1 //===-lto.cpp - LLVM Link Time Optimizer ----------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Devang Patel and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implementes link time optimization library. This library is
11 // intended to be used by linker to optimize code at link time.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Module.h"
16 #include "llvm/PassManager.h"
17 #include "llvm/Linker.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/SymbolTable.h"
21 #include "llvm/Bytecode/Reader.h"
22 #include "llvm/Bytecode/Writer.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/FileUtilities.h"
25 #include "llvm/Support/SystemUtils.h"
26 #include "llvm/Support/Mangler.h"
27 #include "llvm/System/Program.h"
28 #include "llvm/System/Signals.h"
29 #include "llvm/Analysis/Passes.h"
30 #include "llvm/Analysis/Verifier.h"
31 #include "llvm/Target/SubtargetFeature.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Target/TargetMachine.h"
34 #include "llvm/Target/TargetMachineRegistry.h"
35 #include "llvm/Transforms/IPO.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Analysis/LoadValueNumbering.h"
38 #include "llvm/LinkTimeOptimizer.h"
45 llvm::LinkTimeOptimizer *createLLVMOptimizer()
47 llvm::LTO *l = new llvm::LTO();
53 /// If symbol is not used then make it internal and let optimizer takes
55 void LLVMSymbol::mayBeNotUsed() {
56 gv->setLinkage(GlobalValue::InternalLinkage);
60 // FIXME : Take advantage of GlobalPrefix from AsmPrinter
61 static const char *addUnderscore(const char *name) {
62 size_t namelen = strlen(name);
63 char *symName = (char*)malloc(namelen+2);
65 strcpy(&symName[1], name);
69 // Map LLVM LinkageType to LTO LinakgeType
70 static LTOLinkageTypes
71 getLTOLinkageType(GlobalValue *v)
74 if (v->hasExternalLinkage())
75 lt = LTOExternalLinkage;
76 else if (v->hasLinkOnceLinkage())
77 lt = LTOLinkOnceLinkage;
78 else if (v->hasWeakLinkage())
81 // Otherwise it is internal linkage for link time optimizer
82 lt = LTOInternalLinkage;
86 // Find exeternal symbols referenced by VALUE. This is a recursive function.
88 findExternalRefs(Value *value, std::set<std::string> &references,
91 if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) {
92 LTOLinkageTypes lt = getLTOLinkageType(gv);
93 if (lt != LTOInternalLinkage && strncmp (gv->getName().c_str(), "llvm.", 5))
94 references.insert(mangler.getValueName(gv));
97 // GlobalValue, even with InternalLinkage type, may have operands with
98 // ExternalLinkage type. Do not ignore these operands.
99 if (Constant *c = dyn_cast<Constant>(value))
100 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
101 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
102 findExternalRefs(c->getOperand(i), references, mangler);
105 /// If Moduel with InputFilename is available then remove it.
107 LTO::removeModule (const std::string &InputFilename)
109 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
110 if (pos != allModules.end()) {
111 Module *m = allModules[InputFilename.c_str()];
112 allModules.erase(pos);
117 /// InputFilename is a LLVM bytecode file. If Module with InputFilename is
118 /// available then return it. Otherwise parseInputFilename.
120 LTO::getModule(const std::string &InputFilename)
124 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
125 if (pos != allModules.end())
126 m = allModules[InputFilename.c_str()];
128 m = ParseBytecodeFile(InputFilename);
129 allModules[InputFilename.c_str()] = m;
134 /// InputFilename is a LLVM bytecode file. Reade this bytecode file and
135 /// set corresponding target triplet string.
137 LTO::getTargetTriple(const std::string &InputFilename,
138 std::string &targetTriple)
140 Module *m = getModule(InputFilename);
142 targetTriple = m->getTargetTriple();
145 /// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
146 /// Collect global functions and symbol names in symbols vector.
147 /// Collect external references in references vector.
148 /// Return LTO_READ_SUCCESS if there is no error.
150 LTO::readLLVMObjectFile(const std::string &InputFilename,
151 NameToSymbolMap &symbols,
152 std::set<std::string> &references)
154 Module *m = getModule(InputFilename);
156 return LTO_READ_FAILURE;
158 // Use mangler to add GlobalPrefix to names to match linker names.
159 // FIXME : Instead of hard coding "-" use GlobalPrefix.
