ghc-9.14.1: GHC/Tc/Deriv/Infer.hs
{-
(c) The University of Glasgow 2006
(c) The GRASP/AQUA Project, Glasgow University, 1992-1998
-}
{-# LANGUAGE MultiWayIf #-}
-- | Functions for inferring (and simplifying) the context for derived instances.
module GHC.Tc.Deriv.Infer
( inferConstraints
, simplifyInstanceContexts
)
where
import GHC.Prelude
import GHC.Tc.Deriv.Utils
import GHC.Tc.Errors.Types ( ErrCtxtMsg(..) )
import GHC.Tc.Utils.Env
import GHC.Tc.Deriv.Generate
import GHC.Tc.Deriv.Functor
import GHC.Tc.Deriv.Generics
import GHC.Tc.Utils.TcMType
import GHC.Tc.Utils.Monad
import GHC.Tc.Types.Origin
import GHC.Tc.Types.Constraint
import GHC.Tc.Utils.TcType
import GHC.Tc.Solver( simplifyTopImplic )
import GHC.Tc.Solver.Solve( solveWanteds )
import GHC.Tc.Solver.Monad ( runTcS )
import GHC.Tc.Validity (validDerivPred)
import GHC.Tc.Utils.Unify (buildImplicationFor)
import GHC.Tc.Zonk.TcType ( zonkWC )
import GHC.Tc.Zonk.Env ( ZonkFlexi(..), initZonkEnv )
import GHC.Core.Class
import GHC.Core.DataCon
import GHC.Core.TyCon
import GHC.Core.TyCo.Ppr (pprTyVars)
import GHC.Core.Type
import GHC.Core.Predicate
import GHC.Core.Unify (tcUnifyTy)
import GHC.Data.Pair
import GHC.Builtin.Names
import GHC.Builtin.Types (mkConstraintTupleTy, typeToTypeKind)
import GHC.Utils.Outputable
import GHC.Utils.Panic
import GHC.Utils.Misc
import GHC.Types.Basic
import GHC.Types.Var
import GHC.Data.Bag
import Control.Monad
import Control.Monad.Trans.Class (lift)
import Control.Monad.Trans.Reader (ask)
import Data.Function (on)
import Data.Functor.Classes (liftEq)
import Data.List (sortBy)
import Data.Maybe
----------------------
inferConstraints :: DerivSpecMechanism -> DerivEnv
-> DerivM (ThetaSpec, [TyVar], [TcType], DerivSpecMechanism)
-- inferConstraints figures out the constraints needed for the
-- instance declaration generated by a 'deriving' clause on a
-- data type declaration. It also returns the new in-scope type
-- variables and instance types, in case they were changed due to
-- the presence of functor-like constraints.
-- See Note [Inferring the instance context]
-- e.g. inferConstraints
-- C Int (T [a]) -- Class and inst_tys
-- :RTList a -- Rep tycon and its arg tys
-- where T [a] ~R :RTList a
--
-- Generate a sufficiently large set of constraints that typechecking the
-- generated method definitions should succeed. This set will be simplified
-- before being used in the instance declaration
inferConstraints mechanism (DerivEnv { denv_ctxt = ctxt
, denv_tvs = tvs
, denv_cls = main_cls
, denv_inst_tys = inst_tys })
= do { let wildcard = isStandaloneWildcardDeriv ctxt
infer_constraints :: DerivM (ThetaSpec, [TyVar], [TcType], DerivSpecMechanism)
infer_constraints =
case mechanism of
DerivSpecStock{dsm_stock_dit = dit}
-> do (thetas, tvs, inst_tys, dit') <- inferConstraintsStock dit
pure ( thetas, tvs, inst_tys
, mechanism{dsm_stock_dit = dit'} )
DerivSpecAnyClass
-> infer_constraints_simple inferConstraintsAnyclass
DerivSpecNewtype { dsm_newtype_dit =
DerivInstTys{dit_cls_tys = cls_tys}
, dsm_newtype_rep_ty = rep_ty }
-> infer_constraints_simple $
inferConstraintsCoerceBased cls_tys rep_ty
DerivSpecVia { dsm_via_cls_tys = cls_tys
, dsm_via_ty = via_ty }
-> infer_constraints_simple $
inferConstraintsCoerceBased cls_tys via_ty
-- Most deriving strategies do not need to do anything special to
-- the type variables and arguments to the class in the derived
-- instance, so they can pass through unchanged. The exception to
-- this rule is stock deriving. See
-- Note [Inferring the instance context].
