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::LinkTimeOptimizer *l = new llvm::LinkTimeOptimizer();
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));
96 else if (Constant *c = dyn_cast<Constant>(value))
97 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
98 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
99 findExternalRefs(c->getOperand(i), references, mangler);
102 /// InputFilename is a LLVM bytecode file. If Module with InputFilename is
103 /// available then return it. Otherwise parseInputFilename.
105 LinkTimeOptimizer::getModule(const std::string &InputFilename)
109 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
110 if (pos != allModules.end())
111 m = allModules[InputFilename.c_str()];
113 m = ParseBytecodeFile(InputFilename);
114 allModules[InputFilename.c_str()] = m;
119 /// InputFilename is a LLVM bytecode file. Reade this bytecode file and
120 /// set corresponding target triplet string.
122 LinkTimeOptimizer::getTargetTriple(const std::string &InputFilename,
123 std::string &targetTriple)
125 Module *m = getModule(InputFilename);
127 targetTriple = m->getTargetTriple();
130 /// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
131 /// Collect global functions and symbol names in symbols vector.
132 /// Collect external references in references vector.
133 /// Return LTO_READ_SUCCESS if there is no error.
135 LinkTimeOptimizer::readLLVMObjectFile(const std::string &InputFilename,
136 NameToSymbolMap &symbols,
137 std::set<std::string> &references)
139 Module *m = getModule(InputFilename);
141 return LTO_READ_FAILURE;
143 // Use mangler to add GlobalPrefix to names to match linker names.
144 // FIXME : Instead of hard coding "-" use GlobalPrefix.
145 Mangler mangler(*m, "_");
147 modules.push_back(m);
149 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
151 LTOLinkageTypes lt = getLTOLinkageType(f);
153 if (!f->isExternal() && lt != LTOInternalLinkage
154 && strncmp (f->getName().c_str(), "llvm.", 5)) {
155 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
156 mangler.getValueName(f));
157 symbols[newSymbol->getMangledName()] = newSymbol;
158 allSymbols[newSymbol->getMangledName()] = newSymbol;
161 // Collect external symbols referenced by this function.
162 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
163 for (BasicBlock::iterator i = b->begin(), be = b->end();
165 for (unsigned count = 0, total = i->getNumOperands();
166 count != total; ++count)
167 findExternalRefs(i->getOperand(count), references, mangler);
170 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
172 LTOLinkageTypes lt = getLTOLinkageType(v);
173 if (!v->isExternal() && lt != LTOInternalLinkage
174 && strncmp (v->getName().c_str(), "llvm.", 5)) {
175 LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
176 mangler.getValueName(v));
177 symbols[newSymbol->getMangledName()] = newSymbol;
178 allSymbols[newSymbol->getMangledName()] = newSymbol;
180 for (unsigned count = 0, total = v->getNumOperands();
181 count != total; ++count)
182 findExternalRefs(v->getOperand(count), references, mangler);
187 return LTO_READ_SUCCESS;
190 /// Optimize module M using various IPO passes. Use exportList to
191 /// internalize selected symbols. Target platform is selected
192 /// based on information available to module M. No new target
193 /// features are selected.
194 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
195 std::vector<const char *> &exportList)
197 // Instantiate the pass manager to organize the passes.
200 // Collect Target info
202 const TargetMachineRegistry::Entry* March =
203 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
206 return LTO_NO_TARGET;
209 std::string Features;
210 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
212 return LTO_NO_TARGET;
214 TargetMachine &Target = *target.get();
216 // Start off with a verification pass.
217 Passes.add(createVerifierPass());
219 // Add an appropriate TargetData instance for this module...
220 Passes.add(new TargetData(*Target.getTargetData()));
222 // Often if the programmer does not specify proper prototypes for the
223 // functions they are calling, they end up calling a vararg version of the
224 // function that does not get a body filled in (the real function has typed
225 // arguments). This pass merges the two functions.
226 Passes.add(createFunctionResolvingPass());
228 // Internalize symbols if export list is nonemty
229 if (!exportList.empty())
230 Passes.add(createInternalizePass(exportList));
232 // Now that we internalized some globals, see if we can hack on them!
233 Passes.add(createGlobalOptimizerPass());
235 // Linking modules together can lead to duplicated global constants, only
236 // keep one copy of each constant...
237 Passes.add(createConstantMergePass());
239 // If the -s command line option was specified, strip the symbols out of the
240 // resulting program to make it smaller. -s is a GLD option that we are
242 Passes.add(createStripSymbolsPass());
244 // Propagate constants at call sites into the functions they call.
245 Passes.add(createIPConstantPropagationPass());
247 // Remove unused arguments from functions...
248 Passes.add(createDeadArgEliminationPass());
250 Passes.add(createFunctionInliningPass()); // Inline small functions
252 Passes.add(createPruneEHPass()); // Remove dead EH info
254 Passes.add(createGlobalDCEPass()); // Remove dead functions
256 // If we didn't decide to inline a function, check to see if we can
257 // transform it to pass arguments by value instead of by reference.
258 Passes.add(createArgumentPromotionPass());
260 // The IPO passes may leave cruft around. Clean up after them.
261 Passes.add(createInstructionCombiningPass());
263 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
265 // Run a few AA driven optimizations here and now, to cleanup the code.
266 Passes.add(createGlobalsModRefPass()); // IP alias analysis
268 Passes.add(createLICMPass()); // Hoist loop invariants
269 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
270 Passes.add(createGCSEPass()); // Remove common subexprs
271 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
273 // Cleanup and simplify the code after the scalar optimizations.
274 Passes.add(createInstructionCombiningPass());
276 // Delete basic blocks, which optimization passes may have killed...
277 Passes.add(createCFGSimplificationPass());
279 // Now that we have optimized the program, discard unreachable functions...
280 Passes.add(createGlobalDCEPass());
282 // Make sure everything is still good.
283 Passes.add(createVerifierPass());
285 FunctionPassManager *CodeGenPasses =
286 new FunctionPassManager(new ExistingModuleProvider(M));
288 CodeGenPasses->add(new TargetData(*Target.getTargetData()));
289 Target.addPassesToEmitFile(*CodeGenPasses, Out, TargetMachine::AssemblyFile,
292 // Run our queue of passes all at once now, efficiently.
295 // Run the code generator, if present.
296 CodeGenPasses->doInitialization();
297 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
298 if (!I->isExternal())
299 CodeGenPasses->run(*I);
301 CodeGenPasses->doFinalization();
303 return LTO_OPT_SUCCESS;
306 ///Link all modules together and optimize them using IPO. Generate
307 /// native object file using OutputFilename
308 /// Return appropriate LTOStatus.
310 LinkTimeOptimizer::optimizeModules(const std::string &OutputFilename,
311 std::vector<const char *> &exportList,
312 std::string &targetTriple)
317 std::ios::openmode io_mode =
318 std::ios::out | std::ios::trunc | std::ios::binary;
319 std::string *errMsg = NULL;
320 Module *bigOne = modules[0];
321 Linker theLinker("LinkTimeOptimizer", bigOne, false);
322 for (unsigned i = 1, e = modules.size(); i != e; ++i)
323 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
324 return LTO_MODULE_MERGE_FAILURE;
327 // Enable this when -save-temps is used
328 std::ofstream Out("big.bc", io_mode);
329 WriteBytecodeToFile(bigOne, Out, true);
332 // Strip leading underscore because it was added to match names
334 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
335 const char *name = exportList[i];
336 NameToSymbolMap::iterator itr = allSymbols.find(name);
337 if (itr != allSymbols.end())
338 exportList[i] = allSymbols[name]->getName();
341 sys::Path tmpAsmFilePath("/tmp/");
343 if (tmpAsmFilePath.createTemporaryFileOnDisk(&ErrMsg)) {
344 std::cerr << "lto: " << ErrMsg << "\n";
345 return LTO_WRITE_FAILURE;
347 sys::RemoveFileOnSignal(tmpAsmFilePath);
349 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
350 if (!asmFile.is_open() || asmFile.bad()) {
351 if (tmpAsmFilePath.exists())
352 tmpAsmFilePath.eraseFromDisk();
353 return LTO_WRITE_FAILURE;
356 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
358 if (status != LTO_OPT_SUCCESS) {
359 tmpAsmFilePath.eraseFromDisk();
363 targetTriple = bigOne->getTargetTriple();
365 // Run GCC to assemble and link the program into native code.
368 // We can't just assemble and link the file with the system assembler
369 // and linker because we don't know where to put the _start symbol.
370 // GCC mysteriously knows how to do it.
371 const sys::Path gcc = FindExecutable("gcc", "/");
373 tmpAsmFilePath.eraseFromDisk();
374 return LTO_ASM_FAILURE;
377 std::vector<const char*> args;
378 args.push_back(gcc.c_str());
379 args.push_back("-c");
380 args.push_back("-x");
381 args.push_back("assembler");
382 args.push_back("-o");
383 args.push_back(OutputFilename.c_str());
384 args.push_back(tmpAsmFilePath.c_str());
387 sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1);
389 tmpAsmFilePath.eraseFromDisk();
391 return LTO_OPT_SUCCESS;