RelaxRelocations. The clang driver can also pass FunctionSections, |
144 | // DataSections and DebuggerTuning via command line flags. |
145 | AddUnsigned(Conf.Options.MCOptions.X86RelaxRelocations); |
146 | AddUnsigned(Conf.Options.FunctionSections); |
147 | AddUnsigned(Conf.Options.DataSections); |
148 | AddUnsigned((unsigned)Conf.Options.DebuggerTuning); |
149 | for (auto &A : Conf.MAttrs) |
150 | AddString(A); |
151 | if (Conf.RelocModel) |
152 | AddUnsigned(*Conf.RelocModel); |
153 | else |
154 | AddUnsigned(-1); |
155 | if (Conf.CodeModel) |
156 | AddUnsigned(*Conf.CodeModel); |
157 | else |
158 | AddUnsigned(-1); |
159 | for (const auto &S : Conf.MllvmArgs) |
160 | AddString(S); |
161 | AddUnsigned(static_cast<int>(Conf.CGOptLevel)); |
162 | AddUnsigned(static_cast<int>(Conf.CGFileType)); |
163 | AddUnsigned(Conf.OptLevel); |
164 | AddUnsigned(Conf.Freestanding); |
165 | AddString(Conf.OptPipeline); |
166 | AddString(Conf.AAPipeline); |
167 | AddString(Conf.OverrideTriple); |
168 | AddString(Conf.DefaultTriple); |
169 | AddString(Conf.DwoDir); |
170 | |
171 | // Include the hash for the current module |
172 | auto ModHash = Index.getModuleHash(ModPath: ModuleID); |
173 | Hasher.update(Data: ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); |
174 | |
175 | // TODO: `ExportList` is determined by `ImportList`. Since `ImportList` is |
176 | // used to compute cache key, we could omit hashing `ExportList` here. |
177 | std::vector<uint64_t> ExportsGUID; |
178 | ExportsGUID.reserve(n: ExportList.size()); |
179 | for (const auto &VI : ExportList) |
180 | ExportsGUID.push_back(x: VI.getGUID()); |
181 | |
182 | // Sort the export list elements GUIDs. |
183 | llvm::sort(C&: ExportsGUID); |
184 | for (auto GUID : ExportsGUID) |
185 | Hasher.update(Data: ArrayRef<uint8_t>((uint8_t *)&GUID, sizeof(GUID))); |
186 | |
187 | // Order using module hash, to be both independent of module name and |
188 | // module order. |
189 | auto Comp = [&](const std::pair<StringRef, GlobalValue::GUID> &L, |
190 | const std::pair<StringRef, GlobalValue::GUID> &R) { |
191 | return std::make_pair(x: Index.getModule(ModPath: L.first)->second, y: L.second) < |
192 | std::make_pair(x: Index.getModule(ModPath: R.first)->second, y: R.second); |
193 | }; |
194 | FunctionImporter::SortedImportList SortedImportList(ImportList, Comp); |
195 | |
196 | // Count the number of imports for each source module. |
197 | DenseMap<StringRef, unsigned> ModuleToNumImports; |
198 | for (const auto &[FromModule, GUID, Type] : SortedImportList) |
199 | ++ModuleToNumImports[FromModule]; |
200 | |
201 | std::optional<StringRef> LastModule; |
202 | for (const auto &[FromModule, GUID, Type] : SortedImportList) { |
203 | if (LastModule != FromModule) { |
204 | // Include the hash for every module we import functions from. The set of |
205 | // imported symbols for each module may affect code generation and is |
206 | // sensitive to link order, so include that as well. |
207 | LastModule = FromModule; |
208 | auto ModHash = Index.getModule(ModPath: FromModule)->second; |
209 | Hasher.update(Data: ArrayRef<uint8_t>((uint8_t *)&ModHash[0], sizeof(ModHash))); |
210 | AddUint64(ModuleToNumImports[FromModule]); |
211 | } |
212 | AddUint64(GUID); |
213 | AddUint8(Type); |
214 | } |
215 | |
216 | // Include the hash for the resolved ODR. |
217 | for (auto &Entry : ResolvedODR) { |
218 | Hasher.update(Data: ArrayRef<uint8_t>((const uint8_t *)&Entry.first, |
219 | sizeof(GlobalValue::GUID))); |
220 | Hasher.update(Data: ArrayRef<uint8_t>((const uint8_t *)&Entry.second, |
221 | sizeof(GlobalValue::LinkageTypes))); |
222 | } |
223 | |
224 | // Members of CfiFunctionDefs and CfiFunctionDecls that are referenced or |
225 | // defined in this module. |
226 | std::set<GlobalValue::GUID> UsedCfiDefs; |
227 | std::set<GlobalValue::GUID> UsedCfiDecls; |
228 | |
229 | // Typeids used in this module. |
230 | std::set<GlobalValue::GUID> UsedTypeIds; |
231 | |
232 | auto AddUsedCfiGlobal = [&](GlobalValue::GUID ValueGUID) { |
233 | if (CfiFunctionDefs.contains(V: ValueGUID)) |
234 | UsedCfiDefs.insert(x: ValueGUID); |
235 | if (CfiFunctionDecls.contains(V: ValueGUID)) |
236 | UsedCfiDecls.insert(x: ValueGUID); |
237 | }; |
238 | |
239 | auto AddUsedThings = [&](GlobalValueSummary *GS) { |
240 | if (!GS) return; |
241 | AddUnsigned(GS->getVisibility()); |
242 | AddUnsigned(GS->isLive()); |
243 | AddUnsigned(GS->canAutoHide()); |
244 | for (const ValueInfo &VI : GS->refs()) { |
245 | AddUnsigned(VI.isDSOLocal(WithDSOLocalPropagation: Index.withDSOLocalPropagation())); |
246 | AddUsedCfiGlobal(VI.getGUID()); |
247 | } |
248 | if (auto *GVS = dyn_cast<GlobalVarSummary>(Val: GS)) { |
249 | AddUnsigned(GVS->maybeReadOnly()); |
250 | AddUnsigned(GVS->maybeWriteOnly()); |
251 | } |
252 | if (auto *FS = dyn_cast<FunctionSummary>(Val: GS)) { |
253 | for (auto &TT : FS->type_tests()) |
254 | UsedTypeIds.insert(x: TT); |
255 | for (auto &TT : FS->type_test_assume_vcalls()) |
256 | UsedTypeIds.insert(x: TT.GUID); |
257 | for (auto &TT : FS->type_checked_load_vcalls()) |
258 | UsedTypeIds.insert(x: TT.GUID); |
259 | for (auto &TT : FS->type_test_assume_const_vcalls()) |
260 | UsedTypeIds.insert(x: TT.VFunc.GUID); |
261 | for (auto &TT : FS->type_checked_load_const_vcalls()) |
262 | UsedTypeIds.insert(x: TT.VFunc.GUID); |
263 | for (auto &ET : FS->calls()) { |
264 | AddUnsigned(ET.first.isDSOLocal(WithDSOLocalPropagation: Index.withDSOLocalPropagation())); |
265 | AddUsedCfiGlobal(ET.first.getGUID()); |
266 | } |
267 | } |
268 | }; |
269 | |
270 | // Include the hash for the linkage type to reflect internalization and weak |
271 | // resolution, and collect any used type identifier resolutions. |
272 | for (auto &GS : DefinedGlobals) { |
273 | GlobalValue::LinkageTypes Linkage = GS.second->linkage(); |
274 | Hasher.update( |
275 | Data: ArrayRef<uint8_t>((const uint8_t *)&Linkage, sizeof(Linkage))); |
276 | AddUsedCfiGlobal(GS.first); |
277 | AddUsedThings(GS.second); |
278 | } |
279 | |
280 | // Imported functions may introduce new uses of type identifier resolutions, |
281 | // so we need to collect their used resolutions as well. |
282 | for (const auto &[FromModule, GUID, Type] : SortedImportList) { |
283 | GlobalValueSummary *S = Index.findSummaryInModule(ValueGUID: GUID, ModuleId: FromModule); |
284 | AddUsedThings(S); |
285 | // If this is an alias, we also care about any types/etc. that the aliasee |
286 | // may reference. |
287 | if (auto *AS = dyn_cast_or_null<AliasSummary>(Val: S)) |
288 | AddUsedThings(AS->getBaseObject()); |
289 | } |
290 | |
291 | auto AddTypeIdSummary = [&](StringRef TId, const TypeIdSummary &S) { |
292 | AddString(TId); |
293 | |
294 | AddUnsigned(S.TTRes.TheKind); |
295 | AddUnsigned(S.TTRes.SizeM1BitWidth); |
296 | |
297 | AddUint64(S.TTRes.AlignLog2); |
298 | AddUint64(S.TTRes.SizeM1); |
299 | AddUint64(S.TTRes.BitMask); |
300 | AddUint64(S.TTRes.InlineBits); |
301 | |
302 | AddUint64(S.WPDRes.size()); |
303 | for (auto &WPD : S.WPDRes) { |
304 | AddUnsigned(WPD.first); |
305 | AddUnsigned(WPD.second.TheKind); |
306 | AddString(WPD.second.SingleImplName); |
307 | |
308 | AddUint64(WPD.second.ResByArg.size()); |
309 | for (auto &ByArg : WPD.second.ResByArg) { |
310 | AddUint64(ByArg.first.size()); |
311 | for (uint64_t Arg : ByArg.first) |
312 | AddUint64(Arg); |
313 | AddUnsigned(ByArg.second.TheKind); |
314 | AddUint64(ByArg.second.Info); |
315 | AddUnsigned(ByArg.second.Byte); |
316 | AddUnsigned(ByArg.second.Bit); |
317 | } |
318 | } |
319 | }; |
320 | |
321 | // Include the hash for all type identifiers used by this module. |
322 | for (GlobalValue::GUID TId : UsedTypeIds) { |
323 | auto TidIter = Index.typeIds().equal_range(x: TId); |
324 | for (const auto &I : make_range(p: TidIter)) |
325 | AddTypeIdSummary(I.second.first, I.second.second); |
326 | } |
327 | |
328 | AddUnsigned(UsedCfiDefs.size()); |
329 | for (auto &V : UsedCfiDefs) |
330 | AddUint64(V); |
331 | |
332 | AddUnsigned(UsedCfiDecls.size()); |
333 | for (auto &V : UsedCfiDecls) |
334 | AddUint64(V); |
335 | |
336 | if (!Conf.SampleProfile.empty()) { |
337 | auto FileOrErr = MemoryBuffer::getFile(Filename: Conf.SampleProfile); |
338 | if (FileOrErr) { |
339 | Hasher.update(Str: FileOrErr.get()->getBuffer()); |
340 | |
341 | if (!Conf.ProfileRemapping.empty()) { |
342 | FileOrErr = MemoryBuffer::getFile(Filename: Conf.ProfileRemapping); |
343 | if (FileOrErr) |
344 | Hasher.update(Str: FileOrErr.get()->getBuffer()); |
345 | } |
346 | } |
347 | } |
348 | |
349 | return toHex(Input: Hasher.result()); |
350 | } |
351 | |
352 | std::string llvm::recomputeLTOCacheKey(const std::string &Key, |
353 | StringRef ) { |
354 | SHA1 Hasher; |
355 | |
356 | auto AddString = [&](StringRef Str) { |
357 | Hasher.update(Str); |
358 | Hasher.update(Data: ArrayRef<uint8_t>{0}); |
359 | }; |
360 | AddString(Key); |
361 | AddString(ExtraID); |
362 | |
363 | return toHex(Input: Hasher.result()); |
364 | } |
365 | |
366 | static void thinLTOResolvePrevailingGUID( |
367 | const Config &C, ValueInfo VI, |
368 | DenseSet<GlobalValueSummary *> &GlobalInvolvedWithAlias, |
369 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
370 | isPrevailing, |
371 | function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> |
372 | recordNewLinkage, |
373 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { |
374 | GlobalValue::VisibilityTypes Visibility = |
375 | C.VisibilityScheme == Config::ELF ? VI.getELFVisibility() |
376 | : GlobalValue::DefaultVisibility; |
377 | for (auto &S : VI.getSummaryList()) { |
378 | GlobalValue::LinkageTypes OriginalLinkage = S->linkage(); |
379 | // Ignore local and appending linkage values since the linker |
380 | // doesn't resolve them. |
381 | if (GlobalValue::isLocalLinkage(Linkage: OriginalLinkage) || |
382 | GlobalValue::isAppendingLinkage(Linkage: S->linkage())) |
383 | continue; |
384 | // We need to emit only one of these. The prevailing module will keep it, |
385 | // but turned into a weak, while the others will drop it when possible. |
386 | // This is both a compile-time optimization and a correctness |
387 | // transformation. This is necessary for correctness when we have exported |
388 | // a reference - we need to convert the linkonce to weak to |
389 | // ensure a copy is kept to satisfy the exported reference. |
390 | // FIXME: We may want to split the compile time and correctness |
391 | // aspects into separate routines. |
392 | if (isPrevailing(VI.getGUID(), S.get())) { |
393 | if (GlobalValue::isLinkOnceLinkage(Linkage: OriginalLinkage)) { |
394 | S->setLinkage(GlobalValue::getWeakLinkage( |
395 | ODR: GlobalValue::isLinkOnceODRLinkage(Linkage: OriginalLinkage))); |
396 | // The kept copy is eligible for auto-hiding (hidden visibility) if all |
397 | // copies were (i.e. they were all linkonce_odr global unnamed addr). |
398 | // If any copy is not (e.g. it was originally weak_odr), then the symbol |
399 | // must remain externally available (e.g. a weak_odr from an explicitly |
400 | // instantiated template). Additionally, if it is in the |
401 | // GUIDPreservedSymbols set, that means that it is visibile outside |
402 | // the summary (e.g. in a native object or a bitcode file without |
403 | // summary), and in that case we cannot hide it as it isn't possible to |
404 | // check all copies. |
405 | S->setCanAutoHide(VI.canAutoHide() && |
406 | !GUIDPreservedSymbols.count(V: VI.getGUID())); |
407 | } |
408 | if (C.VisibilityScheme == Config::FromPrevailing) |
409 | Visibility = S->getVisibility(); |
410 | } |
411 | // Alias and aliasee can't be turned into available_externally. |
412 | // When force-import-all is used, it indicates that object linking is not |
413 | // supported by the target. In this case, we can't change the linkage as |
414 | // well in case the global is converted to declaration. |
415 | else if (!isa<AliasSummary>(Val: S.get()) && |
416 | !GlobalInvolvedWithAlias.count(V: S.get()) && !ForceImportAll) |
417 | S->setLinkage(GlobalValue::AvailableExternallyLinkage); |
418 | |
419 | // For ELF, set visibility to the computed visibility from summaries. We |
420 | // don't track visibility from declarations so this may be more relaxed than |
421 | // the most constraining one. |
422 | if (C.VisibilityScheme == Config::ELF) |
423 | S->setVisibility(Visibility); |
424 | |
425 | if (S->linkage() != OriginalLinkage) |
426 | recordNewLinkage(S->modulePath(), VI.getGUID(), S->linkage()); |
427 | } |
428 | |
429 | if (C.VisibilityScheme == Config::FromPrevailing) { |
430 | for (auto &S : VI.getSummaryList()) { |
431 | GlobalValue::LinkageTypes OriginalLinkage = S->linkage(); |
432 | if (GlobalValue::isLocalLinkage(Linkage: OriginalLinkage) || |
433 | GlobalValue::isAppendingLinkage(Linkage: S->linkage())) |
434 | continue; |
435 | S->setVisibility(Visibility); |
436 | } |
437 | } |
438 | } |
439 | |
440 | /// Resolve linkage for prevailing symbols in the \p Index. |
441 | // |
442 | // We'd like to drop these functions if they are no longer referenced in the |
443 | // current module. However there is a chance that another module is still |
444 | // referencing them because of the import. We make sure we always emit at least |
445 | // one copy. |
446 | void llvm::thinLTOResolvePrevailingInIndex( |
447 | const Config &C, ModuleSummaryIndex &Index, |
448 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
449 | isPrevailing, |
450 | function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> |
451 | recordNewLinkage, |
452 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { |
453 | // We won't optimize the globals that are referenced by an alias for now |
454 | // Ideally we should turn the alias into a global and duplicate the definition |
455 | // when needed. |
456 | DenseSet<GlobalValueSummary *> GlobalInvolvedWithAlias; |
457 | for (auto &I : Index) |
458 | for (auto &S : I.second.SummaryList) |
459 | if (auto AS = dyn_cast<AliasSummary>(Val: S.get())) |
460 | GlobalInvolvedWithAlias.insert(V: &AS->getAliasee()); |
461 | |
462 | for (auto &I : Index) |
463 | thinLTOResolvePrevailingGUID(C, VI: Index.getValueInfo(R: I), |
464 | GlobalInvolvedWithAlias, isPrevailing, |
465 | recordNewLinkage, GUIDPreservedSymbols); |
466 | } |
467 | |
468 | static void thinLTOInternalizeAndPromoteGUID( |
469 | ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported, |
470 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
471 | isPrevailing) { |
472 | auto ExternallyVisibleCopies = |
473 | llvm::count_if(Range: VI.getSummaryList(), |
474 | P: [](const std::unique_ptr<GlobalValueSummary> &Summary) { |
475 | return !GlobalValue::isLocalLinkage(Linkage: Summary->linkage()); |
476 | }); |
477 | |
478 | for (auto &S : VI.getSummaryList()) { |
479 | // First see if we need to promote an internal value because it is not |
480 | // exported. |
481 | if (isExported(S->modulePath(), VI)) { |
482 | if (GlobalValue::isLocalLinkage(Linkage: S->linkage())) |
483 | S->setLinkage(GlobalValue::ExternalLinkage); |
484 | continue; |
485 | } |
486 | |
487 | // Otherwise, see if we can internalize. |
488 | if (!EnableLTOInternalization) |
489 | continue; |
490 | |
491 | // Non-exported values with external linkage can be internalized. |
492 | if (GlobalValue::isExternalLinkage(Linkage: S->linkage())) { |
493 | S->setLinkage(GlobalValue::InternalLinkage); |
494 | continue; |
495 | } |
496 | |
497 | // Non-exported function and variable definitions with a weak-for-linker |
498 | // linkage can be internalized in certain cases. The minimum legality |
499 | // requirements would be that they are not address taken to ensure that we |
500 | // don't break pointer equality checks, and that variables are either read- |
501 | // or write-only. For functions, this is the case if either all copies are |
502 | // [local_]unnamed_addr, or we can propagate reference edge attributes |
503 | // (which is how this is guaranteed for variables, when analyzing whether |
504 | // they are read or write-only). |
505 | // |
506 | // However, we only get to this code for weak-for-linkage values in one of |
507 | // two cases: |
508 | // 1) The prevailing copy is not in IR (it is in native code). |
509 | // 2) The prevailing copy in IR is not exported from its module. |
510 | // Additionally, at least for the new LTO API, case 2 will only happen if |
511 | // there is exactly one definition of the value (i.e. in exactly one |
512 | // module), as duplicate defs are result in the value being marked exported. |
513 | // Likely, users of the legacy LTO API are similar, however, currently there |
514 | // are llvm-lto based tests of the legacy LTO API that do not mark |
515 | // duplicate linkonce_odr copies as exported via the tool, so we need |
516 | // to handle that case below by checking the number of copies. |
517 | // |
518 | // Generally, we only want to internalize a weak-for-linker value in case |
519 | // 2, because in case 1 we cannot see how the value is used to know if it |
520 | // is read or write-only. We also don't want to bloat the binary with |
521 | // multiple internalized copies of non-prevailing linkonce/weak functions. |
522 | // Note if we don't internalize, we will convert non-prevailing copies to |
523 | // available_externally anyway, so that we drop them after inlining. The |
524 | // only reason to internalize such a function is if we indeed have a single |
525 | // copy, because internalizing it won't increase binary size, and enables |
526 | // use of inliner heuristics that are more aggressive in the face of a |
527 | // single call to a static (local). For variables, internalizing a read or |
528 | // write only variable can enable more aggressive optimization. However, we |
529 | // already perform this elsewhere in the ThinLTO backend handling for |
530 | // read or write-only variables (processGlobalForThinLTO). |
531 | // |
532 | // Therefore, only internalize linkonce/weak if there is a single copy, that |
533 | // is prevailing in this IR module. We can do so aggressively, without |
534 | // requiring the address to be insignificant, or that a variable be read or |
535 | // write-only. |
536 | if (!GlobalValue::isWeakForLinker(Linkage: S->linkage()) || |
537 | GlobalValue::isExternalWeakLinkage(Linkage: S->linkage())) |
538 | continue; |
539 | |
540 | if (isPrevailing(VI.getGUID(), S.get()) && ExternallyVisibleCopies == 1) |
541 | S->setLinkage(GlobalValue::InternalLinkage); |
542 | } |
543 | } |
544 | |
545 | // Update the linkages in the given \p Index to mark exported values |
546 | // as external and non-exported values as internal. |
547 | void llvm::thinLTOInternalizeAndPromoteInIndex( |
548 | ModuleSummaryIndex &Index, |
549 | function_ref<bool(StringRef, ValueInfo)> isExported, |
550 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
551 | isPrevailing) { |
552 | for (auto &I : Index) |
553 | thinLTOInternalizeAndPromoteGUID(VI: Index.getValueInfo(R: I), isExported, |
554 | isPrevailing); |
555 | } |
556 | |
557 | // Requires a destructor for std::vector<InputModule>. |
558 | InputFile::~InputFile() = default; |
559 | |
560 | Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) { |
561 | std::unique_ptr<InputFile> File(new InputFile); |
562 | |
563 | Expected<IRSymtabFile> FOrErr = readIRSymtab(MBRef: Object); |
564 | if (!FOrErr) |
565 | return FOrErr.takeError(); |
566 | |
567 | File->TargetTriple = FOrErr->TheReader.getTargetTriple(); |
568 | File->SourceFileName = FOrErr->TheReader.getSourceFileName(); |
569 | File->COFFLinkerOpts = FOrErr->TheReader.getCOFFLinkerOpts(); |
570 | File->DependentLibraries = FOrErr->TheReader.getDependentLibraries(); |
571 | File->ComdatTable = FOrErr->TheReader.getComdatTable(); |
572 | |
573 | for (unsigned I = 0; I != FOrErr->Mods.size(); ++I) { |
574 | size_t Begin = File->Symbols.size(); |
575 | for (const irsymtab::Reader::SymbolRef &Sym : |
576 | FOrErr->TheReader.module_symbols(I)) |
577 | // Skip symbols that are irrelevant to LTO. Note that this condition needs |
578 | // to match the one in Skip() in LTO::addRegularLTO(). |
579 | if (Sym.isGlobal() && !Sym.isFormatSpecific()) |
580 | File->Symbols.push_back(x: Sym); |
581 | File->ModuleSymIndices.push_back(x: {Begin, File->Symbols.size()}); |
582 | } |
583 | |
584 | File->Mods = FOrErr->Mods; |
585 | File->Strtab = std::move(FOrErr->Strtab); |
586 | return std::move(File); |
587 | } |
588 | |
589 | StringRef InputFile::getName() const { |
590 | return Mods[0].getModuleIdentifier(); |
591 | } |
592 | |
593 | BitcodeModule &InputFile::getSingleBitcodeModule() { |
594 | assert(Mods.size() == 1 && "Expect only one bitcode module" ); |
595 | return Mods[0]; |
596 | } |
597 | |
598 | LTO::RegularLTOState::RegularLTOState(unsigned ParallelCodeGenParallelismLevel, |
599 | const Config &Conf) |
600 | : ParallelCodeGenParallelismLevel(ParallelCodeGenParallelismLevel), |
601 | Ctx(Conf), CombinedModule(std::make_unique<Module>(args: "ld-temp.o" , args&: Ctx)), |
602 | Mover(std::make_unique<IRMover>(args&: *CombinedModule)) {} |
603 | |
604 | LTO::ThinLTOState::ThinLTOState(ThinBackend BackendParam) |
605 | : Backend(std::move(BackendParam)), CombinedIndex(/*HaveGVs*/ false) { |
606 | if (!Backend.isValid()) |
607 | Backend = |
608 | createInProcessThinBackend(Parallelism: llvm::heavyweight_hardware_concurrency()); |
609 | } |
610 | |
611 | LTO::LTO(Config Conf, ThinBackend Backend, |
612 | unsigned ParallelCodeGenParallelismLevel, LTOKind LTOMode) |
613 | : Conf(std::move(Conf)), |
614 | RegularLTO(ParallelCodeGenParallelismLevel, this->Conf), |
615 | ThinLTO(std::move(Backend)), |
616 | GlobalResolutions( |
617 | std::make_unique<DenseMap<StringRef, GlobalResolution>>()), |
618 | LTOMode(LTOMode) { |
619 | if (Conf.KeepSymbolNameCopies || LTOKeepSymbolCopies) { |
620 | Alloc = std::make_unique<BumpPtrAllocator>(); |
621 | GlobalResolutionSymbolSaver = std::make_unique<llvm::StringSaver>(args&: *Alloc); |
622 | } |
623 | } |
624 | |
625 | // Requires a destructor for MapVector<BitcodeModule>. |
626 | LTO::~LTO() = default; |
627 | |
628 | // Add the symbols in the given module to the GlobalResolutions map, and resolve |
629 | // their partitions. |
630 | void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms, |
631 | ArrayRef<SymbolResolution> Res, |
632 | unsigned Partition, bool InSummary) { |
633 | auto *ResI = Res.begin(); |
634 | auto *ResE = Res.end(); |
635 | (void)ResE; |
636 | for (const InputFile::Symbol &Sym : Syms) { |
637 | assert(ResI != ResE); |
638 | SymbolResolution Res = *ResI++; |
639 | |
640 | StringRef SymbolName = Sym.getName(); |
641 | // Keep copies of symbols if the client of LTO says so. |
642 | if (GlobalResolutionSymbolSaver && !GlobalResolutions->contains(Val: SymbolName)) |
643 | SymbolName = GlobalResolutionSymbolSaver->save(S: SymbolName); |
644 | |
645 | auto &GlobalRes = (*GlobalResolutions)[SymbolName]; |
646 | GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr(); |
647 | if (Res.Prevailing) { |
648 | assert(!GlobalRes.Prevailing && |
649 | "Multiple prevailing defs are not allowed" ); |
650 | GlobalRes.Prevailing = true; |
651 | GlobalRes.IRName = std::string(Sym.getIRName()); |
652 | } else if (!GlobalRes.Prevailing && GlobalRes.IRName.empty()) { |
653 | // Sometimes it can be two copies of symbol in a module and prevailing |
654 | // symbol can have no IR name. That might happen if symbol is defined in |
655 | // module level inline asm block. In case we have multiple modules with |
656 | // the same symbol we want to use IR name of the prevailing symbol. |
657 | // Otherwise, if we haven't seen a prevailing symbol, set the name so that |
658 | // we can later use it to check if there is any prevailing copy in IR. |
659 | GlobalRes.IRName = std::string(Sym.getIRName()); |
660 | } |
661 | |
662 | // In rare occasion, the symbol used to initialize GlobalRes has a different |
663 | // IRName from the inspected Symbol. This can happen on macOS + iOS, when a |
664 | // symbol is referenced through its mangled name, say @"\01_symbol" while |
665 | // the IRName is @symbol (the prefix underscore comes from MachO mangling). |
666 | // In that case, we have the same actual Symbol that can get two different |
667 | // GUID, leading to some invalid internalization. Workaround this by marking |
668 | // the GlobalRes external. |
669 | |
670 | // FIXME: instead of this check, it would be desirable to compute GUIDs |
671 | // based on mangled name, but this requires an access to the Target Triple |
672 | // and would be relatively invasive on the codebase. |
673 | if (GlobalRes.IRName != Sym.getIRName()) { |
674 | GlobalRes.Partition = GlobalResolution::External; |
675 | GlobalRes.VisibleOutsideSummary = true; |
676 | } |
677 | |
678 | // Set the partition to external if we know it is re-defined by the linker |
679 | // with -defsym or -wrap options, used elsewhere, e.g. it is visible to a |
680 | // regular object, is referenced from llvm.compiler.used/llvm.used, or was |
681 | // already recorded as being referenced from a different partition. |
682 | if (Res.LinkerRedefined || Res.VisibleToRegularObj || Sym.isUsed() || |
683 | (GlobalRes.Partition != GlobalResolution::Unknown && |
684 | GlobalRes.Partition != Partition)) { |
685 | GlobalRes.Partition = GlobalResolution::External; |
686 | } else |
687 | // First recorded reference, save the current partition. |
688 | GlobalRes.Partition = Partition; |
689 | |
690 | // Flag as visible outside of summary if visible from a regular object or |
691 | // from a module that does not have a summary. |
692 | GlobalRes.VisibleOutsideSummary |= |
693 | (Res.VisibleToRegularObj || Sym.isUsed() || !InSummary); |
694 | |
695 | GlobalRes.ExportDynamic |= Res.ExportDynamic; |
696 | } |
697 | } |
698 | |
699 | void LTO::releaseGlobalResolutionsMemory() { |
700 | // Release GlobalResolutions dense-map itself. |
701 | GlobalResolutions.reset(); |
702 | // Release the string saver memory. |
703 | GlobalResolutionSymbolSaver.reset(); |
704 | Alloc.reset(); |
705 | } |
706 | |
707 | static void writeToResolutionFile(raw_ostream &OS, InputFile *Input, |
708 | ArrayRef<SymbolResolution> Res) { |
709 | StringRef Path = Input->getName(); |
710 | OS << Path << '\n'; |
711 | auto ResI = Res.begin(); |
712 | for (const InputFile::Symbol &Sym : Input->symbols()) { |
713 | assert(ResI != Res.end()); |
714 | SymbolResolution Res = *ResI++; |
715 | |
716 | OS << "-r=" << Path << ',' << Sym.getName() << ','; |
717 | if (Res.Prevailing) |
718 | OS << 'p'; |
719 | if (Res.FinalDefinitionInLinkageUnit) |
720 | OS << 'l'; |
721 | if (Res.VisibleToRegularObj) |
722 | OS << 'x'; |
723 | if (Res.LinkerRedefined) |
724 | OS << 'r'; |
725 | OS << '\n'; |
726 | } |
727 | OS.flush(); |
728 | assert(ResI == Res.end()); |
729 | } |
730 | |
731 | Error LTO::add(std::unique_ptr<InputFile> Input, |
732 | ArrayRef<SymbolResolution> Res) { |
733 | assert(!CalledGetMaxTasks); |
734 | |
735 | if (Conf.ResolutionFile) |
736 | writeToResolutionFile(OS&: *Conf.ResolutionFile, Input: Input.get(), Res); |
737 | |
738 | if (RegularLTO.CombinedModule->getTargetTriple().empty()) { |
739 | Triple InputTriple(Input->getTargetTriple()); |
