[cfe-commits] [PATCH] Implements support to run standalone tools

Manuel Klimek klimek at google.com
Tue Jan 31 12:07:35 PST 2012


On Tue, Jan 31, 2012 at 8:31 PM, David Blaikie <dblaikie at gmail.com> wrote:
> On Tue, Jan 31, 2012 at 11:16 AM, Manuel Klimek <klimek at google.com> wrote:
>> On Tue, Jan 31, 2012 at 4:34 PM, nobled <nobled at dreamwidth.org> wrote:
>>> On Mon, Jan 30, 2012 at 2:49 PM, Manuel Klimek <klimek at google.com> wrote:
>>>> Ping + a re-based version of the patch.
>>>>
>>>
>>>
>>> This looks really cool. Aside from the create()/New() half-rename confusion,
>>> just some minor nits:
>>>
>>>> +/// \brief Returns a new FrontendActionFactory for any type that provides an
>>>> +/// implementation of NewFrontendAction().
>>>> +///
>>>> +/// FactoryT must implement: FrontendAction *NewFrontendAction().
>>>> +///
>>>> +/// Example:
>>>> +/// struct ProvidesFrontendActions {
>>>> +///   FrontendActionFactory *NewFrontendAction();
>>> You meant "FrontendAction *" here, right?
>>
>> Yep, done. Thx for the catch.
>>
>>>> +/// } Factory;
>>>> +/// FrontendActionFactory *FactoryAdapter =
>>>> +///   newFrontendActionFactory(&Factory);
>>>> +template <typename FactoryT>
>>>> +FrontendActionFactory *newFrontendActionFactory(FactoryT *ActionFactory);
>>>> +
>>>> +/// \brief Runs (and deletes) the tool on 'Code' with the -fsynatx-only flag.
>>> syntax*
>>
>> Done.
>>
>>>> +///
>>>> +/// \param ToolAction The action to run over the code.
>>>> +/// \param Code C++ code.
>>>> +///
>>>> +/// \return - True if 'ToolAction' was successfully executed.
>>>> +bool runSyntaxOnlyToolOnCode(
>>>> +    clang::FrontendAction *ToolAction, llvm::StringRef Code);
>>>> +
>>>> +/// \brief Converts a vector<string> into a vector<char*> suitable to pass
>>>> +/// to main-style functions taking (int Argc, char *Argv[]).
>>>> +std::vector<char*> commandLineToArgv(const std::vector<std::string> *Command);
>>> This looks like it can just take an ArrayRef<const std::string>...
>>
>> Done.
>>
>>>> +/// \see JsonCompileCommandLineDatabase
>>>> +CompileCommand findCompileArgsInJsonDatabase(
>>>> +    llvm::StringRef FileName, llvm::StringRef JsonDatabase,
>>>> +    std::string &ErrorMessage);
>>> Just a note on the whole patch: you don't need the llvm:: prefix on StringRef
>>> or ArrayRef, or a few other common types in clang code anymore. You just have
>>> to include "clang/Basic/LLVM.h" to use them inside the clang namespace.
>>
>> Cool, done.
>>
>>>> +  /// \brief Returns the file manager used in the tool.
>>>> +  ///
>>>> +  /// The file manager is shared between all translation units.
>>>> +  FileManager &getFiles() { return Files; }
>>>> +
>>>> + private:
>>>> +  /// \brief Add translation units to run the tool over.
>>>> +  ///
>>>> +  /// Translation units not found in JsonDatabaseDirectory (see constructore)
>>> constructor*
>>
>> Done.
>>
>>>> +  /// will be skipped.
>>>> +  void addTranslationUnits(
>>>> +      llvm::StringRef JsonDatabaseDirectory,
>>>> +      llvm::ArrayRef<std::string> TranslationUnits);
>>>> +
>>>> +  // We store command lines as pair (file name, command line).
>>>> +  typedef std::pair< std::string, std::vector<std::string> > CommandLine;
>>>> +  std::vector<CommandLine> CommandLines;
>>>> +
>>>> +  FileManager Files;
>>>> +  // Maps <file name> -> <file content>.
>>>> +  std::map<std::string, std::string> MappedFileContents;
>>>> +};
>>>> +
>>>> +template <typename T>
>>>> +FrontendActionFactory *newFrontendActionFactory() {
>>>> +  class SimpleFrontendActionFactory : public FrontendActionFactory {
>>>> +  public:
>>>> +    virtual clang::FrontendAction *New() { return new T; }
