The server is multi-threaded and will spawn a new thread for each new request.
Thus, the completers need not manage their own threads or even provide async
APIs; we _want_ them to block because now were implementing the request-response
networking API.
The client gets the async API through the network (i.e., it can do something
else while the request is pending).
Everything that needs access to filetype_map_ has been split into a new class
called IdentifierDatabase. This class is thread-safe. Previously, multiple
threads could access filetype_map_ at the same time, some reading things from
it, others writing to it. WTF was I doing? My best guess is that I introduced
this vile stupidity when I added the second thread to IdentifierCompleter;
previously it was impossible for multiple threads to stomp on filetype_map_
because only one thread ever accessed it. I changed that some time ago and
forgot to protect filetype_map_.
Idiot.
The answer was to use boost::move() on the functors, but since there is AFAIK no
move emulation support for Boost.Function, the fix may only work on C++11
compilers. That needs to be looked into.
This change should fix the random hangs and segfaults when using the clang
completer. Also, assertion errors printed to the console on vim exit should go
away too, same thing with segfaults on vim exit. These "on exit" errors were
caused by not cleanly shutting down the background threads; both the identifier
completer and the clang one now join the threads on destruction. This results in
a clean shutdown.
The new clang completer architecture now uses only one clang thread (again)
instead of a completion and parsing thread. Since the parsing task needs to wait
on the completion task if it was started first (and vice-versa) there's no point
to using two threads. The desired "simplicity" of using two threads for these
two tasks actually created needless complexity (and bugs). Sigh. Such is life.
A TranslationUnit abstraction was also created and this in turn also reduces the
complexity of the clang completer.
The clang completer now also has some (very) basic tests.
We limit the number of candidates returned to Vim to 20 and also make sure that
we are not returning any duplicate candidates. This provides a noticeable
improvement in latency.
First off, we don't block the GUI thread anymore for ClangCompleter (that was
always temporary). Secondly, now ClangCompleter will cache the data coming from
clang so that query-based filtering of members is fast.
This will make it easy to use the same Candidates for both the
IdentifierCompleter and the ClangCompleter, thereby reducing memory consumption
and increasing performance.
