{-
(c) The University of Glasgow 2006
(c) The AQUA Project, Glasgow University, 1994-1998


Core-syntax unfoldings

Unfoldings (which can travel across module boundaries) are in Core
syntax (namely @CoreExpr@s).

The type @Unfolding@ sits ``above'' simply-Core-expressions
unfoldings, capturing ``higher-level'' things we know about a binding,
usually things that the simplifier found out (e.g., ``it's a
literal'').  In the corner of a @CoreUnfolding@ unfolding, you will
find, unsurprisingly, a Core expression.
-}



module GHC.Core.Unfold (
        Unfolding, UnfoldingGuidance,   -- Abstract types

        ExprSize(..), sizeExpr,

        ArgSummary(..), nonTriv,
        CallCtxt(..),

        UnfoldingOpts (..), defaultUnfoldingOpts,
        updateCreationThreshold, updateUseThreshold,
        updateFunAppDiscount, updateDictDiscount,
        updateVeryAggressive, updateCaseScaling,
        updateCaseThreshold, updateReportPrefix,

        inlineBoringOk, calcUnfoldingGuidance
    ) where

import GHC.Prelude

import GHC.Core
import GHC.Core.Utils
import GHC.Types.Id
import GHC.Core.DataCon
import GHC.Types.Literal
import GHC.Builtin.PrimOps
import GHC.Types.Id.Info
import GHC.Types.RepType ( isZeroBitTy )
import GHC.Types.Basic  ( Arity, RecFlag )
import GHC.Core.Type
import GHC.Builtin.Names
import GHC.Data.Bag
import GHC.Utils.Misc
import GHC.Utils.Outputable
import GHC.Types.ForeignCall
import GHC.Types.Tickish

import qualified Data.ByteString as BS

-- | Unfolding options
data UnfoldingOpts = UnfoldingOpts
   { UnfoldingOpts -> Int
unfoldingCreationThreshold :: !Int
      -- ^ Threshold above which unfoldings are not *created*

   , UnfoldingOpts -> Int
unfoldingUseThreshold :: !Int
      -- ^ Threshold above which unfoldings are not *inlined*

   , UnfoldingOpts -> Int
unfoldingFunAppDiscount :: !Int
      -- ^ Discount for lambdas that are used (applied)

   , UnfoldingOpts -> Int
unfoldingDictDiscount :: !Int
      -- ^ Discount for dictionaries

   , UnfoldingOpts -> Bool
unfoldingVeryAggressive :: !Bool
      -- ^ Force inlining in many more cases

   , UnfoldingOpts -> Int
unfoldingCaseThreshold :: !Int
      -- ^ Don't consider depth up to x

   , UnfoldingOpts -> Int
unfoldingCaseScaling :: !Int
      -- ^ Penalize depth with 1/x

   , UnfoldingOpts -> Maybe String
unfoldingReportPrefix :: !(Maybe String)
      -- ^ Only report inlining decisions for names with this prefix
   }

defaultUnfoldingOpts :: UnfoldingOpts
defaultUnfoldingOpts :: UnfoldingOpts
defaultUnfoldingOpts = UnfoldingOpts
   { unfoldingCreationThreshold :: Int
unfoldingCreationThreshold = Int
750
      -- The unfoldingCreationThreshold threshold must be reasonably high
      -- to take account of possible discounts.
      -- E.g. 450 is not enough in 'fulsom' for Interval.sqr to
      -- inline into Csg.calc (The unfolding for sqr never makes it
      -- into the interface file.)

   , unfoldingUseThreshold :: Int
unfoldingUseThreshold   = Int
90
      -- Last adjusted upwards in #18282, when I reduced
      -- the result discount for constructors.

   , unfoldingFunAppDiscount :: Int
unfoldingFunAppDiscount = Int
60
      -- Be fairly keen to inline a function if that means
      -- we'll be able to pick the right method from a dictionary

   , unfoldingDictDiscount :: Int
unfoldingDictDiscount   = Int
30
      -- Be fairly keen to inline a function if that means
      -- we'll be able to pick the right method from a dictionary

   , unfoldingVeryAggressive :: Bool
unfoldingVeryAggressive = Bool
False

      -- Only apply scaling once we are deeper than threshold cases
      -- in an RHS.
   , unfoldingCaseThreshold :: Int
unfoldingCaseThreshold = Int
2

      -- Penalize depth with (size*depth)/scaling
   , unfoldingCaseScaling :: Int
unfoldingCaseScaling = Int
30

      -- Don't filter inlining decision reports
   , unfoldingReportPrefix :: Maybe String
unfoldingReportPrefix = Maybe String
forall a. Maybe a
Nothing
   }

