Version 65 (modified by simonpj, 8 years ago) (diff)


DataParallel/ClosureConversion Up?

Closure conversion without classes

The following scheme approaches the problem of mixing converted and unconverted code from the point of view of GHC's Core representation, avoiding the use of classes as much as possible. In particular, the scheme gracefully handles any declarations that themselves cannot be converted, but occur in a converted module. The two essential ideas are that (1) we move between converted and unconverted values/code using a conversion isomorphism and (2) we treat unconverted declarations differently depending on whether or not they involve arrows; e.g., the definition of Int by way of unboxed values (which we cannot convert) doesn't prevent us from using Ints as is in converted code.

The closure type

We represent closures by

data a :-> b = forall e. !(e -> a -> b) :$ e

and define closure creation and application as

lam :: (a -> b) -> (a :-> b)
lam f = const f :$ ()

($:) :: (a :-> b) -> a -> b
(f :$ e) $: x = f e x

So, we have (->)_CC == (:->).

Overview and invariants

The meta-function "^", written postfix, converts from normal code to closure-converted code.

  • For each function f::ty, create a closure-converted function f_cc::ty^, where ty^ is the closure-converted version of ty.
  • For each data type T, create a closure-converted data type T_CC, whose constructors use (:->) instead of (->).
  • The value iso<ty> is a pair of functions, converting to and fro between ty and ty^.


  e :: ty     implies       e^ :: ty^

  to iso<ty> :: ty -> ty^
  fr iso<ty> :: ty^ -> ty

Conversion status

All TyCons, DataCons, and Ids have a conversion status that determines how occurences of these entities are treated during conversion. For an Id named v, we have two alternatives:

  1. The binding of v was compiled without conversion and we have to use v itself in converted code, which requires the use of an in-place conversion function.
  2. Otherwise, we have a converted variant v_CC, and we use v_CC instead of v in converted code.

For a type constructor T and its data constructors C, we have three alternatives:

  1. The declaration introducing T and its constructors was compiled without conversion or we were unable to convert it, as it uses some language feature that prevents conversion.
  2. A converted variant T_CC exists, but coincides with T (e.g., because T neither directly nor indirectly involves arrows).
  3. A converted variant T_CC exists and differs from T.

In the last two cases, we also have a conversion constructor isoT whose type and meaning is described below.

An example of a feature that prevents conversion are unboxed values. We cannot make a closure from a function that has an unboxed argument, as we can neither instantiate the parametric polymorphic closure type with unboxed types, nor can we put unboxed values into the existentially quantified environment of a closure.

Converting types

Conversion of type terms

We determine the converted type t^ of t as follows:

T^            = T_CC , if T_CC exists
              = T    , otherwise
a^            = a_CC
(t1 -> t2)^   = t1 -> t2   , if kindOf t1 == #
                             or kindOf t2 == #
              = t1^ :-> t2^, otherwise
(t1 t2)^      = t1^ t2^
(forall a.t)^ = forall a_CC.t^

Here some examples,

(Int -> Int)^           = Int :-> Int
(forall a. [a] -> [a])^ = [a] :-> [a]
([Int -> Int] -> Int)^  = [Int :-> Int] :-> Int
(Int# -> Int# -> Int#)^ = Int# -> Int# -> Int#
((Int -> Int) -> Int#)^ = (Int -> Int) -> Int#
(Int -> Int -> Int#)^   = Int :-> (Int -> Int#)

Why do we use (t1 -> t2)^ = t1 -> t2 when either argument type is unboxed, instead of producing t1^ -> t2^? Because we want to avoid creating conversion constructors (see below) for such types. After all, the conversion constructor isoArr for function arrows works only for arrows of kind *->*->*.

Conversion constructors

To move between t and t^ we use conversion functions. And to deal with type constructors, we need conversion constructors; i.e., functions that map conversion functions for type arguments to conversion functions for compound types.

Conversion pairs

Conversion functions come in pairs, which we wrap with the following data type for convenience:

data a :<->: b = (:<->:) {to :: a -> b, fr ::b -> a}

The functions witness the isomorphism between the two representations, as usual.

Types of convercion constructors

The type of a conversion constructor depends on the kind of the converted type constructor:

isoTy (t::k1->k2) = forall a a_CC.
                      isoTy (a::k1) -> isoTy (t a::k2)
isoTy (t::*)      = t :<->: t^

where type conversion t^ is defined below.

