Initial, rough kinda working version ClangComplete

Many things need to be ironed out still.

autoload/youcompleteme.vim

@@ -21,16 +21,19 @@ set cpo&vim
 
 " This needs to be called outside of a function
 let s:script_folder_path = escape( expand( '<sfile>:p:h' ), '\' )
-let s:old_cursor_text = ''
+let s:searched_and_no_results_found = 0
+let s:should_use_clang = 0
 let g:ycm_min_num_of_chars_for_completion = 2
 
-" Set up the plugin, load all our modules, bind our keys etc.
+" TODO: check for a non-broken version of Vim and stop executing (with an error
+" message) if the detected version is too old
+
 function! youcompleteme#Enable()
 
   augroup youcompleteme
     autocmd!
     autocmd CursorMovedI * call s:OnMovedI()
-    " BufWinEnter/Leave?
+    " TODO: BufWinEnter/Leave?
     autocmd BufRead,BufEnter * call s:OnBufferVisit()
     autocmd CursorHold,CursorHoldI * call s:OnCursorHold()
   augroup END
@@ -62,6 +65,7 @@ function! youcompleteme#Enable()
   exe 'python sys.path = sys.path + ["' . s:script_folder_path . '/../python"]'
   py import ycm
   py csystem = ycm.CompletionSystem()
+  py clangcomp = ycm.ClangComplete()
 endfunction
 
 
@@ -96,8 +100,8 @@ endfunction
 
 
 function! s:AddIdentifierIfNeeded()
-  py vim.command( "let should_add_identifier = '" +
+  py vim.command( "let should_add_identifier = " +
-        \ str( int( ycm.ShouldAddIdentifier() ) ) + "'" )
+        \ str( int( ycm.ShouldAddIdentifier() ) ) )
   if should_add_identifier != 1
     return
   endif
@@ -125,37 +129,38 @@ function! s:InvokeCompletion()
     return
   endif
 
-  py vim.command( "let cursor_text = '" + ycm.CurrentCursorTextVim() + "'" )
+  " This is tricky. First, having 'refresh' set to 'always' in the dictionary
-
+  " that our completion function returns makes sure that our completion function
-  " infinite loops are bad, mkay?
+  " is called on every keystroke when the completion menu is showing
-  if cursor_text == '' || cursor_text == s:old_cursor_text
+  " (pumvisible() == true). So there's no point in invoking the completion menu
+  " with our feedkeys call then.
+  " Secondly, when the sequence of characters the user typed produces no
+  " results in our search an infinite loop can occur. The problem is that our
+  " feedkeys call triggers the OnCursorMovedI event which we are tied to.
+  " So we solve this with the searched_and_no_results_found script-scope
+  " variable that prevents this infinite loop from starting.
+  if pumvisible() || s:searched_and_no_results_found
+    let s:searched_and_no_results_found = 0
     return
   endif
 
   " <c-x><c-u> invokes the user's completion function (which we have set to
-  " youcompleteme#Complete), and <c-p> tells vim to select the previous
+  " youcompleteme#Complete), and <c-p> tells Vim to select the previous
-  " completion candidate. This is necessary because by default, vim selects the
+  " completion candidate. This is necessary because by default, Vim selects the
   " first candidate when completion is invoked, and selecting a candidate
-  " automatically replaces the current text with it. Calling <c-p> forces vim to
+  " automatically replaces the current text with it. Calling <c-p> forces Vim to
   " deselect the first candidate and in turn preserve the user's current text
   " until he explicitly chooses to replace it with a completion.
   call feedkeys( "\<C-X>\<C-U>\<C-P>", 'n' )
 endfunction
 
 
-" This is our main entry point. This is what vim calls to get completions.
+function! s:IdentifierCompletion(query)
-function! youcompleteme#Complete(findstart, base)
+  if strlen( a:query ) < g:ycm_min_num_of_chars_for_completion
-  if a:findstart
-    py vim.command( 'let start_column = ' + str(
-                   \ ycm.CompletionStartColumn() ) )
-    return start_column
-  else
-    let s:old_cursor_text = a:base
-    if strlen( a:base ) < g:ycm_min_num_of_chars_for_completion
     return []
   endif
 