-- Helpers for "GHC.Driver.Session"

updateCreationThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCreationThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCreationThreshold Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingCreationThreshold = n }

updateUseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateUseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateUseThreshold Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingUseThreshold = n }

updateFunAppDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts
updateFunAppDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts
updateFunAppDiscount Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingFunAppDiscount = n }

updateDictDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts
updateDictDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts
updateDictDiscount Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingDictDiscount = n }

updateVeryAggressive :: Bool -> UnfoldingOpts -> UnfoldingOpts
updateVeryAggressive :: Bool -> UnfoldingOpts -> UnfoldingOpts
updateVeryAggressive Bool
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingVeryAggressive = n }


updateCaseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCaseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCaseThreshold Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingCaseThreshold = n }

updateCaseScaling :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCaseScaling :: Int -> UnfoldingOpts -> UnfoldingOpts
updateCaseScaling Int
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingCaseScaling = n }

updateReportPrefix :: Maybe String -> UnfoldingOpts -> UnfoldingOpts
updateReportPrefix :: Maybe String -> UnfoldingOpts -> UnfoldingOpts
updateReportPrefix Maybe String
n UnfoldingOpts
opts = UnfoldingOpts
opts { unfoldingReportPrefix = n }

data ArgSummary = TrivArg       -- Nothing interesting
                | NonTrivArg    -- Arg has structure
                | ValueArg      -- Arg is a con-app or PAP
                                -- ..or con-like. Note [Conlike is interesting]

instance Outputable ArgSummary where
  ppr :: ArgSummary -> SDoc
ppr ArgSummary
TrivArg    = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"TrivArg"
  ppr ArgSummary
NonTrivArg = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"NonTrivArg"
  ppr ArgSummary
ValueArg   = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"ValueArg"

nonTriv ::  ArgSummary -> Bool
nonTriv :: ArgSummary -> Bool
nonTriv ArgSummary
TrivArg = Bool
False
nonTriv ArgSummary
_       = Bool
True

data CallCtxt
  = BoringCtxt
  | RhsCtxt RecFlag     -- Rhs of a let-binding; see Note [RHS of lets]
  | DiscArgCtxt         -- Argument of a function with non-zero arg discount
  | RuleArgCtxt         -- We are somewhere in the argument of a function with rules

  | ValAppCtxt          -- We're applied to at least one value arg
                        -- This arises when we have ((f x |> co) y)
                        -- Then the (f x) has argument 'x' but in a ValAppCtxt

  | CaseCtxt            -- We're the scrutinee of a case
                        -- that decomposes its scrutinee

instance Outputable CallCtxt where
  ppr :: CallCtxt -> SDoc
ppr CallCtxt
CaseCtxt    = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"CaseCtxt"
  ppr CallCtxt
ValAppCtxt  = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"ValAppCtxt"
  ppr CallCtxt
BoringCtxt  = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"BoringCtxt"
  ppr (RhsCtxt RecFlag
ir)= String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"RhsCtxt" SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc
parens (RecFlag -> SDoc
forall a. Outputable a => a -> SDoc
ppr RecFlag
ir)
  ppr CallCtxt
DiscArgCtxt = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"DiscArgCtxt"
  ppr CallCtxt
RuleArgCtxt = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"RuleArgCtxt"

{-
Note [Calculate unfolding guidance on the non-occ-anal'd expression]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Notice that we give the non-occur-analysed expression to
calcUnfoldingGuidance.  In some ways it'd be better to occur-analyse
first; for example, sometimes during simplification, there's a large
let-bound thing which has been substituted, and so is now dead; so
'expr' contains two copies of the thing while the occurrence-analysed
expression doesn't.

Nevertheless, we *don't* and *must not* occ-analyse before computing
the size because

a) The size computation bales out after a while, whereas occurrence
   analysis does not.

b) Residency increases sharply if you occ-anal first.  I'm not
   100% sure why, but it's a large effect.  Compiling Cabal went
   from residency of 534M to over 800M with this one change.