As an example, consider

data T (f::*->*) = T1 (f Int) | T2 (f Bool)

The type of the conversion constructor is as follows :

isoTy (T::(*->*)->*) =
  forall f f_CC. 
    (forall a a_CC. 
       (a :<->: a_CC) -> (f a :<->: f_CC a_CC)) ->
    T f :<->: T_CC f_CC

The conversion constructor might be implemented as

isoT isof = toT :<->: frT
    toT (T1 x) = T1 (to (isof isoInt ) x)
    toT (T2 y) = T2 (to (isof isoBool) y)
    frT (T1 x) = T1 (fr (isof isoInt ) x)
    frT (T2 y) = T2 (fr (isof isoBool) y)

where isoInt and isoBool are the conversion constructors for Ints and Bools.

Moreover, the conversion constructor for function arrows is

isoArr :: a :<->: a_CC   -- argument conversion
       -> b :<->: b_CC   -- result conversion
       -> (a -> b) :<->: (a_CC :-> b_CC)
isoArr (toa :<->: fra) (tob :<->: frb) = toArr :<->: frArr
    toArr f        = const (tob . f . fra) :$ ()
    frArr (f :$ e) = frb . f e . toa



To perform the actual conversion of values of a type t::*, we generate a conversion iso<t> of type t :<->: t^ as follows:

iso<T>          = isoT           , if T_CC exists
                = idIso<*>       , otherwise
iso<a::k>       = idIso<k>
iso<t1 -> t2>   = idIso<*>       , if kindOf t1 == #
                                   or kindOf t2 == #
                = isoArr         , otherwise 
                    iso<t1> iso<t2>
iso<t1 t2>      = iso<t1> iso<t2>
iso<forall a.t> = toIso :<->: frIso
    toIso (x::forall a.t) = /\a. to (iso<t>)@a x@a
    frIso (x::forall a.t) = /\a. fr (iso<t>)@a x@a


idIso<*>      = id :<->: id

OPEN PROBLEM: What should idIso<k> do for k other than *? We might think

idIso<k1->k2> = \_ -> (idIso<k2>)

would work, but it doesn't always. Take this example

data T a = MkT a
unit :: T a -> a
unit (MkT x) = x

-- converted
foo :: f (Int -> Int) -> (forall a. f a -> a) -> Int
foo t u = u t 1

-- not converted
bar = foo unit

Here, we will have to convert the first argument to foo and that conversion needs to convert the embedded Int -> Int to Int :-> Int, which is hard to do in foo = fr iso<...> foo_CC as we don't know anything about f::*->*.


Here some example conversions:

iso<Int -> Int>     = isoArr isoInt isoInt
iso<Int -> Int#>    = id :<->: id
iso<[a -> a]>       = isoList (isoArr (id :<->: id) 
                                      (id :<->: id))
iso<f (Int -> Int)> = ???

Converting type declarations

Conversion rules

If a type declaration for constructor T occurs in a converted module, we need to decide whether to convert the declaration of T. We decide this as follows:

  1. If the declaration of T mentions another algebraic type constructor S for which there is no S_CC, then we cannot convert T.
  2. If all algebraic type constructors S mentioned in T's definiton have a conversion S_CC == S, we do not convert T, but set T_CC == T and generate a suitable conversion constructor isoT. (NB: The condition implies that T does not mention any function arrows.)
  3. If the declaration of T uses any features that we cannot (or for the moment, don't want to) convert, simply don't convert it.
  4. Otherwise, we generate a converted type declaration T_CC together with a conversion constructor isoT. Conversion proceeds by converting all data constructors (see below).

Moreover, we handle other forms of type constructors as follows:

  • FunTyCon: We have (->)_CC = (:->).
  • TupleTyCon: We have (,..,)_CC = (,..,). We may either have a (long) list of conversion constructors iso(,..,) pre-defined or need to generate them inline by generating a suitable case expression where needed.
  • SynTyCon: Closure conversion operates on coreView; hence, we will see no synonyms. (Well, we may see synonym families, but will treat them as not convertible for the moment.)
  • PrimTyCon: We essentially ignore primitive types during conversion, assuming that their converted and unconverted forms coincide. As they cannot contain values of other types, we need no conversion constructor.
  • CoercionTyCon and SuperKindTyCon: They don't categorise values and are ignored during conversion.

Conversion constructor

Whenever we have a converted type constructor T_CC, we also need to generate a conversion constructor isoT. If T has one or more arguments, the conversion is non-trivial, even for T_CC == T.

Converting data constructors

We convert a data constructor C :: t1 -> ... -> tn by generating a converted constructor C_CC :: t1^ -> .. -> tn^. This includes the generation of a corresponding new worker Id. For example, if the original worker has the type signature

MkT :: (Int -> Int) -> Int

the converted worker is

MkT_CC :: (Int :-> Int) -> Int

As a consequence, whenever we convert a partial worker application in an expression, we need to introduce a closure on the spot. (Simon pointed out that this is a rare case anyway.)

