module GHC.Core.Opt.Exitify ( exitifyProgram ) where

{-
Note [Exitification]
~~~~~~~~~~~~~~~~~~~~

This module implements Exitification. The goal is to pull as much code out of
recursive functions as possible, as the simplifier is better at inlining into
call-sites that are not in recursive functions.

Example:

  let t = foo bar
  joinrec go 0     x y = t (x*x)
          go (n-1) x y = jump go (n-1) (x+y)
  in …

We’d like to inline `t`, but that does not happen: Because t is a thunk and is
used in a recursive function, doing so might lose sharing in general. In
this case, however, `t` is on the _exit path_ of `go`, so called at most once.
How do we make this clearly visible to the simplifier?

A code path (i.e., an expression in a tail-recursive position) in a recursive
function is an exit path if it does not contain a recursive call. We can bind
this expression outside the recursive function, as a join-point.

Example result:

  let t = foo bar
  join exit x = t (x*x)
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)
  in …

Now `t` is no longer in a recursive function, and good things happen!
-}

import GHC.Prelude
import GHC.Builtin.Uniques
import GHC.Core
import GHC.Core.Utils
import GHC.Core.FVs
import GHC.Core.Type

import GHC.Types.Var
import GHC.Types.Id
import GHC.Types.Id.Info
import GHC.Types.Var.Set
import GHC.Types.Var.Env
import GHC.Types.Basic( JoinPointHood(..) )

import GHC.Utils.Monad.State.Strict
import GHC.Utils.Misc( mapSnd )

import GHC.Data.FastString

import Data.Bifunctor
import Control.Monad

-- | Traverses the AST, simply to find all joinrecs and call 'exitify' on them.
-- The really interesting function is exitifyRec
exitifyProgram :: CoreProgram -> CoreProgram
exitifyProgram :: CoreProgram -> CoreProgram
exitifyProgram CoreProgram
binds = (Bind JoinId -> Bind JoinId) -> CoreProgram -> CoreProgram
forall a b. (a -> b) -> [a] -> [b]
map Bind JoinId -> Bind JoinId
goTopLvl CoreProgram
binds
  where
    goTopLvl :: Bind JoinId -> Bind JoinId
goTopLvl (NonRec JoinId
v CoreExpr
e) = JoinId -> CoreExpr -> Bind JoinId
forall b. b -> Expr b -> Bind b
NonRec JoinId
v (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope_toplvl CoreExpr
e)
    goTopLvl (Rec [(JoinId, CoreExpr)]
pairs) = [(JoinId, CoreExpr)] -> Bind JoinId
forall b. [(b, Expr b)] -> Bind b
Rec (((JoinId, CoreExpr) -> (JoinId, CoreExpr))
-> [(JoinId, CoreExpr)] -> [(JoinId, CoreExpr)]
forall a b. (a -> b) -> [a] -> [b]
map ((CoreExpr -> CoreExpr) -> (JoinId, CoreExpr) -> (JoinId, CoreExpr)
forall b c a. (b -> c) -> (a, b) -> (a, c)
forall (p :: * -> * -> *) b c a.
Bifunctor p =>
(b -> c) -> p a b -> p a c
second (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope_toplvl)) [(JoinId, CoreExpr)]
pairs)
      -- Top-level bindings are never join points

    in_scope_toplvl :: InScopeSet
in_scope_toplvl = InScopeSet
emptyInScopeSet InScopeSet -> CoreProgram -> InScopeSet
`extendInScopeSetBndrs` CoreProgram
binds

    go :: InScopeSet -> CoreExpr -> CoreExpr
    go :: InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
_    e :: CoreExpr
e@(Var{})       = CoreExpr
e
    go InScopeSet
_    e :: CoreExpr
e@(Lit {})      = CoreExpr
e
    go InScopeSet
_    e :: CoreExpr
e@(Type {})     = CoreExpr
e
    go InScopeSet
_    e :: CoreExpr
e@(Coercion {}) = CoreExpr
e
    go InScopeSet
in_scope (Cast CoreExpr
e' CoercionR
c) = CoreExpr -> CoercionR -> CoreExpr
forall b. Expr b -> CoercionR -> Expr b
Cast (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
e') CoercionR
c
    go InScopeSet
in_scope (Tick CoreTickish
t CoreExpr
e') = CoreTickish -> CoreExpr -> CoreExpr
forall b. CoreTickish -> Expr b -> Expr b
Tick CoreTickish
t (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
e')
    go InScopeSet
in_scope (App CoreExpr
e1 CoreExpr
e2) = CoreExpr -> CoreExpr -> CoreExpr
forall b. Expr b -> Expr b -> Expr b
App (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
e1) (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
e2)

    go InScopeSet
in_scope (Lam JoinId
v CoreExpr
e')
      = JoinId -> CoreExpr -> CoreExpr
forall b. b -> Expr b -> Expr b
Lam JoinId
v (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope' CoreExpr
e')
      where in_scope' :: InScopeSet
in_scope' = InScopeSet
in_scope InScopeSet -> JoinId -> InScopeSet
`extendInScopeSet` JoinId
v

    go InScopeSet
in_scope (Case CoreExpr
scrut JoinId
bndr Type
ty [Alt JoinId]
alts)
      = CoreExpr -> JoinId -> Type -> [Alt JoinId] -> CoreExpr
forall b. Expr b -> b -> Type -> [Alt b] -> Expr b
Case (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
scrut) JoinId
bndr Type
ty ((Alt JoinId -> Alt JoinId) -> [Alt JoinId] -> [Alt JoinId]
forall a b. (a -> b) -> [a] -> [b]
map Alt JoinId -> Alt JoinId
go_alt [Alt JoinId]
alts)
      where
        in_scope1 :: InScopeSet
in_scope1 = InScopeSet
in_scope InScopeSet -> JoinId -> InScopeSet
`extendInScopeSet` JoinId
bndr
        go_alt :: Alt JoinId -> Alt JoinId
go_alt (Alt AltCon
dc [JoinId]
pats CoreExpr
rhs) = AltCon -> [JoinId] -> CoreExpr -> Alt JoinId
forall b. AltCon -> [b] -> Expr b -> Alt b
Alt AltCon
dc [JoinId]
pats (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope' CoreExpr
rhs)
           where in_scope' :: InScopeSet
in_scope' = InScopeSet
in_scope1 InScopeSet -> [JoinId] -> InScopeSet
`extendInScopeSetList` [JoinId]
pats

    go InScopeSet
in_scope (Let (NonRec JoinId
bndr CoreExpr
rhs) CoreExpr
body)
      = Bind JoinId -> CoreExpr -> CoreExpr
forall b. Bind b -> Expr b -> Expr b
Let (JoinId -> CoreExpr -> Bind JoinId
forall b. b -> Expr b -> Bind b
NonRec JoinId
bndr (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope CoreExpr
rhs)) (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope' CoreExpr
body)
      where
        in_scope' :: InScopeSet
in_scope' = InScopeSet
in_scope InScopeSet -> JoinId -> InScopeSet
`extendInScopeSet` JoinId
bndr

    go InScopeSet
in_scope (Let (Rec [(JoinId, CoreExpr)]
pairs) CoreExpr
body)
      | Bool
is_join_rec = CoreProgram -> CoreExpr -> CoreExpr
forall b. [Bind b] -> Expr b -> Expr b
mkLets (InScopeSet -> [(JoinId, CoreExpr)] -> CoreProgram
exitifyRec InScopeSet
in_scope' [(JoinId, CoreExpr)]
pairs') CoreExpr
body'
      | Bool
otherwise   = Bind JoinId -> CoreExpr -> CoreExpr
forall b. Bind b -> Expr b -> Expr b
Let ([(JoinId, CoreExpr)] -> Bind JoinId
forall b. [(b, Expr b)] -> Bind b
Rec [(JoinId, CoreExpr)]
pairs') CoreExpr
body'
      where
        is_join_rec :: Bool
is_join_rec = ((JoinId, CoreExpr) -> Bool) -> [(JoinId, CoreExpr)] -> Bool
forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool
any (JoinId -> Bool
isJoinId (JoinId -> Bool)
-> ((JoinId, CoreExpr) -> JoinId) -> (JoinId, CoreExpr) -> Bool
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (JoinId, CoreExpr) -> JoinId
forall a b. (a, b) -> a
fst) [(JoinId, CoreExpr)]
pairs
        in_scope' :: InScopeSet
in_scope'   = InScopeSet
in_scope InScopeSet -> Bind JoinId -> InScopeSet
`extendInScopeSetBind` ([(JoinId, CoreExpr)] -> Bind JoinId
forall b. [(b, Expr b)] -> Bind b
Rec [(JoinId, CoreExpr)]
pairs)
        pairs' :: [(JoinId, CoreExpr)]
pairs'      = (CoreExpr -> CoreExpr)
-> [(JoinId, CoreExpr)] -> [(JoinId, CoreExpr)]
forall (f :: * -> *) b c a.
Functor f =>
(b -> c) -> f (a, b) -> f (a, c)
mapSnd (InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope') [(JoinId, CoreExpr)]
pairs
        body' :: CoreExpr
body'       = InScopeSet -> CoreExpr -> CoreExpr
go InScopeSet
in_scope' CoreExpr
body


-- | State Monad used inside `exitify`
type ExitifyM =  State [(JoinId, CoreExpr)]

-- | Given a recursive group of a joinrec, identifies “exit paths” and binds them as
--   join-points outside the joinrec.
exitifyRec :: InScopeSet -> [(Var,CoreExpr)] -> [CoreBind]
exitifyRec :: InScopeSet -> [(JoinId, CoreExpr)] -> CoreProgram
exitifyRec InScopeSet
in_scope [(JoinId, CoreExpr)]
pairs
  = [ JoinId -> CoreExpr -> Bind JoinId
forall b. b -> Expr b -> Bind b
NonRec JoinId
xid CoreExpr
rhs | (JoinId
xid,CoreExpr
rhs) <- [(JoinId, CoreExpr)]
exits ] CoreProgram -> CoreProgram -> CoreProgram
forall a. [a] -> [a] -> [a]
++ [[(JoinId, CoreExpr)] -> Bind JoinId
forall b. [(b, Expr b)] -> Bind b
Rec [(JoinId, CoreExpr)]
pairs']
  where
    -- We need the set of free variables of many subexpressions here, so
    -- annotate the AST with them
    -- see Note [Calculating free variables]
    ann_pairs :: [(JoinId, CoreExprWithFVs)]
ann_pairs = ((JoinId, CoreExpr) -> (JoinId, CoreExprWithFVs))
-> [(JoinId, CoreExpr)] -> [(JoinId, CoreExprWithFVs)]
forall a b. (a -> b) -> [a] -> [b]
map ((CoreExpr -> CoreExprWithFVs)
-> (JoinId, CoreExpr) -> (JoinId, CoreExprWithFVs)
forall b c a. (b -> c) -> (a, b) -> (a, c)
forall (p :: * -> * -> *) b c a.
Bifunctor p =>
(b -> c) -> p a b -> p a c
second CoreExpr -> CoreExprWithFVs
freeVars) [(JoinId, CoreExpr)]
pairs

    -- Which are the recursive calls?
    recursive_calls :: VarSet
recursive_calls = [JoinId] -> VarSet
mkVarSet ([JoinId] -> VarSet) -> [JoinId] -> VarSet
forall a b. (a -> b) -> a -> b
$ ((JoinId, CoreExpr) -> JoinId) -> [(JoinId, CoreExpr)] -> [JoinId]
forall a b. (a -> b) -> [a] -> [b]
map (JoinId, CoreExpr) -> JoinId
forall a b. (a, b) -> a
fst [(JoinId, CoreExpr)]
pairs

    ([(JoinId, CoreExpr)]
pairs',[(JoinId, CoreExpr)]
exits) = (State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
-> [(JoinId, CoreExpr)]
-> ([(JoinId, CoreExpr)], [(JoinId, CoreExpr)])
forall s a. State s a -> s -> (a, s)
`runState` []) (State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
 -> ([(JoinId, CoreExpr)], [(JoinId, CoreExpr)]))
-> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
-> ([(JoinId, CoreExpr)], [(JoinId, CoreExpr)])
forall a b. (a -> b) -> a -> b
$
        [(JoinId, CoreExprWithFVs)]
-> ((JoinId, CoreExprWithFVs)
    -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
-> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
forall (t :: * -> *) (m :: * -> *) a b.
(Traversable t, Monad m) =>
t a -> (a -> m b) -> m (t b)
forM [(JoinId, CoreExprWithFVs)]
ann_pairs (((JoinId, CoreExprWithFVs)
  -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
 -> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)])
-> ((JoinId, CoreExprWithFVs)
    -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
-> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
forall a b. (a -> b) -> a -> b
$ \(JoinId
x,CoreExprWithFVs
rhs) -> do
            -- go past the lambdas of the join point
            let ([JoinId]
args, CoreExprWithFVs
body) = Int -> CoreExprWithFVs -> ([JoinId], CoreExprWithFVs)
forall bndr annot.
Int -> AnnExpr bndr annot -> ([bndr], AnnExpr bndr annot)
collectNAnnBndrs (JoinId -> Int
idJoinArity JoinId
x) CoreExprWithFVs
rhs
            body' <- [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go [JoinId]
args CoreExprWithFVs
body
            let rhs' = [JoinId] -> CoreExpr -> CoreExpr
forall b. [b] -> Expr b -> Expr b
mkLams [JoinId]
args CoreExpr
body'
            return (x, rhs')

    ---------------------
    -- 'go' is the main working function.
    -- It goes through the RHS (tail-call positions only),
    -- checks if there are no more recursive calls, if so, abstracts over
    -- variables bound on the way and lifts it out as a join point.
    --
    -- ExitifyM is a state monad to keep track of floated binds
    go :: [Var]           -- Variables that are in-scope here, but
                          -- not in scope at the joinrec; that is,
                          -- we must potentially abstract over them.
                          -- Invariant: they are kept in dependency order
       -> CoreExprWithFVs -- Current expression in tail position
       -> ExitifyM CoreExpr

    -- We first look at the expression (no matter what it shape is)
