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/System/Program.h"
27 #include "llvm/System/Signals.h"
28 #include "llvm/Analysis/Passes.h"
29 #include "llvm/Analysis/Verifier.h"
30 #include "llvm/Target/SubtargetFeature.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetMachineRegistry.h"
34 #include "llvm/Transforms/IPO.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Analysis/LoadValueNumbering.h"
37 #include "llvm/LinkTimeOptimizer.h"
44 llvm::LinkTimeOptimizer *createLLVMOptimizer()
46 llvm::LinkTimeOptimizer *l = new llvm::LinkTimeOptimizer();
52 /// If symbol is not used then make it internal and let optimizer takes
54 void LLVMSymbol::mayBeNotUsed() {
55 gv->setLinkage(GlobalValue::InternalLinkage);
59 // FIXME : Take advantage of GlobalPrefix from AsmPrinter
60 static const char *addUnderscore(const char *name) {
61 size_t namelen = strlen(name);
62 char *symName = (char*)malloc(namelen+2);
64 strcpy(&symName[1], name);
68 // Map LLVM LinkageType to LTO LinakgeType
69 static LTOLinkageTypes
70 getLTOLinkageType(GlobalValue *v)
73 if (v->hasExternalLinkage())
74 lt = LTOExternalLinkage;
75 else if (v->hasLinkOnceLinkage())
76 lt = LTOLinkOnceLinkage;
77 else if (v->hasWeakLinkage())
80 // Otherwise it is internal linkage for link time optimizer
81 lt = LTOInternalLinkage;
85 // Find exeternal symbols referenced by VALUE. This is a recursive function.
87 findExternalRefs(Value *value, std::set<const char *> &references) {
89 if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) {
90 LTOLinkageTypes lt = getLTOLinkageType(gv);
91 if (lt != LTOInternalLinkage && strncmp (gv->getName().c_str(), "llvm.", 5))
92 references.insert(addUnderscore(gv->getName().c_str()));
94 else if (Constant *c = dyn_cast<Constant>(value))
95 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
96 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
97 findExternalRefs(c->getOperand(i), references);
100 /// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
101 /// Collect global functions and symbol names in symbols vector.
102 /// Collect external references in references vector.
103 /// Return LTO_READ_SUCCESS if there is no error.
105 LinkTimeOptimizer::readLLVMObjectFile(const std::string &InputFilename,
106 NameToSymbolMap &symbols,
107 std::set<const char *> &references)
109 Module *m = ParseBytecodeFile(InputFilename);
111 return LTO_READ_FAILURE;
113 modules.push_back(m);
115 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
117 LTOLinkageTypes lt = getLTOLinkageType(f);
119 if (!f->isExternal() && lt != LTOInternalLinkage
120 && strncmp (f->getName().c_str(), "llvm.", 5)) {
121 const char *name = addUnderscore(f->getName().c_str());
122 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f);
123 symbols[name] = newSymbol;
124 allSymbols[name] = newSymbol;
127 // Collect external symbols referenced by this function.
128 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
129 for (BasicBlock::iterator i = b->begin(), be = b->end();
131 for (unsigned count = 0, total = i->getNumOperands();
132 count != total; ++count)
133 findExternalRefs(i->getOperand(count), references);
136 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
138 LTOLinkageTypes lt = getLTOLinkageType(v);
139 if (!v->isExternal() && lt != LTOInternalLinkage
140 && strncmp (v->getName().c_str(), "llvm.", 5)) {
141 const char *name = addUnderscore(v->getName().c_str());
142 LLVMSymbol *newSymbol = new LLVMSymbol(lt,v);
143 symbols[name] = newSymbol;
145 for (unsigned count = 0, total = v->getNumOperands();
146 count != total; ++count)
147 findExternalRefs(v->getOperand(count), references);
152 return LTO_READ_SUCCESS;
155 /// Optimize module M using various IPO passes. Use exportList to
156 /// internalize selected symbols. Target platform is selected
157 /// based on information available to module M. No new target
158 /// features are selected.
159 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
160 std::vector<const char *> &exportList)
162 // Instantiate the pass manager to organize the passes.
165 // Collect Target info
167 const TargetMachineRegistry::Entry* March =
168 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
171 return LTO_NO_TARGET;
174 std::string Features;
175 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
177 return LTO_NO_TARGET;
179 TargetMachine &Target = *target.get();
181 // Start off with a verification pass.
182 Passes.add(createVerifierPass());
184 // Add an appropriate TargetData instance for this module...
185 Passes.add(new TargetData(*Target.getTargetData()));
187 // Often if the programmer does not specify proper prototypes for the
188 // functions they are calling, they end up calling a vararg version of the
189 // function that does not get a body filled in (the real function has typed
190 // arguments). This pass merges the two functions.
