{-
(c) The GRASP/AQUA Project, Glasgow University, 1993-1998


                        -----------------
                        A demand analysis
                        -----------------
-}


module GHC.Core.Opt.DmdAnal
   ( DmdAnalOpts(..)
   , dmdAnalProgram
   )
where

import GHC.Prelude

import GHC.Types.Demand   -- All of it

import GHC.Core
import GHC.Core.DataCon
import GHC.Core.Utils
import GHC.Core.TyCon
import GHC.Core.Type
import GHC.Core.Predicate( isEqualityClass, isCTupleClass )
import GHC.Core.FVs      ( rulesRhsFreeIds, bndrRuleAndUnfoldingIds )
import GHC.Core.Coercion ( Coercion )
import GHC.Core.TyCo.FVs     ( coVarsOfCos )
import GHC.Core.TyCo.Compare ( eqType )
import GHC.Core.Multiplicity ( scaledThing )
import GHC.Core.FamInstEnv
import GHC.Core.Opt.Arity ( typeArity )
import GHC.Core.Opt.WorkWrap.Utils

import GHC.Builtin.PrimOps
import GHC.Builtin.Types.Prim ( realWorldStatePrimTy )

import GHC.Types.Unique.Set
import GHC.Types.Unique.MemoFun
import GHC.Types.RepType
import GHC.Types.ForeignCall ( isSafeForeignCall )
import GHC.Types.Id
import GHC.Types.Var.Env
import GHC.Types.Var.Set
import GHC.Types.Basic

import GHC.Utils.Misc
import GHC.Utils.Panic
import GHC.Utils.Panic.Plain
import GHC.Utils.Outputable

import Data.List        ( mapAccumL )

{-
************************************************************************
*                                                                      *
\subsection{Top level stuff}
*                                                                      *
************************************************************************
-}

-- | Options for the demand analysis
data DmdAnalOpts = DmdAnalOpts
   { DmdAnalOpts -> Bool
dmd_strict_dicts    :: !Bool
   -- ^ Value of `-fdicts-strict` (on by default).
   -- When set, all functons are implicitly strict in dictionary args.
   , DmdAnalOpts -> Bool
dmd_do_boxity       :: !Bool
   -- ^ Governs whether the analysis should update boxity signatures.
   -- See Note [Don't change boxity without worker/wrapper].
   , DmdAnalOpts -> Arity
dmd_unbox_width     :: !Int
   -- ^ Value of `-fdmd-unbox-width`.
   -- See Note [Unboxed demand on function bodies returning small products]
   , DmdAnalOpts -> Arity
dmd_max_worker_args :: !Int
   -- ^ Value of `-fmax-worker-args`.
   -- Don't unbox anything if we end up with more than this many args.
   }

-- This is a strict alternative to (,)
-- See Note [Space Leaks in Demand Analysis]
data WithDmdType a = WithDmdType !DmdType !a

getAnnotated :: WithDmdType a -> a
getAnnotated :: forall a. WithDmdType a -> a
getAnnotated (WithDmdType DmdType
_ a
a) = a
a

data DmdResult a b = R !a !b

-- | Outputs a new copy of the Core program in which binders have been annotated
-- with demand and strictness information.
--
-- Note: use `seqBinds` on the result to avoid leaks due to lazyness (cf Note
-- [Stamp out space leaks in demand analysis])
dmdAnalProgram :: DmdAnalOpts -> FamInstEnvs -> [CoreRule] -> CoreProgram -> CoreProgram
dmdAnalProgram :: DmdAnalOpts
-> FamInstEnvs -> [CoreRule] -> CoreProgram -> CoreProgram
dmdAnalProgram DmdAnalOpts
opts FamInstEnvs
fam_envs [CoreRule]
rules CoreProgram
binds
  = WithDmdType CoreProgram -> CoreProgram
forall a. WithDmdType a -> a
getAnnotated (WithDmdType CoreProgram -> CoreProgram)
-> WithDmdType CoreProgram -> CoreProgram
forall a b. (a -> b) -> a -> b
$ AnalEnv -> CoreProgram -> WithDmdType CoreProgram
go (DmdAnalOpts -> FamInstEnvs -> AnalEnv
emptyAnalEnv DmdAnalOpts
opts FamInstEnvs
fam_envs) CoreProgram
binds
  where
    -- See Note [Analysing top-level bindings]
    -- and Note [Why care for top-level demand annotations?]
    go :: AnalEnv -> CoreProgram -> WithDmdType CoreProgram
go AnalEnv
_   []     = DmdType -> CoreProgram -> WithDmdType CoreProgram
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
nopDmdType []
    go AnalEnv
env (Bind Id
b:CoreProgram
bs) = WithDmdType (DmdResult (Bind Id) CoreProgram)
-> WithDmdType CoreProgram
forall b. WithDmdType (DmdResult b [b]) -> WithDmdType [b]
cons_up (WithDmdType (DmdResult (Bind Id) CoreProgram)
 -> WithDmdType CoreProgram)
-> WithDmdType (DmdResult (Bind Id) CoreProgram)
-> WithDmdType CoreProgram
forall a b. (a -> b) -> a -> b
$ TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType CoreProgram)
-> WithDmdType (DmdResult (Bind Id) CoreProgram)
forall a.
TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBind TopLevelFlag
TopLevel AnalEnv
env SubDemand
topSubDmd Bind Id
b AnalEnv -> WithDmdType CoreProgram
anal_body
      where
        anal_body :: AnalEnv -> WithDmdType CoreProgram
anal_body AnalEnv
env'
          | WithDmdType DmdType
body_ty CoreProgram
bs' <- AnalEnv -> CoreProgram -> WithDmdType CoreProgram
go AnalEnv
env' CoreProgram
bs
          = DmdType -> CoreProgram -> WithDmdType CoreProgram
forall a. DmdType -> a -> WithDmdType a
WithDmdType (DmdType
body_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` AnalEnv -> [Id] -> DmdEnv
keep_alive_roots AnalEnv
env' (Bind Id -> [Id]
forall b. Bind b -> [b]
bindersOf Bind Id
b)) CoreProgram
bs'

    cons_up :: WithDmdType (DmdResult b [b]) -> WithDmdType [b]
    cons_up :: forall b. WithDmdType (DmdResult b [b]) -> WithDmdType [b]
cons_up (WithDmdType DmdType
dmd_ty (R b
b' [b]
bs')) = DmdType -> [b] -> WithDmdType [b]
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
dmd_ty (b
b' b -> [b] -> [b]
forall a. a -> [a] -> [a]
: [b]
bs')

    keep_alive_roots :: AnalEnv -> [Id] -> DmdEnv
    -- See Note [Absence analysis for stable unfoldings and RULES]
    -- Here we keep alive "roots", e.g., exported ids and stuff mentioned in
    -- orphan RULES
    keep_alive_roots :: AnalEnv -> [Id] -> DmdEnv
keep_alive_roots AnalEnv
env [Id]
ids = [DmdEnv] -> DmdEnv
plusDmdEnvs ((Id -> DmdEnv) -> [Id] -> [DmdEnv]
forall a b. (a -> b) -> [a] -> [b]
map (AnalEnv -> Id -> DmdEnv
demandRoot AnalEnv
env) ((Id -> Bool) -> [Id] -> [Id]
forall a. (a -> Bool) -> [a] -> [a]
filter Id -> Bool
is_root [Id]
ids))

    is_root :: Id -> Bool
    is_root :: Id -> Bool
is_root Id
id = Id -> Bool
isExportedId Id
id Bool -> Bool -> Bool
|| Id -> VarSet -> Bool
elemVarSet Id
id VarSet
rule_fvs

    rule_fvs :: IdSet
    rule_fvs :: VarSet
rule_fvs = [CoreRule] -> VarSet
rulesRhsFreeIds [CoreRule]
rules

demandRoot :: AnalEnv -> Id -> DmdEnv
-- See Note [Absence analysis for stable unfoldings and RULES]
demandRoot :: AnalEnv -> Id -> DmdEnv
demandRoot AnalEnv
env Id
id = (DmdEnv, CoreExpr) -> DmdEnv
forall a b. (a, b) -> a
fst (AnalEnv -> Demand -> CoreExpr -> (DmdEnv, CoreExpr)
dmdAnalStar AnalEnv
env Demand
topDmd (Id -> CoreExpr
forall b. Id -> Expr b
Var Id
id))

demandRoots :: AnalEnv -> [Id] -> DmdEnv
-- See Note [Absence analysis for stable unfoldings and RULES]
demandRoots :: AnalEnv -> [Id] -> DmdEnv
demandRoots AnalEnv
env [Id]
roots = [DmdEnv] -> DmdEnv
plusDmdEnvs ((Id -> DmdEnv) -> [Id] -> [DmdEnv]
forall a b. (a -> b) -> [a] -> [b]
map (AnalEnv -> Id -> DmdEnv
demandRoot AnalEnv
env) [Id]
roots)

demandRootSet :: AnalEnv -> IdSet -> DmdEnv
demandRootSet :: AnalEnv -> VarSet -> DmdEnv
demandRootSet AnalEnv
env VarSet
ids = AnalEnv -> [Id] -> DmdEnv
demandRoots AnalEnv
env (VarSet -> [Id]
forall elt. UniqSet elt -> [elt]
nonDetEltsUniqSet VarSet
ids)
  -- It's OK to use nonDetEltsUniqSet here because plusDmdType is commutative

-- | We attach useful (e.g. not 'topDmd') 'idDemandInfo' to top-level bindings
-- that satisfy this function.
--
-- Basically, we want to know how top-level *functions* are *used*
-- (e.g. called). The information will always be lazy.
-- Any other top-level bindings are boring.
--
-- See also Note [Why care for top-level demand annotations?].
isInterestingTopLevelFn :: Id -> Bool
-- SG tried to set this to True and got a +2% ghc/alloc regression in T5642
-- (which is dominated by the Simplifier) at no gain in analysis precision.
-- If there was a gain, that regression might be acceptable.
-- Plus, we could use LetUp for thunks and share some code with local let
-- bindings.
isInterestingTopLevelFn :: Id -> Bool
isInterestingTopLevelFn Id
id = Type -> Arity
typeArity (Id -> Type
idType Id
id) Arity -> Arity -> Bool
forall a. Ord a => a -> a -> Bool
> Arity
0

{- Note [Stamp out space leaks in demand analysis]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The demand analysis pass outputs a new copy of the Core program in
which binders have been annotated with demand and strictness
information. It's tiresome to ensure that this information is fully
evaluated everywhere that we produce it, so we just run a single
seqBinds over the output before returning it, to ensure that there are
no references holding on to the input Core program.

This makes a ~30% reduction in peak memory usage when compiling
DynFlags (cf #9675 and #13426).

This is particularly important when we are doing late demand analysis,
since we don't do a seqBinds at any point thereafter. Hence code
generation would hold on to an extra copy of the Core program, via
unforced thunks in demand or strictness information; and it is the
most memory-intensive part of the compilation process, so this added
seqBinds makes a big difference in peak memory usage.

Note [Don't change boxity without worker/wrapper]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider (T21754)
  f n = n+1
  {-# NOINLINE f #-}
With `-fno-worker-wrapper`, we should not give `f` a boxity signature that says
that it unboxes its argument! Client modules would never be able to cancel away
the box for n. Likewise we shouldn't give `f` the CPR property.

Similarly, in the last run of DmdAnal before codegen (which does not have a
worker/wrapper phase) we should not change boxity in any way. Remember: an
earlier result of the demand analyser, complete with worker/wrapper, has aleady
given a demand signature (with boxity info) to the function.
(The "last run" is mainly there to attach demanded-once info to let-bindings.)

In general, we should not run Note [Boxity analysis] unless worker/wrapper
follows to exploit the boxity and make sure that calling modules can observe the
reported boxity.

Hence DmdAnal is configured by a flag `dmd_do_boxity` that is True only
if worker/wrapper follows after DmdAnal. If it is not set, and the signature
is not subject to Note [Boxity for bottoming functions], DmdAnal tries
to transfer over the previous boxity to the new demand signature, in
`setIdDmdAndBoxSig`.

Why isn't CprAnal configured with a similar flag? Because if we aren't going to
do worker/wrapper we don't run CPR analysis at all. (see GHC.Core.Opt.Pipeline)

It might be surprising that we only try to preserve *arg* boxity, not boxity on
FVs. But FV demands won't make it into interface files anyway, so it's a waste
of energy.
Besides, W/W zaps the `DmdEnv` portion of a signature, so we don't know the old
boxity to begin with; see Note [Zapping DmdEnv after Demand Analyzer].

Note [Analysing top-level bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider a CoreProgram like
  e1 = ...
  n1 = ...
  e2 = \a b -> ... fst (n1 a b) ...
  n2 = \c d -> ... snd (e2 c d) ...
  ...
where e* are exported, but n* are not.
Intuitively, we can see that @n1@ is only ever called with two arguments
and in every call site, the first component of the result of the call
is evaluated. Thus, we'd like it to have idDemandInfo @LC(L,C(M,P(1L,A))@.
NB: We may *not* give e2 a similar annotation, because it is exported and
external callers might use it in arbitrary ways, expressed by 'topDmd'.
This can then be exploited by Nested CPR and eta-expansion,
see Note [Why care for top-level demand annotations?].

How do we get this result? Answer: By analysing the program as if it was a let
expression of this form:
  let e1 = ... in
  let n1 = ... in
  let e2 = ... in
  let n2 = ... in
  (e1,e2, ...)
E.g. putting all bindings in nested lets and returning all exported binders in a tuple.
Of course, we will not actually build that CoreExpr! Instead we faithfully
simulate analysis of said expression by adding the free variable 'DmdEnv'
of @e*@'s strictness signatures to the 'DmdType' we get from analysing the
nested bindings.

And even then the above form blows up analysis performance in T10370:
If @e1@ uses many free variables, we'll unnecessarily carry their demands around
with us from the moment we analyse the pair to the moment we bubble back up to
the binding for @e1@. So instead we analyse as if we had
  let e1 = ... in
  (e1, let n1 = ... in
  (    let e2 = ... in
  (e2, let n2 = ... in
  (    ...))))
That is, a series of right-nested pairs, where the @fst@ are the exported
binders of the last enclosing let binding and @snd@ continues the nested
lets.

Variables occurring free in RULE RHSs are to be handled the same as exported Ids.
See also Note [Absence analysis for stable unfoldings and RULES].

Note [Why care for top-level demand annotations?]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Reading Note [Analysing top-level bindings], you might think that we go through
quite some trouble to get useful demands for top-level bindings. They can never
be strict, for example, so why bother?

First, we get to eta-expand top-level bindings that we weren't able to
eta-expand before without Call Arity. From T18894b:
  module T18894b (f) where
  eta :: Int -> Int -> Int
  eta x = if fst (expensive x) == 13 then \y -> ... else \y -> ...
  f m = ... eta m 2 ... eta 2 m ...
Since only @f@ is exported, we see all call sites of @eta@ and can eta-expand to
arity 2.

The call demands we get for some top-level bindings will also allow Nested CPR
to unbox deeper. From T18894:
  module T18894 (h) where
  g m n = (2 * m, 2 `div` n)
  {-# NOINLINE g #-}
  h :: Int -> Int
  h m = ... snd (g m 2) ... uncurry (+) (g 2 m) ...
Only @h@ is exported, hence we see that @g@ is always called in contexts were we
also force the division in the second component of the pair returned by @g@.
This allows Nested CPR to evaluate the division eagerly and return an I# in its
position.
-}

{-
************************************************************************
*                                                                      *
\subsection{The analyser itself}
*                                                                      *
************************************************************************
-}

-- | Analyse a binding group and its \"body\", e.g. where it is in scope.
--
-- It calls a function that knows how to analyse this \"body\" given
-- an 'AnalEnv' with updated demand signatures for the binding group
-- (reflecting their 'idDmdSigInfo') and expects to receive a
-- 'DmdType' in return, which it uses to annotate the binding group with their
-- 'idDemandInfo'.
dmdAnalBind
  :: TopLevelFlag
  -> AnalEnv
  -> SubDemand                 -- ^ Demand put on the "body"
                               --   (important for join points)
  -> CoreBind
  -> (AnalEnv -> WithDmdType a) -- ^ How to analyse the "body", e.g.
                               --   where the binding is in scope
  -> WithDmdType (DmdResult CoreBind a)
dmdAnalBind :: forall a.
TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBind TopLevelFlag
top_lvl AnalEnv
env SubDemand
dmd Bind Id
bind AnalEnv -> WithDmdType a
anal_body = case Bind Id
bind of
  NonRec Id
id CoreExpr
rhs
    | TopLevelFlag -> Id -> Bool
useLetUp TopLevelFlag
top_lvl Id
id
    -> TopLevelFlag
-> AnalEnv
-> Id
-> CoreExpr
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
forall a.
TopLevelFlag
-> AnalEnv
-> Id
-> CoreExpr
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBindLetUp   TopLevelFlag
top_lvl AnalEnv
env_rhs     Id
id CoreExpr
rhs AnalEnv -> WithDmdType a
anal_body
  Bind Id
_ -> TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
forall a.
TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBindLetDown TopLevelFlag
top_lvl AnalEnv
env_rhs SubDemand
dmd Bind Id
bind   AnalEnv -> WithDmdType a
anal_body
  where
    env_rhs :: AnalEnv
env_rhs = Bind Id -> AnalEnv -> AnalEnv
enterDFun Bind Id
bind AnalEnv
env

-- | Annotates uninteresting top level functions ('isInterestingTopLevelFn')
-- with 'topDmd', the rest with the given demand.
setBindIdDemandInfo :: TopLevelFlag -> Id -> Demand -> Id
setBindIdDemandInfo :: TopLevelFlag -> Id -> Demand -> Id
setBindIdDemandInfo TopLevelFlag
top_lvl Id
id Demand
dmd = Id -> Demand -> Id
setIdDemandInfo Id
id (Demand -> Id) -> Demand -> Id
forall a b. (a -> b) -> a -> b
$ case TopLevelFlag
top_lvl of
  TopLevelFlag
TopLevel | Bool -> Bool
not (Id -> Bool
isInterestingTopLevelFn Id
id) -> Demand
topDmd
  TopLevelFlag
_                                           -> Demand
dmd

-- | Update the demand signature, but be careful not to change boxity info if
-- `dmd_do_boxity` is True or if the signature is bottom.
-- See Note [Don't change boxity without worker/wrapper]
-- and Note [Boxity for bottoming functions].
setIdDmdAndBoxSig :: DmdAnalOpts -> Id -> DmdSig -> Id
setIdDmdAndBoxSig :: DmdAnalOpts -> Id -> DmdSig -> Id
setIdDmdAndBoxSig DmdAnalOpts
opts Id
id DmdSig
sig = Id -> DmdSig -> Id
setIdDmdSig Id
id (DmdSig -> Id) -> DmdSig -> Id
forall a b. (a -> b) -> a -> b
$
  if DmdAnalOpts -> Bool
dmd_do_boxity DmdAnalOpts
opts Bool -> Bool -> Bool
|| DmdSig -> Bool
isBottomingSig DmdSig
sig
    then DmdSig
sig
    else DmdSig -> DmdSig -> DmdSig
transferArgBoxityDmdSig (Id -> DmdSig
idDmdSig Id
id) DmdSig
sig

-- | Let bindings can be processed in two ways:
-- Down (RHS before body) or Up (body before RHS).
-- This function handles the up variant.
--
-- It is very simple. For  let x = rhs in body
--   * Demand-analyse 'body' in the current environment
--   * Find the demand, 'rhs_dmd' placed on 'x' by 'body'
--   * Demand-analyse 'rhs' in 'rhs_dmd'
--
-- This is used for a non-recursive local let without manifest lambdas (see
-- 'useLetUp').
--
-- This is the LetUp rule in the paper “Higher-Order Cardinality Analysis”.
dmdAnalBindLetUp :: TopLevelFlag
                 -> AnalEnv
                 -> Id
                 -> CoreExpr
                 -> (AnalEnv -> WithDmdType a)
                 -> WithDmdType (DmdResult CoreBind a)
dmdAnalBindLetUp :: forall a.
TopLevelFlag
-> AnalEnv
-> Id
-> CoreExpr
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBindLetUp TopLevelFlag
top_lvl AnalEnv
env Id
id CoreExpr
rhs AnalEnv -> WithDmdType a
anal_body = DmdType
-> DmdResult (Bind Id) a -> WithDmdType (DmdResult (Bind Id) a)
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
final_ty (Bind Id -> a -> DmdResult (Bind Id) a
forall a b. a -> b -> DmdResult a b
R (Id -> CoreExpr -> Bind Id
forall b. b -> Expr b -> Bind b
NonRec Id
id' CoreExpr
rhs') (a
body'))
  where
    WithDmdType DmdType
body_ty a
body'   = AnalEnv -> WithDmdType a
anal_body (AnalEnv -> Id -> AnalEnv
addInScopeAnalEnv AnalEnv
env Id
id)
    -- See Note [Bringing a new variable into scope]
    WithDmdType DmdType
body_ty' Demand
id_dmd = AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
body_ty Id
id
    -- See Note [Finalising boxity for demand signatures]

    id_dmd' :: Demand
id_dmd'            = AnalEnv -> Type -> Demand -> Demand
finaliseLetBoxity AnalEnv
env (Id -> Type
idType Id
id) Demand
id_dmd
    !id' :: Id
id'               = TopLevelFlag -> Id -> Demand -> Id
setBindIdDemandInfo TopLevelFlag
top_lvl Id
id Demand
id_dmd'
    (DmdEnv
rhs_ty, CoreExpr
rhs')     = AnalEnv -> Demand -> CoreExpr -> (DmdEnv, CoreExpr)
dmdAnalStar AnalEnv
env Demand
id_dmd' CoreExpr
rhs

    -- See Note [Absence analysis for stable unfoldings and RULES]
    rule_fvs :: VarSet
rule_fvs           = Id -> VarSet
bndrRuleAndUnfoldingIds Id
id
    final_ty :: DmdType
final_ty           = DmdType
body_ty' DmdType -> DmdEnv -> DmdType
`plusDmdType` DmdEnv
rhs_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` AnalEnv -> VarSet -> DmdEnv
demandRootSet AnalEnv
env VarSet
rule_fvs

-- | Let bindings can be processed in two ways:
-- Down (RHS before body) or Up (body before RHS).
-- This function handles the down variant.
--
-- It computes a demand signature (by means of 'dmdAnalRhsSig') and uses
-- that at call sites in the body.
--
-- It is used for toplevel definitions, recursive definitions and local
-- non-recursive definitions that have manifest lambdas (cf. 'useLetUp').
-- Local non-recursive definitions without a lambda are handled with LetUp.
--
-- This is the LetDown rule in the paper “Higher-Order Cardinality Analysis”.
dmdAnalBindLetDown :: TopLevelFlag -> AnalEnv -> SubDemand -> CoreBind -> (AnalEnv -> WithDmdType a) -> WithDmdType (DmdResult CoreBind a)
dmdAnalBindLetDown :: forall a.
TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBindLetDown TopLevelFlag
top_lvl AnalEnv
env SubDemand
dmd Bind Id
bind AnalEnv -> WithDmdType a
anal_body = case Bind Id
bind of
  NonRec Id
id CoreExpr
rhs
    | (AnalEnv
env', WeakDmds
weak_fv, Id
id1, CoreExpr
rhs1) <-
        TopLevelFlag
-> RecFlag
-> AnalEnv
-> SubDemand
-> Id
-> CoreExpr
-> (AnalEnv, WeakDmds, Id, CoreExpr)
dmdAnalRhsSig TopLevelFlag
top_lvl RecFlag
NonRecursive AnalEnv
env SubDemand
dmd Id
id CoreExpr
rhs
    -> AnalEnv
-> WeakDmds
-> [(Id, CoreExpr)]
-> ([(Id, CoreExpr)] -> Bind Id)
-> WithDmdType (DmdResult (Bind Id) a)
do_rest AnalEnv
env' WeakDmds
weak_fv [(Id
id1, CoreExpr
rhs1)] ((Id -> CoreExpr -> Bind Id) -> (Id, CoreExpr) -> Bind Id
forall a b c. (a -> b -> c) -> (a, b) -> c
uncurry Id -> CoreExpr -> Bind Id
forall b. b -> Expr b -> Bind b
NonRec ((Id, CoreExpr) -> Bind Id)
-> ([(Id, CoreExpr)] -> (Id, CoreExpr))
-> [(Id, CoreExpr)]
-> Bind Id
forall b c a. (b -> c) -> (a -> b) -> a -> c
. [(Id, CoreExpr)] -> (Id, CoreExpr)
forall a. [a] -> a
only)
  Rec [(Id, CoreExpr)]
pairs
    | (AnalEnv
env', WeakDmds
weak_fv, [(Id, CoreExpr)]
pairs') <- TopLevelFlag
-> AnalEnv
-> SubDemand
-> [(Id, CoreExpr)]
-> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
dmdFix TopLevelFlag
top_lvl AnalEnv
env SubDemand
dmd [(Id, CoreExpr)]
pairs
    -> AnalEnv
-> WeakDmds
-> [(Id, CoreExpr)]
-> ([(Id, CoreExpr)] -> Bind Id)
-> WithDmdType (DmdResult (Bind Id) a)
do_rest AnalEnv
env' WeakDmds
weak_fv [(Id, CoreExpr)]
pairs' [(Id, CoreExpr)] -> Bind Id
forall b. [(b, Expr b)] -> Bind b
Rec
  where
    do_rest :: AnalEnv
-> WeakDmds
-> [(Id, CoreExpr)]
-> ([(Id, CoreExpr)] -> Bind Id)
-> WithDmdType (DmdResult (Bind Id) a)
do_rest AnalEnv
env' WeakDmds
weak_fv [(Id, CoreExpr)]
pairs1 [(Id, CoreExpr)] -> Bind Id
build_bind = DmdType
-> DmdResult (Bind Id) a -> WithDmdType (DmdResult (Bind Id) a)
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
final_ty (Bind Id -> a -> DmdResult (Bind Id) a
forall a b. a -> b -> DmdResult a b
R ([(Id, CoreExpr)] -> Bind Id
build_bind [(Id, CoreExpr)]
pairs2) a
body')
      where
        WithDmdType DmdType
body_ty a
body'        = AnalEnv -> WithDmdType a
anal_body AnalEnv
env'
        -- see Note [Lazy and unleashable free variables]
        dmd_ty :: DmdType
dmd_ty                          = DmdType -> WeakDmds -> DmdType
addWeakFVs DmdType
body_ty WeakDmds
weak_fv
        WithDmdType DmdType
final_ty [Demand]
id_dmds    = AnalEnv -> DmdType -> [Id] -> WithDmdType [Demand]
findBndrsDmds AnalEnv
env' DmdType
dmd_ty (((Id, CoreExpr) -> Id) -> [(Id, CoreExpr)] -> [Id]
forall a b. (a -> b) -> [a] -> [b]
strictMap (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst [(Id, CoreExpr)]
pairs1)
        -- Important to force this as build_bind might not force it.
        !pairs2 :: [(Id, CoreExpr)]
pairs2                         = ((Id, CoreExpr) -> Demand -> (Id, CoreExpr))
-> [(Id, CoreExpr)] -> [Demand] -> [(Id, CoreExpr)]
forall a b c. (a -> b -> c) -> [a] -> [b] -> [c]
strictZipWith (Id, CoreExpr) -> Demand -> (Id, CoreExpr)
do_one [(Id, CoreExpr)]
pairs1 [Demand]
id_dmds
        do_one :: (Id, CoreExpr) -> Demand -> (Id, CoreExpr)
do_one (Id
id', CoreExpr
rhs') Demand
dmd          = ((,) (Id -> CoreExpr -> (Id, CoreExpr))
-> Id -> CoreExpr -> (Id, CoreExpr)
forall a b. (a -> b) -> a -> b
$! TopLevelFlag -> Id -> Demand -> Id
setBindIdDemandInfo TopLevelFlag
top_lvl Id
id' Demand
dmd) (CoreExpr -> (Id, CoreExpr)) -> CoreExpr -> (Id, CoreExpr)
forall a b. (a -> b) -> a -> b
$! CoreExpr
rhs'
        -- If the actual demand is better than the vanilla call
        -- demand, you might think that we might do better to re-analyse
        -- the RHS with the stronger demand.
        -- But (a) That seldom happens, because it means that *every* path in
        --         the body of the let has to use that stronger demand
        -- (b) It often happens temporarily in when fixpointing, because
        --     the recursive function at first seems to place a massive demand.
        --     But we don't want to go to extra work when the function will
        --     probably iterate to something less demanding.
        -- In practice, all the times the actual demand on id2 is more than
        -- the vanilla call demand seem to be due to (b).  So we don't
        -- bother to re-analyse the RHS.

-- | Mimic the effect of 'GHC.Core.Prep.mkFloat', turning non-trivial argument
-- expressions/RHSs into a proper let-bound thunk (lifted) or a case (with
-- unlifted scrutinee).
anticipateANF :: CoreExpr -> Card -> Card
anticipateANF :: CoreExpr -> Card -> Card
anticipateANF CoreExpr
e Card
n
  | CoreExpr -> Bool
exprIsTrivial CoreExpr
e                               = Card
n -- trivial expr won't have a binding
  | Type -> Bool
definitelyUnliftedType (HasDebugCallStack => CoreExpr -> Type
CoreExpr -> Type
exprType CoreExpr
e)
  , Bool -> Bool
not (Card -> Bool
isAbs Card
n Bool -> Bool -> Bool
&& CoreExpr -> Bool
exprOkForSpeculation CoreExpr
e)       = Card -> Card
forall {p}. p -> Card
case_bind Card
n
  | Bool
otherwise                                     = Card -> Card
let_bind  Card
n
  where
    case_bind :: p -> Card
case_bind p
_ = Card
C_11       -- evaluated exactly once
    let_bind :: Card -> Card
let_bind    = Card -> Card
oneifyCard -- evaluated at most once

-- Do not process absent demands
-- Otherwise act like in a normal demand analysis
-- See ↦* relation in the Cardinality Analysis paper
dmdAnalStar :: AnalEnv
            -> Demand   -- This one takes a *Demand*
            -> CoreExpr
            -> (DmdEnv, CoreExpr)
dmdAnalStar :: AnalEnv -> Demand -> CoreExpr -> (DmdEnv, CoreExpr)
dmdAnalStar AnalEnv
env (Card
n :* SubDemand
sd) CoreExpr
e
  -- NB: (:*) expands AbsDmd and BotDmd as needed
  | WithDmdType DmdType
dmd_ty CoreExpr
e' <- AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
sd CoreExpr
e
  , Card
n' <- CoreExpr -> Card -> Card
anticipateANF CoreExpr
e Card
n
      -- See Note [Anticipating ANF in demand analysis]
      -- and Note [Analysing with absent demand]
  = (DmdType -> DmdEnv
discardArgDmds (DmdType -> DmdEnv) -> DmdType -> DmdEnv
forall a b. (a -> b) -> a -> b
$ Card -> DmdType -> DmdType
multDmdType Card
n' DmdType
dmd_ty, CoreExpr
e')

-- Main Demand Analysis machinery
dmdAnal, dmdAnal' :: AnalEnv
        -> SubDemand         -- The main one takes a *SubDemand*
        -> CoreExpr -> WithDmdType CoreExpr

dmdAnal :: AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
d CoreExpr
e = -- pprTrace "dmdAnal" (ppr d <+> ppr e) $
                  AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal' AnalEnv
env SubDemand
d CoreExpr
e

dmdAnal' :: AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal' AnalEnv
_ SubDemand
_ (Lit Literal
lit)     = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
nopDmdType (Literal -> CoreExpr
forall b. Literal -> Expr b
Lit Literal
lit)
dmdAnal' AnalEnv
_ SubDemand
_ (Type Type
ty)     = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
nopDmdType (Type -> CoreExpr
forall b. Type -> Expr b
Type Type
ty) -- Doesn't happen, in fact
dmdAnal' AnalEnv
_ SubDemand
_ (Coercion Coercion
co)
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType (DmdEnv -> DmdType
noArgsDmdType (Coercion -> DmdEnv
coercionDmdEnv Coercion
co)) (Coercion -> CoreExpr
forall b. Coercion -> Expr b
Coercion Coercion
co)

dmdAnal' AnalEnv
env SubDemand
dmd (Var Id
var)
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType (AnalEnv -> Id -> SubDemand -> DmdType
dmdTransform AnalEnv
env Id
var SubDemand
dmd) (Id -> CoreExpr
forall b. Id -> Expr b
Var Id
var)

dmdAnal' AnalEnv
env SubDemand
dmd (Cast CoreExpr
e Coercion
co)
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType (DmdType
dmd_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` Coercion -> DmdEnv
coercionDmdEnv Coercion
co) (CoreExpr -> Coercion -> CoreExpr
forall b. Expr b -> Coercion -> Expr b
Cast CoreExpr
e' Coercion
co)
  where
    WithDmdType DmdType
dmd_ty CoreExpr
e' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
dmd CoreExpr
e

dmdAnal' AnalEnv
env SubDemand
dmd (Tick CoreTickish
t CoreExpr
e)
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
dmd_ty (CoreTickish -> CoreExpr -> CoreExpr
forall b. CoreTickish -> Expr b -> Expr b
Tick CoreTickish
t CoreExpr
e')
  where
    WithDmdType DmdType
dmd_ty CoreExpr
e' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
dmd CoreExpr
e

dmdAnal' AnalEnv
env SubDemand
dmd (App CoreExpr
fun (Type Type
ty))
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
fun_ty (CoreExpr -> CoreExpr -> CoreExpr
forall b. Expr b -> Expr b -> Expr b
App CoreExpr
fun' (Type -> CoreExpr
forall b. Type -> Expr b
Type Type
ty))
  where
    WithDmdType DmdType
fun_ty CoreExpr
fun' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
dmd CoreExpr
fun

-- Lots of the other code is there to make this
-- beautiful, compositional, application rule :-)
dmdAnal' AnalEnv
env SubDemand
dmd (App CoreExpr
fun CoreExpr
arg)
  = -- This case handles value arguments (type args handled above)
    -- Crucially, coercions /are/ handled here, because they are
    -- value arguments (#10288)
    let
        call_dmd :: SubDemand
call_dmd          = SubDemand -> SubDemand
mkCalledOnceDmd SubDemand
dmd
        WithDmdType DmdType
fun_ty CoreExpr
fun' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
call_dmd CoreExpr
fun
        (Demand
arg_dmd, DmdType
res_ty) = DmdType -> (Demand, DmdType)
splitDmdTy DmdType
fun_ty
        (DmdEnv
arg_ty, CoreExpr
arg')    = AnalEnv -> Demand -> CoreExpr -> (DmdEnv, CoreExpr)
dmdAnalStar AnalEnv
env Demand
arg_dmd CoreExpr
arg
    in
--    pprTrace "dmdAnal:app" (vcat
--         [ text "dmd =" <+> ppr dmd
--         , text "expr =" <+> ppr (App fun arg)
--         , text "fun dmd_ty =" <+> ppr fun_ty
--         , text "arg dmd =" <+> ppr arg_dmd
--         , text "arg dmd_ty =" <+> ppr arg_ty
--         , text "res dmd_ty =" <+> ppr res_ty
--         , text "overall res dmd_ty =" <+> ppr (res_ty `plusDmdType` arg_ty) ])
    DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType (DmdType
res_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` DmdEnv
arg_ty) (CoreExpr -> CoreExpr -> CoreExpr
forall b. Expr b -> Expr b -> Expr b
App CoreExpr
fun' CoreExpr
arg')

dmdAnal' AnalEnv
env SubDemand
dmd (Lam Id
var CoreExpr
body)
  | Id -> Bool
isTyVar Id
var
  = let
        WithDmdType DmdType
body_ty CoreExpr
body' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal (AnalEnv -> Id -> AnalEnv
addInScopeAnalEnv AnalEnv
env Id
var) SubDemand
dmd CoreExpr
body
        -- See Note [Bringing a new variable into scope]
    in
    DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
body_ty (Id -> CoreExpr -> CoreExpr
forall b. b -> Expr b -> Expr b
Lam Id
var CoreExpr
body')

  | Bool
otherwise
  = let (Card
n, SubDemand
body_dmd)    = SubDemand -> (Card, SubDemand)
peelCallDmd SubDemand
dmd
          -- body_dmd: a demand to analyze the body

        WithDmdType DmdType
body_ty CoreExpr
body' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal (AnalEnv -> Id -> AnalEnv
addInScopeAnalEnv AnalEnv
env Id
var) SubDemand
body_dmd CoreExpr
body
        -- See Note [Bringing a new variable into scope]
        WithDmdType DmdType
lam_ty Id
var'   = AnalEnv -> DmdType -> Id -> WithDmdType Id
annotateLamIdBndr AnalEnv
env DmdType
body_ty Id
var
        new_dmd_type :: DmdType
new_dmd_type = Card -> DmdType -> DmdType
multDmdType Card
n DmdType
lam_ty
    in
    DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
new_dmd_type (Id -> CoreExpr -> CoreExpr
forall b. b -> Expr b -> Expr b
Lam Id
var' CoreExpr
body')

dmdAnal' AnalEnv
env SubDemand
dmd (Case CoreExpr
scrut Id
case_bndr Type
ty [Alt AltCon
alt_con [Id]
bndrs CoreExpr
rhs])
  -- Only one alternative.
  -- If it's a DataAlt, it should be the only constructor of the type and we
  -- can consider its field demands when analysing the scrutinee.
  | AltCon -> Bool
want_precise_field_dmds AltCon
alt_con
  = let
        rhs_env :: AnalEnv
rhs_env = AnalEnv -> [Id] -> AnalEnv
addInScopeAnalEnvs AnalEnv
env (Id
case_bndrId -> [Id] -> [Id]
forall a. a -> [a] -> [a]
:[Id]
bndrs)
        -- See Note [Bringing a new variable into scope]
        WithDmdType DmdType
rhs_ty CoreExpr
rhs'           = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
rhs_env SubDemand
dmd CoreExpr
rhs
        WithDmdType DmdType
alt_ty1 [Demand]
fld_dmds      = AnalEnv -> DmdType -> [Id] -> WithDmdType [Demand]
findBndrsDmds AnalEnv
env DmdType
rhs_ty [Id]
bndrs
        WithDmdType DmdType
alt_ty2 Demand
case_bndr_dmd = AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
alt_ty1 Id
case_bndr
        !case_bndr' :: Id
case_bndr'                       = Id -> Demand -> Id
setIdDemandInfo Id
case_bndr Demand
case_bndr_dmd

        -- Evaluation cardinality on the case binder is irrelevant and a no-op.
        -- What matters is its nested sub-demand!
        -- NB: If case_bndr_dmd is absDmd, boxity will say Unboxed, which is
        -- what we want, because then `seq` will put a `seqDmd` on its scrut.
        (Card
_ :* SubDemand
case_bndr_sd) = Demand -> Demand
strictifyDmd Demand
case_bndr_dmd

        -- Compute demand on the scrutinee
        -- FORCE the result, otherwise thunks will end up retaining the
        -- whole DmdEnv
        !(![Id]
bndrs', !SubDemand
scrut_sd)
          | DataAlt DataCon
_ <- AltCon
alt_con
          -- See Note [Demand on the scrutinee of a product case]
          , let !scrut_sd :: SubDemand
scrut_sd = SubDemand -> [Demand] -> SubDemand
scrutSubDmd SubDemand
case_bndr_sd [Demand]
fld_dmds
          -- See Note [Demand on case-alternative binders]
          , let !fld_dmds' :: [Demand]
fld_dmds' = SubDemand -> Arity -> [Demand]
fieldBndrDmds SubDemand
scrut_sd ([Demand] -> Arity
forall a. [a] -> Arity
forall (t :: * -> *) a. Foldable t => t a -> Arity
length [Demand]
fld_dmds)
          , let !bndrs' :: [Id]
bndrs' = HasCallStack => [Id] -> [Demand] -> [Id]
[Id] -> [Demand] -> [Id]
setBndrsDemandInfo [Id]
bndrs [Demand]
fld_dmds'
          = ([Id]
bndrs', SubDemand
scrut_sd)
          | Bool
otherwise
          -- DEFAULT alts. Simply add demands and discard the evaluation
          -- cardinality, as we evaluate the scrutinee exactly once.
          = Bool -> ([Id], SubDemand) -> ([Id], SubDemand)
forall a. HasCallStack => Bool -> a -> a
assert ([Id] -> Bool
forall a. [a] -> Bool
forall (t :: * -> *) a. Foldable t => t a -> Bool
null [Id]
bndrs) ([Id]
bndrs, SubDemand
case_bndr_sd)

        alt_ty3 :: DmdType
alt_ty3
          -- See Note [Precise exceptions and strictness analysis] in "GHC.Types.Demand"
          | FamInstEnvs -> CoreExpr -> Bool
exprMayThrowPreciseException (AnalEnv -> FamInstEnvs
ae_fam_envs AnalEnv
env) CoreExpr
scrut
          = DmdType -> DmdType
deferAfterPreciseException DmdType
alt_ty2
          | Bool
otherwise
          = DmdType
alt_ty2

        WithDmdType DmdType
scrut_ty CoreExpr
scrut' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
scrut_sd CoreExpr
scrut
        res_ty :: DmdType
res_ty             = DmdType
alt_ty3 DmdType -> DmdEnv -> DmdType
`plusDmdType` DmdType -> DmdEnv
discardArgDmds DmdType
scrut_ty
    in
--    pprTrace "dmdAnal:Case1" (vcat [ text "scrut" <+> ppr scrut
--                                   , text "dmd" <+> ppr dmd
--                                   , text "case_bndr_dmd" <+> ppr (idDemandInfo case_bndr')
--                                   , text "scrut_sd" <+> ppr scrut_sd
--                                   , text "scrut_ty" <+> ppr scrut_ty
--                                   , text "alt_ty" <+> ppr alt_ty2
--                                   , text "res_ty" <+> ppr res_ty ]) $
    DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
res_ty (CoreExpr -> Id -> Type -> [Alt Id] -> CoreExpr
forall b. Expr b -> b -> Type -> [Alt b] -> Expr b
Case CoreExpr
scrut' Id
case_bndr' Type
ty [AltCon -> [Id] -> CoreExpr -> Alt Id
forall b. AltCon -> [b] -> Expr b -> Alt b
Alt AltCon
alt_con [Id]
bndrs' CoreExpr
rhs'])
    where
      want_precise_field_dmds :: AltCon -> Bool
want_precise_field_dmds (DataAlt DataCon
dc)
        | Maybe DataCon
Nothing <- TyCon -> Maybe DataCon
tyConSingleAlgDataCon_maybe (TyCon -> Maybe DataCon) -> TyCon -> Maybe DataCon
forall a b. (a -> b) -> a -> b
$ DataCon -> TyCon
dataConTyCon DataCon
dc
        = Bool
False    -- Not a product type, even though this is the
                   -- only remaining possible data constructor
        | IsRecDataConResult
DefinitelyRecursive <- AnalEnv -> DataCon -> IsRecDataConResult
ae_rec_dc AnalEnv
env DataCon
dc
        = Bool
False     -- See Note [Demand analysis for recursive data constructors]
        | Bool
otherwise
        = Bool
True
      want_precise_field_dmds (LitAlt {}) = Bool
False  -- Like the non-product datacon above
      want_precise_field_dmds AltCon
DEFAULT     = Bool
True

dmdAnal' AnalEnv
env SubDemand
dmd (Case CoreExpr
scrut Id
case_bndr Type
ty [Alt Id]
alts)
  = let      -- Case expression with multiple alternatives
        WithDmdType DmdType
scrut_ty CoreExpr
scrut' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
topSubDmd CoreExpr
scrut

        WithDmdType DmdType
alt_ty1 Demand
case_bndr_dmd = AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
alt_ty Id
case_bndr
        !case_bndr' :: Id
case_bndr'                       = Id -> Demand -> Id
setIdDemandInfo Id
case_bndr Demand
case_bndr_dmd
        WithDmdType DmdType
alt_ty [Alt Id]
alts'          = AnalEnv -> SubDemand -> Id -> [Alt Id] -> WithDmdType [Alt Id]
dmdAnalSumAlts AnalEnv
env SubDemand
dmd Id
case_bndr [Alt Id]
alts

        fam_envs :: FamInstEnvs
fam_envs             = AnalEnv -> FamInstEnvs
ae_fam_envs AnalEnv
env
        alt_ty2 :: DmdType
alt_ty2
          -- See Note [Precise exceptions and strictness analysis] in "GHC.Types.Demand"
          | FamInstEnvs -> CoreExpr -> Bool
exprMayThrowPreciseException FamInstEnvs
fam_envs CoreExpr
scrut
          = DmdType -> DmdType
deferAfterPreciseException DmdType
alt_ty1
          | Bool
otherwise
          = DmdType
alt_ty1
        res_ty :: DmdType
res_ty               = DmdType
scrut_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` DmdType -> DmdEnv
discardArgDmds DmdType
alt_ty2

    in
--    pprTrace "dmdAnal:Case2" (vcat [ text "scrut" <+> ppr scrut
--                                   , text "scrut_ty" <+> ppr scrut_ty
--                                   , text "alt_ty1" <+> ppr alt_ty1
--                                   , text "alt_ty2" <+> ppr alt_ty2
--                                   , text "res_ty" <+> ppr res_ty ]) $
    DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
res_ty (CoreExpr -> Id -> Type -> [Alt Id] -> CoreExpr
forall b. Expr b -> b -> Type -> [Alt b] -> Expr b
Case CoreExpr
scrut' Id
case_bndr' Type
ty [Alt Id]
alts')

dmdAnal' AnalEnv
env SubDemand
dmd (Let Bind Id
bind CoreExpr
body)
  = DmdType -> CoreExpr -> WithDmdType CoreExpr
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
final_ty (Bind Id -> CoreExpr -> CoreExpr
forall b. Bind b -> Expr b -> Expr b
Let Bind Id
bind' CoreExpr
body')
  where
    !(WithDmdType DmdType
final_ty (R Bind Id
bind' CoreExpr
body')) = TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType CoreExpr)
-> WithDmdType (DmdResult (Bind Id) CoreExpr)
forall a.
TopLevelFlag
-> AnalEnv
-> SubDemand
-> Bind Id
-> (AnalEnv -> WithDmdType a)
-> WithDmdType (DmdResult (Bind Id) a)
dmdAnalBind TopLevelFlag
NotTopLevel AnalEnv
env SubDemand
dmd Bind Id
bind AnalEnv -> WithDmdType CoreExpr
go'
    go' :: AnalEnv -> WithDmdType CoreExpr
go' !AnalEnv
env'                 = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env' SubDemand
dmd CoreExpr
body

-- | A simple, syntactic analysis of whether an expression MAY throw a precise
-- exception when evaluated. It's always sound to return 'True'.
-- See Note [Which scrutinees may throw precise exceptions].
exprMayThrowPreciseException :: FamInstEnvs -> CoreExpr -> Bool
exprMayThrowPreciseException :: FamInstEnvs -> CoreExpr -> Bool
exprMayThrowPreciseException FamInstEnvs
envs CoreExpr
e
  | Bool -> Bool
not (FamInstEnvs -> Type -> Bool
forcesRealWorld FamInstEnvs
envs (HasDebugCallStack => CoreExpr -> Type
CoreExpr -> Type
exprType CoreExpr
e))
  = Bool
False -- 1. in the Note
  | (Var Id
f, [CoreExpr]
_) <- CoreExpr -> (CoreExpr, [CoreExpr])
forall b. Expr b -> (Expr b, [Expr b])
collectArgs CoreExpr
e
  , Just PrimOp
op    <- Id -> Maybe PrimOp
isPrimOpId_maybe Id
f
  , PrimOp
op PrimOp -> PrimOp -> Bool
forall a. Eq a => a -> a -> Bool
/= PrimOp
RaiseIOOp
  = Bool
False -- 2. in the Note
  | (Var Id
f, [CoreExpr]
_) <- CoreExpr -> (CoreExpr, [CoreExpr])
forall b. Expr b -> (Expr b, [Expr b])
collectArgs CoreExpr
e
  , Just ForeignCall
fcall <- Id -> Maybe ForeignCall
isFCallId_maybe Id
f
  , Bool -> Bool
not (ForeignCall -> Bool
isSafeForeignCall ForeignCall
fcall)
  = Bool
False -- 3. in the Note
  | Bool
otherwise
  = Bool
True  -- _. in the Note

-- | Recognises types that are
--    * @State# RealWorld@
--    * Unboxed tuples with a @State# RealWorld@ field
-- modulo coercions. This will detect 'IO' actions (even post Nested CPR! See
-- T13380e) and user-written variants thereof by their type.
forcesRealWorld :: FamInstEnvs -> Type -> Bool
forcesRealWorld :: FamInstEnvs -> Type -> Bool
forcesRealWorld FamInstEnvs
fam_envs Type
ty
  | Type
ty Type -> Type -> Bool
`eqType` Type
realWorldStatePrimTy
  = Bool
True
  | Just (TyCon
tc, [Type]
tc_args, Coercion
_co)  <- FamInstEnvs -> Type -> Maybe (TyCon, [Type], Coercion)
normSplitTyConApp_maybe FamInstEnvs
fam_envs Type
ty
  , TyCon -> Bool
isUnboxedTupleTyCon TyCon
tc
  , let field_tys :: [Scaled Type]
field_tys = DataCon -> [Type] -> [Scaled Type]
dataConInstArgTys (TyCon -> DataCon
tyConSingleDataCon TyCon
tc) [Type]
tc_args
  = (Scaled Type -> Bool) -> [Scaled Type] -> Bool
forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool
any (Type -> Type -> Bool
eqType Type
realWorldStatePrimTy (Type -> Bool) -> (Scaled Type -> Type) -> Scaled Type -> Bool
forall b c a. (b -> c) -> (a -> b) -> a -> c
. Scaled Type -> Type
forall a. Scaled a -> a
scaledThing) [Scaled Type]
field_tys
  | Bool
otherwise
  = Bool
False

dmdAnalSumAlts :: AnalEnv -> SubDemand -> Id -> [CoreAlt] -> WithDmdType [CoreAlt]
dmdAnalSumAlts :: AnalEnv -> SubDemand -> Id -> [Alt Id] -> WithDmdType [Alt Id]
dmdAnalSumAlts AnalEnv
_ SubDemand
_ Id
_ [] = DmdType -> [Alt Id] -> WithDmdType [Alt Id]
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
botDmdType []
  -- Base case is botDmdType, for empty case alternatives
  -- This is a unit for lubDmdType, and the right result
  -- when there really are no alternatives
dmdAnalSumAlts AnalEnv
env SubDemand
dmd Id
case_bndr (Alt Id
alt:[Alt Id]
alts)
  = let
      WithDmdType DmdType
cur_ty  Alt Id
alt'  = AnalEnv -> SubDemand -> Id -> Alt Id -> WithDmdType (Alt Id)
dmdAnalSumAlt AnalEnv
env SubDemand
dmd Id
case_bndr Alt Id
alt
      WithDmdType DmdType
rest_ty [Alt Id]
alts' = AnalEnv -> SubDemand -> Id -> [Alt Id] -> WithDmdType [Alt Id]
dmdAnalSumAlts AnalEnv
env SubDemand
dmd Id
case_bndr [Alt Id]
alts
    in DmdType -> [Alt Id] -> WithDmdType [Alt Id]
forall a. DmdType -> a -> WithDmdType a
WithDmdType (DmdType -> DmdType -> DmdType
lubDmdType DmdType
cur_ty DmdType
rest_ty) (Alt Id
alt'Alt Id -> [Alt Id] -> [Alt Id]
forall a. a -> [a] -> [a]
:[Alt Id]
alts')


dmdAnalSumAlt :: AnalEnv -> SubDemand -> Id -> CoreAlt -> WithDmdType CoreAlt
dmdAnalSumAlt :: AnalEnv -> SubDemand -> Id -> Alt Id -> WithDmdType (Alt Id)
dmdAnalSumAlt AnalEnv
env SubDemand
dmd Id
case_bndr (Alt AltCon
con [Id]
bndrs CoreExpr
rhs)
  | let rhs_env :: AnalEnv
rhs_env = AnalEnv -> [Id] -> AnalEnv
addInScopeAnalEnvs AnalEnv
env (Id
case_bndrId -> [Id] -> [Id]
forall a. a -> [a] -> [a]
:[Id]
bndrs)
    -- See Note [Bringing a new variable into scope]
  , WithDmdType DmdType
rhs_ty CoreExpr
rhs' <- AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
rhs_env SubDemand
dmd CoreExpr
rhs
  , WithDmdType DmdType
alt_ty [Demand]
dmds <- AnalEnv -> DmdType -> [Id] -> WithDmdType [Demand]
findBndrsDmds AnalEnv
env DmdType
rhs_ty [Id]
bndrs
  , let (Card
_ :* SubDemand
case_bndr_sd) = DmdType -> Id -> Demand
findIdDemand DmdType
alt_ty Id
case_bndr
        -- See Note [Demand on case-alternative binders]
        -- we can't use the scrut_sd, because it says 'Prod' and we'll use
        -- topSubDmd anyway for scrutinees of sum types.
        scrut_sd :: SubDemand
scrut_sd = SubDemand -> [Demand] -> SubDemand
scrutSubDmd SubDemand
case_bndr_sd [Demand]
dmds
        dmds' :: [Demand]
dmds' = SubDemand -> Arity -> [Demand]
fieldBndrDmds SubDemand
scrut_sd ([Demand] -> Arity
forall a. [a] -> Arity
forall (t :: * -> *) a. Foldable t => t a -> Arity
length [Demand]
dmds)
        -- Do not put a thunk into the Alt
        !new_ids :: [Id]
new_ids            = HasCallStack => [Id] -> [Demand] -> [Id]
[Id] -> [Demand] -> [Id]
setBndrsDemandInfo [Id]
bndrs [Demand]
dmds'
  = -- pprTrace "dmdAnalSumAlt" (ppr con $$ ppr case_bndr $$ ppr dmd $$ ppr alt_ty) $
    DmdType -> Alt Id -> WithDmdType (Alt Id)
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
alt_ty (AltCon -> [Id] -> CoreExpr -> Alt Id
forall b. AltCon -> [b] -> Expr b -> Alt b
Alt AltCon
con [Id]
new_ids CoreExpr
rhs')

-- See Note [Demand on the scrutinee of a product case]
scrutSubDmd :: SubDemand -> [Demand] -> SubDemand
scrutSubDmd :: SubDemand -> [Demand] -> SubDemand
scrutSubDmd SubDemand
case_sd [Demand]
fld_dmds =
  -- pprTraceWith "scrutSubDmd" (\scrut_sd -> ppr case_sd $$ ppr fld_dmds $$ ppr scrut_sd) $
  SubDemand
case_sd SubDemand -> SubDemand -> SubDemand
`plusSubDmd` Boxity -> [Demand] -> SubDemand
mkProd Boxity
Unboxed [Demand]
fld_dmds

-- See Note [Demand on case-alternative binders]
fieldBndrDmds :: SubDemand -- on the scrutinee
              -> Arity
              -> [Demand]  -- Final demands for the components of the DataCon
fieldBndrDmds :: SubDemand -> Arity -> [Demand]
fieldBndrDmds SubDemand
scrut_sd Arity
n_flds =
  case Arity -> SubDemand -> Maybe (Boxity, [Demand])
viewProd Arity
n_flds SubDemand
scrut_sd of
    Just (Boxity
_, [Demand]
ds) -> [Demand]
ds
    Maybe (Boxity, [Demand])
Nothing      -> Arity -> Demand -> [Demand]
forall a. Arity -> a -> [a]
replicate Arity
n_flds Demand
topDmd
                      -- Either an arity mismatch or scrut_sd was a call demand.
                      -- See Note [Untyped demand on case-alternative binders]

{-
Note [Anticipating ANF in demand analysis]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When analysing non-complex (e.g., trivial) thunks and complex function
arguments, we have to pretend that the expression is really in administrative
normal form (ANF), the conversion to which is done by CorePrep.

Consider
```
f x = let y = x |> co in y `seq` y `seq` ()
```
E.g., 'y' is a let-binding with a trivial RHS. That may occur if 'y' can't be
inlined, for example. Now, is 'x' used once? It may appear as if that is the
case, since its only occurrence is in 'y's memoised RHS. But actually, CorePrep
will *not* allocate a thunk for 'y', because it is trivial and could just
re-use the memoisation mechanism of 'x'! By saying that 'x' is used once it
becomes a single-entry thunk and a call to 'f' will evaluate it twice.
The same applies to trivial arguments, e.g., `f z` really evaluates `z` twice.

So, somewhat counter-intuitively, trivial arguments and let RHSs will *not* be
memoised. On the other hand, evaluation of non-trivial arguments and let RHSs
*will* be memoised. In fact, consider the effect of conversion to ANF on complex
function arguments (as done by 'GHC.Core.Prep.mkFloat'):
```
f2 (g2 x) ===> let y = g2 x in f2 y                   (if `y` is lifted)
f3 (g3 x) ===> case g3 x of y { __DEFAULT -> f3 y }   (if `y` is not lifted)
```
So if a lifted argument like `g2 x` is complex enough, it will be memoised.
Regardless how many times 'f2' evaluates its parameter, the argument will be
evaluated at most once to WHNF.
Similarly, when an unlifted argument like `g3 x` is complex enough, we will
evaluate it *exactly* once to WHNF, no matter how 'f3' evaluates its parameter.

Note that any evaluation beyond WHNF is not affected by memoisation. So this
Note affects the outer 'Card' of a 'Demand', but not its nested 'SubDemand'.
'anticipateANF' predicts the effect of case-binding and let-binding complex
arguments, as well as the lack of memoisation for trivial let RHSs.
In particular, this takes care of the gripes in
Note [Analysing with absent demand] relating to unlifted types.

Note [Analysing with absent demand]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose we analyse an expression with demand A.  The "A" means
"absent", so this expression will never be needed. What should happen?
There are several wrinkles:

* We *do* want to analyse the expression regardless.
  Reason: Note [Always analyse in virgin pass]

  But we can post-process the results to ignore all the usage
  demands coming back. This is done by 'multDmdType' with the appropriate
  (absent) evaluation cardinality A or B.

* Nevertheless, which sub-demand should we pick for analysis?
  Since the demand was absent, any would do. Worker/wrapper will replace
  absent bindings with an absent filler anyway, so annotations in the RHS
  of an absent binding don't matter much.
  Picking 'botSubDmd' would be the most useful, but would also look a bit
  misleading in the Core output of DmdAnal, because all nested annotations would
  be bottoming. Better pick 'seqSubDmd', so that we annotate many of those
  nested bindings with A themselves.

* Since we allow unlifted arguments that are not ok-for-speculation,
  we need to be extra careful in the following situation, because unlifted
  values are evaluated even if they are not used. Example from #9254:
     f :: (() -> (# Int#, () #)) -> ()
          -- Strictness signature is
          --    <1C(1,P(A,1L))>
          -- I.e. calls k, but discards first component of result
     f k = case k () of (# _, r #) -> r

     g :: Int -> ()
     g y = f (\n -> (# case y of I# y2 -> y2, n #))

  Here, f's strictness signature says (correctly) that it calls its argument
  function and ignores the first component of its result.

  But in function g, we *will* evaluate the 'case y of ...', because it has type
  Int#. So in the program as written, 'y' will be evaluated. Hence we must
  record this usage of 'y', else 'g' will say 'y' is absent, and will w/w so
  that 'y' is bound to an absent filler (see Note [Absent fillers]), leading
  to a crash when 'y' is evaluated.

  Now, worker/wrapper could be smarter and replace `case y of I# y2 -> y2`
  with a suitable absent filler such as `RUBBISH[IntRep] @Int#`.
  But as long as worker/wrapper isn't equipped to do so, we must be cautious,
  and follow Note [Anticipating ANF in demand analysis]. That is, in
  'dmdAnalStar', we will set the evaluation cardinality to C_11, anticipating
  the case binding of the complex argument `case y of I# y2 -> y2`. This
  cardinlities' only effect is in the call to 'multDmdType', where it makes sure
  that the demand on the arg's free variable 'y' is not absent and strict, so
  that it is ultimately passed unboxed to 'g'.

Note [Always analyse in virgin pass]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Tricky point: make sure that we analyse in the 'virgin' pass. Consider
   rec { f acc x True  = f (...rec { g y = ...g... }...)
         f acc x False = acc }
In the virgin pass for 'f' we'll give 'f' a very strict (bottom) type.
That might mean that we analyse the sub-expression containing the
E = "...rec g..." stuff in a bottom demand.  Suppose we *didn't analyse*
E, but just returned botType.

Then in the *next* (non-virgin) iteration for 'f', we might analyse E
in a weaker demand, and that will trigger doing a fixpoint iteration
for g.  But *because it's not the virgin pass* we won't start g's
iteration at bottom.  Disaster.  (This happened in $sfibToList' of
nofib/spectral/fibheaps.)

So in the virgin pass we make sure that we do analyse the expression
at least once, to initialise its signatures.

Note [Which scrutinees may throw precise exceptions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This is the specification of 'exprMayThrowPreciseExceptions',
which is important for Scenario 2 of
Note [Precise exceptions and strictness analysis] in GHC.Types.Demand.

For an expression @f a1 ... an :: ty@ we determine that
  1. False  If ty is *not* @State# RealWorld@ or an unboxed tuple thereof.
            This check is done by 'forcesRealWorld'.
            (Why not simply unboxed pairs as above? This is motivated by
            T13380{d,e}.)
  2. False  If f is a PrimOp, and it is *not* raiseIO#
  3. False  If f is an unsafe FFI call ('PlayRisky')
  _. True   Otherwise "give up".

It is sound to return False in those cases, because
  1. We don't give any guarantees for unsafePerformIO, so no precise exceptions
     from pure code.
  2. raiseIO# is the only primop that may throw a precise exception.
  3. Unsafe FFI calls may not interact with the RTS (to throw, for example).
     See haddock on GHC.Types.ForeignCall.PlayRisky.

We *need* to return False in those cases, because
  1. We would lose too much strictness in pure code, all over the place.
  2. We would lose strictness for primops like getMaskingState#, which
     introduces a substantial regression in
     GHC.IO.Handle.Internals.wantReadableHandle.
  3. We would lose strictness for code like GHC.Fingerprint.fingerprintData,
     where an intermittent FFI call to c_MD5Init would otherwise lose
     strictness on the arguments len and buf, leading to regressions in T9203
     (2%) and i386's haddock.base (5%). Tested by T13380f.

In !3014 we tried a more sophisticated analysis by introducing ConOrDiv (nic)
to the Divergence lattice, but in practice it turned out to be hard to untaint
from 'topDiv' to 'conDiv', leading to bugs, performance regressions and
complexity that didn't justify the single fixed testcase T13380c.

Note [Demand analysis for recursive data constructors]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
T11545 features a single-product, recursive data type
  data A = A A A ... A
    deriving Eq
Naturally, `(==)` is deeply strict in `A` and in fact will never terminate. That
leads to very large (exponential in the depth) demand signatures and fruitless
churn in boxity analysis, demand analysis and worker/wrapper.

So we detect `A` as a recursive data constructor (see
Note [Detecting recursive data constructors]) analysing `case x of A ...`
and simply assume L for the demand on field binders, which is the same code
path as we take for sum types. This code happens in want_precise_field_dmds
in the Case equation for dmdAnal.

Combined with the B demand on the case binder, we get the very small demand
signature <1S><1S>b on `(==)`. This improves ghc/alloc performance on T11545
tenfold! See also Note [CPR for recursive data constructors] which describes the
sibling mechanism in CPR analysis.

Note [Demand on the scrutinee of a product case]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When figuring out the demand on the scrutinee of a product case,
we use the demands of the case alternative, i.e. id_dmds.
But note that these include the demand on the case binder;
see Note [Demand on case-alternative binders].
This is crucial. Example:
   f x = case x of y { (a,b) -> k y a }
If we just take scrut_demand = 1P(L,A), then we won't pass x to the
worker, so the worker will rebuild
     x = (a, absent-error)
and that'll crash.

Note [Demand on case-alternative binders]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The demand on a binder in a case alternative comes
  (a) From the demand on the binder itself
  (b) From the demand on the case binder
Forgetting (b) led directly to #10148.

Example. Source code:
  f x@(p,_) = if p then foo x else True

  foo (p,True) = True
  foo (p,q)    = foo (q,p)

After strictness analysis, forgetting (b):
  f = \ (x_an1 [Dmd=1P(1L,ML)] :: (Bool, Bool)) ->
      case x_an1
      of wild_X7 [Dmd=MP(ML,ML)]
      { (p_an2 [Dmd=1L], ds_dnz [Dmd=A]) ->
      case p_an2 of _ {
        False -> GHC.Types.True;
        True -> foo wild_X7 }

Note that ds_dnz is syntactically dead, but the expression bound to it is
reachable through the case binder wild_X7. Now watch what happens if we inline
foo's wrapper:
  f = \ (x_an1 [Dmd=1P(1L,ML)] :: (Bool, Bool)) ->
      case x_an1
      of _ [Dmd=MP(ML,ML)]
      { (p_an2 [Dmd=1L], ds_dnz [Dmd=A]) ->
      case p_an2 of _ {
        False -> GHC.Types.True;
        True -> $wfoo_soq GHC.Types.True ds_dnz }

Look at that! ds_dnz has come back to life in the call to $wfoo_soq! A second
run of demand analysis would no longer infer ds_dnz to be absent.
But unlike occurrence analysis, which infers properties of the *syntactic*
shape of the program, the results of demand analysis describe expressions
*semantically* and are supposed to be mostly stable across Simplification.
That's why we should better account for (b).
In #10148, we ended up emitting a single-entry thunk instead of an updateable
thunk for a let binder that was an an absent case-alt binder during DmdAnal.

This is needed even for non-product types, in case the case-binder
is used but the components of the case alternative are not.

Note [Untyped demand on case-alternative binders]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
With unsafeCoerce, #8037 and #22039 taught us that the demand on the case binder
may be a call demand or have a different number of fields than the constructor
of the case alternative it is used in. From T22039:

  blarg :: (Int, Int) -> Int
  blarg (x,y) = x+y
  -- blarg :: <1!P(1L,1L)>

  f :: Either Int Int -> Int
  f Left{} = 0
  f e = blarg (unsafeCoerce e)
  ==> { desugars to }
  f = \ (ds_d1nV :: Either Int Int) ->
      case ds_d1nV of wild_X1 {
        Left ds_d1oV -> lvl_s1Q6;
        Right ipv_s1Pl ->
          blarg
            (case unsafeEqualityProof @(*) @(Either Int Int) @(Int, Int) of
             { UnsafeRefl co_a1oT ->
             wild_X1 `cast` (Sub (Sym co_a1oT) :: Either Int Int ~R# (Int, Int))
             })
      }

The case binder `e`/`wild_X1` has demand 1!P(1L,1L), with two fields, from the call
to `blarg`, but `Right` only has one field. Although the code will crash when
executed, we must be able to analyse it in 'fieldBndrDmds' and conservatively
approximate with Top instead of panicking because of the mismatch.
In #22039, this kind of code was guarded behind a safe `cast` and thus dead
code, but nevertheless led to a panic of the compiler.

You might wonder why the same problem doesn't come up when scrutinising a
product type instead of a sum type. It appears that for products, `wild_X1`
will be inlined before DmdAnal.

See also Note [mkWWstr and unsafeCoerce] for a related issue.

Note [Aggregated demand for cardinality]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
FIXME: This Note should be named [LetUp vs. LetDown] and probably predates
said separation. SG

We use different strategies for strictness and usage/cardinality to
"unleash" demands captured on free variables by bindings. Let us
consider the example:

f1 y = let {-# NOINLINE h #-}
           h = y
       in  (h, h)

We are interested in obtaining cardinality demand U1 on |y|, as it is
used only in a thunk, and, therefore, is not going to be updated any
more. Therefore, the demand on |y|, captured and unleashed by usage of
|h| is U1. However, if we unleash this demand every time |h| is used,
and then sum up the effects, the ultimate demand on |y| will be U1 +
U1 = U. In order to avoid it, we *first* collect the aggregate demand
on |h| in the body of let-expression, and only then apply the demand
transformer:

transf[x](U) = {y |-> U1}

so the resulting demand on |y| is U1.

The situation is, however, different for strictness, where this
aggregating approach exhibits worse results because of the nature of
|both| operation for strictness. Consider the example:

f y c =
  let h x = y |seq| x
   in case of
        True  -> h True
        False -> y

It is clear that |f| is strict in |y|, however, the suggested analysis
will infer from the body of |let| that |h| is used lazily (as it is
used in one branch only), therefore lazy demand will be put on its
free variable |y|. Conversely, if the demand on |h| is unleashed right
on the spot, we will get the desired result, namely, that |f| is
strict in |y|.


************************************************************************
*                                                                      *
                    Demand transformer
*                                                                      *
************************************************************************
-}

dmdTransform :: AnalEnv   -- ^ The analysis environment
             -> Id        -- ^ The variable
             -> SubDemand -- ^ The evaluation context of the var
             -> DmdType   -- ^ The demand type unleashed by the variable in this
                          -- context. The returned DmdEnv includes the demand on
                          -- this function plus demand on its free variables
-- See Note [What are demand signatures?] in "GHC.Types.Demand"
dmdTransform :: AnalEnv -> Id -> SubDemand -> DmdType
dmdTransform AnalEnv
env Id
var SubDemand
sd
  -- Data constructors
  | Just DataCon
con <- Id -> Maybe DataCon
isDataConWorkId_maybe Id
var
  = -- pprTraceWith "dmdTransform:DataCon" (\ty -> ppr con $$ ppr sd $$ ppr ty) $
    [StrictnessMark] -> SubDemand -> DmdType
dmdTransformDataConSig (DataCon -> [StrictnessMark]
dataConRepStrictness DataCon
con) SubDemand
sd
  -- See Note [DmdAnal for DataCon wrappers]
  | Id -> Bool
isDataConWrapId Id
var, let rhs :: CoreExpr
rhs = Unfolding -> CoreExpr
uf_tmpl (Id -> Unfolding
realIdUnfolding Id
var)
  , WithDmdType DmdType
dmd_ty CoreExpr
_rhs' <- AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
sd CoreExpr
rhs
  = DmdType
dmd_ty
  -- Dictionary component selectors
  -- Used to be controlled by a flag.
  -- See #18429 for some perf measurements.
  | Just Class
_ <- Id -> Maybe Class
isClassOpId_maybe Id
var
  = -- pprTrace "dmdTransform:DictSel" (ppr var $$ ppr (idDmdSig var) $$ ppr sd) $
    DmdSig -> SubDemand -> DmdType
dmdTransformDictSelSig (Id -> DmdSig
idDmdSig Id
var) SubDemand
sd
  -- Imported functions
  | Id -> Bool
isGlobalId Id
var
  , let res :: DmdType
res = DmdSig -> SubDemand -> DmdType
dmdTransformSig (Id -> DmdSig
idDmdSig Id
var) SubDemand
sd
  = -- pprTrace "dmdTransform:import" (vcat [ppr var, ppr (idDmdSig var), ppr sd, ppr res])
    DmdType
res
  -- Top-level or local let-bound thing for which we use LetDown ('useLetUp').
  -- In that case, we have a strictness signature to unleash in our AnalEnv.
  | Just (DmdSig
sig, TopLevelFlag
top_lvl) <- AnalEnv -> Id -> Maybe (DmdSig, TopLevelFlag)
lookupSigEnv AnalEnv
env Id
var
  , let fn_ty :: DmdType
fn_ty = DmdSig -> SubDemand -> DmdType
dmdTransformSig DmdSig
sig SubDemand
sd
  = -- pprTrace "dmdTransform:LetDown" (vcat [ppr var, ppr sig, ppr sd, ppr fn_ty]) $
    case TopLevelFlag
top_lvl of
      TopLevelFlag
NotTopLevel -> DmdType -> Id -> Demand -> DmdType
addVarDmd DmdType
fn_ty Id
var (Card
C_11 HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* SubDemand
sd)
      TopLevelFlag
TopLevel
        | Id -> Bool
isInterestingTopLevelFn Id
var
        -- Top-level things will be used multiple times or not at
        -- all anyway, hence the multDmd below: It means we don't
        -- have to track whether @var@ is used strictly or at most
        -- once, because ultimately it never will.
        -> DmdType -> Id -> Demand -> DmdType
addVarDmd DmdType
fn_ty Id
var (Card
C_0N Card -> Demand -> Demand
`multDmd` (Card
C_11 HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* SubDemand
sd)) -- discard strictness
        | Bool
otherwise
        -> DmdType
fn_ty -- don't bother tracking; just annotate with 'topDmd' later
  -- Everything else:
  --   * Local let binders for which we use LetUp (cf. 'useLetUp')
  --   * Lambda binders
  --   * Case and constructor field binders
  | Bool
otherwise
  = -- pprTrace "dmdTransform:other" (vcat [ppr var, ppr boxity, ppr sd]) $
    DmdEnv -> DmdType
noArgsDmdType (DmdEnv -> Id -> Demand -> DmdEnv
addVarDmdEnv DmdEnv
nopDmdEnv Id
var (Card
C_11 HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* SubDemand
sd))

{- *********************************************************************
*                                                                      *
                      Binding right-hand sides
*                                                                      *
********************************************************************* -}

-- | An environment in which all demands are weak according to 'isWeakDmd'.
-- See Note [Lazy and unleashable free variables].
type WeakDmds = VarEnv Demand

-- | @dmdAnalRhsSig@ analyses the given RHS to compute a demand signature
-- for the LetDown rule. It works as follows:
--
--  * assuming the weakest possible body sub-demand, L
--  * looking at the definition
--  * determining a strictness signature
--
-- Since it assumed a body sub-demand of L, the resulting signature is
-- applicable at any call site.
dmdAnalRhsSig
  :: TopLevelFlag
  -> RecFlag
  -> AnalEnv -> SubDemand
  -> Id -> CoreExpr
  -> (AnalEnv, WeakDmds, Id, CoreExpr)
-- Process the RHS of the binding, add the strictness signature
-- to the Id, and augment the environment with the signature as well.
-- See Note [NOINLINE and strictness]
dmdAnalRhsSig :: TopLevelFlag
-> RecFlag
-> AnalEnv
-> SubDemand
-> Id
-> CoreExpr
-> (AnalEnv, WeakDmds, Id, CoreExpr)
dmdAnalRhsSig TopLevelFlag
top_lvl RecFlag
rec_flag AnalEnv
env SubDemand
let_dmd Id
id CoreExpr
rhs
  = -- pprTrace "dmdAnalRhsSig" (ppr id $$ ppr let_dmd $$ ppr rhs_dmds $$ ppr sig $$ ppr weak_fvs) $
    (AnalEnv
final_env, WeakDmds
weak_fvs, Id
final_id, CoreExpr
final_rhs)
  where
    threshold_arity :: Arity
threshold_arity = Id -> CoreExpr -> Arity
thresholdArity Id
id CoreExpr
rhs

    rhs_dmd :: SubDemand
rhs_dmd = Arity -> SubDemand -> SubDemand
mkCalledOnceDmds Arity
threshold_arity SubDemand
body_dmd

    body_dmd :: SubDemand
body_dmd
      | Id -> Bool
isJoinId Id
id
      -- See Note [Demand analysis for join points]
      -- See Note [Invariants on join points] invariant 2b, in GHC.Core
      --     threshold_arity matches the join arity of the join point
      -- See Note [Unboxed demand on function bodies returning small products]
      = AnalEnv -> RecFlag -> Maybe Type -> SubDemand -> SubDemand
unboxedWhenSmall AnalEnv
env RecFlag
rec_flag (Id -> Maybe Type
resultType_maybe Id
id) SubDemand
let_dmd
      | Bool
otherwise
      -- See Note [Unboxed demand on function bodies returning small products]
      = AnalEnv -> RecFlag -> Maybe Type -> SubDemand -> SubDemand
unboxedWhenSmall AnalEnv
env RecFlag
rec_flag (Id -> Maybe Type
resultType_maybe Id
id) SubDemand
topSubDmd

    WithDmdType DmdType
rhs_dmd_ty CoreExpr
rhs' = AnalEnv -> SubDemand -> CoreExpr -> WithDmdType CoreExpr
dmdAnal AnalEnv
env SubDemand
rhs_dmd CoreExpr
rhs
    DmdType DmdEnv
rhs_env [Demand]
rhs_dmds = DmdType
rhs_dmd_ty
    ([Demand]
final_rhs_dmds, CoreExpr
final_rhs) = AnalEnv
-> Id
-> Arity
-> [Demand]
-> Divergence
-> CoreExpr
-> ([Demand], CoreExpr)
finaliseArgBoxities AnalEnv
env Id
id Arity
threshold_arity
                                                      [Demand]
rhs_dmds (DmdEnv -> Divergence
de_div DmdEnv
rhs_env) CoreExpr
rhs'

    sig :: DmdSig
sig = Arity -> DmdType -> DmdSig
mkDmdSigForArity Arity
threshold_arity (DmdEnv -> [Demand] -> DmdType
DmdType DmdEnv
sig_env [Demand]
final_rhs_dmds)

    opts :: DmdAnalOpts
opts       = AnalEnv -> DmdAnalOpts
ae_opts AnalEnv
env
    final_id :: Id
final_id   = DmdAnalOpts -> Id -> DmdSig -> Id
setIdDmdAndBoxSig DmdAnalOpts
opts Id
id DmdSig
sig
    !final_env :: AnalEnv
final_env = TopLevelFlag -> AnalEnv -> Id -> DmdSig -> AnalEnv
extendAnalEnv TopLevelFlag
top_lvl AnalEnv
env Id
final_id DmdSig
sig

    -- See Note [Aggregated demand for cardinality]
    -- FIXME: That Note doesn't explain the following lines at all. The reason
    --        is really much different: When we have a recursive function, we'd
    --        have to also consider the free vars of the strictness signature
    --        when checking whether we found a fixed-point. That is expensive;
    --        we only want to check whether argument demands of the sig changed.
    --        reuseEnv makes it so that the FV results are stable as long as the
    --        last argument demands were. Strictness won't change. But used-once
    --        might turn into used-many even if the signature was stable and
    --        we'd have to do an additional iteration. reuseEnv makes sure that
    --        we never get used-once info for FVs of recursive functions.
    --        See #14816 where we try to get rid of reuseEnv.
    rhs_env1 :: DmdEnv
rhs_env1 = case RecFlag
rec_flag of
                RecFlag
Recursive    -> DmdEnv -> DmdEnv
reuseEnv DmdEnv
rhs_env
                RecFlag
NonRecursive -> DmdEnv
rhs_env

    -- See Note [Absence analysis for stable unfoldings and RULES]
    rhs_env2 :: DmdEnv
rhs_env2 = DmdEnv
rhs_env1 DmdEnv -> DmdEnv -> DmdEnv
`plusDmdEnv` AnalEnv -> VarSet -> DmdEnv
demandRootSet AnalEnv
env (Id -> VarSet
bndrRuleAndUnfoldingIds Id
id)

    -- See Note [Lazy and unleashable free variables]
    !(!DmdEnv
sig_env, !WeakDmds
weak_fvs) = DmdEnv -> (DmdEnv, WeakDmds)
splitWeakDmds DmdEnv
rhs_env2

splitWeakDmds :: DmdEnv -> (DmdEnv, WeakDmds)
splitWeakDmds :: DmdEnv -> (DmdEnv, WeakDmds)
splitWeakDmds (DE WeakDmds
fvs Divergence
div) = (WeakDmds -> Divergence -> DmdEnv
DE WeakDmds
sig_fvs Divergence
div, WeakDmds
weak_fvs)
  where (!WeakDmds
weak_fvs, !WeakDmds
sig_fvs) = (Demand -> Bool) -> WeakDmds -> (WeakDmds, WeakDmds)
forall a. (a -> Bool) -> VarEnv a -> (VarEnv a, VarEnv a)
partitionVarEnv Demand -> Bool
isWeakDmd WeakDmds
fvs

thresholdArity :: Id -> CoreExpr -> Arity
-- See Note [Demand signatures are computed for a threshold arity based on idArity]
thresholdArity :: Id -> CoreExpr -> Arity
thresholdArity Id
fn CoreExpr
rhs
  = case Id -> Maybe Arity
isJoinId_maybe Id
fn of
      Just Arity
join_arity -> (Id -> Bool) -> [Id] -> Arity
forall a. (a -> Bool) -> [a] -> Arity
count Id -> Bool
isId ([Id] -> Arity) -> [Id] -> Arity
forall a b. (a -> b) -> a -> b
$ ([Id], CoreExpr) -> [Id]
forall a b. (a, b) -> a
fst (([Id], CoreExpr) -> [Id]) -> ([Id], CoreExpr) -> [Id]
forall a b. (a -> b) -> a -> b
$ Arity -> CoreExpr -> ([Id], CoreExpr)
forall b. Arity -> Expr b -> ([b], Expr b)
collectNBinders Arity
join_arity CoreExpr
rhs
      Maybe Arity
Nothing         -> Id -> Arity
idArity Id
fn

-- | The result type after applying 'idArity' many arguments. Returns 'Nothing'
-- when the type doesn't have exactly 'idArity' many arrows.
resultType_maybe :: Id -> Maybe Type
resultType_maybe :: Id -> Maybe Type
resultType_maybe Id
id
  | ([PiTyBinder]
pis,Type
ret_ty) <- Type -> ([PiTyBinder], Type)
splitPiTys (Id -> Type
idType Id
id)
  , (PiTyBinder -> Bool) -> [PiTyBinder] -> Arity
forall a. (a -> Bool) -> [a] -> Arity
count PiTyBinder -> Bool
isAnonPiTyBinder [PiTyBinder]
pis Arity -> Arity -> Bool
forall a. Eq a => a -> a -> Bool
== Id -> Arity
idArity Id
id
  = Type -> Maybe Type
forall a. a -> Maybe a
Just (Type -> Maybe Type) -> Type -> Maybe Type
forall a b. (a -> b) -> a -> b
$! Type
ret_ty
  | Bool
otherwise
  = Maybe Type
forall a. Maybe a
Nothing

unboxedWhenSmall :: AnalEnv -> RecFlag -> Maybe Type -> SubDemand -> SubDemand
-- See Note [Unboxed demand on function bodies returning small products]
unboxedWhenSmall :: AnalEnv -> RecFlag -> Maybe Type -> SubDemand -> SubDemand
unboxedWhenSmall AnalEnv
_   RecFlag
_        Maybe Type
Nothing       SubDemand
sd = SubDemand
sd
unboxedWhenSmall AnalEnv
env RecFlag
rec_flag (Just Type
ret_ty) SubDemand
sd = Arity -> Type -> SubDemand -> SubDemand
go Arity
1 Type
ret_ty SubDemand
sd
  where
    -- Magic constant, bounding the depth of optimistic 'Unboxed' flags. We
    -- might want to minmax in the future.
    max_depth :: Arity
max_depth | RecFlag -> Bool
isRec RecFlag
rec_flag = Arity
3 -- So we get at most something as deep as !P(L!P(L!L))
              | Bool
otherwise      = Arity
1 -- Otherwise be unbox too deep in T18109, T18174 and others and get a bunch of stack overflows
    go :: Int -> Type -> SubDemand -> SubDemand
    go :: Arity -> Type -> SubDemand -> SubDemand
go Arity
depth Type
ty SubDemand
sd
      | Arity
depth Arity -> Arity -> Bool
forall a. Ord a => a -> a -> Bool
<= Arity
max_depth
      , Just (TyCon
tc, [Type]
tc_args, Coercion
_co) <- FamInstEnvs -> Type -> Maybe (TyCon, [Type], Coercion)
normSplitTyConApp_maybe (AnalEnv -> FamInstEnvs
ae_fam_envs AnalEnv
env) Type
ty
      , Just DataCon
dc <- TyCon -> Maybe DataCon
tyConSingleAlgDataCon_maybe TyCon
tc
      , [Id] -> Bool
forall a. [a] -> Bool
forall (t :: * -> *) a. Foldable t => t a -> Bool
null (DataCon -> [Id]
dataConExTyCoVars DataCon
dc) -- Can't unbox results with existentials
      , DataCon -> Arity
dataConRepArity DataCon
dc Arity -> Arity -> Bool
forall a. Ord a => a -> a -> Bool
<= DmdAnalOpts -> Arity
dmd_unbox_width (AnalEnv -> DmdAnalOpts
ae_opts AnalEnv
env)
      , Just (Boxity
_, [Demand]
ds) <- Arity -> SubDemand -> Maybe (Boxity, [Demand])
viewProd (DataCon -> Arity
dataConRepArity DataCon
dc) SubDemand
sd
      , [Type]
arg_tys <- (Scaled Type -> Type) -> [Scaled Type] -> [Type]
forall a b. (a -> b) -> [a] -> [b]
map Scaled Type -> Type
forall a. Scaled a -> a
scaledThing ([Scaled Type] -> [Type]) -> [Scaled Type] -> [Type]
forall a b. (a -> b) -> a -> b
$ DataCon -> [Type] -> [Scaled Type]
dataConInstArgTys DataCon
dc [Type]
tc_args
      , [Demand] -> [Type] -> Bool
forall a b. [a] -> [b] -> Bool
equalLength [Demand]
ds [Type]
arg_tys
      = Boxity -> [Demand] -> SubDemand
mkProd Boxity
Unboxed ([Demand] -> SubDemand) -> [Demand] -> SubDemand
forall a b. (a -> b) -> a -> b
$! (Type -> Demand -> Demand) -> [Type] -> [Demand] -> [Demand]
forall a b c. (a -> b -> c) -> [a] -> [b] -> [c]
strictZipWith (Arity -> Type -> Demand -> Demand
go_dmd (Arity
depthArity -> Arity -> Arity
forall a. Num a => a -> a -> a
+Arity
1)) [Type]
arg_tys [Demand]
ds
      | Bool
otherwise
      = SubDemand
sd

    go_dmd :: Int -> Type -> Demand -> Demand
    go_dmd :: Arity -> Type -> Demand -> Demand
go_dmd Arity
depth Type
ty Demand
dmd = case Demand
dmd of
      Demand
AbsDmd  -> Demand
AbsDmd
      Demand
BotDmd  -> Demand
BotDmd
      Card
n :* SubDemand
sd -> Card
n HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* Arity -> Type -> SubDemand -> SubDemand
go Arity
depth Type
ty SubDemand
sd

-- | If given the (local, non-recursive) let-bound 'Id', 'useLetUp' determines
-- whether we should process the binding up (body before rhs) or down (rhs
-- before body).
--
-- We use LetDown if there is a chance to get a useful strictness signature to
-- unleash at call sites. LetDown is generally more precise than LetUp if we can
-- correctly guess how it will be used in the body, that is, for which incoming
-- demand the strictness signature should be computed, which allows us to
-- unleash higher-order demands on arguments at call sites. This is mostly the
-- case when
--
--   * The binding takes any arguments before performing meaningful work (cf.
--     'idArity'), in which case we are interested to see how it uses them.
--   * The binding is a join point, hence acting like a function, not a value.
--     As a big plus, we know *precisely* how it will be used in the body; since
--     it's always tail-called, we can directly unleash the incoming demand of
--     the let binding on its RHS when computing a strictness signature. See
--     [Demand analysis for join points].
--
-- Thus, if the binding is not a join point and its arity is 0, we have a thunk
-- and use LetUp, implying that we have no usable demand signature available
-- when we analyse the let body.
--
-- Since thunk evaluation is memoised, we want to unleash its 'DmdEnv' of free
-- vars at most once, regardless of how many times it was forced in the body.
-- This makes a real difference wrt. usage demands. The other reason is being
-- able to unleash a more precise product demand on its RHS once we know how the
-- thunk was used in the let body.
--
-- Characteristic examples, always assuming a single evaluation:
--
--   * @let x = 2*y in x + x@ => LetUp. Compared to LetDown, we find out that
--     the expression uses @y@ at most once.
--   * @let x = (a,b) in fst x@ => LetUp. Compared to LetDown, we find out that
--     @b@ is absent.
--   * @let f x = x*2 in f y@ => LetDown. Compared to LetUp, we find out that
--     the expression uses @y@ strictly, because we have @f@'s demand signature
--     available at the call site.
--   * @join exit = 2*y in if a then exit else if b then exit else 3*y@ =>
--     LetDown. Compared to LetUp, we find out that the expression uses @y@
--     strictly, because we can unleash @exit@'s signature at each call site.
--   * For a more convincing example with join points, see Note [Demand analysis
--     for join points].
--
useLetUp :: TopLevelFlag -> Var -> Bool
useLetUp :: TopLevelFlag -> Id -> Bool
useLetUp TopLevelFlag
top_lvl Id
f = TopLevelFlag -> Bool
isNotTopLevel TopLevelFlag
top_lvl Bool -> Bool -> Bool
&& Id -> Arity
idArity Id
f Arity -> Arity -> Bool
forall a. Eq a => a -> a -> Bool
== Arity
0 Bool -> Bool -> Bool
&& Bool -> Bool
not (Id -> Bool
isJoinId Id
f)

{- Note [Demand analysis for join points]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
   g :: (Int,Int) -> Int
   g (p,q) = p+q

   f :: T -> Int -> Int
   f x p = g (join j y = (p,y)
              in case x of
                   A -> j 3
                   B -> j 4
                   C -> (p,7))

If j was a vanilla function definition, we'd analyse its body with
evalDmd, and think that it was lazy in p.  But for join points we can
do better!  We know that j's body will (if called at all) be evaluated
with the demand that consumes the entire join-binding, in this case
the argument demand from g.  Whizzo!  g evaluates both components of
its argument pair, so p will certainly be evaluated if j is called.

For f to be strict in p, we need /all/ paths to evaluate p; in this
case the C branch does so too, so we are fine.  So, as usual, we need
to transport demands on free variables to the call site(s).  Compare
Note [Lazy and unleashable free variables].

The implementation is easy.  When analysing a join point, we can
analyse its body with the demand from the entire join-binding (written
let_dmd here).