    -- and determine if we can turn it into a exit join point
    go :: [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go [JoinId]
captured CoreExprWithFVs
ann_e
        | -- An exit expression has no recursive calls
          let fvs :: VarSet
fvs = DVarSet -> VarSet
dVarSetToVarSet (CoreExprWithFVs -> DVarSet
freeVarsOf CoreExprWithFVs
ann_e)
        , VarSet -> VarSet -> Bool
disjointVarSet VarSet
fvs VarSet
recursive_calls
        = [JoinId] -> CoreExpr -> VarSet -> ExitifyM CoreExpr
go_exit [JoinId]
captured (CoreExprWithFVs -> CoreExpr
forall bndr annot. AnnExpr bndr annot -> Expr bndr
deAnnotate CoreExprWithFVs
ann_e) VarSet
fvs

    -- We could not turn it into a exit join point. So now recurse
    -- into all expression where eligible exit join points might sit,
    -- i.e. into all tail-call positions:

    -- Case right hand sides are in tail-call position
    go [JoinId]
captured (DVarSet
_, AnnCase CoreExprWithFVs
scrut JoinId
bndr Type
ty [AnnAlt JoinId DVarSet]
alts) = do
        alts' <- [AnnAlt JoinId DVarSet]
-> (AnnAlt JoinId DVarSet
    -> State [(JoinId, CoreExpr)] (Alt JoinId))
-> State [(JoinId, CoreExpr)] [Alt JoinId]
forall (t :: * -> *) (m :: * -> *) a b.
(Traversable t, Monad m) =>
t a -> (a -> m b) -> m (t b)
forM [AnnAlt JoinId DVarSet]
alts ((AnnAlt JoinId DVarSet -> State [(JoinId, CoreExpr)] (Alt JoinId))
 -> State [(JoinId, CoreExpr)] [Alt JoinId])
-> (AnnAlt JoinId DVarSet
    -> State [(JoinId, CoreExpr)] (Alt JoinId))
-> State [(JoinId, CoreExpr)] [Alt JoinId]
forall a b. (a -> b) -> a -> b
$ \(AnnAlt AltCon
dc [JoinId]
pats CoreExprWithFVs
rhs) -> do
            rhs' <- [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go ([JoinId]
captured [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ [JoinId
bndr] [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ [JoinId]
pats) CoreExprWithFVs
rhs
            return (Alt dc pats rhs')
        return $ Case (deAnnotate scrut) bndr ty alts'

    go [JoinId]
captured (DVarSet
_, AnnLet AnnBind JoinId DVarSet
ann_bind CoreExprWithFVs
body)
        -- join point, RHS and body are in tail-call position
        | AnnNonRec JoinId
j CoreExprWithFVs
rhs <- AnnBind JoinId DVarSet
ann_bind
        , JoinPoint Int
join_arity <- JoinId -> JoinPointHood
idJoinPointHood JoinId
j
        = do let ([JoinId]
params, CoreExprWithFVs
join_body) = Int -> CoreExprWithFVs -> ([JoinId], CoreExprWithFVs)
forall bndr annot.
Int -> AnnExpr bndr annot -> ([bndr], AnnExpr bndr annot)
collectNAnnBndrs Int
join_arity CoreExprWithFVs
rhs
             join_body' <- [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go ([JoinId]
captured [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ [JoinId]
params) CoreExprWithFVs
join_body
             let rhs' = [JoinId] -> CoreExpr -> CoreExpr
forall b. [b] -> Expr b -> Expr b
mkLams [JoinId]
params CoreExpr
join_body'
             body' <- go (captured ++ [j]) body
             return $ Let (NonRec j rhs') body'

        -- rec join point, RHSs and body are in tail-call position
        | AnnRec [(JoinId, CoreExprWithFVs)]
pairs <- AnnBind JoinId DVarSet
ann_bind
        , JoinId -> Bool
isJoinId ((JoinId, CoreExprWithFVs) -> JoinId
forall a b. (a, b) -> a
fst ([(JoinId, CoreExprWithFVs)] -> (JoinId, CoreExprWithFVs)
forall a. HasCallStack => [a] -> a
head [(JoinId, CoreExprWithFVs)]
pairs))
        = do let js :: [JoinId]
js = ((JoinId, CoreExprWithFVs) -> JoinId)
-> [(JoinId, CoreExprWithFVs)] -> [JoinId]
forall a b. (a -> b) -> [a] -> [b]
map (JoinId, CoreExprWithFVs) -> JoinId
forall a b. (a, b) -> a
fst [(JoinId, CoreExprWithFVs)]
pairs
             pairs' <- [(JoinId, CoreExprWithFVs)]
-> ((JoinId, CoreExprWithFVs)
    -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
-> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
forall (t :: * -> *) (m :: * -> *) a b.
(Traversable t, Monad m) =>
t a -> (a -> m b) -> m (t b)
forM [(JoinId, CoreExprWithFVs)]
