lto.cpp revision 21b70b237d404ebdde190a072d90f630db92f691
1//===-lto.cpp - 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 implements the Link Time Optimization library. This library is 11// intended to be used by linker to optimize code at link time. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Module.h" 16#include "llvm/PassManager.h" 17#include "llvm/Linker.h" 18#include "llvm/Constants.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/ModuleProvider.h" 21#include "llvm/Bitcode/ReaderWriter.h" 22#include "llvm/Support/CommandLine.h" 23#include "llvm/Support/FileUtilities.h" 24#include "llvm/Support/SystemUtils.h" 25#include "llvm/Support/Mangler.h" 26#include "llvm/Support/MemoryBuffer.h" 27#include "llvm/System/Program.h" 28#include "llvm/System/Signals.h" 29#include "llvm/Analysis/Passes.h" 30#include "llvm/Analysis/LoopPass.h" 31#include "llvm/Analysis/Verifier.h" 32#include "llvm/CodeGen/FileWriters.h" 33#include "llvm/Target/SubtargetFeature.h" 34#include "llvm/Target/TargetOptions.h" 35#include "llvm/Target/TargetData.h" 36#include "llvm/Target/TargetMachine.h" 37#include "llvm/Target/TargetMachineRegistry.h" 38#include "llvm/Target/TargetAsmInfo.h" 39#include "llvm/Transforms/IPO.h" 40#include "llvm/Transforms/Scalar.h" 41#include "llvm/Analysis/LoadValueNumbering.h" 42#include "llvm/Support/MathExtras.h" 43#include "llvm/LinkTimeOptimizer.h" 44#include <fstream> 45#include <ostream> 46using namespace llvm; 47 48extern "C" 49llvm::LinkTimeOptimizer *createLLVMOptimizer(unsigned VERSION) 50{ 51 // Linker records LLVM_LTO_VERSION based on llvm optimizer available 52 // during linker build. Match linker's recorded LTO VERSION number 53 // with installed llvm optimizer version. If these numbers do not match 54 // then linker may not be able to use llvm optimizer dynamically. 55 if (VERSION != LLVM_LTO_VERSION) 56 return NULL; 57 58 llvm::LTO *l = new llvm::LTO(); 59 return l; 60} 61 62/// If symbol is not used then make it internal and let optimizer takes 63/// care of it. 64void LLVMSymbol::mayBeNotUsed() { 65 gv->setLinkage(GlobalValue::InternalLinkage); 66} 67 68// Map LLVM LinkageType to LTO LinakgeType 69static LTOLinkageTypes 70getLTOLinkageType(GlobalValue *v) 71{ 72 LTOLinkageTypes lt; 73 if (v->hasExternalLinkage()) 74 lt = LTOExternalLinkage; 75 else if (v->hasLinkOnceLinkage()) 76 lt = LTOLinkOnceLinkage; 77 else if (v->hasWeakLinkage()) 78 lt = LTOWeakLinkage; 79 else 80 // Otherwise it is internal linkage for link time optimizer 81 lt = LTOInternalLinkage; 82 return lt; 83} 84 85// MAP LLVM VisibilityType to LTO VisibilityType 86static LTOVisibilityTypes 87getLTOVisibilityType(GlobalValue *v) 88{ 89 LTOVisibilityTypes vis; 90 if (v->hasHiddenVisibility()) 91 vis = LTOHiddenVisibility; 92 else if (v->hasProtectedVisibility()) 93 vis = LTOProtectedVisibility; 94 else 95 vis = LTODefaultVisibility; 96 return vis; 97} 98 99// Find exeternal symbols referenced by VALUE. This is a recursive function. 100static void 101findExternalRefs(Value *value, std::set<std::string> &references, 102 Mangler &mangler) { 103 104 if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) { 105 LTOLinkageTypes lt = getLTOLinkageType(gv); 106 if (lt != LTOInternalLinkage && strncmp (gv->getName().c_str(), "llvm.", 5)) 107 references.insert(mangler.getValueName(gv)); 108 } 109 110 // GlobalValue, even with InternalLinkage type, may have operands with 111 // ExternalLinkage type. Do not ignore these operands. 