infer_constraints_simple
:: DerivM ThetaSpec
-> DerivM (ThetaSpec, [TyVar], [TcType], DerivSpecMechanism)
infer_constraints_simple infer_thetas = do
thetas <- infer_thetas
pure (thetas, tvs, inst_tys, mechanism)
-- Constraints arising from superclasses
-- See Note [Superclasses of derived instance]
cls_tvs = classTyVars main_cls
sc_constraints = assertPpr (equalLength cls_tvs inst_tys)
(ppr main_cls <+> ppr inst_tys) $
mkDirectThetaSpec
(mkDerivOrigin wildcard) TypeLevel
(substTheta cls_subst (classSCTheta main_cls))
cls_subst = assert (equalLength cls_tvs inst_tys) $
zipTvSubst cls_tvs inst_tys
; (inferred_constraints, tvs', inst_tys', mechanism')
<- infer_constraints
; lift $ traceTc "inferConstraints" $ vcat
[ ppr main_cls <+> ppr inst_tys'
, ppr inferred_constraints
]
; return ( sc_constraints ++ inferred_constraints
, tvs', inst_tys', mechanism' ) }
-- | Like 'inferConstraints', but used only in the case of the @stock@ deriving
-- strategy. The constraints are inferred by inspecting the fields of each data
-- constructor. In this example:
--
-- > data Foo = MkFoo Int Char deriving Show
--
-- We would infer the following constraints ('ThetaSpec's):
--
-- > (Show Int, Show Char)
--
-- Note that this function also returns the type variables ('TyVar's) and
-- class arguments ('TcType's) for the resulting instance. This is because
-- when deriving 'Functor'-like classes, we must sometimes perform kind
-- substitutions to ensure the resulting instance is well kinded, which may
-- affect the type variables and class arguments. In this example:
--
-- > newtype Compose (f :: k -> Type) (g :: Type -> k) (a :: Type) =
-- > Compose (f (g a)) deriving stock Functor
--
-- We must unify @k@ with @Type@ in order for the resulting 'Functor' instance
-- to be well kinded, so we return @[]@/@[Type, f, g]@ for the
-- 'TyVar's/'TcType's, /not/ @[k]@/@[k, f, g]@.
-- See Note [Inferring the instance context].
inferConstraintsStock :: DerivInstTys
-> DerivM (ThetaSpec, [TyVar], [TcType], DerivInstTys)
inferConstraintsStock dit@(DerivInstTys { dit_cls_tys = cls_tys
, dit_tc = tc
, dit_tc_args = tc_args
, dit_rep_tc = rep_tc
, dit_rep_tc_args = rep_tc_args })
= do DerivEnv { denv_ctxt = ctxt
, denv_tvs = tvs
, denv_cls = main_cls
, denv_inst_tys = inst_tys } <- ask
let wildcard = isStandaloneWildcardDeriv ctxt
inst_ty = mkTyConApp tc tc_args
tc_binders = tyConBinders rep_tc
choose_level bndr
| isNamedTyConBinder bndr = KindLevel
| otherwise = TypeLevel
t_or_ks = map choose_level tc_binders ++ repeat TypeLevel
-- want to report *kind* errors when possible
-- Constraints arising from the arguments of each constructor
con_arg_constraints
:: ([TyVar] -> CtOrigin
-> TypeOrKind
-> Type
-> [(ThetaSpec, Maybe Subst)])
-> (ThetaSpec, [TyVar], [TcType], DerivInstTys)
con_arg_constraints get_arg_constraints
= let -- Constraints from the fields of each data constructor.
(predss, mbSubsts) = unzip
[ preds_and_mbSubst
| data_con <- tyConDataCons rep_tc
, (arg_n, arg_t_or_k, arg_ty)
<- zip3 [1..] t_or_ks $
derivDataConInstArgTys data_con dit
-- No constraints for unlifted types
-- See Note [Deriving and unboxed types]
, not (isUnliftedType arg_ty)
, let orig = DerivOriginDC data_con arg_n wildcard
, preds_and_mbSubst
<- get_arg_constraints (dataConUnivTyVars data_con)
orig arg_t_or_k arg_ty
]
-- Stupid constraints from DatatypeContexts. Note that we
-- must gather these constraints from the data constructors,
-- not from the parent type constructor, as the latter could
-- lead to redundant constraints due to thinning.
-- See Note [The stupid context] in GHC.Core.DataCon.
stupid_theta =
[ substTyWith (dataConUnivTyVars data_con)
(dataConInstUnivs data_con rep_tc_args)
stupid_pred
| data_con <- tyConDataCons rep_tc
, stupid_pred <- dataConStupidTheta data_con
]
preds = concat predss
-- If the constraints require a subtype to be of kind
-- (* -> *) (which is the case for functor-like
-- constraints), then we explicitly unify the subtype's
-- kinds with (* -> *).
-- See Note [Inferring the instance context]
subst = foldl' composeTCvSubst
emptySubst (catMaybes mbSubsts)
unmapped_tvs = filter (\v -> v `notElemSubst` subst
&& not (v `isInScope` subst)) tvs
(subst', _) = substTyVarBndrs subst unmapped_tvs
stupid_theta_origin = mkDirectThetaSpec
deriv_origin TypeLevel
(substTheta subst' stupid_theta)
preds' = map (substPredSpec subst') preds
inst_tys' = substTys subst' inst_tys
dit' = substDerivInstTys subst' dit
tvs' = tyCoVarsOfTypesWellScoped inst_tys'
in ( stupid_theta_origin ++ preds'
, tvs', inst_tys', dit' )
is_generic = main_cls `hasKey` genClassKey
is_generic1 = main_cls `hasKey` gen1ClassKey
-- is_functor_like: see Note [Inferring the instance context]
is_functor_like = typeKind inst_ty `tcEqKind` typeToTypeKind
|| is_generic1
get_gen1_constraints ::
Class
-> [TyVar] -- The universally quantified type variables for the
-- data constructor
-> CtOrigin -> TypeOrKind -> Type
-> [(ThetaSpec, Maybe Subst)]
get_gen1_constraints functor_cls dc_univs orig t_or_k ty
= mk_functor_like_constraints orig t_or_k functor_cls $
get_gen1_constrained_tys last_dc_univ ty
where
-- If we are deriving an instance of 'Generic1' and have made
-- it this far, then there should be at least one universal type
-- variable, making this use of 'last' safe.