160 Mangler mangler(*m, "_");
162 modules.push_back(m);
164 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
166 LTOLinkageTypes lt = getLTOLinkageType(f);
168 if (!f->isExternal() && lt != LTOInternalLinkage
169 && strncmp (f->getName().c_str(), "llvm.", 5)) {
170 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
171 mangler.getValueName(f));
172 symbols[newSymbol->getMangledName()] = newSymbol;
173 allSymbols[newSymbol->getMangledName()] = newSymbol;
176 // Collect external symbols referenced by this function.
177 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
178 for (BasicBlock::iterator i = b->begin(), be = b->end();
180 for (unsigned count = 0, total = i->getNumOperands();
181 count != total; ++count)
182 findExternalRefs(i->getOperand(count), references, mangler);
185 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
187 LTOLinkageTypes lt = getLTOLinkageType(v);
188 if (!v->isExternal() && lt != LTOInternalLinkage
189 && strncmp (v->getName().c_str(), "llvm.", 5)) {
190 LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
191 mangler.getValueName(v));
192 symbols[newSymbol->getMangledName()] = newSymbol;
193 allSymbols[newSymbol->getMangledName()] = newSymbol;
195 for (unsigned count = 0, total = v->getNumOperands();
196 count != total; ++count)
197 findExternalRefs(v->getOperand(count), references, mangler);
202 return LTO_READ_SUCCESS;
205 /// Optimize module M using various IPO passes. Use exportList to
206 /// internalize selected symbols. Target platform is selected
207 /// based on information available to module M. No new target
208 /// features are selected.
209 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
210 std::vector<const char *> &exportList)
212 // Instantiate the pass manager to organize the passes.
215 // Collect Target info
217 const TargetMachineRegistry::Entry* March =
218 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
221 return LTO_NO_TARGET;
224 std::string Features;
225 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
227 return LTO_NO_TARGET;
229 TargetMachine &Target = *target.get();
231 // Start off with a verification pass.
232 Passes.add(createVerifierPass());
234 // Add an appropriate TargetData instance for this module...
235 Passes.add(new TargetData(*Target.getTargetData()));
237 // Often if the programmer does not specify proper prototypes for the
238 // functions they are calling, they end up calling a vararg version of the
239 // function that does not get a body filled in (the real function has typed
240 // arguments). This pass merges the two functions.
241 Passes.add(createFunctionResolvingPass());
243 // Internalize symbols if export list is nonemty
244 if (!exportList.empty())
245 Passes.add(createInternalizePass(exportList));
247 // Now that we internalized some globals, see if we can hack on them!
248 Passes.add(createGlobalOptimizerPass());
250 // Linking modules together can lead to duplicated global constants, only
251 // keep one copy of each constant...
252 Passes.add(createConstantMergePass());
254 // If the -s command line option was specified, strip the symbols out of the
255 // resulting program to make it smaller. -s is a GLD option that we are
257 Passes.add(createStripSymbolsPass());
259 // Propagate constants at call sites into the functions they call.
260 Passes.add(createIPConstantPropagationPass());
262 // Remove unused arguments from functions...
263 Passes.add(createDeadArgEliminationPass());
265 Passes.add(createFunctionInliningPass()); // Inline small functions
267 Passes.add(createPruneEHPass()); // Remove dead EH info
269 Passes.add(createGlobalDCEPass()); // Remove dead functions
271 // If we didn't decide to inline a function, check to see if we can
272 // transform it to pass arguments by value instead of by reference.
273 Passes.add(createArgumentPromotionPass());
275 // The IPO passes may leave cruft around. Clean up after them.
276 Passes.add(createInstructionCombiningPass());
278 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
280 // Run a few AA driven optimizations here and now, to cleanup the code.
281 Passes.add(createGlobalsModRefPass()); // IP alias analysis
283 Passes.add(createLICMPass()); // Hoist loop invariants
284 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
285 Passes.add(createGCSEPass()); // Remove common subexprs
286 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
288 // Cleanup and simplify the code after the scalar optimizations.
289 Passes.add(createInstructionCombiningPass());
291 // Delete basic blocks, which optimization passes may have killed...
292 Passes.add(createCFGSimplificationPass());
294 // Now that we have optimized the program, discard unreachable functions...
295 Passes.add(createGlobalDCEPass());
297 // Make sure everything is still good.
298 Passes.add(createVerifierPass());
300 FunctionPassManager *CodeGenPasses =
301 new FunctionPassManager(new ExistingModuleProvider(M));
303 CodeGenPasses->add(new TargetData(*Target.getTargetData()));
304 Target.addPassesToEmitFile(*CodeGenPasses, Out, TargetMachine::AssemblyFile,
307 // Run our queue of passes all at once now, efficiently.
310 // Run the code generator, if present.
311 CodeGenPasses->doInitialization();
312 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
313 if (!I->isExternal())
314 CodeGenPasses->run(*I);
316 CodeGenPasses->doFinalization();
318 return LTO_OPT_SUCCESS;
321 ///Link all modules together and optimize them using IPO. Generate
322 /// native object file using OutputFilename
323 /// Return appropriate LTOStatus.
325 LTO::optimizeModules(const std::string &OutputFilename,
326 std::vector<const char *> &exportList,
327 std::string &targetTriple)
332 std::ios::openmode io_mode =
333 std::ios::out | std::ios::trunc | std::ios::binary;
334 std::string *errMsg = NULL;
335 Module *bigOne = modules[0];
336 Linker theLinker("LinkTimeOptimizer", bigOne, false);
337 for (unsigned i = 1, e = modules.size(); i != e; ++i)
338 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
339 return LTO_MODULE_MERGE_FAILURE;
342 // Enable this when -save-temps is used
343 std::ofstream Out("big.bc", io_mode);
344 WriteBytecodeToFile(bigOne, Out, true);
347 // Strip leading underscore because it was added to match names
349 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
350 const char *name = exportList[i];
351 NameToSymbolMap::iterator itr = allSymbols.find(name);
352 if (itr != allSymbols.end())
353 exportList[i] = allSymbols[name]->getName();
358 sys::Path TempDir = sys::Path::GetTemporaryDirectory(&ErrMsg);
359 if (TempDir.isEmpty()) {
360 std::cerr << "lto: " << ErrMsg << "\n";
361 return LTO_WRITE_FAILURE;
363 sys::Path tmpAsmFilePath(TempDir);
364 if (!tmpAsmFilePath.appendComponent("lto")) {
365 std::cerr << "lto: " << ErrMsg << "\n";
366 TempDir.eraseFromDisk(true);
367 return LTO_WRITE_FAILURE;
369 if (tmpAsmFilePath.createTemporaryFileOnDisk(&ErrMsg)) {
370 std::cerr << "lto: " << ErrMsg << "\n";
371 TempDir.eraseFromDisk(true);
372 return LTO_WRITE_FAILURE;
374 sys::RemoveFileOnSignal(tmpAsmFilePath);
376 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
377 if (!asmFile.is_open() || asmFile.bad()) {
378 if (tmpAsmFilePath.exists()) {
379 tmpAsmFilePath.eraseFromDisk();
380 TempDir.eraseFromDisk(true);
382 return LTO_WRITE_FAILURE;
385 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
387 if (status != LTO_OPT_SUCCESS) {
388 tmpAsmFilePath.eraseFromDisk();
389 TempDir.eraseFromDisk(true);
393 targetTriple = bigOne->getTargetTriple();
395 // Run GCC to assemble and link the program into native code.
398 // We can't just assemble and link the file with the system assembler
399 // and linker because we don't know where to put the _start symbol.
400 // GCC mysteriously knows how to do it.
401 const sys::Path gcc = sys::Program::FindProgramByName("gcc");
403 tmpAsmFilePath.eraseFromDisk();
404 TempDir.eraseFromDisk(true);
405 return LTO_ASM_FAILURE;
408 std::vector<const char*> args;
409 args.push_back(gcc.c_str());
410 args.push_back("-c");
411 args.push_back("-x");
412 args.push_back("assembler");
413 args.push_back("-o");
414 args.push_back(OutputFilename.c_str());
415 args.push_back(tmpAsmFilePath.c_str());
418 if (sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1, &ErrMsg)) {
419 std::cerr << "lto: " << ErrMsg << "\n";
420 return LTO_ASM_FAILURE;
423 tmpAsmFilePath.eraseFromDisk();
424 TempDir.eraseFromDisk(true);
426 return LTO_OPT_SUCCESS;