Support for Generic Ghc Api Ast Traversals

This Wiki page is a spin off of GhcApiStatus, focussing on information related to generic traversals (transformations/queries) over the Ghc Ast types.

Data.Traversable

David Waern has been working on deriving Data.Traversable for the Ast types, to get mapM for use in Haddock 2. He points out that, by default, SYB-style traversals do not seem to support type-changing maps.

Data.Generics

Since Ghc does support (standalone) deriving of Data and Typeable, it seems natural to want those classes instantiated for the Ghc Api types. In case that would lead to blowup of standard releases, it would be even better if those instances could be provided as an add-on to a release.

- until recently, standalone deriving of Data/Typeable was broken (#2378, fixed in HEAD, 07/01/08)

- standalone deriving of Typeable fails if data constructors are not in scope - this looks like a bug (#2433)

- standalone deriving of Data fails if data constructors are not in scope - this means that Data instances are in conflict with the intended data abstraction for some Ghc Ast types!

- using the internal CPP hackery of Data.Generics for Typeable, standalone deriving of Data for non-abstract types, and the traditional manual dummy Data instances for abstract types, it seems we can get initial Data.Generics support over the Ghc Ast types!

- since I want to be able to identify the abstract types, I've replaced the traditional toConstr _ = error "toConstr":

abstractConstr   n = mkConstr (abstractDataType n) ("{abstract:"++n++"}") [] Prefix
abstractDataType n = mkDataType n [abstractConstr n]

-- TypeableINSTANCE_TYPEABLE0(SrcSpan,srcSpanTc,"SrcSpan")
instance Data SrcSpan where
  toConstr _   = abstractConstr "SrcSpan"
  gunfold _ _  = error "gunfold"
  dataTypeOf _ = mkNorepType "SrcSpan"

TODO: Is that ok, or is it going to cause problems?

- putting all those instances in a single file seems to trigger a memory bug in the compiler (#2438), but splitting them into two files appears to avoid that. See the attached Instances and Instances0

- for starters, here's a Data-based show that shows the constructors/abstract types instead of pretty-printing them:

showData :: Data a => Int -> a -> String
showData n = generic `ext1Q` list `extQ` string 
                     `extQ` occName `extQ` moduleName `extQ` srcSpan
                     `extQ` postTcType `extQ` fixity
  where generic :: Data a => a -> String
        generic t = "\n" ++ replicate n ' ' 
                 ++ "(" ++ showConstr (toConstr t)
                 ++ space (concat (intersperse " " (gmapQ (showData (n+1)) t))) ++ ")"
        space "" = ""
        space s  = ' ':s
        string     = show :: String -> String
        list l     = "[" ++ concat (intersperse "," (map (showData (n+1)) l)) ++ "]"
        occName    = ("{OccName: "++) . (++"}") .  OccName.occNameString 
        srcSpan    = ("{"++) . (++"}") . showSDoc . ppr :: SrcSpan -> String
        moduleName = ("{"++) . (++"}") . showSDoc . ppr :: ModuleName -> String
        postTcType = const "{!type placeholder here?!}" :: PostTcType -> String
        fixity     = const "{!fixity placeholder here?!}" :: GHC.Fixity -> String

For example usage, see the attached APISybTesting: it parses a TestModule, prettyprints and shows, does an identity transform, and an example query (extract classes and family declarations).

- some of the predefined instances in Data.Generics.Instances are dummies, preventing traversals of substructures or pretending that non-concrete types are Data - this should probably be changed by splitting that module and not re-exporting the dummy part from Data.Generics by default. See  http://www.haskell.org/pipermail/generics/2008-June/000347.html,  http://www.haskell.org/pipermail/generics/2008-June/000346.html,  http://www.haskell.org/pipermail/generics/2008-June/000342.html for more info.

- it seems as if the issue of type-changing traversals can be addressed within SYB, although those dummy instances cause trouble here.

•  Traversable Functor Data,or: X marks the spot, a technique based on gfoldl plus one nasty and two non-nasty uses of unsafeCoerce. The nasty use interacts badly with the dummy instances.

•  gMap in SYB1, a technique based on gfoldl, gunfoldl, and some Dynamic types in the middle. Avoids unsafeCoerce, and uses the fact that the dummy instances are slightly more honest about gunfoldl than about gfoldl, but gMap cannot distinguish functor parameters from equivalent types (no fmap) and its type is too general for direct use (runtime type errors instead of compile time type errors)

•  http://www.haskell.org/pipermail/generics/2008-July/000351.html, combining the advantages of the previous two techniques to define a Data-based fmap. The remaining uses of unsafeCoerce only distinguish the functor parameter type from equivalent types occuring elsewhere in the functor, so seem to be ok?