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     }
 244     args.push_back("-c");
 245     args.push_back("-x");
 246     args.push_back("assembler");
 247     args.push_back("-o");
 248     args.push_back(objPath.c_str());
 249     args.push_back(asmPath.c_str());
 250     args.push_back(0);
 251
 252     // invoke assembler
 253     if ( sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 0, 0, &errMsg) ) {
 254         errMsg = "error in assembly";
 255         return true;
 256     }
 257     return false; // success
 258 }
 259
 260
 261
 262 bool LTOCodeGenerator::determineTarget(std::string& errMsg)
 263 {
 264     if ( _target == NULL ) {
 265         // create target machine from info for merged modules
 266         Module* mergedModule = _linker.getModule();
 267         const TargetMachineRegistry::entry* march =
 268           TargetMachineRegistry::getClosestStaticTargetForModule(
 269                                                        *mergedModule, errMsg);
 270         if ( march == NULL )
 271             return true;
 272
 273         // construct LTModule, hand over ownership of module and target
 274         std::string FeatureStr =
 275           getFeatureString(_linker.getModule()->getTargetTriple().c_str());
 276         _target = march->CtorFn(*mergedModule, FeatureStr.c_str());
 277     }
 278     return false;
 279 }
 280
 281 void LTOCodeGenerator::applyScopeRestrictions()
 282 {
 283     if ( !_scopeRestrictionsDone ) {
 284         Module* mergedModule = _linker.getModule();
 285
 286         // Start off with a verification pass.
 287         PassManager passes;
 288         passes.add(createVerifierPass());
 289
 290         // mark which symbols can not be internalized
 291         if ( !_mustPreserveSymbols.empty() ) {
 292             Mangler mangler(*mergedModule,
 293                                 _target->getTargetAsmInfo()->getGlobalPrefix());
 294             std::vector<const char*> mustPreserveList;
 295             for (Module::iterator f = mergedModule->begin(),
 296                                         e = mergedModule->end(); f != e; ++f) {
 297                 if ( !f->isDeclaration()
 298                   && _mustPreserveSymbols.count(mangler.getValueName(f)) )
 299                     mustPreserveList.push_back(::strdup(f->getName().c_str()));
 300             }
 301             for (Module::global_iterator v = mergedModule->global_begin(),
 302                                  e = mergedModule->global_end(); v !=  e; ++v) {
 303                 if ( !v->isDeclaration()
 304                   && _mustPreserveSymbols.count(mangler.getValueName(v)) )
 305                     mustPreserveList.push_back(::strdup(v->getName().c_str()));
 306             }
 307             passes.add(createInternalizePass(mustPreserveList));
 308         }
 309         // apply scope restrictions
 310         passes.run(*mergedModule);
 311
 312         _scopeRestrictionsDone = true;
 313     }
 314 }
 315
 316 /// Optimize merged modules using various IPO passes
 317 bool LTOCodeGenerator::generateAssemblyCode(raw_ostream& out,
 318                                             std::string& errMsg)
 319 {
 320     if (  this->determineTarget(errMsg) )
 321         return true;
 322
 323     // mark which symbols can not be internalized
 324     this->applyScopeRestrictions();
 325
 326     Module* mergedModule = _linker.getModule();
 327
 328      // If target supports exception handling then enable it now.
 329     if ( _target->getTargetAsmInfo()->doesSupportExceptionHandling() )
 330         llvm::ExceptionHandling = true;
 331
 332     // set codegen model
 333     switch( _codeModel ) {
 334         case LTO_CODEGEN_PIC_MODEL_STATIC:
 335             _target->setRelocationModel(Reloc::Static);
 336             break;
 337         case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
 338             _target->setRelocationModel(Reloc::PIC_);
 339             break;
 340         case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
 341             _target->setRelocationModel(Reloc::DynamicNoPIC);
 342             break;
 343     }
 344
 345     // if options were requested, set them
 346     if ( !_codegenOptions.empty() )
 347         cl::ParseCommandLineOptions(_codegenOptions.size(),
 348                                                 (char**)&_codegenOptions[0]);
 349
 350     // Instantiate the pass manager to organize the passes.
 351     PassManager passes;
 352
 353     // Start off with a verification pass.
 354     passes.add(createVerifierPass());
 355
 356     // Add an appropriate TargetData instance for this module...
 357     passes.add(new TargetData(*_target->getTargetData()));
 358
 359     // Propagate constants at call sites into the functions they call.  This
 360     // opens opportunities for globalopt (and inlining) by substituting function
 361     // pointers passed as arguments to direct uses of functions.
 362     passes.add(createIPSCCPPass());
 363
 364     // Now that we internalized some globals, see if we can hack on them!
