Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 1 | //===-LTO.h - LLVM Link Time Optimizer ------------------------------------===// |
| 2 | // |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame^] | 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 |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 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. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #ifndef LLVM_LTO_LTO_H |
| 16 | #define LLVM_LTO_LTO_H |
| 17 | |
| 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/LTO/Config.h" |
| 24 | #include "llvm/Linker/IRMover.h" |
| 25 | #include "llvm/Object/IRSymtab.h" |
| 26 | #include "llvm/Support/Error.h" |
| 27 | #include "llvm/Support/ToolOutputFile.h" |
| 28 | #include "llvm/Support/thread.h" |
| 29 | #include "llvm/Target/TargetOptions.h" |
| 30 | #include "llvm/Transforms/IPO/FunctionImport.h" |
| 31 | |
| 32 | namespace llvm { |
| 33 | |
| 34 | class BitcodeModule; |
| 35 | class Error; |
| 36 | class LLVMContext; |
| 37 | class MemoryBufferRef; |
| 38 | class Module; |
| 39 | class Target; |
| 40 | class raw_pwrite_stream; |
| 41 | |
Andrew Scull | 0372a57 | 2018-11-16 15:47:06 +0000 | [diff] [blame] | 42 | /// Resolve linkage for prevailing symbols in the \p Index. Linkage changes |
| 43 | /// recorded in the index and the ThinLTO backends must apply the changes to |
| 44 | /// the module via thinLTOResolvePrevailingInModule. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 45 | /// |
| 46 | /// This is done for correctness (if value exported, ensure we always |
| 47 | /// emit a copy), and compile-time optimization (allow drop of duplicates). |
Andrew Scull | 0372a57 | 2018-11-16 15:47:06 +0000 | [diff] [blame] | 48 | void thinLTOResolvePrevailingInIndex( |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 49 | ModuleSummaryIndex &Index, |
| 50 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 51 | isPrevailing, |
| 52 | function_ref<void(StringRef, GlobalValue::GUID, GlobalValue::LinkageTypes)> |
| 53 | recordNewLinkage); |
| 54 | |
| 55 | /// Update the linkages in the given \p Index to mark exported values |
| 56 | /// as external and non-exported values as internal. The ThinLTO backends |
| 57 | /// must apply the changes to the Module via thinLTOInternalizeModule. |
| 58 | void thinLTOInternalizeAndPromoteInIndex( |
| 59 | ModuleSummaryIndex &Index, |
| 60 | function_ref<bool(StringRef, GlobalValue::GUID)> isExported); |
| 61 | |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame^] | 62 | /// Computes a unique hash for the Module considering the current list of |
| 63 | /// export/import and other global analysis results. |
| 64 | /// The hash is produced in \p Key. |
| 65 | void computeLTOCacheKey( |
| 66 | SmallString<40> &Key, const lto::Config &Conf, |
| 67 | const ModuleSummaryIndex &Index, StringRef ModuleID, |
| 68 | const FunctionImporter::ImportMapTy &ImportList, |
| 69 | const FunctionImporter::ExportSetTy &ExportList, |
| 70 | const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR, |
| 71 | const GVSummaryMapTy &DefinedGlobals, |
| 72 | const std::set<GlobalValue::GUID> &CfiFunctionDefs = {}, |
| 73 | const std::set<GlobalValue::GUID> &CfiFunctionDecls = {}); |
| 74 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 75 | namespace lto { |
| 76 | |
| 77 | /// Given the original \p Path to an output file, replace any path |
| 78 | /// prefix matching \p OldPrefix with \p NewPrefix. Also, create the |
| 79 | /// resulting directory if it does not yet exist. |
| 80 | std::string getThinLTOOutputFile(const std::string &Path, |
| 81 | const std::string &OldPrefix, |
| 82 | const std::string &NewPrefix); |
| 83 | |
| 84 | /// Setup optimization remarks. |
| 85 | Expected<std::unique_ptr<ToolOutputFile>> |
| 86 | setupOptimizationRemarks(LLVMContext &Context, StringRef LTORemarksFilename, |
| 87 | bool LTOPassRemarksWithHotness, int Count = -1); |
| 88 | |
| 89 | class LTO; |
| 90 | struct SymbolResolution; |
| 91 | class ThinBackendProc; |
| 92 | |
| 93 | /// An input file. This is a symbol table wrapper that only exposes the |
| 94 | /// information that an LTO client should need in order to do symbol resolution. |
| 95 | class InputFile { |
| 96 | public: |
| 97 | class Symbol; |
| 98 | |
| 99 | private: |
| 100 | // FIXME: Remove LTO class friendship once we have bitcode symbol tables. |
| 101 | friend LTO; |
| 102 | InputFile() = default; |
| 103 | |
| 104 | std::vector<BitcodeModule> Mods; |
| 105 | SmallVector<char, 0> Strtab; |
| 106 | std::vector<Symbol> Symbols; |
| 107 | |
| 108 | // [begin, end) for each module |
| 109 | std::vector<std::pair<size_t, size_t>> ModuleSymIndices; |
| 110 | |
| 111 | StringRef TargetTriple, SourceFileName, COFFLinkerOpts; |
| 112 | std::vector<StringRef> ComdatTable; |
| 113 | |
| 114 | public: |
| 115 | ~InputFile(); |
| 116 | |
| 117 | /// Create an InputFile. |
| 118 | static Expected<std::unique_ptr<InputFile>> create(MemoryBufferRef Object); |
| 119 | |
| 120 | /// The purpose of this class is to only expose the symbol information that an |
| 121 | /// LTO client should need in order to do symbol resolution. |
| 122 | class Symbol : irsymtab::Symbol { |
| 123 | friend LTO; |
| 124 | |
| 125 | public: |
| 126 | Symbol(const irsymtab::Symbol &S) : irsymtab::Symbol(S) {} |
| 127 | |
| 128 | using irsymtab::Symbol::isUndefined; |
| 129 | using irsymtab::Symbol::isCommon; |
| 130 | using irsymtab::Symbol::isWeak; |
| 131 | using irsymtab::Symbol::isIndirect; |
| 132 | using irsymtab::Symbol::getName; |
| 133 | using irsymtab::Symbol::getVisibility; |
| 134 | using irsymtab::Symbol::canBeOmittedFromSymbolTable; |
| 135 | using irsymtab::Symbol::isTLS; |
| 136 | using irsymtab::Symbol::getComdatIndex; |
| 137 | using irsymtab::Symbol::getCommonSize; |
| 138 | using irsymtab::Symbol::getCommonAlignment; |
| 139 | using irsymtab::Symbol::getCOFFWeakExternalFallback; |
| 140 | using irsymtab::Symbol::getSectionName; |
| 141 | using irsymtab::Symbol::isExecutable; |
| 142 | }; |
| 143 | |
| 144 | /// A range over the symbols in this InputFile. |
| 145 | ArrayRef<Symbol> symbols() const { return Symbols; } |
| 146 | |
| 147 | /// Returns linker options specified in the input file. |
| 148 | StringRef getCOFFLinkerOpts() const { return COFFLinkerOpts; } |
| 149 | |
| 150 | /// Returns the path to the InputFile. |
| 151 | StringRef getName() const; |
| 152 | |
| 153 | /// Returns the input file's target triple. |
| 154 | StringRef getTargetTriple() const { return TargetTriple; } |
| 155 | |
| 156 | /// Returns the source file path specified at compile time. |
| 157 | StringRef getSourceFileName() const { return SourceFileName; } |
| 158 | |
| 159 | // Returns a table with all the comdats used by this file. |
| 160 | ArrayRef<StringRef> getComdatTable() const { return ComdatTable; } |
| 161 | |
| 162 | private: |
| 163 | ArrayRef<Symbol> module_symbols(unsigned I) const { |
| 164 | const auto &Indices = ModuleSymIndices[I]; |
| 165 | return {Symbols.data() + Indices.first, Symbols.data() + Indices.second}; |
| 166 | } |
| 167 | }; |
| 168 | |
| 169 | /// This class wraps an output stream for a native object. Most clients should |
| 170 | /// just be able to return an instance of this base class from the stream |
| 171 | /// callback, but if a client needs to perform some action after the stream is |
| 172 | /// written to, that can be done by deriving from this class and overriding the |
| 173 | /// destructor. |
| 174 | class NativeObjectStream { |
| 175 | public: |
| 176 | NativeObjectStream(std::unique_ptr<raw_pwrite_stream> OS) : OS(std::move(OS)) {} |
| 177 | std::unique_ptr<raw_pwrite_stream> OS; |
| 178 | virtual ~NativeObjectStream() = default; |
| 179 | }; |
| 180 | |
| 181 | /// This type defines the callback to add a native object that is generated on |
| 182 | /// the fly. |
| 183 | /// |
| 184 | /// Stream callbacks must be thread safe. |
| 185 | typedef std::function<std::unique_ptr<NativeObjectStream>(unsigned Task)> |
| 186 | AddStreamFn; |
| 187 | |
| 188 | /// This is the type of a native object cache. To request an item from the |
| 189 | /// cache, pass a unique string as the Key. For hits, the cached file will be |
| 190 | /// added to the link and this function will return AddStreamFn(). For misses, |
