tools/lto/LTOCodeGenerator.cpp

   1 //===-LTOCodeGenerator.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 "LTOModule.h"
  16 #include "LTOCodeGenerator.h"
  17
  18
  19 #include "llvm/Module.h"
  20 #include "llvm/PassManager.h"
  21 #include "llvm/Linker.h"
  22 #include "llvm/Constants.h"
  23 #include "llvm/DerivedTypes.h"
  24 #include "llvm/ModuleProvider.h"
  25 #include "llvm/Bitcode/ReaderWriter.h"
  26 #include "llvm/Support/CommandLine.h"
  27 #include "llvm/Support/SystemUtils.h"
  28 #include "llvm/Support/Mangler.h"
  29 #include "llvm/Support/MemoryBuffer.h"
  30 #include "llvm/Support/raw_ostream.h"
  31 #include "llvm/System/Signals.h"
  32 #include "llvm/Analysis/Passes.h"
  33 #include "llvm/Analysis/LoopPass.h"
  34 #include "llvm/Analysis/Verifier.h"
  35 #include "llvm/CodeGen/FileWriters.h"
  36 #include "llvm/Target/SubtargetFeature.h"
  37 #include "llvm/Target/TargetOptions.h"
  38 #include "llvm/Target/TargetData.h"
  39 #include "llvm/Target/TargetMachine.h"
  40 #include "llvm/Target/TargetMachineRegistry.h"
  41 #include "llvm/Target/TargetAsmInfo.h"
  42 #include "llvm/Transforms/IPO.h"
  43 #include "llvm/Transforms/Scalar.h"
  44 #include "llvm/ADT/StringExtras.h"
  45 #include "llvm/Config/config.h"
  46
  47
  48 #include <fstream>
  49 #include <unistd.h>
  50 #include <stdlib.h>
  51 #include <fcntl.h>
  52
  53
  54 using namespace llvm;
  55
  56 static cl::opt<bool> DisableInline("disable-inlining",
  57   cl::desc("Do not run the inliner pass"));
  58
  59
  60 const char* LTOCodeGenerator::getVersionString()
  61 {
  62 #ifdef LLVM_VERSION_INFO
  63     return PACKAGE_NAME " version " PACKAGE_VERSION ", " LLVM_VERSION_INFO;
  64 #else
  65     return PACKAGE_NAME " version " PACKAGE_VERSION;
  66 #endif
  67 }
  68
  69
  70 LTOCodeGenerator::LTOCodeGenerator()
  71     : _linker("LinkTimeOptimizer", "ld-temp.o"), _target(NULL),
  72       _emitDwarfDebugInfo(false), _scopeRestrictionsDone(false),
  73       _codeModel(LTO_CODEGEN_PIC_MODEL_DYNAMIC),
  74       _nativeObjectFile(NULL)
  75 {
  76
  77 }
  78
  79 LTOCodeGenerator::~LTOCodeGenerator()
  80 {
  81     delete _target;
  82     delete _nativeObjectFile;
  83 }
  84
  85
  86
  87 bool LTOCodeGenerator::addModule(LTOModule* mod, std::string& errMsg)
  88 {
  89     return _linker.LinkInModule(mod->getLLVVMModule(), &errMsg);
  90 }
  91
  92
  93 bool LTOCodeGenerator::setDebugInfo(lto_debug_model debug, std::string& errMsg)
  94 {
  95     switch (debug) {
  96         case LTO_DEBUG_MODEL_NONE:
  97             _emitDwarfDebugInfo = false;
  98             return false;
  99
 100         case LTO_DEBUG_MODEL_DWARF:
 101             _emitDwarfDebugInfo = true;
 102             return false;
 103     }
 104     errMsg = "unknown debug format";
 105     return true;
 106 }
 107
 108
 109 bool LTOCodeGenerator::setCodePICModel(lto_codegen_model model,
 110                                                         std::string& errMsg)
 111 {
 112     switch (model) {
 113         case LTO_CODEGEN_PIC_MODEL_STATIC:
 114         case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
 115         case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
 116             _codeModel = model;
 117             return false;
 118     }
 119     errMsg = "unknown pic model";
 120     return true;
 121 }
 122
 123 void LTOCodeGenerator::addMustPreserveSymbol(const char* sym)
 124 {
 125     _mustPreserveSymbols[sym] = 1;
