Follow up:
Argh! I have just seen that I have a bunch of test failures
related to my MR (which, needless to say, it's still WIP).
For example:
run/T9140.run.stdout.normalised 2021-03-31 09:35:48.000000000 +0200
@@ -1,12 +1,4 @@
-<interactive>:2:5:
- You can't mix polymorphic and unlifted bindings: a = (# 1 #)
- Probable fix: add a type signature
-
-<interactive>:3:5:
- You can't mix polymorphic and unlifted bindings: a = (# 1, 3 #)
- Probable fix: add a type signature
-
So it looks like some diagnostic is now not being reported and,
surprise surprise, this was emitted from the DsM monad.
I have the suspect that indeed Richard was right (like he always
is :) ) -- when we go from a DsM to a TcM monad (See `initDsTc`)
for example, I think we also need to carry into the new monad all
the diagnostics we collected so far.
This implies indeed a mutual dependency (as Simon pointed out, heh).
So I think my cunning plan of embedding is crumbling -- I suspect
we would end up with a type `TcRnDsMessage` which captures the
dependency.
Sorry for not seeing it sooner!
On Wed, 31 Mar 2021 at 08:05, Alfredo Di Napoli
<alfredo.dinap...@gmail.com <mailto:alfredo.dinap...@gmail.com>>
wrote:
Morning all,
*Richard*: sorry! Unfortunately MR !4798 is the cornerstone
of this refactoring work but it's also gargantuan. Let's
discuss a plan to attack it, but fundamentally there is a
critical mass of changes that needs to happen atomically or
it wouldn't make much sense, and alas this doesn't play in
our favour when it comes to MR size and ease of review.
However, to quickly reply to your remak: currently (for the
sake of the "minimum-viable-product") I am trying to
stabilise the external interfaces, by which I mean giving
functions their final type signature while I do what's
easiest to make things typecheck. In this phase what I think
is the easiest is to wrap the majority of diagnostics into
the `xxUnknownxx` constructor, and change them gradually
later. A fair warning, though: you say "I would think that a
DsMessage would later be wrapped in an envelope." This might
be true for Ds messages (didn't actually invest any brain
cycles to check that) but in general we have to turn a
message into an envelope as soon as we have a chance to do
so, because we need to grab the `SrcSpan` and the `DynFlags`
*at the point of creation* of the diagnostics. Carrying
around a message and make it bubble up at some random point
won't be a good plan (even for Ds messages). Having said
that, I clearly have very little knowledge about this area of
GHC, so feel free to disagree :)
*John*: Although it's a bit hard to predict how well this is
going to evolve, my current embedding, to refresh everyone's
memory, is the following:
data DsMessage =
DsUnknownMessage !DiagnosticMessage
-- ^ Stop-gap constructor to ease the migration.
| DsLiftedTcRnMessage !TcRnMessage
-- ^ A diagnostic coming straight from the Typecheck-renamer.
-- More messages added in the future, of course
At first I thought this was the wrong way around, due to
Simon's comment, but this actually creates pleasant external
interfaces. To give you a bunch of examples from MR !4798:
deSugar :: HscEnv -> ModLocation -> TcGblEnv -> IO (Messages
DsMessage, Maybe ModGuts)
deSugarExpr :: HscEnv -> LHsExpr GhcTc -> IO (Messages
DsMessage, Maybe CoreExpr)
Note something interesting: the second function actually
calls `runTcInteractive` inside the body, but thanks to the
`DsLiftedTcRnMessage` we can still expose to the consumer an
opaque `DsMessage` , which is what I would expect to see from
a function called "deSugarExpr". Conversely, I would be
puzzled to find those functions returning a `TcRnDsMessage`.
Having said all of that, I am not advocating this design is
"the best". I am sure we will iterate on it. I am just
reporting that even this baseline seems to be decent from an
API perspective :)
On Wed, 31 Mar 2021 at 05:45, John Ericson
<john.ericson@obsidian.systems>
<mailto:john.ericson@obsidian.systems> wrote:
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/412>
/ https://github.com/ghc-proposals/ghc-proposals/pull/243
<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
<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://www.youtube.com/watch?v=nUvKoG_V_U0> /
https://github.com/gelisam/klister
<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 3/30/21 10:14 AM, Richard Eisenberg wrote:
On Mar 30, 2021, at 4:57 AM, Alfredo Di Napoli
<alfredo.dinap...@gmail.com
<mailto:alfredo.dinap...@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 <mailto:ghc-devs@haskell.org>
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
<http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs>
_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org <mailto:ghc-devs@haskell.org>
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
<http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs>
