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/Verifier.h"
21 #include "llvm/Bytecode/WriteBytecodePass.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/Transforms/IPO.h"
24 #include "llvm/Transforms/Scalar.h"
25 #include "llvm/Transforms/Utils/Linker.h"
26 #include "Support/SystemUtils.h"
27 #include "Support/CommandLine.h"
32 DisableInline("disable-inlining", cl::desc("Do not run the inliner pass"));
35 Verify("verify", cl::desc("Verify intermediate results of all passes"));
38 DisableOptimizations("disable-opt",
39 cl::desc("Do not run any optimization passes"));
42 /// CopyEnv - This function takes an array of environment variables and makes a
43 /// copy of it. This copy can then be manipulated any way the caller likes
44 /// without affecting the process's real environment.
47 /// envp - An array of C strings containing an environment.
50 /// NULL - An error occurred.
52 /// Otherwise, a pointer to a new array of C strings is returned. Every string
53 /// in the array is a duplicate of the one in the original array (i.e. we do
54 /// not copy the char *'s from one array to another).
56 static char ** CopyEnv(char ** const envp) {
57 // Count the number of entries in the old list;
58 unsigned entries; // The number of entries in the old environment list
59 for (entries = 0; envp[entries] != NULL; entries++)
62 // Add one more entry for the NULL pointer that ends the list.
65 // If there are no entries at all, just return NULL.
69 // Allocate a new environment list.
70 char **newenv = new char* [entries];
71 if ((newenv = new char* [entries]) == NULL)
74 // Make a copy of the list. Don't forget the NULL that ends the list.
76 while (envp[entries] != NULL) {
77 newenv[entries] = new char[strlen (envp[entries]) + 1];
78 strcpy (newenv[entries], envp[entries]);
81 newenv[entries] = NULL;
87 /// RemoveEnv - Remove the specified environment variable from the environment
91 /// name - The name of the variable to remove. It cannot be NULL.
92 /// envp - The array of environment variables. It cannot be NULL.
95 /// This is mainly done because functions to remove items from the environment
96 /// are not available across all platforms. In particular, Solaris does not
97 /// seem to have an unsetenv() function or a setenv() function (or they are
98 /// undocumented if they do exist).
100 static void RemoveEnv(const char * name, char ** const envp) {
101 for (unsigned index=0; envp[index] != NULL; index++) {
102 // Find the first equals sign in the array and make it an EOS character.
103 char *p = strchr (envp[index], '=');
109 // Compare the two strings. If they are equal, zap this string.
110 // Otherwise, restore it.
111 if (!strcmp(name, envp[index]))
120 static inline void addPass(PassManager &PM, Pass *P) {
121 // Add the pass to the pass manager...
124 // If we are verifying all of the intermediate steps, add the verifier...
125 if (Verify) PM.add(createVerifierPass());
128 /// GenerateBytecode - generates a bytecode file from the specified module.
131 /// M - The module for which bytecode should be generated.
132 /// Strip - Flags whether symbols should be stripped from the output.
133 /// Internalize - Flags whether all symbols should be marked internal.
134 /// Out - Pointer to file stream to which to write the output.
136 /// Returns non-zero value on error.
138 int llvm::GenerateBytecode(Module *M, bool Strip, bool Internalize,
140 // In addition to just linking the input from GCC, we also want to spiff it up
141 // a little bit. Do this now.
144 if (Verify) Passes.add(createVerifierPass());
146 // Add an appropriate TargetData instance for this module...
147 addPass(Passes, new TargetData("gccld", M));
149 // Often if the programmer does not specify proper prototypes for the
150 // functions they are calling, they end up calling a vararg version of the
151 // function that does not get a body filled in (the real function has typed
152 // arguments). This pass merges the two functions.
153 addPass(Passes, createFunctionResolvingPass());
155 if (!DisableOptimizations) {
157 // Now that composite has been compiled, scan through the module, looking
158 // for a main function. If main is defined, mark all other functions
160 addPass(Passes, createInternalizePass());
163 // Now that we internalized some globals, see if we can mark any globals as
165 addPass(Passes, createGlobalConstifierPass());
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, createGlobalDCEPass()); // Remove dead functions
189 // If we didn't decide to inline a function, check to see if we can
190 // transform it to pass arguments by value instead of by reference.
191 addPass(Passes, createArgumentPromotionPass());
193 // The IPO passes may leave cruft around. Clean up after them.
194 addPass(Passes, createInstructionCombiningPass());
196 addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
198 // Run a few AA driven optimizations here and now, to cleanup the code.
199 // Eventually we should put an IP AA in place here.
201 addPass(Passes, createLICMPass()); // Hoist loop invariants
202 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
203 addPass(Passes, createGCSEPass()); // Remove common subexprs
205 // Cleanup and simplify the code after the scalar optimizations.
206 addPass(Passes, createInstructionCombiningPass());
208 // Delete basic blocks, which optimization passes may have killed...
209 addPass(Passes, createCFGSimplificationPass());
211 // Now that we have optimized the program, discard unreachable functions...
212 addPass(Passes, createGlobalDCEPass());
215 // Make sure everything is still good.
216 Passes.add(createVerifierPass());
218 // Add the pass that writes bytecode to the output file...
219 addPass(Passes, new WriteBytecodePass(Out));
221 // Run our queue of passes all at once now, efficiently.
227 /// GenerateAssembly - generates a native assembly language source file from the
228 /// specified bytecode file.
231 /// InputFilename - The name of the output bytecode file.
232 /// OutputFilename - The name of the file to generate.
233 /// llc - The pathname to use for LLC.
234 /// envp - The environment to use when running LLC.
236 /// Return non-zero value on error.
238 int llvm::GenerateAssembly(const std::string &OutputFilename,
239 const std::string &InputFilename,
240 const std::string &llc,
241 char ** const envp) {
242 // Run LLC to convert the bytecode file into assembly code.
244 cmd[0] = llc.c_str();
247 cmd[3] = OutputFilename.c_str();
248 cmd[4] = InputFilename.c_str();
251 return ExecWait(cmd, envp);
254 /// GenerateAssembly - generates a native assembly language source file from the
255 /// specified bytecode file.
256 int llvm::GenerateCFile(const std::string &OutputFile,
257 const std::string &InputFile,
258 const std::string &llc, char ** const envp) {
259 // Run LLC to convert the bytecode file into C.
262 cmd[0] = llc.c_str();
266 cmd[4] = OutputFile.c_str();
267 cmd[5] = InputFile.c_str();
269 return ExecWait(cmd, envp);
272 /// GenerateNative - generates a native assembly language source file from the
273 /// specified assembly source file.
276 /// InputFilename - The name of the output bytecode file.
277 /// OutputFilename - The name of the file to generate.
278 /// Libraries - The list of libraries with which to link.
279 /// LibPaths - The list of directories in which to find libraries.
280 /// gcc - The pathname to use for GGC.
281 /// envp - A copy of the process's current environment.
286 /// Returns non-zero value on error.
288 int llvm::GenerateNative(const std::string &OutputFilename,
289 const std::string &InputFilename,
290 const std::vector<std::string> &Libraries,
291 const std::vector<std::string> &LibPaths,
292 const std::string &gcc, char ** const envp) {
293 // Remove these environment variables from the environment of the
294 // programs that we will execute. It appears that GCC sets these
295 // environment variables so that the programs it uses can configure
296 // themselves identically.
298 // However, when we invoke GCC below, we want it to use its normal
299 // configuration. Hence, we must sanitize its environment.
300 char ** clean_env = CopyEnv(envp);
301 if (clean_env == NULL)
303 RemoveEnv("LIBRARY_PATH", clean_env);
304 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
305 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
306 RemoveEnv("COMPILER_PATH", clean_env);
307 RemoveEnv("COLLECT_GCC", clean_env);
309 std::vector<const char *> cmd;
311 // Run GCC to assemble and link the program into native code.
314 // We can't just assemble and link the file with the system assembler
315 // and linker because we don't know where to put the _start symbol.
316 // GCC mysteriously knows how to do it.
317 cmd.push_back(gcc.c_str());
318 cmd.push_back("-fno-strict-aliasing");
319 cmd.push_back("-O3");
321 cmd.push_back(OutputFilename.c_str());
322 cmd.push_back(InputFilename.c_str());
324 // Adding the library paths creates a problem for native generation. If we
325 // include the search paths from llvmgcc, then we'll be telling normal gcc
326 // to look inside of llvmgcc's library directories for libraries. This is
327 // bad because those libraries hold only bytecode files (not native object
328 // files). In the end, we attempt to link the bytecode libgcc into a native
331 // Add in the library path options.
332 for (unsigned index=0; index < LibPaths.size(); index++) {
334 cmd.push_back(LibPaths[index].c_str());
338 // Add in the libraries to link.
339 std::vector<std::string> Libs(Libraries);
340 for (unsigned index = 0; index < Libs.size(); index++) {
341 if (Libs[index] != "crtend") {
342 Libs[index] = "-l" + Libs[index];
343 cmd.push_back(Libs[index].c_str());
348 // Run the compiler to assembly and link together the program.
349 return ExecWait(&(cmd[0]), clean_env);