Another win for join points!  #13543.

However, note that the strictness signature for a join point can
look a little puzzling.  E.g.

    (join j x = \y. error "urk")
    (in case v of              )
    (     A -> j 3             )  x
    (     B -> j 4             )
    (     C -> \y. blah        )

The entire thing is in a C(1,L) context, so j's strictness signature
will be    [A]b
meaning one absent argument, returns bottom.  That seems odd because
there's a \y inside.  But it's right because when consumed in a C(1,L)
context the RHS of the join point is indeed bottom.

Note [Demand signatures are computed for a threshold arity based on idArity]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Given a binding { f = rhs }, we compute a "theshold arity", and do demand
analysis based on a call with that many value arguments.

The threshold we use is

* Ordinary bindings: idArity f.
  Why idArity arguments? Because that's a conservative estimate of how many
  arguments we must feed a function before it does anything interesting with
  them.  Also it elegantly subsumes the trivial RHS and PAP case.  E.g. for
      f = g
  we want to use a threshold arity based on g, not 0!

  idArity is /at least/ the number of manifest lambdas, but might be higher for
  PAPs and trivial RHS (see Note [Demand analysis for trivial right-hand sides]).

* Join points: the value-binder subset of the JoinArity.  This can
  be less than the number of visible lambdas; e.g.
     join j x = \y. blah
     in ...(jump j 2)....(jump j 3)....
  We know that j will never be applied to more than 1 arg (its join
  arity, and we don't eta-expand join points, so here a threshold
  of 1 is the best we can do.

Note that the idArity of a function varies independently of its cardinality
properties (cf. Note [idArity varies independently of dmdTypeDepth]), so we
implicitly encode the arity for when a demand signature is sound to unleash
in its 'dmdTypeDepth', not in its idArity (cf. Note [Understanding DmdType
and DmdSig] in GHC.Types.Demand). It is unsound to unleash a demand
signature when the incoming number of arguments is less than that. See
GHC.Types.Demand Note [What are demand signatures?]  for more details on
soundness.

Note that there might, in principle, be functions for which we might want to
analyse for more incoming arguments than idArity. Example:

  f x =
    if expensive
      then \y -> ... y ...
      else \y -> ... y ...

We'd analyse `f` under a unary call demand C(1,L), corresponding to idArity
being 1. That's enough to look under the manifest lambda and find out how a
unary call would use `x`, but not enough to look into the lambdas in the if
branches.

On the other hand, if we analysed for call demand C(1,C(1,L)), we'd get useful
strictness info for `y` (and more precise info on `x`) and possibly CPR
information, but

  * We would no longer be able to unleash the signature at unary call sites

  * Performing the worker/wrapper split based on this information would be
    implicitly eta-expanding `f`, playing fast and loose with divergence and
    even being unsound in the presence of newtypes, so we refrain from doing so.
    Also see Note [Don't eta expand in w/w] in GHC.Core.Opt.WorkWrap.

Since we only compute one signature, we do so for arity 1. Computing multiple
signatures for different arities (i.e., polyvariance) would be entirely
possible, if it weren't for the additional runtime and implementation
complexity.

Note [idArity varies independently of dmdTypeDepth]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In general, an Id `f` has two independently varying attributes:

* f's idArity, and
* the dmdTypeDepth of f's demand signature

For example, if f's demand signature is <L><L>, f's arity could be
greater than, or less than 2. Why?  Because both are conservative
approximations:

* Arity n means "does no expensive work until applied to at least n args"
  (e.g. (f x1..xm) is cheap to bring to HNF for m<n)

* Dmd sig with n args means "here is how to transform the incoming demand
  when applied to n args".  This is /semantic/ property, unrelated to
  arity. See GHC.Types.Demand Note [Understanding DmdType and DmdSig]

We used to check in GHC.Core.Lint that dmdTypeDepth <= idArity for a let-bound
identifier. But that means we would have to zap demand signatures every time we
reset or decrease arity.

For example, consider the following expression:

    (let go x y = `x` seq ... in go) |> co

`go` might have a strictness signature of `<1L><L>`. The simplifier will identify
`go` as a nullary join point through `joinPointBinding_maybe` and float the
coercion into the binding, leading to an arity decrease:

    join go = (\x y -> `x` seq ...) |> co in go

With the CoreLint check, we would have to zap `go`'s perfectly viable strictness
signature.

However, in the case of a /bottoming/ signature, f : <L><L>b, we /can/
say that f's arity is no greater than 2, because it'd be false to say
that f does no work when applied to 3 args.  Lint checks this constraint,
in `GHC.Core.Lint.lintLetBind`.

Note [Demand analysis for trivial right-hand sides]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
    foo = plusInt |> co
where plusInt is an arity-2 function with known strictness.  Clearly
we want plusInt's strictness to propagate to foo!  But because it has
no manifest lambdas, it won't do so automatically, and indeed 'co' might
have type (Int->Int->Int) ~ T.

Fortunately, GHC.Core.Opt.Arity gives 'foo' arity 2, which is enough for LetDown to
forward plusInt's demand signature, and all is well (see Note [Newtype arity] in
GHC.Core.Opt.Arity)! A small example is the test case NewtypeArity.

Note [Absence analysis for stable unfoldings and RULES]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Among others, tickets #18638 and #23208 show that it's really important to treat
stable unfoldings as demanded. Consider

   g = blah

   f = \x.  ...no use of g....
   {- f's stable unfolding is f = \x. ...g... -}

If f is ever inlined we use 'g'. But f's current RHS makes no use
of 'g', so if we don't look at the unfolding we'll mark g as Absent,
and transform to

   g = error "Entered absent value"
   f = \x. ...
   {- f's stable unfolding is f = \x. ...g... -}

Now if f is subsequently inlined, we'll use 'g' and ... disaster.

SOLUTION: if f has a stable unfolding, treat every free variable as a
/demand root/, that is: Analyse it as if it was a variable occuring in a
'topDmd' context. This is done in `demandRoot` (which we also use for exported
top-level ids). Do the same for Ids free in the RHS of any RULES for f.

Wrinkles:

  (W1) You may wonder how it can be that f's optimised RHS has somehow
    discarded 'g', but when f is inlined we /don't/ discard g in the same
    way. I think a simple example is
       g = (a,b)
       f = \x.  fst g
       {-# INLINE f #-}

    Now f's optimised RHS will be \x.a, but if we change g to (error "..")
    (since it is apparently Absent) and then inline (\x. fst g) we get
    disaster.  But regardless, #18638 was a more complicated version of
    this, that actually happened in practice.

  (W2) You might wonder why we don't simply take the free vars of the
    unfolding/RULE and map them to topDmd. The reason is that any of the free vars
    might have demand signatures themselves that in turn demand transitive free
    variables and that we hence need to unleash! This came up in #23208.
    Consider

       err :: Int -> b
       err = error "really important message"

       sg :: Int -> Int
       sg _ = case err of {}  -- Str=<1B>b {err:->S}

       g :: a -> a  -- g is exported
       g x = x
       {-# RULES "g" g @Int = sg #-}

    Here, `err` is only demanded by `sg`'s demand signature: It doesn't occur
    in the weak_fvs of `sg`'s RHS at all. Hence when we `demandRoots` `sg`
    because it occurs in the RULEs of `g` (which is exported), we better unleash
    the demand signature of `sg`, too! Before #23208 we simply added a 'topDmd'
    for `sg`, failing to unleash the signature and hence observed an absent
    error instead of the `really important message`.

Note [DmdAnal for DataCon wrappers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We give DataCon wrappers a (necessarily flat) demand signature in
`GHC.Types.Id.Make.mkDataConRep`, so that passes such as the Simplifier can
exploit it via the call to `GHC.Core.Opt.Simplify.Utils.isStrictArgInfo` in
`GHC.Core.Opt.Simplify.Iteration.rebuildCall`. But during DmdAnal, we *ignore*
the demand signature of a DataCon wrapper, and instead analyse its unfolding at
every call site.

The reason is that DataCon *worker*s have very precise demand transformers,
computed by `dmdTransformDataConSig`. It would be awkward if DataCon *wrappers*
would behave much less precisely during DmdAnal. Example:

   data T1 = MkT1 { get_x1 :: Int,  get_y1 :: Int }
   data T2 = MkT2 { get_x2 :: !Int, get_y2 :: Int }
   f1 x y = get_x1 (MkT1 x y)
   f2 x y = get_x2 (MkT2 x y)

Here `MkT1` has no wrapper. `get_x1` puts a demand `!P(1!L,A)` on its argument,
and `dmdTransformDataConSig` will transform that demand to an absent demand on
`y` in `f1` and an unboxing demand on `x`.
But `MkT2` has a wrapper (to evaluate the first field). If demand analysis deals
with `MkT2` only through its demand signature, demand signatures can't transform
an incoming demand `P(1!L,A)` in a useful way, so we won't get an absent demand
on `y` in `f2` or see that `x` can be unboxed. That's a serious loss.

The example above will not actually occur, because $WMkT2 would be inlined.
Nevertheless, we can get interesting sub-demands on DataCon wrapper
applications in boring contexts; see T22241.

You might worry about the efficiency cost of demand-analysing datacon wrappers
at every call site. But in fact they are inlined /anyway/ in the Final phase,
which happens before DmdAnal, so few wrappers remain. And analysing the
unfoldings for the remaining calls (which are those in a boring context) will be
exactly as (in)efficent as if we'd inlined those calls. It turns out to be not
measurable in practice.

See also Note [CPR for DataCon wrappers] in `GHC.Core.Opt.CprAnal`.

Note [Boxity for bottoming functions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider (A)
    indexError :: Show a => (a, a) -> a -> String -> b
    -- Str=<..><1!P(S,S)><1S><S>b
    indexError rng i s = error (show rng ++ show i ++ show s)

    get :: (Int, Int) -> Int -> [a] -> a
    get p@(l,u) i xs
      | l <= i, i < u = xs !! (i-u)
      | otherwise     = indexError p i "get"

The hot path of `get` certainly wants to unbox `p` as well as `l` and
`u`, but the unimportant, diverging error path needs `l::a` and `u::a`
boxed, since `indexError` can't unbox them because they are polymorphic.
This pattern often occurs in performance sensitive code that does
bounds-checking.

So we want to give `indexError` a signature like `<1!P(!S,!S)><1!S><S!S>b`
where the !S (meaning Poly Unboxed C1N) says that the polymorphic arguments
are unboxed (recursively).  The wrapper for `indexError` won't /actually/
unbox them (because their polymorphic type doesn't allow that) but when
demand-analysing /callers/, we'll behave as if that call needs the args
unboxed.

Then at call sites of `indexError`, we will end up doing some
reboxing, because `$windexError` still takes boxed arguments. This
reboxing should usually float into the slow, diverging code path; but
sometimes (sadly) it doesn't: see Note [Reboxed crud for bottoming calls].

Here is another important case (B):
    f x = Just x  -- Suppose f is not inlined for some reason
                  -- Main point: f takes its argument boxed

    wombat x = error (show (f x))

    g :: Bool -> Int -> a
    g True  x = x+1
    g False x = wombat x

Again we want `wombat` to pretend to take its Int-typed argument unboxed,
even though it has to pass it boxed to `f`, so that `g` can take its
argument unboxed (and rebox it before calling `wombat`).

So here's what we do: while summarising `indexError`'s boxity signature in
`finaliseArgBoxities`:

* To address (B), for bottoming functions, we start by using `unboxDeeplyDmd`
  to make all its argument demands unboxed, right to the leaves; regardless
  of what the analysis said.

* To address (A), for bottoming functions, in the DontUnbox case when the
  argument is a type variable, we /refrain/ from using trimBoxity.
  (Remember the previous bullet: we have already doen `unboxDeeplyDmd`.)

Wrinkle:

* Remember Note [No lazy, Unboxed demands in demand signature]. So
  unboxDeeplyDmd doesn't recurse into lazy demands.  It's extremely unusual
  to have lazy demands in the arguments of a bottoming function anyway.
  But it can happen, when the demand analyser gives up because it
  encounters a recursive data type; see Note [Demand analysis for recursive
  data constructors].

Note [Reboxed crud for bottoming calls]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
For functions like `get` in Note [Boxity for bottoming functions], it's clear
that the reboxed crud will be floated inside to the call site of `$windexError`.
But here's an example where that is not the case:
```hs
import GHC.Ix

theresCrud :: Int -> Int -> Int
theresCrud x y = go x
  where
    go 0 = index (0,y) 0
    go 1 = index (x,y) 1
    go n = go (n-1)
    {-# NOINLINE theresCrud #-}
```
If you look at the Core, you'll see that `y` will be reboxed and used in the
two exit join points for the `$windexError` calls, while `x` is only reboxed in the
exit join point for `index (x,y) 1` (happens in lvl below):
```
$wtheresCrud = \ ww ww1 ->
      let { y = I# ww1 } in
      join { lvl2 = ... case lvl1 ww y of wild { }; ... } in
      join { lvl3 = ... case lvl y of wild { }; ... } in
      ...
```
This is currently a bug that we willingly accept and it's documented in #21128.

See also Note [indexError] in base:GHC.Ix, which describes how we use
SPECIALISE to mitigate this problem for indexError.
-}

{- *********************************************************************
*                                                                      *
             Finalising boxity
*                                                                      *
********************************************************************* -}

{- Note [Finalising boxity for demand signatures]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The worker/wrapper pass must strictly adhere to the boxity decisions
encoded in the demand signature, because that is the information that
demand analysis propagates throughout the program. Failing to
implement the strategy laid out in the signature can result in
reboxing in unexpected places. Hence, we must completely anticipate
unboxing decisions during demand analysis and reflect these decisions
in demand annotations. That is the job of 'finaliseArgBoxities',
which is defined here and called from demand analysis.

Here is a list of different Notes it has to take care of:

  * Note [No lazy, Unboxed demands in demand signature] such as `L!P(L)` in
    general, but still allow Note [Unboxing evaluated arguments]
  * Note [No nested Unboxed inside Boxed in demand signature] such as `1P(1!L)`
  * Note [mkWWstr and unsafeCoerce]

NB: Then, the worker/wrapper blindly trusts the boxity info in the
demand signature; that is why 'canUnboxArg' does not look at
strictness -- it is redundant to do so.

Note [Finalising boxity for let-bound Ids]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
  let x = e in body
where the demand on 'x' is 1!P(blah).  We want to unbox x according to
Note [Thunk splitting] in GHC.Core.Opt.WorkWrap.  We must do this because
worker/wrapper ignores strictness and looks only at boxity flags; so if
x's demand is L!P(blah) we might still split it (wrongly).  We want to
switch to Boxed on any lazy demand.

That is what finaliseLetBoxity does.  It has no worker-arg budget, so it
is much simpler than finaliseArgBoxities.

Note [No nested Unboxed inside Boxed in demand signature]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
```
f p@(x,y)
  | even (x+y) = []
  | otherwise  = [p]
```
Demand analysis will infer that the function body puts a demand of `1P(1!L,1!L)`
on 'p', e.g., Boxed on the outside but Unboxed on the inside. But worker/wrapper
can't unbox the pair components without unboxing the pair! So we better say
`1P(1L,1L)` in the demand signature in order not to spread wrong Boxity info.
That happens via the call to trimBoxity in 'finaliseArgBoxities'/'finaliseLetBoxity'.

Note [No lazy, Unboxed demands in demand signature]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider T19407:

  data Huge = Huge Bool () ... () -- think: DynFlags
  data T = T { h :: Huge, n :: Int }
  f t@(T h _) = g h t
  g (H b _ ... _) t = if b then 1 else n t

The body of `g` puts (approx.) demand `L!P(A,1)` on `t`. But we better
not put that demand in `g`'s demand signature, because worker/wrapper will not
in general unbox a lazy-and-unboxed demand like `L!P(..)`.
(The exception are known-to-be-evaluated arguments like strict fields,
see Note [Unboxing evaluated arguments].)

The program above is an example where spreading misinformed boxity through the
signature is particularly egregious. If we give `g` that signature, then `f`
puts demand `S!P(1!P(1L,A,..),ML)` on `t`. Now we will unbox `t` in `f` it and
we get

  f (T (H b _ ... _) n) = $wf b n
  $wf b n = $wg b (T (H b x ... x) n)
  $wg = ...

Massive reboxing in `$wf`! Solution: Trim boxity on lazy demands in
'trimBoxity', modulo Note [Unboxing evaluated arguments].

Note [Unboxing evaluated arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this program (due to Roman):

    data X a = X !a

    foo :: X Int -> Int -> Int
    foo x@(X a) n = go 0
     where
       go i | i < n     = a + go (i+1)
            | otherwise = 0

We want the worker for 'foo' to look like this:

    $wfoo :: Int# -> Int# -> Int#

with the first argument unboxed, so that it is not eval'd each time around the
'go' loop (which would otherwise happen, since 'foo' is not strict in 'a'). It
is sound for the wrapper to pass an unboxed arg because X is strict
(see Note [Strictness and Unboxing] in "GHC.Core.Opt.DmdAnal"), so its argument
must be evaluated. And if we *don't* pass an unboxed argument, we can't even
repair it by adding a `seq` thus:

    foo (X a) n = a `seq` go 0

because the seq is discarded (very early) since X is strict!

So here's what we do

* Since this has nothing to do with how 'foo' uses 'a', we leave demand
  analysis alone, but account for the additional evaluatedness when
  annotating the binder 'finaliseArgBoxities', which will retain the Unboxed
  boxity on 'a' in the definition of 'foo' in the demand 'L!P(L)'; meaning
  it's used lazily but unboxed nonetheless. This seems to contradict Note
  [No lazy, Unboxed demands in demand signature], but we know that 'a' is
  evaluated and thus can be unboxed.

* When 'finaliseArgBoxities' decides to unbox a record, it will zip the field demands
  together with the respective 'StrictnessMark'. In case of 'x', it will pair
  up the lazy field demand 'L!P(L)' on 'a' with 'MarkedStrict' to account for
  the strict field.

* Said 'StrictnessMark' is passed to the recursive invocation of 'go_args' in
  'finaliseArgBoxities' when deciding whether to unbox 'a'. 'a' was used lazily, but
  since it also says 'MarkedStrict', we'll retain the 'Unboxed' boxity on 'a'.

* Worker/wrapper will consult 'canUnboxArg' for its unboxing decision. It will
  /not/ look at the strictness bits of the demand, only at Boxity flags. As such,
  it will happily unbox 'a' despite the lazy demand on it.

The net effect is that boxity analysis and the w/w transformation are more
aggressive about unboxing the strict arguments of a data constructor than when
looking at strictness info exclusively. It is very much like (Nested) CPR, which
needs its nested fields to be evaluated in order for it to unbox nestedly.

There is the usual danger of reboxing, which as usual we ignore. But
if X is monomorphic, and has an UNPACK pragma, then this optimisation
is even more important.  We don't want the wrapper to rebox an unboxed
argument, and pass an Int to $wfoo!

This works in nested situations like T10482

    data family Bar a
    data instance Bar (a, b) = BarPair !(Bar a) !(Bar b)
    newtype instance Bar Int = Bar Int

    foo :: Bar ((Int, Int), Int) -> Int -> Int
    foo f k = case f of BarPair x y ->
              case burble of
                 True -> case x of
                           BarPair p q -> ...
                 False -> ...

The extra eagerness lets us produce a worker of type:
     $wfoo :: Int# -> Int# -> Int# -> Int -> Int
     $wfoo p# q# y# = ...

even though the `case x` is only lazily evaluated.

--------- Historical note ------------
We used to add data-con strictness demands when demand analysing case
expression. However, it was noticed in #15696 that this misses some cases. For
instance, consider the program (from T10482)

    data family Bar a
    data instance Bar (a, b) = BarPair !(Bar a) !(Bar b)
    newtype instance Bar Int = Bar Int

    foo :: Bar ((Int, Int), Int) -> Int -> Int
    foo f k =
      case f of
        BarPair x y -> case burble of
                          True -> case x of
                                    BarPair p q -> ...
                          False -> ...

We really should be able to assume that `p` is already evaluated since it came
from a strict field of BarPair. This strictness would allow us to produce a
worker of type:

    $wfoo :: Int# -> Int# -> Int# -> Int -> Int
    $wfoo p# q# y# = ...

even though the `case x` is only lazily evaluated

Indeed before we fixed #15696 this would happen since we would float the inner
`case x` through the `case burble` to get:

    foo f k =
      case f of
        BarPair x y -> case x of
                          BarPair p q -> case burble of
                                          True -> ...
                                          False -> ...

However, after fixing #15696 this could no longer happen (for the reasons
discussed in ticket:15696#comment:76). This means that the demand placed on `f`
would then be significantly weaker (since the False branch of the case on
`burble` is not strict in `p` or `q`).

Consequently, we now instead account for data-con strictness in mkWWstr_one,
applying the strictness demands to the final result of DmdAnal. The result is
that we get the strict demand signature we wanted even if we can't float
the case on `x` up through the case on `burble`.

Note [Worker argument budget]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In 'finaliseArgBoxities' we don't want to generate workers with zillions of
argument when, say given a strict record with zillions of fields.  So we
limit the maximum number of worker args ('max_wkr_args') to the maximum of
  - -fmax-worker-args=N
  - The number of args in the original function; if it already has has
    zillions of arguments we don't want to seek /fewer/ args in the worker.
(Maybe we should /add/ them instead of maxing?)

We pursue a "layered" strategy for unboxing: we unbox the top level of the
argument(s), subject to budget; if there are any arguments left we unbox the
next layer, using that depleted budget.
Unboxing an argument *increases* the budget for the inner layer roughly
according to how many registers that argument takes (unboxed tuples take
multiple registers, see below), as determined by 'unariseArity'.
Budget is spent when we have to pass a non-absent field as a parameter.

To achieve this, we use the classic almost-circular programming technique in
which we we write one pass that takes a lazy list of the Budgets for every
layer. The effect is that of a breadth-first search (over argument type and
demand structure) to compute Budgets followed by a depth-first search to
construct the product demands, but laziness allows us to do it all in one
pass and without intermediate data structures.