112 if (Constant *c = dyn_cast<Constant>(value)) 113 // Handle ConstantExpr, ConstantStruct, ConstantArry etc.. 114 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i) 115 findExternalRefs(c->getOperand(i), references, mangler); 116} 117 118/// If Module with InputFilename is available then remove it from allModules 119/// and call delete on it. 120void 121LTO::removeModule (const std::string &InputFilename) 122{ 123 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str()); 124 if (pos == allModules.end()) 125 return; 126 127 Module *m = pos->second; 128 allModules.erase(pos); 129 delete m; 130} 131 132/// InputFilename is a LLVM bitcode file. If Module with InputFilename is 133/// available then return it. Otherwise parseInputFilename. 134Module * 135LTO::getModule(const std::string &InputFilename) 136{ 137 Module *m = NULL; 138 139 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str()); 140 if (pos != allModules.end()) 141 m = allModules[InputFilename.c_str()]; 142 else { 143 if (MemoryBuffer *Buffer 144 = MemoryBuffer::getFile(&InputFilename[0], InputFilename.size())) { 145 m = ParseBitcodeFile(Buffer); 146 delete Buffer; 147 } 148 allModules[InputFilename.c_str()] = m; 149 } 150 return m; 151} 152 153/// InputFilename is a LLVM bitcode file. Reade this bitcode file and 154/// set corresponding target triplet string. 155void 156LTO::getTargetTriple(const std::string &InputFilename, 157 std::string &targetTriple) 158{ 159 Module *m = getModule(InputFilename); 160 if (m) 161 targetTriple = m->getTargetTriple(); 162} 163 164/// InputFilename is a LLVM bitcode file. Read it using bitcode reader. 165/// Collect global functions and symbol names in symbols vector. 166/// Collect external references in references vector. 167/// Return LTO_READ_SUCCESS if there is no error. 168enum LTOStatus 169LTO::readLLVMObjectFile(const std::string &InputFilename, 170 NameToSymbolMap &symbols, 171 std::set<std::string> &references) 172{ 173 Module *m = getModule(InputFilename); 174 if (!m) 175 return LTO_READ_FAILURE; 176 177 // Collect Target info 178 getTarget(m); 179 180 if (!Target) 181 return LTO_READ_FAILURE; 182 183 // Use mangler to add GlobalPrefix to names to match linker names. 184 // FIXME : Instead of hard coding "-" use GlobalPrefix. 185 Mangler mangler(*m, Target->getTargetAsmInfo()->getGlobalPrefix()); 186 modules.push_back(m); 187 188 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) { 189 LTOLinkageTypes lt = getLTOLinkageType(f); 190 LTOVisibilityTypes vis = getLTOVisibilityType(f); 191 if (!f->isDeclaration() && lt != LTOInternalLinkage 192 && strncmp (f->getName().c_str(), "llvm.", 5)) { 193 int alignment = ( 16 > f->getAlignment() ? 16 : f->getAlignment()); 194 LLVMSymbol *newSymbol = new LLVMSymbol(lt, vis, f, f->getName(), 195 mangler.getValueName(f), 196 Log2_32(alignment)); 197 symbols[newSymbol->getMangledName()] = newSymbol; 198 allSymbols[newSymbol->getMangledName()] = newSymbol; 199 } 200 201 // Collect external symbols referenced by this function. 