small improvement to roles mechanism

Simon Peyton-Jones simonpj at microsoft.com
Sat Oct 12 12:00:11 UTC 2013


I think perhaps we are going a bit overboard here!  Some thoughts:


*         Coercible is an experimental feature.  If you don't use it, it won't hurt you.  I don't think there is anything wrong with having signposted experimental features in a released compiler. And there is some merit: we will get useful feedback about the design from people who go ahead and use it anyway.  (We did this for type functions a couple of releases ago.)
I don't really understand what the suggestion "disallow Coercible" below means.

*         The open design issue for GND and for Coercible is the same.  It is simply this: if s is coercible to t, is (T s) coercible to (T t)?  By "coercible" I mean representationally-equivalent, ~R.  For GND we ask this question about the types of class methods; in Coercible, it arises when simplifying Coercible constraints.  But it is the same question.

*         The current proposal, which I am happy with, is to express the answer to this question in T's role annotation.  There is design wiggle-room here:

o   If there is no user-supplied annotation, what should the default be?  The inferred (lease-restrictive) role, or the Nominal (most restrictive) role?

o   How expressive is the role annotation language, especially where type constructors are concerned?
So the design is not fully stable here, but I think the principle, that T's role ascription answers the question, is sound.

*         The current, fairly inexpressive role annotation language makes three Hackage packages (a tiny number) fail to compile because their use of GND is rejected.  They have fairly few dependencies, so in all 18 packages break.  Each of the three can be fixed with very modest changes. This does not sound like a lot of pain to me.

That said, if you think it best, I'd be perfectly happy to make GND failure into a warning for one release cycle.  (I wouldn't do that myself, but I'm quite willing to be guided on this one.)



In short:

*         Coercible is opt-in anyway

*         If you prefer, GND failure can be a warning
What's not to like?  Am I missing anything?
Simon



From: ghc-devs [mailto:ghc-devs-bounces at haskell.org] On Behalf Of Richard Eisenberg
Sent: 12 October 2013 03:20
To: Edward Kmett
Cc: Stephanie Weirich; Joachim Breitner; ghc-devs at haskell.org Devs
Subject: Re: small improvement to roles mechanism

I tend to agree with the above remarks and have been silencing the little voice in the back of my head that has suggested we pull roles out of 7.8. Ever since the debate about default roles vis-a-vis abstract datatypes started, I've been wondering if we're being a little hasty here.

As we think about how to proceed, I think it's worth teasing apart the two different strains in all of this: enforcing type safety and enforcing abstraction.

Roles, as they are, do a quite good job of enforcing type safety. When compiling all of Hackage, the majority of the breakage from roles was due to real type goofs in programs. A total of 18 packages (out of the 3,234 that compile with 7.6.3) fail to compile because of deficiencies in the role system, including transitive breakage (i.e., packages that depend on broken packages). A small change in the GND check discussed in this forum would fix 11 of these, leaving only 7. Given that type systems are necessarily conservative, I'm not too displeased with this result -- the 3 packages that would need to be changed would require only a few lines of code. Given this result, I disagree with Edward's claim that roles will provide the biggest pain in upgrading from 7.6 to 7.8.

That said, I of course would like to do better and would prefer to 0 broken packages due to deficiencies in the role system.

On the other hand, roles as they are do a poor job of enforcing abstraction. But, before roles came along, there wasn't anything resembling a way of enforcing abstraction in the context of GND. So, this isn't exactly a new problem. The only thing (I think) exacerbating the problem is that Coercible now makes it easier than ever to break abstraction.

That said, I of course would like to fix this problem, too.

So, how to proceed? There are a few options:

1. I'll change the GND check to be a little more liberal, and we release roles in 7.8. I could notify the authors of the three packages that need to be updated because of the lack of role abstraction.

2. Like #1, but disallow Coercible. This way, the abstraction problem is no worse than it was before. (Apologies to Joachim if he minds this suggestion.)

3. Pull roles out of 7.8, giving us a little more time to Get It Right.

It's a little hard for me to choose between these options, but I think it's good to be conservative in language design of a language in real use, so I lean slightly toward option #3. Doing this is very feasible from a technical standpoint -- it's easy to turn off the checks.

If we go with #3, given the flux in the design, I'm even uncertain about the idea of warnings in 7.8. It would all feel a little silly if we issue these warnings and then change the design between 7.8 and 7.10 just enough to make the warnings wrong. I think warnings in 7.8.2 is a better idea.

As for Johan's remark that the feature should have been vetted more thoroughly -- I completely agree, and I'll take responsibility for that decision. I did do some ad-hoc testing against packages on Hackage known to use GND, but the testing was not as thorough as it could have been. That said, I'm not sure anything would have played out too differently had I done more extensive testing; the results were about as I expected.

I do greatly appreciate everyone's feedback and interest in this!