>>> You mean "create()"?
>>
>> Yep, missing tests... Added.
>>
>>>> +  };
>>>> +
>>>> +  return new SimpleFrontendActionFactory;
>>>> +}
>>>> +
>>>> +template <typename FactoryT>
>>>> +FrontendActionFactory *newFrontendActionFactory(FactoryT *ActionFactory) {
>>>> +  class FrontendActionFactoryAdapter : public FrontendActionFactory {
>>>> +  public:
>>>> +    explicit FrontendActionFactoryAdapter(FactoryT *ActionFactory)
>>>> +      : ActionFactory(ActionFactory) {}
>>>> +
>>>> +    virtual clang::FrontendAction *New() {
>>> "create()" here, too... It looks like these templates aren't covered by the
>>> unittest/getting instantiated at all.
>>
>> Yep, correct. This is tested by other tests in our branch, but it
>> makes sense to test them independently as part of this layer, too.
>> Added a unit test.
>>
>>> diff --git a/lib/Tooling/CMakeLists.txt b/lib/Tooling/CMakeLists.txt
>>> new file mode 100644
>>> index 0000000..b0e1235
>>> --- /dev/null
>>>> +++ b/lib/Tooling/CMakeLists.txt
>>> @@ -0,0 +1,6 @@
>>>> +set(LLVM_LINK_COMPONENTS support)
>>>> +SET(LLVM_USED_LIBS clangBasic clangFrontend clangAST clangRewrite)
>>> Here and in unittest/Tooling/Makefile, you set a dependency on the Rewrite lib,
>>> but I don't see any Rewrite/ includes. Is there a reason for that?
>>
>> I forgot to rip them out when I separated the dependencies from the
>> refactoring library in our branch (upcoming super awesome feature ;).
>> Removed.
>>
>>>> +bool ToolInvocation::run() {
>>>> +  const std::vector<char*> Argv = commandLineToArgv(&CommandLine);
>>>> +  const char *const BinaryName = Argv[0];
>>>> +  DiagnosticOptions DefaultDiagnosticOptions;
>>>> +  TextDiagnosticPrinter DiagnosticPrinter(
>>>> +      llvm::errs(), DefaultDiagnosticOptions);
>>>> +  DiagnosticsEngine Diagnostics(llvm::IntrusiveRefCntPtr<clang::DiagnosticIDs>(
>>>> +      new DiagnosticIDs()), &DiagnosticPrinter, false);
>>>> +
>>>> +  const llvm::OwningPtr<clang::driver::Driver> Driver(
>>>> +      newDriver(&Diagnostics, BinaryName));
>>>> +  // Since the input might only be virtual, don't check whether it exists.
>>>> +  Driver->setCheckInputsExist(false);
>>>> +  const llvm::OwningPtr<clang::driver::Compilation> Compilation(
>>>> +      Driver->BuildCompilation(llvm::ArrayRef<const char*>(
>>>> +          &Argv[0], Argv.size() - 1)));
>>> (The fact that the Driver entrypoint need a song and dance to produce
>>> a char * array
>>> is annoying, but that's obviously a pre-existing problem with the
>>> API... I'm hoping
>>> to fix that in the future.)
>>>
>>>> +ClangTool::ClangTool(int argc, char **argv)
>>>> +    : Files((FileSystemOptions())) {
>>>> +  if (argc < 3) {
>>>> +    llvm::outs() << "Usage: " << argv[0] << " <cmake-output-dir> "
>>>> +                 << "<file1> <file2> ...\n";
>>>> +    exit(1);
>>> This looks like a good place to use llvm::report_fatal_error() instead. Same
>>> for the two other exit(1) calls below.
>>
>> Done.
>>
>>>
>>>> +  }
>>>> +  addTranslationUnits(argv[1], std::vector<std::string>(argv + 2, argv + argc));
>>>> +}
>>>> +
>>>> +void ClangTool::addTranslationUnits(
>>>> +    llvm::StringRef JsonDatabaseDirectory,
>>>> +    llvm::ArrayRef<std::string> TranslationUnits) {
>>> TranslationUnits can just be an ArrayRef<StringRef> here.
>>
>> We're handing in a vector<string> which we created from argv, argc in
>> the ClangTool constructor. Is there a magic way to create a
>> ArrayRef<StringRef> from a vector<string> that can't figure out?
>
> I think the implication is that you'd create a vector<StringRef>
> instead, to match the ArrayRef<StringRef> parameter. Does that
> work/make sense?

Heh, yea, that obviously makes sense... I should go home now, it's
getting late :)

Please find the updated patch attached.

Cheers,
/Manuel

>
>>
>>>> +  llvm::SmallString<1024> JsonDatabasePath(JsonDatabaseDirectory);
>>>> +  llvm::sys::path::append(JsonDatabasePath, "compile_commands.json");
>>>> +  llvm::OwningPtr<llvm::MemoryBuffer> JsonDatabase;
>>>> +  llvm::error_code Result =
>>>> +      llvm::MemoryBuffer::getFile(JsonDatabasePath, JsonDatabase);
>>>> +  if (Result != 0) {
>>>> +    llvm::outs() << "Error while opening JSON database: " << Result.message()
>>>> +                 << "\n";
>>>> +    exit(1);
>>>> +  }
>>>> +  llvm::StringRef BaseDirectory(::getenv("PWD"));
>>>> +  for (unsigned I = 0; I < TranslationUnits.size(); ++I) {
>>>> +    llvm::SmallString<1024> File(getAbsolutePath(
>>>> +        TranslationUnits[I], BaseDirectory));
>>>> +    std::string ErrorMessage;
>>>> +    clang::tooling::CompileCommand LookupResult =
>>>> +        clang::tooling::findCompileArgsInJsonDatabase(
>>>> +            File.str(), JsonDatabase->getBuffer(), ErrorMessage);
>>>> +    if (!ErrorMessage.empty()) {
>>>> +      llvm::outs() << "Error while parsing JSON database: " << ErrorMessage
>>>> +                   << "\n";
>>>> +      exit(1);
>>>> +    }
>>>> +    if (!LookupResult.CommandLine.empty()) {
>>>> +      if (!LookupResult.Directory.empty()) {
>>>> +        // FIXME: What should happen if CommandLine includes -working-directory
>>>> +        // as well?
>>>> +        LookupResult.CommandLine.push_back(
>>>> +            "-working-directory=" + LookupResult.Directory);
>>>> +      }
>>>> +      CommandLines.push_back(make_pair(File.str(), LookupResult.CommandLine));
>>>> +    } else {
>>>> +      // FIXME: There are two use cases here: doing a fuzzy
>>>> +      // "find . -name '*.cc' |xargs tool" match, where as a user I don't care
>>>> +      // about the .cc files that were not found, and the use case where I
>>>> +      // specify all files I want to run over explicitly, where this should
>>>> +      // be an error. We'll want to add an option for this.
>>>> +      llvm::outs() << "Skipping " << File << ". Command line not found.\n";
>>>> +    }
>>>> +  }
>>>> +}
>>>> +
>>>> +void ClangTool::mapVirtualFiles(
>>>> +    const std::map<std::string, std::string> &FileContents) {
>>> At first glance, MappedFileContents in both ToolInvocation and ClangTool
>>> looks like it can just be a std::map<StringRef,StringRef>. Does it really need
>>> to have ownership of its own malloc'd copy of the file contents?
>>
>> An excellent point. Changed.
>>
>>>> +  MappedFileContents.insert(FileContents.begin(), FileContents.end());
>>>> +}
>>>> +
>>>> +int ClangTool::run(FrontendActionFactory *ActionFactory) {
>>> It also looks like the ClangTool class isn't covered by the unittest, either.
>>
>> I added ClangCheck.cpp and an integration test. I first wanted to keep
>> this change smaller, but leaving out clang-check and an integration
>> test was probably a little over the top...
>>
>>> Looking forward to this landing!
>>
>> Me too :) Thanks a lot for reviewing!
