Alfredo also replied to this pointing his embedding plan. I also prefer that, because I really wish TH didn't smear together the phases so much. Moreover, I hope with

 - GHC proposals https://github.com/ghc-proposals/ghc-proposals/pull/412 / https://github.com/ghc-proposals/ghc-proposals/pull/243

 - The parallelism work currently be planned in https://gitlab.haskell.org/ghc/ghc/-/wikis/Plan-for-increased-parallelism-and-more-detailed-intermediate-output

we might actually have an opportunity/extra motivation to do that. Splices and quotes will still induce intricate inter-phase dependencies, but I hope that could be mediated by the driver rather than just baked into each phase.

(One final step would be the "stuck macros" technique of https://www.youtube.com/watch?v=nUvKoG_V_U0 / https://github.com/gelisam/klister, where TH splices would be able to making "blocking queries" of the the compiler in ways that induce more of these fine-grained dependencies.)

Anyways, while we could also do a "RnTsDsError" and split later, I hope Alfredo's alternative of embedding won't be too much harder and prepare us for these exciting areas of exploration.

John

On Mar 30, 2021, at 4:57 AM, Alfredo Di Napoli <alfredo.dinapoli@gmail.com> wrote:

I'll explore the idea of adding a second IORef.

Renaming/type-checking is already mutually recursive. (The renamer must call the type-checker in order to rename -- that is, evaluate -- untyped splices. I actually can't recall why the type-checker needs to call the renamer.) So we will have a TcRnError. Now we see that the desugarer ends up mixed in, too. We could proceed how Alfredo suggests, by adding a second IORef. Or we could just make TcRnDsError (maybe renaming that).

What's the disadvantage? Clients will have to potentially know about all the different error forms with either approach (that is, using my combined type or using multiple IORefs). The big advantage to separating is maybe module dependencies? But my guess is that the dependencies won't be an issue here, due to the fact that these components are already leaning on each other. Maybe the advantage is just in having smaller types? Maybe.

I don't have a great sense as to what to do here, but I would want a clear reason that e.g. the TcRn monad would have two IORefs, while other monads will work with GhcMessage (instead of a whole bunch of IORefs).

Richard

_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs