tools/lto2/LTOCodeGenerator.cpp

//===-LTOCodeGenerator.cpp - LLVM Link Time Optimizer ---------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
// 
//===----------------------------------------------------------------------===//
//
// This file implements the Link Time Optimization library. This library is 
// intended to be used by linker to optimize code at link time.
//
//===----------------------------------------------------------------------===//

#include "LTOModule.h"
#include "LTOCodeGenerator.h"


#include "llvm/Module.h"
#include "llvm/PassManager.h"
#include "llvm/Linker.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/ModuleProvider.h"
#include "llvm/Bitcode/ReaderWriter.h"
#include "llvm/Support/SystemUtils.h"
#include "llvm/Support/Mangler.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/System/Signals.h"
#include "llvm/Analysis/Passes.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/Verifier.h"
#include "llvm/Analysis/LoadValueNumbering.h"
#include "llvm/CodeGen/FileWriters.h"
#include "llvm/Target/SubtargetFeature.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetMachineRegistry.h"
#include "llvm/Target/TargetAsmInfo.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Config/config.h"


#include <fstream>
#include <unistd.h>
#include <stdlib.h>
#include <fcntl.h>


using namespace llvm;



const char* LTOCodeGenerator::getVersionString()
{
#ifdef LLVM_VERSION_INFO
    return PACKAGE_NAME " version " PACKAGE_VERSION ", " LLVM_VERSION_INFO;
#else
    return PACKAGE_NAME " version " PACKAGE_VERSION;
#endif
}


LTOCodeGenerator::LTOCodeGenerator() 
    : _linker("LinkTimeOptimizer", "ld-temp.o"), _target(NULL),
      _emitDwarfDebugInfo(false), _scopeRestrictionsDone(false),
      _codeModel(LTO_CODEGEN_PIC_MODEL_DYNAMIC),
      _nativeObjectFile(NULL)
{

}

LTOCodeGenerator::~LTOCodeGenerator()
{
    delete _target;
    delete _nativeObjectFile;
}



bool LTOCodeGenerator::addModule(LTOModule* mod, std::string& errMsg)
{
    return _linker.LinkInModule(mod->getLLVVMModule(), &errMsg);
}
    

bool LTOCodeGenerator::setDebugInfo(lto_debug_model debug, std::string& errMsg)
{
    switch (debug) {
        case LTO_DEBUG_MODEL_NONE:
            _emitDwarfDebugInfo = false;
            return false;
            
        case LTO_DEBUG_MODEL_DWARF:
            _emitDwarfDebugInfo = true;
            return false;
    }
    errMsg = "unknown debug format";
    return true;
}


bool LTOCodeGenerator::setCodePICModel(lto_codegen_model model, 
                                                        std::string& errMsg)
{
    switch (model) {
        case LTO_CODEGEN_PIC_MODEL_STATIC:
        case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
        case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
            _codeModel = model;
            return false;
    }
    errMsg = "unknown pic model";
    return true;
}


void LTOCodeGenerator::addMustPreserveSymbol(const char* sym)
{
    _mustPreserveSymbols[sym] = 1;
}


bool LTOCodeGenerator::writeMergedModules(const char* path, std::string& errMsg)
{
    if ( this->determineTarget(errMsg) ) 
        return true;

    // mark which symbols can not be internalized 
    this->applyScopeRestrictions();

    // create output file
    std::ofstream out(path, std::ios_base::out|std::ios::trunc|std::ios::binary);
    if ( out.fail() ) {
        errMsg = "could not open bitcode file for writing: ";
        errMsg += path;
        return true;
    }
    
    // write bitcode to it
    WriteBitcodeToFile(_linker.getModule(), out);
    if ( out.fail() ) {
        errMsg = "could not write bitcode file: ";
        errMsg += path;
        return true;
    }
    
    return false;
}


const void* LTOCodeGenerator::compile(size_t* length, std::string& errMsg)
{
    // make unique temp .s file to put generated assembly code
    sys::Path uniqueAsmPath("lto-llvm.s");
    if ( uniqueAsmPath.createTemporaryFileOnDisk(true, &errMsg) )
        return NULL;
    sys::RemoveFileOnSignal(uniqueAsmPath);
       
    // generate assembly code
    std::ofstream asmFile(uniqueAsmPath.c_str());
    bool genResult = this->generateAssemblyCode(asmFile, errMsg);
    asmFile.close();
    if ( genResult ) {
        if ( uniqueAsmPath.exists() )
            uniqueAsmPath.eraseFromDisk();
        return NULL;
    }
    
    // make unique temp .o file to put generated object file
    sys::PathWithStatus uniqueObjPath("lto-llvm.o");
    if ( uniqueObjPath.createTemporaryFileOnDisk(true, &errMsg) ) {
        if ( uniqueAsmPath.exists() )
            uniqueAsmPath.eraseFromDisk();
        return NULL;
    }
    sys::RemoveFileOnSignal(uniqueObjPath);

    // assemble the assembly code
    const std::string& uniqueObjStr = uniqueObjPath.toString();
    bool asmResult = this->assemble(uniqueAsmPath.toString(), 
                                                        uniqueObjStr, errMsg);
    if ( !asmResult ) {
        // remove old buffer if compile() called twice
        delete _nativeObjectFile;
        
        // read .o file into memory buffer
        _nativeObjectFile = MemoryBuffer::getFile(uniqueObjStr.c_str(),&errMsg);
    }

    // remove temp files
    uniqueAsmPath.eraseFromDisk();
    uniqueObjPath.eraseFromDisk();

    // return buffer, unless error
    if ( _nativeObjectFile == NULL )
        return NULL;
    *length = _nativeObjectFile->getBufferSize();
    return _nativeObjectFile->getBufferStart();
}


bool LTOCodeGenerator::assemble(const std::string& asmPath, 
                                const std::string& objPath, std::string& errMsg)
{
    // find compiler driver
    const sys::Path gcc = sys::Program::FindProgramByName("gcc");
    if ( gcc.isEmpty() ) {
        errMsg = "can't locate gcc";
        return true;
    }

    // build argument list
    std::vector<const char*> args;
    std::string targetTriple = _linker.getModule()->getTargetTriple();
    args.push_back(gcc.c_str());
    if ( targetTriple.find("darwin") != targetTriple.size() ) {
        if (strncmp(targetTriple.c_str(), "i686-apple-", 11) == 0) {
            args.push_back("-arch");
            args.push_back("i386");
        }
        else if (strncmp(targetTriple.c_str(), "x86_64-apple-", 13) == 0) {
            args.push_back("-arch");
            args.push_back("x86_64");
        }
        else if (strncmp(targetTriple.c_str(), "powerpc-apple-", 14) == 0) {
            args.push_back("-arch");
            args.push_back("ppc");
        }
        else if (strncmp(targetTriple.c_str(), "powerpc64-apple-", 16) == 0) {
            args.push_back("-arch");
            args.push_back("ppc64");
        }
    }
    args.push_back("-c");
    args.push_back("-x");
    args.push_back("assembler");
    args.push_back("-o");
    args.push_back(objPath.c_str());
    args.push_back(asmPath.c_str());
    args.push_back(0);

    // invoke assembler
    if ( sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 0, 0, &errMsg) ) {
        errMsg = "error in assembly";    
        return true;
    }
    return false; // success
}



bool LTOCodeGenerator::determineTarget(std::string& errMsg)
{
    if ( _target == NULL ) {
        // create target machine from info for merged modules
        Module* mergedModule = _linker.getModule();
        const TargetMachineRegistry::entry* march = 
          TargetMachineRegistry::getClosestStaticTargetForModule(
                                                       *mergedModule, errMsg);
        if ( march == NULL )
            return true;

        // construct LTModule, hand over ownership of module and target
        //
        // FIXME: This is an inelegant way of specifying the features of a
        // subtarget. It would be better if we could encode this information
        // into the IR. See <rdar://5972456>.
        SubtargetFeatures Features;
        std::string FeatureStr;
        std::string TargetTriple = _linker.getModule()->getTargetTriple();

        if (strncmp(TargetTriple.c_str(), "powerpc-apple-", 14) == 0) {
          Features.AddFeature("altivec", true);
        } else if (strncmp(TargetTriple.c_str(), "powerpc64-apple-", 16) == 0) {
          Features.AddFeature("64bit", true);
          Features.AddFeature("altivec", true);
        }

        _target = march->CtorFn(*mergedModule, Features.getString());
    }
    return false;
}

void LTOCodeGenerator::applyScopeRestrictions()
{
    if ( !_scopeRestrictionsDone ) {
        Module* mergedModule = _linker.getModule();

        // Start off with a verification pass.
        PassManager passes;
        passes.add(createVerifierPass());

        // mark which symbols can not be internalized 
        if ( !_mustPreserveSymbols.empty() ) {
            Mangler mangler(*mergedModule, 
                                _target->getTargetAsmInfo()->getGlobalPrefix());
            std::vector<const char*> mustPreserveList;
            for (Module::iterator f = mergedModule->begin(), 
                                        e = mergedModule->end(); f != e; ++f) {
                if ( !f->isDeclaration() 
                  && _mustPreserveSymbols.count(mangler.getValueName(f)) )
                    mustPreserveList.push_back(::strdup(f->getName().c_str()));
            }
            for (Module::global_iterator v = mergedModule->global_begin(), 
                                 e = mergedModule->global_end(); v !=  e; ++v) {
                if ( !v->isDeclaration()
                  && _mustPreserveSymbols.count(mangler.getValueName(v)) )
                    mustPreserveList.push_back(::strdup(v->getName().c_str()));
            }
            passes.add(createInternalizePass(mustPreserveList));
        }
        // apply scope restrictions
        passes.run(*mergedModule);
        
