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. Read it using bytecode reader.
103 /// Collect global functions and symbol names in symbols vector.
104 /// Collect external references in references vector.
105 /// Return LTO_READ_SUCCESS if there is no error.
107 LinkTimeOptimizer::readLLVMObjectFile(const std::string &InputFilename,
108 NameToSymbolMap &symbols,
109 std::set<std::string> &references)
111 Module *m = ParseBytecodeFile(InputFilename);
113 return LTO_READ_FAILURE;
115 // Use mangler to add GlobalPrefix to names to match linker names.
116 // FIXME : Instead of hard coding "-" use GlobalPrefix.
117 Mangler mangler(*m, "_");
119 modules.push_back(m);
121 for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
123 LTOLinkageTypes lt = getLTOLinkageType(f);
125 if (!f->isExternal() && lt != LTOInternalLinkage
126 && strncmp (f->getName().c_str(), "llvm.", 5)) {
127 LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
128 mangler.getValueName(f));
129 symbols[newSymbol->getMangledName()] = newSymbol;
130 allSymbols[newSymbol->getMangledName()] = newSymbol;
133 // Collect external symbols referenced by this function.
134 for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
135 for (BasicBlock::iterator i = b->begin(), be = b->end();
137 for (unsigned count = 0, total = i->getNumOperands();
138 count != total; ++count)
139 findExternalRefs(i->getOperand(count), references, mangler);
142 for (Module::global_iterator v = m->global_begin(), e = m->global_end();
144 LTOLinkageTypes lt = getLTOLinkageType(v);
145 if (!v->isExternal() && lt != LTOInternalLinkage
146 && strncmp (v->getName().c_str(), "llvm.", 5)) {
147 LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
148 mangler.getValueName(v));
149 symbols[newSymbol->getMangledName()] = newSymbol;
151 for (unsigned count = 0, total = v->getNumOperands();
152 count != total; ++count)
153 findExternalRefs(v->getOperand(count), references, mangler);
158 return LTO_READ_SUCCESS;
161 /// Optimize module M using various IPO passes. Use exportList to
162 /// internalize selected symbols. Target platform is selected
163 /// based on information available to module M. No new target
164 /// features are selected.
165 static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
166 std::vector<const char *> &exportList)
168 // Instantiate the pass manager to organize the passes.
171 // Collect Target info
173 const TargetMachineRegistry::Entry* March =
174 TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
177 return LTO_NO_TARGET;
180 std::string Features;
181 std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
183 return LTO_NO_TARGET;
185 TargetMachine &Target = *target.get();
187 // Start off with a verification pass.
188 Passes.add(createVerifierPass());
190 // Add an appropriate TargetData instance for this module...
191 Passes.add(new TargetData(*Target.getTargetData()));
193 // Often if the programmer does not specify proper prototypes for the
194 // functions they are calling, they end up calling a vararg version of the
195 // function that does not get a body filled in (the real function has typed
196 // arguments). This pass merges the two functions.
197 Passes.add(createFunctionResolvingPass());
199 // Internalize symbols if export list is nonemty
200 if (!exportList.empty())
201 Passes.add(createInternalizePass(exportList));
203 // Now that we internalized some globals, see if we can hack on them!
204 Passes.add(createGlobalOptimizerPass());
206 // Linking modules together can lead to duplicated global constants, only
207 // keep one copy of each constant...
208 Passes.add(createConstantMergePass());
210 // If the -s command line option was specified, strip the symbols out of the
211 // resulting program to make it smaller. -s is a GLD option that we are
213 Passes.add(createStripSymbolsPass());
215 // Propagate constants at call sites into the functions they call.
216 Passes.add(createIPConstantPropagationPass());
218 // Remove unused arguments from functions...
