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), _gccPath(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::setGccPath(const char* path)
 124 {
 125     if ( _gccPath )
 126         delete _gccPath;
 127     _gccPath = new sys::Path(path);
 128 }
 129
 130 void LTOCodeGenerator::addMustPreserveSymbol(const char* sym)
 131 {
 132     _mustPreserveSymbols[sym] = 1;
 133 }
 134
 135
 136 bool LTOCodeGenerator::writeMergedModules(const char* path, std::string& errMsg)
 137 {
 138     if ( this->determineTarget(errMsg) )
 139         return true;
 140
 141     // mark which symbols can not be internalized
 142     this->applyScopeRestrictions();
 143
 144     // create output file
 145     std::ofstream out(path, std::ios_base::out|std::ios::trunc|std::ios::binary);
 146     if ( out.fail() ) {
 147         errMsg = "could not open bitcode file for writing: ";
 148         errMsg += path;
 149         return true;
 150     }
 151
 152     // write bitcode to it
 153     WriteBitcodeToFile(_linker.getModule(), out);
 154     if ( out.fail() ) {
 155         errMsg = "could not write bitcode file: ";
 156         errMsg += path;
 157         return true;
 158     }
 159
 160     return false;
 161 }
 162
 163
 164 const void* LTOCodeGenerator::compile(size_t* length, std::string& errMsg)
 165 {
 166     // make unique temp .s file to put generated assembly code
 167     sys::Path uniqueAsmPath("lto-llvm.s");
 168     if ( uniqueAsmPath.createTemporaryFileOnDisk(true, &errMsg) )
 169         return NULL;
 170     sys::RemoveFileOnSignal(uniqueAsmPath);
 171
 172     // generate assembly code
 173     bool genResult = false;
 174     {
 175       raw_fd_ostream asmFile(uniqueAsmPath.c_str(), false, errMsg);
 176       if (!errMsg.empty())
 177         return NULL;
 178       genResult = this->generateAssemblyCode(asmFile, errMsg);
 179     }
 180     if ( genResult ) {
 181         if ( uniqueAsmPath.exists() )
 182             uniqueAsmPath.eraseFromDisk();
 183         return NULL;
 184     }
 185
 186     // make unique temp .o file to put generated object file
 187     sys::PathWithStatus uniqueObjPath("lto-llvm.o");
 188     if ( uniqueObjPath.createTemporaryFileOnDisk(true, &errMsg) ) {
 189         if ( uniqueAsmPath.exists() )
 190             uniqueAsmPath.eraseFromDisk();
 191         return NULL;
 192     }
 193     sys::RemoveFileOnSignal(uniqueObjPath);
 194
 195     // assemble the assembly code
 196     const std::string& uniqueObjStr = uniqueObjPath.toString();
 197     bool asmResult = this->assemble(uniqueAsmPath.toString(),
 198                                                         uniqueObjStr, errMsg);
 199     if ( !asmResult ) {
 200         // remove old buffer if compile() called twice
 201         delete _nativeObjectFile;
 202
 203         // read .o file into memory buffer
 204         _nativeObjectFile = MemoryBuffer::getFile(uniqueObjStr.c_str(),&errMsg);
 205     }
 206
 207     // remove temp files
 208     uniqueAsmPath.eraseFromDisk();
 209     uniqueObjPath.eraseFromDisk();
 210
 211     // return buffer, unless error
 212     if ( _nativeObjectFile == NULL )
 213         return NULL;
 214     *length = _nativeObjectFile->getBufferSize();
 215     return _nativeObjectFile->getBufferStart();
 216 }
 217
 218
 219 bool LTOCodeGenerator::assemble(const std::string& asmPath,
 220                                 const std::string& objPath, std::string& errMsg)
 221 {
 222     sys::Path gcc;
 223     if ( _gccPath ) {
 224         gcc = *_gccPath;
 225     } else {
 226         // find compiler driver
 227         gcc = sys::Program::FindProgramByName("gcc");
 228         if ( gcc.isEmpty() ) {
 229             errMsg = "can't locate gcc";
 230             return true;
 231         }
 232     }
 233
 234     // build argument list
 235     std::vector<const char*> args;
 236     std::string targetTriple = _linker.getModule()->getTargetTriple();
 237     args.push_back(gcc.c_str());
 238     if ( targetTriple.find("darwin") != targetTriple.size() ) {
 239         if (strncmp(targetTriple.c_str(), "i386-apple-", 11) == 0) {
 240             args.push_back("-arch");
 241             args.push_back("i386");
 242         }
 243         else if (strncmp(targetTriple.c_str(), "x86_64-apple-", 13) == 0) {
 244             args.push_back("-arch");
 245             args.push_back("x86_64");
 246         }
 247         else if (strncmp(targetTriple.c_str(), "powerpc-apple-", 14) == 0) {
 248             args.push_back("-arch");
 249             args.push_back("ppc");
 250         }
 251         else if (strncmp(targetTriple.c_str(), "powerpc64-apple-", 16) == 0) {
 252             args.push_back("-arch");
 253             args.push_back("ppc64");
 254         }
 255         else if (strncmp(targetTriple.c_str(), "arm-apple-", 10) == 0) {
 256             args.push_back("-arch");
 257             args.push_back("arm");
 258         }
 259         else if ((strncmp(targetTriple.c_str(), "armv4t-apple-", 13) == 0) ||
 260                  (strncmp(targetTriple.c_str(), "thumbv4t-apple-", 15) == 0)) {
 261             args.push_back("-arch");
 262             args.push_back("armv4t");
 263         }
 264         else if ((strncmp(targetTriple.c_str(), "armv5-apple-", 12) == 0) ||
 265                  (strncmp(targetTriple.c_str(), "armv5e-apple-", 13) == 0) ||
 266                  (strncmp(targetTriple.c_str(), "thumbv5-apple-", 14) == 0) ||
 267                  (strncmp(targetTriple.c_str(), "thumbv5e-apple-", 15) == 0)) {
 268             args.push_back("-arch");
 269             args.push_back("armv5");
 270         }
 271         else if ((strncmp(targetTriple.c_str(), "armv6-apple-", 12) == 0) ||
 272                  (strncmp(targetTriple.c_str(), "thumbv6-apple-", 14) == 0)) {
 273             args.push_back("-arch");
 274             args.push_back("armv6");
 275         }
 276     }
 277     args.push_back("-c");
 278     args.push_back("-x");
 279     args.push_back("assembler");
 280     args.push_back("-o");
 281     args.push_back(objPath.c_str());
 282     args.push_back(asmPath.c_str());
 283     args.push_back(0);
 284
 285     // invoke assembler
 286     if ( sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 0, 0, &errMsg) ) {
 287         errMsg = "error in assembly";
 288         return true;
 289     }
 290     return false; // success
 291 }
 292
 293
 294
 295 bool LTOCodeGenerator::determineTarget(std::string& errMsg)
 296 {
 297     if ( _target == NULL ) {
 298         // create target machine from info for merged modules
 299         Module* mergedModule = _linker.getModule();
 300         const TargetMachineRegistry::entry* march =
 301           TargetMachineRegistry::getClosestStaticTargetForModule(
 302                                                        *mergedModule, errMsg);
 303         if ( march == NULL )
 304             return true;
 305
 306         // construct LTModule, hand over ownership of module and target
 307         std::string FeatureStr =
 308           getFeatureString(_linker.getModule()->getTargetTriple().c_str());
 309         _target = march->CtorFn(*mergedModule, FeatureStr.c_str());
 310     }
 311     return false;
 312 }
 313
 314 void LTOCodeGenerator::applyScopeRestrictions()
 315 {
 316     if ( !_scopeRestrictionsDone ) {
 317         Module* mergedModule = _linker.getModule();
 318
 319         // Start off with a verification pass.