 126 }
 127
 128
 129 bool LTOCodeGenerator::writeMergedModules(const char* path, std::string& errMsg)
 130 {
 131     if ( this->determineTarget(errMsg) )
 132         return true;
 133
 134     // mark which symbols can not be internalized
 135     this->applyScopeRestrictions();
 136
 137     // create output file
 138     std::ofstream out(path, std::ios_base::out|std::ios::trunc|std::ios::binary);
 139     if ( out.fail() ) {
 140         errMsg = "could not open bitcode file for writing: ";
 141         errMsg += path;
 142         return true;
 143     }
 144
 145     // write bitcode to it
 146     WriteBitcodeToFile(_linker.getModule(), out);
 147     if ( out.fail() ) {
 148         errMsg = "could not write bitcode file: ";
 149         errMsg += path;
 150         return true;
 151     }
 152
 153     return false;
 154 }
 155
 156
 157 const void* LTOCodeGenerator::compile(size_t* length, std::string& errMsg)
 158 {
 159     // make unique temp .s file to put generated assembly code
 160     sys::Path uniqueAsmPath("lto-llvm.s");
 161     if ( uniqueAsmPath.createTemporaryFileOnDisk(true, &errMsg) )
 162         return NULL;
 163     sys::RemoveFileOnSignal(uniqueAsmPath);
 164
 165     // generate assembly code
 166     bool genResult = false;
 167     {
 168       raw_fd_ostream asmFile(uniqueAsmPath.c_str(), false, errMsg);
 169       if (!errMsg.empty())
 170         return NULL;
 171       genResult = this->generateAssemblyCode(asmFile, errMsg);
 172     }
 173     if ( genResult ) {
 174         if ( uniqueAsmPath.exists() )
 175             uniqueAsmPath.eraseFromDisk();
 176         return NULL;
 177     }
 178
 179     // make unique temp .o file to put generated object file
 180     sys::PathWithStatus uniqueObjPath("lto-llvm.o");
 181     if ( uniqueObjPath.createTemporaryFileOnDisk(true, &errMsg) ) {
 182         if ( uniqueAsmPath.exists() )
 183             uniqueAsmPath.eraseFromDisk();
 184         return NULL;
 185     }
 186     sys::RemoveFileOnSignal(uniqueObjPath);
 187
 188     // assemble the assembly code
 189     const std::string& uniqueObjStr = uniqueObjPath.toString();
 190     bool asmResult = this->assemble(uniqueAsmPath.toString(),
 191                                                         uniqueObjStr, errMsg);
 192     if ( !asmResult ) {
 193         // remove old buffer if compile() called twice
 194         delete _nativeObjectFile;
 195
 196         // read .o file into memory buffer
 197         _nativeObjectFile = MemoryBuffer::getFile(uniqueObjStr.c_str(),&errMsg);
 198     }
 199
 200     // remove temp files
 201     uniqueAsmPath.eraseFromDisk();
 202     uniqueObjPath.eraseFromDisk();
 203
 204     // return buffer, unless error
 205     if ( _nativeObjectFile == NULL )
 206         return NULL;
 207     *length = _nativeObjectFile->getBufferSize();
 208     return _nativeObjectFile->getBufferStart();
 209 }
 210
 211
 212 bool LTOCodeGenerator::assemble(const std::string& asmPath,
 213                                 const std::string& objPath, std::string& errMsg)
 214 {
 215     // find compiler driver
 216     const sys::Path gcc = sys::Program::FindProgramByName("gcc");
 217     if ( gcc.isEmpty() ) {
 218         errMsg = "can't locate gcc";
 219         return true;
 220     }
 221
 222     // build argument list
 223     std::vector<const char*> args;
 224     std::string targetTriple = _linker.getModule()->getTargetTriple();
 225     args.push_back(gcc.c_str());
 226     if ( targetTriple.find("darwin") != targetTriple.size() ) {
 227         if (strncmp(targetTriple.c_str(), "i386-apple-", 11) == 0) {
 228             args.push_back("-arch");
 229             args.push_back("i386");
 230         }
 231         else if (strncmp(targetTriple.c_str(), "x86_64-apple-", 13) == 0) {
 232             args.push_back("-arch");
 233             args.push_back("x86_64");
 234         }
 235         else if (strncmp(targetTriple.c_str(), "powerpc-apple-", 14) == 0) {
 236             args.push_back("-arch");
 237             args.push_back("ppc");
 238         }
 239         else if (strncmp(targetTriple.c_str(), "powerpc64-apple-", 16) == 0) {
 240             args.push_back("-arch");
 241             args.push_back("ppc64");
 242         }
 243         else if (strncmp(targetTriple.c_str(), "arm-apple-", 10) == 0) {
 244             args.push_back("-arch");
 245             args.push_back("arm");
 246         }
 247         else if ((strncmp(targetTriple.c_str(), "armv4t-apple-", 13) == 0) ||
 248                  (strncmp(targetTriple.c_str(), "thumbv4t-apple-", 15) == 0)) {
 249             args.push_back("-arch");
 250             args.push_back("armv4t");
 251         }
 252         else if ((strncmp(targetTriple.c_str(), "armv5-apple-", 12) == 0) ||
 253                  (strncmp(targetTriple.c_str(), "armv5e-apple-", 13) == 0) ||
 254                  (strncmp(targetTriple.c_str(), "thumbv5-apple-", 14) == 0) ||
 255                  (strncmp(targetTriple.c_str(), "thumbv5e-apple-", 15) == 0)) {
 256             args.push_back("-arch");
 257             args.push_back("armv5");
 258         }
 259         else if ((strncmp(targetTriple.c_str(), "armv6-apple-", 12) == 0) ||
 260                  (strncmp(targetTriple.c_str(), "thumbv6-apple-", 14) == 0)) {
 261             args.push_back("-arch");
 262             args.push_back("armv6");
 263         }
 264     }
 265     args.push_back("-c");
 266     args.push_back("-x");
 267     args.push_back("assembler");
 268     args.push_back("-o");
 269     args.push_back(objPath.c_str());
 270     args.push_back(asmPath.c_str());
 271     args.push_back(0);
 272
 273     // invoke assembler
 274     if ( sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 0, 0, &errMsg) ) {
 275         errMsg = "error in assembly";
 276         return true;
 277     }
 278     return false; // success
 279 }
 280
 281
 282
 283 bool LTOCodeGenerator::determineTarget(std::string& errMsg)
 284 {
 285     if ( _target == NULL ) {
 286         // create target machine from info for merged modules
 287         Module* mergedModule = _linker.getModule();
 288         const TargetMachineRegistry::entry* march =
 289           TargetMachineRegistry::getClosestStaticTargetForModule(
 290                                                        *mergedModule, errMsg);
 291         if ( march == NULL )
 292             return true;
 293
 294         // construct LTModule, hand over ownership of module and target
 295         std::string FeatureStr =
 296           getFeatureString(_linker.getModule()->getTargetTriple().c_str());
 297         _target = march->CtorFn(*mergedModule, FeatureStr.c_str());
 298     }
 299     return false;
 300 }
 301
 302 void LTOCodeGenerator::applyScopeRestrictions()
 303 {
 304     if ( !_scopeRestrictionsDone ) {
 305         Module* mergedModule = _linker.getModule();
 306
 307         // Start off with a verification pass.