last_dc_univ = assert (not (null dc_univs)) $
last dc_univs
get_std_constrained_tys ::
[TyVar] -- The universally quantified type variables for the
-- data constructor
-> CtOrigin -> TypeOrKind -> Type
-> [(ThetaSpec, Maybe Subst)]
get_std_constrained_tys dc_univs orig t_or_k ty
| is_functor_like
= mk_functor_like_constraints orig t_or_k main_cls $
deepSubtypesContaining last_dc_univ ty
| otherwise
= [( [mk_cls_pred orig t_or_k main_cls ty]
, Nothing )]
where
-- If 'is_functor_like' holds, then there should be at least one
-- universal type variable, making this use of 'last' safe.
last_dc_univ = assert (not (null dc_univs)) $
last dc_univs
mk_functor_like_constraints :: CtOrigin -> TypeOrKind
-> Class -> [Type]
-> [(ThetaSpec, Maybe Subst)]
-- 'cls' is usually main_cls (Functor or Traversable etc), but if
-- main_cls = Generic1, then 'cls' can be Functor; see
-- get_gen1_constraints
--
-- For each type, generate two constraints,
-- [cls ty, kind(ty) ~ (*->*)], and a kind substitution that results
-- from unifying kind(ty) with * -> *. If the unification is
-- successful, it will ensure that the resulting instance is well
-- kinded. If not, the second constraint will result in an error
-- message which points out the kind mismatch.
-- See Note [Inferring the instance context]
mk_functor_like_constraints orig t_or_k cls
= map $ \ty -> let ki = typeKind ty in
( [ mk_cls_pred orig t_or_k cls ty
, SimplePredSpec
{ sps_pred = mkNomEqPred ki typeToTypeKind
, sps_origin = orig
, sps_type_or_kind = KindLevel
}
]
, tcUnifyTy ki typeToTypeKind
)
-- Extra Data constraints
-- The Data class (only) requires that for
-- instance (...) => Data (T t1 t2)
-- IF t1:*, t2:*
-- THEN (Data t1, Data t2) are among the (...) constraints
-- Reason: when the IF holds, we generate a method
-- dataCast2 f = gcast2 f
-- and we need the Data constraints to typecheck the method
extra_constraints
| main_cls `hasKey` dataClassKey
, all (isLiftedTypeKind . typeKind) rep_tc_args
= [ mk_cls_pred deriv_origin t_or_k main_cls ty
| (t_or_k, ty) <- zip t_or_ks rep_tc_args]
| otherwise
= []
mk_cls_pred orig t_or_k cls ty
-- Don't forget to apply to cls_tys' too
= SimplePredSpec
{ sps_pred = mkClassPred cls (cls_tys' ++ [ty])
, sps_origin = orig
, sps_type_or_kind = t_or_k
}
cls_tys' | is_generic1 = []
-- In the awkward Generic1 case, cls_tys' should be
-- empty, since we are applying the class Functor.
| otherwise = cls_tys
deriv_origin = mkDerivOrigin wildcard
if -- Generic constraints are easy
| is_generic
-> return ([], tvs, inst_tys, dit)
-- Generic1 needs Functor
-- See Note [Getting base classes]
| is_generic1
-> assert (tyConTyVars rep_tc `lengthExceeds` 0) $
-- Generic1 has a single kind variable
assert (cls_tys `lengthIs` 1) $
do { functorClass <- lift $ tcLookupClass functorClassName
; pure $ con_arg_constraints
$ get_gen1_constraints functorClass }
-- The others are a bit more complicated
| otherwise
-> do { let (arg_constraints, tvs', inst_tys', dit')
= con_arg_constraints get_std_constrained_tys
; lift $ traceTc "inferConstraintsStock" $ vcat
[ ppr main_cls <+> ppr inst_tys'
, ppr arg_constraints
]
; return ( extra_constraints ++ arg_constraints
, tvs', inst_tys', dit' ) }
-- | Like 'inferConstraints', but used only in the case of @DeriveAnyClass@,
-- which gathers its constraints based on the type signatures of the class's
-- methods instead of the types of the data constructor's field.
--
-- See Note [Gathering and simplifying constraints for DeriveAnyClass]
-- for an explanation of how these constraints are used to determine the
-- derived instance context.
inferConstraintsAnyclass :: DerivM ThetaSpec
inferConstraintsAnyclass
= do { DerivEnv { denv_ctxt = ctxt
, denv_cls = cls
, denv_inst_tys = inst_tys } <- ask
; let gen_dms = [ (sel_id, dm_ty)
| (sel_id, Just (_, GenericDM dm_ty)) <- classOpItems cls ]
; let wildcard = isStandaloneWildcardDeriv ctxt
meth_pred :: (Id, Type) -> PredSpec
-- (Id,Type) are the selector Id and the generic default method type
-- NB: the latter is /not/ quantified over the class variables
-- See Note [Gathering and simplifying constraints for DeriveAnyClass]
meth_pred (sel_id, gen_dm_ty)
= let (sel_tvs, _cls_pred, meth_ty) = tcSplitMethodTy (varType sel_id)
meth_ty' = substTyWith sel_tvs inst_tys meth_ty
gen_dm_ty' = substTyWith sel_tvs inst_tys gen_dm_ty in
-- This is the only place where a SubTypePredSpec is
-- constructed instead of a SimplePredSpec. See
-- Note [Gathering and simplifying constraints for DeriveAnyClass]
-- for a more in-depth explanation.