This can occasionally mean that the guidance is very pessimistic;
it gets fixed up next round.  And it should be rare, because large
let-bound things that are dead are usually caught by preInlineUnconditionally


************************************************************************
*                                                                      *
\subsection{The UnfoldingGuidance type}
*                                                                      *
************************************************************************
-}

inlineBoringOk :: CoreExpr -> Bool
-- See Note [INLINE for small functions]
-- True => the result of inlining the expression is
--         no bigger than the expression itself
--     eg      (\x y -> f y x)
-- This is a quick and dirty version. It doesn't attempt
-- to deal with  (\x y z -> x (y z))
-- The really important one is (x `cast` c)
inlineBoringOk :: CoreExpr -> Bool
inlineBoringOk CoreExpr
e
  = Int -> CoreExpr -> Bool
go Int
0 CoreExpr
e
  where
    go :: Int -> CoreExpr -> Bool
    go :: Int -> CoreExpr -> Bool
go Int
credit (Lam Id
x CoreExpr
e) | Id -> Bool
isId Id
x           = Int -> CoreExpr -> Bool
go (Int
creditInt -> Int -> Int
forall a. Num a => a -> a -> a
+Int
1) CoreExpr
e
                        | Bool
otherwise        = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
e
        -- See Note [Count coercion arguments in boring contexts]
    go Int
credit (App CoreExpr
f (Type {}))            = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
f
    go Int
credit (App CoreExpr
f CoreExpr
a) | Int
credit Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
0
                        , CoreExpr -> Bool
exprIsTrivial CoreExpr
a  = Int -> CoreExpr -> Bool
go (Int
creditInt -> Int -> Int
forall a. Num a => a -> a -> a
-Int
1) CoreExpr
f
    go Int
credit (Tick CoreTickish
_ CoreExpr
e)                   = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
e -- dubious
    go Int
credit (Cast CoreExpr
e CoercionR
_)                   = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
e
    go Int
credit (Case CoreExpr
e Id
b Type
_ [Alt Id]
alts)
      | [Alt Id] -> Bool
forall a. [a] -> Bool
forall (t :: * -> *) a. Foldable t => t a -> Bool
null [Alt Id]
alts
      = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
e   -- EmptyCase is like e
      | Just CoreExpr
rhs <- CoreExpr -> Id -> [Alt Id] -> Maybe CoreExpr
isUnsafeEqualityCase CoreExpr
e Id
b [Alt Id]
alts
      = Int -> CoreExpr -> Bool
go Int
credit CoreExpr
rhs -- See Note [Inline unsafeCoerce]
    go Int
_      (Var {})                     = Bool
boringCxtOk
    go Int
_      (Lit Literal
l)                      = Literal -> Bool
litIsTrivial Literal
l Bool -> Bool -> Bool
&& Bool
boringCxtOk
    go Int
_      CoreExpr
_                            = Bool
boringCxtNotOk

calcUnfoldingGuidance
        :: UnfoldingOpts
        -> Bool          -- Definitely a top-level, bottoming binding
        -> CoreExpr      -- Expression to look at
        -> UnfoldingGuidance
calcUnfoldingGuidance :: UnfoldingOpts -> Bool -> CoreExpr -> UnfoldingGuidance
calcUnfoldingGuidance UnfoldingOpts
opts Bool
is_top_bottoming (Tick CoreTickish
t CoreExpr
expr)
  | Bool -> Bool
not (CoreTickish -> Bool
forall (pass :: TickishPass). GenTickish pass -> Bool
tickishIsCode CoreTickish
t)  -- non-code ticks don't matter for unfolding
  = UnfoldingOpts -> Bool -> CoreExpr -> UnfoldingGuidance
calcUnfoldingGuidance UnfoldingOpts
opts Bool
is_top_bottoming CoreExpr
expr
calcUnfoldingGuidance UnfoldingOpts
opts Bool
is_top_bottoming CoreExpr
expr
  = case UnfoldingOpts -> Int -> [Id] -> CoreExpr -> ExprSize
sizeExpr UnfoldingOpts
opts Int
bOMB_OUT_SIZE [Id]
val_bndrs CoreExpr
body of
      ExprSize
TooBig -> UnfoldingGuidance
UnfNever
      SizeIs Int
size Bag (Id, Int)
cased_bndrs Int
scrut_discount
        | CoreExpr -> Int -> Int -> Bool
uncondInline CoreExpr
expr Int
n_val_bndrs Int
size
        -> UnfWhen { ug_unsat_ok :: Bool
ug_unsat_ok = Bool
unSaturatedOk
                   , ug_boring_ok :: Bool
ug_boring_ok =  Bool
boringCxtOk
                   , ug_arity :: Int
ug_arity = Int
n_val_bndrs }   -- Note [INLINE for small functions]

        | Bool
is_top_bottoming
        -> UnfoldingGuidance
UnfNever   -- See Note [Do not inline top-level bottoming functions]

        | Bool
otherwise
        -> UnfIfGoodArgs { ug_args :: [Int]
ug_args  = (Id -> Int) -> [Id] -> [Int]
forall a b. (a -> b) -> [a] -> [b]
map (Bag (Id, Int) -> Id -> Int
mk_discount Bag (Id, Int)
cased_bndrs) [Id]
val_bndrs
                         , ug_size :: Int
ug_size  = Int
size
                         , ug_res :: Int
ug_res   = Int
scrut_discount }