 308         PassManager passes;
 309         passes.add(createVerifierPass());
 310
 311         // mark which symbols can not be internalized
 312         if ( !_mustPreserveSymbols.empty() ) {
 313             Mangler mangler(*mergedModule,
 314                                 _target->getTargetAsmInfo()->getGlobalPrefix());
 315             std::vector<const char*> mustPreserveList;
 316             for (Module::iterator f = mergedModule->begin(),
 317                                         e = mergedModule->end(); f != e; ++f) {
 318                 if ( !f->isDeclaration()
 319                   && _mustPreserveSymbols.count(mangler.getValueName(f)) )
 320                     mustPreserveList.push_back(::strdup(f->getName().c_str()));
 321             }
 322             for (Module::global_iterator v = mergedModule->global_begin(),
 323                                  e = mergedModule->global_end(); v !=  e; ++v) {
 324                 if ( !v->isDeclaration()
 325                   && _mustPreserveSymbols.count(mangler.getValueName(v)) )
 326                     mustPreserveList.push_back(::strdup(v->getName().c_str()));
 327             }
 328             passes.add(createInternalizePass(mustPreserveList));
 329         }
 330         // apply scope restrictions
 331         passes.run(*mergedModule);
 332
 333         _scopeRestrictionsDone = true;
 334     }
 335 }
 336
 337 /// Optimize merged modules using various IPO passes
 338 bool LTOCodeGenerator::generateAssemblyCode(raw_ostream& out,
 339                                             std::string& errMsg)
 340 {
 341     if (  this->determineTarget(errMsg) )
 342         return true;
 343
 344     // mark which symbols can not be internalized
 345     this->applyScopeRestrictions();
 346
 347     Module* mergedModule = _linker.getModule();
 348
 349      // If target supports exception handling then enable it now.
 350     if ( _target->getTargetAsmInfo()->doesSupportExceptionHandling() )
 351         llvm::ExceptionHandling = true;
 352
 353     // set codegen model
 354     switch( _codeModel ) {
 355         case LTO_CODEGEN_PIC_MODEL_STATIC:
 356             _target->setRelocationModel(Reloc::Static);
 357             break;
 358         case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
 359             _target->setRelocationModel(Reloc::PIC_);
 360             break;
 361         case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
 362             _target->setRelocationModel(Reloc::DynamicNoPIC);
 363             break;
 364     }
 365
 366     // if options were requested, set them
 367     if ( !_codegenOptions.empty() )
 368         cl::ParseCommandLineOptions(_codegenOptions.size(),
 369                                                 (char**)&_codegenOptions[0]);
 370
 371     // Instantiate the pass manager to organize the passes.
 372     PassManager passes;
 373
 374     // Start off with a verification pass.
 375     passes.add(createVerifierPass());
 376
 377     // Add an appropriate TargetData instance for this module...
 378     passes.add(new TargetData(*_target->getTargetData()));
 379
 380     // Propagate constants at call sites into the functions they call.  This
 381     // opens opportunities for globalopt (and inlining) by substituting function
 382     // pointers passed as arguments to direct uses of functions.
 383     passes.add(createIPSCCPPass());
 384
 385     // Now that we internalized some globals, see if we can hack on them!
 386     passes.add(createGlobalOptimizerPass());
 387
 388     // Linking modules together can lead to duplicated global constants, only
 389     // keep one copy of each constant...
 390     passes.add(createConstantMergePass());
 391
 392     // Remove unused arguments from functions...
 393     passes.add(createDeadArgEliminationPass());
 394
 395     // Reduce the code after globalopt and ipsccp.  Both can open up significant
 396     // simplification opportunities, and both can propagate functions through
 397     // function pointers.  When this happens, we often have to resolve varargs
 398     // calls, etc, so let instcombine do this.
