Broken Data.Data instances

Richard Eisenberg eir at cis.upenn.edu
Fri Aug 15 15:27:20 UTC 2014


Simon,

I've been encouraging the type family approach. See https://phabricator.haskell.org/D157

Thanks,
Richard

On Aug 15, 2014, at 11:17 AM, Simon Peyton Jones <simonpj at microsoft.com> wrote:

> Eek.  Glancing at this I see that every single data type has an extra type parameter.  To me this feels like a sledgehammer to crack a nut.  What is wrong with the type-function approach?
>  
> Simon
>  
> From: Alan & Kim Zimmerman [mailto:alan.zimm at gmail.com] 
> Sent: 13 August 2014 07:50
> To: Philip K.F. Hölzenspies
> Cc: Simon Peyton Jones; Edward Kmett; ghc-devs at haskell.org
> Subject: Re: Broken Data.Data instances
>  
> And I dipped my toes into the phabricator water, and uploaded a diff to https://phabricator.haskell.org/D153
> 
> I left the lines long for now, so that it is clear that I simply added parameters to existing type signatures.
> 
>  
> 
> On Tue, Aug 12, 2014 at 10:51 PM, Alan & Kim Zimmerman <alan.zimm at gmail.com> wrote:
> 
> Status update
> 
> I have worked through a proof of concept update to the GHC AST whereby the type is provided as a parameter to each data type. This was basically a mechanical process of changing type signatures, and required very little actual code changes, being only to initialise the placeholder types.
> 
> The enabling types are
> 
> 
>     type PostTcType = Type        -- Used for slots in the abstract syntax
>                     -- where we want to keep slot for a type
>                     -- to be added by the type checker...but
>                     -- [before typechecking it's just bogus]
> 
>     type PreTcType = ()             -- used before typechecking
> 
> 
>     class PlaceHolderType a where
>       placeHolderType :: a
> 
>     instance PlaceHolderType PostTcType where
> 
> 
>       placeHolderType  = panic "Evaluated the place holder for a PostTcType"
> 
>     instance PlaceHolderType PreTcType where
>       placeHolderType = ()
> 
> These are used to replace all instances of PostTcType in the hsSyn types.
> 
> The change was applied against HEAD as of last friday, and can be found here
> 
> https://github.com/alanz/ghc/tree/wip/landmine-param
> https://github.com/alanz/haddock/tree/wip/landmine-param
> 
> They pass 'sh validate' with GHC 7.6.3, and compile against GHC 7.8.3. I have not tried to validate that yet, have no reason to expect failure.
> 
> 
> Can I please get some feedback as to whether this is a worthwhile change?
> 
> 
> It is the first step to getting a generic traversal safe AST
> 
> Regards
> 
>   Alan
> 
>  
> 
> On Mon, Jul 28, 2014 at 5:45 PM, Alan & Kim Zimmerman <alan.zimm at gmail.com> wrote:
> 
> FYI I edited the paste at http://lpaste.net/108262 to show the problem
> 
>  
> 
> On Mon, Jul 28, 2014 at 5:41 PM, Alan & Kim Zimmerman <alan.zimm at gmail.com> wrote:
> 
> I already tried that, the syntax does not seem to allow it.
> 
> I suspect some higher form of sorcery will be required, as alluded to herehttp://stackoverflow.com/questions/14133121/can-i-constrain-a-type-family
> 
> Alan
> 
>  
> 
> On Mon, Jul 28, 2014 at 4:55 PM, <p.k.f.holzenspies at utwente.nl> wrote:
> 
> Dear Alan,
>  
> I would think you would want to constrain the result, i.e.
>  
> type family (Data (PostTcType a)) => PostTcType a where …
>  
> The Data-instance of ‘a’ doesn’t give you much if you have a ‘PostTcType a’.
>  
> Your point about SYB-recognition of WrongPhase is, of course, a good one ;)
>  
> Regards,
> Philip
>  
>  
>  
> From: Alan & Kim Zimmerman [mailto:alan.zimm at gmail.com] 
> Sent: maandag 28 juli 2014 14:10
> To: Holzenspies, P.K.F. (EWI)
> Cc: Simon Peyton Jones; Edward Kmett; ghc-devs at haskell.org
> 
> Subject: Re: Broken Data.Data instances
>  
> Philip
> 