SubTypePredSpec { stps_ty_actual = gen_dm_ty'
, stps_ty_expected = meth_ty'
, stps_origin = mkDerivOrigin wildcard
}
; pure $ map meth_pred gen_dms }
-- Like 'inferConstraints', but used only for @GeneralizedNewtypeDeriving@ and
-- @DerivingVia@. Since both strategies generate code involving 'coerce', the
-- inferred constraints set up the scaffolding needed to typecheck those uses
-- of 'coerce'. In this example:
--
-- > newtype Age = MkAge Int deriving newtype Num
--
-- We would infer the following constraints ('ThetaSpec'):
--
-- > (Num Int, Coercible Age Int)
inferConstraintsCoerceBased :: [Type] -> Type
-> DerivM ThetaSpec
inferConstraintsCoerceBased cls_tys rep_ty = do
DerivEnv { denv_ctxt = ctxt
, denv_tvs = tvs
, denv_cls = cls
, denv_inst_tys = inst_tys } <- ask
let -- rep_ty might come from:
-- GeneralizedNewtypeDeriving / DerivSpecNewtype:
-- the underlying type of the newtype ()
-- DerivingVia / DerivSpecVia
-- the `via` type
rep_pred_o = SimplePredSpec { sps_pred = mkClassPred cls (cls_tys ++ [rep_ty])
, sps_origin = deriv_origin
, sps_type_or_kind = TypeLevel
}
-- rep_pred is the representation dictionary, from where
-- we are going to get all the methods for the final
-- dictionary
deriv_origin = mkDerivOrigin sa_wildcard
sa_wildcard = isStandaloneWildcardDeriv ctxt
-- Next we collect constraints for the class methods
-- If there are no methods, we don't need any constraints
-- Otherwise we need (C rep_ty), for the representation methods,
-- and constraints to coerce each individual method
meth_preds :: ThetaSpec
meth_preds | null meths = [] -- No methods => no constraints (#12814)
| otherwise = rep_pred_o : coercible_constraints
meths = classMethods cls
coercible_constraints
= [ SimplePredSpec
{ sps_pred =
assertPpr (tvs1 == tvs2) (ppr t1 $$ ppr t2) $
-- assert: mkCoerceClassMethEqn returns two
-- foralls with the very same forall-binders
tcMkDFunSigmaTy tvs2 theta2 $
mkConstraintTupleTy $ mkReprEqPred tau1 tau2 : theta1
-- The two method types (tau1, tau2) must be coercible.
-- Also, if there are constraints, the constraints
-- provided to the derived method (theta2) must be
-- sufficient to solve the constraints required by the
-- method being coerced (theta1).
-- See Note [Inferred contexts from method constraints]
, sps_origin = DerivOriginCoerce meth t1 t2 sa_wildcard
, sps_type_or_kind = TypeLevel
}
| meth <- meths
, let (Pair t1 t2) = mkCoerceClassMethEqn cls tvs
inst_tys rep_ty meth
-- If we have class C a b c where { op :: op_ty }
-- and inst_tys = [t1, t2, t3]
-- then t1 = op_ty{t1,t2,rep_ty/a,b,c]
-- t2 = op_ty{t1,t2,t3/a,b,c]
, let (tvs1, theta1, tau1) = tcSplitSigmaTy t1
, let (tvs2, theta2, tau2) = tcSplitSigmaTy t2
]
pure meth_preds
{- Note [Inferring the instance context]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
There are two sorts of 'deriving', as represented by the two constructors
for DerivContext:
* InferContext mb_wildcard: This can either be:
- The deriving clause for a data type.
(e.g, data T a = T1 a deriving( Eq ))
In this case, mb_wildcard = Nothing.
- A standalone declaration with an extra-constraints wildcard
(e.g., deriving instance _ => Eq (Foo a))
In this case, mb_wildcard = Just loc, where loc is the location
of the extra-constraints wildcard.
Here we must infer an instance context,
and generate instance declaration
instance Eq a => Eq (T a) where ...
* SupplyContext theta: standalone deriving
deriving instance Eq a => Eq (T a)
Here we only need to fill in the bindings;
the instance context (theta) is user-supplied
For the InferContext case, we must figure out the
instance context (inferConstraintsStock). Suppose we are inferring
the instance context for
C t1 .. tn (T s1 .. sm)
There are two cases
* (T s1 .. sm) :: * (the normal case)
Then we behave like Eq and guess (C t1 .. tn t)
for each data constructor arg of type t. More
details below.
* (T s1 .. sm) :: * -> * (the functor-like case)
Then we behave like Functor.
In both cases we produce a bunch of un-simplified constraints
and them simplify them in simplifyInstanceContexts; see
Note [Simplifying the instance context].
In the functor-like case, we may need to unify some kind variables with * in
order for the generated instance to be well-kinded. An example from #10524:
newtype Compose (f :: k2 -> *) (g :: k1 -> k2) (a :: k1)
= Compose (f (g a)) deriving Functor
Earlier in the deriving pipeline, GHC unifies the kind of Compose f g
(k1 -> *) with the kind of Functor's argument (* -> *), so k1 := *. But this
alone isn't enough, since k2 wasn't unified with *:
instance (Functor (f :: k2 -> *), Functor (g :: * -> k2)) =>
Functor (Compose f g) where ...
