1 //===-LTO.h - LLVM Link Time Optimizer ------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file declares functions and classes used to support LTO. It is intended
10 // to be used both by LTO classes as well as by clients (gold-plugin) that
11 // don't utilize the LTO code generator interfaces.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_LTO_LTO_H
16 #define LLVM_LTO_LTO_H
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/StringSet.h"
21 #include "llvm/IR/DiagnosticInfo.h"
22 #include "llvm/IR/ModuleSummaryIndex.h"
23 #include "llvm/IR/RemarkStreamer.h"
24 #include "llvm/LTO/Config.h"
25 #include "llvm/Linker/IRMover.h"
26 #include "llvm/Object/IRSymtab.h"
27 #include "llvm/Support/Error.h"
28 #include "llvm/Support/ToolOutputFile.h"
29 #include "llvm/Support/thread.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Transforms/IPO/FunctionImport.h"
38 class MemoryBufferRef;
41 class raw_pwrite_stream;
43 /// Resolve linkage for prevailing symbols in the \p Index. Linkage changes
44 /// recorded in the index and the ThinLTO backends must apply the changes to
45 /// the module via thinLTOResolvePrevailingInModule.
47 /// This is done for correctness (if value exported, ensure we always
48 /// emit a copy), and compile-time optimization (allow drop of duplicates).
49 void thinLTOResolvePrevailingInIndex(
50 ModuleSummaryIndex &Index,
51 function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)>
53 function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)>
55 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols);
57 /// Update the linkages in the given \p Index to mark exported values
58 /// as external and non-exported values as internal. The ThinLTO backends
59 /// must apply the changes to the Module via thinLTOInternalizeModule.
60 void thinLTOInternalizeAndPromoteInIndex(
61 ModuleSummaryIndex &Index,
62 function_ref<bool(StringRef, GlobalValue::GUID)> isExported);
64 /// Computes a unique hash for the Module considering the current list of
65 /// export/import and other global analysis results.
66 /// The hash is produced in \p Key.
67 void computeLTOCacheKey(
68 SmallString<40> &Key, const lto::Config &Conf,
69 const ModuleSummaryIndex &Index, StringRef ModuleID,
70 const FunctionImporter::ImportMapTy &ImportList,
71 const FunctionImporter::ExportSetTy &ExportList,
72 const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
73 const GVSummaryMapTy &DefinedGlobals,
74 const std::set<GlobalValue::GUID> &CfiFunctionDefs = {},
75 const std::set<GlobalValue::GUID> &CfiFunctionDecls = {});
79 /// Given the original \p Path to an output file, replace any path
80 /// prefix matching \p OldPrefix with \p NewPrefix. Also, create the
81 /// resulting directory if it does not yet exist.
82 std::string getThinLTOOutputFile(const std::string &Path,
83 const std::string &OldPrefix,
84 const std::string &NewPrefix);
86 /// Setup optimization remarks.
87 Expected<std::unique_ptr<ToolOutputFile>>
88 setupOptimizationRemarks(LLVMContext &Context, StringRef RemarksFilename,
89 StringRef RemarksPasses, StringRef RemarksFormat,
90 bool RemarksWithHotness, int Count = -1);
92 /// Setups the output file for saving statistics.
93 Expected<std::unique_ptr<ToolOutputFile>>
94 setupStatsFile(StringRef StatsFilename);
97 struct SymbolResolution;
98 class ThinBackendProc;
100 /// An input file. This is a symbol table wrapper that only exposes the
101 /// information that an LTO client should need in order to do symbol resolution.
107 // FIXME: Remove LTO class friendship once we have bitcode symbol tables.
109 InputFile() = default;
111 std::vector<BitcodeModule> Mods;
112 SmallVector<char, 0> Strtab;
113 std::vector<Symbol> Symbols;
115 // [begin, end) for each module
116 std::vector<std::pair<size_t, size_t>> ModuleSymIndices;
118 StringRef TargetTriple, SourceFileName, COFFLinkerOpts;
119 std::vector<StringRef> DependentLibraries;
120 std::vector<StringRef> ComdatTable;
125 /// Create an InputFile.