>>
>> Cheers,
>> /Manuel
>>
>>>
>>>
>>>> On Wed, Jan 25, 2012 at 3:59 PM, Manuel Klimek <klimek at google.com> wrote:
>>>>> On Tue, Jan 24, 2012 at 4:43 PM, Hal Finkel <hfinkel at anl.gov> wrote:
>>>>>> On Tue, 2012-01-24 at 11:03 +0100, Manuel Klimek wrote:
>>>>>>> The attached patch adds support to run clang tools (FrontendActions)
>>>>>>> as standalone tools, or repeatedly in-memory in a process.
>>>>>>> This is useful for unit testing, map-reduce-style applications, source
>>>>>>> transformation daemons, and command line tools.
>>>>>>
>>>>>> I am also interested in having this kind of functionality. A few quick
>>>>>> comments:
>>>>>>
>>>>>> 1. The coding standards say that function names should begin with a
>>>>>> lower-case letter.
>>>>>
>>>>> Done. I jumped on the opportunity to dogfood refactoring support in
>>>>> our current tooling branch and wrote a script that changed all
>>>>> incorrectly named functions automatically (and created a sed-script to
>>>>> post-produce comment changes, which made me notice a bug in a
>>>>> comment).
>>>>>
>>>>>> 2. The comments contain several references to CMake; what, if anything,
>>>>>> in this patch is tied to CMake, or designed to be compatible with CMake?
>>>>>>
>>>>>> 2b.
>>>>>>
>>>>>>
>>>>>>> +/// \param JsonDatabase A JSON formatted list of compile commands.
>>>>>>> This lookup
>>>>>>> +/// command supports only a subset of the JSON standard as written by
>>>>>>> CMake.
>>>>>>>
>>>>>>
>>>>>> Please be more verbose here. What is not supported?
>>>>>>
>>>>>> Generally, I think that it would be helpful for you to provide a
>>>>>> paragraph or two explaining how this extension is to be used, what kind
>>>>>> of things can be specified in JSON inputs, how this ties into CMake (or
>>>>>> not), etc. with a few small examples. Some of this can be gleaned from
>>>>>> the test case, but some nicely-formatted text (without all of the
>>>>>> escaping) would, IMHO, be earlier to read.
>>>>>
>>>>> Hopefully better expressed now. Please let me know if you want more /
>>>>> different details.
>>>>>
>>>>> Thanks a lot for the review!
>>>>> /Manuel
>>>>>
>>>>>>
>>>>>>  -Hal
>>>>>>
>>>>>>>
>>>>>>> Cheers,
>>>>>>> /Manuel
>>>>>>>
>>>>>>> Rietveld link:
>>>>>>> http://codereview.appspot.com/5570054/
>>>>>>> _______________________________________________
>>>>>>> cfe-commits mailing list
>>>>>>> cfe-commits at cs.uiuc.edu
>>>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>>>>>
>>>>>> --
>>>>>> Hal Finkel
>>>>>> Postdoctoral Appointee
>>>>>> Leadership Computing Facility
>>>>>> Argonne National Laboratory
>>>>>>
>>>>
>>>> _______________________________________________
>>>> cfe-commits mailing list
>>>> cfe-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>>>
>>
>> _______________________________________________
>> cfe-commits mailing list
>> cfe-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: tooling.patch
Type: text/x-patch
Size: 47772 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20120131/4289580b/attachment.bin>


More information about the cfe-commits mailing list