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"
30 #include "llvm/Support/Streams.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 NoCompress("disable-compression", cl::init(true),
46 cl::desc("Don't compress the generated bytecode"));
49 /// CopyEnv - This function takes an array of environment variables and makes a
50 /// copy of it. This copy can then be manipulated any way the caller likes
51 /// without affecting the process's real environment.
54 /// envp - An array of C strings containing an environment.
57 /// NULL - An error occurred.
59 /// Otherwise, a pointer to a new array of C strings is returned. Every string
60 /// in the array is a duplicate of the one in the original array (i.e. we do
61 /// not copy the char *'s from one array to another).
63 static char ** CopyEnv(char ** const envp) {
64 // Count the number of entries in the old list;
65 unsigned entries; // The number of entries in the old environment list
66 for (entries = 0; envp[entries] != NULL; entries++)
69 // Add one more entry for the NULL pointer that ends the list.
72 // If there are no entries at all, just return NULL.
76 // Allocate a new environment list.
77 char **newenv = new char* [entries];
78 if ((newenv = new char* [entries]) == NULL)
81 // Make a copy of the list. Don't forget the NULL that ends the list.
83 while (envp[entries] != NULL) {
84 newenv[entries] = new char[strlen (envp[entries]) + 1];
85 strcpy (newenv[entries], envp[entries]);
88 newenv[entries] = NULL;
94 /// RemoveEnv - Remove the specified environment variable from the environment
98 /// name - The name of the variable to remove. It cannot be NULL.
99 /// envp - The array of environment variables. It cannot be NULL.
102 /// This is mainly done because functions to remove items from the environment
103 /// are not available across all platforms. In particular, Solaris does not
104 /// seem to have an unsetenv() function or a setenv() function (or they are
105 /// undocumented if they do exist).
107 static void RemoveEnv(const char * name, char ** const envp) {
108 for (unsigned index=0; envp[index] != NULL; index++) {
109 // Find the first equals sign in the array and make it an EOS character.
110 char *p = strchr (envp[index], '=');
116 // Compare the two strings. If they are equal, zap this string.
117 // Otherwise, restore it.
118 if (!strcmp(name, envp[index]))
125 static void dumpArgs(const char **args) {
128 cerr << ' ' << *args++;
129 cerr << '\n' << std::flush;
132 static inline void addPass(PassManager &PM, Pass *P) {
133 // Add the pass to the pass manager...
136 // If we are verifying all of the intermediate steps, add the verifier...
137 if (Verify) PM.add(createVerifierPass());
140 static bool isBytecodeLibrary(const sys::Path &FullPath) {
141 // Check for a bytecode file
142 if (FullPath.isBytecodeFile()) return true;
143 // Check for a dynamic library file
144 if (FullPath.isDynamicLibrary()) return false;
145 // Check for a true bytecode archive file
146 if (FullPath.isArchive() ) {
147 std::string ErrorMessage;
148 Archive* ar = Archive::OpenAndLoadSymbols( FullPath, &ErrorMessage );
149 return ar->isBytecodeArchive();
154 static bool isBytecodeLPath(const std::string &LibPath) {
155 sys::Path LPath(LibPath);
157 // Make sure it exists and is a directory
158 sys::FileStatus Status;
159 if (LPath.getFileStatus(Status) || !Status.isDir)
162 // Grab the contents of the -L path
163 std::set<sys::Path> Files;
164 if (LPath.getDirectoryContents(Files, 0))
167 // Iterate over the contents one by one to determine
168 // if this -L path has any bytecode shared libraries
170 std::set<sys::Path>::iterator File = Files.begin();
171 std::string dllsuffix = sys::Path::GetDLLSuffix();
172 for (; File != Files.end(); ++File) {
175 if (File->getFileStatus(Status) || Status.isDir)
178 std::string path = File->toString();
180 // Check for an ending '.dll', '.so' or '.a' suffix as all
181 // other files are not of interest to us here
182 if (path.find(dllsuffix, path.size()-dllsuffix.size()) == std::string::npos
183 && path.find(".a", path.size()-2) == std::string::npos)
186 // Finally, check to see if the file is a true bytecode file
187 if (isBytecodeLibrary(*File))
193 /// GenerateBytecode - generates a bytecode file from the specified module.
196 /// M - The module for which bytecode should be generated.
197 /// StripLevel - 2 if we should strip all symbols, 1 if we should strip
199 /// Internalize - Flags whether all symbols should be marked internal.
200 /// Out - Pointer to file stream to which to write the output.
202 /// Returns non-zero value on error.
204 int llvm::GenerateBytecode(Module *M, int StripLevel, bool Internalize,
206 // In addition to just linking the input from GCC, we also want to spiff it up
207 // a little bit. Do this now.