191 Passes.add(createFunctionResolvingPass());
193 // Internalize symbols if export list is nonemty
194 if (!exportList.empty())
195 Passes.add(createInternalizePass(exportList));
197 // Now that we internalized some globals, see if we can hack on them!
198 Passes.add(createGlobalOptimizerPass());
200 // Linking modules together can lead to duplicated global constants, only
201 // keep one copy of each constant...
202 Passes.add(createConstantMergePass());
204 // If the -s command line option was specified, strip the symbols out of the
205 // resulting program to make it smaller. -s is a GLD option that we are
207 Passes.add(createStripSymbolsPass());
209 // Propagate constants at call sites into the functions they call.
210 Passes.add(createIPConstantPropagationPass());
212 // Remove unused arguments from functions...
213 Passes.add(createDeadArgEliminationPass());
215 Passes.add(createFunctionInliningPass()); // Inline small functions
217 Passes.add(createPruneEHPass()); // Remove dead EH info
219 Passes.add(createGlobalDCEPass()); // Remove dead functions
221 // If we didn't decide to inline a function, check to see if we can
222 // transform it to pass arguments by value instead of by reference.
223 Passes.add(createArgumentPromotionPass());
225 // The IPO passes may leave cruft around. Clean up after them.
226 Passes.add(createInstructionCombiningPass());
228 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
230 // Run a few AA driven optimizations here and now, to cleanup the code.
231 Passes.add(createGlobalsModRefPass()); // IP alias analysis
233 Passes.add(createLICMPass()); // Hoist loop invariants
234 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
235 Passes.add(createGCSEPass()); // Remove common subexprs
236 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
238 // Cleanup and simplify the code after the scalar optimizations.
239 Passes.add(createInstructionCombiningPass());
241 // Delete basic blocks, which optimization passes may have killed...
242 Passes.add(createCFGSimplificationPass());
244 // Now that we have optimized the program, discard unreachable functions...
245 Passes.add(createGlobalDCEPass());
247 // Make sure everything is still good.
248 Passes.add(createVerifierPass());
250 Target.addPassesToEmitFile(Passes, Out, TargetMachine::AssemblyFile, true);
252 // Run our queue of passes all at once now, efficiently.
255 return LTO_OPT_SUCCESS;
258 ///Link all modules together and optimize them using IPO. Generate
259 /// native object file using OutputFilename
260 /// Return appropriate LTOStatus.
262 LinkTimeOptimizer::optimizeModules(const std::string &OutputFilename,
263 std::vector<const char *> &exportList)
268 std::ios::openmode io_mode =
269 std::ios::out | std::ios::trunc | std::ios::binary;
270 std::string *errMsg = NULL;
271 Module *bigOne = modules[0];
272 Linker theLinker("LinkTimeOptimizer", bigOne, false);
273 for (unsigned i = 1, e = modules.size(); i != e; ++i)
274 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
275 return LTO_MODULE_MERGE_FAILURE;
278 // Enable this when -save-temps is used
279 std::ofstream Out("big.bc", io_mode);
280 WriteBytecodeToFile(bigOne, Out, true);
283 // Strip leading underscore because it was added to match names
285 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
286 const char *name = exportList[i];
287 if (strlen(name) > 2 && name[0] == '_')
288 exportList[i] = &name[1];
291 sys::Path tmpAsmFilePath("/tmp/");
292 tmpAsmFilePath.createTemporaryFileOnDisk();
293 sys::RemoveFileOnSignal(tmpAsmFilePath);
295 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
296 if (!asmFile.is_open() || asmFile.bad()) {
297 if (tmpAsmFilePath.exists())
298 tmpAsmFilePath.eraseFromDisk();
299 return LTO_WRITE_FAILURE;
302 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
304 if (status != LTO_OPT_SUCCESS) {
305 tmpAsmFilePath.eraseFromDisk();
309 // Run GCC to assemble and link the program into native code.
312 // We can't just assemble and link the file with the system assembler
313 // and linker because we don't know where to put the _start symbol.
314 // GCC mysteriously knows how to do it.
315 const sys::Path gcc = FindExecutable("gcc", "/");
317 tmpAsmFilePath.eraseFromDisk();
318 return LTO_ASM_FAILURE;
321 std::vector<const char*> args;
322 args.push_back(gcc.c_str());
323 args.push_back("-c");
324 args.push_back("-x");
325 args.push_back("assembler");
326 args.push_back("-o");
327 args.push_back(OutputFilename.c_str());
328 args.push_back(tmpAsmFilePath.c_str());
331 int R1 = sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1);
333 tmpAsmFilePath.eraseFromDisk();
335 return LTO_OPT_SUCCESS;