The two Functor constraints are ill-kinded. To ensure this doesn't happen, we:
1. Collect all of a datatype's subtypes which require functor-like
constraints.
2. For each subtype, create a substitution by unifying the subtype's kind
with (* -> *).
3. Compose all the substitutions into one, then apply that substitution to
all of the in-scope type variables and the instance types.
Note [Getting base classes]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Functor and Typeable are defined in package 'base', and that is not available
when compiling 'ghc-prim'. So we must be careful that 'deriving' for stuff in
ghc-prim does not use Functor or Typeable implicitly via these lookups.
Note [Deriving and unboxed types]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We have some special hacks to support things like
data T = MkT Int# deriving ( Show )
Specifically, we use GHC.Tc.Deriv.Generate.box to box the Int# into an Int
(which we know how to show), and append a '#'. Parentheses are not required
for unboxed values (`MkT -3#` is a valid expression).
Note [Superclasses of derived instance]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In general, a derived instance decl needs the superclasses of the derived
class too. So if we have
data T a = ...deriving( Ord )
then the initial context for Ord (T a) should include Eq (T a). Often this is
redundant; we'll also generate an Ord constraint for each constructor argument,
and that will probably generate enough constraints to make the Eq (T a) constraint
be satisfied too. But not always; consider:
data S a = S
instance Eq (S a)
instance Ord (S a)
data T a = MkT (S a) deriving( Ord )
instance Num a => Eq (T a)
The derived instance for (Ord (T a)) must have a (Num a) constraint!
Similarly consider:
data T a = MkT deriving( Data )
Here there *is* no argument field, but we must nevertheless generate
a context for the Data instances:
instance Typeable a => Data (T a) where ...
Note [Inferred contexts from method constraints]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider the `deriving Alt` part of this example (from the passing part of
T20815a):
class Alt f where
some :: forall a. Applicative f => f a -> f [a]
newtype T f a = T (f a) deriving Alt
We will produce this derived instance declaration:
instance (Alt f, ???) => Alt (T f) where
some :: forall a. Applicative (T f) => T f a -> T f [a]
some @a (d1 :: Applicative (T f))
= coerce @(f a -> f [a])
@(T f a -> T f [a])
(d2 :: Coercible (f a -> f [a]) (T f a -> T f [a]))
(some @f (d3 :: Alt f) @a (d4 :: Applicative f))
(Dictionary abstractions and applications are added here even though they are
not usually visible, or even emitted in the code generated by `deriving`.)
The task of `inferConstraints` is to determine the `???` such that it will be
sufficient to solve the constraints arising from that definition of `some`. We
can write out what the type checker sees as follows:
forall f
[G] Alt f -- Given
[G] ??? -- Given
=>
forall a.
[G] Applicative (T f) -- Also given (as d1)
=>
[W] Coercible (f a -> f [a]) (T f a -> T f [a]) -- Wanted (as d2)
[W] Alt f -- Wanted (as d3)
[W] Applicative f -- Wanted (as d4)
`d3` is trivially provided by the given `Alt f`. The simplest way to ensure that
`d4` and `d2` can be solved is to:
* Generate this "target constraint" (in `inferConstraintsCoerceBased`):
forall a. Applicative (T f)
=> ( Coercible (f a -> f [a]) (T f a -> T f [a])
, Applicative f
)
* Simplify the target constraint (in `simplifyInstanceContexts`, which in turn
calls `simplifyDeriv`). This solves the `Coercible` constraint outright, but
cannot solve the `Applicative f` constraint.
See Note [Simplifying the instance context]
* The leftover, unsolved constraint (here `Applicative f`) becomes the `???` in
the derived instance decl.
The target constraint for GND is created in `inferConstraintsCoerceBased`.
In general, the point here is that the inferred context for a derived instance
must include, for each class method with constraints, a quantified constraint
mapping the provided context for the derived method to both:
- the `Coercible` corresponding to the monotypes of the base and derived
methods, and
- the needed context for the base method.
************************************************************************
* *
Finding the fixed point of deriving equations
* *
************************************************************************
Note [Simplifying the instance context]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
data T a b = C1 (Foo a) (Bar b)
| C2 Int (T b a)
| C3 (T a a)
deriving (Eq)
We want to come up with an instance declaration of the form
instance (Ping a, Pong b, ...) => Eq (T a b) where
x == y = ...
It is pretty easy, albeit tedious, to fill in the code "...". The
trick is to figure out what the context for the instance decl is,
namely Ping, Pong and friends.
Let's call the context reqd for the T instance of class C at types
(a,b, ...) C (T a b). Thus:
Eq (T a b) = (Ping a, Pong b, ...)
Now we can get a (recursive) equation from the data decl. This part
is done by inferConstraintsStock.
Eq (T a b) = Eq (Foo a) u Eq (Bar b) -- From C1
u Eq (T b a) u Eq Int -- From C2
u Eq (T a a) -- From C3
Foo and Bar may have explicit instances for Eq, in which case we can
just substitute for them. Alternatively, either or both may have
their Eq instances given by deriving clauses, in which case they
form part of the system of equations.
Now all we need do is simplify and solve the equations, iterating to
find the least fixpoint. This is done by simplifyInstanceConstraints.
Notice that the order of the arguments can
switch around, as here in the recursive calls to T.
Let's suppose Eq (Foo a) = Eq a, and Eq (Bar b) = Ping b.
