JSON output of GHC Diagnostics

[Ben Gamari]

Ben Bellick via ghc-devs <ghc-devs at haskell.org> writes:

> Hey everyone,
>
> I am working on a summer of Haskell project to implement JSON output of GHC's diagnostics. I have begun working through some of the implementation details but have stumbled across a few road blocks and I wanted to get some feedback on. Discussion surrounding the plan can be found [here](https://github.com/haskellfoundation/tech-proposals/pull/50).
>
> In looking through the current implementation of the -ddump-json​ flag, I found a [note](https://gitlab.haskell.org/ghc/ghc/-/blob/master/compiler/GHC/Utils/Logger.hs#L392) which seems to indicate that a sensible (albeit invasive) approach is to alter the definition of [LogAction](https://gitlab.haskell.org/ghc/ghc/-/blob/master/compiler/GHC/Utils/Logger.hs#L174), reproduced below:
>
...
>
> So instead of taking in an SDoc​, I could alter this signature to take
> a set of Diagnostic​s, though this may introduce other complications.
> In the first place, it seems log action is processing all of the
> diagnostics one at a time, though this particular JSON output should
> process all of them at once in order to place them all in a single
> JSON object. So, I am inclined to believe that this approach may not
> work.
>
Taking all of the diagnostics at once is possible (and, with the
addition of internal state, could even be done without changing the
LogAction type). However, it would mean that downstream users wouldn't
benefit from incremental reporting of errors.

I don't honestly know whether how much value such incrementality has to
the typical consumers you have in mind, but I think it could be easily
preserved with something like the JSON Lines approach that Bryan
suggests.

Admittedly, this would be a change in the semantics of -ddump-json but I
don't consider this to be a problem. Frankly, `-d` flags are intended
primarily for debugging and are generally not considered to be stable
interfaces. In fact, if we are going to improve the JSON output, I think
we should probably rename the flag to something in the `-f` namespace
and deprecate the old `-d` flag.

Cheers,

- Ben