pairs (((JoinId, CoreExprWithFVs)
  -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
 -> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)])
-> ((JoinId, CoreExprWithFVs)
    -> State [(JoinId, CoreExpr)] (JoinId, CoreExpr))
-> State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
forall a b. (a -> b) -> a -> b
$ \(JoinId
j,CoreExprWithFVs
rhs) -> do
                 let join_arity :: Int
join_arity = JoinId -> Int
idJoinArity JoinId
j
                     ([JoinId]
params, CoreExprWithFVs
join_body) = Int -> CoreExprWithFVs -> ([JoinId], CoreExprWithFVs)
forall bndr annot.
Int -> AnnExpr bndr annot -> ([bndr], AnnExpr bndr annot)
collectNAnnBndrs Int
join_arity CoreExprWithFVs
rhs
                 join_body' <- [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go ([JoinId]
captured [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ [JoinId]
js [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ [JoinId]
params) CoreExprWithFVs
join_body
                 let rhs' = [JoinId] -> CoreExpr -> CoreExpr
forall b. [b] -> Expr b -> Expr b
mkLams [JoinId]
params CoreExpr
join_body'
                 return (j, rhs')
             body' <- go (captured ++ js) body
             return $ Let (Rec pairs') body'

        -- normal Let, only the body is in tail-call position
        | Bool
otherwise
        = do body' <- [JoinId] -> CoreExprWithFVs -> ExitifyM CoreExpr
go ([JoinId]
captured [JoinId] -> [JoinId] -> [JoinId]
forall a. [a] -> [a] -> [a]
++ Bind JoinId -> [JoinId]
forall b. Bind b -> [b]
bindersOf Bind JoinId
bind ) CoreExprWithFVs
body
             return $ Let bind body'
      where bind :: Bind JoinId
bind = AnnBind JoinId DVarSet -> Bind JoinId
forall b annot. AnnBind b annot -> Bind b
deAnnBind AnnBind JoinId DVarSet
ann_bind

    -- Cannot be turned into an exit join point, but also has no
    -- tail-call subexpression. Nothing to do here.
    go [JoinId]
_ CoreExprWithFVs
ann_e = CoreExpr -> ExitifyM CoreExpr
forall a. a -> State [(JoinId, CoreExpr)] a
forall (m :: * -> *) a. Monad m => a -> m a
return (CoreExprWithFVs -> CoreExpr
forall bndr annot. AnnExpr bndr annot -> Expr bndr
deAnnotate CoreExprWithFVs
ann_e)

    ---------------------
    go_exit :: [Var]      -- Variables captured locally
            -> CoreExpr   -- An exit expression
            -> VarSet     -- Free vars of the expression
            -> ExitifyM CoreExpr
    -- go_exit deals with a tail expression that is floatable
    -- out as an exit point; that is, it mentions no recursive calls
    go_exit :: [JoinId] -> CoreExpr -> VarSet -> ExitifyM CoreExpr
go_exit [JoinId]
captured CoreExpr
e VarSet
fvs
      -- Do not touch an expression that is already a join jump where all arguments
      -- are captured variables. See Note [Idempotency]
      -- But _do_ float join jumps with interesting arguments.
      -- See Note [Jumps can be interesting]
      | (Var JoinId
f, [CoreExpr]
args) <- CoreExpr -> (CoreExpr, [CoreExpr])
forall b. Expr b -> (Expr b, [Expr b])
collectArgs CoreExpr
e
      , JoinId -> Bool
isJoinId JoinId
f
      , (CoreExpr -> Bool) -> [CoreExpr] -> Bool
forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool
all CoreExpr -> Bool
isCapturedVarArg [CoreExpr]
args
      = CoreExpr -> ExitifyM CoreExpr
forall a. a -> State [(JoinId, CoreExpr)] a
forall (m :: * -> *) a. Monad m => a -> m a
return CoreExpr
e

      -- Do not touch a boring expression (see Note [Interesting expression])
      | Bool -> Bool
not Bool
is_interesting
      = CoreExpr -> ExitifyM CoreExpr
forall a. a -> State [(JoinId, CoreExpr)] a
forall (m :: * -> *) a. Monad m => a -> m a
return CoreExpr
e

      -- Cannot float out if local join points are used, as
      -- we cannot abstract over them
      | Bool
captures_join_points
      = CoreExpr -> ExitifyM CoreExpr
forall a. a -> State [(JoinId, CoreExpr)] a
forall (m :: * -> *) a. Monad m => a -> m a
