DataKinds and package boundaries: "the 'impossible' happened"

[aristidb]

I don't know what is going on there, so I can't really give much more information, sorry.

aristid@sigmoid:~/Code/aws/Examples$ ghc -package-conf ../dist/package.conf.inplace GetObject.hs 
[1 of 1] Compiling Main             ( GetObject.hs, GetObject.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 7.4.2 for x86_64-unknown-linux):
	tyThingTyCon
    Data constructor `aws-0.7.0:Aws.Core.NormalQuery{d r2qt}'

Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

[aristidb]

I forgot to enclose the error in a code block...

aristid@sigmoid:~/Code/aws/Examples$ ghc -package-conf ../dist/package.conf.inplace GetObject.hs 
[1 of 1] Compiling Main             ( GetObject.hs, GetObject.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 7.4.2 for x86_64-unknown-linux):
	tyThingTyCon
    Data constructor `aws-0.7.0:Aws.Core.NormalQuery{d r2qt}'

Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

[aristidb]

So, while I still don't know what is going on, I think this might be related to usage of DataKinds?. I have tracked this down by trying to go back to a commit in my repository where this behavior does not happen, and it's just before I introduced some very simple code that uses DataKinds?. It might be the package boundary that is confusing GHC there?

Please tell me if there is other information that I can attach to help you. Maybe I can cook up a minimal test, but this is fairly hard when you don't know what is going on!

[aristidb]

So I succeeded at creating a minimal test example. It _is_ related to DataKinds?, and it's remarkable how little code is needed to trigger it. I'm going to attach it as a tarfile, and as a small Github repository for easier browsing:  https://github.com/aristidb/data-kinds-haskell-bug

[aristidb]

Minimal test example

[aristidb]

Note that there's a script trigger.sh that should trigger the bug. It does basically the same ghc -package-conf invocation as in the original unminimised example, and yields this output for me:

aristid@sigmoid:~/Code/data-kinds-haskell-bug$ ./trigger.sh 
Resolving dependencies...
Configuring data-kinds-haskell-bug-0.1...
Preprocessing library data-kinds-haskell-bug-0.1...
Building data-kinds-haskell-bug-0.1...
[1 of 1] Compiling IsThisTheBug     ( IsThisTheBug.hs, dist/build/IsThisTheBug.o )
Registering data-kinds-haskell-bug-0.1...
[1 of 1] Compiling Main             ( Importer.hs, Importer.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 7.4.2 for x86_64-unknown-linux):
	tyThingTyCon
    Data constructor `data-kinds-haskell-bug-0.1:IsThisTheBug.KindConstructor1{d r2Q}'

Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

(The fact that Trac doesn't allow editing comments makes it hard for me, because I always think of new stuff after I've already made the comment.)

[simonpj]

Thanks. DataKinds is not an advertised feature of 7.4.2. I've tested your example with HEAD (which will become 7.6) and it's fine. Pending the release you can build from source, or grab a snapshot distribution.

(I won't add a regression test because it's a complicated one, and there are so many possible causes.)

Simon

[aristidb]

OK, I guess I need to implement the "ugly" solution to my problem that doesn't require DataKinds?. For an "unadvertised" feature, word spread pretty quickly, perhaps you should have include a few big red warning texts to tell people that it's considered highly experimental in 7.4.

I should note by the way that no problem was caused as long as the DataKinds? code was loaded as source modules, not as package modules. If that helps pin down a cause for a possible regression text.