[cfe-dev] Question about custom Clang static analyzer checker registration in clang-tidy without using plugin

[Artem Dergachev via cfe-dev]

I still don't understand your approach. There's the current plugin 
system that auto-registers checkers (the trick with "-cc1 -load" 
demonstrated by examples/analyzer-plugin/MainCallChecker.cpp) and 
there's the normal way of writing a checker in clang source tree (like 
all default checkers are written) that semi-automatically (through the 
registerChecker<>() call) puts the checker into the checker registry and 
makes it automatically accessible to clang-tidy when it links to the 
updated checkers library.

Because none of these approaches seem to have any problems with checker 
registration, could you describe what exactly are you doing that causes 
problems with checker registration?

On 13/03/2018 11:11 AM, via cfe-dev wrote:
> Hi George,
>
> Thanks for the fast response. I totally agree with your point of the disadvantages of using plugins. My question was exactly how to avoid using plugins. Maybe I had some misunderstanding, does the "plugin" you talked about not only include dynamic libraries, but also include registering checkers in programs that uses clang as a library?
>
> For the 2 solutions you propsed, they are good suggestions, but unfortunately they won't work in our specific situation. The checker I am adding is very company specific, and involves specific class in company's code base; I believe it is impossible to make it public, because it will reveal company's internal codebase; also it is useless to public because the situation it tries to analyze is very ad-hoc, company specific, thus won't apply to other people's code at all.
>
> As for rebasing against upstream, that's not feasible as well because this involves maintenance cost. I am at a large company where I don't have control or any influence on the process it sync third-party libraries with upstream, neither does the clang-tidy team of our company. That's why they suggest me doing it upstream.
>
> The 3 approaches in my original question was in fact about modifying upstream to expose the checker registration process, so that not only us, but other people could also benefit from it. If there is such an API, then in a program that uses clang as a library, they can call the exposed API to register their own checker, before using AnalysisConsumer to do a static analysis.
>
> Neither of the 3 approaches I proposed uses plugin architecture. It is just caller program who instantiates checker (or some form of checker factory), then hands it over to Clang, therefore there is no need to worry about how Clang can find the libraries.
>
> I am more then happy if there are other ideas and I can contribute my manpower to coming up with some patches.
>
> Best,
> Kan Li
>
> ----- Original Message -----
> From: George Karpenkov <ekarpenkov at apple.com>
> To: Li Kan <likan_999.student at sina.com>
> Cc: cfe-dev at lists.llvm.org
> Subject: Re: [cfe-dev] Question about custom Clang static analyzer checker registration in clang-tidy without using plugin
> Date: 2018-03-14 01:05
>
>
> Hi Li,
> I haven’t used plugins myself, but I can see two solutions side-stepping the problem.
> 1) Could you upstream your checker and put it into “alpha”, and then turn it on via flag?
> 2) If (1) is impossible [though upstreaming into “alpha” should not be too hard], you could fork clang and add your changes there,
> and then (hopefully) regularly rebase vs. trunk.
> The problem with the plugin infrastructure is that Clang does not guarantee a stable API.
> Even though things would probably work, there’s no guarantee they won’t break with a new release,
> and if you use plugins you would get all the associated disadvantages (no straightforward injection, have to care about library being found, etc)
> without the usual advantages (plugin might unexpectedly stop working).
> Though if you absolutely must use plugins, I think it would be possible to come up with an API for doing so.
> George
>> On Mar 12, 2018, at 6:20 PM, Li Kan via cfe-dev <cfe-dev at lists.llvm.org> wrote:
>>
>> Hi,
>>
>> I have a question about how to register custom checkers without using a plugin. The background is, I have been working on a checker and I was able to come up with a solution to integrate into our code review system by using AnalysisConsumer and load custom checkers using plugins. However, the team of our company who works on clang-tidy integration requests me to do that in clang-tidy, instead of a separate set of things.
>>
>> The way clang-tidy is used in the code review system is to create ClangTidyASTConsumerFactory and then create a ClangTidyASTConsumer, which is later used to parse the codebase. Therefore the goal here is to find a way to let the users of clang libraries, especially ClangTidyASTConsumer{,Factory}, to let AnalysisConsumer pickup custom checkers statically linked into the program;
>>
>> Currently the checker registration is very deep in the stack, the only places it looks for checkers are a set of builtin checkers and plugins. There are a few directions I can think of
>> 1) Add some arguments to AnalysisConsumer constructor and pass it down to createCheckerManager, and add logic there to look for checkers passed in; however, this is problematic because all the users of AnalysisConsumer, such as ClangTidyASTConsumer, needs to expose the arguments as well;
>> 2) Have some static variable that stores the list of checkers factories, and expose a function that allows user to add checkers factories to it. Later the ClangCheckerRegistry picks up the checkers in the list. This is the ordinary way, but the existence of a static variable looks ugly and not thread-safe.
>> 3) Like registerBuiltinCheckers, add a weak function registerCustomCheckers, initially empty, and client code is allowed to override it, filled in with their custom checkers. In ClangCheckerRegistry it invokes the function, just after it invokes the registerBuiltinCheckers. This is the least intrusive way but the drawback is if there are multiple places in the system that needs to add a checker, it is not very convenient.
>>
>> What's the best approach, or is there other better approach? I am more than happy to help and send out some patch to add some mechanism. I want to hear the opinions from Clang developers and get blessed by the direction, before I actually start on it, to avoid any arguments.
>>
>>
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev