LTOCodeGenerator.cpp revision cb3718832375a581c5ea23f15918f3ea447a446c
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
54using namespace llvm;
55
56static cl::opt<bool> DisableInline("disable-inlining",
57  cl::desc("Do not run the inliner pass"));
58
59
60const 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
70LTOCodeGenerator::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
79LTOCodeGenerator::~LTOCodeGenerator()
80{
81    delete _target;
82    delete _nativeObjectFile;
83}
84
85
86
87bool LTOCodeGenerator::addModule(LTOModule* mod, std::string& errMsg)
88{
89    return _linker.LinkInModule(mod->getLLVVMModule(), &errMsg);
90}
91
92
93bool 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
109bool 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
123void LTOCodeGenerator::addMustPreserveSymbol(const char* sym)
124{
125    _mustPreserveSymbols[sym] = 1;
126}
127
128
129bool 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
157const 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    std::string error;
167    bool genResult = false;
168    {
169      raw_fd_ostream asmFile(uniqueAsmPath.c_str(), error);
170      genResult = this->generateAssemblyCode(asmFile, errMsg);
171    }
172    if ( genResult ) {
173        if ( uniqueAsmPath.exists() )
174            uniqueAsmPath.eraseFromDisk();
175        return NULL;
176    }
177
178    // make unique temp .o file to put generated object file
179    sys::PathWithStatus uniqueObjPath("lto-llvm.o");
180    if ( uniqueObjPath.createTemporaryFileOnDisk(true, &errMsg) ) {
181        if ( uniqueAsmPath.exists() )
182            uniqueAsmPath.eraseFromDisk();
183        return NULL;
184    }
185    sys::RemoveFileOnSignal(uniqueObjPath);
186
187    // assemble the assembly code
188    const std::string& uniqueObjStr = uniqueObjPath.toString();
189    bool asmResult = this->assemble(uniqueAsmPath.toString(),
190                                                        uniqueObjStr, errMsg);
191    if ( !asmResult ) {
192        // remove old buffer if compile() called twice
193        delete _nativeObjectFile;
194
195        // read .o file into memory buffer
196        _nativeObjectFile = MemoryBuffer::getFile(uniqueObjStr.c_str(),&errMsg);
197    }
198
199    // remove temp files
200    uniqueAsmPath.eraseFromDisk();
201    uniqueObjPath.eraseFromDisk();
202
203    // return buffer, unless error
204    if ( _nativeObjectFile == NULL )
205        return NULL;
206    *length = _nativeObjectFile->getBufferSize();
207    return _nativeObjectFile->getBufferStart();
208}
209
210
211bool LTOCodeGenerator::assemble(const std::string& asmPath,
212                                const std::string& objPath, std::string& errMsg)
213{
214    // find compiler driver
215    const sys::Path gcc = sys::Program::FindProgramByName("gcc");
216    if ( gcc.isEmpty() ) {
217        errMsg = "can't locate gcc";
218        return true;
219    }
220
221    // build argument list
222    std::vector<const char*> args;
223    std::string targetTriple = _linker.getModule()->getTargetTriple();
224    args.push_back(gcc.c_str());
225    if ( targetTriple.find("darwin") != targetTriple.size() ) {
226        if (strncmp(targetTriple.c_str(), "i686-apple-", 11) == 0) {
227            args.push_back("-arch");
228            args.push_back("i386");
229        }
230        else if (strncmp(targetTriple.c_str(), "x86_64-apple-", 13) == 0) {
231            args.push_back("-arch");
232            args.push_back("x86_64");
233        }
234        else if (strncmp(targetTriple.c_str(), "powerpc-apple-", 14) == 0) {
235            args.push_back("-arch");
236            args.push_back("ppc");
237        }
238        else if (strncmp(targetTriple.c_str(), "powerpc64-apple-", 16) == 0) {
239            args.push_back("-arch");
240            args.push_back("ppc64");
241        }
242    }
243    args.push_back("-c");
244    args.push_back("-x");
245    args.push_back("assembler");
246    args.push_back("-o");
247    args.push_back(objPath.c_str());
248    args.push_back(asmPath.c_str());
249    args.push_back(0);
250
251    // invoke assembler
252    if ( sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 0, 0, &errMsg) ) {
253        errMsg = "error in assembly";
254        return true;
255    }
256    return false; // success
257}
258
259
260
261bool LTOCodeGenerator::determineTarget(std::string& errMsg)
262{
263    if ( _target == NULL ) {
264        // create target machine from info for merged modules
265        Module* mergedModule = _linker.getModule();
266        const TargetMachineRegistry::entry* march =
267          TargetMachineRegistry::getClosestStaticTargetForModule(
268                                                       *mergedModule, errMsg);
269        if ( march == NULL )
270            return true;
271
272        // construct LTModule, hand over ownership of module and target
273        std::string FeatureStr =
274          getFeatureString(_linker.getModule()->getTargetTriple().c_str());
275        _target = march->CtorFn(*mergedModule, FeatureStr.c_str());
276    }
277    return false;
278}
279
280void LTOCodeGenerator::applyScopeRestrictions()
281{
282    if ( !_scopeRestrictionsDone ) {
283        Module* mergedModule = _linker.getModule();
284
285        // Start off with a verification pass.