 320         PassManager passes;
 321         passes.add(createVerifierPass());
 322
 323         // mark which symbols can not be internalized
 324         if ( !_mustPreserveSymbols.empty() ) {
 325             Mangler mangler(*mergedModule,
 326                                 _target->getTargetAsmInfo()->getGlobalPrefix());
 327             std::vector<const char*> mustPreserveList;
 328             for (Module::iterator f = mergedModule->begin(),
 329                                         e = mergedModule->end(); f != e; ++f) {
 330                 if ( !f->isDeclaration()
 331                   && _mustPreserveSymbols.count(mangler.getValueName(f)) )
 332                     mustPreserveList.push_back(::strdup(f->getName().c_str()));
 333             }
 334             for (Module::global_iterator v = mergedModule->global_begin(),
 335                                  e = mergedModule->global_end(); v !=  e; ++v) {
 336                 if ( !v->isDeclaration()
 337                   && _mustPreserveSymbols.count(mangler.getValueName(v)) )
 338                     mustPreserveList.push_back(::strdup(v->getName().c_str()));
 339             }
 340             passes.add(createInternalizePass(mustPreserveList));
 341         }
 342         // apply scope restrictions
 343         passes.run(*mergedModule);
 344
 345         _scopeRestrictionsDone = true;
 346     }
 347 }
 348
 349 /// Optimize merged modules using various IPO passes
 350 bool LTOCodeGenerator::generateAssemblyCode(raw_ostream& out,
 351                                             std::string& errMsg)
 352 {
 353     if (  this->determineTarget(errMsg) )
 354         return true;
 355
 356     // mark which symbols can not be internalized
 357     this->applyScopeRestrictions();
 358
 359     Module* mergedModule = _linker.getModule();
 360
 361      // If target supports exception handling then enable it now.
 362     if ( _target->getTargetAsmInfo()->doesSupportExceptionHandling() )
 363         llvm::ExceptionHandling = true;
 364
 365     // set codegen model
 366     switch( _codeModel ) {
 367         case LTO_CODEGEN_PIC_MODEL_STATIC:
 368             _target->setRelocationModel(Reloc::Static);
 369             break;
 370         case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
 371             _target->setRelocationModel(Reloc::PIC_);
 372             break;
 373         case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
 374             _target->setRelocationModel(Reloc::DynamicNoPIC);
 375             break;
 376     }
 377
 378     // if options were requested, set them
 379     if ( !_codegenOptions.empty() )
 380         cl::ParseCommandLineOptions(_codegenOptions.size(),
 381                                                 (char**)&_codegenOptions[0]);
 382
 383     // Instantiate the pass manager to organize the passes.
 384     PassManager passes;
 385
 386     // Start off with a verification pass.
 387     passes.add(createVerifierPass());
 388
 389     // Add an appropriate TargetData instance for this module...
 390     passes.add(new TargetData(*_target->getTargetData()));
 391
 392     // Propagate constants at call sites into the functions they call.  This
 393     // opens opportunities for globalopt (and inlining) by substituting function
 394     // pointers passed as arguments to direct uses of functions.
 395     passes.add(createIPSCCPPass());
 396
 397     // Now that we internalized some globals, see if we can hack on them!
 398     passes.add(createGlobalOptimizerPass());
 399
 400     // Linking modules together can lead to duplicated global constants, only
 401     // keep one copy of each constant...
 402     passes.add(createConstantMergePass());
 403
 404     // Remove unused arguments from functions...
 405     passes.add(createDeadArgEliminationPass());
 406
 407     // Reduce the code after globalopt and ipsccp.  Both can open up significant
 408     // simplification opportunities, and both can propagate functions through
 409     // function pointers.  When this happens, we often have to resolve varargs
 410     // calls, etc, so let instcombine do this.