return CoreExpr
e

      -- We have something to float out!
      | Bool
otherwise
      = do { -- Assemble the RHS of the exit join point
             let rhs :: CoreExpr
rhs   = [JoinId] -> CoreExpr -> CoreExpr
forall b. [b] -> Expr b -> Expr b
mkLams [JoinId]
abs_vars CoreExpr
e
                 avoid :: InScopeSet
avoid = InScopeSet
in_scope InScopeSet -> [JoinId] -> InScopeSet
`extendInScopeSetList` [JoinId]
captured
             -- Remember this binding under a suitable name
           ; v <- InScopeSet -> Int -> CoreExpr -> ExitifyM JoinId
addExit InScopeSet
avoid ([JoinId] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [JoinId]
abs_vars) CoreExpr
rhs
             -- And jump to it from here
           ; return $ mkVarApps (Var v) abs_vars }

      where
        -- Used to detect exit expressions that are already proper exit jumps
        isCapturedVarArg :: CoreExpr -> Bool
isCapturedVarArg (Var JoinId
v) = JoinId
v JoinId -> [JoinId] -> Bool
forall a. Eq a => a -> [a] -> Bool
forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool
`elem` [JoinId]
captured
        isCapturedVarArg CoreExpr
_ = Bool
False

        -- An interesting exit expression has free, non-imported
        -- variables from outside the recursive group
        -- See Note [Interesting expression]
        is_interesting :: Bool
is_interesting = (JoinId -> Bool) -> VarSet -> Bool
anyVarSet JoinId -> Bool
isLocalId (VarSet -> Bool) -> VarSet -> Bool
forall a b. (a -> b) -> a -> b
$
                         VarSet
fvs VarSet -> VarSet -> VarSet
`minusVarSet` [JoinId] -> VarSet
mkVarSet [JoinId]
captured

        -- The arguments of this exit join point
        -- See Note [Picking arguments to abstract over]
        abs_vars :: [JoinId]
abs_vars = (VarSet, [JoinId]) -> [JoinId]
forall a b. (a, b) -> b
snd ((VarSet, [JoinId]) -> [JoinId]) -> (VarSet, [JoinId]) -> [JoinId]
forall a b. (a -> b) -> a -> b
$ (JoinId -> (VarSet, [JoinId]) -> (VarSet, [JoinId]))
-> (VarSet, [JoinId]) -> [JoinId] -> (VarSet, [JoinId])
forall a b. (a -> b -> b) -> b -> [a] -> b
forall (t :: * -> *) a b.
Foldable t =>
(a -> b -> b) -> b -> t a -> b
foldr JoinId -> (VarSet, [JoinId]) -> (VarSet, [JoinId])
pick (VarSet
fvs, []) [JoinId]
captured
          where
            pick :: JoinId -> (VarSet, [JoinId]) -> (VarSet, [JoinId])
pick JoinId
v (VarSet
fvs', [JoinId]
acc) | JoinId
v JoinId -> VarSet -> Bool
`elemVarSet` VarSet
fvs' = (VarSet
fvs' VarSet -> JoinId -> VarSet
`delVarSet` JoinId
v, JoinId -> JoinId
zap JoinId
v JoinId -> [JoinId] -> [JoinId]
forall a. a -> [a] -> [a]
: [JoinId]
acc)
                               | Bool
otherwise           = (VarSet
fvs',               [JoinId]
acc)

        -- We are going to abstract over these variables, so we must
        -- zap any IdInfo they have; see #15005
        -- cf. GHC.Core.Opt.SetLevels.abstractVars
        zap :: JoinId -> JoinId
zap JoinId
v | JoinId -> Bool
isId JoinId
v = JoinId -> IdInfo -> JoinId
setIdInfo JoinId
v IdInfo
vanillaIdInfo
              | Bool
otherwise = JoinId
v

        -- We cannot abstract over join points
        captures_join_points :: Bool
captures_join_points = (JoinId -> Bool) -> [JoinId] -> Bool
forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool
any JoinId -> Bool
isJoinId [JoinId]
abs_vars


-- Picks a new unique, which is disjoint from
--  * the free variables of the whole joinrec
--  * any bound variables (captured)
--  * any exit join points created so far.
mkExitJoinId :: InScopeSet -> Type -> JoinArity -> ExitifyM JoinId
mkExitJoinId :: InScopeSet -> Type -> Int -> ExitifyM JoinId
mkExitJoinId InScopeSet
in_scope Type
ty Int
join_arity = do
    fs <- State [(JoinId, CoreExpr)] [(JoinId, CoreExpr)]
forall s. State s s
get
    let avoid = InScopeSet
in_scope InScopeSet -> [JoinId] -> InScopeSet
`extendInScopeSetList` (((JoinId, CoreExpr) -> JoinId) -> [(JoinId, CoreExpr)] -> [JoinId]
forall a b. (a -> b) -> [a] -> [b]
