1 //===- GenerateCode.cpp - Functions for generating executable files ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains functions for generating executable files once linking
11 // has finished. This includes generating a shell script to run the JIT or
12 // a native executable derived from the bytecode.
14 //===----------------------------------------------------------------------===//
17 #include "llvm/System/Program.h"
18 #include "llvm/Module.h"
19 #include "llvm/PassManager.h"
20 #include "llvm/Analysis/LoadValueNumbering.h"
21 #include "llvm/Analysis/Passes.h"
22 #include "llvm/Analysis/Verifier.h"
23 #include "llvm/Bytecode/Archive.h"
24 #include "llvm/Bytecode/WriteBytecodePass.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Transforms/IPO.h"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/Support/SystemUtils.h"
29 #include "llvm/Support/CommandLine.h"
35 DisableInline("disable-inlining", cl::desc("Do not run the inliner pass"));
38 Verify("verify", cl::desc("Verify intermediate results of all passes"));
41 DisableOptimizations("disable-opt",
42 cl::desc("Do not run any optimization passes"));
45 /// CopyEnv - This function takes an array of environment variables and makes a
46 /// copy of it. This copy can then be manipulated any way the caller likes
47 /// without affecting the process's real environment.
50 /// envp - An array of C strings containing an environment.
53 /// NULL - An error occurred.
55 /// Otherwise, a pointer to a new array of C strings is returned. Every string
56 /// in the array is a duplicate of the one in the original array (i.e. we do
57 /// not copy the char *'s from one array to another).
59 static char ** CopyEnv(char ** const envp) {
60 // Count the number of entries in the old list;
61 unsigned entries; // The number of entries in the old environment list
62 for (entries = 0; envp[entries] != NULL; entries++)
65 // Add one more entry for the NULL pointer that ends the list.
68 // If there are no entries at all, just return NULL.
72 // Allocate a new environment list.
73 char **newenv = new char* [entries];
74 if ((newenv = new char* [entries]) == NULL)
77 // Make a copy of the list. Don't forget the NULL that ends the list.
79 while (envp[entries] != NULL) {
80 newenv[entries] = new char[strlen (envp[entries]) + 1];
81 strcpy (newenv[entries], envp[entries]);
84 newenv[entries] = NULL;
90 /// RemoveEnv - Remove the specified environment variable from the environment
94 /// name - The name of the variable to remove. It cannot be NULL.
95 /// envp - The array of environment variables. It cannot be NULL.
98 /// This is mainly done because functions to remove items from the environment
99 /// are not available across all platforms. In particular, Solaris does not
100 /// seem to have an unsetenv() function or a setenv() function (or they are
101 /// undocumented if they do exist).
103 static void RemoveEnv(const char * name, char ** const envp) {
104 for (unsigned index=0; envp[index] != NULL; index++) {
105 // Find the first equals sign in the array and make it an EOS character.
106 char *p = strchr (envp[index], '=');
112 // Compare the two strings. If they are equal, zap this string.
113 // Otherwise, restore it.
114 if (!strcmp(name, envp[index]))
121 static void dumpArgs(const char **args) {
122 std::cerr << *args++;
124 std::cerr << ' ' << *args++;
125 std::cerr << '\n' << std::flush;
128 static inline void addPass(PassManager &PM, Pass *P) {
129 // Add the pass to the pass manager...
132 // If we are verifying all of the intermediate steps, add the verifier...
133 if (Verify) PM.add(createVerifierPass());
136 static bool isBytecodeLibrary(const sys::Path &FullPath) {
137 // Check for a bytecode file
138 if (FullPath.isBytecodeFile()) return true;
139 // Check for a dynamic library file
140 if (FullPath.isDynamicLibrary()) return false;
141 // Check for a true bytecode archive file
142 if (FullPath.isArchive() ) {
143 std::string ErrorMessage;
144 Archive* ar = Archive::OpenAndLoadSymbols( FullPath, &ErrorMessage );
145 return ar->isBytecodeArchive();
150 static bool isBytecodeLPath(const std::string &LibPath) {
151 sys::Path LPath(LibPath);
153 // Make sure it exists and is a directory
154 sys::FileStatus Status;
155 if (LPath.getFileStatus(Status) || !Status.isDir)
158 // Grab the contents of the -L path
159 std::set<sys::Path> Files;
160 if (LPath.getDirectoryContents(Files, 0))
163 // Iterate over the contents one by one to determine
164 // if this -L path has any bytecode shared libraries
166 std::set<sys::Path>::iterator File = Files.begin();
167 std::string dllsuffix = sys::Path::GetDLLSuffix();
168 for (; File != Files.end(); ++File) {
171 if (File->getFileStatus(Status) || Status.isDir)
174 std::string path = File->toString();
176 // Check for an ending '.dll', '.so' or '.a' suffix as all
177 // other files are not of interest to us here
178 if (path.find(dllsuffix, path.size()-dllsuffix.size()) == std::string::npos
179 && path.find(".a", path.size()-2) == std::string::npos)
182 // Finally, check to see if the file is a true bytecode file
183 if (isBytecodeLibrary(*File))
189 /// GenerateBytecode - generates a bytecode file from the specified module.
