ThinLTOCodeGenerator.h revision 327952
1//===-ThinLTOCodeGenerator.h - LLVM Link Time Optimizer -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file declares the ThinLTOCodeGenerator class, similar to the 11// LTOCodeGenerator but for the ThinLTO scheme. It provides an interface for 12// linker plugin. 13// 14//===----------------------------------------------------------------------===// 15 16#ifndef LLVM_LTO_THINLTOCODEGENERATOR_H 17#define LLVM_LTO_THINLTOCODEGENERATOR_H 18 19#include "llvm-c/lto.h" 20#include "llvm/ADT/StringSet.h" 21#include "llvm/ADT/Triple.h" 22#include "llvm/IR/ModuleSummaryIndex.h" 23#include "llvm/Support/CachePruning.h" 24#include "llvm/Support/CodeGen.h" 25#include "llvm/Support/MemoryBuffer.h" 26#include "llvm/Target/TargetOptions.h" 27 28#include <string> 29 30namespace llvm { 31class StringRef; 32class LLVMContext; 33class TargetMachine; 34 35/// Wrapper around MemoryBufferRef, owning the identifier 36class ThinLTOBuffer { 37 std::string OwnedIdentifier; 38 StringRef Buffer; 39 40public: 41 ThinLTOBuffer(StringRef Buffer, StringRef Identifier) 42 : OwnedIdentifier(Identifier), Buffer(Buffer) {} 43 44 MemoryBufferRef getMemBuffer() const { 45 return MemoryBufferRef(Buffer, 46 {OwnedIdentifier.c_str(), OwnedIdentifier.size()}); 47 } 48 StringRef getBuffer() const { return Buffer; } 49 StringRef getBufferIdentifier() const { return OwnedIdentifier; } 50}; 51 52/// Helper to gather options relevant to the target machine creation 53struct TargetMachineBuilder { 54 Triple TheTriple; 55 std::string MCpu; 56 std::string MAttr; 57 TargetOptions Options; 58 Optional<Reloc::Model> RelocModel; 59 CodeGenOpt::Level CGOptLevel = CodeGenOpt::Aggressive; 60 61 std::unique_ptr<TargetMachine> create() const; 62}; 63 64/// This class define an interface similar to the LTOCodeGenerator, but adapted 65/// for ThinLTO processing. 66/// The ThinLTOCodeGenerator is not intended to be reuse for multiple 67/// compilation: the model is that the client adds modules to the generator and 68/// ask to perform the ThinLTO optimizations / codegen, and finally destroys the 69/// codegenerator. 70class ThinLTOCodeGenerator { 71public: 72 /// Add given module to the code generator. 73 void addModule(StringRef Identifier, StringRef Data); 74 75 /** 76 * Adds to a list of all global symbols that must exist in the final generated 77 * code. If a symbol is not listed there, it will be optimized away if it is 78 * inlined into every usage. 79 */ 80 void preserveSymbol(StringRef Name); 81 82 /** 83 * Adds to a list of all global symbols that are cross-referenced between 84 * ThinLTO files. If the ThinLTO CodeGenerator can ensure that every 85 * references from a ThinLTO module to this symbol is optimized away, then 86 * the symbol can be discarded. 87 */ 88 void crossReferenceSymbol(StringRef Name); 89 90 /** 91 * Process all the modules that were added to the code generator in parallel. 92 * 93 * Client can access the resulting object files using getProducedBinaries(), 94 * unless setGeneratedObjectsDirectory() has been called, in which case 95 * results are available through getProducedBinaryFiles(). 96 */ 97 void run(); 98 99 /** 100 * Return the "in memory" binaries produced by the code generator. This is 101 * filled after run() unless setGeneratedObjectsDirectory() has been 102 * called, in which case results are available through 103 * getProducedBinaryFiles(). 104 */ 105 std::vector<std::unique_ptr<MemoryBuffer>> &getProducedBinaries() { 106 return ProducedBinaries; 107 } 108 109 /** 110 * Return the "on-disk" binaries produced by the code generator. This is 111 * filled after run() when setGeneratedObjectsDirectory() has been 112 * called, in which case results are available through getProducedBinaries(). 113 */ 114 std::vector<std::string> &getProducedBinaryFiles() { 115 return ProducedBinaryFiles; 116 } 117 118 /** 119 * \defgroup Options setters 120 * @{ 121 */ 122 123 /** 124 * \defgroup Cache controlling options 125 * 126 * These entry points control the ThinLTO cache. The cache is intended to 127 * support incremental build, and thus needs to be persistent accross build. 128 * The client enabled the cache by supplying a path to an existing directory. 129 * The code generator will use this to store objects files that may be reused 130 * during a subsequent build. 131 * To avoid filling the disk space, a few knobs are provided: 132 * - The pruning interval limit the frequency at which the garbage collector 133 * will try to scan the cache directory to prune it from expired entries. 134 * Setting to -1 disable the pruning (default). 135 * - The pruning expiration time indicates to the garbage collector how old 136 * an entry needs to be to be removed. 137 * - Finally, the garbage collector can be instructed to prune the cache till 138 * the occupied space goes below a threshold. 139 * @{ 140 */ 141 142 struct CachingOptions { 143 std::string Path; // Path to the cache, empty to disable. 144 CachePruningPolicy Policy; 145 }; 146 147 /// Provide a path to a directory where to store the cached files for 148 /// incremental build. 149 void setCacheDir(std::string Path) { CacheOptions.Path = std::move(Path); } 150 151 /// Cache policy: interval (seconds) between two prunes of the cache. Set to a 152 /// negative value to disable pruning. A value of 0 will be ignored. 153 void setCachePruningInterval(int Interval) { 154 if (Interval == 0) 155 return; 156 if(Interval < 0) 157 CacheOptions.Policy.Interval.reset(); 158 else 159 CacheOptions.Policy.Interval = std::chrono::seconds(Interval); 160 } 161 162 /// Cache policy: expiration (in seconds) for an entry. 163 /// A value of 0 will be ignored. 164 void setCacheEntryExpiration(unsigned Expiration) { 165 if (Expiration) 166 CacheOptions.Policy.Expiration = std::chrono::seconds(Expiration); 167 } 168 169 /** 170 * Sets the maximum cache size that can be persistent across build, in terms 171 * of percentage of the available space on the the disk. Set to 100 to 172 * indicate no limit, 50 to indicate that the cache size will not be left over 173 * half the available space. A value over 100 will be reduced to 100, and a 174 * value of 0 will be ignored. 175 * 176 * 177 * The formula looks like: 178 * AvailableSpace = FreeSpace + ExistingCacheSize 179 * NewCacheSize = AvailableSpace * P/100 180 * 181 */ 182 void setMaxCacheSizeRelativeToAvailableSpace(unsigned Percentage) { 183 if (Percentage) 184 CacheOptions.Policy.MaxSizePercentageOfAvailableSpace = Percentage; 185 } 186 187 /**@}*/ 188 189 /// Set the path to a directory where to save temporaries at various stages of 190 /// the processing. 191 void setSaveTempsDir(std::string Path) { SaveTempsDir = std::move(Path); } 192 193 /// Set the path to a directory where to save generated object files. This 194 /// path can be used by a linker to request on-disk files instead of in-memory 195 /// buffers. When set, results are available through getProducedBinaryFiles() 196 /// instead of getProducedBinaries(). 197 void setGeneratedObjectsDirectory(std::string Path) { 198 SavedObjectsDirectoryPath = std::move(Path); 199 } 200 201 /// CPU to use to initialize the TargetMachine 202 void setCpu(std::string Cpu) { TMBuilder.MCpu = std::move(Cpu); } 203 204 /// Subtarget attributes 205 void setAttr(std::string MAttr) { TMBuilder.MAttr = std::move(MAttr); } 206 207 /// TargetMachine options 208 void setTargetOptions(TargetOptions Options) { 209 TMBuilder.Options = std::move(Options); 210 } 211 212 /// Enable the Freestanding mode: indicate that the optimizer should not 213 /// assume builtins are present on the target. 214 void setFreestanding(bool Enabled) { Freestanding = Enabled; } 215 216 /// CodeModel 217 void setCodePICModel(Optional<Reloc::Model> Model) { 218 TMBuilder.RelocModel = Model; 219 } 220 221 /// CodeGen optimization level 222 void setCodeGenOptLevel(CodeGenOpt::Level CGOptLevel) { 223 TMBuilder.CGOptLevel = CGOptLevel; 224 } 225 226 /// IR optimization level: from 0 to 3. 227 void setOptLevel(unsigned NewOptLevel) { 228 OptLevel = (NewOptLevel > 3) ? 3 : NewOptLevel; 229 } 230 231 /// Disable CodeGen, only run the stages till codegen and stop. The output 232 /// will be bitcode. 233 void disableCodeGen(bool Disable) { DisableCodeGen = Disable; } 234 235 /// Perform CodeGen only: disable all other stages. 236 void setCodeGenOnly(bool CGOnly) { CodeGenOnly = CGOnly; } 237 238 /**@}*/ 239 240 /** 241 * \defgroup Set of APIs to run individual stages in isolation. 242 * @{ 243 */ 244 245 /** 246 * Produce the combined summary index from all the bitcode files: 247 * "thin-link". 248 */ 249 std::unique_ptr<ModuleSummaryIndex> linkCombinedIndex(); 250 251 /** 252 * Perform promotion and renaming of exported internal functions, 253 * and additionally resolve weak and linkonce symbols. 254 * Index is updated to reflect linkage changes from weak resolution. 255 */ 256 void promote(Module &Module, ModuleSummaryIndex &Index); 257 258 /** 259 * Compute and emit the imported files for module at \p ModulePath. 260 */ 261 static void emitImports(StringRef ModulePath, StringRef OutputName, 262 ModuleSummaryIndex &Index); 263 264 /** 265 * Perform cross-module importing for the module identified by 266 * ModuleIdentifier. 267 */ 268 void crossModuleImport(Module &Module, ModuleSummaryIndex &Index); 269 270 /** 271 * Compute the list of summaries needed for importing into module. 272 */ 273 static void gatherImportedSummariesForModule( 274 StringRef ModulePath, ModuleSummaryIndex &Index, 275 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex); 276 277 /** 278 * Perform internalization. Index is updated to reflect linkage changes. 279 */ 280 void internalize(Module &Module, ModuleSummaryIndex &Index); 281 282 /** 283 * Perform post-importing ThinLTO optimizations. 284 */ 285 void optimize(Module &Module); 286 287 /** 288 * Perform ThinLTO CodeGen. 289 */ 290 std::unique_ptr<MemoryBuffer> codegen(Module &Module); 291 292 /**@}*/ 293 294private: 295 /// Helper factory to build a TargetMachine 296 TargetMachineBuilder TMBuilder; 297 298 /// Vector holding the in-memory buffer containing the produced binaries, when 299 /// SavedObjectsDirectoryPath isn't set. 300 std::vector<std::unique_ptr<MemoryBuffer>> ProducedBinaries; 301 302 /// Path to generated files in the supplied SavedObjectsDirectoryPath if any. 303 std::vector<std::string> ProducedBinaryFiles; 304 305 /// Vector holding the input buffers containing the bitcode modules to 306 /// process. 307 std::vector<ThinLTOBuffer> Modules; 308 309 /// Set of symbols that need to be preserved outside of the set of bitcode 310 /// files. 311 StringSet<> PreservedSymbols; 312 313 /// Set of symbols that are cross-referenced between bitcode files. 314 StringSet<> CrossReferencedSymbols; 315 316 /// Control the caching behavior. 317 CachingOptions CacheOptions; 318 319 /// Path to a directory to save the temporary bitcode files. 320 std::string SaveTempsDir; 321 322 /// Path to a directory to save the generated object files. 323 std::string SavedObjectsDirectoryPath; 324 325 /// Flag to enable/disable CodeGen. When set to true, the process stops after 326 /// optimizations and a bitcode is produced. 327 bool DisableCodeGen = false; 328 329 /// Flag to indicate that only the CodeGen will be performed, no cross-module 330 /// importing or optimization. 331 bool CodeGenOnly = false; 332 333 /// Flag to indicate that the optimizer should not assume builtins are present 334 /// on the target. 335 bool Freestanding = false; 336 337 /// IR Optimization Level [0-3]. 338 unsigned OptLevel = 3; 339}; 340} 341#endif 342