1 //===- gccld.cpp - LLVM 'ld' compatible linker ----------------------------===//
3 // This utility is intended to be compatible with GCC, and follows standard
4 // system 'ld' conventions. As such, the default output file is ./a.out.
5 // Additionally, this program outputs a shell script that is used to invoke LLI
6 // to execute the program. In this manner, the generated executable (a.out for
7 // example), is directly executable, whereas the bytecode file actually lives in
8 // the a.out.bc file generated by this program. Also, Force is on by default.
10 // Note that if someone (or a script) deletes the executable program generated,
11 // the .bc file will be left around. Considering that this is a temporary hack,
12 // I'm not to worried about this.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/Utils/Linker.h"
17 #include "llvm/Module.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/Bytecode/Reader.h"
20 #include "llvm/Bytecode/WriteBytecodePass.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/Transforms/IPO.h"
23 #include "llvm/Transforms/Scalar.h"
24 #include "Support/CommandLine.h"
25 #include "Support/Signals.h"
30 #include <sys/types.h> // For FileExists
35 InputFilenames(cl::Positional, cl::desc("<input bytecode files>"),
39 OutputFilename("o", cl::desc("Override output filename"), cl::init("a.out"),
40 cl::value_desc("filename"));
43 Verbose("v", cl::desc("Print information about actions taken"));
46 LibPaths("L", cl::desc("Specify a library search path"), cl::Prefix,
47 cl::value_desc("directory"));
50 Libraries("l", cl::desc("Specify libraries to link to"), cl::Prefix,
51 cl::value_desc("library prefix"));
54 Strip("s", cl::desc("Strip symbol info from executable"));
57 NoInternalize("disable-internalize",
58 cl::desc("Do not mark all symbols as internal"));
61 LinkAsLibrary("link-as-library", cl::desc("Link the .bc files together as a"
62 " library, not an executable"));
64 // Compatibility options that are ignored, but support by LD
66 CO3("soname", cl::Hidden, cl::desc("Compatibility option: ignored"));
68 CO4("version-script", cl::Hidden, cl::desc("Compatibility option: ignored"));
70 CO5("eh-frame-hdr", cl::Hidden, cl::desc("Compatibility option: ignored"));
73 // FileExists - Return true if the specified string is an openable file...
74 static inline bool FileExists(const std::string &FN) {
76 return stat(FN.c_str(), &StatBuf) != -1;
80 // LoadObject - Read the specified "object file", which should not search the
81 // library path to find it.
82 static inline std::auto_ptr<Module> LoadObject(std::string FN,
83 std::string &OutErrorMessage) {
84 if (Verbose) std::cerr << "Loading '" << FN << "'\n";
85 if (!FileExists(FN)) {
86 // Attempt to load from the LLVM_LIB_SEARCH_PATH directory... if we would
87 // otherwise fail. This is used to locate objects like crtend.o.
89 char *SearchPath = getenv("LLVM_LIB_SEARCH_PATH");
90 if (SearchPath && FileExists(std::string(SearchPath)+"/"+FN))
91 FN = std::string(SearchPath)+"/"+FN;
93 OutErrorMessage = "could not find input file '" + FN + "'!";
94 return std::auto_ptr<Module>();
98 std::string ErrorMessage;
99 Module *Result = ParseBytecodeFile(FN, &ErrorMessage);
100 if (Result) return std::auto_ptr<Module>(Result);
102 OutErrorMessage = "Bytecode file '" + FN + "' corrupt!";
103 if (ErrorMessage.size()) OutErrorMessage += ": " + ErrorMessage;
104 return std::auto_ptr<Module>();
108 static Module *LoadSingleLibraryObject(const std::string &Filename) {
109 std::string ErrorMessage;
110 std::auto_ptr<Module> M = LoadObject(Filename, ErrorMessage);
111 if (M.get() == 0 && Verbose) {
112 std::cerr << "Error loading '" + Filename + "'";
113 if (!ErrorMessage.empty()) std::cerr << ": " << ErrorMessage;
121 // LoadLibraryFromDirectory - This looks for a .a, .so, or .bc file in a
122 // particular directory. It returns true if no library is found, otherwise it
123 // puts the loaded modules into the Objects list, and sets isArchive to true if
124 // a .a file was loaded.
126 static inline bool LoadLibraryFromDirectory(const std::string &LibName,
127 const std::string &Directory,
128 std::vector<Module*> &Objects,
130 if (FileExists(Directory + "lib" + LibName + ".a")) {
131 std::string ErrorMessage;
132 if (Verbose) std::cerr << " Loading '" << Directory << "lib"
133 << LibName << ".a'\n";
134 if (!ReadArchiveFile(Directory + "lib" + LibName + ".a", Objects,
135 &ErrorMessage)) { // Read the archive file
137 return false; // Success!