740 | RegularLTO.CombinedModule->setTargetTriple(InputTriple); |
741 | if (InputTriple.isOSBinFormatELF()) |
742 | Conf.VisibilityScheme = Config::ELF; |
743 | } |
744 | |
745 | const SymbolResolution *ResI = Res.begin(); |
746 | for (unsigned I = 0; I != Input->Mods.size(); ++I) |
747 | if (Error Err = addModule(Input&: *Input, ModI: I, ResI, ResE: Res.end())) |
748 | return Err; |
749 | |
750 | assert(ResI == Res.end()); |
751 | return Error::success(); |
752 | } |
753 | |
754 | Error LTO::addModule(InputFile &Input, unsigned ModI, |
755 | const SymbolResolution *&ResI, |
756 | const SymbolResolution *ResE) { |
757 | Expected<BitcodeLTOInfo> LTOInfo = Input.Mods[ModI].getLTOInfo(); |
758 | if (!LTOInfo) |
759 | return LTOInfo.takeError(); |
760 | |
761 | if (EnableSplitLTOUnit) { |
762 | // If only some modules were split, flag this in the index so that |
763 | // we can skip or error on optimizations that need consistently split |
764 | // modules (whole program devirt and lower type tests). |
765 | if (*EnableSplitLTOUnit != LTOInfo->EnableSplitLTOUnit) |
766 | ThinLTO.CombinedIndex.setPartiallySplitLTOUnits(); |
767 | } else |
768 | EnableSplitLTOUnit = LTOInfo->EnableSplitLTOUnit; |
769 | |
770 | BitcodeModule BM = Input.Mods[ModI]; |
771 | |
772 | if ((LTOMode == LTOK_UnifiedRegular || LTOMode == LTOK_UnifiedThin) && |
773 | !LTOInfo->UnifiedLTO) |
774 | return make_error<StringError>( |
775 | Args: "unified LTO compilation must use " |
776 | "compatible bitcode modules (use -funified-lto)" , |
777 | Args: inconvertibleErrorCode()); |
778 | |
779 | if (LTOInfo->UnifiedLTO && LTOMode == LTOK_Default) |
780 | LTOMode = LTOK_UnifiedThin; |
781 | |
782 | bool IsThinLTO = LTOInfo->IsThinLTO && (LTOMode != LTOK_UnifiedRegular); |
783 | |
784 | auto ModSyms = Input.module_symbols(I: ModI); |
785 | addModuleToGlobalRes(Syms: ModSyms, Res: {ResI, ResE}, |
786 | Partition: IsThinLTO ? ThinLTO.ModuleMap.size() + 1 : 0, |
787 | InSummary: LTOInfo->HasSummary); |
788 | |
789 | if (IsThinLTO) |
790 | return addThinLTO(BM, Syms: ModSyms, ResI, ResE); |
791 | |
792 | RegularLTO.EmptyCombinedModule = false; |
793 | Expected<RegularLTOState::AddedModule> ModOrErr = |
794 | addRegularLTO(BM, Syms: ModSyms, ResI, ResE); |
795 | if (!ModOrErr) |
796 | return ModOrErr.takeError(); |
797 | |
798 | if (!LTOInfo->HasSummary) |
799 | return linkRegularLTO(Mod: std::move(*ModOrErr), /*LivenessFromIndex=*/false); |
800 | |
801 | // Regular LTO module summaries are added to a dummy module that represents |
802 | // the combined regular LTO module. |
803 | if (Error Err = BM.readSummary(CombinedIndex&: ThinLTO.CombinedIndex, ModulePath: "" )) |
804 | return Err; |
805 | RegularLTO.ModsWithSummaries.push_back(x: std::move(*ModOrErr)); |
806 | return Error::success(); |
807 | } |
808 | |
809 | // Checks whether the given global value is in a non-prevailing comdat |
810 | // (comdat containing values the linker indicated were not prevailing, |
811 | // which we then dropped to available_externally), and if so, removes |
812 | // it from the comdat. This is called for all global values to ensure the |
813 | // comdat is empty rather than leaving an incomplete comdat. It is needed for |
814 | // regular LTO modules, in case we are in a mixed-LTO mode (both regular |
815 | // and thin LTO modules) compilation. Since the regular LTO module will be |
816 | // linked first in the final native link, we want to make sure the linker |
817 | // doesn't select any of these incomplete comdats that would be left |
818 | // in the regular LTO module without this cleanup. |
819 | static void |
820 | handleNonPrevailingComdat(GlobalValue &GV, |
821 | std::set<const Comdat *> &NonPrevailingComdats) { |
822 | Comdat *C = GV.getComdat(); |
823 | if (!C) |
824 | return; |
825 | |
826 | if (!NonPrevailingComdats.count(x: C)) |
827 | return; |
828 | |
829 | // Additionally need to drop all global values from the comdat to |
830 | // available_externally, to satisfy the COMDAT requirement that all members |
831 | // are discarded as a unit. The non-local linkage global values avoid |
832 | // duplicate definition linker errors. |
833 | GV.setLinkage(GlobalValue::AvailableExternallyLinkage); |
834 | |
835 | if (auto GO = dyn_cast<GlobalObject>(Val: &GV)) |
836 | GO->setComdat(nullptr); |
837 | } |
838 | |
839 | // Add a regular LTO object to the link. |
840 | // The resulting module needs to be linked into the combined LTO module with |
841 | // linkRegularLTO. |
842 | Expected<LTO::RegularLTOState::AddedModule> |
843 | LTO::addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, |
844 | const SymbolResolution *&ResI, |
845 | const SymbolResolution *ResE) { |
846 | RegularLTOState::AddedModule Mod; |
847 | Expected<std::unique_ptr<Module>> MOrErr = |
848 | BM.getLazyModule(Context&: RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true, |
849 | /*IsImporting*/ false); |
850 | if (!MOrErr) |
851 | return MOrErr.takeError(); |
852 | Module &M = **MOrErr; |
853 | Mod.M = std::move(*MOrErr); |
854 | |
855 | if (Error Err = M.materializeMetadata()) |
856 | return std::move(Err); |
857 | |
858 | // If cfi.functions is present and we are in regular LTO mode, LowerTypeTests |
859 | // will rename local functions in the merged module as "<function name>.1". |
860 | // This causes linking errors, since other parts of the module expect the |
861 | // original function name. |
862 | if (LTOMode == LTOK_UnifiedRegular) |
863 | if (NamedMDNode *CfiFunctionsMD = M.getNamedMetadata(Name: "cfi.functions" )) |
864 | M.eraseNamedMetadata(NMD: CfiFunctionsMD); |
865 | |
866 | UpgradeDebugInfo(M); |
867 | |
868 | ModuleSymbolTable SymTab; |
869 | SymTab.addModule(M: &M); |
870 | |
871 | for (GlobalVariable &GV : M.globals()) |
872 | if (GV.hasAppendingLinkage()) |
873 | Mod.Keep.push_back(x: &GV); |
874 | |
875 | DenseSet<GlobalObject *> AliasedGlobals; |
876 | for (auto &GA : M.aliases()) |
877 | if (GlobalObject *GO = GA.getAliaseeObject()) |
878 | AliasedGlobals.insert(V: GO); |
879 | |
880 | // In this function we need IR GlobalValues matching the symbols in Syms |
881 | // (which is not backed by a module), so we need to enumerate them in the same |
882 | // order. The symbol enumeration order of a ModuleSymbolTable intentionally |
883 | // matches the order of an irsymtab, but when we read the irsymtab in |
884 | // InputFile::create we omit some symbols that are irrelevant to LTO. The |
885 | // Skip() function skips the same symbols from the module as InputFile does |
886 | // from the symbol table. |
887 | auto MsymI = SymTab.symbols().begin(), MsymE = SymTab.symbols().end(); |
888 | auto Skip = [&]() { |
889 | while (MsymI != MsymE) { |
890 | auto Flags = SymTab.getSymbolFlags(S: *MsymI); |
891 | if ((Flags & object::BasicSymbolRef::SF_Global) && |
892 | !(Flags & object::BasicSymbolRef::SF_FormatSpecific)) |
893 | return; |
894 | ++MsymI; |
895 | } |
896 | }; |
897 | Skip(); |
898 | |
899 | std::set<const Comdat *> NonPrevailingComdats; |
900 | SmallSet<StringRef, 2> NonPrevailingAsmSymbols; |
901 | for (const InputFile::Symbol &Sym : Syms) { |
902 | assert(ResI != ResE); |
903 | SymbolResolution Res = *ResI++; |
904 | |
905 | assert(MsymI != MsymE); |
906 | ModuleSymbolTable::Symbol Msym = *MsymI++; |
907 | Skip(); |
908 | |
909 | if (GlobalValue *GV = dyn_cast_if_present<GlobalValue *>(Val&: Msym)) { |
910 | if (Res.Prevailing) { |
911 | if (Sym.isUndefined()) |
912 | continue; |
913 | Mod.Keep.push_back(x: GV); |
914 | // For symbols re-defined with linker -wrap and -defsym options, |
915 | // set the linkage to weak to inhibit IPO. The linkage will be |
916 | // restored by the linker. |
917 | if (Res.LinkerRedefined) |
918 | GV->setLinkage(GlobalValue::WeakAnyLinkage); |
919 | |
920 | GlobalValue::LinkageTypes OriginalLinkage = GV->getLinkage(); |
921 | if (GlobalValue::isLinkOnceLinkage(Linkage: OriginalLinkage)) |
922 | GV->setLinkage(GlobalValue::getWeakLinkage( |
923 | ODR: GlobalValue::isLinkOnceODRLinkage(Linkage: OriginalLinkage))); |
924 | } else if (isa<GlobalObject>(Val: GV) && |
925 | (GV->hasLinkOnceODRLinkage() || GV->hasWeakODRLinkage() || |
926 | GV->hasAvailableExternallyLinkage()) && |
927 | !AliasedGlobals.count(V: cast<GlobalObject>(Val: GV))) { |
928 | // Any of the above three types of linkage indicates that the |
929 | // chosen prevailing symbol will have the same semantics as this copy of |
930 | // the symbol, so we may be able to link it with available_externally |
931 | // linkage. We will decide later whether to do that when we link this |
932 | // module (in linkRegularLTO), based on whether it is undefined. |
933 | Mod.Keep.push_back(x: GV); |
934 | GV->setLinkage(GlobalValue::AvailableExternallyLinkage); |
935 | if (GV->hasComdat()) |
936 | NonPrevailingComdats.insert(x: GV->getComdat()); |
937 | cast<GlobalObject>(Val: GV)->setComdat(nullptr); |
938 | } |
939 | |
940 | // Set the 'local' flag based on the linker resolution for this symbol. |
941 | if (Res.FinalDefinitionInLinkageUnit) { |
942 | GV->setDSOLocal(true); |
943 | if (GV->hasDLLImportStorageClass()) |
944 | GV->setDLLStorageClass(GlobalValue::DLLStorageClassTypes:: |
945 | DefaultStorageClass); |
946 | } |
947 | } else if (auto *AS = |
948 | dyn_cast_if_present<ModuleSymbolTable::AsmSymbol *>(Val&: Msym)) { |
949 | // Collect non-prevailing symbols. |
950 | if (!Res.Prevailing) |
951 | NonPrevailingAsmSymbols.insert(V: AS->first); |
952 | } else { |
953 | llvm_unreachable("unknown symbol type" ); |
954 | } |
955 | |
956 | // Common resolution: collect the maximum size/alignment over all commons. |
957 | // We also record if we see an instance of a common as prevailing, so that |
958 | // if none is prevailing we can ignore it later. |
959 | if (Sym.isCommon()) { |
960 | // FIXME: We should figure out what to do about commons defined by asm. |
961 | // For now they aren't reported correctly by ModuleSymbolTable. |
962 | auto &CommonRes = RegularLTO.Commons[std::string(Sym.getIRName())]; |
963 | CommonRes.Size = std::max(a: CommonRes.Size, b: Sym.getCommonSize()); |
964 | if (uint32_t SymAlignValue = Sym.getCommonAlignment()) { |
965 | CommonRes.Alignment = |
966 | std::max(a: Align(SymAlignValue), b: CommonRes.Alignment); |
967 | } |
968 | CommonRes.Prevailing |= Res.Prevailing; |
969 | } |
970 | } |
971 | |
972 | if (!M.getComdatSymbolTable().empty()) |
973 | for (GlobalValue &GV : M.global_values()) |
974 | handleNonPrevailingComdat(GV, NonPrevailingComdats); |
975 | |
976 | // Prepend ".lto_discard <sym>, <sym>*" directive to each module inline asm |
977 | // block. |
978 | if (!M.getModuleInlineAsm().empty()) { |
979 | std::string NewIA = ".lto_discard" ; |
980 | if (!NonPrevailingAsmSymbols.empty()) { |
981 | // Don't dicard a symbol if there is a live .symver for it. |
982 | ModuleSymbolTable::CollectAsmSymvers( |
983 | M, AsmSymver: [&](StringRef Name, StringRef Alias) { |
984 | if (!NonPrevailingAsmSymbols.count(V: Alias)) |
985 | NonPrevailingAsmSymbols.erase(V: Name); |
986 | }); |
987 | NewIA += " " + llvm::join(R&: NonPrevailingAsmSymbols, Separator: ", " ); |
988 | } |
989 | NewIA += "\n" ; |
990 | M.setModuleInlineAsm(NewIA + M.getModuleInlineAsm()); |
991 | } |
992 | |
993 | assert(MsymI == MsymE); |
994 | return std::move(Mod); |
995 | } |
996 | |
997 | Error LTO::linkRegularLTO(RegularLTOState::AddedModule Mod, |
998 | bool LivenessFromIndex) { |
999 | std::vector<GlobalValue *> Keep; |
1000 | for (GlobalValue *GV : Mod.Keep) { |
1001 | if (LivenessFromIndex && !ThinLTO.CombinedIndex.isGUIDLive(GUID: GV->getGUID())) { |
1002 | if (Function *F = dyn_cast<Function>(Val: GV)) { |
1003 | if (DiagnosticOutputFile) { |
1004 | if (Error Err = F->materialize()) |
1005 | return Err; |
1006 | OptimizationRemarkEmitter ORE(F, nullptr); |
1007 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "deadfunction" , F) |
1008 | << ore::NV("Function" , F) |
1009 | << " not added to the combined module " ); |
1010 | } |
1011 | } |
1012 | continue; |
1013 | } |
1014 | |
1015 | if (!GV->hasAvailableExternallyLinkage()) { |
1016 | Keep.push_back(x: GV); |
1017 | continue; |
1018 | } |
1019 | |
1020 | // Only link available_externally definitions if we don't already have a |
1021 | // definition. |
1022 | GlobalValue *CombinedGV = |
1023 | RegularLTO.CombinedModule->getNamedValue(Name: GV->getName()); |
1024 | if (CombinedGV && !CombinedGV->isDeclaration()) |
1025 | continue; |
1026 | |
1027 | Keep.push_back(x: GV); |
1028 | } |
1029 | |
1030 | return RegularLTO.Mover->move(Src: std::move(Mod.M), ValuesToLink: Keep, AddLazyFor: nullptr, |
1031 | /* IsPerformingImport */ false); |
1032 | } |
1033 | |
1034 | // Add a ThinLTO module to the link. |
1035 | Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, |
1036 | const SymbolResolution *&ResI, |
1037 | const SymbolResolution *ResE) { |
1038 | const SymbolResolution *ResITmp = ResI; |
1039 | for (const InputFile::Symbol &Sym : Syms) { |
1040 | assert(ResITmp != ResE); |
1041 | SymbolResolution Res = *ResITmp++; |
1042 | |
1043 | if (!Sym.getIRName().empty()) { |
1044 | auto GUID = GlobalValue::getGUIDAssumingExternalLinkage( |
1045 | GlobalName: GlobalValue::getGlobalIdentifier(Name: Sym.getIRName(), |
1046 | Linkage: GlobalValue::ExternalLinkage, FileName: "" )); |
1047 | if (Res.Prevailing) |
1048 | ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier(); |
1049 | } |
1050 | } |
1051 | |
1052 | if (Error Err = |
1053 | BM.readSummary(CombinedIndex&: ThinLTO.CombinedIndex, ModulePath: BM.getModuleIdentifier(), |
1054 | IsPrevailing: [&](GlobalValue::GUID GUID) { |