Suppose we have -fmax-worker-args=4 for the remainder of this Note.
Then consider this example function:

  boxed :: (Int, Int) -> (Int, (Int, Int, Int)) -> Int
  boxed (a,b) (c, (d,e,f)) = a + b + c + d + e + f

With a budget of 4 args to spend (number of args is only 2), we'd be served well
to unbox both pairs, but not the triple. Indeed, that is what the algorithm
computes, and the following pictogram shows how the budget layers are computed.
Each layer is started with `n ~>`, where `n` is the budget at the start of the
layer. We write -n~> when we spend budget (and n is the remaining budget) and
+n~> when we earn budget. We separate unboxed args with ][ and indicate
inner budget threads becoming negative in braces {{}}, so that we see which
unboxing decision we do *not* commit to. Without further ado:

  4 ~> ][     (a,b) -3~>               ][     (c, ...) -2~>
       ][      | |                     ][      |   |
       ][      | +-------------+       ][      |   +-----------------+
       ][      |               |       ][      |                     |
       ][      v               v       ][      v                     v
  2 ~> ][ +3~> a  -2~> ][      b  -1~> ][ +2~> c  -1~> ][        (d, e, f) -0~>
       ][      |       ][      |       ][      |       ][ {{      |  |  |                          }}
       ][      |       ][      |       ][      |       ][ {{      |  |  +----------------+         }}
       ][      v       ][      v       ][      v       ][ {{      v  +------v            v         }}
  0 ~> ][ +1~> I# -0~> ][ +1~> I# -0~> ][ +1~> I# -0~> ][ {{ +1~> d -0~> ][ e -(-1)~> ][ f -(-2)~> }}

Unboxing increments the budget we have on the next layer (because we don't need
to retain the boxed arg), but in turn the inner layer must afford to retain all
non-absent fields, each decrementing the budget. Note how the budget becomes
negative when trying to unbox the triple and the unboxing decision is "rolled
back". This is done by the 'positiveTopBudget' guard.

There's a bit of complication as a result of handling unboxed tuples correctly;
specifically, handling nested unboxed tuples. Consider (#21737)

  unboxed :: (Int, Int) -> (# Int, (# Int, Int, Int #) #) -> Int
  unboxed (a,b) (# c, (# d, e, f #) #) = a + b + c + d + e + f

Recall that unboxed tuples will be flattened to individual arguments during
unarisation. Here, `unboxed` will have 5 arguments at runtime because of the
nested unboxed tuple, which will be flattened to 4 args. So it's best to leave
`(a,b)` boxed (because we already are above our arg threshold), but unbox `c`
through `f` because that doesn't increase the number of args post unarisation.

Note that the challenge is that syntactically, `(# d, e, f #)` occurs in a
deeper layer than `(a, b)`. Treating unboxed tuples as a regular data type, we'd
make the same unboxing decisions as for `boxed` above; although our starting
budget is 5 (Here, the number of args is greater than -fmax-worker-args), it's
not enough to unbox the triple (we'd finish with budget -1). So we'd unbox `a`
through `c`, but not `d` through `f`, which is silly, because then we'd end up
having 6 arguments at runtime, of which `d` through `f` weren't unboxed.

Hence we pretend that the fields of unboxed tuples appear in the same budget
layer as the tuple itself. For example at the top-level, `(# x,y #)` is to be
treated just like two arguments `x` and `y`.
Of course, for that to work, our budget calculations must initialise
'max_wkr_args' to 5, based on the 'unariseArity' of each Core arg: That would be
1 for the pair and 4 for the unboxed pair. Then when we decide whether to unbox
the unboxed pair, we *directly* recurse into the fields, spending our budget
on retaining `c` and (after recursing once more) `d` through `f` as arguments,
depleting our budget completely in the first layer. Pictorially:

  5 ~> ][         (a,b) -4~>             ][         (# c, ... #)
       ][ {{      | |                 }} ][      c  -3~> ][ (# d, e, f #)
       ][ {{      | +-------+         }} ][      |       ][      d  -2~> ][      e  -1~> ][      f  -0~>
       ][ {{      |         |         }} ][      |       ][      |       ][      |       ][      |
       ][ {{      v         v         }} ][      v       ][      v       ][      v       ][      v
  0 ~> ][ {{ +1~> a -0~> ][ b -(-1)~> }} ][ +1~> I# -0~> ][ +1~> I# -0~> ][ +1~> I# -0~> ][ +1~> I# -0~>

As you can see, we have no budget left to justify unboxing `(a,b)` on the second
layer, which is good, because it would increase the number of args. Also note
that we can still unbox `c` through `f` in this layer, because doing so has a
net zero effect on budget.

Note [The OPAQUE pragma and avoiding the reboxing of arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In https://gitlab.haskell.org/ghc/ghc/-/issues/13143 it was identified that when
a function 'f' with a NOINLINE pragma is W/W transformed, then the worker for
'f' should get the NOINLINE annotation, while the wrapper /should/ be inlined.

That's because if the wrapper for 'f' had stayed NOINLINE, then any worker of a
W/W-transformed /caller of/ 'f' would immediately rebox any unboxed arguments
that is applied to the wrapper of 'f'. When the wrapper is inlined, that kind of
reboxing does not happen.

But now we have functions with OPAQUE pragmas, which by definition (See Note
[OPAQUE pragma]) do not get W/W-transformed. So in order to avoid reboxing
workers of any W/W-transformed /callers of/ 'f' we need to strip all boxity
information from 'f' in the demand analysis. This will inform the
W/W-transformation code that boxed arguments of 'f' must definitely be passed
along in boxed form and as such dissuade the creation of reboxing workers.
-}

-- | How many registers does this type take after unarisation?
unariseArity :: Type -> Arity
unariseArity :: Type -> Arity
unariseArity Type
ty = [PrimRep] -> Arity
forall a. [a] -> Arity
forall (t :: * -> *) a. Foldable t => t a -> Arity
length (HasDebugCallStack => Type -> [PrimRep]
Type -> [PrimRep]
typePrimRep Type
ty)

data Budgets = MkB !Arity Budgets   -- An infinite list of arity budgets

earnTopBudget :: Budgets -> Budgets
earnTopBudget :: Budgets -> Budgets
earnTopBudget (MkB Arity
n Budgets
bg) = Arity -> Budgets -> Budgets
MkB (Arity
nArity -> Arity -> Arity
forall a. Num a => a -> a -> a
+Arity
1) Budgets
bg

spendTopBudget :: Arity -> Budgets -> Budgets
spendTopBudget :: Arity -> Budgets -> Budgets
spendTopBudget Arity
m (MkB Arity
n Budgets
bg) = Arity -> Budgets -> Budgets
MkB (Arity
nArity -> Arity -> Arity
forall a. Num a => a -> a -> a
-Arity
m) Budgets
bg

positiveTopBudget :: Budgets -> Bool
positiveTopBudget :: Budgets -> Bool
positiveTopBudget (MkB Arity
n Budgets
_) = Arity
n Arity -> Arity -> Bool
forall a. Ord a => a -> a -> Bool
>= Arity
0

finaliseArgBoxities :: AnalEnv -> Id -> Arity
                    -> [Demand] -> Divergence
                    -> CoreExpr -> ([Demand], CoreExpr)
finaliseArgBoxities :: AnalEnv
-> Id
-> Arity
-> [Demand]
-> Divergence
-> CoreExpr
-> ([Demand], CoreExpr)
finaliseArgBoxities AnalEnv
env Id
fn Arity
threshold_arity [Demand]
rhs_dmds Divergence
div CoreExpr
rhs

  -- Check for an OPAQUE function: see Note [OPAQUE pragma]
  -- In that case, trim off all boxity info from argument demands
  -- and demand info on lambda binders
  -- See Note [The OPAQUE pragma and avoiding the reboxing of arguments]
  | InlinePragma -> Bool
isOpaquePragma (Id -> InlinePragma
idInlinePragma Id
fn)
  , let trimmed_rhs_dmds :: [Demand]
trimmed_rhs_dmds = (Demand -> Demand) -> [Demand] -> [Demand]
forall a b. (a -> b) -> [a] -> [b]
map Demand -> Demand
trimBoxity [Demand]
rhs_dmds
  = ([Demand]
trimmed_rhs_dmds, [Demand] -> CoreExpr -> CoreExpr
set_lam_dmds [Demand]
trimmed_rhs_dmds CoreExpr
rhs)

  -- Check that we have enough visible binders to match the
  -- threshold arity; if not, we won't do worker/wrapper
  -- This happens if we have simply  {f = g} or a PAP {f = h 13}
  -- we simply want to give f the same demand signature as g
  -- How can such bindings arise?  Perhaps from {-# NOLINE[2] f #-},
  -- or if the call to `f` is currently not-applied (map f xs).
  -- It's a bit of a corner case.  Anyway for now we pass on the
  -- unadulterated demands from the RHS, without any boxity trimming.
  | Arity
threshold_arity Arity -> Arity -> Bool
forall a. Ord a => a -> a -> Bool
> (Id -> Bool) -> [Id] -> Arity
forall a. (a -> Bool) -> [a] -> Arity
count Id -> Bool
isId [Id]
bndrs
  = ([Demand]
rhs_dmds, CoreExpr
rhs)

  -- The normal case
  | Bool
otherwise -- NB: threshold_arity might be less than
              -- manifest arity for join points
  = -- pprTrace "finaliseArgBoxities" (
    --   vcat [text "function:" <+> ppr fn
    --        , text "dmds before:" <+> ppr (map idDemandInfo (filter isId bndrs))
    --        , text "dmds after: " <+>  ppr arg_dmds' ]) $
    ([Demand]
arg_dmds', [Demand] -> CoreExpr -> CoreExpr
set_lam_dmds [Demand]
arg_dmds' CoreExpr
rhs)
    -- set_lam_dmds: we must attach the final boxities to the lambda-binders
    -- of the function, both because that's kosher, and because CPR analysis
    -- uses the info on the binders directly.
  where
    opts :: DmdAnalOpts
opts            = AnalEnv -> DmdAnalOpts
ae_opts AnalEnv
env
    ([Id]
bndrs, CoreExpr
_body)  = CoreExpr -> ([Id], CoreExpr)
forall b. Expr b -> ([b], Expr b)
collectBinders CoreExpr
rhs
    unarise_arity :: Arity
unarise_arity   = [Arity] -> Arity
forall a. Num a => [a] -> a
forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a
sum [ Type -> Arity
unariseArity (Id -> Type
idType Id
b) | Id
b <- [Id]
bndrs, Id -> Bool
isId Id
b ]
    max_wkr_args :: Arity
max_wkr_args    = DmdAnalOpts -> Arity
dmd_max_worker_args DmdAnalOpts
opts Arity -> Arity -> Arity
forall a. Ord a => a -> a -> a
`max` Arity
unarise_arity
                      -- This is the budget initialisation step of
                      -- Note [Worker argument budget]

    -- This is the key line, which uses almost-circular programming
    -- The remaining budget from one layer becomes the initial
    -- budget for the next layer down.  See Note [Worker argument budget]
    (Budgets
remaining_budget, [Demand]
arg_dmds') = Budgets -> [(Type, StrictnessMark, Demand)] -> (Budgets, [Demand])
go_args (Arity -> Budgets -> Budgets
MkB Arity
max_wkr_args Budgets
remaining_budget) [(Type, StrictnessMark, Demand)]
arg_triples

    arg_triples :: [(Type, StrictnessMark, Demand)]
    arg_triples :: [(Type, StrictnessMark, Demand)]
arg_triples = Arity
-> [(Type, StrictnessMark, Demand)]
-> [(Type, StrictnessMark, Demand)]
forall a. Arity -> [a] -> [a]
take Arity
threshold_arity ([(Type, StrictnessMark, Demand)]
 -> [(Type, StrictnessMark, Demand)])
-> [(Type, StrictnessMark, Demand)]
-> [(Type, StrictnessMark, Demand)]
forall a b. (a -> b) -> a -> b
$
                  [ (Id -> Type
idType Id
bndr, StrictnessMark
NotMarkedStrict, Id -> Demand
get_dmd Id
bndr)
                  | Id
bndr <- [Id]
bndrs, Id -> Bool
isRuntimeVar Id
bndr ]

    get_dmd :: Id -> Demand
    get_dmd :: Id -> Demand
get_dmd Id
bndr
      | Bool
is_bot_fn = Demand -> Demand
unboxDeeplyDmd Demand
dmd -- See Note [Boxity for bottoming functions],
      | Bool
otherwise = Demand
dmd                --     case (B)
      where
        dmd :: Demand
dmd = Id -> Demand
idDemandInfo Id
bndr

    -- is_bot_fn:  see Note [Boxity for bottoming functions]
    is_bot_fn :: Bool
is_bot_fn = Divergence
div Divergence -> Divergence -> Bool
forall a. Eq a => a -> a -> Bool
== Divergence
botDiv

    go_args :: Budgets -> [(Type,StrictnessMark,Demand)] -> (Budgets, [Demand])
    go_args :: Budgets -> [(Type, StrictnessMark, Demand)] -> (Budgets, [Demand])
go_args Budgets
bg [(Type, StrictnessMark, Demand)]
triples = (Budgets -> (Type, StrictnessMark, Demand) -> (Budgets, Demand))
-> Budgets
-> [(Type, StrictnessMark, Demand)]
-> (Budgets, [Demand])
forall (t :: * -> *) s a b.
Traversable t =>
(s -> a -> (s, b)) -> s -> t a -> (s, t b)
mapAccumL Budgets -> (Type, StrictnessMark, Demand) -> (Budgets, Demand)
go_arg Budgets
bg [(Type, StrictnessMark, Demand)]
triples

    go_arg :: Budgets -> (Type,StrictnessMark,Demand) -> (Budgets, Demand)
    go_arg :: Budgets -> (Type, StrictnessMark, Demand) -> (Budgets, Demand)
go_arg bg :: Budgets
bg@(MkB Arity
bg_top Budgets
bg_inner) (Type
ty, StrictnessMark
str_mark, dmd :: Demand
dmd@(Card
n :* SubDemand
_))
      = case AnalEnv
-> Type
-> StrictnessMark
-> Demand
-> UnboxingDecision [(Type, StrictnessMark, Demand)]
wantToUnboxArg AnalEnv
env Type
ty StrictnessMark
str_mark Demand
dmd of
          UnboxingDecision [(Type, StrictnessMark, Demand)]
DropAbsent -> (Budgets
bg, Demand
dmd)

          UnboxingDecision [(Type, StrictnessMark, Demand)]
DontUnbox | Bool
is_bot_fn, Type -> Bool
isTyVarTy Type
ty -> (Budgets
retain_budget, Demand
dmd)
                    | Bool
otherwise               -> (Budgets
retain_budget, Demand -> Demand
trimBoxity Demand
dmd)
            -- If bot: Keep deep boxity even though WW won't unbox
            -- See Note [Boxity for bottoming functions] case (A)
            -- trimBoxity: see Note [No lazy, Unboxed demands in demand signature]
            where
              retain_budget :: Budgets
retain_budget = Arity -> Budgets -> Budgets
spendTopBudget (Type -> Arity
unariseArity Type
ty) Budgets
bg
                -- spendTopBudget: spend from our budget the cost of the
                -- retaining the arg
                -- The unboxed case does happen here, for example
                --   app g x = g x :: (# Int, Int #)
                -- here, `x` is used `L`azy and thus Boxed

          DoUnbox [(Type, StrictnessMark, Demand)]
triples
            | Type -> Bool
isUnboxedTupleType Type
ty
            , (Budgets
bg', [Demand]
dmds') <- Budgets -> [(Type, StrictnessMark, Demand)] -> (Budgets, [Demand])
go_args Budgets
bg [(Type, StrictnessMark, Demand)]
triples
            -> (Budgets
bg', Card
n HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* (Boxity -> [Demand] -> SubDemand
mkProd Boxity
Unboxed ([Demand] -> SubDemand) -> [Demand] -> SubDemand
forall a b. (a -> b) -> a -> b
$! [Demand]
dmds'))
                     -- See Note [Worker argument budget]
                     -- unboxed tuples are always unboxed, deeply
                     -- NB: Recurse with bg, *not* bg_inner! The unboxed fields
                     -- are at the same budget layer.

            | Type -> Bool
isUnboxedSumType Type
ty
            -> String -> SDoc -> (Budgets, Demand)
forall a. HasCallStack => String -> SDoc -> a
pprPanic String
"Unboxing through unboxed sum" (Id -> SDoc
forall a. Outputable a => a -> SDoc
ppr Id
fn SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<+> Type -> SDoc
forall a. Outputable a => a -> SDoc
ppr Type
ty)
                     -- We currently don't return DoUnbox for unboxed sums.
                     -- But hopefully we will at some point. When that happens,
                     -- it would still be impossible to predict the effect
                     -- of dropping absent fields and unboxing others on the
                     -- unariseArity of the sum without losing sanity.
                     -- We could overwrite bg_top with the one from
                     -- retain_budget while still unboxing inside the alts as in
                     -- the tuple case for a conservative solution, though.

            | Bool
otherwise
            -> (Arity -> Budgets -> Budgets
spendTopBudget Arity
1 (Arity -> Budgets -> Budgets
MkB Arity
bg_top Budgets
final_bg_inner), Demand
final_dmd)
            where
              (Budgets
bg_inner', [Demand]
dmds') = Budgets -> [(Type, StrictnessMark, Demand)] -> (Budgets, [Demand])
go_args (Budgets -> Budgets
earnTopBudget Budgets
bg_inner) [(Type, StrictnessMark, Demand)]
triples
                     -- earnTopBudget: give back the cost of retaining the
                     -- arg we are insted unboxing.
              dmd' :: Demand
dmd' = Card
n HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* (Boxity -> [Demand] -> SubDemand
mkProd Boxity
Unboxed ([Demand] -> SubDemand) -> [Demand] -> SubDemand
forall a b. (a -> b) -> a -> b
$! [Demand]
dmds')
              ~(Budgets
final_bg_inner, Demand
final_dmd) -- "~": This match *must* be lazy!
                 | Budgets -> Bool
positiveTopBudget Budgets
bg_inner' = (Budgets
bg_inner', Demand
dmd')
                 | Bool
otherwise                   = (Budgets
bg_inner,  Demand -> Demand
trimBoxity Demand
dmd)

    set_lam_dmds :: [Demand] -> CoreExpr -> CoreExpr
    -- Attach the demands to the outer lambdas of this expression
    set_lam_dmds :: [Demand] -> CoreExpr -> CoreExpr
set_lam_dmds (Demand
dmd:[Demand]
dmds) (Lam Id
v CoreExpr
e)
      | Id -> Bool
isTyVar Id
v = Id -> CoreExpr -> CoreExpr
forall b. b -> Expr b -> Expr b
Lam Id
v ([Demand] -> CoreExpr -> CoreExpr
set_lam_dmds (Demand
dmdDemand -> [Demand] -> [Demand]
forall a. a -> [a] -> [a]
:[Demand]
dmds) CoreExpr
e)
      | Bool
otherwise = Id -> CoreExpr -> CoreExpr
forall b. b -> Expr b -> Expr b
Lam (Id
v Id -> Demand -> Id
`setIdDemandInfo` Demand
dmd) ([Demand] -> CoreExpr -> CoreExpr
set_lam_dmds [Demand]
dmds CoreExpr
e)
    set_lam_dmds [Demand]
dmds (Cast CoreExpr
e Coercion
co) = CoreExpr -> Coercion -> CoreExpr
forall b. Expr b -> Coercion -> Expr b
Cast ([Demand] -> CoreExpr -> CoreExpr
set_lam_dmds [Demand]
dmds CoreExpr
e) Coercion
co
       -- This case happens for an OPAQUE function, which may look like
       --     f = (\x y. blah) |> co
       -- We give it strictness but no boxity (#22502)
    set_lam_dmds [Demand]
_ CoreExpr
e = CoreExpr
e
       -- In the OPAQUE case, the list of demands at this point might be
       -- non-empty, e.g., when looking at a PAP. Hence don't panic (#22997).

finaliseLetBoxity
  :: AnalEnv
  -> Type                   -- ^ Type of the let-bound Id
  -> Demand                 -- ^ How the Id is used
  -> Demand
-- See Note [Finalising boxity for let-bound Ids]
-- This function is like finaliseArgBoxities, but much simpler because
-- it has no "budget".  It simply unboxes strict demands, and stops
-- when it reaches a lazy one.
finaliseLetBoxity :: AnalEnv -> Type -> Demand -> Demand
finaliseLetBoxity AnalEnv
env Type
ty Demand
dmd
  = (Type, StrictnessMark, Demand) -> Demand
go (Type
ty, StrictnessMark
NotMarkedStrict, Demand
dmd)
  where
    go :: (Type,StrictnessMark,Demand) -> Demand
    go :: (Type, StrictnessMark, Demand) -> Demand
go (Type
ty, StrictnessMark
str, dmd :: Demand
dmd@(Card
n :* SubDemand
_)) =
      case AnalEnv
-> Type
-> StrictnessMark
-> Demand
-> UnboxingDecision [(Type, StrictnessMark, Demand)]
wantToUnboxArg AnalEnv
env Type
ty StrictnessMark
str Demand
dmd of
        UnboxingDecision [(Type, StrictnessMark, Demand)]
DropAbsent      -> Demand
dmd
        UnboxingDecision [(Type, StrictnessMark, Demand)]
DontUnbox       -> Demand -> Demand
trimBoxity Demand
dmd
        DoUnbox [(Type, StrictnessMark, Demand)]
triples -> Card
n HasDebugCallStack => Card -> SubDemand -> Demand
Card -> SubDemand -> Demand
:* (Boxity -> [Demand] -> SubDemand
mkProd Boxity
Unboxed ([Demand] -> SubDemand) -> [Demand] -> SubDemand
forall a b. (a -> b) -> a -> b
$! ((Type, StrictnessMark, Demand) -> Demand)
-> [(Type, StrictnessMark, Demand)] -> [Demand]
forall a b. (a -> b) -> [a] -> [b]
map (Type, StrictnessMark, Demand) -> Demand
go [(Type, StrictnessMark, Demand)]
triples)

wantToUnboxArg :: AnalEnv -> Type -> StrictnessMark -> Demand
               -> UnboxingDecision [(Type, StrictnessMark, Demand)]
wantToUnboxArg :: AnalEnv
-> Type
-> StrictnessMark
-> Demand
-> UnboxingDecision [(Type, StrictnessMark, Demand)]
wantToUnboxArg AnalEnv
env Type
ty StrictnessMark
str_mark dmd :: Demand
dmd@(Card
n :* SubDemand
_)
  = case FamInstEnvs
-> Type -> Demand -> UnboxingDecision (DataConPatContext Demand)
canUnboxArg (AnalEnv -> FamInstEnvs
ae_fam_envs AnalEnv
env) Type
ty Demand
dmd of
      UnboxingDecision (DataConPatContext Demand)
DropAbsent -> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info. UnboxingDecision unboxing_info
DropAbsent
      UnboxingDecision (DataConPatContext Demand)
DontUnbox  -> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info. UnboxingDecision unboxing_info
DontUnbox

      DoUnbox (DataConPatContext{ dcpc_dc :: forall s. DataConPatContext s -> DataCon
dcpc_dc      = DataCon
dc
                                , dcpc_tc_args :: forall s. DataConPatContext s -> [Type]
dcpc_tc_args = [Type]
tc_args
                                , dcpc_args :: forall s. DataConPatContext s -> [s]
dcpc_args    = [Demand]
dmds })
       -- OK, so we /can/ unbox it; but do we /want/ to?
       | Bool -> Bool
not (Card -> Bool
isStrict Card
n Bool -> Bool -> Bool
|| StrictnessMark -> Bool
isMarkedStrict StrictnessMark
str_mark)   -- Don't unbox a lazy field
         -- isMarkedStrict: see Note [Unboxing evaluated arguments] in DmdAnal
       -> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info. UnboxingDecision unboxing_info
DontUnbox

       | Type -> Bool
doNotUnbox Type
ty
       -> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info. UnboxingDecision unboxing_info
DontUnbox  -- See Note [Do not unbox class dictionaries]
                     -- NB: 'ty' has not been normalised, so this will (rightly)
                     --     catch newtype dictionaries too.
                     -- NB: even for bottoming functions, don't unbox dictionaries

       | IsRecDataConResult
DefinitelyRecursive <- AnalEnv -> DataCon -> IsRecDataConResult
ae_rec_dc AnalEnv
env DataCon
dc
         -- See Note [Which types are unboxed?]
         -- and Note [Demand analysis for recursive data constructors]
       -> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info. UnboxingDecision unboxing_info
DontUnbox

       | Bool
otherwise  -- Bad cases dealt with: we want to unbox!
       -> [(Type, StrictnessMark, Demand)]
-> UnboxingDecision [(Type, StrictnessMark, Demand)]
forall unboxing_info.
unboxing_info -> UnboxingDecision unboxing_info
DoUnbox ([Type]
-> [StrictnessMark] -> [Demand] -> [(Type, StrictnessMark, Demand)]
forall a b c. [a] -> [b] -> [c] -> [(a, b, c)]
zip3 (DataCon -> [Type] -> [Type]
dubiousDataConInstArgTys DataCon
dc [Type]
tc_args)
                        (DataCon -> [StrictnessMark]
dataConRepStrictness DataCon
dc)
                        [Demand]
dmds)


doNotUnbox :: Type -> Bool
-- Do not unbox class dictionaries, except equality classes and tuples
-- Note [Do not unbox class dictionaries]
doNotUnbox :: Type -> Bool
doNotUnbox Type
arg_ty
  = case Type -> Maybe TyCon
tyConAppTyCon_maybe Type
arg_ty of
      Just TyCon
tc | Just Class
cls <- TyCon -> Maybe Class
tyConClass_maybe TyCon
tc
              -> Bool -> Bool
not (Class -> Bool
isEqualityClass Class
cls Bool -> Bool -> Bool
|| Class -> Bool
isCTupleClass Class
cls)
       -- See (DNB2) and (DNB1) in Note [Do not unbox class dictionaries]

      Maybe TyCon
_ -> Bool
False

{- Note [Do not unbox class dictionaries]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We never unbox class dictionaries in worker/wrapper.