We do not specially handle wrappers of data constructors or field selectors. They are converted just like any other toplevel function.


For example, when we convert

data Int = I# Int#

we get Int_CC = Int and we have

isoInt :: Int :<->: Int
isoInt = toInt :<->: frInt
    toInt (I# i#) = I# i#
    frInt (I# i#) = I# i#

As another example,

data Maybe a = Nothing | Just a

implies Maybe_CC = Maybe and

isoMaybe :: (a :<->: a_CC) -> (Maybe a :<->: Maybe a_CC)
isoMaybe isoa = toMaybe :<->: frMaybe
    toMaybe isoa Nothing  = Nothing
    toMaybe isoa (Just x) = Just (to isoa x)
    frMaybe isoa Nothing  = Nothing
    frMaybe isoa (Just x) = Just (fr isoa x)

Converting classes and instances

We don't alter class and instance declarations in any way. However, the dictionary type constructors and dfuns are converted in the same way as other data types and value bindings, respectively.

As an example, assume Num Int were defined as

class Num a where
  (+)    :: a -> a -> a
  negate :: a -> a
instance Num Int where
  (+)    = primAddInt
  negate = primNegateInt

with the Core code being

data Num a = 
  Num {
    (+)    :: a -> a -> a,
    negate :: a -> a
dNumInt = Num Int
dNumInt = Num primAddInt primNegateInt

Then, closure conversion gives us

data Num_CC a = Num_CC (a :-> a :-> a) (a :-> a)
(+_CC)    :: Num_CC a :-> a :-> a :-> a
negate_CC :: Num_CC a :-> a :-> a

dNumInt_CC :: Num_CC Int   -- as Int_CC = Int
dNumInt_CC = Num_CC 
               (to isoIntToIntToInt primAddInt) 
               (to isoIntToInt primNegateInt)
    isoIntToIntToInt = isoArr isoInt isoIntToInt
    isoIntToInt      = isoArr isoInt isoInt

Converting value bindings


For every binding

f :: t = e

we generate

f_CC :: t^ = e^


When converting a toplevel binding for f :: t, we generate f_CC :: t^ and redefine f as

f :: t = fr iso<t> f_CC



add :: Num a -> a -> a
add = \dNum x -> (+) dNum x 1

we generate

add :: Num a -> a -> a
add = fr isoFun add_CC
    isoFun = isoNum (id :<->: id) `isoArr`
             (id :<->: id)        `isoArr`
             (id :<->: id)

add_CC :: Num_CC a :-> a :-> a
add_CC = lam $ \dNum -> 
           (\dNum x -> (+_CC) $: dNum $: x $: 1) :$ dNum

If add is used in unconverted code it will still refer to the converted computation add_CC; i.e., we can use converted (and subsequently vectorised) code from unconverted/unvectorised code just by importing a converted/vectorised module as normal into an unconverted module.

Converting terms

cc[[C e1 .. en]] = C_CC e1 .. en   , if C_CC exists 
                                     and arity C = n
                 =                 , if C_CC exists
  lam_k $ \x'0 .. x'k ->           , and arity C = k + 1 + n
    (\(x1, .., xn) x(n+1) -> 
      C_CC x1 xn x(n+1) x'0 .. x'k :$ (e1, .. en)) 
cc[[x::t]]       = x_CC            , if x_CC exists
                 = to iso<t> x_CC  , otherwise
cc[[lit]]        = lit
cc[[e1 e2]]      = cc[[e1]] $: cc[e2]

chak: revision front

Apart from the standard rules, we need to handle the following special cases:

  • We come across a value variable v where idCC v == NoCC whose type is t: we generate convert t v (see below).
  • We come across a case expression where the scrutinised type T has tyConCC T == NoCC: we leave the case expression as is (i.e., unconverted), but make sure that the idCC field of all variables bound by patterns in the alternatives have their idCC field as NoCC. (This implies that the previous case will kick in and convert the (unconverted) values obtained after decomposition.)
  • Whenever we have an FC cast from or to a newtype T, where tyConCC T == NoCC, we need to add a convert tau or trevnoc tau, respectively. We can spot these casts by inspecting the kind of every coercion used in a cast. One side of the equality will have the newtype constructor.
  • We come across a dfun: If its idCC field is NoCC, we keep the selection as is, but apply convert t e from it, where t is the type of the selected method and e the selection expression. If idCC is ConvCC d_CC, and the dfun's class is converted, d_CC is fully converted. If it's class is not converted, we also keep the selection unconverted, but have a bit less to do in convert t e. TODO This needs to be fully worked out.



Have an example with two modules one unconverted, where the converted imports the unconverted.

Also have an example that motivates why we have to vectorise/CC declarations such as Int.