> I think the main reason for the WrongPhase thing is to have something that explicitly has a Data and Typeable instance, to allow generic (SYB) traversal. If we can get by without this so much the better.
> 
> On a related note, is there any way to constrain the 'a' in
> 
> type family PostTcType a where
>   PostTcType Id    = TcType
>   PostTcType other = WrongPhaseTyp
> 
> to have an instance of Data?
> 
> I am experimenting with traversals over my earlier paste, and got stuck here (which is the reason the Show instances were commentet out in the original).
> 
> Alan
>  
>  
> 
> On Mon, Jul 28, 2014 at 12:30 PM, <p.k.f.holzenspies at utwente.nl> wrote:
> Sorry about that… I’m having it out with my terminal server and the server seems to be winning. Here’s another go:
>  
> I always read the () as “there’s nothing meaningful to stick in here, but I have to stick in something” so I don’t necessarily want the WrongPhase-thing. There is very old commentary stating it would be lovely if someone could expose the PostTcType as a parameter of the AST-types, but that there are so many types and constructors, that it’s a boring chore to do. Actually, I was hoping haRe would come up to speed to be able to do this. That being said, I think Simon’s idea to turn PostTcType into a type-family is a better way altogether; it also documents intent, i.e. () may not say so much, but PostTcType RdrName says quite a lot.
>  
> Simon commented that a lot of the internal structures aren’t trees, but cyclic graphs, e.g. the TyCon for Maybe references the DataCons for Just and Nothing, which again refer to the TyCon for Maybe. I was wondering whether it would be possible to make stateful lenses for this. Of course, for specific cases, we could do this, but I wonder if it is also possible to have lenses remember the things they visited and not visit them twice. Any ideas on this, Edward?
>  
> Regards,
> Philip
>  
>  
>  
>  
>  
> From: Alan & Kim Zimmerman [mailto:alan.zimm at gmail.com]
> Sent: maandag 28 juli 2014 11:14
> To: Simon Peyton Jones
> Cc: Edward Kmett; Holzenspies, P.K.F. (EWI); ghc-devs
> 
> Subject: Re: Broken Data.Data instances
>  
> I have made a conceptual example of this here http://lpaste.net/108262
> 
> Alan
>  
> 
> On Mon, Jul 28, 2014 at 9:50 AM, Alan & Kim Zimmerman <alan.zimm at gmail.com> wrote:
> What about creating a specific type with a single constructor for the "not relevant to this phase" type to be used instead of () above? That would also clearly document what was going on.
> 
> Alan
>  
> 
> On Mon, Jul 28, 2014 at 9:14 AM, Simon Peyton Jones <simonpj at microsoft.com> wrote:
> I've had to mangle a bunch of hand-written Data instances and push out patches to a dozen packages that used to be built this way before I convinced the authors to switch to safer versions of Data. Using virtual smart constructors like we do now in containers and Text where needed can be used to preserve internal invariants, etc.
> 
>  
> If the “hand grenades” are the PostTcTypes, etc, then I can explain why they are there.  
>  
> There simply is no sensible type you can put before the type checker runs.  For example one of the constructors  in HsExpr is
>   | HsMultiIf   PostTcType [LGRHS id (LHsExpr id)]
> 
> After type checking we know what type the thing has, but before we have no clue.
>  
> We could get around this by saying
>             type PostTcType = Maybe TcType
> but that would mean that every post-typechecking consumer would need a redundant pattern-match on a Just that would always succeed.
>  
> It’s nothing deeper than that.  Adding Maybes everywhere would be possible, just clunky.
>  
>  
> However we now have type functions, and HsExpr is parameterised by an ‘id’ parameter, which changes from RdrName (after parsing) to Name (after renaming) to Id (after typechecking).  So we could do this:
>   | HsMultiIf   (PostTcType id) [LGRHS id (LHsExpr id)]
