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 static inline void addPass(PassManager &PM, Pass *P) {
43 // Add the pass to the pass manager...
46 // If we are verifying all of the intermediate steps, add the verifier...
47 if (Verify) PM.add(createVerifierPass());
50 /// GenerateBytecode - generates a bytecode file from the specified module.
53 /// M - The module for which bytecode should be generated.
54 /// Strip - Flags whether symbols should be stripped from the output.
55 /// Internalize - Flags whether all symbols should be marked internal.
56 /// Out - Pointer to file stream to which to write the output.
58 /// Returns non-zero value on error.
60 int llvm::GenerateBytecode(Module *M, bool Strip, bool Internalize,
62 // In addition to just linking the input from GCC, we also want to spiff it up
63 // a little bit. Do this now.
66 if (Verify) Passes.add(createVerifierPass());
68 // Add an appropriate TargetData instance for this module...
69 addPass(Passes, new TargetData("gccld", M));
71 // Often if the programmer does not specify proper prototypes for the
72 // functions they are calling, they end up calling a vararg version of the
73 // function that does not get a body filled in (the real function has typed
74 // arguments). This pass merges the two functions.
75 addPass(Passes, createFunctionResolvingPass());
77 if (!DisableOptimizations) {
79 // Now that composite has been compiled, scan through the module, looking
80 // for a main function. If main is defined, mark all other functions
82 addPass(Passes, createInternalizePass());
85 // Now that we internalized some globals, see if we can mark any globals as
87 addPass(Passes, createGlobalConstifierPass());
89 // Linking modules together can lead to duplicated global constants, only
90 // keep one copy of each constant...
91 addPass(Passes, createConstantMergePass());
93 // If the -s command line option was specified, strip the symbols out of the
94 // resulting program to make it smaller. -s is a GCC option that we are
97 addPass(Passes, createSymbolStrippingPass());
99 // Propagate constants at call sites into the functions they call.
100 addPass(Passes, createIPConstantPropagationPass());
102 // Remove unused arguments from functions...
103 addPass(Passes, createDeadArgEliminationPass());
106 addPass(Passes, createFunctionInliningPass()); // Inline small functions
108 addPass(Passes, createPruneEHPass()); // Remove dead EH info
109 addPass(Passes, createGlobalDCEPass()); // Remove dead functions
111 // If we didn't decide to inline a function, check to see if we can
112 // transform it to pass arguments by value instead of by reference.
113 addPass(Passes, createArgumentPromotionPass());
115 // The IPO passes may leave cruft around. Clean up after them.
116 addPass(Passes, createInstructionCombiningPass());
118 addPass(Passes, createScalarReplAggregatesPass()); // Break up allocas
120 // Run a few AA driven optimizations here and now, to cleanup the code.
121 // Eventually we should put an IP AA in place here.
123 addPass(Passes, createLICMPass()); // Hoist loop invariants
124 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
125 addPass(Passes, createGCSEPass()); // Remove common subexprs
127 // Cleanup and simplify the code after the scalar optimizations.
128 addPass(Passes, createInstructionCombiningPass());
130 // Delete basic blocks, which optimization passes may have killed...
131 addPass(Passes, createCFGSimplificationPass());
133 // Now that we have optimized the program, discard unreachable functions...
134 addPass(Passes, createGlobalDCEPass());
137 // Make sure everything is still good.
138 Passes.add(createVerifierPass());
140 // Add the pass that writes bytecode to the output file...
141 addPass(Passes, new WriteBytecodePass(Out));
143 // Run our queue of passes all at once now, efficiently.
149 /// GenerateAssembly - generates a native assembly language source file from the
150 /// specified bytecode file.
153 /// InputFilename - The name of the output bytecode file.
154 /// OutputFilename - The name of the file to generate.
155 /// llc - The pathname to use for LLC.
156 /// envp - The environment to use when running LLC.
158 /// Return non-zero value on error.
160 int llvm::GenerateAssembly(const std::string &OutputFilename,
161 const std::string &InputFilename,
162 const std::string &llc,
163 char ** const envp) {
164 // Run LLC to convert the bytecode file into assembly code.
166 cmd[0] = llc.c_str();
169 cmd[3] = OutputFilename.c_str();
170 cmd[4] = InputFilename.c_str();
173 return ExecWait(cmd, envp);
176 /// GenerateAssembly - generates a native assembly language source file from the
177 /// specified bytecode file.
178 int llvm::GenerateCFile(const std::string &OutputFile,
179 const std::string &InputFile,
180 const std::string &llc, char ** const envp) {
181 // Run LLC to convert the bytecode file into C.
184 cmd[0] = llc.c_str();
188 cmd[4] = OutputFile.c_str();
189 cmd[5] = InputFile.c_str();
191 return ExecWait(cmd, envp);
194 /// GenerateNative - generates a native assembly language source file from the
195 /// specified assembly source file.
198 /// InputFilename - The name of the output bytecode file.
199 /// OutputFilename - The name of the file to generate.
200 /// Libraries - The list of libraries with which to link.
201 /// LibPaths - The list of directories in which to find libraries.
202 /// gcc - The pathname to use for GGC.
203 /// envp - A copy of the process's current environment.
208 /// Returns non-zero value on error.
210 int llvm::GenerateNative(const std::string &OutputFilename,
211 const std::string &InputFilename,
212 const std::vector<std::string> &Libraries,
213 const std::vector<std::string> &LibPaths,
214 const std::string &gcc, char ** const envp) {
215 // Remove these environment variables from the environment of the
216 // programs that we will execute. It appears that GCC sets these
217 // environment variables so that the programs it uses can configure
218 // themselves identically.
220 // However, when we invoke GCC below, we want it to use its normal
221 // configuration. Hence, we must sanitize its environment.
222 char ** clean_env = CopyEnv(envp);
223 if (clean_env == NULL)
225 RemoveEnv("LIBRARY_PATH", clean_env);
226 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
227 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
228 RemoveEnv("COMPILER_PATH", clean_env);
229 RemoveEnv("COLLECT_GCC", clean_env);
231 std::vector<const char *> cmd;
233 // Run GCC to assemble and link the program into native code.
236 // We can't just assemble and link the file with the system assembler
237 // and linker because we don't know where to put the _start symbol.
238 // GCC mysteriously knows how to do it.
239 cmd.push_back(gcc.c_str());
240 cmd.push_back("-fno-strict-aliasing");
241 cmd.push_back("-O3");
243 cmd.push_back(OutputFilename.c_str());
244 cmd.push_back(InputFilename.c_str());
246 // Adding the library paths creates a problem for native generation. If we
247 // include the search paths from llvmgcc, then we'll be telling normal gcc
248 // to look inside of llvmgcc's library directories for libraries. This is
249 // bad because those libraries hold only bytecode files (not native object
250 // files). In the end, we attempt to link the bytecode libgcc into a native
253 // Add in the library path options.
254 for (unsigned index=0; index < LibPaths.size(); index++) {
256 cmd.push_back(LibPaths[index].c_str());
260 // Add in the libraries to link.
261 std::vector<std::string> Libs(Libraries);
262 for (unsigned index = 0; index < Libs.size(); index++) {
263 if (Libs[index] != "crtend") {
264 Libs[index] = "-l" + Libs[index];
265 cmd.push_back(Libs[index].c_str());
270 // Run the compiler to assembly and link together the program.
271 return ExecWait(&(cmd[0]), clean_env);