Version 1 (modified by cactus, 6 months ago) (diff)


Frontend representation

After parsing, and during renaming, pattern synonyms are stored as TyClDecls.


The typechecking pass turns PatSynDecls into a PatSyn and several HsBinds. To fill in the PatSyn, we typecheck the right-hand side of the pattern synonym declaration, then do some extra processing on it to reject as-patterns and optionally compute the reverse of the pattern synonym (for implicitly bidirectional ones). Afterwards, we collect universal & existential type variables and typeclass dictionary variables to be used when creating ConPatOut patterns from pattern synonym occurances, and generate some HsBinds:

  • The PatSyn stores typing information for the pattern synonym, to be consulted when typechecking pattern synonym usage sites.
  • The first HsBind is the binder for the matcher function generated from the pattern synonym. The matcher is used when desugaring pattern synonym usage sites (see below).
  • For bidirectional pattern synonyms, another HsBind called a wrapper is created to be used for pattern synonym usages in expression contexts. It is a wrapper in the same sense as a constructor wrapper.

Pattern synonym occurances in patterns are turned into ConPatOuts just like regular constructor matches. ConPatOut has been changed to store a ConLike instead of a DataCon; the ConLike type is simply the sum of DataCon and PatSyn.


ConLikes are handled uniformly all the way until mkCoAlgCaseMatchResult. There, we have a mixed list of DataCon and PatSyn-based patterns.


This list is grouped so that subsequent DataCon patterns are put in the same group, and PatSyn patterns are all in their own groups. This is needed so that when doing pattern matching per column, given e.g.

data T = MkT1 | MkT2 Bool | MkT3
pattern P x = MkT2 x

and a list of cases

MkT1 _ -> alt1
P True -> alt2
MkT2 _ -> alt3

we don't compile that into

DEFAULT -> ... P ...
MkT1 -> alt1
MkT2 _ -> alt3

since we can't see into P (and we don't want to, since it might be imported from another module). The correct thing to do is to compile that into

DEFAULT -> ... (P, alt2) and (MkT2 _, alt3) ...
MkT1 -> alt1

Consecutive occurances of the same pattern synonym (e.g. if we had `P True and P False` in the previous example) are compiled into a single match; the arguments are then matched in a sub-case.


For each pattern synonym, a matcher function is generated which gets a scrutinee and a success and a failure continuations. Given a type

data T a where
     MkT :: (Cls b) => b -> a -> T a

and a pattern synonym

pattern P x y = MkT x y

we generate the matcher function

P :: forall r a. T a -> (forall b. Cls b => b -> a -> r) -> r -> r
P scrutinee pass fail = case scrutinee of
  MkT x y -> pass x y
  _ -> fail

Occurances of pattern synonyms are then desugared into calls to this matcher function. This allows pattern synonym definitions to be just as opaque as function definitions: their type defines their interface completely. This gives us a story for exporting pattern synonym definitions that is entirely consistent with existing function definition exports.