[cfe-dev] Using clang-query on the whole project

[Peter Collingbourne]

On Thu, Jun 11, 2015 at 11:50:40AM +0000, Manuel Klimek wrote:
> On Thu, Jun 11, 2015 at 1:29 PM Jacob Carlborg <doob at me.com> wrote:
> 
> > On 2015-06-11 11:24, Kevin Funk wrote:
> > > Heya,
> > >
> > > I've found clang-query to be useful to query single files. Matching the
> > Clang
> > > AST is easy enough and fast to learn. I don't see how to use clang-query
> > on a
> > > whole project, though, which would be super useful.
> > >
> > > A simple example where it's effectively useless: Assume some header
> > "global.h"
> > > containing "struct Global {};", every file in the project includes this
> > header.
> > >
> > > Now the idiomatic way to run clang-query over all the files in the
> > project:
> > >    clang-query -p $BUILD $(find . -name "*.cpp")
> > >
> > > Let's try to find record decls named "Global":
> > >    match recordDecl(hasName("Global"), isDefinition())
> > >
> > > => You get one match per file / translation unit. But in fact, this is
> > not
> > > really what you want. You want one match here.
> > >
> > > So of course this is difficult: clang-query is TU-centric, while for a
> > whole
> > > project you'd "somehow" want a global view over the source code and not
> > have
> > > duplicate results for a query like above.
> > >
> > > Questions:
> > > - Is it possible to get clang-query to behave like that?
> > >    - Any research/pointers in that regard?
> > > - Is it possible to filter duplicates in the results without external
> > tools?
> >
> > In theory it should be easy to modify clang-query to check the location
> > of a declaration and only return declarations for unique locations.
> >
> 
> Well, the problem is that a clang-query process is started for each TU, so
> you'd need to somehow tell clang-query which results were in a completely
> different process of clang-query. I don't think this is going to be easy or
> worth it...

This isn't entirely accurate. Each TU is stored in a separate AST, but all
of the ASTs live in a single process. When a match query is run, we simply
iterate over a vector of TU ASTs (see MatchQuery::run in Query.cpp). If you
wanted to deduplicate results, you could probably do it in MatchQuery::run
by pretty printing each result's SourceLocation and using that as a string key.

Thanks,
-- 
Peter