192 /// M - The module for which bytecode should be generated.
193 /// StripLevel - 2 if we should strip all symbols, 1 if we should strip
195 /// Internalize - Flags whether all symbols should be marked internal.
196 /// Out - Pointer to file stream to which to write the output.
198 /// Returns non-zero value on error.
200 int llvm::GenerateBytecode(Module *M, int StripLevel, bool Internalize,
202 // In addition to just linking the input from GCC, we also want to spiff it up
203 // a little bit. Do this now.
206 if (Verify) Passes.add(createVerifierPass());
208 // Add an appropriate TargetData instance for this module...
209 addPass(Passes, new TargetData(M));
211 // Often if the programmer does not specify proper prototypes for the
212 // functions they are calling, they end up calling a vararg version of the
213 // function that does not get a body filled in (the real function has typed
214 // arguments). This pass merges the two functions.
215 addPass(Passes, createFunctionResolvingPass());
217 if (!DisableOptimizations) {
218 // Now that composite has been compiled, scan through the module, looking
219 // for a main function. If main is defined, mark all other functions
221 addPass(Passes, createInternalizePass(Internalize));
223 // Propagate constants at call sites into the functions they call. This
224 // opens opportunities for globalopt (and inlining) by substituting function
225 // pointers passed as arguments to direct uses of functions.
226 addPass(Passes, createIPSCCPPass());
228 // Now that we internalized some globals, see if we can hack on them!
229 addPass(Passes, createGlobalOptimizerPass());
231 // Linking modules together can lead to duplicated global constants, only
232 // keep one copy of each constant...
233 addPass(Passes, createConstantMergePass());
235 // Remove unused arguments from functions...
236 addPass(Passes, createDeadArgEliminationPass());
239 addPass(Passes, createFunctionInliningPass()); // Inline small functions
241 addPass(Passes, createPruneEHPass()); // Remove dead EH info
242 addPass(Passes, createGlobalOptimizerPass()); // Optimize globals again.
243 addPass(Passes, createGlobalDCEPass()); // Remove dead functions
245 // If we didn't decide to inline a function, check to see if we can
246 // transform it to pass arguments by value instead of by reference.
247 addPass(Passes, createArgumentPromotionPass());
249 // The IPO passes may leave cruft around. Clean up after them.
250 addPass(Passes, createInstructionCombiningPass());
252 addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
254 // Run a few AA driven optimizations here and now, to cleanup the code.
255 addPass(Passes, createGlobalsModRefPass()); // IP alias analysis
257 addPass(Passes, createLICMPass()); // Hoist loop invariants
258 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
259 addPass(Passes, createGCSEPass()); // Remove common subexprs
260 addPass(Passes, createDeadStoreEliminationPass()); // Nuke dead stores
262 // Cleanup and simplify the code after the scalar optimizations.
263 addPass(Passes, createInstructionCombiningPass());
265 // Delete basic blocks, which optimization passes may have killed...
266 addPass(Passes, createCFGSimplificationPass());
268 // Now that we have optimized the program, discard unreachable functions...
269 addPass(Passes, createGlobalDCEPass());
272 // If the -s or -S command line options were specified, strip the symbols out
273 // of the resulting program to make it smaller. -s and -S are GLD options
274 // that we are supporting.
276 addPass(Passes, createStripSymbolsPass(StripLevel == 1));
278 // Make sure everything is still good.
279 Passes.add(createVerifierPass());
281 // Add the pass that writes bytecode to the output file...
282 addPass(Passes, new WriteBytecodePass(Out));
284 // Run our queue of passes all at once now, efficiently.
290 /// GenerateAssembly - generates a native assembly language source file from the
291 /// specified bytecode file.
294 /// InputFilename - The name of the output bytecode file.
295 /// OutputFilename - The name of the file to generate.
296 /// llc - The pathname to use for LLC.
298 /// Return non-zero value on error.
300 int llvm::GenerateAssembly(const std::string &OutputFilename,
301 const std::string &InputFilename,
302 const sys::Path &llc,
305 // Run LLC to convert the bytecode file into assembly code.
306 std::vector<const char*> args;
307 args.push_back(llc.c_str());
308 args.push_back("-f");
309 args.push_back("-o");
310 args.push_back(OutputFilename.c_str());
311 args.push_back(InputFilename.c_str());
313 if (Verbose) dumpArgs(&args[0]);
314 return sys::Program::ExecuteAndWait(llc, &args[0],0,0,0,&ErrMsg);
317 /// GenerateCFile - generates a C source file from the specified bytecode file.
318 int llvm::GenerateCFile(const std::string &OutputFile,
319 const std::string &InputFile,
320 const sys::Path &llc,
323 // Run LLC to convert the bytecode file into C.
