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/Module.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/Analysis/LoadValueNumbering.h"
20 #include "llvm/Analysis/Passes.h"
21 #include "llvm/Analysis/Verifier.h"
22 #include "llvm/Bytecode/WriteBytecodePass.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/Transforms/IPO.h"
25 #include "llvm/Transforms/Scalar.h"
26 #include "llvm/Support/Linker.h"
27 #include "llvm/Support/SystemUtils.h"
28 #include "llvm/Support/CommandLine.h"
33 DisableInline("disable-inlining", cl::desc("Do not run the inliner pass"));
36 Verify("verify", cl::desc("Verify intermediate results of all passes"));
39 DisableOptimizations("disable-opt",
40 cl::desc("Do not run any optimization passes"));
43 /// CopyEnv - This function takes an array of environment variables and makes a
44 /// copy of it. This copy can then be manipulated any way the caller likes
45 /// without affecting the process's real environment.
48 /// envp - An array of C strings containing an environment.
51 /// NULL - An error occurred.
53 /// Otherwise, a pointer to a new array of C strings is returned. Every string
54 /// in the array is a duplicate of the one in the original array (i.e. we do
55 /// not copy the char *'s from one array to another).
57 static char ** CopyEnv(char ** const envp) {
58 // Count the number of entries in the old list;
59 unsigned entries; // The number of entries in the old environment list
60 for (entries = 0; envp[entries] != NULL; entries++)
63 // Add one more entry for the NULL pointer that ends the list.
66 // If there are no entries at all, just return NULL.
70 // Allocate a new environment list.
71 char **newenv = new char* [entries];
72 if ((newenv = new char* [entries]) == NULL)
75 // Make a copy of the list. Don't forget the NULL that ends the list.
77 while (envp[entries] != NULL) {
78 newenv[entries] = new char[strlen (envp[entries]) + 1];
79 strcpy (newenv[entries], envp[entries]);
82 newenv[entries] = NULL;
88 /// RemoveEnv - Remove the specified environment variable from the environment
92 /// name - The name of the variable to remove. It cannot be NULL.
93 /// envp - The array of environment variables. It cannot be NULL.
96 /// This is mainly done because functions to remove items from the environment
97 /// are not available across all platforms. In particular, Solaris does not
98 /// seem to have an unsetenv() function or a setenv() function (or they are
99 /// undocumented if they do exist).
101 static void RemoveEnv(const char * name, char ** const envp) {
102 for (unsigned index=0; envp[index] != NULL; index++) {
103 // Find the first equals sign in the array and make it an EOS character.
104 char *p = strchr (envp[index], '=');
110 // Compare the two strings. If they are equal, zap this string.
111 // Otherwise, restore it.
112 if (!strcmp(name, envp[index]))
121 static inline void addPass(PassManager &PM, Pass *P) {
122 // Add the pass to the pass manager...
125 // If we are verifying all of the intermediate steps, add the verifier...
126 if (Verify) PM.add(createVerifierPass());
129 /// GenerateBytecode - generates a bytecode file from the specified module.
132 /// M - The module for which bytecode should be generated.
133 /// Strip - Flags whether symbols should be stripped from the output.
134 /// Internalize - Flags whether all symbols should be marked internal.
135 /// Out - Pointer to file stream to which to write the output.
137 /// Returns non-zero value on error.
139 int llvm::GenerateBytecode(Module *M, bool Strip, bool Internalize,
141 // In addition to just linking the input from GCC, we also want to spiff it up
142 // a little bit. Do this now.
145 if (Verify) Passes.add(createVerifierPass());
147 // Add an appropriate TargetData instance for this module...
148 addPass(Passes, new TargetData("gccld", M));
150 // Often if the programmer does not specify proper prototypes for the
151 // functions they are calling, they end up calling a vararg version of the
152 // function that does not get a body filled in (the real function has typed
153 // arguments). This pass merges the two functions.
154 addPass(Passes, createFunctionResolvingPass());
156 if (!DisableOptimizations) {
158 // Now that composite has been compiled, scan through the module, looking
159 // for a main function. If main is defined, mark all other functions
161 addPass(Passes, createInternalizePass());
164 // Now that we internalized some globals, see if we can hack on them!
165 addPass(Passes, createGlobalOptimizerPass());
167 // Linking modules together can lead to duplicated global constants, only
168 // keep one copy of each constant...
169 addPass(Passes, createConstantMergePass());
171 // If the -s command line option was specified, strip the symbols out of the
172 // resulting program to make it smaller. -s is a GCC option that we are
175 addPass(Passes, createSymbolStrippingPass());
177 // Propagate constants at call sites into the functions they call.
178 addPass(Passes, createIPConstantPropagationPass());
180 // Remove unused arguments from functions...
181 addPass(Passes, createDeadArgEliminationPass());
184 addPass(Passes, createFunctionInliningPass()); // Inline small functions
186 addPass(Passes, createPruneEHPass()); // Remove dead EH info
187 addPass(Passes, createGlobalOptimizerPass()); // Optimize globals again.
188 addPass(Passes, createGlobalDCEPass()); // Remove dead functions
190 // If we didn't decide to inline a function, check to see if we can
191 // transform it to pass arguments by value instead of by reference.