 411     passes.add(createInstructionCombiningPass());
 412     if (!DisableInline)
 413         passes.add(createFunctionInliningPass()); // Inline small functions
 414     passes.add(createPruneEHPass());              // Remove dead EH info
 415     passes.add(createGlobalDCEPass());            // Remove dead functions
 416
 417     // If we didn't decide to inline a function, check to see if we can
 418     // transform it to pass arguments by value instead of by reference.
 419     passes.add(createArgumentPromotionPass());
 420
 421     // The IPO passes may leave cruft around.  Clean up after them.
 422     passes.add(createInstructionCombiningPass());
 423     passes.add(createJumpThreadingPass());        // Thread jumps.
 424     passes.add(createScalarReplAggregatesPass()); // Break up allocas
 425
 426     // Run a few AA driven optimizations here and now, to cleanup the code.
 427     passes.add(createFunctionAttrsPass());        // Add nocapture
 428     passes.add(createGlobalsModRefPass());        // IP alias analysis
 429     passes.add(createLICMPass());                 // Hoist loop invariants
 430     passes.add(createGVNPass());                  // Remove common subexprs
 431     passes.add(createMemCpyOptPass());            // Remove dead memcpy's
 432     passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
 433
 434     // Cleanup and simplify the code after the scalar optimizations.
 435     passes.add(createInstructionCombiningPass());
 436     passes.add(createJumpThreadingPass());        // Thread jumps.
 437     passes.add(createPromoteMemoryToRegisterPass()); // Cleanup after threading.
 438
 439
 440     // Delete basic blocks, which optimization passes may have killed...
 441     passes.add(createCFGSimplificationPass());
 442
 443     // Now that we have optimized the program, discard unreachable functions...
 444     passes.add(createGlobalDCEPass());
 445
 446     // Make sure everything is still good.
 447     passes.add(createVerifierPass());
 448
 449     FunctionPassManager* codeGenPasses =
 450             new FunctionPassManager(new ExistingModuleProvider(mergedModule));
 451
 452     codeGenPasses->add(new TargetData(*_target->getTargetData()));
 453
 454     MachineCodeEmitter* mce = NULL;
 455
 456     switch (_target->addPassesToEmitFile(*codeGenPasses, out,
 457                                          TargetMachine::AssemblyFile,
 458                                          CodeGenOpt::Aggressive)) {
 459         case FileModel::MachOFile:
 460             mce = AddMachOWriter(*codeGenPasses, out, *_target);
 461             break;
 462         case FileModel::ElfFile:
 463             mce = AddELFWriter(*codeGenPasses, out, *_target);
 464             break;
 465         case FileModel::AsmFile:
 466             break;
 467         case FileModel::Error:
 468         case FileModel::None:
 469             errMsg = "target file type not supported";
 470             return true;
 471     }
 472
 473     if (_target->addPassesToEmitFileFinish(*codeGenPasses, mce,
 474                                            CodeGenOpt::Aggressive)) {
 475         errMsg = "target does not support generation of this file type";
 476         return true;
 477     }
 478
 479     // Run our queue of passes all at once now, efficiently.
 480     passes.run(*mergedModule);
 481
 482     // Run the code generator, and write assembly file
 483     codeGenPasses->doInitialization();
 484
 485     for (Module::iterator
 486            it = mergedModule->begin(), e = mergedModule->end(); it != e; ++it)
 487       if (!it->isDeclaration())
 488         codeGenPasses->run(*it);
 489
 490     codeGenPasses->doFinalization();
 491     return false; // success
 492 }
 493
 494
 495 /// Optimize merged modules using various IPO passes
 496 void LTOCodeGenerator::setCodeGenDebugOptions(const char* options)
 497 {
 498     std::string ops(options);
 499     for (std::string o = getToken(ops); !o.empty(); o = getToken(ops)) {
 500         // ParseCommandLineOptions() expects argv[0] to be program name.
 501         // Lazily add that.
 502         if ( _codegenOptions.empty() )
 503             _codegenOptions.push_back("libLTO");
 504         _codegenOptions.push_back(strdup(o.c_str()));
 505     }
 506 }