We start with:
Eq (T a b) = {} -- The empty set
Next iteration:
Eq (T a b) = Eq (Foo a) u Eq (Bar b) -- From C1
u Eq (T b a) u Eq Int -- From C2
u Eq (T a a) -- From C3
After simplification:
= Eq a u Ping b u {} u {} u {}
= Eq a u Ping b
Next iteration:
Eq (T a b) = Eq (Foo a) u Eq (Bar b) -- From C1
u Eq (T b a) u Eq Int -- From C2
u Eq (T a a) -- From C3
After simplification:
= Eq a u Ping b
u (Eq b u Ping a)
u (Eq a u Ping a)
= Eq a u Ping b u Eq b u Ping a
The next iteration gives the same result, so this is the fixpoint. We
need to make a canonical form of the RHS to ensure convergence. We do
this by simplifying the RHS to a form in which
- the classes constrain only tyvars
- the list is sorted by tyvar (major key) and then class (minor key)
- no duplicates, of course
Note [Deterministic simplifyInstanceContexts]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Canonicalisation uses nonDetCmpType which is nondeterministic. Sorting
with nonDetCmpType puts the returned lists in a nondeterministic order.
If we were to return them, we'd get class constraints in
nondeterministic order.
Consider:
data ADT a b = Z a b deriving Eq
The generated code could be either:
instance (Eq a, Eq b) => Eq (Z a b) where
Or:
instance (Eq b, Eq a) => Eq (Z a b) where
To prevent the order from being nondeterministic we only
canonicalize when comparing and return them in the same order as
simplifyDeriv returned them.
See also Note [nonDetCmpType nondeterminism]
-}
simplifyInstanceContexts :: [DerivSpec ThetaSpec]
-> TcM [DerivSpec ThetaType]
-- Used only for deriving clauses or standalone deriving with an
-- extra-constraints wildcard (InferContext)
-- See Note [Simplifying the instance context]
simplifyInstanceContexts [] = return []
simplifyInstanceContexts infer_specs
= do { traceTc "simplifyInstanceContexts" $ vcat (map pprDerivSpec infer_specs)
; final_specs <- iterate_deriv 1 initial_solutions
-- After simplification finishes, zonk the TcTyVars as described
-- in Note [Overlap and deriving].
; initZonkEnv DefaultFlexi $ traverse zonkDerivSpec final_specs }
where
------------------------------------------------------------------
-- The initial solutions for the equations claim that each
-- instance has an empty context; this solution is certainly
-- in canonical form.
initial_solutions :: [ThetaType]
initial_solutions = [ [] | _ <- infer_specs ]
------------------------------------------------------------------
-- iterate_deriv calculates the next batch of solutions,
-- compares it with the current one; finishes if they are the
-- same, otherwise recurses with the new solutions.
-- It fails if any iteration fails
iterate_deriv :: Int -> [ThetaType] -> TcM [DerivSpec ThetaType]
iterate_deriv n current_solns
| n > 20 -- Looks as if we are in an infinite loop
-- This can happen if we have -XUndecidableInstances
-- (See GHC.Tc.Solver.tcSimplifyDeriv.)
= pprPanic "solveDerivEqns: probable loop"
(vcat (map pprDerivSpec infer_specs) $$ ppr current_solns)
| otherwise
= do { -- Extend the inst info from the explicit instance decls
-- with the current set of solutions, and simplify each RHS
inst_specs <- zipWithM (\soln -> newDerivClsInst . setDerivSpecTheta soln)
current_solns infer_specs
; new_solns <- checkNoErrs $
extendLocalInstEnv inst_specs $
mapM simplifyDeriv infer_specs
; if (current_solns `eqSolution` new_solns) then
return [ setDerivSpecTheta soln spec
| (spec, soln) <- zip infer_specs current_solns ]
else
iterate_deriv (n+1) new_solns }
eqSolution = (liftEq . liftEq) eqType `on` canSolution
-- Canonicalise for comparison
-- See Note [Deterministic simplifyInstanceContexts]
canSolution = map (sortBy nonDetCmpType)
{-
***********************************************************************************
* *
* Simplify derived constraints
* *
***********************************************************************************
-}
-- | Given @instance (wanted) => C inst_ty@, simplify 'wanted' as much
-- as possible. Fail if not possible.
simplifyDeriv :: DerivSpec ThetaSpec
-> TcM ThetaType -- ^ Needed constraints (after simplification),
-- i.e. @['PredType']@.
simplifyDeriv (DS { ds_loc = loc, ds_tvs = tvs
, ds_cls = clas, ds_tys = inst_tys, ds_theta = deriv_rhs
, ds_skol_info = skol_info, ds_user_ctxt = user_ctxt })
= setSrcSpan loc $
addErrCtxt (DerivInstCtxt (mkClassPred clas inst_tys)) $
do {
-- See [STEP DAC BUILD]
-- Generate the implication constraints, one for each method, to solve
-- with the skolemized variables. Start "one level down" because
-- we are going to wrap the result in an implication with tvs,
-- in step [DAC RESIDUAL]
; (tc_lvl, wanteds) <- captureThetaSpecConstraints user_ctxt deriv_rhs
; traceTc "simplifyDeriv inputs" $
vcat [ pprTyVars tvs $$ ppr deriv_rhs $$ ppr wanteds, ppr skol_info ]
-- See [STEP DAC SOLVE]
-- Simplify the constraints, starting at the same level at which
-- they are generated (c.f. the call to runTcSWithEvBinds in
-- simplifyInfer)
; (solved_wanteds, _) <- setTcLevel tc_lvl $
runTcS $
solveWanteds wanteds
-- It's not yet zonked! Obviously zonk it before peering at it
; solved_wanteds <- liftZonkM $ zonkWC solved_wanteds
-- See [STEP DAC HOIST]
-- From the simplified constraints extract a subset 'good' that will
-- become the context 'min_theta' for the derived instance.