1055 | return ThinLTO.PrevailingModuleForGUID[GUID] == |
1056 | BM.getModuleIdentifier(); |
1057 | })) |
1058 | return Err; |
1059 | LLVM_DEBUG(dbgs() << "Module " << BM.getModuleIdentifier() << "\n" ); |
1060 | |
1061 | for (const InputFile::Symbol &Sym : Syms) { |
1062 | assert(ResI != ResE); |
1063 | SymbolResolution Res = *ResI++; |
1064 | |
1065 | if (!Sym.getIRName().empty()) { |
1066 | auto GUID = GlobalValue::getGUIDAssumingExternalLinkage( |
1067 | GlobalName: GlobalValue::getGlobalIdentifier(Name: Sym.getIRName(), |
1068 | Linkage: GlobalValue::ExternalLinkage, FileName: "" )); |
1069 | if (Res.Prevailing) { |
1070 | assert(ThinLTO.PrevailingModuleForGUID[GUID] == |
1071 | BM.getModuleIdentifier()); |
1072 | |
1073 | // For linker redefined symbols (via --wrap or --defsym) we want to |
1074 | // switch the linkage to `weak` to prevent IPOs from happening. |
1075 | // Find the summary in the module for this very GV and record the new |
1076 | // linkage so that we can switch it when we import the GV. |
1077 | if (Res.LinkerRedefined) |
1078 | if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( |
1079 | ValueGUID: GUID, ModuleId: BM.getModuleIdentifier())) |
1080 | S->setLinkage(GlobalValue::WeakAnyLinkage); |
1081 | } |
1082 | |
1083 | // If the linker resolved the symbol to a local definition then mark it |
1084 | // as local in the summary for the module we are adding. |
1085 | if (Res.FinalDefinitionInLinkageUnit) { |
1086 | if (auto S = ThinLTO.CombinedIndex.findSummaryInModule( |
1087 | ValueGUID: GUID, ModuleId: BM.getModuleIdentifier())) { |
1088 | S->setDSOLocal(true); |
1089 | } |
1090 | } |
1091 | } |
1092 | } |
1093 | |
1094 | if (!ThinLTO.ModuleMap.insert(KV: {BM.getModuleIdentifier(), BM}).second) |
1095 | return make_error<StringError>( |
1096 | Args: "Expected at most one ThinLTO module per bitcode file" , |
1097 | Args: inconvertibleErrorCode()); |
1098 | |
1099 | if (!Conf.ThinLTOModulesToCompile.empty()) { |
1100 | if (!ThinLTO.ModulesToCompile) |
1101 | ThinLTO.ModulesToCompile = ModuleMapType(); |
1102 | // This is a fuzzy name matching where only modules with name containing the |
1103 | // specified switch values are going to be compiled. |
1104 | for (const std::string &Name : Conf.ThinLTOModulesToCompile) { |
1105 | if (BM.getModuleIdentifier().contains(Other: Name)) { |
1106 | ThinLTO.ModulesToCompile->insert(KV: {BM.getModuleIdentifier(), BM}); |
1107 | LLVM_DEBUG(dbgs() << "[ThinLTO] Selecting " << BM.getModuleIdentifier() |
1108 | << " to compile\n" ); |
1109 | } |
1110 | } |
1111 | } |
1112 | |
1113 | return Error::success(); |
1114 | } |
1115 | |
1116 | unsigned LTO::getMaxTasks() const { |
1117 | CalledGetMaxTasks = true; |
1118 | auto ModuleCount = ThinLTO.ModulesToCompile ? ThinLTO.ModulesToCompile->size() |
1119 | : ThinLTO.ModuleMap.size(); |
1120 | return RegularLTO.ParallelCodeGenParallelismLevel + ModuleCount; |
1121 | } |
1122 | |
1123 | // If only some of the modules were split, we cannot correctly handle |
1124 | // code that contains type tests or type checked loads. |
1125 | Error LTO::checkPartiallySplit() { |
1126 | if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits()) |
1127 | return Error::success(); |
1128 | |
1129 | const Module *Combined = RegularLTO.CombinedModule.get(); |
1130 | Function *TypeTestFunc = |
1131 | Intrinsic::getDeclarationIfExists(M: Combined, id: Intrinsic::type_test); |
1132 | Function *TypeCheckedLoadFunc = |
1133 | Intrinsic::getDeclarationIfExists(M: Combined, id: Intrinsic::type_checked_load); |
1134 | Function *TypeCheckedLoadRelativeFunc = Intrinsic::getDeclarationIfExists( |
1135 | M: Combined, id: Intrinsic::type_checked_load_relative); |
1136 | |
1137 | // First check if there are type tests / type checked loads in the |
1138 | // merged regular LTO module IR. |
1139 | if ((TypeTestFunc && !TypeTestFunc->use_empty()) || |
1140 | (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()) || |
1141 | (TypeCheckedLoadRelativeFunc && |
1142 | !TypeCheckedLoadRelativeFunc->use_empty())) |
1143 | return make_error<StringError>( |
1144 | Args: "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)" , |
1145 | Args: inconvertibleErrorCode()); |
1146 | |
1147 | // Otherwise check if there are any recorded in the combined summary from the |
1148 | // ThinLTO modules. |
1149 | for (auto &P : ThinLTO.CombinedIndex) { |
1150 | for (auto &S : P.second.SummaryList) { |
1151 | auto *FS = dyn_cast<FunctionSummary>(Val: S.get()); |
1152 | if (!FS) |
1153 | continue; |
1154 | if (!FS->type_test_assume_vcalls().empty() || |
1155 | !FS->type_checked_load_vcalls().empty() || |
1156 | !FS->type_test_assume_const_vcalls().empty() || |
1157 | !FS->type_checked_load_const_vcalls().empty() || |
1158 | !FS->type_tests().empty()) |
1159 | return make_error<StringError>( |
1160 | Args: "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)" , |
1161 | Args: inconvertibleErrorCode()); |
1162 | } |
1163 | } |
1164 | return Error::success(); |
1165 | } |
1166 | |
1167 | Error LTO::run(AddStreamFn AddStream, FileCache Cache) { |
1168 | // Compute "dead" symbols, we don't want to import/export these! |
1169 | DenseSet<GlobalValue::GUID> GUIDPreservedSymbols; |
1170 | DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions; |
1171 | for (auto &Res : *GlobalResolutions) { |
1172 | // Normally resolution have IR name of symbol. We can do nothing here |
1173 | // otherwise. See comments in GlobalResolution struct for more details. |
1174 | if (Res.second.IRName.empty()) |
1175 | continue; |
1176 | |
1177 | GlobalValue::GUID GUID = GlobalValue::getGUIDAssumingExternalLinkage( |
1178 | GlobalName: GlobalValue::dropLLVMManglingEscape(Name: Res.second.IRName)); |
1179 | |
1180 | if (Res.second.VisibleOutsideSummary && Res.second.Prevailing) |
1181 | GUIDPreservedSymbols.insert(V: GUID); |
1182 | |
1183 | if (Res.second.ExportDynamic) |
1184 | DynamicExportSymbols.insert(V: GUID); |
1185 | |
1186 | GUIDPrevailingResolutions[GUID] = |
1187 | Res.second.Prevailing ? PrevailingType::Yes : PrevailingType::No; |
1188 | } |
1189 | |
1190 | auto isPrevailing = [&](GlobalValue::GUID G) { |
1191 | auto It = GUIDPrevailingResolutions.find(Val: G); |
1192 | if (It == GUIDPrevailingResolutions.end()) |
1193 | return PrevailingType::Unknown; |
1194 | return It->second; |
1195 | }; |
1196 | computeDeadSymbolsWithConstProp(Index&: ThinLTO.CombinedIndex, GUIDPreservedSymbols, |
1197 | isPrevailing, ImportEnabled: Conf.OptLevel > 0); |
1198 | |
1199 | // Setup output file to emit statistics. |
1200 | auto StatsFileOrErr = setupStatsFile(Conf.StatsFile); |
1201 | if (!StatsFileOrErr) |
1202 | return StatsFileOrErr.takeError(); |
1203 | std::unique_ptr<ToolOutputFile> StatsFile = std::move(StatsFileOrErr.get()); |
1204 | |
1205 | // TODO: Ideally this would be controlled automatically by detecting that we |
1206 | // are linking with an allocator that supports these interfaces, rather than |
1207 | // an internal option (which would still be needed for tests, however). For |
1208 | // example, if the library exported a symbol like __malloc_hot_cold the linker |
1209 | // could recognize that and set a flag in the lto::Config. |
1210 | if (SupportsHotColdNew) |
1211 | ThinLTO.CombinedIndex.setWithSupportsHotColdNew(); |
1212 | |
1213 | Error Result = runRegularLTO(AddStream); |
1214 | if (!Result) |
1215 | // This will reset the GlobalResolutions optional once done with it to |
1216 | // reduce peak memory before importing. |
1217 | Result = runThinLTO(AddStream, Cache, GUIDPreservedSymbols); |
1218 | |
1219 | if (StatsFile) |
1220 | PrintStatisticsJSON(OS&: StatsFile->os()); |
1221 | |
1222 | return Result; |
1223 | } |
1224 | |
1225 | void lto::updateMemProfAttributes(Module &Mod, |
1226 | const ModuleSummaryIndex &Index) { |
1227 | if (Index.withSupportsHotColdNew()) |
1228 | return; |
1229 | |
1230 | // The profile matcher applies hotness attributes directly for allocations, |
1231 | // and those will cause us to generate calls to the hot/cold interfaces |
1232 | // unconditionally. If supports-hot-cold-new was not enabled in the LTO |
1233 | // link then assume we don't want these calls (e.g. not linking with |
1234 | // the appropriate library, or otherwise trying to disable this behavior). |
1235 | for (auto &F : Mod) { |
1236 | for (auto &BB : F) { |
1237 | for (auto &I : BB) { |
1238 | auto *CI = dyn_cast<CallBase>(Val: &I); |
1239 | if (!CI) |
1240 | continue; |
1241 | if (CI->hasFnAttr(Kind: "memprof" )) |
1242 | CI->removeFnAttr(Kind: "memprof" ); |
1243 | // Strip off all memprof metadata as it is no longer needed. |
1244 | // Importantly, this avoids the addition of new memprof attributes |
1245 | // after inlining propagation. |
1246 | // TODO: If we support additional types of MemProf metadata beyond hot |
1247 | // and cold, we will need to update the metadata based on the allocator |
1248 | // APIs supported instead of completely stripping all. |
1249 | CI->setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
1250 | CI->setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
1251 | } |
1252 | } |
1253 | } |
1254 | } |
1255 | |
1256 | Error LTO::runRegularLTO(AddStreamFn AddStream) { |
1257 | // Setup optimization remarks. |
1258 | auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks( |
1259 | Context&: RegularLTO.CombinedModule->getContext(), RemarksFilename: Conf.RemarksFilename, |
1260 | RemarksPasses: Conf.RemarksPasses, RemarksFormat: Conf.RemarksFormat, RemarksWithHotness: Conf.RemarksWithHotness, |
1261 | RemarksHotnessThreshold: Conf.RemarksHotnessThreshold); |
1262 | LLVM_DEBUG(dbgs() << "Running regular LTO\n" ); |
1263 | if (!DiagFileOrErr) |
1264 | return DiagFileOrErr.takeError(); |
1265 | DiagnosticOutputFile = std::move(*DiagFileOrErr); |
1266 | |
1267 | // Finalize linking of regular LTO modules containing summaries now that |
1268 | // we have computed liveness information. |
1269 | for (auto &M : RegularLTO.ModsWithSummaries) |
1270 | if (Error Err = linkRegularLTO(Mod: std::move(M), |
1271 | /*LivenessFromIndex=*/true)) |
1272 | return Err; |
1273 | |
1274 | // Ensure we don't have inconsistently split LTO units with type tests. |
1275 | // FIXME: this checks both LTO and ThinLTO. It happens to work as we take |
1276 | // this path both cases but eventually this should be split into two and |
1277 | // do the ThinLTO checks in `runThinLTO`. |
1278 | if (Error Err = checkPartiallySplit()) |
1279 | return Err; |
1280 | |
1281 | // Make sure commons have the right size/alignment: we kept the largest from |
1282 | // all the prevailing when adding the inputs, and we apply it here. |
1283 | const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout(); |
1284 | for (auto &I : RegularLTO.Commons) { |
1285 | if (!I.second.Prevailing) |
1286 | // Don't do anything if no instance of this common was prevailing. |
1287 | continue; |
1288 | GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(Name: I.first); |
1289 | if (OldGV && DL.getTypeAllocSize(Ty: OldGV->getValueType()) == I.second.Size) { |
1290 | // Don't create a new global if the type is already correct, just make |
1291 | // sure the alignment is correct. |
1292 | OldGV->setAlignment(I.second.Alignment); |
1293 | continue; |
1294 | } |
1295 | ArrayType *Ty = |
1296 | ArrayType::get(ElementType: Type::getInt8Ty(C&: RegularLTO.Ctx), NumElements: I.second.Size); |
1297 | auto *GV = new GlobalVariable(*RegularLTO.CombinedModule, Ty, false, |
1298 | GlobalValue::CommonLinkage, |
1299 | ConstantAggregateZero::get(Ty), "" ); |
1300 | GV->setAlignment(I.second.Alignment); |
1301 | if (OldGV) { |
1302 | OldGV->replaceAllUsesWith(V: GV); |
1303 | GV->takeName(V: OldGV); |
1304 | OldGV->eraseFromParent(); |
1305 | } else { |
1306 | GV->setName(I.first); |
1307 | } |
1308 | } |
1309 | |
1310 | updateMemProfAttributes(Mod&: *RegularLTO.CombinedModule, Index: ThinLTO.CombinedIndex); |
1311 | |
1312 | bool WholeProgramVisibilityEnabledInLTO = |
1313 | Conf.HasWholeProgramVisibility && |
1314 | // If validation is enabled, upgrade visibility only when all vtables |
1315 | // have typeinfos. |
1316 | (!Conf.ValidateAllVtablesHaveTypeInfos || Conf.AllVtablesHaveTypeInfos); |
1317 | |
1318 | // This returns true when the name is local or not defined. Locals are |
1319 | // expected to be handled separately. |
1320 | auto IsVisibleToRegularObj = [&](StringRef name) { |
1321 | auto It = GlobalResolutions->find(Val: name); |
1322 | return (It == GlobalResolutions->end() || |
1323 | It->second.VisibleOutsideSummary || !It->second.Prevailing); |
1324 | }; |
1325 | |
1326 | // If allowed, upgrade public vcall visibility metadata to linkage unit |
1327 | // visibility before whole program devirtualization in the optimizer. |
1328 | updateVCallVisibilityInModule( |
1329 | M&: *RegularLTO.CombinedModule, WholeProgramVisibilityEnabledInLTO, |
1330 | DynamicExportSymbols, ValidateAllVtablesHaveTypeInfos: Conf.ValidateAllVtablesHaveTypeInfos, |
1331 | IsVisibleToRegularObj); |
1332 | updatePublicTypeTestCalls(M&: *RegularLTO.CombinedModule, |
1333 | WholeProgramVisibilityEnabledInLTO); |
1334 | |
1335 | if (Conf.PreOptModuleHook && |
1336 | !Conf.PreOptModuleHook(0, *RegularLTO.CombinedModule)) |
1337 | return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile)); |
1338 | |
1339 | if (!Conf.CodeGenOnly) { |
1340 | for (const auto &R : *GlobalResolutions) { |
1341 | GlobalValue *GV = |
1342 | RegularLTO.CombinedModule->getNamedValue(Name: R.second.IRName); |
1343 | if (!R.second.isPrevailingIRSymbol()) |
1344 | continue; |
