1//===--- Tooling.cpp - Running clang standalone tools ---------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10//  This file implements functions to run clang tools standalone instead
11//  of running them as a plugin.
12//
13//===----------------------------------------------------------------------===//
14
15#include "clang/Tooling/Tooling.h"
16#include "clang/Driver/Compilation.h"
17#include "clang/Driver/Driver.h"
18#include "clang/Driver/Tool.h"
19#include "clang/Frontend/CompilerInstance.h"
20#include "clang/Frontend/FrontendDiagnostic.h"
21#include "clang/Frontend/TextDiagnosticPrinter.h"
22#include "clang/Tooling/ArgumentsAdjusters.h"
23#include "clang/Tooling/CompilationDatabase.h"
24#include "llvm/ADT/STLExtras.h"
25#include "llvm/Option/Option.h"
26#include "llvm/Support/Debug.h"
27#include "llvm/Support/FileSystem.h"
28#include "llvm/Support/Host.h"
29#include "llvm/Support/raw_ostream.h"
30
31// For chdir, see the comment in ClangTool::run for more information.
32#ifdef _WIN32
33#  include <direct.h>
34#else
35#  include <unistd.h>
36#endif
37
38namespace clang {
39namespace tooling {
40
41FrontendActionFactory::~FrontendActionFactory() {}
42
43// FIXME: This file contains structural duplication with other parts of the
44// code that sets up a compiler to run tools on it, and we should refactor
45// it to be based on the same framework.
46
47/// \brief Builds a clang driver initialized for running clang tools.
48static clang::driver::Driver *newDriver(clang::DiagnosticsEngine *Diagnostics,
49                                        const char *BinaryName) {
50  const std::string DefaultOutputName = "a.out";
51  clang::driver::Driver *CompilerDriver = new clang::driver::Driver(
52    BinaryName, llvm::sys::getDefaultTargetTriple(),
53    DefaultOutputName, *Diagnostics);
54  CompilerDriver->setTitle("clang_based_tool");
55  return CompilerDriver;
56}
57
58/// \brief Retrieves the clang CC1 specific flags out of the compilation's jobs.
59///
60/// Returns NULL on error.
61static const llvm::opt::ArgStringList *getCC1Arguments(
62    clang::DiagnosticsEngine *Diagnostics,
63    clang::driver::Compilation *Compilation) {
64  // We expect to get back exactly one Command job, if we didn't something
65  // failed. Extract that job from the Compilation.
66  const clang::driver::JobList &Jobs = Compilation->getJobs();
67  if (Jobs.size() != 1 || !isa<clang::driver::Command>(*Jobs.begin())) {
68    SmallString<256> error_msg;
69    llvm::raw_svector_ostream error_stream(error_msg);
70    Compilation->PrintJob(error_stream, Compilation->getJobs(), "; ", true);
71    Diagnostics->Report(clang::diag::err_fe_expected_compiler_job)
72        << error_stream.str();
73    return NULL;
74  }
75
76  // The one job we find should be to invoke clang again.
77  const clang::driver::Command *Cmd =
78      cast<clang::driver::Command>(*Jobs.begin());
79  if (StringRef(Cmd->getCreator().getName()) != "clang") {
80    Diagnostics->Report(clang::diag::err_fe_expected_clang_command);
81    return NULL;
82  }
83
84  return &Cmd->getArguments();
85}
86
87/// \brief Returns a clang build invocation initialized from the CC1 flags.
88static clang::CompilerInvocation *newInvocation(
89    clang::DiagnosticsEngine *Diagnostics,
90    const llvm::opt::ArgStringList &CC1Args) {
91  assert(!CC1Args.empty() && "Must at least contain the program name!");
92  clang::CompilerInvocation *Invocation = new clang::CompilerInvocation;
93  clang::CompilerInvocation::CreateFromArgs(
94      *Invocation, CC1Args.data() + 1, CC1Args.data() + CC1Args.size(),
95      *Diagnostics);
96  Invocation->getFrontendOpts().DisableFree = false;
97  Invocation->getCodeGenOpts().DisableFree = false;
98  return Invocation;
99}
100
101bool runToolOnCode(clang::FrontendAction *ToolAction, const Twine &Code,
102                   const Twine &FileName) {
103  return runToolOnCodeWithArgs(
104      ToolAction, Code, std::vector<std::string>(), FileName);
105}
106
107bool runToolOnCodeWithArgs(clang::FrontendAction *ToolAction, const Twine &Code,
108                           const std::vector<std::string> &Args,
109                           const Twine &FileName) {
110  SmallString<16> FileNameStorage;
111  StringRef FileNameRef = FileName.toNullTerminatedStringRef(FileNameStorage);
112  std::vector<std::string> Commands;
113  Commands.push_back("clang-tool");
114  Commands.push_back("-fsyntax-only");
115  Commands.insert(Commands.end(), Args.begin(), Args.end());
116  Commands.push_back(FileNameRef.data());
117  FileManager Files((FileSystemOptions()));
118  ToolInvocation Invocation(Commands, ToolAction, &Files);
119
120  SmallString<1024> CodeStorage;
121  Invocation.mapVirtualFile(FileNameRef,
122                            Code.toNullTerminatedStringRef(CodeStorage));
123  return Invocation.run();
124}
125
126std::string getAbsolutePath(StringRef File) {
127  SmallString<1024> BaseDirectory;
128  if (const char *PWD = ::getenv("PWD"))
129    BaseDirectory = PWD;
130  else
131    llvm::sys::fs::current_path(BaseDirectory);
132  SmallString<1024> PathStorage;
133  if (llvm::sys::path::is_absolute(File)) {
134    llvm::sys::path::native(File, PathStorage);
135    return PathStorage.str();
136  }
137  StringRef RelativePath(File);
138  // FIXME: Should '.\\' be accepted on Win32?
