[cfe-dev] Making FrontendActions composable (or not)

[Jordan Rose]

On Nov 18, 2013, at 0:17 , Manuel Klimek <klimek at google.com> wrote:

> On Sun, Nov 17, 2013 at 11:48 AM, Sean Silva <silvas at purdue.edu> wrote:
> It seems weird to be using FrontendAction for clang-tidy (and the analyzer too). clang-tidy isn't a compiler frontend (I mean look at the methods: hasCodeCompletionSupport? getCurrentFileKind? shouldEraseOutputFiles?). AFAICT All you are really need is a vector of PPCallbacks, and a vector of what is effectively std::function<TidyResult(ASTContext &)>. It seems like there should be at most a single FrontendAction which just sets things up so that it can forward everything down into those vectors.
> 
> That is basically what I proposed as "b". The problem with that is that this is not how the static analyzer is currently factored - if I want to pull it apart into a PPCallbacks and an ASTConsumer, I have to duplicate significant amounts of code. Thus, if we want to go that way, the next step is to make the static analyzer more modular: basically pull out a class that can give you both the PPCallbacks and the ASTConsumer. But on the other hand, such a class already exists - it is the static analyzer's FrontendAction; it is exactly what we want, namely a "factory" class (well, in the broader sense) of a PPCallbacks and ASTConsumer instance, which happens to also be able to put the compiler into the state it needs it in.
> 
> Perhaps when you argue on the interface level, you mean "there should really be something in between a FrontendAction and an ASTConsumer or PPCallbacks interface", and perhaps you're right, but that sounds like a much bigger change...
> 
> Cheers,
> /Manuel

It’s been a while since I’ve looked at the frontend-ish parts of the analyzer, but AFAICT it’s not actually using PPCallbacks in any way. I don’t think there’s any problem with jumping directly to the AnalysisConsumer.

That said, if we ever want to add preprocessing-time checks, they wouldn’t be included in clang-tidy without a separate mechanism, but I think I’m okay with that. The current model is certainly intended to behave as an idempotent ASTConsumer.

Jordan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20131118/798e39ec/attachment.html>