219 Passes.add(createDeadArgEliminationPass());
221 Passes.add(createFunctionInliningPass()); // Inline small functions
223 Passes.add(createPruneEHPass()); // Remove dead EH info
225 Passes.add(createGlobalDCEPass()); // Remove dead functions
227 // If we didn't decide to inline a function, check to see if we can
228 // transform it to pass arguments by value instead of by reference.
229 Passes.add(createArgumentPromotionPass());
231 // The IPO passes may leave cruft around. Clean up after them.
232 Passes.add(createInstructionCombiningPass());
234 Passes.add(createScalarReplAggregatesPass()); // Break up allocas
236 // Run a few AA driven optimizations here and now, to cleanup the code.
237 Passes.add(createGlobalsModRefPass()); // IP alias analysis
239 Passes.add(createLICMPass()); // Hoist loop invariants
240 Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
241 Passes.add(createGCSEPass()); // Remove common subexprs
242 Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
244 // Cleanup and simplify the code after the scalar optimizations.
245 Passes.add(createInstructionCombiningPass());
247 // Delete basic blocks, which optimization passes may have killed...
248 Passes.add(createCFGSimplificationPass());
250 // Now that we have optimized the program, discard unreachable functions...
251 Passes.add(createGlobalDCEPass());
253 // Make sure everything is still good.
254 Passes.add(createVerifierPass());
256 Target.addPassesToEmitFile(Passes, Out, TargetMachine::AssemblyFile, true);
258 // Run our queue of passes all at once now, efficiently.
261 return LTO_OPT_SUCCESS;
264 ///Link all modules together and optimize them using IPO. Generate
265 /// native object file using OutputFilename
266 /// Return appropriate LTOStatus.
268 LinkTimeOptimizer::optimizeModules(const std::string &OutputFilename,
269 std::vector<const char *> &exportList)
274 std::ios::openmode io_mode =
275 std::ios::out | std::ios::trunc | std::ios::binary;
276 std::string *errMsg = NULL;
277 Module *bigOne = modules[0];
278 Linker theLinker("LinkTimeOptimizer", bigOne, false);
279 for (unsigned i = 1, e = modules.size(); i != e; ++i)
280 if (theLinker.LinkModules(bigOne, modules[i], errMsg))
281 return LTO_MODULE_MERGE_FAILURE;
284 // Enable this when -save-temps is used
285 std::ofstream Out("big.bc", io_mode);
286 WriteBytecodeToFile(bigOne, Out, true);
289 // Strip leading underscore because it was added to match names
291 for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
292 const char *name = exportList[i];
293 if (strlen(name) > 2 && name[0] == '_')
294 exportList[i] = &name[1];
297 sys::Path tmpAsmFilePath("/tmp/");
298 tmpAsmFilePath.createTemporaryFileOnDisk();
299 sys::RemoveFileOnSignal(tmpAsmFilePath);
301 std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
302 if (!asmFile.is_open() || asmFile.bad()) {
303 if (tmpAsmFilePath.exists())
304 tmpAsmFilePath.eraseFromDisk();
305 return LTO_WRITE_FAILURE;
308 enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
310 if (status != LTO_OPT_SUCCESS) {
311 tmpAsmFilePath.eraseFromDisk();
315 // Run GCC to assemble and link the program into native code.
318 // We can't just assemble and link the file with the system assembler
319 // and linker because we don't know where to put the _start symbol.
320 // GCC mysteriously knows how to do it.
321 const sys::Path gcc = FindExecutable("gcc", "/");
323 tmpAsmFilePath.eraseFromDisk();
324 return LTO_ASM_FAILURE;
327 std::vector<const char*> args;
328 args.push_back(gcc.c_str());
329 args.push_back("-c");
330 args.push_back("-x");
331 args.push_back("assembler");
332 args.push_back("-o");
333 args.push_back(OutputFilename.c_str());
334 args.push_back(tmpAsmFilePath.c_str());
337 int R1 = sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1);
339 tmpAsmFilePath.eraseFromDisk();
341 return LTO_OPT_SUCCESS;