fix: add local imported packages to the packages to analyze when using cache

[ldez]

To reproduce the problem:

git clone git@github.com:gabolaev/gl-issue-reproduction.git && cd $_
golangci-lint cache clean
git checkout HEAD .; golangci-lint run; sed -i 's/some text/SOME OTHER TEXT/g' somepackage/somefile.go; golangci-lint run
# git checkout HEAD .; ./golangci-lint run; sed -i 's/some text/SOME OTHER TEXT/g' somepackage/somefile.go; ./golangci-lint run

If it works you should see:

somepackage/somefile.go:9:2: printf: (*github.com/sirupsen/logrus.Logger).Warnf format %v reads arg #1, but call has 0 args (govet)
        log.Warnf("some text %v") // this line is supposed to be reported by govet printf linter
        ^
somepackage/somefile.go:9:2: printf: (*github.com/sirupsen/logrus.Logger).Warnf format %v reads arg #1, but call has 0 args (govet)
        log.Warnf("SOME OTHER TEXT %v") // this line is supposed to be reported by govet printf linter
        ^

but if it not works:

somepackage/somefile.go:9:2: printf: (*github.com/sirupsen/logrus.Logger).Warnf format %v reads arg #1, but call has 0 args (govet)
        log.Warnf("some text %v") // this line is supposed to be reported by govet printf linter
        ^

This will impact negatively the performance of a run with a cache but it will increase the stability of reports.

Fixes #4423

[bombsimon]

Did you do any tests (manual or otherwise) on a somewhat larger codebase and/or with a lot of local imports? Are we talking millisecond increases or can this become a cardinality bomb and grow much bigger than that?

The fix looks good and is understandable so I'll approve this since I have full trust in your judgement here 😄

[ldez]

Did you do any tests (manual or otherwise) on a somewhat larger codebase and/or with a lot of local imports?

I didn't do a benchmark because this depends on how many direct packages a package depends on, so it's difficult to create accurate benchmarks.

Are we talking millisecond increases or can this become a cardinality bomb and grow much bigger than that?

As it just adds the direct packages related to one analyzed package and not the indirect packages, the cardinality is limited.

But if a package depends on thousands of local packages (I hope nobody does that), the impact will be important on performances (only for the run after the first one) but I don't have a better solution.

EDIT: I have some doubts maybe the real problem is the linter 🤔

[gabolaev]

It works, thank you!

[alexandear]

LGTM