202 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b) 203 for (BasicBlock::iterator i = b->begin(), be = b->end(); 204 i != be; ++i) { 205 for (unsigned count = 0, total = i->getNumOperands(); 206 count != total; ++count) 207 findExternalRefs(i->getOperand(count), references, mangler); 208 } 209 } 210 211 for (Module::global_iterator v = m->global_begin(), e = m->global_end(); 212 v != e; ++v) { 213 LTOLinkageTypes lt = getLTOLinkageType(v); 214 LTOVisibilityTypes vis = getLTOVisibilityType(v); 215 if (!v->isDeclaration() && lt != LTOInternalLinkage 216 && strncmp (v->getName().c_str(), "llvm.", 5)) { 217 const TargetData *TD = Target->getTargetData(); 218 LLVMSymbol *newSymbol = new LLVMSymbol(lt, vis, v, v->getName(), 219 mangler.getValueName(v), 220 TD->getPreferredAlignmentLog(v)); 221 symbols[newSymbol->getMangledName()] = newSymbol; 222 allSymbols[newSymbol->getMangledName()] = newSymbol; 223 224 for (unsigned count = 0, total = v->getNumOperands(); 225 count != total; ++count) 226 findExternalRefs(v->getOperand(count), references, mangler); 227 228 } 229 } 230 231 return LTO_READ_SUCCESS; 232} 233 234/// Get TargetMachine. 235/// Use module M to find appropriate Target. 236void 237LTO::getTarget (Module *M) { 238 239 if (Target) 240 return; 241 242 std::string Err; 243 const TargetMachineRegistry::entry* March = 244 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err); 245 246 if (March == 0) 247 return; 248 249 // Create target 250 std::string Features; 251 Target = March->CtorFn(*M, Features); 252} 253 254/// Optimize module M using various IPO passes. Use exportList to 255/// internalize selected symbols. Target platform is selected 256/// based on information available to module M. No new target 257/// features are selected. 258enum LTOStatus 259LTO::optimize(Module *M, std::ostream &Out, 260 std::vector<const char *> &exportList) 261{ 262 // Instantiate the pass manager to organize the passes. 263 PassManager Passes; 264 265 // Collect Target info 266 getTarget(M); 267 268 if (!Target) 269 return LTO_NO_TARGET; 270 271 // If target supports exception handling then enable it now. 272 if (Target->getTargetAsmInfo()->doesSupportExceptionHandling()) 273 ExceptionHandling = true; 274 275 // Start off with a verification pass. 276 Passes.add(createVerifierPass()); 277 278 // Add an appropriate TargetData instance for this module... 279 Passes.add(new TargetData(*Target->getTargetData())); 280 281 // Internalize symbols if export list is nonemty 282 if (!exportList.empty()) 283 Passes.add(createInternalizePass(exportList)); 284 285 // Now that we internalized some globals, see if we can hack on them! 286 Passes.add(createGlobalOptimizerPass()); 287 288 // Linking modules together can lead to duplicated global constants, only 289 // keep one copy of each constant... 290 Passes.add(createConstantMergePass()); 291 292 // If the -s command line option was specified, strip the symbols out of the 293 // resulting program to make it smaller. -s is a GLD option that we are 294 // supporting. 295 Passes.add(createStripSymbolsPass()); 296 297 // Propagate constants at call sites into the functions they call. 298 Passes.add(createIPConstantPropagationPass()); 299 300 // Remove unused arguments from functions... 301 Passes.add(createDeadArgEliminationPass()); 302 303 Passes.add(createFunctionInliningPass()); // Inline small functions 304 305 Passes.add(createPruneEHPass()); // Remove dead EH info 306 307 Passes.add(createGlobalDCEPass()); // Remove dead functions 308 309 // If we didn't decide to inline a function, check to see if we can 310 // transform it to pass arguments by value instead of by reference. 