1345 | if (R.second.Partition != 0 && |
1346 | R.second.Partition != GlobalResolution::External) |
1347 | continue; |
1348 | |
1349 | // Ignore symbols defined in other partitions. |
1350 | // Also skip declarations, which are not allowed to have internal linkage. |
1351 | if (!GV || GV->hasLocalLinkage() || GV->isDeclaration()) |
1352 | continue; |
1353 | |
1354 | // Symbols that are marked DLLImport or DLLExport should not be |
1355 | // internalized, as they are either externally visible or referencing |
1356 | // external symbols. Symbols that have AvailableExternally or Appending |
1357 | // linkage might be used by future passes and should be kept as is. |
1358 | // These linkages are seen in Unified regular LTO, because the process |
1359 | // of creating split LTO units introduces symbols with that linkage into |
1360 | // one of the created modules. Normally, only the ThinLTO backend would |
1361 | // compile this module, but Unified Regular LTO processes both |
1362 | // modules created by the splitting process as regular LTO modules. |
1363 | if ((LTOMode == LTOKind::LTOK_UnifiedRegular) && |
1364 | ((GV->getDLLStorageClass() != GlobalValue::DefaultStorageClass) || |
1365 | GV->hasAvailableExternallyLinkage() || GV->hasAppendingLinkage())) |
1366 | continue; |
1367 | |
1368 | GV->setUnnamedAddr(R.second.UnnamedAddr ? GlobalValue::UnnamedAddr::Global |
1369 | : GlobalValue::UnnamedAddr::None); |
1370 | if (EnableLTOInternalization && R.second.Partition == 0) |
1371 | GV->setLinkage(GlobalValue::InternalLinkage); |
1372 | } |
1373 | |
1374 | if (Conf.PostInternalizeModuleHook && |
1375 | !Conf.PostInternalizeModuleHook(0, *RegularLTO.CombinedModule)) |
1376 | return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile)); |
1377 | } |
1378 | |
1379 | if (!RegularLTO.EmptyCombinedModule || Conf.AlwaysEmitRegularLTOObj) { |
1380 | if (Error Err = |
1381 | backend(C: Conf, AddStream, ParallelCodeGenParallelismLevel: RegularLTO.ParallelCodeGenParallelismLevel, |
1382 | M&: *RegularLTO.CombinedModule, CombinedIndex&: ThinLTO.CombinedIndex)) |
1383 | return Err; |
1384 | } |
1385 | |
1386 | return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile)); |
1387 | } |
1388 | |
1389 | SmallVector<const char *> LTO::getRuntimeLibcallSymbols(const Triple &TT) { |
1390 | RTLIB::RuntimeLibcallsInfo Libcalls(TT); |
1391 | SmallVector<const char *> LibcallSymbols; |
1392 | ArrayRef<RTLIB::LibcallImpl> LibcallImpls = Libcalls.getLibcallImpls(); |
1393 | LibcallSymbols.reserve(N: LibcallImpls.size()); |
1394 | |
1395 | for (RTLIB::LibcallImpl Impl : LibcallImpls) { |
1396 | if (Impl != RTLIB::Unsupported) |
1397 | LibcallSymbols.push_back(Elt: Libcalls.getLibcallImplName(CallImpl: Impl)); |
1398 | } |
1399 | |
1400 | return LibcallSymbols; |
1401 | } |
1402 | |
1403 | Error ThinBackendProc::emitFiles( |
1404 | const FunctionImporter::ImportMapTy &ImportList, llvm::StringRef ModulePath, |
1405 | const std::string &NewModulePath) const { |
1406 | return emitFiles(ImportList, ModulePath, NewModulePath, |
1407 | SummaryPath: NewModulePath + ".thinlto.bc" , |
1408 | /*ImportsFiles=*/std::nullopt); |
1409 | } |
1410 | |
1411 | Error ThinBackendProc::emitFiles( |
1412 | const FunctionImporter::ImportMapTy &ImportList, llvm::StringRef ModulePath, |
1413 | const std::string &NewModulePath, StringRef SummaryPath, |
1414 | std::optional<std::reference_wrapper<ImportsFilesContainer>> ImportsFiles) |
1415 | const { |
1416 | ModuleToSummariesForIndexTy ModuleToSummariesForIndex; |
1417 | GVSummaryPtrSet DeclarationSummaries; |
1418 | |
1419 | std::error_code EC; |
1420 | gatherImportedSummariesForModule(ModulePath, ModuleToDefinedGVSummaries, |
1421 | ImportList, ModuleToSummariesForIndex, |
1422 | DecSummaries&: DeclarationSummaries); |
1423 | |
1424 | raw_fd_ostream OS(SummaryPath, EC, sys::fs::OpenFlags::OF_None); |
1425 | if (EC) |
1426 | return createFileError(F: "cannot open " + Twine(SummaryPath), EC); |
1427 | |
1428 | writeIndexToFile(Index: CombinedIndex, Out&: OS, ModuleToSummariesForIndex: &ModuleToSummariesForIndex, |
1429 | DecSummaries: &DeclarationSummaries); |
1430 | |
1431 | if (ShouldEmitImportsFiles) { |
1432 | Error ImportsFilesError = EmitImportsFiles( |
1433 | ModulePath, OutputFilename: NewModulePath + ".imports" , ModuleToSummariesForIndex); |
1434 | if (ImportsFilesError) |
1435 | return ImportsFilesError; |
1436 | } |
1437 | |
1438 | // Optionally, store the imports files. |
1439 | if (ImportsFiles) |
1440 | processImportsFiles( |
1441 | ModulePath, ModuleToSummariesForIndex, |
1442 | F: [&](StringRef M) { ImportsFiles->get().push_back(Elt: M.str()); }); |
1443 | |
1444 | return Error::success(); |
1445 | } |
1446 | |
1447 | namespace { |
1448 | /// Base class for ThinLTO backends that perform code generation and insert the |
1449 | /// generated files back into the link. |
1450 | class CGThinBackend : public ThinBackendProc { |
1451 | protected: |
1452 | AddStreamFn AddStream; |
1453 | DenseSet<GlobalValue::GUID> CfiFunctionDefs; |
1454 | DenseSet<GlobalValue::GUID> CfiFunctionDecls; |
1455 | bool ShouldEmitIndexFiles; |
1456 | |
1457 | public: |
1458 | CGThinBackend( |
1459 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1460 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1461 | AddStreamFn AddStream, lto::IndexWriteCallback OnWrite, |
1462 | bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles, |
1463 | ThreadPoolStrategy ThinLTOParallelism) |
1464 | : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries, |
1465 | OnWrite, ShouldEmitImportsFiles, ThinLTOParallelism), |
1466 | AddStream(std::move(AddStream)), |
1467 | ShouldEmitIndexFiles(ShouldEmitIndexFiles) { |
1468 | auto &Defs = CombinedIndex.cfiFunctionDefs(); |
1469 | CfiFunctionDefs.insert_range(R: Defs.guids()); |
1470 | auto &Decls = CombinedIndex.cfiFunctionDecls(); |
1471 | CfiFunctionDecls.insert_range(R: Decls.guids()); |
1472 | } |
1473 | }; |
1474 | |
1475 | /// This backend performs code generation by scheduling a job to run on |
1476 | /// an in-process thread when invoked for each task. |
1477 | class InProcessThinBackend : public CGThinBackend { |
1478 | protected: |
1479 | FileCache Cache; |
1480 | |
1481 | public: |
1482 | InProcessThinBackend( |
1483 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1484 | ThreadPoolStrategy ThinLTOParallelism, |
1485 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1486 | AddStreamFn AddStream, FileCache Cache, lto::IndexWriteCallback OnWrite, |
1487 | bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles) |
1488 | : CGThinBackend(Conf, CombinedIndex, ModuleToDefinedGVSummaries, |
1489 | AddStream, OnWrite, ShouldEmitIndexFiles, |
1490 | ShouldEmitImportsFiles, ThinLTOParallelism), |
1491 | Cache(std::move(Cache)) {} |
1492 | |
1493 | virtual Error runThinLTOBackendThread( |
1494 | AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM, |
1495 | ModuleSummaryIndex &CombinedIndex, |
1496 | const FunctionImporter::ImportMapTy &ImportList, |
1497 | const FunctionImporter::ExportSetTy &ExportList, |
1498 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
1499 | const GVSummaryMapTy &DefinedGlobals, |
1500 | MapVector<StringRef, BitcodeModule> &ModuleMap) { |
1501 | auto RunThinBackend = [&](AddStreamFn AddStream) { |
1502 | LTOLLVMContext BackendContext(Conf); |
1503 | Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(Context&: BackendContext); |
1504 | if (!MOrErr) |
1505 | return MOrErr.takeError(); |
1506 | |
1507 | return thinBackend(C: Conf, Task, AddStream, M&: **MOrErr, CombinedIndex, |
1508 | ImportList, DefinedGlobals, ModuleMap: &ModuleMap, |
1509 | CodeGenOnly: Conf.CodeGenOnly); |
1510 | }; |
1511 | |
1512 | auto ModuleID = BM.getModuleIdentifier(); |
1513 | |
1514 | if (ShouldEmitIndexFiles) { |
1515 | if (auto E = emitFiles(ImportList, ModulePath: ModuleID, NewModulePath: ModuleID.str())) |
1516 | return E; |
1517 | } |
1518 | |
1519 | if (!Cache.isValid() || !CombinedIndex.modulePaths().count(Key: ModuleID) || |
1520 | all_of(Range: CombinedIndex.getModuleHash(ModPath: ModuleID), |
1521 | P: [](uint32_t V) { return V == 0; })) |
1522 | // Cache disabled or no entry for this module in the combined index or |
1523 | // no module hash. |
1524 | return RunThinBackend(AddStream); |
1525 | |
1526 | // The module may be cached, this helps handling it. |
1527 | std::string Key = computeLTOCacheKey( |
1528 | Conf, Index: CombinedIndex, ModuleID, ImportList, ExportList, ResolvedODR, |
1529 | DefinedGlobals, CfiFunctionDefs, CfiFunctionDecls); |
1530 | Expected<AddStreamFn> CacheAddStreamOrErr = Cache(Task, Key, ModuleID); |
1531 | if (Error Err = CacheAddStreamOrErr.takeError()) |
1532 | return Err; |
1533 | AddStreamFn &CacheAddStream = *CacheAddStreamOrErr; |
1534 | if (CacheAddStream) |
1535 | return RunThinBackend(CacheAddStream); |
1536 | |
1537 | return Error::success(); |
1538 | } |
1539 | |
1540 | Error start( |
1541 | unsigned Task, BitcodeModule BM, |
1542 | const FunctionImporter::ImportMapTy &ImportList, |
1543 | const FunctionImporter::ExportSetTy &ExportList, |
1544 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
1545 | MapVector<StringRef, BitcodeModule> &ModuleMap) override { |
1546 | StringRef ModulePath = BM.getModuleIdentifier(); |
1547 | assert(ModuleToDefinedGVSummaries.count(ModulePath)); |
1548 | const GVSummaryMapTy &DefinedGlobals = |
1549 | ModuleToDefinedGVSummaries.find(Val: ModulePath)->second; |
1550 | BackendThreadPool.async( |
1551 | F: [=](BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, |
1552 | const FunctionImporter::ImportMapTy &ImportList, |
1553 | const FunctionImporter::ExportSetTy &ExportList, |
1554 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> |
1555 | &ResolvedODR, |
1556 | const GVSummaryMapTy &DefinedGlobals, |
1557 | MapVector<StringRef, BitcodeModule> &ModuleMap) { |
1558 | if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) |
1559 | timeTraceProfilerInitialize(TimeTraceGranularity: Conf.TimeTraceGranularity, |
1560 | ProcName: "thin backend" ); |
1561 | Error E = runThinLTOBackendThread( |
1562 | AddStream, Cache, Task, BM, CombinedIndex, ImportList, ExportList, |
1563 | ResolvedODR, DefinedGlobals, ModuleMap); |
1564 | if (E) { |
1565 | std::unique_lock<std::mutex> L(ErrMu); |
1566 | if (Err) |
1567 | Err = joinErrors(E1: std::move(*Err), E2: std::move(E)); |
1568 | else |
1569 | Err = std::move(E); |
1570 | } |
1571 | if (LLVM_ENABLE_THREADS && Conf.TimeTraceEnabled) |
1572 | timeTraceProfilerFinishThread(); |
1573 | }, |
1574 | ArgList&: BM, ArgList: std::ref(t&: CombinedIndex), ArgList: std::ref(t: ImportList), ArgList: std::ref(t: ExportList), |
1575 | ArgList: std::ref(t: ResolvedODR), ArgList: std::ref(t: DefinedGlobals), ArgList: std::ref(t&: ModuleMap)); |
1576 | |
1577 | if (OnWrite) |
1578 | OnWrite(std::string(ModulePath)); |
1579 | return Error::success(); |
1580 | } |
1581 | }; |
1582 | |
1583 | /// This backend is utilized in the first round of a two-codegen round process. |
1584 | /// It first saves optimized bitcode files to disk before the codegen process |
1585 | /// begins. After codegen, it stores the resulting object files in a scratch |
1586 | /// buffer. Note the codegen data stored in the scratch buffer will be extracted |
1587 | /// and merged in the subsequent step. |
1588 | class FirstRoundThinBackend : public InProcessThinBackend { |
1589 | AddStreamFn IRAddStream; |
1590 | FileCache IRCache; |
1591 | |
1592 | public: |
1593 | FirstRoundThinBackend( |
1594 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1595 | ThreadPoolStrategy ThinLTOParallelism, |
1596 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1597 | AddStreamFn CGAddStream, FileCache CGCache, AddStreamFn IRAddStream, |
1598 | FileCache IRCache) |
1599 | : InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism, |
1600 | ModuleToDefinedGVSummaries, std::move(CGAddStream), |
1601 | std::move(CGCache), /*OnWrite=*/nullptr, |
1602 | /*ShouldEmitIndexFiles=*/false, |
1603 | /*ShouldEmitImportsFiles=*/false), |
1604 | IRAddStream(std::move(IRAddStream)), IRCache(std::move(IRCache)) {} |
1605 | |
1606 | Error runThinLTOBackendThread( |
1607 | AddStreamFn CGAddStream, FileCache CGCache, unsigned Task, |
1608 | BitcodeModule BM, ModuleSummaryIndex &CombinedIndex, |
1609 | const FunctionImporter::ImportMapTy &ImportList, |
1610 | const FunctionImporter::ExportSetTy &ExportList, |
1611 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
1612 | const GVSummaryMapTy &DefinedGlobals, |
1613 | MapVector<StringRef, BitcodeModule> &ModuleMap) override { |
1614 | auto RunThinBackend = [&](AddStreamFn CGAddStream, |
1615 | AddStreamFn IRAddStream) { |
1616 | LTOLLVMContext BackendContext(Conf); |
1617 | Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(Context&: BackendContext); |
1618 | if (!MOrErr) |
1619 | return MOrErr.takeError(); |
1620 | |
1621 | return thinBackend(C: Conf, Task, AddStream: CGAddStream, M&: **MOrErr, CombinedIndex, |
1622 | ImportList, DefinedGlobals, ModuleMap: &ModuleMap, |
1623 | CodeGenOnly: Conf.CodeGenOnly, IRAddStream); |
1624 | }; |
1625 | |
1626 | auto ModuleID = BM.getModuleIdentifier(); |
1627 | // Like InProcessThinBackend, we produce index files as needed for |
1628 | // FirstRoundThinBackend. However, these files are not generated for |
1629 | // SecondRoundThinBackend. |
1630 | if (ShouldEmitIndexFiles) { |
1631 | if (auto E = emitFiles(ImportList, ModulePath: ModuleID, NewModulePath: ModuleID.str())) |
1632 | return E; |
1633 | } |
1634 | |
1635 | assert((CGCache.isValid() == IRCache.isValid()) && |
