[cfe-dev] [RFC] Adding lifetime analysis to clang

[Gábor Horváth via cfe-dev]

Dmitri Gribenko <gribozavr at gmail.com> ezt írta (időpont: 2019. ápr. 15., H
16:02):

> On Mon, Apr 15, 2019 at 2:26 PM Gábor Horváth <xazax.hun at gmail.com> wrote:
> >
> > Thank you for summarizing the discussion and raising all these points.
> > Based on your response I propose the following plan moving forward:
> >
> > * Start submitting patches regarding the non-controversial parts of the
> analysis, including:
> >   - Adding the annotations to Clang
> >   - Generalize existing statement-local warnings
> >     - Technical question here: should we expect the STL vendors to
> annotate the types, should we hard-code the annotations into the compiler,
> or should we actually implement type inference but restrict it to STL types?
> >   - Adding some extra statement-local warnings
> >   - Adding the flow sensitive analysis
> >
> > Can we add you as a reviewer to those patches?
>
> Happy to help :)
>
> > To start the discussions regarding the type category inference, as far
> as I understand you are more worried about the false negatives that can
> give a bad impression about the analysis rather than the false positives.
>
> I would say my concerns are around the understandability of the
> system, more so than the false negatives of inference.
>
> > One idea we had is to have a refactoring that will automatically
> annotate the user defined types based on the inference we have. This could
> give the users a way to understand the root causes of such false negatives.
> > In case the community will not want automatic type category inference in
> the compiler we are still likely to implement it as a Tidy check with a
> refactoring to make it easier for users to adopt this analysis.
>
> That'd be helpful, however, relying purely on such analysis is not
> much better than inference.  Users should annotate types that conform
> to type categories.  We should also figure out a way to learn if it is
> common for users to have types that are in spirit conforming to type
> categories, but don't conform syntactically (e.g., have named methods
> instead of overloaded operators).  If those are important, we would
> need to build corresponding annotations.
>

Great! So let's move on with upstreaming the non-controversial parts and
return to discussing the inference problem once we have more experience
from users and other data.

Thanks,
Gábor


> Dmitri
>
> --
> main(i,j){for(i=2;;i++){for(j=2;j<i;j++){if(!(i%j)){j=0;break;}}if
> (j){printf("%d\n",i);}}} /*Dmitri Gribenko <gribozavr at gmail.com>*/
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20190415/800957d2/attachment.html>