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"
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"));
45 static inline void addPass(PassManager &PM, Pass *P) {
46 // Add the pass to the pass manager...
49 // If we are verifying all of the intermediate steps, add the verifier...
50 if (Verify) PM.add(createVerifierPass());
53 /// GenerateBytecode - generates a bytecode file from the specified module.
56 /// M - The module for which bytecode should be generated.
57 /// Strip - Flags whether symbols should be stripped from the output.
58 /// Internalize - Flags whether all symbols should be marked internal.
59 /// Out - Pointer to file stream to which to write the output.
64 /// Returns non-zero value on error.
67 GenerateBytecode (Module *M, bool Strip, bool Internalize, std::ostream *Out) {
68 // In addition to just linking the input from GCC, we also want to spiff it up
69 // a little bit. Do this now.
72 if (Verify) Passes.add(createVerifierPass());
74 // Add an appropriate TargetData instance for this module...
75 addPass(Passes, new TargetData("gccld", M));
77 // Often if the programmer does not specify proper prototypes for the
78 // functions they are calling, they end up calling a vararg version of the
79 // function that does not get a body filled in (the real function has typed
80 // arguments). This pass merges the two functions.
81 addPass(Passes, createFunctionResolvingPass());
83 if (!DisableOptimizations) {
84 // Linking modules together can lead to duplicated global constants, only
85 // keep one copy of each constant...
86 addPass(Passes, createConstantMergePass());
88 // If the -s command line option was specified, strip the symbols out of the
89 // resulting program to make it smaller. -s is a GCC option that we are
92 addPass(Passes, createSymbolStrippingPass());
95 // Now that composite has been compiled, scan through the module, looking
96 // for a main function. If main is defined, mark all other functions
98 addPass(Passes, createInternalizePass());
101 // Propagate constants at call sites into the functions they call.
102 addPass(Passes, createIPConstantPropagationPass());
104 // Remove unused arguments from functions...
105 addPass(Passes, createDeadArgEliminationPass());
108 addPass(Passes, createFunctionInliningPass()); // Inline small functions
110 // Run a few AA driven optimizations here and now, to cleanup the code.
111 // Eventually we should put an IP AA in place here.
113 addPass(Passes, createLICMPass()); // Hoist loop invariants
114 addPass(Passes, createLoadValueNumberingPass()); // GVN for load instrs
115 addPass(Passes, createGCSEPass()); // Remove common subexprs
117 // The FuncResolve pass may leave cruft around if functions were prototyped
118 // differently than they were defined. Remove this cruft.
119 addPass(Passes, createInstructionCombiningPass());
121 // Delete basic blocks, which optimization passes may have killed...
122 addPass(Passes, createCFGSimplificationPass());
124 // Now that we have optimized the program, discard unreachable functions...
125 addPass(Passes, createGlobalDCEPass());
128 // Make sure everything is still good.
129 Passes.add(createVerifierPass());
131 // Add the pass that writes bytecode to the output file...
132 addPass(Passes, new WriteBytecodePass(Out));
134 // Run our queue of passes all at once now, efficiently.
140 /// GenerateAssembly - generates a native assembly language source file from the
141 /// specified bytecode file.
144 /// InputFilename - The name of the output bytecode file.
145 /// OutputFilename - The name of the file to generate.
146 /// llc - The pathname to use for LLC.
147 /// envp - The environment to use when running LLC.
152 /// Return non-zero value on error.
155 GenerateAssembly(const std::string &OutputFilename,
156 const std::string &InputFilename,
157 const std::string &llc,
160 // Run LLC to convert the bytecode file into assembly code.
163 cmd[0] = llc.c_str();
166 cmd[3] = OutputFilename.c_str();
167 cmd[4] = InputFilename.c_str();
170 return ExecWait(cmd, envp);
173 /// GenerateNative - generates a native assembly language source file from the
174 /// specified assembly source file.
177 /// InputFilename - The name of the output bytecode file.
178 /// OutputFilename - The name of the file to generate.
179 /// Libraries - The list of libraries with which to link.
180 /// LibPaths - The list of directories in which to find libraries.
181 /// gcc - The pathname to use for GGC.
182 /// envp - A copy of the process's current environment.
187 /// Returns non-zero value on error.
190 GenerateNative(const std::string &OutputFilename,
191 const std::string &InputFilename,
192 const std::vector<std::string> &Libraries,
193 const std::vector<std::string> &LibPaths,
194 const std::string &gcc,
195 char ** const envp) {
196 // Remove these environment variables from the environment of the
197 // programs that we will execute. It appears that GCC sets these
198 // environment variables so that the programs it uses can configure
199 // themselves identically.
201 // However, when we invoke GCC below, we want it to use its normal
202 // configuration. Hence, we must sanitize its environment.
203 char ** clean_env = CopyEnv(envp);
204 if (clean_env == NULL)
206 RemoveEnv("LIBRARY_PATH", clean_env);
207 RemoveEnv("COLLECT_GCC_OPTIONS", clean_env);
208 RemoveEnv("GCC_EXEC_PREFIX", clean_env);
209 RemoveEnv("COMPILER_PATH", clean_env);
210 RemoveEnv("COLLECT_GCC", clean_env);
212 std::vector<const char *> cmd;
214 // Run GCC to assemble and link the program into native code.
217 // We can't just assemble and link the file with the system assembler
218 // and linker because we don't know where to put the _start symbol.
219 // GCC mysteriously knows how to do it.
220 cmd.push_back(gcc.c_str());
222 cmd.push_back(OutputFilename.c_str());
223 cmd.push_back(InputFilename.c_str());
225 // Adding the library paths creates a problem for native generation. If we
226 // include the search paths from llvmgcc, then we'll be telling normal gcc
227 // to look inside of llvmgcc's library directories for libraries. This is
228 // bad because those libraries hold only bytecode files (not native object
229 // files). In the end, we attempt to link the bytecode libgcc into a native
232 // Add in the library path options.
233 for (unsigned index=0; index < LibPaths.size(); index++) {
235 cmd.push_back(LibPaths[index].c_str());
239 // Add in the libraries to link.
240 std::vector<std::string> Libs(Libraries);
241 for (unsigned index = 0; index < Libs.size(); index++) {
242 if (Libs[index] != "crtend") {
243 Libs[index] = "-l" + Libs[index];
244 cmd.push_back(Libs[index].c_str());
249 // Run the compiler to assembly and link together the program.
250 return ExecWait(&(cmd[0]), clean_env);
253 } // End llvm namespace