 399     passes.add(createInstructionCombiningPass());
 400     if (!DisableInline)
 401         passes.add(createFunctionInliningPass()); // Inline small functions
 402     passes.add(createPruneEHPass());              // Remove dead EH info
 403     passes.add(createGlobalDCEPass());            // Remove dead functions
 404
 405     // If we didn't decide to inline a function, check to see if we can
 406     // transform it to pass arguments by value instead of by reference.
 407     passes.add(createArgumentPromotionPass());
 408
 409     // The IPO passes may leave cruft around.  Clean up after them.
 410     passes.add(createInstructionCombiningPass());
 411     passes.add(createJumpThreadingPass());        // Thread jumps.
 412     passes.add(createScalarReplAggregatesPass()); // Break up allocas
 413
 414     // Run a few AA driven optimizations here and now, to cleanup the code.
 415     passes.add(createFunctionAttrsPass());        // Add nocapture
 416     passes.add(createGlobalsModRefPass());        // IP alias analysis
 417     passes.add(createLICMPass());                 // Hoist loop invariants
 418     passes.add(createGVNPass());                  // Remove common subexprs
 419     passes.add(createMemCpyOptPass());            // Remove dead memcpy's
 420     passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
 421
 422     // Cleanup and simplify the code after the scalar optimizations.
 423     passes.add(createInstructionCombiningPass());
 424     passes.add(createJumpThreadingPass());        // Thread jumps.
 425     passes.add(createPromoteMemoryToRegisterPass()); // Cleanup after threading.
 426
 427
 428     // Delete basic blocks, which optimization passes may have killed...
 429     passes.add(createCFGSimplificationPass());
 430
 431     // Now that we have optimized the program, discard unreachable functions...
 432     passes.add(createGlobalDCEPass());
 433
 434     // Make sure everything is still good.
 435     passes.add(createVerifierPass());
 436
 437     FunctionPassManager* codeGenPasses =
 438             new FunctionPassManager(new ExistingModuleProvider(mergedModule));
 439
 440     codeGenPasses->add(new TargetData(*_target->getTargetData()));
 441
 442     MachineCodeEmitter* mce = NULL;
 443
 444     switch (_target->addPassesToEmitFile(*codeGenPasses, out,
 445                                          TargetMachine::AssemblyFile,
 446                                          CodeGenOpt::Aggressive)) {
 447         case FileModel::MachOFile:
 448             mce = AddMachOWriter(*codeGenPasses, out, *_target);
 449             break;
 450         case FileModel::ElfFile:
 451             mce = AddELFWriter(*codeGenPasses, out, *_target);
 452             break;
 453         case FileModel::AsmFile:
 454             break;
 455         case FileModel::Error:
 456         case FileModel::None:
 457             errMsg = "target file type not supported";
 458             return true;
 459     }
 460
 461     if (_target->addPassesToEmitFileFinish(*codeGenPasses, mce,
 462                                            CodeGenOpt::Aggressive)) {
 463         errMsg = "target does not support generation of this file type";
 464         return true;
 465     }
 466
 467     // Run our queue of passes all at once now, efficiently.
 468     passes.run(*mergedModule);
 469
 470     // Run the code generator, and write assembly file
 471     codeGenPasses->doInitialization();
 472
 473     for (Module::iterator
 474            it = mergedModule->begin(), e = mergedModule->end(); it != e; ++it)
 475       if (!it->isDeclaration())
 476         codeGenPasses->run(*it);
 477
 478     codeGenPasses->doFinalization();
 479     return false; // success
 480 }
 481
 482
 483 /// Optimize merged modules using various IPO passes
 484 void LTOCodeGenerator::setCodeGenDebugOptions(const char* options)
 485 {
 486     std::string ops(options);
 487     for (std::string o = getToken(ops); !o.empty(); o = getToken(ops)) {
 488         // ParseCommandLineOptions() expects argv[0] to be program name.
 489         // Lazily add that.
 490         if ( _codegenOptions.empty() )
 491             _codegenOptions.push_back("libLTO");
 492         _codegenOptions.push_back(strdup(o.c_str()));
 493     }
 494 }