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 Module with InputFilename is available then remove it from allModules
106 /// and call delete on it.
108 LTO::removeModule (const std::string &InputFilename)
110 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
111 if (pos == allModules.end())
114 Module *m = pos->second;
115 allModules.erase(pos);
119 /// InputFilename is a LLVM bytecode file. If Module with InputFilename is
120 /// available then return it. Otherwise parseInputFilename.
122 LTO::getModule(const std::string &InputFilename)
126 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
127 if (pos != allModules.end())
128 m = allModules[InputFilename.c_str()];
130 m = ParseBytecodeFile(InputFilename);
131 allModules[InputFilename.c_str()] = m;
136 /// InputFilename is a LLVM bytecode file. Reade this bytecode file and
137 /// set corresponding target triplet string.
139 LTO::getTargetTriple(const std::string &InputFilename,
140 std::string &targetTriple)
142 Module *m = getModule(InputFilename);
144 targetTriple = m->getTargetTriple();
147 /// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
148 /// Collect global functions and symbol names in symbols vector.
149 /// Collect external references in references vector.
150 /// Return LTO_READ_SUCCESS if there is no error.
152 LTO::readLLVMObjectFile(const std::string &InputFilename,
153 NameToSymbolMap &symbols,
154 std::set<std::string> &references)
156 Module *m = getModule(InputFilename);
158 return LTO_READ_FAILURE;
160 // Use mangler to add GlobalPrefix to names to match linker names.
161 // FIXME : Instead of hard coding "-" use GlobalPrefix.
162 Mangler mangler(*m, "_");
164 modules.push_back(m);
166 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
168 LTOLinkageTypes lt = getLTOLinkageType(f);
170 if (!f->isExternal() && lt != LTOInternalLinkage
171 && strncmp (f->getName().c_str(), "llvm.", 5)) {
172 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
173 mangler.getValueName(f));
174 symbols[newSymbol->getMangledName()] = newSymbol;
175 allSymbols[newSymbol->getMangledName()] = newSymbol;
178 // Collect external symbols referenced by this function.
179 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
180 for (BasicBlock::iterator i = b->begin(), be = b->end();
182 for (unsigned count = 0, total = i->getNumOperands();
183 count != total; ++count)
184 findExternalRefs(i->getOperand(count), references, mangler);
187 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
189 LTOLinkageTypes lt = getLTOLinkageType(v);
190 if (!v->isExternal() && lt != LTOInternalLinkage
191 && strncmp (v->getName().c_str(), "llvm.", 5)) {
192 LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
193 mangler.getValueName(v));
194 symbols[newSymbol->getMangledName()] = newSymbol;
195 allSymbols[newSymbol->getMangledName()] = newSymbol;
197 for (unsigned count = 0, total = v->getNumOperands();
198 count != total; ++count)
199 findExternalRefs(v->getOperand(count), references, mangler);
204 return LTO_READ_SUCCESS;
207 /// Optimize module M using various IPO passes. Use exportList to
208 /// internalize selected symbols. Target platform is selected
209 /// based on information available to module M. No new target
210 /// features are selected.
211 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
212 std::vector<const char *> &exportList)
214 // Instantiate the pass manager to organize the passes.
217 // Collect Target info
219 const TargetMachineRegistry::Entry* March =
220 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
223 return LTO_NO_TARGET;
226 std::string Features;
227 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
229 return LTO_NO_TARGET;
231 TargetMachine &Target = *target.get();
233 // Start off with a verification pass.
234 Passes.add(createVerifierPass());
236 // Add an appropriate TargetData instance for this module...
237 Passes.add(new TargetData(*Target.getTargetData()));
239 // Often if the programmer does not specify proper prototypes for the
240 // functions they are calling, they end up calling a vararg version of the
241 // function that does not get a body filled in (the real function has typed
242 // arguments). This pass merges the two functions.
243 Passes.add(createFunctionResolvingPass());
245 // Internalize symbols if export list is nonemty
246 if (!exportList.empty())
247 Passes.add(createInternalizePass(exportList));
249 // Now that we internalized some globals, see if we can hack on them!
250 Passes.add(createGlobalOptimizerPass());
252 // Linking modules together can lead to duplicated global constants, only
253 // keep one copy of each constant...
254 Passes.add(createConstantMergePass());
256 // If the -s command line option was specified, strip the symbols out of the
257 // resulting program to make it smaller. -s is a GLD option that we are
259 Passes.add(createStripSymbolsPass());
261 // Propagate constants at call sites into the functions they call.
262 Passes.add(createIPConstantPropagationPass());
264 // Remove unused arguments from functions...
265 Passes.add(createDeadArgEliminationPass());
267 Passes.add(createFunctionInliningPass()); // Inline small functions
269 Passes.add(createPruneEHPass()); // Remove dead EH info
271 Passes.add(createGlobalDCEPass()); // Remove dead functions
273 // If we didn't decide to inline a function, check to see if we can
274 // transform it to pass arguments by value instead of by reference.