1636 | "Both caches for CG and IR should have matching availability" ); |
1637 | if (!CGCache.isValid() || !CombinedIndex.modulePaths().count(Key: ModuleID) || |
1638 | all_of(Range: CombinedIndex.getModuleHash(ModPath: ModuleID), |
1639 | P: [](uint32_t V) { return V == 0; })) |
1640 | // Cache disabled or no entry for this module in the combined index or |
1641 | // no module hash. |
1642 | return RunThinBackend(CGAddStream, IRAddStream); |
1643 | |
1644 | // Get CGKey for caching object in CGCache. |
1645 | std::string CGKey = computeLTOCacheKey( |
1646 | Conf, Index: CombinedIndex, ModuleID, ImportList, ExportList, ResolvedODR, |
1647 | DefinedGlobals, CfiFunctionDefs, CfiFunctionDecls); |
1648 | Expected<AddStreamFn> CacheCGAddStreamOrErr = |
1649 | CGCache(Task, CGKey, ModuleID); |
1650 | if (Error Err = CacheCGAddStreamOrErr.takeError()) |
1651 | return Err; |
1652 | AddStreamFn &CacheCGAddStream = *CacheCGAddStreamOrErr; |
1653 | |
1654 | // Get IRKey for caching (optimized) IR in IRCache with an extra ID. |
1655 | std::string IRKey = recomputeLTOCacheKey(Key: CGKey, /*ExtraID=*/"IR" ); |
1656 | Expected<AddStreamFn> CacheIRAddStreamOrErr = |
1657 | IRCache(Task, IRKey, ModuleID); |
1658 | if (Error Err = CacheIRAddStreamOrErr.takeError()) |
1659 | return Err; |
1660 | AddStreamFn &CacheIRAddStream = *CacheIRAddStreamOrErr; |
1661 | |
1662 | // Ideally, both CG and IR caching should be synchronized. However, in |
1663 | // practice, their availability may differ due to different expiration |
1664 | // times. Therefore, if either cache is missing, the backend process is |
1665 | // triggered. |
1666 | if (CacheCGAddStream || CacheIRAddStream) { |
1667 | LLVM_DEBUG(dbgs() << "[FirstRound] Cache Miss for " |
1668 | << BM.getModuleIdentifier() << "\n" ); |
1669 | return RunThinBackend(CacheCGAddStream ? CacheCGAddStream : CGAddStream, |
1670 | CacheIRAddStream ? CacheIRAddStream : IRAddStream); |
1671 | } |
1672 | |
1673 | return Error::success(); |
1674 | } |
1675 | }; |
1676 | |
1677 | /// This backend operates in the second round of a two-codegen round process. |
1678 | /// It starts by reading the optimized bitcode files that were saved during the |
1679 | /// first round. The backend then executes the codegen only to further optimize |
1680 | /// the code, utilizing the codegen data merged from the first round. Finally, |
1681 | /// it writes the resulting object files as usual. |
1682 | class SecondRoundThinBackend : public InProcessThinBackend { |
1683 | std::unique_ptr<SmallVector<StringRef>> IRFiles; |
1684 | stable_hash CombinedCGDataHash; |
1685 | |
1686 | public: |
1687 | SecondRoundThinBackend( |
1688 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1689 | ThreadPoolStrategy ThinLTOParallelism, |
1690 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1691 | AddStreamFn AddStream, FileCache Cache, |
1692 | std::unique_ptr<SmallVector<StringRef>> IRFiles, |
1693 | stable_hash CombinedCGDataHash) |
1694 | : InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism, |
1695 | ModuleToDefinedGVSummaries, std::move(AddStream), |
1696 | std::move(Cache), |
1697 | /*OnWrite=*/nullptr, |
1698 | /*ShouldEmitIndexFiles=*/false, |
1699 | /*ShouldEmitImportsFiles=*/false), |
1700 | IRFiles(std::move(IRFiles)), CombinedCGDataHash(CombinedCGDataHash) {} |
1701 | |
1702 | virtual Error runThinLTOBackendThread( |
1703 | AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM, |
1704 | ModuleSummaryIndex &CombinedIndex, |
1705 | const FunctionImporter::ImportMapTy &ImportList, |
1706 | const FunctionImporter::ExportSetTy &ExportList, |
1707 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
1708 | const GVSummaryMapTy &DefinedGlobals, |
1709 | MapVector<StringRef, BitcodeModule> &ModuleMap) override { |
1710 | auto RunThinBackend = [&](AddStreamFn AddStream) { |
1711 | LTOLLVMContext BackendContext(Conf); |
1712 | std::unique_ptr<Module> LoadedModule = |
1713 | cgdata::loadModuleForTwoRounds(OrigModule&: BM, Task, Context&: BackendContext, IRFiles: *IRFiles); |
1714 | |
1715 | return thinBackend(C: Conf, Task, AddStream, M&: *LoadedModule, CombinedIndex, |
1716 | ImportList, DefinedGlobals, ModuleMap: &ModuleMap, |
1717 | /*CodeGenOnly=*/true); |
1718 | }; |
1719 | |
1720 | auto ModuleID = BM.getModuleIdentifier(); |
1721 | if (!Cache.isValid() || !CombinedIndex.modulePaths().count(Key: ModuleID) || |
1722 | all_of(Range: CombinedIndex.getModuleHash(ModPath: ModuleID), |
1723 | P: [](uint32_t V) { return V == 0; })) |
1724 | // Cache disabled or no entry for this module in the combined index or |
1725 | // no module hash. |
1726 | return RunThinBackend(AddStream); |
1727 | |
1728 | // Get Key for caching the final object file in Cache with the combined |
1729 | // CGData hash. |
1730 | std::string Key = computeLTOCacheKey( |
1731 | Conf, Index: CombinedIndex, ModuleID, ImportList, ExportList, ResolvedODR, |
1732 | DefinedGlobals, CfiFunctionDefs, CfiFunctionDecls); |
1733 | Key = recomputeLTOCacheKey(Key, |
1734 | /*ExtraID=*/std::to_string(val: CombinedCGDataHash)); |
1735 | Expected<AddStreamFn> CacheAddStreamOrErr = Cache(Task, Key, ModuleID); |
1736 | if (Error Err = CacheAddStreamOrErr.takeError()) |
1737 | return Err; |
1738 | AddStreamFn &CacheAddStream = *CacheAddStreamOrErr; |
1739 | |
1740 | if (CacheAddStream) { |
1741 | LLVM_DEBUG(dbgs() << "[SecondRound] Cache Miss for " |
1742 | << BM.getModuleIdentifier() << "\n" ); |
1743 | return RunThinBackend(CacheAddStream); |
1744 | } |
1745 | |
1746 | return Error::success(); |
1747 | } |
1748 | }; |
1749 | } // end anonymous namespace |
1750 | |
1751 | ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism, |
1752 | lto::IndexWriteCallback OnWrite, |
1753 | bool ShouldEmitIndexFiles, |
1754 | bool ShouldEmitImportsFiles) { |
1755 | auto Func = |
1756 | [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1757 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1758 | AddStreamFn AddStream, FileCache Cache) { |
1759 | return std::make_unique<InProcessThinBackend>( |
1760 | args: Conf, args&: CombinedIndex, args: Parallelism, args: ModuleToDefinedGVSummaries, |
1761 | args&: AddStream, args&: Cache, args: OnWrite, args: ShouldEmitIndexFiles, |
1762 | args: ShouldEmitImportsFiles); |
1763 | }; |
1764 | return ThinBackend(Func, Parallelism); |
1765 | } |
1766 | |
1767 | StringLiteral lto::getThinLTODefaultCPU(const Triple &TheTriple) { |
1768 | if (!TheTriple.isOSDarwin()) |
1769 | return "" ; |
1770 | if (TheTriple.getArch() == Triple::x86_64) |
1771 | return "core2" ; |
1772 | if (TheTriple.getArch() == Triple::x86) |
1773 | return "yonah" ; |
1774 | if (TheTriple.isArm64e()) |
1775 | return "apple-a12" ; |
1776 | if (TheTriple.getArch() == Triple::aarch64 || |
1777 | TheTriple.getArch() == Triple::aarch64_32) |
1778 | return "cyclone" ; |
1779 | return "" ; |
1780 | } |
1781 | |
1782 | // Given the original \p Path to an output file, replace any path |
1783 | // prefix matching \p OldPrefix with \p NewPrefix. Also, create the |
1784 | // resulting directory if it does not yet exist. |
1785 | std::string lto::getThinLTOOutputFile(StringRef Path, StringRef OldPrefix, |
1786 | StringRef NewPrefix) { |
1787 | if (OldPrefix.empty() && NewPrefix.empty()) |
1788 | return std::string(Path); |
1789 | SmallString<128> NewPath(Path); |
1790 | llvm::sys::path::replace_path_prefix(Path&: NewPath, OldPrefix, NewPrefix); |
1791 | StringRef ParentPath = llvm::sys::path::parent_path(path: NewPath.str()); |
1792 | if (!ParentPath.empty()) { |
1793 | // Make sure the new directory exists, creating it if necessary. |
1794 | if (std::error_code EC = llvm::sys::fs::create_directories(path: ParentPath)) |
1795 | llvm::errs() << "warning: could not create directory '" << ParentPath |
1796 | << "': " << EC.message() << '\n'; |
1797 | } |
1798 | return std::string(NewPath); |
1799 | } |
1800 | |
1801 | namespace { |
1802 | class WriteIndexesThinBackend : public ThinBackendProc { |
1803 | std::string OldPrefix, NewPrefix, NativeObjectPrefix; |
1804 | raw_fd_ostream *LinkedObjectsFile; |
1805 | |
1806 | public: |
1807 | WriteIndexesThinBackend( |
1808 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1809 | ThreadPoolStrategy ThinLTOParallelism, |
1810 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1811 | std::string OldPrefix, std::string NewPrefix, |
1812 | std::string NativeObjectPrefix, bool ShouldEmitImportsFiles, |
1813 | raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite) |
1814 | : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries, |
1815 | OnWrite, ShouldEmitImportsFiles, ThinLTOParallelism), |
1816 | OldPrefix(OldPrefix), NewPrefix(NewPrefix), |
1817 | NativeObjectPrefix(NativeObjectPrefix), |
1818 | LinkedObjectsFile(LinkedObjectsFile) {} |
1819 | |
1820 | Error start( |
1821 | unsigned Task, BitcodeModule BM, |
1822 | const FunctionImporter::ImportMapTy &ImportList, |
1823 | const FunctionImporter::ExportSetTy &ExportList, |
1824 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
1825 | MapVector<StringRef, BitcodeModule> &ModuleMap) override { |
1826 | StringRef ModulePath = BM.getModuleIdentifier(); |
1827 | |
1828 | // The contents of this file may be used as input to a native link, and must |
1829 | // therefore contain the processed modules in a determinstic order that |
1830 | // match the order they are provided on the command line. For that reason, |
1831 | // we cannot include this in the asynchronously executed lambda below. |
1832 | if (LinkedObjectsFile) { |
1833 | std::string ObjectPrefix = |
1834 | NativeObjectPrefix.empty() ? NewPrefix : NativeObjectPrefix; |
1835 | std::string LinkedObjectsFilePath = |
1836 | getThinLTOOutputFile(Path: ModulePath, OldPrefix, NewPrefix: ObjectPrefix); |
1837 | *LinkedObjectsFile << LinkedObjectsFilePath << '\n'; |
1838 | } |
1839 | |
1840 | BackendThreadPool.async( |
1841 | F: [this](const StringRef ModulePath, |
1842 | const FunctionImporter::ImportMapTy &ImportList, |
1843 | const std::string &OldPrefix, const std::string &NewPrefix) { |
1844 | std::string NewModulePath = |
1845 | getThinLTOOutputFile(Path: ModulePath, OldPrefix, NewPrefix); |
1846 | auto E = emitFiles(ImportList, ModulePath, NewModulePath); |
1847 | if (E) { |
1848 | std::unique_lock<std::mutex> L(ErrMu); |
1849 | if (Err) |
1850 | Err = joinErrors(E1: std::move(*Err), E2: std::move(E)); |
1851 | else |
1852 | Err = std::move(E); |
1853 | return; |
1854 | } |
1855 | }, |
1856 | ArgList&: ModulePath, ArgList: ImportList, ArgList&: OldPrefix, ArgList&: NewPrefix); |
1857 | |
1858 | if (OnWrite) |
1859 | OnWrite(std::string(ModulePath)); |
1860 | return Error::success(); |
1861 | } |
1862 | |
1863 | bool isSensitiveToInputOrder() override { |
1864 | // The order which modules are written to LinkedObjectsFile should be |
1865 | // deterministic and match the order they are passed on the command line. |
1866 | return true; |
1867 | } |
1868 | }; |
1869 | } // end anonymous namespace |
1870 | |
1871 | ThinBackend lto::createWriteIndexesThinBackend( |
1872 | ThreadPoolStrategy Parallelism, std::string OldPrefix, |
1873 | std::string NewPrefix, std::string NativeObjectPrefix, |
1874 | bool ShouldEmitImportsFiles, raw_fd_ostream *LinkedObjectsFile, |
1875 | IndexWriteCallback OnWrite) { |
1876 | auto Func = |
1877 | [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
1878 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
1879 | AddStreamFn AddStream, FileCache Cache) { |
1880 | return std::make_unique<WriteIndexesThinBackend>( |
1881 | args: Conf, args&: CombinedIndex, args: Parallelism, args: ModuleToDefinedGVSummaries, |
1882 | args: OldPrefix, args: NewPrefix, args: NativeObjectPrefix, args: ShouldEmitImportsFiles, |
1883 | args: LinkedObjectsFile, args: OnWrite); |
1884 | }; |
1885 | return ThinBackend(Func, Parallelism); |
1886 | } |
1887 | |
1888 | Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache, |
1889 | const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) { |
1890 | LLVM_DEBUG(dbgs() << "Running ThinLTO\n" ); |
1891 | ThinLTO.CombinedIndex.releaseTemporaryMemory(); |
1892 | timeTraceProfilerBegin(Name: "ThinLink" , Detail: StringRef("" )); |
1893 | auto TimeTraceScopeExit = llvm::make_scope_exit(F: []() { |
1894 | if (llvm::timeTraceProfilerEnabled()) |
1895 | llvm::timeTraceProfilerEnd(); |
1896 | }); |
1897 | if (ThinLTO.ModuleMap.empty()) |
1898 | return Error::success(); |
1899 | |
1900 | if (ThinLTO.ModulesToCompile && ThinLTO.ModulesToCompile->empty()) { |
1901 | llvm::errs() << "warning: [ThinLTO] No module compiled\n" ; |
1902 | return Error::success(); |
1903 | } |
1904 | |
1905 | if (Conf.CombinedIndexHook && |
1906 | !Conf.CombinedIndexHook(ThinLTO.CombinedIndex, GUIDPreservedSymbols)) |
1907 | return Error::success(); |
1908 | |
1909 | // Collect for each module the list of function it defines (GUID -> |
1910 | // Summary). |
1911 | DenseMap<StringRef, GVSummaryMapTy> ModuleToDefinedGVSummaries( |
1912 | ThinLTO.ModuleMap.size()); |
1913 | ThinLTO.CombinedIndex.collectDefinedGVSummariesPerModule( |
1914 | ModuleToDefinedGVSummaries); |
1915 | // Create entries for any modules that didn't have any GV summaries |
1916 | // (either they didn't have any GVs to start with, or we suppressed |
1917 | // generation of the summaries because they e.g. had inline assembly |
1918 | // uses that couldn't be promoted/renamed on export). This is so |
1919 | // InProcessThinBackend::start can still launch a backend thread, which |
1920 | // is passed the map of summaries for the module, without any special |
1921 | // handling for this case. |