; let residual_simple = approximateWC False solved_wanteds
-- False: ignore any non-quantifiable constraints,
-- including equalities hidden under Given equalities
head_size = pSizeClassPred clas inst_tys
good = mapMaybeBag get_good residual_simple
-- Returns @Just p@ (where @p@ is the type of the Ct) if a Ct is
-- suitable to be inferred in the context of a derived instance.
-- Returns @Nothing@ if the Ct is too exotic.
-- See (VD2) in Note [Valid 'deriving' predicate] in
-- GHC.Tc.Validity for what constitutes an exotic constraint.
get_good :: Ct -> Maybe PredType
get_good ct | validDerivPred head_size p = Just p
| otherwise = Nothing
where p = ctPred ct
; traceTc "simplifyDeriv outputs" $
vcat [ ppr tvs, ppr residual_simple, ppr good ]
-- Return the good unsolved constraints (unskolemizing on the way out.)
; let min_theta = mkMinimalBySCs id (bagToList good)
-- An important property of mkMinimalBySCs (used above) is that in
-- addition to removing constraints that are made redundant by
-- superclass relationships, it also removes _duplicate_
-- constraints.
-- See Note [Gathering and simplifying constraints for
-- DeriveAnyClass]
-- See [STEP DAC RESIDUAL]
-- Ensure that min_theta is enough to solve /all/ the constraints in
-- solved_wanteds, by solving the implication constraint
--
-- forall tvs. min_theta => solved_wanteds
; min_theta_vars <- mapM newEvVar min_theta
; (leftover_implic, _)
<- buildImplicationFor tc_lvl (getSkolemInfo skol_info) tvs
min_theta_vars solved_wanteds
-- This call to simplifyTop is purely for error reporting
-- See Note [Error reporting for deriving clauses]
-- See also Note [Valid 'deriving' predicate] in GHC.Tc.Validity, as this
-- line of code catches "exotic" constraints like the ones described in
-- (VD2) of that Note.
; simplifyTopImplic leftover_implic
; traceTc "GHC.Tc.Deriv" (ppr deriv_rhs $$ ppr min_theta)
-- Claim: the result instance declaration is guaranteed valid
-- Hence no need to call:
-- checkValidInstance tyvars theta clas inst_tys
-- See Note [Valid 'deriving' predicate] in GHC.Tc.Validity
; return min_theta }
{-
Note [Overlap and deriving]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider some overlapping instances:
instance Show a => Show [a] where ..
instance Show [Char] where ...
Now a data type with deriving:
data T a = MkT [a] deriving( Show )
We want to get the derived instance
instance Show [a] => Show (T a) where...
and NOT
instance Show a => Show (T a) where...
so that the (Show (T Char)) instance does the Right Thing
It's very like the situation when we're inferring the type
of a function
f x = show [x]
and we want to infer
f :: Show [a] => a -> String
As a result, we use vanilla, non-overlappable skolems when inferring the
context for the derived instances. Hence, we instantiate the type variables
using tcInstSkolTyVars, not tcInstSuperSkolTyVars.
We do this skolemisation in GHC.Tc.Deriv.mkEqnHelp, a function which occurs
very early in the deriving pipeline, so that by the time GHC needs to infer the
instance context, all of the types in the computed DerivSpec have been
skolemised appropriately. After the instance context inference has completed,
GHC zonks the TcTyVars in the DerivSpec to ensure that types like
a[sk:1] do not appear in -ddump-deriv output.
All of this is only needed when inferring an instance context, i.e., the
InferContext case. For the SupplyContext case, we don't bother skolemising
at all.
Note [Gathering and simplifying constraints for DeriveAnyClass]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
DeriveAnyClass works quite differently from stock and newtype deriving in
the way it gathers and simplifies constraints to be used in a derived
instance's context. Stock and newtype deriving gather constraints by looking
at the data constructors of the data type for which we are deriving an
instance. But DeriveAnyClass doesn't need to know about a data type's
definition at all!
To see why, consider this example of DeriveAnyClass:
class Foo a where
bar :: forall b. Ix b => a -> b -> String
default bar :: (Show a, Ix c) => a -> c -> String
bar x y = show x ++ show (range (y,y))
baz :: Eq a => a -> a -> Bool
default baz :: (Ord a, Show a) => a -> a -> Bool
baz x y = compare x y == EQ
Because 'bar' and 'baz' have default signatures, this generates a top-level
definition for these generic default methods
$gdm_bar :: forall a. Foo a
=> forall c. (Show a, Ix c)
=> a -> c -> String
$gdm_bar x y = show x ++ show (range (y,y))
(and similarly for baz). Now consider a 'deriving' clause
data Maybe s = ... deriving anyclass Foo
This derives an instance of the form:
instance (CX) => Foo (Maybe s) where
bar = $gdm_bar
baz = $gdm_baz
Now it is GHC's job to fill in a suitable instance context (CX). If
GHC were typechecking the binding
bar = $gdm_bar
it would
* skolemise the expected type of bar
* instantiate the type of $gdm_bar with meta-type variables
* build an implication constraint
[STEP DAC BUILD]
So that's what we do. Fortunately, there is already functionality within GHC
to that does all of the above—namely, tcSubTypeSigma. In the example above,
we want to use tcSubTypeSigma to check the following subtyping relation:
forall c. (Show a, Ix c) => Maybe s -> c -> String -- actual type
<= forall b. (Ix b) => Maybe s -> b -> String -- expected type
That is, we check that the type of $gdm_bar (the actual type) is more
polymorphic than the type of bar (the expected type). We use SubTypePredSpec,
a special form of PredSpec that is only used by DeriveAnyClass, to store
the actual and expected types.