map (JoinId, CoreExpr) -> JoinId
forall a b. (a, b) -> a
fst [(JoinId, CoreExpr)]
fs)
                         InScopeSet -> JoinId -> InScopeSet
`extendInScopeSet` JoinId
exit_id_tmpl -- just cosmetics
    return (uniqAway avoid exit_id_tmpl)
  where
    exit_id_tmpl :: JoinId
exit_id_tmpl = FastString -> Unique -> Type -> Type -> JoinId
mkSysLocal (String -> FastString
fsLit String
"exit") Unique
initExitJoinUnique Type
ManyTy Type
ty
                    JoinId -> Int -> JoinId
`asJoinId` Int
join_arity

addExit :: InScopeSet -> JoinArity -> CoreExpr -> ExitifyM JoinId
addExit :: InScopeSet -> Int -> CoreExpr -> ExitifyM JoinId
addExit InScopeSet
in_scope Int
join_arity CoreExpr
rhs = do
    -- Pick a suitable name
    let ty :: Type
ty = HasDebugCallStack => CoreExpr -> Type
CoreExpr -> Type
exprType CoreExpr
rhs
    v <- InScopeSet -> Type -> Int -> ExitifyM JoinId
mkExitJoinId InScopeSet
in_scope Type
ty Int
join_arity
    fs <- get
    put ((v,rhs):fs)
    return v

{-
Note [Interesting expression]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We do not want this to happen:

  joinrec go 0     x y = x
          go (n-1) x y = jump go (n-1) (x+y)
  in …
==>
  join exit x = x
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)
  in …

because the floated exit path (`x`) is simply a parameter of `go`; there are
not useful interactions exposed this way.

Neither do we want this to happen

  joinrec go 0     x y = x+x
          go (n-1) x y = jump go (n-1) (x+y)
  in …
==>
  join exit x = x+x
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)
  in …

where the floated expression `x+x` is a bit more complicated, but still not
interesting.

Expressions are interesting when they move an occurrence of a variable outside
the recursive `go` that can benefit from being obviously called once, for example:
 * a local thunk that can then be inlined (see example in Note [Exitification])
 * the parameter of a function, where the demand analyzer then can then
   see that it is called at most once, and hence improve the function’s
   strictness signature

So we only hoist an exit expression out if it mentions at least one free,
non-imported variable.

Note [Jumps can be interesting]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
A jump to a join point can be interesting, if its arguments contain free
non-exported variables (z in the following example):

  joinrec go 0     x y = jump j (x+z)
          go (n-1) x y = jump go (n-1) (x+y)
  in …
==>
  join exit x y = jump j (x+z)
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)


The join point itself can be interesting, even if none if its
arguments have free variables free in the joinrec.  For example

  join j p = case p of (x,y) -> x+y
  joinrec go 0     x y = jump j (x,y)
          go (n-1) x y = jump go (n-1) (x+y) y
  in …

Here, `j` would not be inlined because we do not inline something that looks
like an exit join point (see Note [Do not inline exit join points]). But
if we exitify the 'jump j (x,y)' we get

  join j p = case p of (x,y) -> x+y
  join exit x y = jump j (x,y)
  joinrec go 0     x y = jump exit x y
          go (n-1) x y = jump go (n-1) (x+y) y
  in …

and now 'j' can inline, and we get rid of the pair. Here's another
example (assume `g` to be an imported function that, on its own,
does not make this interesting):

  join j y = map f y
  joinrec go 0     x y = jump j (map g x)
          go (n-1) x y = jump go (n-1) (x+y)
  in …

Again, `j` would not be inlined because we do not inline something that looks
like an exit join point (see Note [Do not inline exit join points]).

But after exitification we have

  join j y = map f y
  join exit x = jump j (map g x)
  joinrec go 0     x y = jump j (map g x)
              go (n-1) x y = jump go (n-1) (x+y)
  in …

and now we can inline `j` and this will allow `map/map` to fire.


Note [Idempotency]
~~~~~~~~~~~~~~~~~~

