th-desugar
Package
This package provides the Language.Haskell.TH.Desugar
module, which desugars
Template Haskell's rich encoding of Haskell syntax into a simpler encoding.
This desugaring discards surface syntax information (such as the use of infix
operators) but retains the original meaning of the TH code. The intended use
of this package is as a preprocessor for more advanced code manipulation
tools. Note that the input to any of the ds...
functions should be produced
from a TH quote, using the syntax [| ... |]
. If the input to these functions
is a hand-coded TH syntax tree, the results may be unpredictable. In
particular, it is likely that promoted datatypes will not work as expected.
One explicit goal of this package is to reduce the burden of supporting multiple GHC / TH versions. Thus, the desugared language is the same across all GHC versions, and any inconsistencies are handled internally.
The package was designed for use with the singletons
package, so some design
decisions are based on that use case, when more than one design choice was
possible.
I will try to keep this package up-to-date with respect to changes in GHC. The minimum supported version of GHC is 8.0, which was chosen to avoid various Template Haskell bugs in older GHC versions that affect how this library desugars code. If this choice negatively impacts you, please submit a bug report.
Known limitations
Limited support for kind inference
th-desugar
sometimes has to construct types for certain Haskell entities.
For instance, th-desugar
desugars all Haskell98-style constructors to use
GADT syntax, so the following:
data T (a :: k) = MkT (Proxy a)
Will be desugared to something like this:
data T (a :: k) where
MkT :: forall k (a :: k). Proxy a -> T (a :: k)
Notice that k
is explicitly quantified in the type of MkT
. This is due to
an additional pass that th-desugar
performs over the type variable binders
of T
to extract all implicitly quantified variables and make them explicit.
This makes the desugared types forwards-compatible with a
future version of GHC
that requires all kind variables in a top-level forall
to be explicitly
quantified.
This process of extracting all implicitly quantified kind variables is not
perfect, however. There are some obscure programs that will cause th-desugar
to produce type variable binders that are ill scoped. Here is one example:
data P k (a :: k)
data Foo (a :: Proxy j) (b :: k) c = MkFoo c (P k j)
If you squint hard at MkFoo
, you'll notice that j :: k
. However, this
relationship is not expressed syntactically, which means that th-desugar
will not be aware of it. Therefore, th-desugar
will desugar Foo
to:
data Foo (a :: Proxy j) (b :: k) c where
MkFoo :: forall j k (a :: Proxy j) (b :: k) c.
c -> P k j -> Foo (a :: Proxy j) (b :: k) c
This is incorrect since k
must come before j
in order to be well scoped.
There is a workaround to this issue, however: add more explicit kind
information. If you had instead written this:
data Foo (a :: Proxy (j :: k)) (b :: k) c = MkFoo c (P k j)
Then the fact that j :: k
is expressed directly in the AST, so th-desugar
is able to pick up on it and pick forall k j (a :: Proxy j) (b :: k) c. <...>
as the telescope for the type of MkFoo
.
The following constructs are known to be susceptible to this issue:
- Desugared Haskell98-style constructors
- Locally reified class methods
- Locally reified record selectors
- Locally reified data constructors
- Locally reified type family instances (on GHC 8.8 and later, in which the
Template Haskell AST supports explicit
foralls
in type family equations)
Limited support for linear types
Currently, the th-desugar
AST deliberately makes it impossible to represent
linear types, and desugaring a linear function arrow will simply turn into a
normal function arrow (->)
. This choice is partly motivated by issues in the
way that linear types interact with Template Haskell, which sometimes make it
impossible to tell whether a reified function type is linear or not. See, for
instance, GHC#18378.