| 191 | /// the cache will return a stream callback which must be called at most once to |
| 192 | /// produce content for the stream. The native object stream produced by the |
| 193 | /// stream callback will add the file to the link after the stream is written |
| 194 | /// to. |
| 195 | /// |
| 196 | /// Clients generally look like this: |
| 197 | /// |
| 198 | /// if (AddStreamFn AddStream = Cache(Task, Key)) |
| 199 | /// ProduceContent(AddStream); |
| 200 | typedef std::function<AddStreamFn(unsigned Task, StringRef Key)> |
| 201 | NativeObjectCache; |
| 202 | |
| 203 | /// A ThinBackend defines what happens after the thin-link phase during ThinLTO. |
| 204 | /// The details of this type definition aren't important; clients can only |
| 205 | /// create a ThinBackend using one of the create*ThinBackend() functions below. |
| 206 | typedef std::function<std::unique_ptr<ThinBackendProc>( |
| 207 | Config &C, ModuleSummaryIndex &CombinedIndex, |
| 208 | StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries, |
| 209 | AddStreamFn AddStream, NativeObjectCache Cache)> |
| 210 | ThinBackend; |
| 211 | |
| 212 | /// This ThinBackend runs the individual backend jobs in-process. |
| 213 | ThinBackend createInProcessThinBackend(unsigned ParallelismLevel); |
| 214 | |
| 215 | /// This ThinBackend writes individual module indexes to files, instead of |
| 216 | /// running the individual backend jobs. This backend is for distributed builds |
| 217 | /// where separate processes will invoke the real backends. |
| 218 | /// |
| 219 | /// To find the path to write the index to, the backend checks if the path has a |
| 220 | /// prefix of OldPrefix; if so, it replaces that prefix with NewPrefix. It then |
| 221 | /// appends ".thinlto.bc" and writes the index to that path. If |
| 222 | /// ShouldEmitImportsFiles is true it also writes a list of imported files to a |
| 223 | /// similar path with ".imports" appended instead. |
| 224 | /// LinkedObjectsFile is an output stream to write the list of object files for |
| 225 | /// the final ThinLTO linking. Can be nullptr. |
| 226 | /// OnWrite is callback which receives module identifier and notifies LTO user |
| 227 | /// that index file for the module (and optionally imports file) was created. |
| 228 | using IndexWriteCallback = std::function<void(const std::string &)>; |
| 229 | ThinBackend createWriteIndexesThinBackend(std::string OldPrefix, |
| 230 | std::string NewPrefix, |
| 231 | bool ShouldEmitImportsFiles, |
| 232 | raw_fd_ostream *LinkedObjectsFile, |
| 233 | IndexWriteCallback OnWrite); |
| 234 | |
| 235 | /// This class implements a resolution-based interface to LLVM's LTO |
| 236 | /// functionality. It supports regular LTO, parallel LTO code generation and |
| 237 | /// ThinLTO. You can use it from a linker in the following way: |
| 238 | /// - Set hooks and code generation options (see lto::Config struct defined in |
| 239 | /// Config.h), and use the lto::Config object to create an lto::LTO object. |
| 240 | /// - Create lto::InputFile objects using lto::InputFile::create(), then use |
| 241 | /// the symbols() function to enumerate its symbols and compute a resolution |
| 242 | /// for each symbol (see SymbolResolution below). |
| 243 | /// - After the linker has visited each input file (and each regular object |
| 244 | /// file) and computed a resolution for each symbol, take each lto::InputFile |
| 245 | /// and pass it and an array of symbol resolutions to the add() function. |
| 246 | /// - Call the getMaxTasks() function to get an upper bound on the number of |
| 247 | /// native object files that LTO may add to the link. |
| 248 | /// - Call the run() function. This function will use the supplied AddStream |
| 249 | /// and Cache functions to add up to getMaxTasks() native object files to |
| 250 | /// the link. |
| 251 | class LTO { |
| 252 | friend InputFile; |
| 253 | |
| 254 | public: |
| 255 | /// Create an LTO object. A default constructed LTO object has a reasonable |
| 256 | /// production configuration, but you can customize it by passing arguments to |
| 257 | /// this constructor. |
| 258 | /// FIXME: We do currently require the DiagHandler field to be set in Conf. |