1922 | for (auto &Mod : ThinLTO.ModuleMap) |
1923 | if (!ModuleToDefinedGVSummaries.count(Val: Mod.first)) |
1924 | ModuleToDefinedGVSummaries.try_emplace(Key: Mod.first); |
1925 | |
1926 | FunctionImporter::ImportListsTy ImportLists(ThinLTO.ModuleMap.size()); |
1927 | DenseMap<StringRef, FunctionImporter::ExportSetTy> ExportLists( |
1928 | ThinLTO.ModuleMap.size()); |
1929 | StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR; |
1930 | |
1931 | if (DumpThinCGSCCs) |
1932 | ThinLTO.CombinedIndex.dumpSCCs(OS&: outs()); |
1933 | |
1934 | std::set<GlobalValue::GUID> ExportedGUIDs; |
1935 | |
1936 | bool WholeProgramVisibilityEnabledInLTO = |
1937 | Conf.HasWholeProgramVisibility && |
1938 | // If validation is enabled, upgrade visibility only when all vtables |
1939 | // have typeinfos. |
1940 | (!Conf.ValidateAllVtablesHaveTypeInfos || Conf.AllVtablesHaveTypeInfos); |
1941 | if (hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) |
1942 | ThinLTO.CombinedIndex.setWithWholeProgramVisibility(); |
1943 | |
1944 | // If we're validating, get the vtable symbols that should not be |
1945 | // upgraded because they correspond to typeIDs outside of index-based |
1946 | // WPD info. |
1947 | DenseSet<GlobalValue::GUID> VisibleToRegularObjSymbols; |
1948 | if (WholeProgramVisibilityEnabledInLTO && |
1949 | Conf.ValidateAllVtablesHaveTypeInfos) { |
1950 | // This returns true when the name is local or not defined. Locals are |
1951 | // expected to be handled separately. |
1952 | auto IsVisibleToRegularObj = [&](StringRef name) { |
1953 | auto It = GlobalResolutions->find(Val: name); |
1954 | return (It == GlobalResolutions->end() || |
1955 | It->second.VisibleOutsideSummary || !It->second.Prevailing); |
1956 | }; |
1957 | |
1958 | getVisibleToRegularObjVtableGUIDs(Index&: ThinLTO.CombinedIndex, |
1959 | VisibleToRegularObjSymbols, |
1960 | IsVisibleToRegularObj); |
1961 | } |
1962 | |
1963 | // If allowed, upgrade public vcall visibility to linkage unit visibility in |
1964 | // the summaries before whole program devirtualization below. |
1965 | updateVCallVisibilityInIndex( |
1966 | Index&: ThinLTO.CombinedIndex, WholeProgramVisibilityEnabledInLTO, |
1967 | DynamicExportSymbols, VisibleToRegularObjSymbols); |
1968 | |
1969 | // Perform index-based WPD. This will return immediately if there are |
1970 | // no index entries in the typeIdMetadata map (e.g. if we are instead |
1971 | // performing IR-based WPD in hybrid regular/thin LTO mode). |
1972 | std::map<ValueInfo, std::vector<VTableSlotSummary>> LocalWPDTargetsMap; |
1973 | runWholeProgramDevirtOnIndex(Summary&: ThinLTO.CombinedIndex, ExportedGUIDs, |
1974 | LocalWPDTargetsMap); |
1975 | |
1976 | auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) { |
1977 | return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath(); |
1978 | }; |
1979 | if (EnableMemProfContextDisambiguation) { |
1980 | MemProfContextDisambiguation ContextDisambiguation; |
1981 | ContextDisambiguation.run(Index&: ThinLTO.CombinedIndex, isPrevailing); |
1982 | } |
1983 | |
1984 | // Figure out which symbols need to be internalized. This also needs to happen |
1985 | // at -O0 because summary-based DCE is implemented using internalization, and |
1986 | // we must apply DCE consistently with the full LTO module in order to avoid |
1987 | // undefined references during the final link. |
1988 | for (auto &Res : *GlobalResolutions) { |
1989 | // If the symbol does not have external references or it is not prevailing, |
1990 | // then not need to mark it as exported from a ThinLTO partition. |
1991 | if (Res.second.Partition != GlobalResolution::External || |
1992 | !Res.second.isPrevailingIRSymbol()) |
1993 | continue; |
1994 | auto GUID = GlobalValue::getGUIDAssumingExternalLinkage( |
1995 | GlobalName: GlobalValue::dropLLVMManglingEscape(Name: Res.second.IRName)); |
1996 | // Mark exported unless index-based analysis determined it to be dead. |
1997 | if (ThinLTO.CombinedIndex.isGUIDLive(GUID)) |
1998 | ExportedGUIDs.insert(x: GUID); |
1999 | } |
2000 | |
2001 | // Reset the GlobalResolutions to deallocate the associated memory, as there |
2002 | // are no further accesses. We specifically want to do this before computing |
2003 | // cross module importing, which adds to peak memory via the computed import |
2004 | // and export lists. |
2005 | releaseGlobalResolutionsMemory(); |
2006 | |
2007 | if (Conf.OptLevel > 0) |
2008 | ComputeCrossModuleImport(Index: ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, |
2009 | isPrevailing, ImportLists, ExportLists); |
2010 | |
2011 | // Any functions referenced by the jump table in the regular LTO object must |
2012 | // be exported. |
2013 | auto &Defs = ThinLTO.CombinedIndex.cfiFunctionDefs(); |
2014 | ExportedGUIDs.insert(first: Defs.guid_begin(), last: Defs.guid_end()); |
2015 | auto &Decls = ThinLTO.CombinedIndex.cfiFunctionDecls(); |
2016 | ExportedGUIDs.insert(first: Decls.guid_begin(), last: Decls.guid_end()); |
2017 | |
2018 | auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) { |
2019 | const auto &ExportList = ExportLists.find(Val: ModuleIdentifier); |
2020 | return (ExportList != ExportLists.end() && ExportList->second.count(V: VI)) || |
2021 | ExportedGUIDs.count(x: VI.getGUID()); |
2022 | }; |
2023 | |
2024 | // Update local devirtualized targets that were exported by cross-module |
2025 | // importing or by other devirtualizations marked in the ExportedGUIDs set. |
2026 | updateIndexWPDForExports(Summary&: ThinLTO.CombinedIndex, isExported, |
2027 | LocalWPDTargetsMap); |
2028 | |
2029 | thinLTOInternalizeAndPromoteInIndex(Index&: ThinLTO.CombinedIndex, isExported, |
2030 | isPrevailing); |
2031 | |
2032 | auto recordNewLinkage = [&](StringRef ModuleIdentifier, |
2033 | GlobalValue::GUID GUID, |
2034 | GlobalValue::LinkageTypes NewLinkage) { |
2035 | ResolvedODR[ModuleIdentifier][GUID] = NewLinkage; |
2036 | }; |
2037 | thinLTOResolvePrevailingInIndex(C: Conf, Index&: ThinLTO.CombinedIndex, isPrevailing, |
2038 | recordNewLinkage, GUIDPreservedSymbols); |
2039 | |
2040 | thinLTOPropagateFunctionAttrs(Index&: ThinLTO.CombinedIndex, isPrevailing); |
2041 | |
2042 | generateParamAccessSummary(Index&: ThinLTO.CombinedIndex); |
2043 | |
2044 | if (llvm::timeTraceProfilerEnabled()) |
2045 | llvm::timeTraceProfilerEnd(); |
2046 | |
2047 | TimeTraceScopeExit.release(); |
2048 | |
2049 | auto &ModuleMap = |
2050 | ThinLTO.ModulesToCompile ? *ThinLTO.ModulesToCompile : ThinLTO.ModuleMap; |
2051 | |
2052 | auto RunBackends = [&](ThinBackendProc *BackendProcess) -> Error { |
2053 | auto ProcessOneModule = [&](int I) -> Error { |
2054 | auto &Mod = *(ModuleMap.begin() + I); |
2055 | // Tasks 0 through ParallelCodeGenParallelismLevel-1 are reserved for |
2056 | // combined module and parallel code generation partitions. |
2057 | return BackendProcess->start( |
2058 | Task: RegularLTO.ParallelCodeGenParallelismLevel + I, BM: Mod.second, |
2059 | ImportList: ImportLists[Mod.first], ExportList: ExportLists[Mod.first], |
2060 | ResolvedODR: ResolvedODR[Mod.first], ModuleMap&: ThinLTO.ModuleMap); |
2061 | }; |
2062 | |
2063 | BackendProcess->setup(ThinLTONumTasks: ModuleMap.size(), |
2064 | ThinLTOTaskOffset: RegularLTO.ParallelCodeGenParallelismLevel, |
2065 | Triple: RegularLTO.CombinedModule->getTargetTriple()); |
2066 | |
2067 | if (BackendProcess->getThreadCount() == 1 || |
2068 | BackendProcess->isSensitiveToInputOrder()) { |
2069 | // Process the modules in the order they were provided on the |
2070 | // command-line. It is important for this codepath to be used for |
2071 | // WriteIndexesThinBackend, to ensure the emitted LinkedObjectsFile lists |
2072 | // ThinLTO objects in the same order as the inputs, which otherwise would |
2073 | // affect the final link order. |
2074 | for (int I = 0, E = ModuleMap.size(); I != E; ++I) |
2075 | if (Error E = ProcessOneModule(I)) |
2076 | return E; |
2077 | } else { |
2078 | // When executing in parallel, process largest bitsize modules first to |
2079 | // improve parallelism, and avoid starving the thread pool near the end. |
2080 | // This saves about 15 sec on a 36-core machine while link `clang.exe` |
2081 | // (out of 100 sec). |
2082 | std::vector<BitcodeModule *> ModulesVec; |
2083 | ModulesVec.reserve(n: ModuleMap.size()); |
2084 | for (auto &Mod : ModuleMap) |
2085 | ModulesVec.push_back(x: &Mod.second); |
2086 | for (int I : generateModulesOrdering(R: ModulesVec)) |
2087 | if (Error E = ProcessOneModule(I)) |
2088 | return E; |
2089 | } |
2090 | return BackendProcess->wait(); |
2091 | }; |
2092 | |
2093 | if (!CodeGenDataThinLTOTwoRounds) { |
2094 | std::unique_ptr<ThinBackendProc> BackendProc = |
2095 | ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries, |
2096 | AddStream, Cache); |
2097 | return RunBackends(BackendProc.get()); |
2098 | } |
2099 | |
2100 | // Perform two rounds of code generation for ThinLTO: |
2101 | // 1. First round: Perform optimization and code generation, outputting to |
2102 | // temporary scratch objects. |
2103 | // 2. Merge code generation data extracted from the temporary scratch objects. |
2104 | // 3. Second round: Execute code generation again using the merged data. |
2105 | LLVM_DEBUG(dbgs() << "[TwoRounds] Initializing ThinLTO two-codegen rounds\n" ); |
2106 | |
2107 | unsigned MaxTasks = getMaxTasks(); |
2108 | auto Parallelism = ThinLTO.Backend.getParallelism(); |
2109 | // Set up two additional streams and caches for storing temporary scratch |
2110 | // objects and optimized IRs, using the same cache directory as the original. |
2111 | cgdata::StreamCacheData CG(MaxTasks, Cache, "CG" ), IR(MaxTasks, Cache, "IR" ); |
2112 | |
2113 | // First round: Execute optimization and code generation, outputting to |
2114 | // temporary scratch objects. Serialize the optimized IRs before initiating |
2115 | // code generation. |
2116 | LLVM_DEBUG(dbgs() << "[TwoRounds] Running the first round of codegen\n" ); |
2117 | auto FirstRoundLTO = std::make_unique<FirstRoundThinBackend>( |
2118 | args&: Conf, args&: ThinLTO.CombinedIndex, args&: Parallelism, args&: ModuleToDefinedGVSummaries, |
2119 | args&: CG.AddStream, args&: CG.Cache, args&: IR.AddStream, args&: IR.Cache); |
2120 | if (Error E = RunBackends(FirstRoundLTO.get())) |
2121 | return E; |
2122 | |
2123 | LLVM_DEBUG(dbgs() << "[TwoRounds] Merging codegen data\n" ); |
2124 | auto CombinedHashOrErr = cgdata::mergeCodeGenData(ObjectFiles: *CG.getResult()); |
2125 | if (Error E = CombinedHashOrErr.takeError()) |
2126 | return E; |
2127 | auto CombinedHash = *CombinedHashOrErr; |
2128 | LLVM_DEBUG(dbgs() << "[TwoRounds] CGData hash: " << CombinedHash << "\n" ); |
2129 | |
2130 | // Second round: Read the optimized IRs and execute code generation using the |
2131 | // merged data. |
2132 | LLVM_DEBUG(dbgs() << "[TwoRounds] Running the second round of codegen\n" ); |
2133 | auto SecondRoundLTO = std::make_unique<SecondRoundThinBackend>( |
2134 | args&: Conf, args&: ThinLTO.CombinedIndex, args&: Parallelism, args&: ModuleToDefinedGVSummaries, |
2135 | args&: AddStream, args&: Cache, args: IR.getResult(), args&: CombinedHash); |
2136 | return RunBackends(SecondRoundLTO.get()); |
2137 | } |
2138 | |
2139 | Expected<std::unique_ptr<ToolOutputFile>> lto::( |
2140 | LLVMContext &Context, StringRef , StringRef , |
2141 | StringRef , bool , |
2142 | std::optional<uint64_t> , int Count) { |
2143 | std::string Filename = std::string(RemarksFilename); |
2144 | // For ThinLTO, file.opt.<format> becomes |
2145 | // file.opt.<format>.thin.<num>.<format>. |
2146 | if (!Filename.empty() && Count != -1) |
2147 | Filename = |
2148 | (Twine(Filename) + ".thin." + llvm::utostr(X: Count) + "." + RemarksFormat) |
2149 | .str(); |
2150 | |
2151 | auto ResultOrErr = llvm::setupLLVMOptimizationRemarks( |
2152 | Context, RemarksFilename: Filename, RemarksPasses, RemarksFormat, RemarksWithHotness, |
2153 | RemarksHotnessThreshold); |
2154 | if (Error E = ResultOrErr.takeError()) |
2155 | return std::move(E); |
2156 | |
2157 | if (*ResultOrErr) |
2158 | (*ResultOrErr)->keep(); |
2159 | |
2160 | return ResultOrErr; |
2161 | } |
2162 | |
2163 | Expected<std::unique_ptr<ToolOutputFile>> |
2164 | lto::setupStatsFile(StringRef StatsFilename) { |
2165 | // Setup output file to emit statistics. |
2166 | if (StatsFilename.empty()) |
2167 | return nullptr; |
2168 | |
2169 | llvm::EnableStatistics(DoPrintOnExit: false); |
2170 | std::error_code EC; |
2171 | auto StatsFile = |
2172 | std::make_unique<ToolOutputFile>(args&: StatsFilename, args&: EC, args: sys::fs::OF_None); |
2173 | if (EC) |
2174 | return errorCodeToError(EC); |
2175 | |
2176 | StatsFile->keep(); |
2177 | return std::move(StatsFile); |
2178 | } |
2179 | |
2180 | // Compute the ordering we will process the inputs: the rough heuristic here |
2181 | // is to sort them per size so that the largest module get schedule as soon as |
2182 | // possible. This is purely a compile-time optimization. |
2183 | std::vector<int> lto::generateModulesOrdering(ArrayRef<BitcodeModule *> R) { |
2184 | auto Seq = llvm::seq<int>(Begin: 0, End: R.size()); |
2185 | std::vector<int> ModulesOrdering(Seq.begin(), Seq.end()); |
2186 | llvm::sort(C&: ModulesOrdering, Comp: [&](int LeftIndex, int RightIndex) { |
2187 | auto LSize = R[LeftIndex]->getBuffer().size(); |
2188 | auto RSize = R[RightIndex]->getBuffer().size(); |
2189 | return LSize > RSize; |
2190 | }); |
2191 | return ModulesOrdering; |
2192 | } |
2193 | |
2194 | namespace { |
2195 | /// This out-of-process backend does not perform code generation when invoked |