-	  py csystem.CandidatesForQueryAsync( vim.eval('a:base') )
+  py csystem.CandidatesForQueryAsync( vim.eval('a:query') )
 
   let l:results_ready = 0
   while !l:results_ready
@@ -175,13 +180,57 @@ results = csystem.CandidatesFromStoredRequest()
 if results:
   vim.command( 'let l:results = ' + str( results ) )
 EOF
+
+  let s:searched_and_no_results_found = len( l:results ) == 0
+
   " We need a very recent version of vim for this to work; otherwise, even
   " when we set refresh = always, vim won't call our completefunc on every
   " keystroke. The problem is still present in vim 7.3.390 but is fixed in
   " 7.3.475. It's possible that patch 404 was the one that fixed this issue,
   " but I haven't tested this assumption.
   " A bug in vim causes the '.' register to break when we use set this... sigh
-		return { 'words' : l:results, 'refresh' : 'always'}
+  return { 'words' : l:results, 'refresh' : 'always' }
+endfunction
+
+
+function! s:ClangCompletion( query )
+  py vim.command( 'let l:results = ' +
+        \ str( clangcomp.CandidatesForQuery( vim.eval( 'a:query' ) ) ) )
+
+  let s:searched_and_no_results_found = len( l:results ) == 0
+  return { 'words' : l:results, 'refresh' : 'always' }
+endfunction
+
+
+" This is our main entry point. This is what vim calls to get completions.
+function! youcompleteme#Complete( findstart, base )
+  if a:findstart
+    py << EOF
+start_column = ycm.CompletionStartColumn()
+vim.command( 'let l:start_column = ' + str( start_column ) )
+vim.command( 'let s:should_use_clang = ' +
+              str( int( ycm.ShouldUseClang( start_column ) ) ) )
+EOF
+
+    if ( s:should_use_clang )
+      return start_column
+    else
+      let l:current_column = col('.') - 1
+      let l:query_length = current_column - start_column
+
+      if ( query_length < g:ycm_min_num_of_chars_for_completion )
+        " for vim, -2 means not found but don't trigger an error message
+        " see :h complete-functions
+        return -2
+      endif
+      return start_column
+    endif
+  else
+    if ( s:should_use_clang )
+      return s:ClangCompletion( a:base )
+    else
+      return s:IdentifierCompletion( a:base )
+    endif
   endif
 endfunction
 

cpp/ycm/ClangComplete.cpp

@@ -31,9 +31,10 @@ std::vector< CXUnsavedFile > ToCXUnsavedFiles(
   std::vector< CXUnsavedFile > clang_unsaved_files( unsaved_files.size() );
   for ( uint i = 0; i < unsaved_files.size(); ++i )
   {
-    clang_unsaved_files[ i ].Filename = unsaved_files[ i ].filename_->c_str();
+    // TODO: assert non-null
-    clang_unsaved_files[ i ].Contents = unsaved_files[ i ].contents_->c_str();
+    clang_unsaved_files[ i ].Filename = unsaved_files[ i ].filename_;
-    clang_unsaved_files[ i ].Length   = unsaved_files[ i ].contents_->length();
+    clang_unsaved_files[ i ].Contents = unsaved_files[ i ].contents_;
+    clang_unsaved_files[ i ].Length   = unsaved_files[ i ].length_;
   }
 
   return clang_unsaved_files;
@@ -155,9 +156,15 @@ std::vector< std::string > ClangComplete::CandidatesForLocationInFile(
   std::vector< CXUnsavedFile > cxunsaved_files = ToCXUnsavedFiles(
       unsaved_files );
 
-  // TODO: figure out does codeCompleteAt reparse the TU if the underlying
+  // codeCompleteAt reparses the TU if the underlying source file has changed on
-  // source file has changed on disk since the last time the TU was updated and
+  // disk since the last time the TU was updated and there are no unsaved files.
-  // there are no unsaved files.
+  // If there are unsaved files, then codeCompleteAt will parse the in-memory
+  // file contents we are giving it. In short, it is NEVER a good idea to call
+  // clang_reparseTranslationUnit right before a call to clang_codeCompleteAt.
+  // The only makes clang reparse the whole file TWICE, which has a huge impact
+  // on latency. At the time of writing, it seems that most users of libclang
+  // in the open-source world don't realize this (I checked). Some don't even
+  // call reparse*, but parse* which is even less efficient.
 
   CXCodeCompleteResults *results =
     clang_codeCompleteAt( GetTranslationUnitForFile( filename, unsaved_files ),
@@ -189,7 +196,7 @@ CXTranslationUnit ClangComplete::CreateTranslationUnit(
       flags.size(),
       &cxunsaved_files[ 0 ],
       cxunsaved_files.size(),
-      clang_defaultEditingTranslationUnitOptions());
+      clang_defaultEditingTranslationUnitOptions() );
 }
 
 

cpp/ycm/ClangComplete.h

@@ -31,8 +31,20 @@ namespace YouCompleteMe
 
 struct UnsavedFile
 {
-  std::string *filename_;
+  UnsavedFile() : filename_( NULL ), contents_( NULL ), length_( 0 ) {}
-  std::string *contents_;
+
+  const char *filename_;
+  const char *contents_;
+  unsigned long length_;
+
+  // We need this to be able to export this struct to Python via Boost.Python's
+  // methods. I have no clue why, but it won't compile without it.
+  // TODO: report this problem on the Boost bug tracker, the default equality
+  // operator should be more than adequate here
+  bool operator==( const UnsavedFile &other) const
+  {
+    return filename_ == other.filename_ && contents_ == other.contents_;
+  }
 };
 
 

cpp/ycm/indexer.cpp

@@ -16,6 +16,7 @@
 // along with YouCompleteMe.  If not, see <http://www.gnu.org/licenses/>.
 
 #include "Completer.h"
+#include "ClangComplete.h"
 #include "Future.h"
 
 #include <boost/python.hpp>
@@ -27,17 +28,41 @@ BOOST_PYTHON_MODULE(indexer)
   using namespace boost::python;
   using namespace YouCompleteMe;
 
+	class_< std::vector< std::string > >( "StringVec" )
+		.def( vector_indexing_suite< std::vector< std::string > >() );
+
   class_< Future >( "Future" )
     .def( "ResultsReady", &Future::ResultsReady )
     .def( "GetResults", &Future::GetResults );
 
   class_< Completer, boost::noncopyable >( "Completer" )
     .def( "EnableThreading", &Completer::EnableThreading )
-    // .def( "AddCandidatesToDatabase", actd )
     .def( "AddCandidatesToDatabase", &Completer::AddCandidatesToDatabase )
     .def( "CandidatesForQueryAndTypeAsync",
           &Completer::CandidatesForQueryAndTypeAsync );
 
-	class_< std::vector< std::string > >( "StringVec" )
+  // CAREFUL HERE! For filename and contents we are referring directly to
-		.def( vector_indexing_suite< std::vector< std::string > >() );
+  // Python-allocated and -managed memory since we are accepting pointers to
+  // data members of python objects. We need to ensure that those objects
+  // outlive our UnsavedFile objects.
+	class_< UnsavedFile >( "UnsavedFile" )
+    .add_property( "filename_",
+      make_getter( &UnsavedFile::filename_ ),
+      make_setter( &UnsavedFile::filename_,
+                   return_value_policy< reference_existing_object >() ) )
+    .add_property( "contents_",
+      make_getter( &UnsavedFile::contents_ ),
+      make_setter( &UnsavedFile::contents_,
+                   return_value_policy< reference_existing_object >() ) )
+    .def_readwrite( "length_", &UnsavedFile::length_ );
+
+	class_< std::vector< UnsavedFile > >( "UnsavedFileVec" )
+		.def( vector_indexing_suite< std::vector< UnsavedFile > >() );
+
+  class_< ClangComplete, boost::noncopyable >( "ClangComplete" )
+    .def( "SetGlobalCompileFlags", &ClangComplete::SetGlobalCompileFlags )
+    .def( "SetFileCompileFlags", &ClangComplete::SetFileCompileFlags )
+    .def( "UpdateTranslationUnit", &ClangComplete::UpdateTranslationUnit )
+    .def( "CandidatesForLocationInFile",
+          &ClangComplete::CandidatesForLocationInFile );
 }

python/ycm.py

@@ -23,6 +23,7 @@ import indexer
 
 min_num_chars = int( vim.eval( "g:ycm_min_num_of_chars_for_completion" ) )
 