1. INLINABLE functions
   If we have
      f :: Ord a => [a] -> Int -> a
      {-# INLINABLE f #-}
   and we worker/wrapper f, we'll get a worker with an INLINABLE pragma
   (see Note [Worker/wrapper for INLINABLE functions] in GHC.Core.Opt.WorkWrap),
   which can still be specialised by the type-class specialiser, something like
      fw :: Ord a => [a] -> Int# -> a

   BUT if f is strict in the Ord dictionary, we might unpack it, to get
      fw :: (a->a->Bool) -> [a] -> Int# -> a
   and the type-class specialiser can't specialise that. An example is #6056.

   Historical note: #14955 describes how I got this fix wrong the first time.
   I got aware of the issue in T5075 by the change in boxity of loop between
   demand analysis runs.

2. -fspecialise-aggressively.  As #21286 shows, the same phenomenon can occur
   occur without INLINABLE, when we use -fexpose-all-unfoldings and
   -fspecialise-aggressively to do vigorous cross-module specialisation.

3. #18421 found that unboxing a dictionary can also make the worker less likely
   to inline; the inlining heuristics seem to prefer to inline a function
   applied to a dictionary over a function applied to a bunch of functions.

TL;DR we /never/ unbox class dictionaries. Unboxing the dictionary, and passing
a raft of higher-order functions isn't a huge win anyway -- you really want to
specialise the function.

Wrinkle (DNB1): we /do/ want to unbox tuple dictionaries (#23398)
     f :: (% Eq a, Show a %) => blah
  with -fdicts-strict it is great to unbox to
     $wf :: Eq a => Show a => blah
  (where I have written out the currying explicitly).  Now we can specialise
  $wf on the Eq or Show dictionary.  Nothing is lost.

  And something is gained.  It is possible that `f` will look like this:
     f = /\a. \d:(% Eq a, Show a %). ... f @a (% sel1 d, sel2 d %)...
  where there is a recurive call to `f`, or to another function that takes the
  same tuple dictionary, but where the tuple is built from the components of
  `d`.  The Simplier does not fix this.  But if we unpacked the dictionary
  we'd get
     $wf = /\a. \(d1:Eq a) (d2:Show a). let d = (% d1, d2 %)
             in ...f @a (% sel1 d, sel2 d %)
  and all the tuple building and taking apart will disappear.

Wrinkle (DNB2): we /do/ wnat to unbox equality dictionaries,
  for (~), (~~), and Coercible (#23398).  Their payload is a single unboxed
  coercion.  We never want to specialise on `(t1 ~ t2)`.  All that would do is
  to make a copy of the function's RHS with a particular coercion.  Unlike
  normal class methods, that does not unlock any new optimisation
  opportunities in the specialised RHS.
-}

{- *********************************************************************
*                                                                      *
                      Fixpoints
*                                                                      *
********************************************************************* -}

-- Recursive bindings
dmdFix :: TopLevelFlag
       -> AnalEnv                            -- Does not include bindings for this binding
       -> SubDemand
       -> [(Id,CoreExpr)]
       -> (AnalEnv, WeakDmds, [(Id,CoreExpr)]) -- Binders annotated with strictness info
dmdFix :: TopLevelFlag
-> AnalEnv
-> SubDemand
-> [(Id, CoreExpr)]
-> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
dmdFix TopLevelFlag
top_lvl AnalEnv
env SubDemand
let_dmd [(Id, CoreExpr)]
orig_pairs
  = Arity -> [(Id, CoreExpr)] -> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
loop Arity
1 [(Id, CoreExpr)]
initial_pairs
  where
    opts :: DmdAnalOpts
opts = AnalEnv -> DmdAnalOpts
ae_opts AnalEnv
env
    -- See Note [Initialising strictness]
    initial_pairs :: [(Id, CoreExpr)]
initial_pairs | AnalEnv -> Bool
ae_virgin AnalEnv
env = [(DmdAnalOpts -> Id -> DmdSig -> Id
setIdDmdAndBoxSig DmdAnalOpts
opts Id
id DmdSig
botSig, CoreExpr
rhs) | (Id
id, CoreExpr
rhs) <- [(Id, CoreExpr)]
orig_pairs ]
                  | Bool
otherwise     = [(Id, CoreExpr)]
orig_pairs

    -- If fixed-point iteration does not yield a result we use this instead
    -- See Note [Safe abortion in the fixed-point iteration]
    abort :: (AnalEnv, WeakDmds, [(Id,CoreExpr)])
    abort :: (AnalEnv, WeakDmds, [(Id, CoreExpr)])
abort = (AnalEnv
env, WeakDmds
weak_fv', [(Id, CoreExpr)]
zapped_pairs)
      where (WeakDmds
weak_fv, [(Id, CoreExpr)]
pairs') = Bool -> [(Id, CoreExpr)] -> (WeakDmds, [(Id, CoreExpr)])
step Bool
True ([(Id, CoreExpr)] -> [(Id, CoreExpr)]
zapIdDmdSig [(Id, CoreExpr)]
orig_pairs)
            -- Note [Lazy and unleashable free variables]
            weak_fvs :: WeakDmds
weak_fvs = [WeakDmds] -> WeakDmds
forall a. [VarEnv a] -> VarEnv a
plusVarEnvList ([WeakDmds] -> WeakDmds) -> [WeakDmds] -> WeakDmds
forall a b. (a -> b) -> a -> b
$ ((Id, CoreExpr) -> WeakDmds) -> [(Id, CoreExpr)] -> [WeakDmds]
forall a b. (a -> b) -> [a] -> [b]
map (DmdEnv -> WeakDmds
de_fvs (DmdEnv -> WeakDmds)
-> ((Id, CoreExpr) -> DmdEnv) -> (Id, CoreExpr) -> WeakDmds
forall b c a. (b -> c) -> (a -> b) -> a -> c
. DmdSig -> DmdEnv
dmdSigDmdEnv (DmdSig -> DmdEnv)
-> ((Id, CoreExpr) -> DmdSig) -> (Id, CoreExpr) -> DmdEnv
forall b c a. (b -> c) -> (a -> b) -> a -> c
. Id -> DmdSig
idDmdSig (Id -> DmdSig)
-> ((Id, CoreExpr) -> Id) -> (Id, CoreExpr) -> DmdSig
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst) [(Id, CoreExpr)]
pairs'
            weak_fv' :: WeakDmds
weak_fv'     = (Demand -> Demand -> Demand) -> WeakDmds -> WeakDmds -> WeakDmds
forall a. (a -> a -> a) -> VarEnv a -> VarEnv a -> VarEnv a
plusVarEnv_C Demand -> Demand -> Demand
plusDmd WeakDmds
weak_fv (WeakDmds -> WeakDmds) -> WeakDmds -> WeakDmds
forall a b. (a -> b) -> a -> b
$ (Demand -> Demand) -> WeakDmds -> WeakDmds
forall a b. (a -> b) -> VarEnv a -> VarEnv b
mapVarEnv (Demand -> Demand -> Demand
forall a b. a -> b -> a
const Demand
topDmd) WeakDmds
weak_fvs
            zapped_pairs :: [(Id, CoreExpr)]
zapped_pairs = [(Id, CoreExpr)] -> [(Id, CoreExpr)]
zapIdDmdSig [(Id, CoreExpr)]
pairs'

    -- The fixed-point varies the idDmdSig field of the binders, and terminates if that
    -- annotation does not change any more.
    loop :: Int -> [(Id,CoreExpr)] -> (AnalEnv, WeakDmds, [(Id,CoreExpr)])
    loop :: Arity -> [(Id, CoreExpr)] -> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
loop Arity
n [(Id, CoreExpr)]
pairs = -- pprTrace "dmdFix" (ppr n <+> vcat [ ppr id <+> ppr (idDmdSig id)
                   --                                   | (id,_) <- pairs]) $
                   Arity -> [(Id, CoreExpr)] -> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
loop' Arity
n [(Id, CoreExpr)]
pairs

    loop' :: Arity -> [(Id, CoreExpr)] -> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
loop' Arity
n [(Id, CoreExpr)]
pairs
      | Bool
found_fixpoint = (AnalEnv
final_anal_env, WeakDmds
weak_fv, [(Id, CoreExpr)]
pairs')
      | Arity
n Arity -> Arity -> Bool
forall a. Eq a => a -> a -> Bool
== Arity
10        = (AnalEnv, WeakDmds, [(Id, CoreExpr)])
abort
      | Bool
otherwise      = Arity -> [(Id, CoreExpr)] -> (AnalEnv, WeakDmds, [(Id, CoreExpr)])
loop (Arity
nArity -> Arity -> Arity
forall a. Num a => a -> a -> a
+Arity
1) [(Id, CoreExpr)]
pairs'
      where
        found_fixpoint :: Bool
found_fixpoint    = ((Id, CoreExpr) -> DmdSig) -> [(Id, CoreExpr)] -> [DmdSig]
forall a b. (a -> b) -> [a] -> [b]
map (Id -> DmdSig
idDmdSig (Id -> DmdSig)
-> ((Id, CoreExpr) -> Id) -> (Id, CoreExpr) -> DmdSig
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst) [(Id, CoreExpr)]
pairs' [DmdSig] -> [DmdSig] -> Bool
forall a. Eq a => a -> a -> Bool
== ((Id, CoreExpr) -> DmdSig) -> [(Id, CoreExpr)] -> [DmdSig]
forall a b. (a -> b) -> [a] -> [b]
map (Id -> DmdSig
idDmdSig (Id -> DmdSig)
-> ((Id, CoreExpr) -> Id) -> (Id, CoreExpr) -> DmdSig
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst) [(Id, CoreExpr)]
pairs
        first_round :: Bool
first_round       = Arity
n Arity -> Arity -> Bool
forall a. Eq a => a -> a -> Bool
== Arity
1
        (WeakDmds
weak_fv, [(Id, CoreExpr)]
pairs') = Bool -> [(Id, CoreExpr)] -> (WeakDmds, [(Id, CoreExpr)])
step Bool
first_round [(Id, CoreExpr)]
pairs
        final_anal_env :: AnalEnv
final_anal_env    = TopLevelFlag -> AnalEnv -> [Id] -> AnalEnv
extendAnalEnvs TopLevelFlag
top_lvl AnalEnv
env (((Id, CoreExpr) -> Id) -> [(Id, CoreExpr)] -> [Id]
forall a b. (a -> b) -> [a] -> [b]
map (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst [(Id, CoreExpr)]
pairs')

    step :: Bool -> [(Id, CoreExpr)] -> (WeakDmds, [(Id, CoreExpr)])
    step :: Bool -> [(Id, CoreExpr)] -> (WeakDmds, [(Id, CoreExpr)])
step Bool
first_round [(Id, CoreExpr)]
pairs = (WeakDmds
weak_fv, [(Id, CoreExpr)]
pairs')
      where
        -- In all but the first iteration, delete the virgin flag
        start_env :: AnalEnv
start_env | Bool
first_round = AnalEnv
env
                  | Bool
otherwise   = AnalEnv -> AnalEnv
nonVirgin AnalEnv
env

        start :: (AnalEnv, WeakDmds)
start = (TopLevelFlag -> AnalEnv -> [Id] -> AnalEnv
extendAnalEnvs TopLevelFlag
top_lvl AnalEnv
start_env (((Id, CoreExpr) -> Id) -> [(Id, CoreExpr)] -> [Id]
forall a b. (a -> b) -> [a] -> [b]
map (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst [(Id, CoreExpr)]
pairs), WeakDmds
forall a. VarEnv a
emptyVarEnv)

        !((AnalEnv
_,!WeakDmds
weak_fv), ![(Id, CoreExpr)]
pairs') = ((AnalEnv, WeakDmds)
 -> (Id, CoreExpr) -> ((AnalEnv, WeakDmds), (Id, CoreExpr)))
-> (AnalEnv, WeakDmds)
-> [(Id, CoreExpr)]
-> ((AnalEnv, WeakDmds), [(Id, CoreExpr)])
forall (t :: * -> *) s a b.
Traversable t =>
(s -> a -> (s, b)) -> s -> t a -> (s, t b)
mapAccumL (AnalEnv, WeakDmds)
-> (Id, CoreExpr) -> ((AnalEnv, WeakDmds), (Id, CoreExpr))
my_downRhs (AnalEnv, WeakDmds)
start [(Id, CoreExpr)]
pairs
                -- mapAccumL: Use the new signature to do the next pair
                -- The occurrence analyser has arranged them in a good order
                -- so this can significantly reduce the number of iterations needed

        my_downRhs :: (AnalEnv, WeakDmds)
-> (Id, CoreExpr) -> ((AnalEnv, WeakDmds), (Id, CoreExpr))
my_downRhs (AnalEnv
env, WeakDmds
weak_fv) (Id
id,CoreExpr
rhs)
          = -- pprTrace "my_downRhs" (ppr id $$ ppr (idDmdSig id) $$ ppr sig) $
            ((AnalEnv
env', WeakDmds
weak_fv'), (Id
id', CoreExpr
rhs'))
          where
            !(!AnalEnv
env', !WeakDmds
weak_fv1, !Id
id', !CoreExpr
rhs') = TopLevelFlag
-> RecFlag
-> AnalEnv
-> SubDemand
-> Id
-> CoreExpr
-> (AnalEnv, WeakDmds, Id, CoreExpr)
dmdAnalRhsSig TopLevelFlag
top_lvl RecFlag
Recursive AnalEnv
env SubDemand
let_dmd Id
id CoreExpr
rhs
            !weak_fv' :: WeakDmds
weak_fv'                    = (Demand -> Demand -> Demand) -> WeakDmds -> WeakDmds -> WeakDmds
forall a. (a -> a -> a) -> VarEnv a -> VarEnv a -> VarEnv a
plusVarEnv_C Demand -> Demand -> Demand
plusDmd WeakDmds
weak_fv WeakDmds
weak_fv1

    zapIdDmdSig :: [(Id, CoreExpr)] -> [(Id, CoreExpr)]
    zapIdDmdSig :: [(Id, CoreExpr)] -> [(Id, CoreExpr)]
zapIdDmdSig [(Id, CoreExpr)]
pairs = [(Id -> DmdSig -> Id
setIdDmdSig Id
id DmdSig
nopSig, CoreExpr
rhs) | (Id
id, CoreExpr
rhs) <- [(Id, CoreExpr)]
pairs ]

{- Note [Safe abortion in the fixed-point iteration]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Fixed-point iteration may fail to terminate. But we cannot simply give up and
return the environment and code unchanged! We still need to do one additional
round, for two reasons:

 * To get information on used free variables (both lazy and strict!)
   (see Note [Lazy and unleashable free variables])
 * To ensure that all expressions have been traversed at least once, and any left-over
   strictness annotations have been updated.

This final iteration does not add the variables to the strictness signature
environment, which effectively assigns them 'nopSig' (see "getStrictness")

Note [Trimming a demand to a type]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
There are two reasons we sometimes trim a demand to match a type.
  1. GADTs
  2. Recursive products and widening

More on both below.  But the bottom line is: we really don't want to
have a binder whose demand is more deeply-nested than its type
"allows". So in findBndrDmd we call trimToType and findTypeShape to
trim the demand on the binder to a form that matches the type

Now to the reasons. For (1) consider
  f :: a -> Bool
  f x = case ... of
          A g1 -> case (x |> g1) of (p,q) -> ...
          B    -> error "urk"

where A,B are the constructors of a GADT.  We'll get a 1P(L,L) demand
on x from the A branch, but that's a stupid demand for x itself, which
has type 'a'. Indeed we get ASSERTs going off (notably in
splitUseProdDmd, #8569).

For (2) consider
  data T = MkT Int T    -- A recursive product
  f :: Int -> T -> Int
  f 0 _         = 0
  f _ (MkT n t) = f n t

Here f is lazy in T, but its *usage* is infinite: P(L,P(L,P(L, ...))).
Notice that this happens because T is a product type, and is recursive.
If we are not careful, we'll fail to iterate to a fixpoint in dmdFix,
and bale out entirely, which is inefficient and over-conservative.

Worse, as we discovered in #18304, the size of the usages we compute
can grow /exponentially/, so even 10 iterations costs far too much.
Especially since we then discard the result.

To avoid this we use the same findTypeShape function as for (1), but
arrange that it trims the demand if it encounters the same type constructor
twice (or three times, etc).  We use our standard RecTcChecker mechanism
for this -- see GHC.Core.Opt.WorkWrap.Utils.findTypeShape.

This is usually call "widening".  We could do it just in dmdFix, but
since are doing this findTypeShape business /anyway/ because of (1),
and it has all the right information to hand, it's extremely
convenient to do it there.