| 259 | /// Until that is fixed, a Config argument is required. |
| 260 | LTO(Config Conf, ThinBackend Backend = nullptr, |
| 261 | unsigned ParallelCodeGenParallelismLevel = 1); |
| 262 | ~LTO(); |
| 263 | |
| 264 | /// Add an input file to the LTO link, using the provided symbol resolutions. |
| 265 | /// The symbol resolutions must appear in the enumeration order given by |
| 266 | /// InputFile::symbols(). |
| 267 | Error add(std::unique_ptr<InputFile> Obj, ArrayRef<SymbolResolution> Res); |
| 268 | |
| 269 | /// Returns an upper bound on the number of tasks that the client may expect. |
| 270 | /// This may only be called after all IR object files have been added. For a |
| 271 | /// full description of tasks see LTOBackend.h. |
| 272 | unsigned getMaxTasks() const; |
| 273 | |
| 274 | /// Runs the LTO pipeline. This function calls the supplied AddStream |
| 275 | /// function to add native object files to the link. |
| 276 | /// |
| 277 | /// The Cache parameter is optional. If supplied, it will be used to cache |
| 278 | /// native object files and add them to the link. |
| 279 | /// |
| 280 | /// The client will receive at most one callback (via either AddStream or |
| 281 | /// Cache) for each task identifier. |
| 282 | Error run(AddStreamFn AddStream, NativeObjectCache Cache = nullptr); |
| 283 | |
| 284 | private: |
| 285 | Config Conf; |
| 286 | |
| 287 | struct RegularLTOState { |
| 288 | RegularLTOState(unsigned ParallelCodeGenParallelismLevel, Config &Conf); |
| 289 | struct CommonResolution { |
| 290 | uint64_t Size = 0; |
| 291 | unsigned Align = 0; |
| 292 | /// Record if at least one instance of the common was marked as prevailing |
| 293 | bool Prevailing = false; |
| 294 | }; |
| 295 | std::map<std::string, CommonResolution> Commons; |
| 296 | |
| 297 | unsigned ParallelCodeGenParallelismLevel; |
| 298 | LTOLLVMContext Ctx; |
| 299 | std::unique_ptr<Module> CombinedModule; |
| 300 | std::unique_ptr<IRMover> Mover; |
| 301 | |
| 302 | // This stores the information about a regular LTO module that we have added |
| 303 | // to the link. It will either be linked immediately (for modules without |
| 304 | // summaries) or after summary-based dead stripping (for modules with |
| 305 | // summaries). |
| 306 | struct AddedModule { |
| 307 | std::unique_ptr<Module> M; |
| 308 | std::vector<GlobalValue *> Keep; |
| 309 | }; |
| 310 | std::vector<AddedModule> ModsWithSummaries; |
| 311 | } RegularLTO; |
| 312 | |
| 313 | struct ThinLTOState { |
| 314 | ThinLTOState(ThinBackend Backend); |
| 315 | |
| 316 | ThinBackend Backend; |
| 317 | ModuleSummaryIndex CombinedIndex; |
| 318 | MapVector<StringRef, BitcodeModule> ModuleMap; |
| 319 | DenseMap<GlobalValue::GUID, StringRef> PrevailingModuleForGUID; |
| 320 | } ThinLTO; |
| 321 | |
| 322 | // The global resolution for a particular (mangled) symbol name. This is in |
| 323 | // particular necessary to track whether each symbol can be internalized. |
| 324 | // Because any input file may introduce a new cross-partition reference, we |
| 325 | // cannot make any final internalization decisions until all input files have |
| 326 | // been added and the client has called run(). During run() we apply |
| 327 | // internalization decisions either directly to the module (for regular LTO) |
| 328 | // or to the combined index (for ThinLTO). |
| 329 | struct GlobalResolution { |
| 330 | /// The unmangled name of the global. |
| 331 | std::string IRName; |
| 332 | |
| 333 | /// Keep track if the symbol is visible outside of a module with a summary |
| 334 | /// (i.e. in either a regular object or a regular LTO module without a |
| 335 | /// summary). |
| 336 | bool VisibleOutsideSummary = false; |
| 337 | |
| 338 | bool UnnamedAddr = true; |
| 339 | |
| 340 | /// True if module contains the prevailing definition. |
| 341 | bool Prevailing = false; |
| 342 | |
| 343 | /// Returns true if module contains the prevailing definition and symbol is |
| 344 | /// an IR symbol. For example when module-level inline asm block is used, |
| 345 | /// symbol can be prevailing in module but have no IR name. |
