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));
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 /// InputFilename is a LLVM bytecode file. If Module with InputFilename is
106 /// available then return it. Otherwise parseInputFilename.
108 LinkTimeOptimizer::getModule(const std::string &InputFilename)
112 NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
113 if (pos != allModules.end())
114 m = allModules[InputFilename.c_str()];
116 m = ParseBytecodeFile(InputFilename);
117 allModules[InputFilename.c_str()] = m;
122 /// InputFilename is a LLVM bytecode file. Reade this bytecode file and
123 /// set corresponding target triplet string.
125 LinkTimeOptimizer::getTargetTriple(const std::string &InputFilename,
126 std::string &targetTriple)
128 Module *m = getModule(InputFilename);
130 targetTriple = m->getTargetTriple();
133 /// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
134 /// Collect global functions and symbol names in symbols vector.
135 /// Collect external references in references vector.
136 /// Return LTO_READ_SUCCESS if there is no error.
138 LinkTimeOptimizer::readLLVMObjectFile(const std::string &InputFilename,
139 NameToSymbolMap &symbols,
140 std::set<std::string> &references)
142 Module *m = getModule(InputFilename);
144 return LTO_READ_FAILURE;
146 // Use mangler to add GlobalPrefix to names to match linker names.
147 // FIXME : Instead of hard coding "-" use GlobalPrefix.
148 Mangler mangler(*m, "_");
150 modules.push_back(m);
152 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
154 LTOLinkageTypes lt = getLTOLinkageType(f);
156 if (!f->isExternal() && lt != LTOInternalLinkage
157 && strncmp (f->getName().c_str(), "llvm.", 5)) {
158 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
159 mangler.getValueName(f));
160 symbols[newSymbol->getMangledName()] = newSymbol;
161 allSymbols[newSymbol->getMangledName()] = newSymbol;
164 // Collect external symbols referenced by this function.
165 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
166 for (BasicBlock::iterator i = b->begin(), be = b->end();
168 for (unsigned count = 0, total = i->getNumOperands();
169 count != total; ++count)
170 findExternalRefs(i->getOperand(count), references, mangler);
173 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
175 LTOLinkageTypes lt = getLTOLinkageType(v);
176 if (!v->isExternal() && lt != LTOInternalLinkage
177 && strncmp (v->getName().c_str(), "llvm.", 5)) {
178 LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
179 mangler.getValueName(v));
180 symbols[newSymbol->getMangledName()] = newSymbol;
181 allSymbols[newSymbol->getMangledName()] = newSymbol;
183 for (unsigned count = 0, total = v->getNumOperands();
184 count != total; ++count)
185 findExternalRefs(v->getOperand(count), references, mangler);
190 return LTO_READ_SUCCESS;
193 /// Optimize module M using various IPO passes. Use exportList to
194 /// internalize selected symbols. Target platform is selected
195 /// based on information available to module M. No new target
196 /// features are selected.
197 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
198 std::vector<const char *> &exportList)
200 // Instantiate the pass manager to organize the passes.
203 // Collect Target info
205 const TargetMachineRegistry::Entry* March =
206 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
209 return LTO_NO_TARGET;
212 std::string Features;
213 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
215 return LTO_NO_TARGET;
217 TargetMachine &Target = *target.get();
219 // Start off with a verification pass.
220 Passes.add(createVerifierPass());
222 // Add an appropriate TargetData instance for this module...
223 Passes.add(new TargetData(*Target.getTargetData()));
225 // Often if the programmer does not specify proper prototypes for the
226 // functions they are calling, they end up calling a vararg version of the
227 // function that does not get a body filled in (the real function has typed
228 // arguments). This pass merges the two functions.
229 Passes.add(createFunctionResolvingPass());
231 // Internalize symbols if export list is nonemty
232 if (!exportList.empty())
233 Passes.add(createInternalizePass(exportList));
235 // Now that we internalized some globals, see if we can hack on them!
236 Passes.add(createGlobalOptimizerPass());
238 // Linking modules together can lead to duplicated global constants, only
239 // keep one copy of each constant...
240 Passes.add(createConstantMergePass());
242 // If the -s command line option was specified, strip the symbols out of the
243 // resulting program to make it smaller. -s is a GLD option that we are
245 Passes.add(createStripSymbolsPass());
247 // Propagate constants at call sites into the functions they call.
248 Passes.add(createIPConstantPropagationPass());
250 // Remove unused arguments from functions...
251 Passes.add(createDeadArgEliminationPass());
253 Passes.add(createFunctionInliningPass()); // Inline small functions
255 Passes.add(createPruneEHPass()); // Remove dead EH info
257 Passes.add(createGlobalDCEPass()); // Remove dead functions
259 // If we didn't decide to inline a function, check to see if we can
260 // transform it to pass arguments by value instead of by reference.
261 Passes.add(createArgumentPromotionPass());
263 // The IPO passes may leave cruft around. Clean up after them.
264 Passes.add(createInstructionCombiningPass());
266 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
268 // Run a few AA driven optimizations here and now, to cleanup the code.
269 Passes.add(createGlobalsModRefPass()); // IP alias analysis
271 Passes.add(createLICMPass()); // Hoist loop invariants
272 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
273 Passes.add(createGCSEPass()); // Remove common subexprs
274 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
276 // Cleanup and simplify the code after the scalar optimizations.
277 Passes.add(createInstructionCombiningPass());
279 // Delete basic blocks, which optimization passes may have killed...
280 Passes.add(createCFGSimplificationPass());
282 // Now that we have optimized the program, discard unreachable functions...
283 Passes.add(createGlobalDCEPass());
285 // Make sure everything is still good.
286 Passes.add(createVerifierPass());
288 FunctionPassManager *CodeGenPasses =
289 new FunctionPassManager(new ExistingModuleProvider(M));
291 CodeGenPasses->add(new TargetData(*Target.getTargetData()));
292 Target.addPassesToEmitFile(*CodeGenPasses, Out, TargetMachine::AssemblyFile,
295 // Run our queue of passes all at once now, efficiently.
298 // Run the code generator, if present.
299 CodeGenPasses->doInitialization();
300 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
301 if (!I->isExternal())
302 CodeGenPasses->run(*I);
304 CodeGenPasses->doFinalization();
306 return LTO_OPT_SUCCESS;
309 ///Link all modules together and optimize them using IPO. Generate
310 /// native object file using OutputFilename
311 /// Return appropriate LTOStatus.
313 LinkTimeOptimizer::optimizeModules(const std::string &OutputFilename,
314 std::vector<const char *> &exportList,
315 std::string &targetTriple)
320 std::ios::openmode io_mode =
321 std::ios::out | std::ios::trunc | std::ios::binary;
322 std::string *errMsg = NULL;
323 Module *bigOne = modules[0];
324 Linker theLinker("LinkTimeOptimizer", bigOne, false);
325 for (unsigned i = 1, e = modules.size(); i != e; ++i)
326 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
327 return LTO_MODULE_MERGE_FAILURE;
330 // Enable this when -save-temps is used
331 std::ofstream Out("big.bc", io_mode);
332 WriteBytecodeToFile(bigOne, Out, true);
335 // Strip leading underscore because it was added to match names
337 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
338 const char *name = exportList[i];
339 NameToSymbolMap::iterator itr = allSymbols.find(name);
340 if (itr != allSymbols.end())
341 exportList[i] = allSymbols[name]->getName();
344 sys::Path tmpAsmFilePath("/tmp/");
346 if (tmpAsmFilePath.createTemporaryFileOnDisk(&ErrMsg)) {
347 std::cerr << "lto: " << ErrMsg << "\n";
348 return LTO_WRITE_FAILURE;
350 sys::RemoveFileOnSignal(tmpAsmFilePath);
352 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
353 if (!asmFile.is_open() || asmFile.bad()) {
354 if (tmpAsmFilePath.exists())
355 tmpAsmFilePath.eraseFromDisk();
356 return LTO_WRITE_FAILURE;
359 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
361 if (status != LTO_OPT_SUCCESS) {
362 tmpAsmFilePath.eraseFromDisk();
366 targetTriple = bigOne->getTargetTriple();
368 // Run GCC to assemble and link the program into native code.
371 // We can't just assemble and link the file with the system assembler
372 // and linker because we don't know where to put the _start symbol.
373 // GCC mysteriously knows how to do it.
374 const sys::Path gcc = FindExecutable("gcc", "/");
376 tmpAsmFilePath.eraseFromDisk();
377 return LTO_ASM_FAILURE;
380 std::vector<const char*> args;
381 args.push_back(gcc.c_str());
382 args.push_back("-c");
383 args.push_back("-x");
384 args.push_back("assembler");
385 args.push_back("-o");
386 args.push_back(OutputFilename.c_str());
387 args.push_back(tmpAsmFilePath.c_str());
390 sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1);
392 tmpAsmFilePath.eraseFromDisk();
394 return LTO_OPT_SUCCESS;