+clang_filetypes = set( [ 'c', 'cpp', 'objc', 'objcpp' ] )
 
 class CompletionSystem( object ):
   def __init__( self ):
@@ -87,6 +88,34 @@ class CompletionSystem( object ):
                                             filepath,
                                             True )
 
+class ClangComplete( object ):
+  def __init__( self ):
+    self.completer = indexer.ClangComplete()
+
+  def CandidatesForQuery( self, query ):
+    buffer = vim.current.buffer
+
+    # CAREFUL HERE! For UnsavedFile filename and contents we are referring
+    # directly to Python-allocated and -managed memory since we are accepting
+    # pointers to data members of python objects. We need to ensure that those
+    # objects outlive our UnsavedFile objects.
+    # We do this to avoid an extra copy of the entire file contents.
+    contents = '\n'.join( buffer )
+    unsaved_file = indexer.UnsavedFile()
+    unsaved_file.contents_ = contents
+    unsaved_file.length_ = len( contents )
+    unsaved_file.filename_ = buffer.name
+
+    files = indexer.UnsavedFileVec()
+    files.append( unsaved_file )
+
+    line = int( vim.eval( "line('.')" ) )
+    column = CompletionStartColumn() + 1
+    results = self.completer.CandidatesForLocationInFile( buffer.name,
+                                                          line,
+                                                          column,
+                                                          files )
+    return list( results )
 
 def CurrentColumn():
   """Do NOT access the CurrentColumn in vim.current.line. It doesn't exist yet.
@@ -103,23 +132,48 @@ def CurrentLineAndColumn():
   return line_num, column_num
 
 
+def ShouldUseClang( start_column ):
+  filetype = vim.eval( "&filetype" )
+  if filetype not in clang_filetypes:
+    return False
+
+  line = vim.current.line
+  previous_char_index = start_column - 1
+  if ( not len( line ) or
+       previous_char_index < 0 or
+       previous_char_index >= len( line ) ):
+    return False
+
+  if line[ previous_char_index ] == '.':
+    return True
+
+  if previous_char_index - 1 < 0:
+    return False
+
+  two_previous_chars = line[ previous_char_index - 1 : start_column ]
+  if ( two_previous_chars == '->' or two_previous_chars == '::' ):
+    return True
+
+  return False
+
+
 def IsIdentifierChar( char ):
   return char.isalnum() or char == '_'
 
 
 def CompletionStartColumn():
+  """Returns the 0-based index where the completion string should start. So if
+  the user enters:
+    foo.bar^
+  with the cursor being at the location of the caret, then the starting column
+  would be the index of the letter 'b'.
+  """
+
   line = vim.current.line
-  current_column = CurrentColumn()
+  start_column = CurrentColumn()
-  start_column = current_column
 
   while start_column > 0 and IsIdentifierChar( line[ start_column - 1 ] ):
     start_column -= 1
-
-  if current_column - start_column < min_num_chars:
-    # for vim, -2 means not found but don't trigger an error message
-    # see :h complete-functions
-    return -2
-
   return start_column
 
 
@@ -158,20 +212,6 @@ def PreviousIdentifier():
   return line[ start_column : end_column ]
 
 
-def CurrentCursorText():
-  start_column = CompletionStartColumn()
-  current_column = CurrentColumn()
-
-  if current_column - start_column < min_num_chars:
-    return ""
-
-  return vim.current.line[ start_column : current_column ]
-
-
-def CurrentCursorTextVim():
-  return EscapeForVim( CurrentCursorText() )
-
-
 def ShouldAddIdentifier():
   current_column = CurrentColumn()
   previous_char_index = current_column - 1