139  if (RelativePath.startswith("./")) {
140    RelativePath = RelativePath.substr(strlen("./"));
141  }
142  SmallString<1024> AbsolutePath(BaseDirectory);
143  llvm::sys::path::append(AbsolutePath, RelativePath);
144  llvm::sys::path::native(Twine(AbsolutePath), PathStorage);
145  return PathStorage.str();
146}
147
148ToolInvocation::ToolInvocation(
149    ArrayRef<std::string> CommandLine, FrontendAction *ToolAction,
150    FileManager *Files)
151    : CommandLine(CommandLine.vec()), ToolAction(ToolAction), Files(Files) {
152}
153
154void ToolInvocation::mapVirtualFile(StringRef FilePath, StringRef Content) {
155  SmallString<1024> PathStorage;
156  llvm::sys::path::native(FilePath, PathStorage);
157  MappedFileContents[PathStorage] = Content;
158}
159
160bool ToolInvocation::run() {
161  std::vector<const char*> Argv;
162  for (int I = 0, E = CommandLine.size(); I != E; ++I)
163    Argv.push_back(CommandLine[I].c_str());
164  const char *const BinaryName = Argv[0];
165  IntrusiveRefCntPtr<DiagnosticOptions> DiagOpts = new DiagnosticOptions();
166  TextDiagnosticPrinter DiagnosticPrinter(
167      llvm::errs(), &*DiagOpts);
168  DiagnosticsEngine Diagnostics(
169    IntrusiveRefCntPtr<clang::DiagnosticIDs>(new DiagnosticIDs()),
170    &*DiagOpts, &DiagnosticPrinter, false);
171
172  const OwningPtr<clang::driver::Driver> Driver(
173      newDriver(&Diagnostics, BinaryName));
174  // Since the input might only be virtual, don't check whether it exists.
175  Driver->setCheckInputsExist(false);
176  const OwningPtr<clang::driver::Compilation> Compilation(
177      Driver->BuildCompilation(llvm::makeArrayRef(Argv)));
178  const llvm::opt::ArgStringList *const CC1Args = getCC1Arguments(
179      &Diagnostics, Compilation.get());
180  if (CC1Args == NULL) {
181    return false;
182  }
183  OwningPtr<clang::CompilerInvocation> Invocation(
184      newInvocation(&Diagnostics, *CC1Args));
185  return runInvocation(BinaryName, Compilation.get(), Invocation.take());
186}
187
188bool ToolInvocation::runInvocation(
189    const char *BinaryName,
190    clang::driver::Compilation *Compilation,
191    clang::CompilerInvocation *Invocation) {
192  // Show the invocation, with -v.
193  if (Invocation->getHeaderSearchOpts().Verbose) {
194    llvm::errs() << "clang Invocation:\n";
195    Compilation->PrintJob(llvm::errs(), Compilation->getJobs(), "\n", true);
196    llvm::errs() << "\n";
197  }
198
199  // Create a compiler instance to handle the actual work.
200  clang::CompilerInstance Compiler;
201  Compiler.setInvocation(Invocation);
202  Compiler.setFileManager(Files);
203  // FIXME: What about LangOpts?
204
205  // ToolAction can have lifetime requirements for Compiler or its members, and
206  // we need to ensure it's deleted earlier than Compiler. So we pass it to an
207  // OwningPtr declared after the Compiler variable.
208  OwningPtr<FrontendAction> ScopedToolAction(ToolAction.take());
209
210  // Create the compilers actual diagnostics engine.
211  Compiler.createDiagnostics();
212  if (!Compiler.hasDiagnostics())
213    return false;
214
215  Compiler.createSourceManager(*Files);
216  addFileMappingsTo(Compiler.getSourceManager());
217
218  const bool Success = Compiler.ExecuteAction(*ScopedToolAction);
219
220  Compiler.resetAndLeakFileManager();
221  Files->clearStatCaches();
222  return Success;
223}
224
225void ToolInvocation::addFileMappingsTo(SourceManager &Sources) {
226  for (llvm::StringMap<StringRef>::const_iterator
227           It = MappedFileContents.begin(), End = MappedFileContents.end();
228       It != End; ++It) {
229    // Inject the code as the given file name into the preprocessor options.