324 std::vector<const char*> args;
325 args.push_back(llc.c_str());
326 args.push_back("-march=c");
327 args.push_back("-f");
328 args.push_back("-o");
329 args.push_back(OutputFile.c_str());
330 args.push_back(InputFile.c_str());
332 if (Verbose) dumpArgs(&args[0]);
333 return sys::Program::ExecuteAndWait(llc, &args[0],0,0,0,&ErrMsg);
336 /// GenerateNative - generates a native executable file from the specified
337 /// assembly source file.
340 /// InputFilename - The name of the output bytecode file.
341 /// OutputFilename - The name of the file to generate.
342 /// Libraries - The list of libraries with which to link.
343 /// gcc - The pathname to use for GGC.
344 /// envp - A copy of the process's current environment.
349 /// Returns non-zero value on error.
351 int llvm::GenerateNative(const std::string &OutputFilename,
352 const std::string &InputFilename,
353 const std::vector<std::string> &LibPaths,
354 const std::vector<std::string> &Libraries,
355 const sys::Path &gcc, char ** const envp,
357 bool ExportAllAsDynamic,
358 const std::vector<std::string> &RPaths,
359 const std::string &SOName,
362 // Remove these environment variables from the environment of the
363 // programs that we will execute. It appears that GCC sets these
364 // environment variables so that the programs it uses can configure
365 // themselves identically.
367 // However, when we invoke GCC below, we want it to use its normal
368 // configuration. Hence, we must sanitize its environment.
369 char ** clean_env = CopyEnv(envp);
370 if (clean_env == NULL)
372 RemoveEnv("LIBRARY_PATH", clean_env);
373 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
374 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
375 RemoveEnv("COMPILER_PATH", clean_env);
376 RemoveEnv("COLLECT_GCC", clean_env);
379 // Run GCC to assemble and link the program into native code.
382 // We can't just assemble and link the file with the system assembler
383 // and linker because we don't know where to put the _start symbol.
384 // GCC mysteriously knows how to do it.
385 std::vector<const char*> args;
386 args.push_back(gcc.c_str());
387 args.push_back("-fno-strict-aliasing");
388 args.push_back("-O3");
389 args.push_back("-o");
390 args.push_back(OutputFilename.c_str());
391 args.push_back(InputFilename.c_str());
393 // StringsToDelete - We don't want to call c_str() on temporary strings.
394 // If we need a temporary string, copy it here so that the memory is not
395 // reclaimed until after the exec call. All of these strings are allocated
397 std::vector<char*> StringsToDelete;
399 if (Shared) args.push_back("-shared");
400 if (ExportAllAsDynamic) args.push_back("-export-dynamic");
401 if (!RPaths.empty()) {
402 for (std::vector<std::string>::const_iterator I = RPaths.begin(),
403 E = RPaths.end(); I != E; I++) {
404 std::string rp = "-Wl,-rpath," + *I;
405 StringsToDelete.push_back(strdup(rp.c_str()));
406 args.push_back(StringsToDelete.back());
409 if (!SOName.empty()) {
410 std::string so = "-Wl,-soname," + SOName;
411 StringsToDelete.push_back(strdup(so.c_str()));
412 args.push_back(StringsToDelete.back());
415 // Add in the libpaths to find the libraries.
418 // When gccld is called from the llvm-gxx frontends, the -L paths for
419 // the LLVM cfrontend install paths are appended. We don't want the
420 // native linker to use these -L paths as they contain bytecode files.
421 // Further, we don't want any -L paths that contain bytecode shared
422 // libraries or true bytecode archive files. We omit them in all such
424 for (unsigned index = 0; index < LibPaths.size(); index++)
425 if (!isBytecodeLPath(LibPaths[index])) {
426 std::string Tmp = "-L"+LibPaths[index];
427 StringsToDelete.push_back(strdup(Tmp.c_str()));
428 args.push_back(StringsToDelete.back());
431 // Add in the libraries to link.
432 for (unsigned index = 0; index < Libraries.size(); index++)
433 // HACK: If this is libg, discard it. This gets added by the compiler
434 // driver when doing: 'llvm-gcc main.c -Wl,-native -o a.out -g'. Note that
435 // this should really be fixed by changing the llvm-gcc compiler driver.
436 if (Libraries[index] != "crtend" && Libraries[index] != "g") {
437 std::string Tmp = "-l"+Libraries[index];
438 StringsToDelete.push_back(strdup(Tmp.c_str()));
439 args.push_back(StringsToDelete.back());
441 args.push_back(0); // Null terminate.
443 // Run the compiler to assembly and link together the program.
444 if (Verbose) dumpArgs(&args[0]);
445 int Res = sys::Program::ExecuteAndWait(
446 gcc, &args[0],(const char**)clean_env,0,0,&ErrMsg);
450 while (!StringsToDelete.empty()) {
451 free(StringsToDelete.back());
452 StringsToDelete.pop_back();