210 if (Verify) Passes.add(createVerifierPass());
212 // Add an appropriate TargetData instance for this module...
213 addPass(Passes, new TargetData(M));
215 // Often if the programmer does not specify proper prototypes for the
216 // functions they are calling, they end up calling a vararg version of the
217 // function that does not get a body filled in (the real function has typed
218 // arguments). This pass merges the two functions.
219 addPass(Passes, createFunctionResolvingPass());
221 if (!DisableOptimizations) {
222 // Now that composite has been compiled, scan through the module, looking
223 // for a main function. If main is defined, mark all other functions
225 addPass(Passes, createInternalizePass(Internalize));
227 // Propagate constants at call sites into the functions they call. This
228 // opens opportunities for globalopt (and inlining) by substituting function
229 // pointers passed as arguments to direct uses of functions.
230 addPass(Passes, createIPSCCPPass());
232 // Now that we internalized some globals, see if we can hack on them!
233 addPass(Passes, createGlobalOptimizerPass());
235 // Linking modules together can lead to duplicated global constants, only
236 // keep one copy of each constant...
237 addPass(Passes, createConstantMergePass());
239 // Remove unused arguments from functions...
240 addPass(Passes, createDeadArgEliminationPass());
242 // Reduce the code after globalopt and ipsccp. Both can open up significant
243 // simplification opportunities, and both can propagate functions through
244 // function pointers. When this happens, we often have to resolve varargs
245 // calls, etc, so let instcombine do this.
246 addPass(Passes, createInstructionCombiningPass());
249 addPass(Passes, createFunctionInliningPass()); // Inline small functions
251 addPass(Passes, createPruneEHPass()); // Remove dead EH info
252 addPass(Passes, createGlobalOptimizerPass()); // Optimize globals again.
253 addPass(Passes, createGlobalDCEPass()); // Remove dead functions
255 // If we didn't decide to inline a function, check to see if we can
256 // transform it to pass arguments by value instead of by reference.
257 addPass(Passes, createArgumentPromotionPass());
259 // The IPO passes may leave cruft around. Clean up after them.
260 addPass(Passes, createInstructionCombiningPass());
262 addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
264 // Run a few AA driven optimizations here and now, to cleanup the code.
265 addPass(Passes, createGlobalsModRefPass()); // IP alias analysis
267 addPass(Passes, createLICMPass()); // Hoist loop invariants
268 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
269 addPass(Passes, createGCSEPass()); // Remove common subexprs
270 addPass(Passes, createDeadStoreEliminationPass()); // Nuke dead stores
272 // Cleanup and simplify the code after the scalar optimizations.
273 addPass(Passes, createInstructionCombiningPass());
275 // Delete basic blocks, which optimization passes may have killed...
276 addPass(Passes, createCFGSimplificationPass());
278 // Now that we have optimized the program, discard unreachable functions...
279 addPass(Passes, createGlobalDCEPass());
282 // If the -s or -S command line options were specified, strip the symbols out
283 // of the resulting program to make it smaller. -s and -S are GLD options
284 // that we are supporting.
286 addPass(Passes, createStripSymbolsPass(StripLevel == 1));
288 // Make sure everything is still good.
289 Passes.add(createVerifierPass());
291 // Add the pass that writes bytecode to the output file...
293 addPass(Passes, new WriteBytecodePass(&L, false, !NoCompress));
295 // Run our queue of passes all at once now, efficiently.
301 /// GenerateAssembly - generates a native assembly language source file from the
302 /// specified bytecode file.
305 /// InputFilename - The name of the output bytecode file.
306 /// OutputFilename - The name of the file to generate.
307 /// llc - The pathname to use for LLC.
309 /// Return non-zero value on error.
311 int llvm::GenerateAssembly(const std::string &OutputFilename,
312 const std::string &InputFilename,
313 const sys::Path &llc,
316 // Run LLC to convert the bytecode file into assembly code.
317 std::vector<const char*> args;
318 args.push_back(llc.c_str());
319 args.push_back("-f");
320 args.push_back("-o");
321 args.push_back(OutputFilename.c_str());
322 args.push_back(InputFilename.c_str());
324 if (Verbose) dumpArgs(&args[0]);
325 return sys::Program::ExecuteAndWait(llc, &args[0],0,0,0,&ErrMsg);
328 /// GenerateCFile - generates a C source file from the specified bytecode file.
329 int llvm::GenerateCFile(const std::string &OutputFile,
330 const std::string &InputFile,
331 const sys::Path &llc,
334 // Run LLC to convert the bytecode file into C.