230    const llvm::MemoryBuffer *Input =
231        llvm::MemoryBuffer::getMemBuffer(It->getValue());
232    // FIXME: figure out what '0' stands for.
233    const FileEntry *FromFile = Files->getVirtualFile(
234        It->getKey(), Input->getBufferSize(), 0);
235    Sources.overrideFileContents(FromFile, Input);
236  }
237}
238
239ClangTool::ClangTool(const CompilationDatabase &Compilations,
240                     ArrayRef<std::string> SourcePaths)
241    : Files((FileSystemOptions())) {
242  ArgsAdjusters.push_back(new ClangStripOutputAdjuster());
243  ArgsAdjusters.push_back(new ClangSyntaxOnlyAdjuster());
244  for (unsigned I = 0, E = SourcePaths.size(); I != E; ++I) {
245    SmallString<1024> File(getAbsolutePath(SourcePaths[I]));
246
247    std::vector<CompileCommand> CompileCommandsForFile =
248      Compilations.getCompileCommands(File.str());
249    if (!CompileCommandsForFile.empty()) {
250      for (int I = 0, E = CompileCommandsForFile.size(); I != E; ++I) {
251        CompileCommands.push_back(std::make_pair(File.str(),
252                                  CompileCommandsForFile[I]));
253      }
254    } else {
255      // FIXME: There are two use cases here: doing a fuzzy
256      // "find . -name '*.cc' |xargs tool" match, where as a user I don't care
257      // about the .cc files that were not found, and the use case where I
258      // specify all files I want to run over explicitly, where this should
259      // be an error. We'll want to add an option for this.
260      llvm::outs() << "Skipping " << File << ". Command line not found.\n";
261    }
262  }
263}
264
265void ClangTool::mapVirtualFile(StringRef FilePath, StringRef Content) {
266  MappedFileContents.push_back(std::make_pair(FilePath, Content));
267}
268
269void ClangTool::setArgumentsAdjuster(ArgumentsAdjuster *Adjuster) {
270  clearArgumentsAdjusters();
271  appendArgumentsAdjuster(Adjuster);
272}
273
274void ClangTool::appendArgumentsAdjuster(ArgumentsAdjuster *Adjuster) {
275  ArgsAdjusters.push_back(Adjuster);
276}
277
278void ClangTool::clearArgumentsAdjusters() {
279  for (unsigned I = 0, E = ArgsAdjusters.size(); I != E; ++I)
280    delete ArgsAdjusters[I];
281  ArgsAdjusters.clear();
282}
283
284int ClangTool::run(FrontendActionFactory *ActionFactory) {
285  // Exists solely for the purpose of lookup of the resource path.
286  // This just needs to be some symbol in the binary.
287  static int StaticSymbol;
288  // The driver detects the builtin header path based on the path of the
289  // executable.
290  // FIXME: On linux, GetMainExecutable is independent of the value of the
291  // first argument, thus allowing ClangTool and runToolOnCode to just
292  // pass in made-up names here. Make sure this works on other platforms.
293  std::string MainExecutable =
294      llvm::sys::fs::getMainExecutable("clang_tool", &StaticSymbol);
295
296  bool ProcessingFailed = false;
297  for (unsigned I = 0; I < CompileCommands.size(); ++I) {
298    std::string File = CompileCommands[I].first;
299    // FIXME: chdir is thread hostile; on the other hand, creating the same
300    // behavior as chdir is complex: chdir resolves the path once, thus
301    // guaranteeing that all subsequent relative path operations work
302    // on the same path the original chdir resulted in. This makes a difference
303    // for example on network filesystems, where symlinks might be switched
304    // during runtime of the tool. Fixing this depends on having a file system
305    // abstraction that allows openat() style interactions.
306    if (chdir(CompileCommands[I].second.Directory.c_str()))
307      llvm::report_fatal_error("Cannot chdir into \"" +
308                               CompileCommands[I].second.Directory + "\n!");
309    std::vector<std::string> CommandLine = CompileCommands[I].second.CommandLine;
310    for (unsigned I = 0, E = ArgsAdjusters.size(); I != E; ++I)
311      CommandLine = ArgsAdjusters[I]->Adjust(CommandLine);
312    assert(!CommandLine.empty());
313    CommandLine[0] = MainExecutable;
314    // FIXME: We need a callback mechanism for the tool writer to output a
315    // customized message for each file.
316    DEBUG({
317      llvm::dbgs() << "Processing: " << File << ".\n";
318    });
319    ToolInvocation Invocation(CommandLine, ActionFactory->create(), &Files);
320    for (int I = 0, E = MappedFileContents.size(); I != E; ++I) {
321      Invocation.mapVirtualFile(MappedFileContents[I].first,
322                                MappedFileContents[I].second);
323    }
324    if (!Invocation.run()) {
325      // FIXME: Diagnostics should be used instead.
326      llvm::errs() << "Error while processing " << File << ".\n";
327      ProcessingFailed = true;
328    }
329  }
330  return ProcessingFailed ? 1 : 0;
331}
332
333} // end namespace tooling
334} // end namespace clang
335