        _scopeRestrictionsDone = true;
    }
}

/// Optimize merged modules using various IPO passes
bool LTOCodeGenerator::generateAssemblyCode(std::ostream& out, std::string& errMsg)
{
    if (  this->determineTarget(errMsg) ) 
        return true;

    // mark which symbols can not be internalized 
    this->applyScopeRestrictions();

    Module* mergedModule = _linker.getModule();

     // If target supports exception handling then enable it now.
    if ( _target->getTargetAsmInfo()->doesSupportExceptionHandling() )
        llvm::ExceptionHandling = true;

    // set codegen model
    switch( _codeModel ) {
        case LTO_CODEGEN_PIC_MODEL_STATIC:
            _target->setRelocationModel(Reloc::Static);
            break;
        case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
            _target->setRelocationModel(Reloc::PIC_);
            break;
        case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
            _target->setRelocationModel(Reloc::DynamicNoPIC);
            break;
    }

    // Instantiate the pass manager to organize the passes.
    PassManager passes;

    // Start off with a verification pass.
    passes.add(createVerifierPass());

    // Add an appropriate TargetData instance for this module...
    passes.add(new TargetData(*_target->getTargetData()));
    
    // Propagate constants at call sites into the functions they call.  This
    // opens opportunities for globalopt (and inlining) by substituting function
    // pointers passed as arguments to direct uses of functions.  
    passes.add(createIPSCCPPass());

    // Now that we internalized some globals, see if we can hack on them!
    passes.add(createGlobalOptimizerPass());

    // Linking modules together can lead to duplicated global constants, only
    // keep one copy of each constant...
    passes.add(createConstantMergePass());

    // Remove unused arguments from functions...
    passes.add(createDeadArgEliminationPass());

    // Reduce the code after globalopt and ipsccp.  Both can open up significant
    // simplification opportunities, and both can propagate functions through
    // function pointers.  When this happens, we often have to resolve varargs
    // calls, etc, so let instcombine do this.
    passes.add(createInstructionCombiningPass());
    passes.add(createFunctionInliningPass());     // Inline small functions
    passes.add(createPruneEHPass());              // Remove dead EH info
    passes.add(createGlobalDCEPass());            // Remove dead functions

    // If we didn't decide to inline a function, check to see if we can
    // transform it to pass arguments by value instead of by reference.
    passes.add(createArgumentPromotionPass());

    // The IPO passes may leave cruft around.  Clean up after them.
    passes.add(createInstructionCombiningPass());
    passes.add(createJumpThreadingPass());        // Thread jumps.
    passes.add(createScalarReplAggregatesPass()); // Break up allocas

    // Run a few AA driven optimizations here and now, to cleanup the code.
    passes.add(createGlobalsModRefPass());        // IP alias analysis
    passes.add(createLICMPass());                 // Hoist loop invariants
    passes.add(createGVNPass());                  // Remove common subexprs
    passes.add(createMemCpyOptPass());            // Remove dead memcpy's
    passes.add(createDeadStoreEliminationPass()); // Nuke dead stores

    // Cleanup and simplify the code after the scalar optimizations.
    passes.add(createInstructionCombiningPass());
    passes.add(createJumpThreadingPass());        // Thread jumps.

    // Delete basic blocks, which optimization passes may have killed...
    passes.add(createCFGSimplificationPass());

    // Now that we have optimized the program, discard unreachable functions...
    passes.add(createGlobalDCEPass());

    // Make sure everything is still good.
    passes.add(createVerifierPass());

    FunctionPassManager* codeGenPasses =
            new FunctionPassManager(new ExistingModuleProvider(mergedModule));

    codeGenPasses->add(new TargetData(*_target->getTargetData()));

    MachineCodeEmitter* mce = NULL;

    switch (_target->addPassesToEmitFile(*codeGenPasses, out,
                                      TargetMachine::AssemblyFile, true)) {
        case FileModel::MachOFile:
            mce = AddMachOWriter(*codeGenPasses, out, *_target);
            break;
        case FileModel::ElfFile:
            mce = AddELFWriter(*codeGenPasses, out, *_target);
            break;
        case FileModel::AsmFile:
            break;
        case FileModel::Error:
        case FileModel::None:
            errMsg = "target file type not supported";
            return true;
    }

    if (_target->addPassesToEmitFileFinish(*codeGenPasses, mce, true)) {
        errMsg = "target does not support generation of this file type";
        return true;
    }

    // Run our queue of passes all at once now, efficiently.
    passes.run(*mergedModule);

    // Run the code generator, and write assembly file
    codeGenPasses->doInitialization();

    for (Module::iterator
           it = mergedModule->begin(), e = mergedModule->end(); it != e; ++it)
      if (!it->isDeclaration())
        codeGenPasses->run(*it);

    codeGenPasses->doFinalization();
    return false; // success
}