141 std::cerr << " Error loading archive '" + Directory +"lib"+LibName+".a'";
142 if (!ErrorMessage.empty()) std::cerr << ": " << ErrorMessage;
147 if (FileExists(Directory + "lib" + LibName + ".so"))
148 if (Module *M = LoadSingleLibraryObject(Directory + "lib" + LibName+".so")){
150 Objects.push_back(M);
154 if (FileExists(Directory + "lib" + LibName + ".bc"))
155 if (Module *M = LoadSingleLibraryObject(Directory + "lib" + LibName+".bc")){
157 Objects.push_back(M);
163 // LoadLibrary - This searches for a .a, .so, or .bc file which provides the
164 // LLVM bytecode for the library. It returns true if no library is found,
165 // otherwise it puts the loaded modules into the Objects list, and sets
166 // isArchive to true if a .a file was loaded.
168 static inline bool LoadLibrary(const std::string &LibName,
169 std::vector<Module*> &Objects, bool &isArchive,
170 std::string &ErrorMessage) {
171 std::string Directory;
172 unsigned NextLibPathIdx = 0;
175 // Try loading from the current directory...
176 if (Verbose) std::cerr << " Looking in directory '" << Directory << "'\n";
177 if (!LoadLibraryFromDirectory(LibName, Directory, Objects, isArchive))
180 if (NextLibPathIdx == LibPaths.size()) break;
181 Directory = LibPaths[NextLibPathIdx++]+"/";
184 ErrorMessage = "error linking library '-l" + LibName+ "': library not found!";
188 static void GetAllDefinedSymbols(Module *M,
189 std::set<std::string> &DefinedSymbols) {
190 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
191 if (I->hasName() && !I->isExternal() && !I->hasInternalLinkage())
192 DefinedSymbols.insert(I->getName());
193 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
194 if (I->hasName() && !I->isExternal() && !I->hasInternalLinkage())
195 DefinedSymbols.insert(I->getName());
198 // GetAllUndefinedSymbols - This calculates the set of undefined symbols that
199 // still exist in an LLVM module. This is a bit tricky because there may be two
200 // symbols with the same name, but different LLVM types that will be resolved to
201 // each other, but aren't currently (thus we need to treat it as resolved).
203 static void GetAllUndefinedSymbols(Module *M,
204 std::set<std::string> &UndefinedSymbols) {
205 std::set<std::string> DefinedSymbols;
206 UndefinedSymbols.clear(); // Start out empty
208 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
211 UndefinedSymbols.insert(I->getName());
212 else if (!I->hasInternalLinkage())
213 DefinedSymbols.insert(I->getName());
215 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
218 UndefinedSymbols.insert(I->getName());
219 else if (!I->hasInternalLinkage())
220 DefinedSymbols.insert(I->getName());
223 // Prune out any defined symbols from the undefined symbols set...
224 for (std::set<std::string>::iterator I = UndefinedSymbols.begin();
225 I != UndefinedSymbols.end(); )
226 if (DefinedSymbols.count(*I))
227 UndefinedSymbols.erase(I++); // This symbol really is defined!
229 ++I; // Keep this symbol in the undefined symbols list
233 static bool LinkLibrary(Module *M, const std::string &LibName,
234 std::string &ErrorMessage) {
235 std::set<std::string> UndefinedSymbols;
236 GetAllUndefinedSymbols(M, UndefinedSymbols);
237 if (UndefinedSymbols.empty()) {
238 if (Verbose) std::cerr << " No symbols undefined, don't link library!\n";
239 return false; // No need to link anything in!
242 std::vector<Module*> Objects;
244 if (LoadLibrary(LibName, Objects, isArchive, ErrorMessage)) return true;
246 // Figure out which symbols are defined by all of the modules in the .a file
247 std::vector<std::set<std::string> > DefinedSymbols;
248 DefinedSymbols.resize(Objects.size());
249 for (unsigned i = 0; i != Objects.size(); ++i)
250 GetAllDefinedSymbols(Objects[i], DefinedSymbols[i]);
253 while (Linked) { // While we are linking in object files, loop.
256 for (unsigned i = 0; i != Objects.size(); ++i) {
257 // Consider whether we need to link in this module... we only need to
258 // link it in if it defines some symbol which is so far undefined.
260 const std::set<std::string> &DefSymbols = DefinedSymbols[i];
262 bool ObjectRequired = false;
263 for (std::set<std::string>::iterator I = UndefinedSymbols.begin(),
264 E = UndefinedSymbols.end(); I != E; ++I)
265 if (DefSymbols.count(*I)) {
267 std::cerr << " Found object providing symbol '" << *I << "'...\n";
268 ObjectRequired = true;
272 // We DO need to link this object into the program...
