|Version 7 (modified by chak, 9 years ago) (diff)|
Type Functions: Renaming
GHC is organised such that class and type declarations are processed (during renaming and type checking) before any instance declarations are considered. In the presence of associated types, instance declarations may contain type definitions. In particular, the data constructors introduced by associated data declarations need to be brought into scope before we can rename any expressions. Otherwise, the intution wrt. to phasing is that kind signatures are handled in conjunction with vanilla algebraic data type declarations and instances of indexed types are handled together with instance heads. (NB: GHC splits the treatment of instances into two phases, where instances heads are processed in the first and member function declarations in the second phase.)
Renaming of indexed types
Kind signatures are renamed by RnSource.rnTySig, which is parametrised by a function that handles the binders (i.e., index variables) of the declaration. This is so that we can use the same code for toplevel signatures and those in classes. In the former case, the variables are in a defining position, whereas in classes they are in a usage position (as all index variables must be class parameters).
Before the actual renaming starts, GHC builds an environment of all names bound by a declaration group. These names go into the global RdrName environment and are determined by the function RnNames.getLocalDeclBinders. We need to be careful in handling data and new type instances in this context, as all the data constructors they define need to go into this environment.
Type instances of associated types
To get hold of the data constructors of associated types, RnNames.getLocalDeclBinders now also traverses all instance declarations and descends into the associated type instances.
Family names in type instances
Family names in the heads of type instances are usage occurences, not binders. However, we cannot just use RnNames.lookupGlobalOccRn as we are only now determining the binders to construct the global RdrName environment. On the other hand, we cannot use RnNames.newTopSrcBinder either, although it produces the same name when called multiple times. It does not handle the case, where we define an instance for an imported family. Hence, we introduced a new function RnEnv.lookupFamInstDeclBndr that first attempts a lookup, similar to RnEnv.lookupGlobalOccRn, but if that fails, instead of raising an error, calls newTopSrcBinder. If after all, there is no family declaration in scope, this will be picked up and properly reported during renaming by RnSource.rnTyClDecl.
Despite the effort we go to, to get the right Name for the family RdrName, getLocalDeclBinders does not return that name as part of the binders declared by the current binding group - otherwise, we would get a duplicate declaration error. However, we use the Name to specify the correct name parent for the data constructors of the instance (see bwlow).
Renaming of type instances
There is little extra that needs to be done for type instances. The main difference between vanilla synonyms and data/newtype declarations and the indexed variants is that the tcdTyPats field is not Nothing. We simply call rnTyPats on these fields, which traverses them in the usual way. Moreover, we need to be careful with the family type constructor in the instance head: it is in an occurence position and needs to be looked up suitably, such that we get a nice "out of scope" error if no appropriate family is in scope. By the same token, the family names needs to go into the set of free variables. Finally, RnSource.rnSrcInstDecl invokes RnSource.rnTyClDecl on all associated types of an instance to rename them.
Name parents & importing and exporting
GHC, in RnNames, derives its knowledge of which names may appear in parenthesis after a type or class name in an import or export list by querying the name parent relation, as encoded in RdrName.gre_par of RdrName.GlobalRdrElt. Hence, it is crucial that all the data constructors defined in instances of a family get the family name, not the name of the representation tycon, as their name parent. We go to some effort in RnNames.getLocalDeclBinders to achieve this when renaming source. We also , when sucking declarations from interface files in LoadIface.loadDecl, make the family the name parent for all implicit things of the declaration.
An extra complication with the name parent relation that arises due to associated types is that the name parent relation can have a depth of two (not just one as in H98). Here is an example:
class GMapKey a where data GMap a :: * -> * instance GMapKey Int where data GMap Int v = GMapInt ...
The data constructor GMapInt's parent is GMap whose parent in turn is the class GMapKey; ie, GMapKey is GMapInt's grand parent. In H98, data types have no parents (which is in some places in the code represented by making them their own parent).
We address this in RnNames.filterImports by extending the occurences environment occ_env by an extra, but optional parent component. In the case of associated types, this new component gives the class in which the associated types has been declared.
Finally, to ensure that family instances put into interface files and associated type declarations are not duplicated on the top-level of an interface file, we need to take care in TyCon.isImplicitTyCon to guarantee that (a) associated family declarations are implicit and (b) family instance representation tycons are not.