(Aside: having a separate SubTypePredSpec is not strictly necessary, as we
could theoretically construct this implication constraint by hand and store it
in a SimplePredSpec. In fact, GHC used to do this. However, this is easier
said than done, and there were numerous subtle bugs that resulted from getting
this step wrong, such as #20719. Ultimately, we decided that special-casing a
PredSpec specifically for DeriveAnyClass was worth it.)
tcSubTypeSigma will ultimately spit out an implication constraint, which will
look something like this (call it C1):
forall[2] b. Ix b => (Show (Maybe s), Ix cc,
Maybe s -> b -> String
~ Maybe s -> cc -> String)
Here:
* The level of this forall constraint is forall[2], because we are later
going to wrap it in a forall[1] in [STEP DAC RESIDUAL]
* The 'b' comes from the quantified type variable in the expected type
of bar. The 'cc' is a unification variable that comes from instantiating the
quantified type variable 'c' in $gdm_bar's type. The finer details of
skolemisation and metavariable instantiation are handled behind the scenes
by tcSubTypeSigma.
* It is important that `b` be distinct from `cc`. In this example, this is
clearly the case, but it is not always so obvious when the type variables are
hidden behind type synonyms. Suppose the example were written like this,
for example:
type Method a = forall b. Ix b => a -> b -> String
class Foo a where
bar :: Method a
default bar :: Show a => Method a
bar = ...
Both method signatures quantify a `b` once the `Method` type synonym is
expanded. To ensure that GHC doesn't confuse the two `b`s during
typechecking, tcSubTypeSigma instantiates the `b` in the original signature
with a fresh skolem and the `b` in the default signature with a fresh
unification variable. Doing so prevents #20719 from happening.
* The (Ix b) constraint comes from the context of bar's type. The
(Show (Maybe s)) and (Ix cc) constraints come from the context of $gdm_bar's
type.
* The equality constraint (Maybe s -> b -> String) ~ (Maybe s -> cc -> String)
comes from marrying up the instantiated type of $gdm_bar with the specified
type of bar. Notice that the type variables from the instance, 's' in this
case, are global to this constraint.
Note that it is vital that we instantiate the `c` in $gdm_bar's type with a new
unification variable for each iteration of simplifyDeriv. If we re-use the same
unification variable across multiple iterations, then bad things can happen,
such as #14933.
Similarly for 'baz', tcSubTypeSigma gives the constraint C2
forall[2]. Eq (Maybe s) => (Ord a, Show a,
Maybe s -> Maybe s -> Bool
~ Maybe s -> Maybe s -> Bool)
In this case baz has no local quantification, so the implication
constraint has no local skolems and there are no unification
variables.
[STEP DAC SOLVE]
We can combine these two implication constraints into a single
constraint (C1, C2), and simplify, unifying cc:=b, to get:
forall[2] b. Ix b => Show a
/\
forall[2]. Eq (Maybe s) => (Ord a, Show a)
[STEP DAC HOIST]
Let's call that (C1', C2'). Now we need to hoist the unsolved
constraints out of the implications to become our candidate for
(CX). That is done by approximateWC, which will return:
(Show a, Ord a, Show a)
Now we can use mkMinimalBySCs to remove superclasses and duplicates, giving
(Show a, Ord a)
And that's what GHC uses for CX.
[STEP DAC RESIDUAL]
In this case we have solved all the leftover constraints, but what if
we don't? Simple! We just form the final residual constraint
forall[1] s. CX => (C1',C2')
and simplify that. In simple cases it'll succeed easily, because CX
literally contains the constraints in C1', C2', but if there is anything
more complicated it will be reported in a civilised way.
Note [Error reporting for deriving clauses]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
A surprisingly tricky aspect of deriving to get right is reporting sensible
error messages. In particular, if simplifyDeriv reaches a constraint that it
cannot solve, which might include:
1. Insoluble constraints
2. "Exotic" constraints (See Note [Valid 'deriving' predicate] in
GHC.Tc.Validity)
Then we report an error immediately in simplifyDeriv.
Another possible choice is to punt and let another part of the typechecker
(e.g., simplifyInstanceContexts) catch the errors. But this tends to lead
to worse error messages, so we do it directly in simplifyDeriv.
simplifyDeriv checks for errors in a clever way. If the deriving machinery
infers the context (Foo a)--that is, if this instance is to be generated:
instance Foo a => ...
Then we form an implication of the form:
forall a. Foo a => <residual_wanted_constraints>
And pass it to the simplifier. If the context (Foo a) is enough to discharge
all the constraints in <residual_wanted_constraints>, then everything is
hunky-dory. But if <residual_wanted_constraints> contains, say, an insoluble
constraint, then (Foo a) won't be able to solve it, causing GHC to error.
-}