| 346 | bool isPrevailingIRSymbol() const { return Prevailing && !IRName.empty(); } |
| 347 | |
| 348 | /// This field keeps track of the partition number of this global. The |
| 349 | /// regular LTO object is partition 0, while each ThinLTO object has its own |
| 350 | /// partition number from 1 onwards. |
| 351 | /// |
| 352 | /// Any global that is defined or used by more than one partition, or that |
| 353 | /// is referenced externally, may not be internalized. |
| 354 | /// |
| 355 | /// Partitions generally have a one-to-one correspondence with tasks, except |
| 356 | /// that we use partition 0 for all parallel LTO code generation partitions. |
| 357 | /// Any partitioning of the combined LTO object is done internally by the |
| 358 | /// LTO backend. |
| 359 | unsigned Partition = Unknown; |
| 360 | |
| 361 | /// Special partition numbers. |
| 362 | enum : unsigned { |
| 363 | /// A partition number has not yet been assigned to this global. |
| 364 | Unknown = -1u, |
| 365 | |
| 366 | /// This global is either used by more than one partition or has an |
| 367 | /// external reference, and therefore cannot be internalized. |
| 368 | External = -2u, |
| 369 | |
| 370 | /// The RegularLTO partition |
| 371 | RegularLTO = 0, |
| 372 | }; |
| 373 | }; |
| 374 | |
| 375 | // Global mapping from mangled symbol names to resolutions. |
| 376 | StringMap<GlobalResolution> GlobalResolutions; |
| 377 | |
| 378 | void addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms, |
| 379 | ArrayRef<SymbolResolution> Res, unsigned Partition, |
| 380 | bool InSummary); |
| 381 | |
| 382 | // These functions take a range of symbol resolutions [ResI, ResE) and consume |
| 383 | // the resolutions used by a single input module by incrementing ResI. After |
| 384 | // these functions return, [ResI, ResE) will refer to the resolution range for |
| 385 | // the remaining modules in the InputFile. |
| 386 | Error addModule(InputFile &Input, unsigned ModI, |
| 387 | const SymbolResolution *&ResI, const SymbolResolution *ResE); |
| 388 | |
| 389 | Expected<RegularLTOState::AddedModule> |
| 390 | addRegularLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, |
| 391 | const SymbolResolution *&ResI, const SymbolResolution *ResE); |
| 392 | Error linkRegularLTO(RegularLTOState::AddedModule Mod, |
| 393 | bool LivenessFromIndex); |
| 394 | |
| 395 | Error addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, |
| 396 | const SymbolResolution *&ResI, const SymbolResolution *ResE); |
| 397 | |
| 398 | Error runRegularLTO(AddStreamFn AddStream); |
| 399 | Error runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache); |
| 400 | |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame^] | 401 | Error checkPartiallySplit(); |
| 402 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 403 | mutable bool CalledGetMaxTasks = false; |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame^] | 404 | |
| 405 | // Use Optional to distinguish false from not yet initialized. |
| 406 | Optional<bool> EnableSplitLTOUnit; |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 407 | }; |
| 408 | |
| 409 | /// The resolution for a symbol. The linker must provide a SymbolResolution for |
| 410 | /// each global symbol based on its internal resolution of that symbol. |
| 411 | struct SymbolResolution { |
| 412 | SymbolResolution() |
| 413 | : Prevailing(0), FinalDefinitionInLinkageUnit(0), VisibleToRegularObj(0), |
| 414 | LinkerRedefined(0) {} |
| 415 | |
| 416 | /// The linker has chosen this definition of the symbol. |
| 417 | unsigned Prevailing : 1; |
| 418 | |
| 419 | /// The definition of this symbol is unpreemptable at runtime and is known to |
| 420 | /// be in this linkage unit. |
| 421 | unsigned FinalDefinitionInLinkageUnit : 1; |
| 422 | |
| 423 | /// The definition of this symbol is visible outside of the LTO unit. |
| 424 | unsigned VisibleToRegularObj : 1; |
| 425 | |
| 426 | /// Linker redefined version of the symbol which appeared in -wrap or -defsym |
| 427 | /// linker option. |
| 428 | unsigned LinkerRedefined : 1; |
| 429 | }; |
| 430 | |
| 431 | } // namespace lto |
| 432 | } // namespace llvm |
| 433 | |
| 434 | #endif |