273 if (ObjectRequired) {
274 if (LinkModules(M, Objects[i], &ErrorMessage))
275 return true; // Couldn't link in the right object file...
277 // Since we have linked in this object, delete it from the list of
278 // objects to consider in this archive file.
279 std::swap(Objects[i], Objects.back());
280 std::swap(DefinedSymbols[i], DefinedSymbols.back());
282 DefinedSymbols.pop_back();
283 --i; // Do not skip an entry
285 // The undefined symbols set should have shrunk.
286 GetAllUndefinedSymbols(M, UndefinedSymbols);
287 Linked = true; // We have linked something in!
295 static int PrintAndReturn(const char *progname, const std::string &Message,
296 const std::string &Extra = "") {
297 std::cerr << progname << Extra << ": " << Message << "\n";
302 int main(int argc, char **argv) {
303 cl::ParseCommandLineOptions(argc, argv, " llvm linker for GCC\n");
305 std::string ErrorMessage;
306 std::auto_ptr<Module> Composite(LoadObject(InputFilenames[0], ErrorMessage));
307 if (Composite.get() == 0)
308 return PrintAndReturn(argv[0], ErrorMessage);
310 // If the user specied an extra search path in their environment, respect it.
311 if (char *SearchPath = getenv("LLVM_LIB_SEARCH_PATH"))
312 LibPaths.push_back(SearchPath);
314 for (unsigned i = 1; i < InputFilenames.size(); ++i) {
315 std::auto_ptr<Module> M(LoadObject(InputFilenames[i], ErrorMessage));
317 return PrintAndReturn(argv[0], ErrorMessage);
319 if (Verbose) std::cerr << "Linking in '" << InputFilenames[i] << "'\n";
321 if (LinkModules(Composite.get(), M.get(), &ErrorMessage))
322 return PrintAndReturn(argv[0], ErrorMessage,
323 ": error linking in '" + InputFilenames[i] + "'");
326 // Remove any consecutive duplicates of the same library...
327 Libraries.erase(std::unique(Libraries.begin(), Libraries.end()),
330 // Link in all of the libraries next...
331 for (unsigned i = 0; i != Libraries.size(); ++i) {
332 if (Verbose) std::cerr << "Linking in library: -l" << Libraries[i] << "\n";
333 if (LinkLibrary(Composite.get(), Libraries[i], ErrorMessage))
334 return PrintAndReturn(argv[0], ErrorMessage);
337 // In addition to just linking the input from GCC, we also want to spiff it up
338 // a little bit. Do this now.
342 // Add an appropriate TargetData instance for this module...
343 Passes.add(new TargetData("gccas", Composite.get()));
345 // Linking modules together can lead to duplicated global constants, only keep
346 // one copy of each constant...
348 Passes.add(createConstantMergePass());
350 // If the -s command line option was specified, strip the symbols out of the
351 // resulting program to make it smaller. -s is a GCC option that we are
355 Passes.add(createSymbolStrippingPass());
357 // Often if the programmer does not specify proper prototypes for the
358 // functions they are calling, they end up calling a vararg version of the
359 // function that does not get a body filled in (the real function has typed
360 // arguments). This pass merges the two functions.
362 Passes.add(createFunctionResolvingPass());
364 if (!NoInternalize) {
365 // Now that composite has been compiled, scan through the module, looking
366 // for a main function. If main is defined, mark all other functions
369 Passes.add(createInternalizePass());
372 // Now that we have optimized the program, discard unreachable functions...
374 Passes.add(createGlobalDCEPass());
376 // Add the pass that writes bytecode to the output file...
377 std::string RealBytecodeOutput = OutputFilename;
378 if (!LinkAsLibrary) RealBytecodeOutput += ".bc";
379 std::ofstream Out(RealBytecodeOutput.c_str());
381 return PrintAndReturn(argv[0], "error opening '" + RealBytecodeOutput +
383 Passes.add(new WriteBytecodePass(&Out)); // Write bytecode to file...
385 // Make sure that the Out file gets unlink'd from the disk if we get a SIGINT
386 RemoveFileOnSignal(RealBytecodeOutput);
388 // Run our queue of passes all at once now, efficiently.
389 Passes.run(*Composite.get());
392 if (!LinkAsLibrary) {
393 // Output the script to start the program...
394 std::ofstream Out2(OutputFilename.c_str());
396 return PrintAndReturn(argv[0], "error opening '" + OutputFilename +
398 Out2 << "#!/bin/sh\nlli -q -abort-on-exception $0.bc $*\n";
401 // Make the script executable...
402 chmod(OutputFilename.c_str(), 0755);