Richard

On Oct 11, 2013, at 7:55 PM, Edward Kmett wrote:


I have to agree that I'm somewhat disturbed by the fact that we're pushing this out and we're still finding issues with it this close to release. =(

It strikes me that the role machinery is going to be the cause of the majority of the pain users have upgrading to 7.8, and if I try to pretend to be Mark Lentczner for a bit, it makes it seem highly likely that it'd be the kind of thing that keeps 7.8 from going into a Haskell Platform, causing the groundhog to see his shadow, leaving us with another year of 7.4 or 7.6.x.

I know we're at the 5 yard line, but to metaphorically throw a bunch of metaphors in a blender, if we had to make the uncomfortable decision to perform triage and ask if it should be put off (at least enforcing) the roles machinery to 7.10, so we can know we have it right, how much fallout would there be? Off the top of my head, of course Joachim's work on Coercible would be affected. What else?

One option might be to pull the teeth of role inference for 7.8 with regards to GND, and turn bad roles use into a warning for a release cycle. That would give the community a year to get role annotations in place before generalized newtype deriving for their code just stopped working.

If we ship with this the way it stands, I don't foresee the community reaction being good.

With its teeth pulled, then GND could proceed as before, but with the added detailed warnings from the dictionary coercions helping to guide folks to make the change. By the time we'd be enforcing correct role annotations most folks would have them in place to silence the warnings.

-Edward

On Fri, Oct 11, 2013 at 7:15 PM, Johan Tibell <johan.tibell at gmail.com<mailto:johan.tibell at gmail.com>> wrote:
Oh and let me add: it would have been nice to have the people actually making these change to have done an impact analysis on Hackage, instead of discovering potential issues a week or two before the release. Lets try to do that next time.

On Fri, Oct 11, 2013 at 4:14 PM, Johan Tibell <johan.tibell at gmail.com<mailto:johan.tibell at gmail.com>> wrote:
Let me start by saying that I'm happy we're trying to fix the GND problem. Thanks for working on that.

That being said: is this ready for mainstream consumption? We're forcing this on everyone without any language pragma or flags to opt-in/out. That is bad if we're not sure we're doing the right thing in some cases or if we're causing spurious failures. At ICFP I got the impression that very few people will be affected, but Bryan's result suggests there are more people than I thought.


On Thu, Oct 10, 2013 at 8:26 PM, Richard Eisenberg <eir at cis.upenn.edu<mailto:eir at cis.upenn.edu>> wrote:
In Bryan's recent test of GHC 7.8 against all of Hackage, there were three spurious errors caused by lack of role abstraction. Here are the class definitions where a nominal parameter is inferred, probably against the wishes of the author:

from logict-0.2.3/Control.Monad.Logic.Class:
> class (MonadPlus m) => MonadLogic m where
>     msplit     :: m a -> m (Maybe (a, m a))

from monadLib-3.5.2/MonadLib:
> class (Monad m) => ReaderM m i | m -> i where
>   ask :: m i

from base/Control.Arrow:
> class Arrow a => ArrowApply a where
>     app :: a (a b c, b) c

In each of these, the last parameter of the class is given a nominal role because it appears as the parameter of a type variable. However, in each case, it appears as the parameter of a *class* type variable. This means that, if we somehow knew that the class author wanted the class to be usable with GND, we could simply check every instance declaration for that class to make sure that the relevant concrete instantiation has the right role. For example, when the user writes, for example

> instance ArrowApply Foo where ...

we check that Foo's first parameter has a representational role. If it doesn't, then the instance is rejected.

An alternative, somewhat heavier idea would be to represent roles as class constraints. We could have

> class NextParamNominal (c :: k)
> class NextParamRepresentational (c :: k)

GHC could "generate" instances for every datatype definition. For example:

> type role Map nominal representational
> data Map k v = ...

would induce

> instance NextParamNominal Map
> instance NextParamRepresentational (Map k)

Users would not be able to write these instances -- they would have to be generated by GHC. (Alternatively, there could be no instances, just a little magic in the constraint solver. Somewhat like Coercible.)

Then, the classes above would just have to add a superclass, like this:

> class (Arrow a, NextParamRepresentational a) => ArrowApply a where
>   app :: a (a b c, b) c

The role inference mechanism would be made aware of role constraints and use this one to derive that ArrowApply is OK for GND.

This "heavier" approach has a similar upshot to the first idea of just checking at instance declarations, but it is more customizable and transparent to users (I think).


I'm not sure I'm advocating for this change (or volunteering to implement before the release candidate), but I wanted to document the idea and get any feedback that is out there. This would fix the breakage we've seen without totally changing the kind system.

Thanks,
Richard

PS: Due credit is to migmit for suggesting the type-class idea on glasgow-haskell-users.
_______________________________________________
ghc-devs mailing list
ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
http://www.haskell.org/mailman/listinfo/ghc-devs



_______________________________________________
ghc-devs mailing list
ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
http://www.haskell.org/mailman/listinfo/ghc-devs


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/ghc-devs/attachments/20131012/53c03fbf/attachment.html>


More information about the ghc-devs mailing list