311 Passes.add(createArgumentPromotionPass()); 312 313 // The IPO passes may leave cruft around. Clean up after them. 314 Passes.add(createInstructionCombiningPass()); 315 316 Passes.add(createScalarReplAggregatesPass()); // Break up allocas 317 318 // Run a few AA driven optimizations here and now, to cleanup the code. 319 Passes.add(createGlobalsModRefPass()); // IP alias analysis 320 321 Passes.add(createLICMPass()); // Hoist loop invariants 322 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs 323 Passes.add(createGCSEPass()); // Remove common subexprs 324 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores 325 326 // Cleanup and simplify the code after the scalar optimizations. 327 Passes.add(createInstructionCombiningPass()); 328 329 // Delete basic blocks, which optimization passes may have killed... 330 Passes.add(createCFGSimplificationPass()); 331 332 // Now that we have optimized the program, discard unreachable functions... 333 Passes.add(createGlobalDCEPass()); 334 335 // Make sure everything is still good. 336 Passes.add(createVerifierPass()); 337 338 FunctionPassManager *CodeGenPasses = 339 new FunctionPassManager(new ExistingModuleProvider(M)); 340 341 CodeGenPasses->add(new TargetData(*Target->getTargetData())); 342 343 MachineCodeEmitter *MCE = 0; 344 345 switch (Target->addPassesToEmitFile(*CodeGenPasses, Out, 346 TargetMachine::AssemblyFile, true)) { 347 default: 348 case FileModel::Error: 349 return LTO_WRITE_FAILURE; 350 case FileModel::AsmFile: 351 break; 352 case FileModel::MachOFile: 353 MCE = AddMachOWriter(*CodeGenPasses, Out, *Target); 354 break; 355 case FileModel::ElfFile: 356 MCE = AddELFWriter(*CodeGenPasses, Out, *Target); 357 break; 358 } 359 360 if (Target->addPassesToEmitFileFinish(*CodeGenPasses, MCE, true)) 361 return LTO_WRITE_FAILURE; 362 363 // Run our queue of passes all at once now, efficiently. 364 Passes.run(*M); 365 366 // Run the code generator, if present. 367 CodeGenPasses->doInitialization(); 368 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) { 369 if (!I->isDeclaration()) 370 CodeGenPasses->run(*I); 371 } 372 CodeGenPasses->doFinalization(); 373 374 return LTO_OPT_SUCCESS; 375} 376 377///Link all modules together and optimize them using IPO. Generate 378/// native object file using OutputFilename 379/// Return appropriate LTOStatus. 380enum LTOStatus 381LTO::optimizeModules(const std::string &OutputFilename, 382 std::vector<const char *> &exportList, 383 std::string &targetTriple, 384 bool saveTemps, const char *FinalOutputFilename) 385{ 386 if (modules.empty()) 387 return LTO_NO_WORK; 388 389 std::ios::openmode io_mode = 390 std::ios::out | std::ios::trunc | std::ios::binary; 391 std::string *errMsg = NULL; 392 Module *bigOne = modules[0]; 393 Linker theLinker("LinkTimeOptimizer", bigOne, false); 394 for (unsigned i = 1, e = modules.size(); i != e; ++i) 395 if (theLinker.LinkModules(bigOne, modules[i], errMsg)) 396 return LTO_MODULE_MERGE_FAILURE; 397 // all modules have been handed off to the linker. 398 modules.clear(); 399 400 sys::Path FinalOutputPath(FinalOutputFilename); 401 FinalOutputPath.eraseSuffix(); 402 403 switch(CGModel) { 404 case LTO_CGM_Dynamic: 405 Target->setRelocationModel(Reloc::PIC_); 406 break; 407 case LTO_CGM_DynamicNoPIC: 408 Target->setRelocationModel(Reloc::DynamicNoPIC); 409 break; 410 case LTO_CGM_Static: 411 Target->setRelocationModel(Reloc::Static); 412 break; 413 } 414 415 if (saveTemps) { 416 std::string tempFileName(FinalOutputPath.c_str()); 417 tempFileName += "0.bc"; 418 std::ofstream Out(tempFileName.c_str(), io_mode); 419 WriteBitcodeToFile(bigOne, Out); 420 } 421 422 // Strip leading underscore because it was added to match names 423 // seen by linker. 