286        PassManager passes;
287        passes.add(createVerifierPass());
288
289        // mark which symbols can not be internalized
290        if ( !_mustPreserveSymbols.empty() ) {
291            Mangler mangler(*mergedModule,
292                                _target->getTargetAsmInfo()->getGlobalPrefix());
293            std::vector<const char*> mustPreserveList;
294            for (Module::iterator f = mergedModule->begin(),
295                                        e = mergedModule->end(); f != e; ++f) {
296                if ( !f->isDeclaration()
297                  && _mustPreserveSymbols.count(mangler.getValueName(f)) )
298                    mustPreserveList.push_back(::strdup(f->getName().c_str()));
299            }
300            for (Module::global_iterator v = mergedModule->global_begin(),
301                                 e = mergedModule->global_end(); v !=  e; ++v) {
302                if ( !v->isDeclaration()
303                  && _mustPreserveSymbols.count(mangler.getValueName(v)) )
304                    mustPreserveList.push_back(::strdup(v->getName().c_str()));
305            }
306            passes.add(createInternalizePass(mustPreserveList));
307        }
308        // apply scope restrictions
309        passes.run(*mergedModule);
310
311        _scopeRestrictionsDone = true;
312    }
313}
314
315/// Optimize merged modules using various IPO passes
316bool LTOCodeGenerator::generateAssemblyCode(raw_ostream& out,
317                                            std::string& errMsg)
318{
319    if (  this->determineTarget(errMsg) )
320        return true;
321
322    // mark which symbols can not be internalized
323    this->applyScopeRestrictions();
324
325    Module* mergedModule = _linker.getModule();
326
327     // If target supports exception handling then enable it now.
328    if ( _target->getTargetAsmInfo()->doesSupportExceptionHandling() )
329        llvm::ExceptionHandling = true;
330
331    // set codegen model
332    switch( _codeModel ) {
333        case LTO_CODEGEN_PIC_MODEL_STATIC:
334            _target->setRelocationModel(Reloc::Static);
335            break;
336        case LTO_CODEGEN_PIC_MODEL_DYNAMIC:
337            _target->setRelocationModel(Reloc::PIC_);
338            break;
339        case LTO_CODEGEN_PIC_MODEL_DYNAMIC_NO_PIC:
340            _target->setRelocationModel(Reloc::DynamicNoPIC);
341            break;
342    }
343
344    // if options were requested, set them
345    if ( !_codegenOptions.empty() )
346        cl::ParseCommandLineOptions(_codegenOptions.size(),
347                                                (char**)&_codegenOptions[0]);
348
349    // Instantiate the pass manager to organize the passes.
350    PassManager passes;
351
352    // Start off with a verification pass.
353    passes.add(createVerifierPass());
354
355    // Add an appropriate TargetData instance for this module...
356    passes.add(new TargetData(*_target->getTargetData()));
357
358    // Propagate constants at call sites into the functions they call.  This
359    // opens opportunities for globalopt (and inlining) by substituting function
360    // pointers passed as arguments to direct uses of functions.
361    passes.add(createIPSCCPPass());
362
363    // Now that we internalized some globals, see if we can hack on them!
