[ghc-steering-committee] #287: Simplify subsumption , recommendation: accept

[Eric Seidel]

I support acceptance as well. Changing the semantics of the program as part of type-checking is just bizarre and it seems like the impact in terms of required code changes is quite limited.

Given how easy the fixes appear to be, and the fact that the fixed program will compile just fine with older GHCs, I'm inclined to support option (2): ripping the code out immediately. I've left a comment on the PR asking if we can do so while providing helpful fixit hints to unaware users.

On Fri, Jan 17, 2020, at 08:49, Spiwack, Arnaud wrote:
> Dear all,
> 
> As the shepherd for proposal #287 [ 
> https://github.com/ghc-proposals/ghc-proposals/pull/287 ], I recommend 
> acceptance
> 
> Summary
> 
> The proposal is to remove a bunch of typing rules meant to make 
> RankNTypes a bit more powerful (covariance and contravariance of the 
> arrow constructor, as well as deep instantiation and deep 
> skolemisation). The problem with these rules is that 1/ they complicate 
> the type checker 2/ They are implemented by performing eta-expansion 
> during elaboration, which changes the semantics of program 3/ Dropping 
> (some of) them is beneficial for the quick-look story (proposal 
> currently on hold).
> 
> The authors have tried this change on all of Stackage, and could come 
> up with simple fixes (~300 lines of which) for all but 2 packages 
> (namely, singleton, because of template haskell stuff, and labels for 
> reason not well identified).
> 
> Rationale
> 
> The common point of all these features is that they jiggle the position 
> of foralls around. Which sounds reasonable at first glance. But really, 
> they tend to interact badly with other stuff (I don't remember the 
> circumstances, but there was a discussion of deep instantiation and 
> deep skolemisation in the context of linear types as well). Basically, 
> they are rather fragile features, I'd rather we did without them.
> 
> I think that 300 lines of simple fixes for all of Stackage is a very 
> small number. So it is legitimate to consider this change to have low 
> impact. Regarding the singleton, I'm sure Richard will be happy to fix 
> it ;-) . Only labels remains. It shouldn't be a blocker.
> 
> Remaining question
> 
> There is one open question on the pull request thread. Namely, how to 
> schedule the removal. Two paths are being discussed
>  1. Add a `-XLessSubsumption` flag, turn it off by default, then turn 
> it on by default in version n+1, then remove the subsumption rules 
> altogether in version n+3
>  2. Simply remove the code, preferably early in the release cycle, 
> communicate profusely.
> The thread participants seem to lean towards (2). The reason being that 
> this is a change where it is easy to write code compatible with 
> previous versions of the compiler (everything which compiles after this 
> change also compiles before this change). Also, there is a preference 
> for removing the relevant code sooner rather than later. But I'd like 
> the opinion of the committee.
> 
> Best,
> Arnaud
> _______________________________________________
> ghc-steering-committee mailing list
> ghc-steering-committee at haskell.org
> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>