126 static Expected<std::unique_ptr<InputFile>> create(MemoryBufferRef Object);
128 /// The purpose of this class is to only expose the symbol information that an
129 /// LTO client should need in order to do symbol resolution.
130 class Symbol : irsymtab::Symbol {
134 Symbol(const irsymtab::Symbol &S) : irsymtab::Symbol(S) {}
136 using irsymtab::Symbol::isUndefined;
137 using irsymtab::Symbol::isCommon;
138 using irsymtab::Symbol::isWeak;
139 using irsymtab::Symbol::isIndirect;
140 using irsymtab::Symbol::getName;
141 using irsymtab::Symbol::getIRName;
142 using irsymtab::Symbol::getVisibility;
143 using irsymtab::Symbol::canBeOmittedFromSymbolTable;
144 using irsymtab::Symbol::isTLS;
145 using irsymtab::Symbol::getComdatIndex;
146 using irsymtab::Symbol::getCommonSize;
147 using irsymtab::Symbol::getCommonAlignment;
148 using irsymtab::Symbol::getCOFFWeakExternalFallback;
149 using irsymtab::Symbol::getSectionName;
150 using irsymtab::Symbol::isExecutable;
151 using irsymtab::Symbol::isUsed;
154 /// A range over the symbols in this InputFile.
155 ArrayRef<Symbol> symbols() const { return Symbols; }
157 /// Returns linker options specified in the input file.
158 StringRef getCOFFLinkerOpts() const { return COFFLinkerOpts; }
160 /// Returns dependent library specifiers from the input file.
161 ArrayRef<StringRef> getDependentLibraries() const { return DependentLibraries; }
163 /// Returns the path to the InputFile.
164 StringRef getName() const;
166 /// Returns the input file's target triple.
167 StringRef getTargetTriple() const { return TargetTriple; }
169 /// Returns the source file path specified at compile time.
170 StringRef getSourceFileName() const { return SourceFileName; }
172 // Returns a table with all the comdats used by this file.
173 ArrayRef<StringRef> getComdatTable() const { return ComdatTable; }
175 // Returns the only BitcodeModule from InputFile.
176 BitcodeModule &getSingleBitcodeModule();
179 ArrayRef<Symbol> module_symbols(unsigned I) const {
180 const auto &Indices = ModuleSymIndices[I];
181 return {Symbols.data() + Indices.first, Symbols.data() + Indices.second};
185 /// This class wraps an output stream for a native object. Most clients should
186 /// just be able to return an instance of this base class from the stream
187 /// callback, but if a client needs to perform some action after the stream is
188 /// written to, that can be done by deriving from this class and overriding the
190 class NativeObjectStream {
192 NativeObjectStream(std::unique_ptr<raw_pwrite_stream> OS) : OS(std::move(OS)) {}
193 std::unique_ptr<raw_pwrite_stream> OS;
194 virtual ~NativeObjectStream() = default;
197 /// This type defines the callback to add a native object that is generated on
200 /// Stream callbacks must be thread safe.
202 std::function<std::unique_ptr<NativeObjectStream>(unsigned Task)>;
204 /// This is the type of a native object cache. To request an item from the
205 /// cache, pass a unique string as the Key. For hits, the cached file will be
206 /// added to the link and this function will return AddStreamFn(). For misses,
207 /// the cache will return a stream callback which must be called at most once to
208 /// produce content for the stream. The native object stream produced by the
209 /// stream callback will add the file to the link after the stream is written
212 /// Clients generally look like this:
214 /// if (AddStreamFn AddStream = Cache(Task, Key))
215 /// ProduceContent(AddStream);
216 using NativeObjectCache =
217 std::function<AddStreamFn(unsigned Task, StringRef Key)>;
219 /// A ThinBackend defines what happens after the thin-link phase during ThinLTO.
220 /// The details of this type definition aren't important; clients can only
221 /// create a ThinBackend using one of the create*ThinBackend() functions below.
222 using ThinBackend = std::function<std::unique_ptr<ThinBackendProc>(
223 Config &C, ModuleSummaryIndex &CombinedIndex,
224 StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
225 AddStreamFn AddStream, NativeObjectCache Cache)>;
227 /// This ThinBackend runs the individual backend jobs in-process.
228 ThinBackend createInProcessThinBackend(unsigned ParallelismLevel);
230 /// This ThinBackend writes individual module indexes to files, instead of
231 /// running the individual backend jobs. This backend is for distributed builds
232 /// where separate processes will invoke the real backends.
234 /// To find the path to write the index to, the backend checks if the path has a
235 /// prefix of OldPrefix; if so, it replaces that prefix with NewPrefix. It then
236 /// appends ".thinlto.bc" and writes the index to that path. If
237 /// ShouldEmitImportsFiles is true it also writes a list of imported files to a
238 /// similar path with ".imports" appended instead.
239 /// LinkedObjectsFile is an output stream to write the list of object files for
240 /// the final ThinLTO linking. Can be nullptr.
241 /// OnWrite is callback which receives module identifier and notifies LTO user
242 /// that index file for the module (and optionally imports file) was created.
243 using IndexWriteCallback = std::function<void(const std::string &)>;
244 ThinBackend createWriteIndexesThinBackend(std::string OldPrefix,
245 std::string NewPrefix,
246 bool ShouldEmitImportsFiles,
247 raw_fd_ostream *LinkedObjectsFile,
248 IndexWriteCallback OnWrite);
250 /// This class implements a resolution-based interface to LLVM's LTO
251 /// functionality. It supports regular LTO, parallel LTO code generation and
252 /// ThinLTO. You can use it from a linker in the following way:
253 /// - Set hooks and code generation options (see lto::Config struct defined in
254 /// Config.h), and use the lto::Config object to create an lto::LTO object.
255 /// - Create lto::InputFile objects using lto::InputFile::create(), then use
256 /// the symbols() function to enumerate its symbols and compute a resolution
257 /// for each symbol (see SymbolResolution below).
258 /// - After the linker has visited each input file (and each regular object
259 /// file) and computed a resolution for each symbol, take each lto::InputFile
260 /// and pass it and an array of symbol resolutions to the add() function.
261 /// - Call the getMaxTasks() function to get an upper bound on the number of
262 /// native object files that LTO may add to the link.
263 /// - Call the run() function. This function will use the supplied AddStream
264 /// and Cache functions to add up to getMaxTasks() native object files to
270 /// Create an LTO object. A default constructed LTO object has a reasonable
271 /// production configuration, but you can customize it by passing arguments to
272 /// this constructor.
273 /// FIXME: We do currently require the DiagHandler field to be set in Conf.
274 /// Until that is fixed, a Config argument is required.
275 LTO(Config Conf, ThinBackend Backend = nullptr,
276 unsigned ParallelCodeGenParallelismLevel = 1);
279 /// Add an input file to the LTO link, using the provided symbol resolutions.
280 /// The symbol resolutions must appear in the enumeration order given by
281 /// InputFile::symbols().
282 Error add(std::unique_ptr<InputFile> Obj, ArrayRef<SymbolResolution> Res);
284 /// Returns an upper bound on the number of tasks that the client may expect.
285 /// This may only be called after all IR object files have been added. For a
286 /// full description of tasks see LTOBackend.h.
287 unsigned getMaxTasks() const;
289 /// Runs the LTO pipeline. This function calls the supplied AddStream
290 /// function to add native object files to the link.
292 /// The Cache parameter is optional. If supplied, it will be used to cache
293 /// native object files and add them to the link.
295 /// The client will receive at most one callback (via either AddStream or
296 /// Cache) for each task identifier.
297 Error run(AddStreamFn AddStream, NativeObjectCache Cache = nullptr);
302 struct RegularLTOState {
303 RegularLTOState(unsigned ParallelCodeGenParallelismLevel, Config &Conf);
304 struct CommonResolution {
307 /// Record if at least one instance of the common was marked as prevailing
308 bool Prevailing = false;
310 std::map<std::string, CommonResolution> Commons;
312 unsigned ParallelCodeGenParallelismLevel;
314 std::unique_ptr<Module> CombinedModule;
315 std::unique_ptr<IRMover> Mover;
317 // This stores the information about a regular LTO module that we have added
318 // to the link. It will either be linked immediately (for modules without
319 // summaries) or after summary-based dead stripping (for modules with
322 std::unique_ptr<Module> M;
323 std::vector<GlobalValue *> Keep;
325 std::vector<AddedModule> ModsWithSummaries;
328 struct ThinLTOState {
329 ThinLTOState(ThinBackend Backend);
332 ModuleSummaryIndex CombinedIndex;
333 MapVector<StringRef, BitcodeModule> ModuleMap;
334 DenseMap<GlobalValue::GUID, StringRef> PrevailingModuleForGUID;
337 // The global resolution for a particular (mangled) symbol name. This is in
338 // particular necessary to track whether each symbol can be internalized.
339 // Because any input file may introduce a new cross-partition reference, we
340 // cannot make any final internalization decisions until all input files have
341 // been added and the client has called run(). During run() we apply
342 // internalization decisions either directly to the module (for regular LTO)
343 // or to the combined index (for ThinLTO).
344 struct GlobalResolution {
345 /// The unmangled name of the global.
348 /// Keep track if the symbol is visible outside of a module with a summary
349 /// (i.e. in either a regular object or a regular LTO module without a
351 bool VisibleOutsideSummary = false;
353 bool UnnamedAddr = true;
355 /// True if module contains the prevailing definition.
356 bool Prevailing = false;
358 /// Returns true if module contains the prevailing definition and symbol is
359 /// an IR symbol. For example when module-level inline asm block is used,
360 /// symbol can be prevailing in module but have no IR name.
361 bool isPrevailingIRSymbol() const { return Prevailing && !IRName.empty(); }
363 /// This field keeps track of the partition number of this global. The
364 /// regular LTO object is partition 0, while each ThinLTO object has its own
365 /// partition number from 1 onwards.
367 /// Any global that is defined or used by more than one partition, or that
368 /// is referenced externally, may not be internalized.
370 /// Partitions generally have a one-to-one correspondence with tasks, except
371 /// that we use partition 0 for all parallel LTO code generation partitions.
372 /// Any partitioning of the combined LTO object is done internally by the
374 unsigned Partition = Unknown;
376 /// Special partition numbers.
378 /// A partition number has not yet been assigned to this global.
381 /// This global is either used by more than one partition or has an
382 /// external reference, and therefore cannot be internalized.
385 /// The RegularLTO partition
390 // Global mapping from mangled symbol names to resolutions.
391 StringMap<GlobalResolution> GlobalResolutions;
393 void addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
394 ArrayRef<SymbolResolution> Res, unsigned Partition,
397 // These functions take a range of symbol resolutions [ResI, ResE) and consume
398 // the resolutions used by a single input module by incrementing ResI. After
399 // these functions return, [ResI, ResE) will refer to the resolution range for
400 // the remaining modules in the InputFile.
401 Error addModule(InputFile &Input, unsigned ModI,
402 const SymbolResolution *&ResI, const SymbolResolution *ResE);
404 Expected<RegularLTOState::AddedModule>
405 addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
406 const SymbolResolution *&ResI, const SymbolResolution *ResE);
407 Error linkRegularLTO(RegularLTOState::AddedModule Mod,
408 bool LivenessFromIndex);
410 Error addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
411 const SymbolResolution *&ResI, const SymbolResolution *ResE);
413 Error runRegularLTO(AddStreamFn AddStream);
414 Error runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache,
415 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols);
417 Error checkPartiallySplit();
419 mutable bool CalledGetMaxTasks = false;
421 // Use Optional to distinguish false from not yet initialized.
422 Optional<bool> EnableSplitLTOUnit;
425 /// The resolution for a symbol. The linker must provide a SymbolResolution for
426 /// each global symbol based on its internal resolution of that symbol.
427 struct SymbolResolution {
429 : Prevailing(0), FinalDefinitionInLinkageUnit(0), VisibleToRegularObj(0),
430 LinkerRedefined(0) {}
432 /// The linker has chosen this definition of the symbol.
433 unsigned Prevailing : 1;
435 /// The definition of this symbol is unpreemptable at runtime and is known to
436 /// be in this linkage unit.
437 unsigned FinalDefinitionInLinkageUnit : 1;
439 /// The definition of this symbol is visible outside of the LTO unit.
440 unsigned VisibleToRegularObj : 1;
442 /// Linker redefined version of the symbol which appeared in -wrap or -defsym
444 unsigned LinkerRedefined : 1;