424 for (unsigned i = 0, e = exportList.size(); i != e; ++i) { 425 const char *name = exportList[i]; 426 NameToSymbolMap::iterator itr = allSymbols.find(name); 427 if (itr != allSymbols.end()) 428 exportList[i] = allSymbols[name]->getName(); 429 } 430 431 432 std::string ErrMsg; 433 sys::Path TempDir = sys::Path::GetTemporaryDirectory(&ErrMsg); 434 if (TempDir.isEmpty()) { 435 cerr << "lto: " << ErrMsg << "\n"; 436 return LTO_WRITE_FAILURE; 437 } 438 sys::Path tmpAsmFilePath(TempDir); 439 if (!tmpAsmFilePath.appendComponent("lto")) { 440 cerr << "lto: " << ErrMsg << "\n"; 441 TempDir.eraseFromDisk(true); 442 return LTO_WRITE_FAILURE; 443 } 444 if (tmpAsmFilePath.createTemporaryFileOnDisk(true, &ErrMsg)) { 445 cerr << "lto: " << ErrMsg << "\n"; 446 TempDir.eraseFromDisk(true); 447 return LTO_WRITE_FAILURE; 448 } 449 sys::RemoveFileOnSignal(tmpAsmFilePath); 450 451 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode); 452 if (!asmFile.is_open() || asmFile.bad()) { 453 if (tmpAsmFilePath.exists()) { 454 tmpAsmFilePath.eraseFromDisk(); 455 TempDir.eraseFromDisk(true); 456 } 457 return LTO_WRITE_FAILURE; 458 } 459 460 enum LTOStatus status = optimize(bigOne, asmFile, exportList); 461 asmFile.close(); 462 if (status != LTO_OPT_SUCCESS) { 463 tmpAsmFilePath.eraseFromDisk(); 464 TempDir.eraseFromDisk(true); 465 return status; 466 } 467 468 if (saveTemps) { 469 std::string tempFileName(FinalOutputPath.c_str()); 470 tempFileName += "1.bc"; 471 std::ofstream Out(tempFileName.c_str(), io_mode); 472 WriteBitcodeToFile(bigOne, Out); 473 } 474 475 targetTriple = bigOne->getTargetTriple(); 476 477 // Run GCC to assemble and link the program into native code. 478 // 479 // Note: 480 // We can't just assemble and link the file with the system assembler 481 // and linker because we don't know where to put the _start symbol. 482 // GCC mysteriously knows how to do it. 483 const sys::Path gcc = sys::Program::FindProgramByName("gcc"); 484 if (gcc.isEmpty()) { 485 tmpAsmFilePath.eraseFromDisk(); 486 TempDir.eraseFromDisk(true); 487 return LTO_ASM_FAILURE; 488 } 489 490 std::vector<const char*> args; 491 args.push_back(gcc.c_str()); 492 args.push_back("-c"); 493 args.push_back("-x"); 494 args.push_back("assembler"); 495 args.push_back("-o"); 496 args.push_back(OutputFilename.c_str()); 497 args.push_back(tmpAsmFilePath.c_str()); 498 args.push_back(0); 499 500 if (sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1, 0, &ErrMsg)) { 501 cerr << "lto: " << ErrMsg << "\n"; 502 return LTO_ASM_FAILURE; 503 } 504 505 tmpAsmFilePath.eraseFromDisk(); 506 TempDir.eraseFromDisk(true); 507 508 return LTO_OPT_SUCCESS; 509} 510 511void LTO::printVersion() { 512 cl::PrintVersionMessage(); 513} 514 515/// Unused pure-virtual destructor. Must remain empty. 516LinkTimeOptimizer::~LinkTimeOptimizer() {} 517 518/// Destruct LTO. Delete all modules, symbols and target. 519LTO::~LTO() { 520 521 for (std::vector<Module *>::iterator itr = modules.begin(), e = modules.end(); 522 itr != e; ++itr) 523 delete *itr; 524 525 modules.clear(); 526 527 for (NameToSymbolMap::iterator itr = allSymbols.begin(), e = allSymbols.end(); 528 itr != e; ++itr) 529 delete itr->second; 530 531 allSymbols.clear(); 532 533 delete Target; 534} 535