192 addPass(Passes, createArgumentPromotionPass());
194 // The IPO passes may leave cruft around. Clean up after them.
195 addPass(Passes, createInstructionCombiningPass());
197 addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
199 // Run a few AA driven optimizations here and now, to cleanup the code.
200 addPass(Passes, createGlobalsModRefPass()); // IP alias analysis
202 addPass(Passes, createLICMPass()); // Hoist loop invariants
203 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
204 addPass(Passes, createGCSEPass()); // Remove common subexprs
205 addPass(Passes, createDeadStoreEliminationPass()); // Nuke dead stores
207 // Cleanup and simplify the code after the scalar optimizations.
208 addPass(Passes, createInstructionCombiningPass());
210 // Delete basic blocks, which optimization passes may have killed...
211 addPass(Passes, createCFGSimplificationPass());
213 // Now that we have optimized the program, discard unreachable functions...
214 addPass(Passes, createGlobalDCEPass());
217 // Make sure everything is still good.
218 Passes.add(createVerifierPass());
220 // Add the pass that writes bytecode to the output file...
221 addPass(Passes, new WriteBytecodePass(Out));
223 // Run our queue of passes all at once now, efficiently.
229 /// GenerateAssembly - generates a native assembly language source file from the
230 /// specified bytecode file.
233 /// InputFilename - The name of the output bytecode file.
234 /// OutputFilename - The name of the file to generate.
235 /// llc - The pathname to use for LLC.
236 /// envp - The environment to use when running LLC.
238 /// Return non-zero value on error.
240 int llvm::GenerateAssembly(const std::string &OutputFilename,
241 const std::string &InputFilename,
242 const std::string &llc,
243 char ** const envp) {
244 // Run LLC to convert the bytecode file into assembly code.
246 cmd[0] = llc.c_str();
249 cmd[3] = OutputFilename.c_str();
250 cmd[4] = InputFilename.c_str();
253 return ExecWait(cmd, envp);
256 /// GenerateAssembly - generates a native assembly language source file from the
257 /// specified bytecode file.
258 int llvm::GenerateCFile(const std::string &OutputFile,
259 const std::string &InputFile,
260 const std::string &llc, char ** const envp) {
261 // Run LLC to convert the bytecode file into C.
264 cmd[0] = llc.c_str();
268 cmd[4] = OutputFile.c_str();
269 cmd[5] = InputFile.c_str();
271 return ExecWait(cmd, envp);
274 /// GenerateNative - generates a native assembly language source file from the
275 /// specified assembly source file.
278 /// InputFilename - The name of the output bytecode file.
279 /// OutputFilename - The name of the file to generate.
280 /// Libraries - The list of libraries with which to link.
281 /// LibPaths - The list of directories in which to find libraries.
282 /// gcc - The pathname to use for GGC.
283 /// envp - A copy of the process's current environment.
288 /// Returns non-zero value on error.
290 int llvm::GenerateNative(const std::string &OutputFilename,
291 const std::string &InputFilename,
292 const std::vector<std::string> &Libraries,
293 const std::vector<std::string> &LibPaths,
294 const std::string &gcc, char ** const envp) {
295 // Remove these environment variables from the environment of the
296 // programs that we will execute. It appears that GCC sets these
297 // environment variables so that the programs it uses can configure
298 // themselves identically.
300 // However, when we invoke GCC below, we want it to use its normal
301 // configuration. Hence, we must sanitize its environment.
302 char ** clean_env = CopyEnv(envp);
303 if (clean_env == NULL)
305 RemoveEnv("LIBRARY_PATH", clean_env);
306 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
307 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
308 RemoveEnv("COMPILER_PATH", clean_env);
309 RemoveEnv("COLLECT_GCC", clean_env);
311 std::vector<const char *> cmd;
313 // Run GCC to assemble and link the program into native code.
316 // We can't just assemble and link the file with the system assembler
317 // and linker because we don't know where to put the _start symbol.
318 // GCC mysteriously knows how to do it.
319 cmd.push_back(gcc.c_str());
320 cmd.push_back("-fno-strict-aliasing");
321 cmd.push_back("-O3");
323 cmd.push_back(OutputFilename.c_str());
324 cmd.push_back(InputFilename.c_str());
326 // Adding the library paths creates a problem for native generation. If we
327 // include the search paths from llvmgcc, then we'll be telling normal gcc
328 // to look inside of llvmgcc's library directories for libraries. This is
329 // bad because those libraries hold only bytecode files (not native object
330 // files). In the end, we attempt to link the bytecode libgcc into a native
333 // Add in the library path options.
334 for (unsigned index=0; index < LibPaths.size(); index++) {
336 cmd.push_back(LibPaths[index].c_str());
340 // Add in the libraries to link.
341 std::vector<std::string> Libs(Libraries);
342 for (unsigned index = 0; index < Libs.size(); index++) {
343 if (Libs[index] != "crtend") {
344 Libs[index] = "-l" + Libs[index];
345 cmd.push_back(Libs[index].c_str());
350 // Run the compiler to assembly and link together the program.
351 return ExecWait(&(cmd[0]), clean_env);