Hi, On 2017-04-02 at 01:23:52 +0200, David Feuer wrote:
The fact that GHC needs to have its own containers shouldn't prevent a consistent package set from using a different one. The only hypothetical problem would be if the GHC API leaked containers types, but I'm pretty sure it doesn't: they're all wrapped up in newtypes and exported abstractly, unless someone's made a mistake.
Fwiw, I just opened up the GHC api haddocks and arbitrarily picked a module and sure enough I was able to find a use of an exposed container type rightaway, specificially `IntMap`: https://downloads.haskell.org/~ghc/latest/docs/html/libraries/ghc-8.0.1/HscT... I'd expect to easily find more leaked types from containers and other dependencies. Relatedly, I don't think it's a good idea to implement support for internal/private dependencies at this point yet w/o first having the tools to verify that types that are supposed to stay hidden don't leak through the exposed API. Otoh, tools for doing this kind of static analysis are also needed to verify the existing API versioning contract we already use now. And once we have such go-to tooling, it should be easy to build on top of that. However, if we wanted to decouple `ghc` from `containers` (and possibly other packages, such as process/transformers/unix/bytestring/binary) by declaring the use of its `container` types private/internal to `ghc`'s API, we could just go back to do what we did years ago (e.g. in GHC 7.0.4) by using renamed packages like `ghc-binary` (but also revisit the reasons why we don't do this anymore...). Another approach would be to make the `ghc` package recompilable by `cabal`. I briefly tried that[1] but got side-tracked, and I still believe it should be possible to get the `ghc` package to the point where it's an ordinary `cabal`-buildable package. [1]: Not the least, so I could upload it to Hackage and have its Haddocks being generated without too much magic; but more importantly so that cabal would be less constrained when solving for install-plans when `ghc` is involved, as well as GHC HQ being able to more easily pick up new dependencies for `ghc` w/o having to introduce yet another congestion point in terms of dependency constraints (i.e. packages depended upon by `ghc` are effectively frozen to those picked by a given GHC release once `ghc` enters an install-plan).