275 Passes.add(createArgumentPromotionPass());
277 // The IPO passes may leave cruft around. Clean up after them.
278 Passes.add(createInstructionCombiningPass());
280 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
282 // Run a few AA driven optimizations here and now, to cleanup the code.
283 Passes.add(createGlobalsModRefPass()); // IP alias analysis
285 Passes.add(createLICMPass()); // Hoist loop invariants
286 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
287 Passes.add(createGCSEPass()); // Remove common subexprs
288 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
290 // Cleanup and simplify the code after the scalar optimizations.
291 Passes.add(createInstructionCombiningPass());
293 // Delete basic blocks, which optimization passes may have killed...
294 Passes.add(createCFGSimplificationPass());
296 // Now that we have optimized the program, discard unreachable functions...
297 Passes.add(createGlobalDCEPass());
299 // Make sure everything is still good.
300 Passes.add(createVerifierPass());
302 FunctionPassManager *CodeGenPasses =
303 new FunctionPassManager(new ExistingModuleProvider(M));
305 CodeGenPasses->add(new TargetData(*Target.getTargetData()));
306 Target.addPassesToEmitFile(*CodeGenPasses, Out, TargetMachine::AssemblyFile,
309 // Run our queue of passes all at once now, efficiently.
312 // Run the code generator, if present.
313 CodeGenPasses->doInitialization();
314 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
315 if (!I->isExternal())
316 CodeGenPasses->run(*I);
318 CodeGenPasses->doFinalization();
320 return LTO_OPT_SUCCESS;
323 ///Link all modules together and optimize them using IPO. Generate
324 /// native object file using OutputFilename
325 /// Return appropriate LTOStatus.
327 LTO::optimizeModules(const std::string &OutputFilename,
328 std::vector<const char *> &exportList,
329 std::string &targetTriple)
334 std::ios::openmode io_mode =
335 std::ios::out | std::ios::trunc | std::ios::binary;
336 std::string *errMsg = NULL;
337 Module *bigOne = modules[0];
338 Linker theLinker("LinkTimeOptimizer", bigOne, false);
339 for (unsigned i = 1, e = modules.size(); i != e; ++i)
340 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
341 return LTO_MODULE_MERGE_FAILURE;
344 // Enable this when -save-temps is used
345 std::ofstream Out("big.bc", io_mode);
346 WriteBytecodeToFile(bigOne, Out, true);
349 // Strip leading underscore because it was added to match names
351 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
352 const char *name = exportList[i];
353 NameToSymbolMap::iterator itr = allSymbols.find(name);
354 if (itr != allSymbols.end())
355 exportList[i] = allSymbols[name]->getName();
360 sys::Path TempDir = sys::Path::GetTemporaryDirectory(&ErrMsg);
361 if (TempDir.isEmpty()) {
362 std::cerr << "lto: " << ErrMsg << "\n";
363 return LTO_WRITE_FAILURE;
365 sys::Path tmpAsmFilePath(TempDir);
366 if (!tmpAsmFilePath.appendComponent("lto")) {
367 std::cerr << "lto: " << ErrMsg << "\n";
368 TempDir.eraseFromDisk(true);
369 return LTO_WRITE_FAILURE;
371 if (tmpAsmFilePath.createTemporaryFileOnDisk(&ErrMsg)) {
372 std::cerr << "lto: " << ErrMsg << "\n";
373 TempDir.eraseFromDisk(true);
374 return LTO_WRITE_FAILURE;
376 sys::RemoveFileOnSignal(tmpAsmFilePath);
378 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
379 if (!asmFile.is_open() || asmFile.bad()) {
380 if (tmpAsmFilePath.exists()) {
381 tmpAsmFilePath.eraseFromDisk();
382 TempDir.eraseFromDisk(true);
384 return LTO_WRITE_FAILURE;
387 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
389 if (status != LTO_OPT_SUCCESS) {
390 tmpAsmFilePath.eraseFromDisk();
391 TempDir.eraseFromDisk(true);
395 targetTriple = bigOne->getTargetTriple();
397 // Run GCC to assemble and link the program into native code.
400 // We can't just assemble and link the file with the system assembler
401 // and linker because we don't know where to put the _start symbol.
402 // GCC mysteriously knows how to do it.
403 const sys::Path gcc = sys::Program::FindProgramByName("gcc");
405 tmpAsmFilePath.eraseFromDisk();
406 TempDir.eraseFromDisk(true);
407 return LTO_ASM_FAILURE;
410 std::vector<const char*> args;
411 args.push_back(gcc.c_str());
412 args.push_back("-c");
413 args.push_back("-x");
414 args.push_back("assembler");
415 args.push_back("-o");
416 args.push_back(OutputFilename.c_str());
417 args.push_back(tmpAsmFilePath.c_str());
420 if (sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1, &ErrMsg)) {
421 std::cerr << "lto: " << ErrMsg << "\n";
422 return LTO_ASM_FAILURE;
425 tmpAsmFilePath.eraseFromDisk();
426 TempDir.eraseFromDisk(true);
428 return LTO_OPT_SUCCESS;