2196 | /// for each task. Instead, it generates the necessary information (e.g., the |
2197 | /// summary index shard, import list, etc.) to enable code generation to be |
2198 | /// performed externally, similar to WriteIndexesThinBackend. The backend's |
2199 | /// `wait` function then invokes an external distributor process to carry out |
2200 | /// the backend compilations. |
2201 | class OutOfProcessThinBackend : public CGThinBackend { |
2202 | using SString = SmallString<128>; |
2203 | |
2204 | BumpPtrAllocator Alloc; |
2205 | StringSaver Saver{Alloc}; |
2206 | |
2207 | SString LinkerOutputFile; |
2208 | |
2209 | SString DistributorPath; |
2210 | ArrayRef<StringRef> DistributorArgs; |
2211 | |
2212 | SString RemoteCompiler; |
2213 | ArrayRef<StringRef> RemoteCompilerArgs; |
2214 | |
2215 | bool SaveTemps; |
2216 | |
2217 | SmallVector<StringRef, 0> CodegenOptions; |
2218 | DenseSet<StringRef> CommonInputs; |
2219 | |
2220 | // Information specific to individual backend compilation job. |
2221 | struct Job { |
2222 | unsigned Task; |
2223 | StringRef ModuleID; |
2224 | StringRef NativeObjectPath; |
2225 | StringRef SummaryIndexPath; |
2226 | ImportsFilesContainer ImportsFiles; |
2227 | }; |
2228 | // The set of backend compilations jobs. |
2229 | SmallVector<Job> Jobs; |
2230 | |
2231 | // A unique string to identify the current link. |
2232 | SmallString<8> UID; |
2233 | |
2234 | // The offset to the first ThinLTO task. |
2235 | unsigned ThinLTOTaskOffset; |
2236 | |
2237 | // The target triple to supply for backend compilations. |
2238 | llvm::Triple Triple; |
2239 | |
2240 | public: |
2241 | OutOfProcessThinBackend( |
2242 | const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
2243 | ThreadPoolStrategy ThinLTOParallelism, |
2244 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
2245 | AddStreamFn AddStream, lto::IndexWriteCallback OnWrite, |
2246 | bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles, |
2247 | StringRef LinkerOutputFile, StringRef Distributor, |
2248 | ArrayRef<StringRef> DistributorArgs, StringRef RemoteCompiler, |
2249 | ArrayRef<StringRef> RemoteCompilerArgs, bool SaveTemps) |
2250 | : CGThinBackend(Conf, CombinedIndex, ModuleToDefinedGVSummaries, |
2251 | AddStream, OnWrite, ShouldEmitIndexFiles, |
2252 | ShouldEmitImportsFiles, ThinLTOParallelism), |
2253 | LinkerOutputFile(LinkerOutputFile), DistributorPath(Distributor), |
2254 | DistributorArgs(DistributorArgs), RemoteCompiler(RemoteCompiler), |
2255 | RemoteCompilerArgs(RemoteCompilerArgs), SaveTemps(SaveTemps) {} |
2256 | |
2257 | virtual void setup(unsigned ThinLTONumTasks, unsigned ThinLTOTaskOffset, |
2258 | llvm::Triple Triple) override { |
2259 | UID = itostr(X: sys::Process::getProcessId()); |
2260 | Jobs.resize(N: (size_t)ThinLTONumTasks); |
2261 | this->ThinLTOTaskOffset = ThinLTOTaskOffset; |
2262 | this->Triple = Triple; |
2263 | } |
2264 | |
2265 | Error start( |
2266 | unsigned Task, BitcodeModule BM, |
2267 | const FunctionImporter::ImportMapTy &ImportList, |
2268 | const FunctionImporter::ExportSetTy &ExportList, |
2269 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
2270 | MapVector<StringRef, BitcodeModule> &ModuleMap) override { |
2271 | |
2272 | StringRef ModulePath = BM.getModuleIdentifier(); |
2273 | |
2274 | SString ObjFilePath = sys::path::parent_path(path: LinkerOutputFile); |
2275 | sys::path::append(path&: ObjFilePath, a: sys::path::stem(path: ModulePath) + "." + |
2276 | itostr(X: Task) + "." + UID + ".native.o" ); |
2277 | |
2278 | Job &J = Jobs[Task - ThinLTOTaskOffset]; |
2279 | J = { |
2280 | .Task: Task, |
2281 | .ModuleID: ModulePath, |
2282 | .NativeObjectPath: Saver.save(S: ObjFilePath.str()), |
2283 | .SummaryIndexPath: Saver.save(S: ObjFilePath.str() + ".thinlto.bc" ), |
2284 | .ImportsFiles: {} // Filled in by emitFiles below. |
2285 | }; |
2286 | |
2287 | assert(ModuleToDefinedGVSummaries.count(ModulePath)); |
2288 | |
2289 | // The BackendThreadPool is only used here to write the sharded index files |
2290 | // (similar to WriteIndexesThinBackend). |
2291 | BackendThreadPool.async( |
2292 | F: [=](Job &J, const FunctionImporter::ImportMapTy &ImportList) { |
2293 | if (auto E = emitFiles(ImportList, ModulePath: J.ModuleID, NewModulePath: J.ModuleID.str(), |
2294 | SummaryPath: J.SummaryIndexPath, ImportsFiles: J.ImportsFiles)) { |
2295 | std::unique_lock<std::mutex> L(ErrMu); |
2296 | if (Err) |
2297 | Err = joinErrors(E1: std::move(*Err), E2: std::move(E)); |
2298 | else |
2299 | Err = std::move(E); |
2300 | } |
2301 | }, |
2302 | ArgList: std::ref(t&: J), ArgList: std::ref(t: ImportList)); |
2303 | |
2304 | return Error::success(); |
2305 | } |
2306 | |
2307 | // Derive a set of Clang options that will be shared/common for all DTLTO |
2308 | // backend compilations. We are intentionally minimal here as these options |
2309 | // must remain synchronized with the behavior of Clang. DTLTO does not support |
2310 | // all the features available with in-process LTO. More features are expected |
2311 | // to be added over time. Users can specify Clang options directly if a |
2312 | // feature is not supported. Note that explicitly specified options that imply |
2313 | // additional input or output file dependencies must be communicated to the |
2314 | // distribution system, potentially by setting extra options on the |
2315 | // distributor program. |
2316 | void buildCommonRemoteCompilerOptions() { |
2317 | const lto::Config &C = Conf; |
2318 | auto &Ops = CodegenOptions; |
2319 | |
2320 | Ops.push_back(Elt: Saver.save(S: "-O" + Twine(C.OptLevel))); |
2321 | |
2322 | if (C.Options.EmitAddrsig) |
2323 | Ops.push_back(Elt: "-faddrsig" ); |
2324 | if (C.Options.FunctionSections) |
2325 | Ops.push_back(Elt: "-ffunction-sections" ); |
2326 | if (C.Options.DataSections) |
2327 | Ops.push_back(Elt: "-fdata-sections" ); |
2328 | |
2329 | if (C.RelocModel == Reloc::PIC_) |
2330 | // Clang doesn't have -fpic for all triples. |
2331 | if (!Triple.isOSBinFormatCOFF()) |
2332 | Ops.push_back(Elt: "-fpic" ); |
2333 | |
2334 | // Turn on/off warnings about profile cfg mismatch (default on) |
2335 | // --lto-pgo-warn-mismatch. |
2336 | if (!C.PGOWarnMismatch) { |
2337 | Ops.push_back(Elt: "-mllvm" ); |
2338 | Ops.push_back(Elt: "-no-pgo-warn-mismatch" ); |
2339 | } |
2340 | |
2341 | // Enable sample-based profile guided optimizations. |
2342 | // Sample profile file path --lto-sample-profile=<value>. |
2343 | if (!C.SampleProfile.empty()) { |
2344 | Ops.push_back( |
2345 | Elt: Saver.save(S: "-fprofile-sample-use=" + Twine(C.SampleProfile))); |
2346 | CommonInputs.insert(V: C.SampleProfile); |
2347 | } |
2348 | |
2349 | // We don't know which of options will be used by Clang. |
2350 | Ops.push_back(Elt: "-Wno-unused-command-line-argument" ); |
2351 | |
2352 | // Forward any supplied options. |
2353 | if (!RemoteCompilerArgs.empty()) |
2354 | for (auto &a : RemoteCompilerArgs) |
2355 | Ops.push_back(Elt: a); |
2356 | } |
2357 | |
2358 | // Generates a JSON file describing the backend compilations, for the |
2359 | // distributor. |
2360 | bool emitDistributorJson(StringRef DistributorJson) { |
2361 | using json::Array; |
2362 | std::error_code EC; |
2363 | raw_fd_ostream OS(DistributorJson, EC); |
2364 | if (EC) |
2365 | return false; |
2366 | |
2367 | json::OStream JOS(OS); |
2368 | JOS.object(Contents: [&]() { |
2369 | // Information common to all jobs. |
2370 | JOS.attributeObject(Key: "common" , Contents: [&]() { |
2371 | JOS.attribute(Key: "linker_output" , Contents: LinkerOutputFile); |
2372 | |
2373 | JOS.attributeArray(Key: "args" , Contents: [&]() { |
2374 | JOS.value(V: RemoteCompiler); |
2375 | |
2376 | JOS.value(V: "-c" ); |
2377 | |
2378 | JOS.value(V: Saver.save(S: "--target=" + Triple.str())); |
2379 | |
2380 | for (const auto &A : CodegenOptions) |
2381 | JOS.value(V: A); |
2382 | }); |
2383 | |
2384 | JOS.attribute(Key: "inputs" , Contents: Array(CommonInputs)); |
2385 | }); |
2386 | |
2387 | // Per-compilation-job information. |
2388 | JOS.attributeArray(Key: "jobs" , Contents: [&]() { |
2389 | for (const auto &J : Jobs) { |
2390 | assert(J.Task != 0); |
2391 | |
2392 | SmallVector<StringRef, 2> Inputs; |
2393 | SmallVector<StringRef, 1> Outputs; |
2394 | |
2395 | JOS.object(Contents: [&]() { |
2396 | JOS.attributeArray(Key: "args" , Contents: [&]() { |
2397 | JOS.value(V: J.ModuleID); |
2398 | Inputs.push_back(Elt: J.ModuleID); |
2399 | |
2400 | JOS.value( |
2401 | V: Saver.save(S: "-fthinlto-index=" + Twine(J.SummaryIndexPath))); |
2402 | Inputs.push_back(Elt: J.SummaryIndexPath); |
2403 | |
2404 | JOS.value(V: "-o" ); |
2405 | JOS.value(V: J.NativeObjectPath); |
2406 | Outputs.push_back(Elt: J.NativeObjectPath); |
2407 | }); |
2408 | |
2409 | // Add the bitcode files from which imports will be made. These do |
2410 | // not explicitly appear on the backend compilation command lines |
2411 | // but are recorded in the summary index shards. |
2412 | llvm::append_range(C&: Inputs, R: J.ImportsFiles); |
2413 | JOS.attribute(Key: "inputs" , Contents: Array(Inputs)); |
2414 | |
2415 | JOS.attribute(Key: "outputs" , Contents: Array(Outputs)); |
2416 | }); |
2417 | } |
2418 | }); |
2419 | }); |
2420 | |
2421 | return true; |
2422 | } |
2423 | |
2424 | void removeFile(StringRef FileName) { |
2425 | std::error_code EC = sys::fs::remove(path: FileName, IgnoreNonExisting: true); |
2426 | if (EC && EC != std::make_error_code(e: std::errc::no_such_file_or_directory)) |
2427 | errs() << "warning: could not remove the file '" << FileName |
2428 | << "': " << EC.message() << "\n" ; |
2429 | } |
2430 | |
2431 | Error wait() override { |
2432 | // Wait for the information on the required backend compilations to be |
2433 | // gathered. |
2434 | BackendThreadPool.wait(); |
2435 | if (Err) |
2436 | return std::move(*Err); |
2437 | |
2438 | auto CleanPerJobFiles = llvm::make_scope_exit(F: [&] { |
2439 | if (!SaveTemps) |
2440 | for (auto &Job : Jobs) { |
2441 | removeFile(FileName: Job.NativeObjectPath); |
2442 | if (!ShouldEmitIndexFiles) |
2443 | removeFile(FileName: Job.SummaryIndexPath); |
2444 | } |
2445 | }); |
2446 | |
2447 | const StringRef BCError = "DTLTO backend compilation: " ; |
2448 | |
2449 | buildCommonRemoteCompilerOptions(); |
2450 | |
2451 | SString JsonFile = sys::path::parent_path(path: LinkerOutputFile); |
2452 | sys::path::append(path&: JsonFile, a: sys::path::stem(path: LinkerOutputFile) + "." + UID + |
2453 | ".dist-file.json" ); |
2454 | if (!emitDistributorJson(DistributorJson: JsonFile)) |
2455 | return make_error<StringError>( |
2456 | Args: BCError + "failed to generate distributor JSON script: " + JsonFile, |
2457 | Args: inconvertibleErrorCode()); |
2458 | auto CleanJson = llvm::make_scope_exit(F: [&] { |
2459 | if (!SaveTemps) |
2460 | removeFile(FileName: JsonFile); |
2461 | }); |
2462 | |
2463 | SmallVector<StringRef, 3> Args = {DistributorPath}; |
2464 | llvm::append_range(C&: Args, R&: DistributorArgs); |
2465 | Args.push_back(Elt: JsonFile); |
2466 | std::string ErrMsg; |
2467 | if (sys::ExecuteAndWait(Program: Args[0], Args, |
2468 | /*Env=*/std::nullopt, /*Redirects=*/{}, |
2469 | /*SecondsToWait=*/0, /*MemoryLimit=*/0, ErrMsg: &ErrMsg)) { |
2470 | return make_error<StringError>( |
2471 | Args: BCError + "distributor execution failed" + |
2472 | (!ErrMsg.empty() ? ": " + ErrMsg + Twine("." ) : Twine("." )), |
2473 | Args: inconvertibleErrorCode()); |
2474 | } |
2475 | |
2476 | for (auto &Job : Jobs) { |
2477 | // Load the native object from a file into a memory buffer |
2478 | // and store its contents in the output buffer. |
2479 | auto ObjFileMbOrErr = |
2480 | MemoryBuffer::getFile(Filename: Job.NativeObjectPath, /*IsText=*/false, |
2481 | /*RequiresNullTerminator=*/false); |
2482 | if (std::error_code EC = ObjFileMbOrErr.getError()) |
2483 | return make_error<StringError>( |
2484 | Args: BCError + "cannot open native object file: " + |
2485 | Job.NativeObjectPath + ": " + EC.message(), |
2486 | Args: inconvertibleErrorCode()); |
2487 | auto StreamOrErr = AddStream(Job.Task, Job.ModuleID); |
2488 | if (Error Err = StreamOrErr.takeError()) |
2489 | report_fatal_error(Err: std::move(Err)); |
2490 | auto &Stream = *StreamOrErr->get(); |
2491 | *Stream.OS << ObjFileMbOrErr->get()->getMemBufferRef().getBuffer(); |
2492 | if (Error Err = Stream.commit()) |
2493 | report_fatal_error(Err: std::move(Err)); |
2494 | } |
2495 | |
2496 | return Error::success(); |
2497 | } |
2498 | }; |
2499 | } // end anonymous namespace |
2500 | |
2501 | ThinBackend lto::createOutOfProcessThinBackend( |
2502 | ThreadPoolStrategy Parallelism, lto::IndexWriteCallback OnWrite, |
2503 | bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles, |
2504 | StringRef LinkerOutputFile, StringRef Distributor, |
2505 | ArrayRef<StringRef> DistributorArgs, StringRef RemoteCompiler, |
2506 | ArrayRef<StringRef> RemoteCompilerArgs, bool SaveTemps) { |
2507 | auto Func = |
2508 | [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex, |
2509 | const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
2510 | AddStreamFn AddStream, FileCache /*Cache*/) { |
2511 | return std::make_unique<OutOfProcessThinBackend>( |
2512 | args: Conf, args&: CombinedIndex, args: Parallelism, args: ModuleToDefinedGVSummaries, |
2513 | args&: AddStream, args: OnWrite, args: ShouldEmitIndexFiles, args: ShouldEmitImportsFiles, |
2514 | args: LinkerOutputFile, args: Distributor, args: DistributorArgs, args: RemoteCompiler, |
2515 | args: RemoteCompilerArgs, args: SaveTemps); |
2516 | }; |
2517 | return ThinBackend(Func, Parallelism); |
2518 | } |
2519 | |