> 
> and define PostTcType as a closed type family thus
>  
>      type family PostTcType a where
> 
>           PostTcType Id = TcType
> 
>           PostTcType other = ()
> 
>  
> That would be better than filling it with bottoms.  But it might not help with generic programming, because there’d be a component whose type wasn’t fixed.  I have no idea how generics and type functions interact.
>  
> Simon
>  
> From: Edward Kmett [mailto:ekmett at gmail.com] 
> Sent: 27 July 2014 18:27
> To: p.k.f.holzenspies at utwente.nl
> Cc: alan.zimm at gmail.com; Simon Peyton Jones; ghc-devs
> 
> Subject: Re: Broken Data.Data instances
>  
> Philip, Alan, 
> 
>  
> 
> If you need a hand, I'm happy to pitch in guidance. 
> 
>  
> 
> I've had to mangle a bunch of hand-written Data instances and push out patches to a dozen packages that used to be built this way before I convinced the authors to switch to safer versions of Data. Using virtual smart constructors like we do now in containers and Text where needed can be used to preserve internal invariants, etc.
> 
>  
> 
> This works far better for users of the API than just randomly throwing them a live hand grenade. As I recall, these little grenades in generic programming over the GHC API have been a constant source of pain for libraries like haddock.
> 
>  
> 
> Simon,
> 
>  
> 
> It seems to me that regarding circular data structures, nothing prevents you from walking a circular data structure with Data.Data. You can generate a new one productively that looks just like the old with the contents swapped out, it is indistinguishable to an observer if the fixed point is lost, and a clever observer can use observable sharing to get it back, supposing that they are allowed to try.
> 
>  
> 
> Alternately, we could use the 'virtual constructor' trick there to break the cycle and reintroduce it, but I'm less enthusiastic about that idea, even if it is simpler in many ways.
> 
>  
> 
> -Edward
> 
>  
> 
> On Sun, Jul 27, 2014 at 10:17 AM, <p.k.f.holzenspies at utwente.nl> wrote:
> 
> Alan,
> 
> In that case, let's have a short feedback-loop between the two of us. It seems many of these files (Name.lhs, for example) are really stable through the repo-history. It would be nice to have one bigger refactoring all in one go (some of the code could use a polish, a lot of code seems removable).
> 
> Regards,
> Philip
> 
> Van: Alan & Kim Zimmerman [alan.zimm at gmail.com]
> Verzonden: vrijdag 25 juli 2014 13:44
> Aan: Simon Peyton Jones
> CC: Holzenspies, P.K.F. (EWI); ghc-devs at haskell.org
> Onderwerp: Re: Broken Data.Data instances
> 
> By the way, I would be happy to attempt this task, if the concept is viable.
> 
>  
> 
> On Thu, Jul 24, 2014 at 11:23 PM, Alan & Kim Zimmerman <alan.zimm at gmail.com> wrote:
> 
> While we are talking about fixing traversals, how about getting rid of the phase specific panic initialisers for placeHolderType, placeHolderKind and friends?
> 
> In order to safely traverse with SYB, the following needs to be inserted into all the SYB schemes (see 
> https://github.com/alanz/HaRe/blob/master/src/Language/Haskell/Refact/Utils/GhcUtils.hs)
> 
> -- Check the Typeable items
> checkItemStage1 :: (Typeable a) => SYB.Stage -> a -> Bool
> checkItemStage1 stage x = (const False `SYB.extQ` postTcType `SYB.extQ` fixity `SYB.extQ` nameSet) x
>   where nameSet     = const (stage `elem` [SYB.Parser,SYB.TypeChecker]) :: GHC.NameSet       -> Bool
>         postTcType  = const (stage < SYB.TypeChecker                  ) :: GHC.PostTcType    -> Bool
>         fixity      = const (stage < SYB.Renamer                      ) :: GHC.Fixity        -> Bool
> 
> And in addition HsCmdTop and ParStmtBlock are initialised with explicit 'undefined values.
> 
> Perhaps use an initialiser that can have its panic turned off when called via the GHC API?
> 
> Regards
> 
>   Alan
> 
>  
> 
>  
> 
> On Thu, Jul 24, 2014 at 11:06 PM, Simon Peyton Jones <simonpj at microsoft.com> wrote:
> 
> So... does anyone object to me changing these "broken" instances with the ones given by DeriveDataTypeable?
> 
> That’s fine with me provided (a) the default behaviour is not immediate divergence (which it might well be), and (b) the pitfalls are documented.
>  
> Simon
>  
> From: "Philip K.F. Hölzenspies" [mailto:p.k.f.holzenspies at utwente.nl] 
> Sent: 24 July 2014 18:42
> To: Simon Peyton Jones
> Cc: ghc-devs at haskell.org
> Subject: Re: Broken Data.Data instances
>  
> Dear Simon, et al,
> 
> These are very good points to make for people writing such traversals and queries. I would be more than happy to write a page on the pitfalls etc. on the wiki, but in my experience so far, exploring the innards of GHC is tremendously helped by trying small things out and showing (bits of) the intermediate structures. For me, personally, this has always been hindered by the absence of good instances of Data and/or Show (not having to bring DynFlags and not just visualising with the pretty printer are very helpful).
> 
> So... does anyone object to me changing these "broken" instances with the ones given by DeriveDataTypeable?
> 
> Also, many of these internal data structures could be provided with useful lenses to improve such traversals further. Anyone ever go at that? Would be people be interested?
> 
> Regards,
> Philip
> 
> <image001.jpg>
> Simon Peyton Jones
> 24 Jul 2014 18:22
> GHC’s data structures are often mutually recursive. e.g.
> ·        The TyCon for Maybe contains the DataCon for Just
> 
> ·        The DataCon For just contains Just’s type
> 
> ·        Just’s type contains the TyCon for Maybe
> 
>  
> So any attempt to recursively walk over all these structures, as you would a tree, will fail. 
>  
> Also there’s a lot of sharing.  For example, every occurrence of ‘map’ is a Var, and inside that Var is map’s type, its strictness, its rewrite RULE, etc etc.  In walking over a term you may not want to walk over all that stuff at every occurrence of map.
>  
> Maybe that’s it; I’m not certain since I did not write the Data instances for any of GHC’s types
>  
> Simon
>  
> From: ghc-devs [mailto:ghc-devs-bounces at haskell.org] On Behalf Ofp.k.f.holzenspies at utwente.nl
> Sent: 24 July 2014 16:42
> To: ghc-devs at haskell.org
> Subject: Broken Data.Data instances
>  
> Dear GHC-ers,
>  
> Is there a reason for explicitly broken Data.Data instances? Case in point:
>  
> > instance Data Var where
> >   -- don't traverse?
> >   toConstr _   = abstractConstr "Var"
> >   gunfold _ _  = error "gunfold"
> >   dataTypeOf _ = mkNoRepType "Var"
>  
> I understand (vaguely) arguments about abstract data types, but this also excludes convenient queries that can, e.g. extract all types from a CoreExpr. I had hoped to do stuff like this:
>  
> > collect :: (Typeable b, Data a, MonadPlus m) => a -> m b
> > collect = everything mplus $ mkQ mzero return
> > 
> > allTypes :: CoreExpr -> [Type]
> > allTypes = collect
>  
> Especially when still exploring (parts of) the GHC API, being able to extract things in this fashion is very helpful. SYB’s “everything” being broken by these instances, not so much.
>  
> Would a patch “fixing” these instances be acceptable?
>  
> Regards,
> Philip
>  
>  
>  
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://www.haskell.org/mailman/listinfo/ghc-devs
> 
>  
>  
> 
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://www.haskell.org/mailman/listinfo/ghc-devs
> 
>  
>  
>  
>  
>  
>  
>  
>  
> _______________________________________________
> ghc-devs mailing list
> 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/20140815/aaefeafd/attachment-0001.html>


More information about the ghc-devs mailing list