 365     passes.add(createGlobalOptimizerPass());
 366
 367     // Linking modules together can lead to duplicated global constants, only
 368     // keep one copy of each constant...
 369     passes.add(createConstantMergePass());
 370
 371     // Remove unused arguments from functions...
 372     passes.add(createDeadArgEliminationPass());
 373
 374     // Reduce the code after globalopt and ipsccp.  Both can open up significant
 375     // simplification opportunities, and both can propagate functions through
 376     // function pointers.  When this happens, we often have to resolve varargs
 377     // calls, etc, so let instcombine do this.
 378     passes.add(createInstructionCombiningPass());
 379     if (!DisableInline)
 380         passes.add(createFunctionInliningPass()); // Inline small functions
 381     passes.add(createPruneEHPass());              // Remove dead EH info
 382     passes.add(createGlobalDCEPass());            // Remove dead functions
 383
 384     // If we didn't decide to inline a function, check to see if we can
 385     // transform it to pass arguments by value instead of by reference.
 386     passes.add(createArgumentPromotionPass());
 387
 388     // The IPO passes may leave cruft around.  Clean up after them.
 389     passes.add(createInstructionCombiningPass());
 390     passes.add(createJumpThreadingPass());        // Thread jumps.
 391     passes.add(createScalarReplAggregatesPass()); // Break up allocas
 392
 393     // Run a few AA driven optimizations here and now, to cleanup the code.
 394     passes.add(createFunctionAttrsPass());        // Add nocapture
 395     passes.add(createGlobalsModRefPass());        // IP alias analysis
 396     passes.add(createLICMPass());                 // Hoist loop invariants
 397     passes.add(createGVNPass());                  // Remove common subexprs
 398     passes.add(createMemCpyOptPass());            // Remove dead memcpy's
 399     passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
 400
 401     // Cleanup and simplify the code after the scalar optimizations.
 402     passes.add(createInstructionCombiningPass());
 403     passes.add(createJumpThreadingPass());        // Thread jumps.
 404     passes.add(createPromoteMemoryToRegisterPass()); // Cleanup after threading.
 405
 406
 407     // Delete basic blocks, which optimization passes may have killed...
 408     passes.add(createCFGSimplificationPass());
 409
 410     // Now that we have optimized the program, discard unreachable functions...
 411     passes.add(createGlobalDCEPass());
 412
 413     // Make sure everything is still good.
 414     passes.add(createVerifierPass());
 415
 416     FunctionPassManager* codeGenPasses =
 417             new FunctionPassManager(new ExistingModuleProvider(mergedModule));
 418
 419     codeGenPasses->add(new TargetData(*_target->getTargetData()));
 420
 421     MachineCodeEmitter* mce = NULL;
 422
 423     switch (_target->addPassesToEmitFile(*codeGenPasses, out,
 424                                       TargetMachine::AssemblyFile, false)) {
 425         case FileModel::MachOFile:
 426             mce = AddMachOWriter(*codeGenPasses, out, *_target);
 427             break;
 428         case FileModel::ElfFile:
 429             mce = AddELFWriter(*codeGenPasses, out, *_target);
 430             break;
 431         case FileModel::AsmFile:
 432             break;
 433         case FileModel::Error:
 434         case FileModel::None:
 435             errMsg = "target file type not supported";
 436             return true;
 437     }
 438
 439     if (_target->addPassesToEmitFileFinish(*codeGenPasses, mce, false)) {
 440         errMsg = "target does not support generation of this file type";
 441         return true;
 442     }
 443
 444     // Run our queue of passes all at once now, efficiently.
 445     passes.run(*mergedModule);
 446
 447     // Run the code generator, and write assembly file
 448     codeGenPasses->doInitialization();
 449
 450     for (Module::iterator
 451            it = mergedModule->begin(), e = mergedModule->end(); it != e; ++it)
 452       if (!it->isDeclaration())
 453         codeGenPasses->run(*it);
 454
 455     codeGenPasses->doFinalization();
 456     return false; // success
 457 }
 458
 459
 460 /// Optimize merged modules using various IPO passes
 461 void LTOCodeGenerator::setCodeGenDebugOptions(const char* options)
 462 {
 463     std::string ops(options);
 464     for (std::string o = getToken(ops); !o.empty(); o = getToken(ops)) {
 465         // ParseCommandLineOptions() expects argv[0] to be program name.
 466         // Lazily add that.
 467         if ( _codegenOptions.empty() )
 468             _codegenOptions.push_back("libLTO");
 469         _codegenOptions.push_back(strdup(o.c_str()));
 470     }
 471 }