We do not want this to happen, where we replace the floated expression with
essentially the same expression:

  join exit x = t (x*x)
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)
  in …
==>
  join exit x = t (x*x)
  join exit' x = jump exit x
  joinrec go 0     x y = jump exit' x
          go (n-1) x y = jump go (n-1) (x+y)
  in …

So when the RHS is a join jump, and all of its arguments are captured variables,
then we leave it in place.

Note that `jump exit x` in this example looks interesting, as `exit` is a free
variable. Therefore, idempotency does not simply follow from floating only
interesting expressions.

Note [Calculating free variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We have two options where to annotate the tree with free variables:

 A) The whole tree.
 B) Each individual joinrec as we come across it.

Downside of A: We pay the price on the whole module, even outside any joinrecs.
Downside of B: We pay the price per joinrec, possibly multiple times when
joinrecs are nested.

Further downside of A: If the exitify function returns annotated expressions,
it would have to ensure that the annotations are correct.

We therefore choose B, and calculate the free variables in `exitify`.


Note [Do not inline exit join points]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When we have

  let t = foo bar
  join exit x = t (x*x)
  joinrec go 0     x y = jump exit x
          go (n-1) x y = jump go (n-1) (x+y)
  in …

we do not want the simplifier to simply inline `exit` back in (which it happily
would).

To prevent this, we need to recognize exit join points, and then disable
inlining.

Exit join points, recognizable using `isExitJoinId` are join points with an
occurrence in a recursive group, and can be recognized (after the occurrence
analyzer ran!) using `isExitJoinId`.
This function detects joinpoints with `occ_in_lam (idOccinfo id) == True`,
because the lambdas of a non-recursive join point are not considered for
`occ_in_lam`.  For example, in the following code, `j1` is /not/ marked
occ_in_lam, because `j2` is called only once.

  join j1 x = x+1
  join j2 y = join j1 (y+2)

To prevent inlining, we check for isExitJoinId
* In `preInlineUnconditionally` directly.
* In `simplLetUnfolding` we simply give exit join points no unfolding, which
  prevents inlining in `postInlineUnconditionally` and call sites.

Note [Placement of the exitification pass]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
I (Joachim) experimented with multiple positions for the Exitification pass in
the Core2Core pipeline:

 A) Before the `simpl_phases`
 B) Between the `simpl_phases` and the "main" simplifier pass
 C) After demand_analyser
 D) Before the final simplification phase

Here is the table (this is without inlining join exit points in the final
simplifier run):

        Program |                       Allocs                      |                      Instrs
                | ABCD.log     A.log     B.log     C.log     D.log  | ABCD.log     A.log     B.log     C.log     D.log
----------------|---------------------------------------------------|-------------------------------------------------
 fannkuch-redux |   -99.9%     +0.0%    -99.9%    -99.9%    -99.9%  |    -3.9%     +0.5%     -3.0%     -3.9%     -3.9%
          fasta |    -0.0%     +0.0%     +0.0%     -0.0%     -0.0%  |    -8.5%     +0.0%     +0.0%     -0.0%     -8.5%
            fem |     0.0%      0.0%      0.0%      0.0%     +0.0%  |    -2.2%     -0.1%     -0.1%     -2.1%     -2.1%
           fish |     0.0%      0.0%      0.0%      0.0%     +0.0%  |    -3.1%     +0.0%     -1.1%     -1.1%     -0.0%
   k-nucleotide |   -91.3%    -91.0%    -91.0%    -91.3%    -91.3%  |    -6.3%    +11.4%    +11.4%     -6.3%     -6.2%
            scs |    -0.0%     -0.0%     -0.0%     -0.0%     -0.0%  |    -3.4%     -3.0%     -3.1%     -3.3%     -3.3%
         simple |    -6.0%      0.0%     -6.0%     -6.0%     +0.0%  |    -3.4%     +0.0%     -5.2%     -3.4%     -0.1%
  spectral-norm |    -0.0%      0.0%      0.0%     -0.0%     +0.0%  |    -2.7%     +0.0%     -2.7%     -5.4%     -5.4%
----------------|---------------------------------------------------|-------------------------------------------------
            Min |   -95.0%    -91.0%    -95.0%    -95.0%    -95.0%  |    -8.5%     -3.0%     -5.2%     -6.3%     -8.5%
            Max |    +0.2%     +0.2%     +0.2%     +0.2%     +1.5%  |    +0.4%    +11.4%    +11.4%     +0.4%     +1.5%
 Geometric Mean |    -4.7%     -2.1%     -4.7%     -4.7%     -4.6%  |    -0.4%     +0.1%     -0.1%     -0.3%     -0.2%

Position A is disqualified, as it does not get rid of the allocations in
fannkuch-redux.
Position A and B are disqualified because it increases instructions in k-nucleotide.
Positions C and D have their advantages: C decreases allocations in simpl, but D instructions in fasta.

Assuming we have a budget of _one_ run of Exitification, then C wins (but we
could get more from running it multiple times, as seen in fish).

Note [Picking arguments to abstract over]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

When we create an exit join point, so we need to abstract over those of its
free variables that are be out-of-scope at the destination of the exit join
point. So we go through the list `captured` and pick those that are actually
free variables of the join point.

We do not just `filter (`elemVarSet` fvs) captured`, as there might be
shadowing, and `captured` may contain multiple variables with the same Unique. I
these cases we want to abstract only over the last occurrence, hence the `foldr`
(with emphasis on the `r`). This is #15110.

-}