goldfirere/th-desugar

Desugars Template Haskell abstract syntax to a simpler format without changing semantics

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

Desugaring depends on language extensions of use sites

Suppose you quote some Template Haskell declarations in module A:

{-# LANGUAGE ... #-}
module A where

decs :: Q [Dec]
decs = [d| ... |]

And later desugar the declarations with th-desugar in module B:

{-# LANGUAGE ... #-}
module B where

import A (decs)
import Language.Haskell.TH.Desugar (dsDecs)

$(do desugaredDecs <- dsDecs decs
     ...)

There are some situations where th-desugar's desugaring depends on which language extensions are enabled, such as:

• MonadFailDesugaring (for desugaring partial pattern matches in do notation)
• NoFieldSelectors (for determining if a record field can be reified as a field selector with lookupValueNameWithLocals)

Somewhat counterintuitively, th-desugar will consult the language extensions in module B (the site where the decs are used) for this process, not module A (where the decs were defined). This is really a Template Haskell limitation, since Template Haskell does not offer any way to reify which language extensions were enabled at the time the declarations were defined. As a result, th-desugar can only check for language extensions at use sites.

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:

1. Desugared Haskell98-style constructors
2. Locally reified class methods
3. Locally reified record selectors
4. Locally reified data constructors
5. 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.

Limitations of support for desugaring guards

th-desugar supports guards in the sense that it will desugar guards to equivalent code that instead uses case expressions. For example, this code:

f (x, y)
  | x == "hello" = x
  | otherwise = y

Will be desugared to this code:

f arg =
  case arg of
    (x, y) ->
      case x2 == "hello" of
        True  -> x
        False -> y

This has the advantage that it saves users from needing to care about the complexities of guards. It does have some drawbacks, however, which we describe below.

Desugaring guards can result in quadratic code size

If you desugar this program involving guards:

data T = A Int | B Int | C Int

f :: T -> T -> Maybe Int
f (A x1) (A x2)
  | x1 == x2
  = Just x1
f (B x1) (B x2)
  | x1 == x2
  = Just x1
f (C x1) (C x2)
  | x1 == x2
  = Just x1
f _ _ = Nothing

You will end up with:

f :: T -> T -> Maybe Int
f arg1 arg2 =
  case (# arg1, arg2 #) of
    (# A x1, A x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          case (# arg1, arg2 #) of
            (# B y1, B y2 #) ->
              case y1 == y2 of
                True ->
                  Just y1
                False ->
                  case (# arg1, arg2 #) of
                    (# C z1, C z2 #) ->
                      case z1 == z2 of
                        True ->
                          Just z1
                        False ->
                          case (# arg1, arg2 #) of
                            (# _, _ #) ->
                              Nothing
                    (# _, _ #) ->
                      Nothing
            (# C y1, C y2 #) ->
              case y1 == y2 of
                True ->
                  Just y1
                False ->
                  case (# arg1, arg2 #) of
                    (# _, _ #) ->
                      Nothing
            (# _, _ #) ->
              Nothing
    (# B x1, B x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          case (# arg1, arg2 #) of
            (# C y1, C y2 #) ->
              case y1 == y2 of
                True ->
                  Just y1
                False ->
                  case (# arg1, arg2 #) of
                    (# _, _ #) ->
                      Nothing
            (# _, _ #) ->
              Nothing
    (# C x1, C x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          case (# arg1, arg2 #) of
            (# _, _ #) ->
              Nothing
    (# _, _ #) ->
      Nothing

That is signficantly more code. In the worst case, the algorithm that th-desugar uses for desugaring guards can lead to a quadratic increase in code size. One way to avoid this is avoid having incomplete guards that fall through to later clauses. That is, if you rewrite the original code to this:

f :: T -> T -> Maybe Int
f (A x1) (A x2)
  | x1 == x2
  = Just x1
  | otherwise
  = Nothing
f (B x1) (B x2)
  | x1 == x2
  = Just x1
  | otherwise
  = Nothing
f (C x1) (C x2)
  | x1 == x2
  = Just x1
  | otherwise
  = Nothing

Then th-desugar will desugar it to:

f :: T -> T -> Maybe Int
f arg1 arg2 =
  case (# arg1, arg2 #) of
    (# A x1, A x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          Nothing
    (# B x1, B x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          Nothing
    (# C x1, C x2 #) ->
      case x1 == x2 of
        True ->
          Just x1
        False ->
          Nothing

This code, while still more verbose than the original, uses a constant amount of extra code per clause.

Desugaring guards can produce more warnings than the original code

The approach that th-desugar uses to desugar guards can result in code that produces GHC compiler warnings (if -fenable-th-splice-warnings is enabled) where the original code does not. For example, consider the example from above:

data T = A Int | B Int | C Int

f :: T -> T -> Maybe Int
f (A x1) (A x2)
  | x1 == x2
  = Just x1
f (B x1) (B x2)
  | x1 == x2
  = Just x1
f (C x1) (C x2)
  | x1 == x2
  = Just x1
f _ _ = Nothing

This code compiles without any GHC warnings. If you desugar this code using th-desugar, however, it will produce these warnings:

warning: [-Woverlapping-patterns]
    Pattern match is redundant
    In a case alternative: (# B y1, B y2 #) -> ...
   |
   |             (# B y1, B y2 #) ->
   |             ^^^^^^^^^^^^^^^^^^^...

warning: [-Woverlapping-patterns]
    Pattern match is redundant
    In a case alternative: (# C y1, C y2 #) -> ...
   |
   |             (# C y1, C y2 #) ->
   |             ^^^^^^^^^^^^^^^^^^^...

warning: [-Woverlapping-patterns]
    Pattern match is redundant
    In a case alternative: (# C y1, C y2 #) -> ...
   |
   |             (# C y1, C y2 #) ->
   |             ^^^^^^^^^^^^^^^^^^^...

GHC is correct here: these matches are wholly redundant. th-desugar could potentially recognize this and perform a more sophisticated analysis to detect and remove such matches when desugaring guards, but it currently doesn't do such an analysis.

No support for view patterns

th-desugar does not support desugaring view patterns. An alternative to using view patterns in the patterns of a function is to use pattern guards. Currently, there is not a viable workaround for using view patterns in pattern synonym definitions—see this th-desugar issue.

No support for ApplicativeDo

th-desugar does not take the ApplicativeDo extension into account when desugaring do notation. For example, if you desugar this:

{-# LANGUAGE ApplicativeDo #-}

f x y = do
  x' <- x
  y' <- y
  return (x' ++ y')

Then th-desugar will translate the uses of <- in the do block to uses of Monad operations (e.g., (>>=)) rather than Applicative operations (e.g., (<*>)). See this th-desugar issue.

No support for RecursiveDo

th-desugar does not support the RecursiveDo extension at all, so it cannot desugar any uses of mdo expressions or rec statements.

No support for unresolved infix operators

th-desugar does not support desugaring unresolved infix operators, such as UInfixE. You are unlikely to encounter this limitation when dealing with Template Haskell quotes, since quoted infix operators will translate to uses of InfixE rather than UInfixE. Rather, this limitation would only be encountered if you manually construct a Template Haskell Exp using UInfixE.