364    passes.add(createGlobalOptimizerPass());
365
366    // Linking modules together can lead to duplicated global constants, only
367    // keep one copy of each constant...
368    passes.add(createConstantMergePass());
369
370    // Remove unused arguments from functions...
371    passes.add(createDeadArgEliminationPass());
372
373    // Reduce the code after globalopt and ipsccp.  Both can open up significant
374    // simplification opportunities, and both can propagate functions through
375    // function pointers.  When this happens, we often have to resolve varargs
376    // calls, etc, so let instcombine do this.
377    passes.add(createInstructionCombiningPass());
378    if (!DisableInline)
379        passes.add(createFunctionInliningPass()); // Inline small functions
380    passes.add(createPruneEHPass());              // Remove dead EH info
381    passes.add(createGlobalDCEPass());            // Remove dead functions
382
383    // If we didn't decide to inline a function, check to see if we can
384    // transform it to pass arguments by value instead of by reference.
385    passes.add(createArgumentPromotionPass());
386
387    // The IPO passes may leave cruft around.  Clean up after them.
388    passes.add(createInstructionCombiningPass());
389    passes.add(createJumpThreadingPass());        // Thread jumps.
390    passes.add(createScalarReplAggregatesPass()); // Break up allocas
391
392    // Run a few AA driven optimizations here and now, to cleanup the code.
393    passes.add(createGlobalsModRefPass());        // IP alias analysis
394    passes.add(createLICMPass());                 // Hoist loop invariants
395    passes.add(createGVNPass());                  // Remove common subexprs
396    passes.add(createMemCpyOptPass());            // Remove dead memcpy's
397    passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
398
399    // Cleanup and simplify the code after the scalar optimizations.
400    passes.add(createInstructionCombiningPass());
401    passes.add(createJumpThreadingPass());        // Thread jumps.
402    passes.add(createPromoteMemoryToRegisterPass()); // Cleanup after threading.
403
404
405    // Delete basic blocks, which optimization passes may have killed...
406    passes.add(createCFGSimplificationPass());
407
408    // Now that we have optimized the program, discard unreachable functions...
409    passes.add(createGlobalDCEPass());
410
411    // Make sure everything is still good.
412    passes.add(createVerifierPass());
413
414    FunctionPassManager* codeGenPasses =
415            new FunctionPassManager(new ExistingModuleProvider(mergedModule));
416
417    codeGenPasses->add(new TargetData(*_target->getTargetData()));
418
419    MachineCodeEmitter* mce = NULL;
420
421    switch (_target->addPassesToEmitFile(*codeGenPasses, out,
422                                      TargetMachine::AssemblyFile, true)) {
423        case FileModel::MachOFile:
424            mce = AddMachOWriter(*codeGenPasses, out, *_target);
425            break;
426        case FileModel::ElfFile:
427            mce = AddELFWriter(*codeGenPasses, out, *_target);
428            break;
429        case FileModel::AsmFile:
430            break;
431        case FileModel::Error:
432        case FileModel::None:
433            errMsg = "target file type not supported";
434            return true;
435    }
436
437    if (_target->addPassesToEmitFileFinish(*codeGenPasses, mce, true)) {
438        errMsg = "target does not support generation of this file type";
439        return true;
440    }
441
442    // Run our queue of passes all at once now, efficiently.
443    passes.run(*mergedModule);
444
445    // Run the code generator, and write assembly file
446    codeGenPasses->doInitialization();
447
448    for (Module::iterator
449           it = mergedModule->begin(), e = mergedModule->end(); it != e; ++it)
450      if (!it->isDeclaration())
451        codeGenPasses->run(*it);
452
453    codeGenPasses->doFinalization();
454    return false; // success
455}
456
457
458/// Optimize merged modules using various IPO passes
459void LTOCodeGenerator::setCodeGenDebugOptions(const char* options)
460{
461    std::string ops(options);
462    for (std::string o = getToken(ops); !o.empty(); o = getToken(ops)) {
463        // ParseCommandLineOptions() expects argv[0] to be program name.
464        // Lazily add that.
465        if ( _codegenOptions.empty() )
466            _codegenOptions.push_back("libLTO");
467        _codegenOptions.push_back(strdup(o.c_str()));
468    }
469}
470