335 std::vector<const char*> args;
336 args.push_back(llc.c_str());
337 args.push_back("-march=c");
338 args.push_back("-f");
339 args.push_back("-o");
340 args.push_back(OutputFile.c_str());
341 args.push_back(InputFile.c_str());
343 if (Verbose) dumpArgs(&args[0]);
344 return sys::Program::ExecuteAndWait(llc, &args[0],0,0,0,&ErrMsg);
347 /// GenerateNative - generates a native executable file from the specified
348 /// assembly source file.
351 /// InputFilename - The name of the output bytecode file.
352 /// OutputFilename - The name of the file to generate.
353 /// Libraries - The list of libraries with which to link.
354 /// gcc - The pathname to use for GGC.
355 /// envp - A copy of the process's current environment.
360 /// Returns non-zero value on error.
362 int llvm::GenerateNative(const std::string &OutputFilename,
363 const std::string &InputFilename,
364 const std::vector<std::string> &LibPaths,
365 const std::vector<std::string> &Libraries,
366 const sys::Path &gcc, char ** const envp,
368 bool ExportAllAsDynamic,
369 const std::vector<std::string> &RPaths,
370 const std::string &SOName,
373 // Remove these environment variables from the environment of the
374 // programs that we will execute. It appears that GCC sets these
375 // environment variables so that the programs it uses can configure
376 // themselves identically.
378 // However, when we invoke GCC below, we want it to use its normal
379 // configuration. Hence, we must sanitize its environment.
380 char ** clean_env = CopyEnv(envp);
381 if (clean_env == NULL)
383 RemoveEnv("LIBRARY_PATH", clean_env);
384 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
385 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
386 RemoveEnv("COMPILER_PATH", clean_env);
387 RemoveEnv("COLLECT_GCC", clean_env);
390 // Run GCC to assemble and link the program into native code.
393 // We can't just assemble and link the file with the system assembler
394 // and linker because we don't know where to put the _start symbol.
395 // GCC mysteriously knows how to do it.
396 std::vector<const char*> args;
397 args.push_back(gcc.c_str());
398 args.push_back("-fno-strict-aliasing");
399 args.push_back("-O3");
400 args.push_back("-o");
401 args.push_back(OutputFilename.c_str());
402 args.push_back(InputFilename.c_str());
404 // StringsToDelete - We don't want to call c_str() on temporary strings.
405 // If we need a temporary string, copy it here so that the memory is not
406 // reclaimed until after the exec call. All of these strings are allocated
408 std::vector<char*> StringsToDelete;
410 if (Shared) args.push_back("-shared");
411 if (ExportAllAsDynamic) args.push_back("-export-dynamic");
412 if (!RPaths.empty()) {
413 for (std::vector<std::string>::const_iterator I = RPaths.begin(),
414 E = RPaths.end(); I != E; I++) {
415 std::string rp = "-Wl,-rpath," + *I;
416 StringsToDelete.push_back(strdup(rp.c_str()));
417 args.push_back(StringsToDelete.back());
420 if (!SOName.empty()) {
421 std::string so = "-Wl,-soname," + SOName;
422 StringsToDelete.push_back(strdup(so.c_str()));
423 args.push_back(StringsToDelete.back());
426 // Add in the libpaths to find the libraries.
429 // When gccld is called from the llvm-gxx frontends, the -L paths for
430 // the LLVM cfrontend install paths are appended. We don't want the
431 // native linker to use these -L paths as they contain bytecode files.
432 // Further, we don't want any -L paths that contain bytecode shared
433 // libraries or true bytecode archive files. We omit them in all such
435 for (unsigned index = 0; index < LibPaths.size(); index++)
436 if (!isBytecodeLPath(LibPaths[index])) {
437 std::string Tmp = "-L"+LibPaths[index];
438 StringsToDelete.push_back(strdup(Tmp.c_str()));
439 args.push_back(StringsToDelete.back());
442 // Add in the libraries to link.
443 for (unsigned index = 0; index < Libraries.size(); index++)
444 // HACK: If this is libg, discard it. This gets added by the compiler
445 // driver when doing: 'llvm-gcc main.c -Wl,-native -o a.out -g'. Note that
446 // this should really be fixed by changing the llvm-gcc compiler driver.
447 if (Libraries[index] != "crtend" && Libraries[index] != "g") {
448 std::string Tmp = "-l"+Libraries[index];
449 StringsToDelete.push_back(strdup(Tmp.c_str()));
450 args.push_back(StringsToDelete.back());
452 args.push_back(0); // Null terminate.
454 // Run the compiler to assembly and link together the program.
455 if (Verbose) dumpArgs(&args[0]);
456 int Res = sys::Program::ExecuteAndWait(
457 gcc, &args[0],(const char**)clean_env,0,0,&ErrMsg);
461 while (!StringsToDelete.empty()) {
462 free(StringsToDelete.back());
463 StringsToDelete.pop_back();