-}

{- *********************************************************************
*                                                                      *
                 Strictness signatures and types
*                                                                      *
********************************************************************* -}

noArgsDmdType :: DmdEnv -> DmdType
noArgsDmdType :: DmdEnv -> DmdType
noArgsDmdType DmdEnv
dmd_env = DmdEnv -> [Demand] -> DmdType
DmdType DmdEnv
dmd_env []

coercionDmdEnv :: Coercion -> DmdEnv
coercionDmdEnv :: Coercion -> DmdEnv
coercionDmdEnv Coercion
co = [Coercion] -> DmdEnv
coercionsDmdEnv [Coercion
co]

coercionsDmdEnv :: [Coercion] -> DmdEnv
coercionsDmdEnv :: [Coercion] -> DmdEnv
coercionsDmdEnv [Coercion]
cos
  = WeakDmds -> DmdEnv
mkTermDmdEnv (WeakDmds -> DmdEnv) -> WeakDmds -> DmdEnv
forall a b. (a -> b) -> a -> b
$ (Id -> Demand) -> VarEnv Id -> WeakDmds
forall a b. (a -> b) -> VarEnv a -> VarEnv b
mapVarEnv (Demand -> Id -> Demand
forall a b. a -> b -> a
const Demand
topDmd) (VarEnv Id -> WeakDmds) -> VarEnv Id -> WeakDmds
forall a b. (a -> b) -> a -> b
$ VarSet -> VarEnv Id
forall a. UniqSet a -> UniqFM a a
getUniqSet (VarSet -> VarEnv Id) -> VarSet -> VarEnv Id
forall a b. (a -> b) -> a -> b
$ [Coercion] -> VarSet
coVarsOfCos [Coercion]
cos
  -- The VarSet from coVarsOfCos is really a VarEnv Var

addVarDmd :: DmdType -> Var -> Demand -> DmdType
addVarDmd :: DmdType -> Id -> Demand -> DmdType
addVarDmd (DmdType DmdEnv
fv [Demand]
ds) Id
var Demand
dmd
  = DmdEnv -> [Demand] -> DmdType
DmdType (DmdEnv -> Id -> Demand -> DmdEnv
addVarDmdEnv DmdEnv
fv Id
var Demand
dmd) [Demand]
ds

addWeakFVs :: DmdType -> WeakDmds -> DmdType
addWeakFVs :: DmdType -> WeakDmds -> DmdType
addWeakFVs DmdType
dmd_ty WeakDmds
weak_fvs
  = DmdType
dmd_ty DmdType -> DmdEnv -> DmdType
`plusDmdType` WeakDmds -> DmdEnv
mkTermDmdEnv WeakDmds
weak_fvs
        -- Using plusDmdType (rather than just plus'ing the envs)
        -- is vital.  Consider
        --      let f = \x -> (x,y)
        --      in  error (f 3)
        -- Here, y is treated as a lazy-fv of f, but we must `plusDmd` that L
        -- demand with the bottom coming up from 'error'
        --
        -- I got a loop in the fixpointer without this, due to an interaction
        -- with the weak_fv filtering in dmdAnalRhsSig.  Roughly, it was
        --      letrec f n x
        --          = letrec g y = x `fatbar`
        --                         letrec h z = z + ...g...
        --                         in h (f (n-1) x)
        --      in ...
        -- In the initial iteration for f, f=Bot
        -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
        -- is lazy.  Now consider the fixpoint iteration for g, esp the demands it
        -- places on its free variables.  Suppose it places none.  Then the
        --      x `fatbar` ...call to h...
        -- will give a x->V demand for x.  That turns into a L demand for x,
        -- which floats out of the defn for h.  Without the modifyEnv, that
        -- L demand doesn't get both'd with the Bot coming up from the inner
        -- call to f.  So we just get an L demand for x for g.

setBndrsDemandInfo :: HasCallStack => [Var] -> [Demand] -> [Var]
setBndrsDemandInfo :: HasCallStack => [Id] -> [Demand] -> [Id]
setBndrsDemandInfo (Id
b:[Id]
bs) [Demand]
ds
  | Id -> Bool
isTyVar Id
b = Id
b Id -> [Id] -> [Id]
forall a. a -> [a] -> [a]
: HasCallStack => [Id] -> [Demand] -> [Id]
[Id] -> [Demand] -> [Id]
setBndrsDemandInfo [Id]
bs [Demand]
ds
setBndrsDemandInfo (Id
b:[Id]
bs) (Demand
d:[Demand]
ds) =
    let !new_info :: Id
new_info = Id -> Demand -> Id
setIdDemandInfo Id
b Demand
d
        !vars :: [Id]
vars = HasCallStack => [Id] -> [Demand] -> [Id]
[Id] -> [Demand] -> [Id]
setBndrsDemandInfo [Id]
bs [Demand]
ds
    in Id
new_info Id -> [Id] -> [Id]
forall a. a -> [a] -> [a]
: [Id]
vars
setBndrsDemandInfo [] [Demand]
ds = Bool -> [Id] -> [Id]
forall a. HasCallStack => Bool -> a -> a
assert ([Demand] -> Bool
forall a. [a] -> Bool
forall (t :: * -> *) a. Foldable t => t a -> Bool
null [Demand]
ds) []
setBndrsDemandInfo [Id]
bs [Demand]
_  = String -> SDoc -> [Id]
forall a. HasCallStack => String -> SDoc -> a
pprPanic String
"setBndrsDemandInfo" ([Id] -> SDoc
forall a. Outputable a => a -> SDoc
ppr [Id]
bs)

annotateLamIdBndr :: AnalEnv
                  -> DmdType    -- Demand type of body
                  -> Id         -- Lambda binder
                  -> WithDmdType Id  -- Demand type of lambda
                                     -- and binder annotated with demand

annotateLamIdBndr :: AnalEnv -> DmdType -> Id -> WithDmdType Id
annotateLamIdBndr AnalEnv
env DmdType
dmd_ty Id
id
-- For lambdas we add the demand to the argument demands
-- Only called for Ids
  = Bool -> WithDmdType Id -> WithDmdType Id
forall a. HasCallStack => Bool -> a -> a
assert (Id -> Bool
isId Id
id) (WithDmdType Id -> WithDmdType Id)
-> WithDmdType Id -> WithDmdType Id
forall a b. (a -> b) -> a -> b
$
    -- pprTrace "annLamBndr" (vcat [ppr id, ppr dmd_ty, ppr final_ty]) $
    DmdType -> Id -> WithDmdType Id
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
main_ty Id
new_id
  where
    new_id :: Id
new_id  = Id -> Demand -> Id
setIdDemandInfo Id
id Demand
dmd
    main_ty :: DmdType
main_ty = Demand -> DmdType -> DmdType
addDemand Demand
dmd DmdType
dmd_ty'
    WithDmdType DmdType
dmd_ty' Demand
dmd = AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
dmd_ty Id
id

{- Note [NOINLINE and strictness]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
At one point we disabled strictness for NOINLINE functions, on the
grounds that they should be entirely opaque.  But that lost lots of
useful semantic strictness information, so now we analyse them like
any other function, and pin strictness information on them.

That in turn forces us to worker/wrapper them; see
Note [Worker/wrapper for NOINLINE functions] in GHC.Core.Opt.WorkWrap.


Note [Lazy and unleashable free variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We put the strict and once-used FVs in the DmdType of the Id, so
that at its call sites we unleash demands on its strict fvs.
An example is 'roll' in imaginary/wheel-sieve2
Something like this:
        roll x = letrec
                     go y = if ... then roll (x-1) else x+1
                 in
                 go ms
We want to see that roll is strict in x, which is because
go is called.   So we put the DmdEnv for x in go's DmdType.

Another example:

        f :: Int -> Int -> Int
        f x y = let t = x+1
            h z = if z==0 then t else
                  if z==1 then x+1 else
                  x + h (z-1)
        in h y

Calling h does indeed evaluate x, but we can only see
that if we unleash a demand on x at the call site for t.

Incidentally, here's a place where lambda-lifting h would
lose the cigar --- we couldn't see the joint strictness in t/x

        ON THE OTHER HAND

We don't want to put *all* the fv's from the RHS into the
DmdType. Because

 * it makes the strictness signatures larger, and hence slows down fixpointing

and

 * it is useless information at the call site anyways:
   For lazy, used-many times fv's we will never get any better result than
   that, no matter how good the actual demand on the function at the call site
   is (unless it is always absent, but then the whole binder is useless).

Therefore we exclude lazy multiple-used fv's from the environment in the
DmdType.

But now the signature lies! (Missing variables are assumed to be absent.) To
make up for this, the code that analyses the binding keeps the demand on those
variable separate (usually called "weak_fv") and adds it to the demand of the
whole binding later.

What if we decide _not_ to store a strictness signature for a binding at all, as
we do when aborting a fixed-point iteration? The we risk losing the information
that the strict variables are being used. In that case, we take all free variables
mentioned in the (unsound) strictness signature, conservatively approximate the
demand put on them (topDmd), and add that to the "weak_fv" returned by "dmdFix".


************************************************************************
*                                                                      *
\subsection{Strictness signatures}
*                                                                      *
************************************************************************
-}


data AnalEnv = AE
  { AnalEnv -> DmdAnalOpts
ae_opts      :: !DmdAnalOpts
  -- ^ Analysis options
  , AnalEnv -> SigEnv
ae_sigs      :: !SigEnv
  , AnalEnv -> Bool
ae_virgin    :: !Bool
  -- ^ True on first iteration only. See Note [Initialising strictness]
  , AnalEnv -> FamInstEnvs
ae_fam_envs  :: !FamInstEnvs
  , AnalEnv -> DataCon -> IsRecDataConResult
ae_rec_dc    :: DataCon -> IsRecDataConResult
  -- ^ Memoised result of 'GHC.Core.Opt.WorkWrap.Utils.isRecDataCon'
  }

        -- We use the se_env to tell us whether to
        -- record info about a variable in the DmdEnv
        -- We do so if it's a LocalId, but not top-level
        --
        -- The DmdEnv gives the demand on the free vars of the function
        -- when it is given enough args to satisfy the strictness signature

type SigEnv = VarEnv (DmdSig, TopLevelFlag)

instance Outputable AnalEnv where
  ppr :: AnalEnv -> SDoc
ppr AnalEnv
env = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"AE" SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<+> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc
braces ([SDoc] -> SDoc
forall doc. IsDoc doc => [doc] -> doc
vcat
         [ String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"ae_virgin =" SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<+> Bool -> SDoc
forall a. Outputable a => a -> SDoc
ppr (AnalEnv -> Bool
ae_virgin AnalEnv
env)
         , String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"ae_sigs =" SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<+> SigEnv -> SDoc
forall a. Outputable a => a -> SDoc
ppr (AnalEnv -> SigEnv
ae_sigs AnalEnv
env)
         ])

emptyAnalEnv :: DmdAnalOpts -> FamInstEnvs -> AnalEnv
emptyAnalEnv :: DmdAnalOpts -> FamInstEnvs -> AnalEnv
emptyAnalEnv DmdAnalOpts
opts FamInstEnvs
fam_envs
    = AE { ae_opts :: DmdAnalOpts
ae_opts         = DmdAnalOpts
opts
         , ae_sigs :: SigEnv
ae_sigs         = SigEnv
emptySigEnv
         , ae_virgin :: Bool
ae_virgin       = Bool
True
         , ae_fam_envs :: FamInstEnvs
ae_fam_envs     = FamInstEnvs
fam_envs
         , ae_rec_dc :: DataCon -> IsRecDataConResult
ae_rec_dc       = (DataCon -> IsRecDataConResult) -> DataCon -> IsRecDataConResult
forall k a. Uniquable k => (k -> a) -> k -> a
memoiseUniqueFun (FamInstEnvs -> IntWithInf -> DataCon -> IsRecDataConResult
isRecDataCon FamInstEnvs
fam_envs IntWithInf
3)
         }

-- | Unset the 'dmd_strict_dicts' flag if any of the given bindings is a DFun
-- binding. Part of the mechanism that detects
-- Note [Do not strictify a DFun's parameter dictionaries].
enterDFun :: CoreBind -> AnalEnv -> AnalEnv
enterDFun :: Bind Id -> AnalEnv -> AnalEnv
enterDFun Bind Id
bind AnalEnv
env
  | (Id -> Bool) -> [Id] -> Bool
forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool
any Id -> Bool
isDFunId (Bind Id -> [Id]
forall b. Bind b -> [b]
bindersOf Bind Id
bind)
  = AnalEnv
env { ae_opts = (ae_opts env) { dmd_strict_dicts = False } }
  | Bool
otherwise
  = AnalEnv
env

emptySigEnv :: SigEnv
emptySigEnv :: SigEnv
emptySigEnv = SigEnv
forall a. VarEnv a
emptyVarEnv

-- | Extend an environment with the strictness sigs attached to the Ids
extendAnalEnvs :: TopLevelFlag -> AnalEnv -> [Id] -> AnalEnv
extendAnalEnvs :: TopLevelFlag -> AnalEnv -> [Id] -> AnalEnv
extendAnalEnvs TopLevelFlag
top_lvl AnalEnv
env [Id]
vars
  = AnalEnv
env { ae_sigs = extendSigEnvs top_lvl (ae_sigs env) vars }

extendSigEnvs :: TopLevelFlag -> SigEnv -> [Id] -> SigEnv
extendSigEnvs :: TopLevelFlag -> SigEnv -> [Id] -> SigEnv
extendSigEnvs TopLevelFlag
top_lvl SigEnv
sigs [Id]
vars
  = SigEnv -> [(Id, (DmdSig, TopLevelFlag))] -> SigEnv
forall a. VarEnv a -> [(Id, a)] -> VarEnv a
extendVarEnvList SigEnv
sigs [ (Id
var, (Id -> DmdSig
idDmdSig Id
var, TopLevelFlag
top_lvl)) | Id
var <- [Id]
vars]

extendAnalEnv :: TopLevelFlag -> AnalEnv -> Id -> DmdSig -> AnalEnv
extendAnalEnv :: TopLevelFlag -> AnalEnv -> Id -> DmdSig -> AnalEnv
extendAnalEnv TopLevelFlag
top_lvl AnalEnv
env Id
var DmdSig
sig
  = AnalEnv
env { ae_sigs = extendSigEnv top_lvl (ae_sigs env) var sig }

extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> DmdSig -> SigEnv
extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> DmdSig -> SigEnv
extendSigEnv TopLevelFlag
top_lvl SigEnv
sigs Id
var DmdSig
sig = SigEnv -> Id -> (DmdSig, TopLevelFlag) -> SigEnv
forall a. VarEnv a -> Id -> a -> VarEnv a
extendVarEnv SigEnv
sigs Id
var (DmdSig
sig, TopLevelFlag
top_lvl)

lookupSigEnv :: AnalEnv -> Id -> Maybe (DmdSig, TopLevelFlag)
lookupSigEnv :: AnalEnv -> Id -> Maybe (DmdSig, TopLevelFlag)
lookupSigEnv AnalEnv
env Id
id = SigEnv -> Id -> Maybe (DmdSig, TopLevelFlag)
forall a. VarEnv a -> Id -> Maybe a
lookupVarEnv (AnalEnv -> SigEnv
ae_sigs AnalEnv
env) Id
id

addInScopeAnalEnv :: AnalEnv -> Var -> AnalEnv
addInScopeAnalEnv :: AnalEnv -> Id -> AnalEnv
addInScopeAnalEnv AnalEnv
env Id
id = AnalEnv
env { ae_sigs = delVarEnv (ae_sigs env) id }

addInScopeAnalEnvs :: AnalEnv -> [Var] -> AnalEnv
addInScopeAnalEnvs :: AnalEnv -> [Id] -> AnalEnv
addInScopeAnalEnvs AnalEnv
env [Id]
ids = AnalEnv
env { ae_sigs = delVarEnvList (ae_sigs env) ids }

nonVirgin :: AnalEnv -> AnalEnv
nonVirgin :: AnalEnv -> AnalEnv
nonVirgin AnalEnv
env = AnalEnv
env { ae_virgin = False }

findBndrsDmds :: AnalEnv -> DmdType -> [Var] -> WithDmdType [Demand]
-- Return the demands on the Ids in the [Var]
findBndrsDmds :: AnalEnv -> DmdType -> [Id] -> WithDmdType [Demand]
findBndrsDmds AnalEnv
env DmdType
dmd_ty [Id]
bndrs
  = DmdType -> [Id] -> WithDmdType [Demand]
go DmdType
dmd_ty [Id]
bndrs
  where
    go :: DmdType -> [Id] -> WithDmdType [Demand]
go DmdType
dmd_ty []  = DmdType -> [Demand] -> WithDmdType [Demand]
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
dmd_ty []
    go DmdType
dmd_ty (Id
b:[Id]
bs)
      | Id -> Bool
isId Id
b    = let WithDmdType DmdType
dmd_ty1 [Demand]
dmds = DmdType -> [Id] -> WithDmdType [Demand]
go DmdType
dmd_ty [Id]
bs
                        WithDmdType DmdType
dmd_ty2 Demand
dmd  = AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
dmd_ty1 Id
b
                    in DmdType -> [Demand] -> WithDmdType [Demand]
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
dmd_ty2  (Demand
dmd Demand -> [Demand] -> [Demand]
forall a. a -> [a] -> [a]
: [Demand]
dmds)
      | Bool
otherwise = DmdType -> [Id] -> WithDmdType [Demand]
go DmdType
dmd_ty [Id]
bs

findBndrDmd :: AnalEnv -> DmdType -> Id -> WithDmdType Demand
-- See Note [Trimming a demand to a type]
findBndrDmd :: AnalEnv -> DmdType -> Id -> WithDmdType Demand
findBndrDmd AnalEnv
env DmdType
dmd_ty Id
id
  = -- pprTrace "findBndrDmd" (ppr id $$ ppr dmd_ty $$ ppr starting_dmd $$ ppr dmd') $
    DmdType -> Demand -> WithDmdType Demand
forall a. DmdType -> a -> WithDmdType a
WithDmdType DmdType
dmd_ty' Demand
dmd'
  where
    dmd' :: Demand
dmd' = Demand -> Demand
strictify (Demand -> Demand) -> Demand -> Demand
forall a b. (a -> b) -> a -> b
$
           Demand -> TypeShape -> Demand
trimToType Demand
starting_dmd (FamInstEnvs -> Type -> TypeShape
findTypeShape FamInstEnvs
fam_envs Type
id_ty)

    (DmdType
dmd_ty', Demand
starting_dmd) = DmdType -> Id -> (DmdType, Demand)
peelFV DmdType
dmd_ty Id
id

    id_ty :: Type
id_ty = Id -> Type
idType Id
id

    strictify :: Demand -> Demand
strictify Demand
dmd
      -- See Note [Making dictionary parameters strict]
      -- and Note [Do not strictify a DFun's parameter dictionaries]
      | DmdAnalOpts -> Bool
dmd_strict_dicts (AnalEnv -> DmdAnalOpts
ae_opts AnalEnv
env)
      = Type -> Demand -> Demand
strictifyDictDmd Type
id_ty Demand
dmd
      | Bool
otherwise
      = Demand
dmd

    fam_envs :: FamInstEnvs
fam_envs = AnalEnv -> FamInstEnvs
ae_fam_envs AnalEnv
env

{- Note [Bringing a new variable into scope]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
   f x = blah
   g = ...(\f. ...f...)...

In the body of the '\f', any occurrence of `f` refers to the lambda-bound `f`,
not the top-level `f` (which will be in `ae_sigs`).  So it's very important
to delete `f` from `ae_sigs` when we pass a lambda/case/let-up binding of `f`.
Otherwise chaos results (#22718).

Note [Making dictionary parameters strict]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The Opt_DictsStrict flag makes GHC use call-by-value for dictionaries.  Why?

* Generally CBV is more efficient.

* A datatype dictionary is always non-bottom and never takes much work to
  compute.  E.g. a DFun from an instance decl always returns a dictionary
  record immediately.  See DFunUnfolding in CoreSyn.
  See also Note [Recursive superclasses] in TcInstDcls.

See #17758 for more background and perf numbers.

Wrinkles:

* A newtype dictionary is *not* always non-bottom.  E.g.
      class C a where op :: a -> a
      instance C Int where op = error "urk"
  Now a value of type (C Int) is just a newtype wrapper (a cast) around
  the error thunk.  Don't strictify these!

* Strictifying DFuns risks destroying the invariant that DFuns never take much
  work to compute, so we don't do it.
  See Note [Do not strictify a DFun's parameter dictionaries] for details.

* Although worker/wrapper *could* unbox strictly used dictionaries, we do not do
  so; see Note [Do not unbox class dictionaries].

The implementation is extremely simple: just make the strictness
analyser strictify the demand on a dictionary binder in
'findBndrDmd' if the binder does not belong to a DFun.

Note [Do not strictify a DFun's parameter dictionaries]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The typechecker can tie recursive knots involving (non-recursive) DFuns, so
we must not strictify a DFun's parameter dictionaries (#22549).
T22549 has an example involving undecidable instances that <<loop>>s when we
strictify the DFun of, e.g., `$fEqSeqT`:

  Main.$fEqSeqT
    = \@m @a ($dEq :: Eq (m (ViewT m a))) ($dMonad :: Monad m) ->
        GHC.Classes.C:Eq @(SeqT m a) ($c== @m @a $dEq $dMonad)
                                     ($c/= @m @a $dEq $dMonad)

  Rec {
    $dEq_a = Main.$fEqSeqT @Identity @Int $dEq_b Main.$fMonadIdentity
    $dEq_b = ... $dEq_a ... <another strict context due to DFun>
  }

If we make `$fEqSeqT` strict in `$dEq`, we'll collapse the Rec group into a
giant, <<loop>>ing thunk.

To prevent that, we never strictify dictionary params when inside a DFun.
That is implemented by unsetting 'dmd_strict_dicts' when entering a DFun.

See also Note [Speculative evaluation] in GHC.CoreToStg.Prep which has a rather
similar example in #20836. We may never speculate *arguments* of (recursive)
DFun calls, likewise we should not mark *formal parameters* of recursive DFuns
as strict.

Note [Initialising strictness]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
See section 9.2 (Finding fixpoints) of the paper.

Our basic plan is to initialise the strictness of each Id in a
recursive group to "bottom", and find a fixpoint from there.  However,
this group B might be inside an *enclosing* recursive group A, in
which case we'll do the entire fixpoint shebang on for each iteration
of A. This can be illustrated by the following example:

Example:

  f [] = []
  f (x:xs) = let g []     = f xs
                 g (y:ys) = y+1 : g ys
              in g (h x)

At each iteration of the fixpoint for f, the analyser has to find a
fixpoint for the enclosed function g. In the meantime, the demand
values for g at each iteration for f are *greater* than those we
encountered in the previous iteration for f. Therefore, we can begin
the fixpoint for g not with the bottom value but rather with the
result of the previous analysis. I.e., when beginning the fixpoint
process for g, we can start from the demand signature computed for g
previously and attached to the binding occurrence of g.

To speed things up, we initialise each iteration of A (the enclosing
one) from the result of the last one, which is neatly recorded in each
binder.  That way we make use of earlier iterations of the fixpoint
algorithm. (Cunning plan.)

But on the *first* iteration we want to *ignore* the current strictness
of the Id, and start from "bottom".  Nowadays the Id can have a current
strictness, because interface files record strictness for nested bindings.
To know when we are in the first iteration, we look at the ae_virgin
field of the AnalEnv.


Note [Final Demand Analyser run]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Some of the information that the demand analyser determines is not always
preserved by the simplifier.  For example, the simplifier will happily rewrite
  \y [Demand=MU] let x = y in x + x
to
  \y [Demand=MU] y + y
which is quite a lie: Now y occurs more than just once.

The once-used information is (currently) only used by the code
generator, though.  So:

 * We zap the used-once info in the worker-wrapper;
   see Note [Zapping Used Once info in WorkWrap] in
   GHC.Core.Opt.WorkWrap.
   If it's not reliable, it's better not to have it at all.

 * Just before TidyCore, we add a pass of the demand analyser,
      but WITHOUT subsequent worker/wrapper and simplifier,
   right before TidyCore.  See SimplCore.getCoreToDo.

   This way, correct information finds its way into the module interface
   (strictness signatures!) and the code generator (single-entry thunks!)

Note that, in contrast, the single-call information (C(M,..)) /can/ be
relied upon, as the simplifier tends to be very careful about not
duplicating actual function calls.

Also see #11731.

Note [Space Leaks in Demand Analysis]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Ticket: #15455
MR: !5399

In the past the result of demand analysis was not forced until the whole module
had finished being analysed. In big programs, this led to a big build up of thunks
which were all ultimately forced at the end of the analysis.

This was because the return type of the analysis was a lazy pair:
  dmdAnal :: AnalEnv -> SubDemand -> CoreExpr -> (DmdType, CoreExpr)
To avoid space leaks we added extra bangs to evaluate the DmdType component eagerly; but
we were never sure we had added enough.
The easiest way to systematically fix this was to use a strict pair type for the
return value of the analysis so that we can be more confident that the result
is incrementally computed rather than all at the end.

A second, only loosely related point is that
the updating of Ids was not forced because the result of updating
an Id was placed into a lazy field in CoreExpr. This meant that until the end of
demand analysis, the unforced Ids would retain the DmdEnv which the demand information
was fetch from. Now we are quite careful to force Ids before putting them
back into core expressions so that we can garbage-collect the environments more eagerly.
For example see the `Case` branch of `dmdAnal'` where `case_bndr'` is forced
or `dmdAnalSumAlt`.

The net result of all these improvements is the peak live memory usage of compiling
jsaddle-dom decreases about 4GB (from 6.5G to 2.5G). A bunch of bytes allocated benchmarks also
decrease because we allocate a lot fewer thunks which we immediately overwrite and
also runtime for the pass is faster! Overall, good wins.

-}