  where
    ([Id]
bndrs, CoreExpr
body) = CoreExpr -> ([Id], CoreExpr)
forall b. Expr b -> ([b], Expr b)
collectBinders CoreExpr
expr
    bOMB_OUT_SIZE :: Int
bOMB_OUT_SIZE = UnfoldingOpts -> Int
unfoldingCreationThreshold UnfoldingOpts
opts
           -- Bomb out if size gets bigger than this
    val_bndrs :: [Id]
val_bndrs   = (Id -> Bool) -> [Id] -> [Id]
forall a. (a -> Bool) -> [a] -> [a]
filter Id -> Bool
isId [Id]
bndrs
    n_val_bndrs :: Int
n_val_bndrs = [Id] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [Id]
val_bndrs

    mk_discount :: Bag (Id,Int) -> Id -> Int
    mk_discount :: Bag (Id, Int) -> Id -> Int
mk_discount Bag (Id, Int)
cbs Id
bndr = (Int -> (Id, Int) -> Int) -> Int -> Bag (Id, Int) -> Int
forall b a. (b -> a -> b) -> b -> Bag a -> b
forall (t :: * -> *) b a.
Foldable t =>
(b -> a -> b) -> b -> t a -> b
foldl' Int -> (Id, Int) -> Int
combine Int
0 Bag (Id, Int)
cbs
           where
             combine :: Int -> (Id, Int) -> Int
combine Int
acc (Id
bndr', Int
disc)
               | Id
bndr Id -> Id -> Bool
forall a. Eq a => a -> a -> Bool
== Id
bndr' = Int
acc Int -> Int -> Int
`plus_disc` Int
disc
               | Bool
otherwise     = Int
acc

             plus_disc :: Int -> Int -> Int
             plus_disc :: Int -> Int -> Int
plus_disc | Type -> Bool
isFunTy (Id -> Type
idType Id
bndr) = Int -> Int -> Int
forall a. Ord a => a -> a -> a
max
                       | Bool
otherwise             = Int -> Int -> Int
forall a. Num a => a -> a -> a
(+)
             -- See Note [Function and non-function discounts]

{- Note [Inline unsafeCoerce]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We really want to inline unsafeCoerce, even when applied to boring
arguments.  It doesn't look as if its RHS is smaller than the call
   unsafeCoerce x = case unsafeEqualityProof @a @b of UnsafeRefl -> x
but that case is discarded in CoreToStg -- see Note [Implementing unsafeCoerce]
in base:Unsafe.Coerce.

Moreover, if we /don't/ inline it, we may be left with
          f (unsafeCoerce x)
which will build a thunk -- bad, bad, bad.

Conclusion: we really want inlineBoringOk to be True of the RHS of
unsafeCoerce. And it really is, because we regard
  case unsafeEqualityProof @a @b of UnsafeRefl -> rhs
as trivial iff rhs is. This is (U4) in Note [Implementing unsafeCoerce].

Note [Computing the size of an expression]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The basic idea of sizeExpr is obvious enough: count nodes.  But getting the
heuristics right has taken a long time.  Here's the basic strategy:

    * Variables, literals: 0
      (Exception for string literals, see litSize.)

    * Function applications (f e1 .. en): 1 + #value args

    * Constructor applications: 1, regardless of #args

    * Let(rec): 1 + size of components

    * Note, cast: 0

Examples

  Size  Term
  --------------
    0     42#
    0     x
    0     True
    2     f x
    1     Just x
    4     f (g x)

Notice that 'x' counts 0, while (f x) counts 2.  That's deliberate: there's
a function call to account for.  Notice also that constructor applications
are very cheap, because exposing them to a caller is so valuable.

[25/5/11] All sizes are now multiplied by 10, except for primops
(which have sizes like 1 or 4.  This makes primops look fantastically
cheap, and seems to be almost universally beneficial.  Done partly as a
result of #4978.

Note [Do not inline top-level bottoming functions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The FloatOut pass has gone to some trouble to float out calls to 'error'
and similar friends.  See Note [Bottoming floats] in GHC.Core.Opt.SetLevels.
Do not re-inline them!  But we *do* still inline if they are very small
(the uncondInline stuff).

Note [INLINE for small functions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider        {-# INLINE f #-}
                f x = Just x
                g y = f y
Then f's RHS is no larger than its LHS, so we should inline it into
even the most boring context.  In general, f the function is
sufficiently small that its body is as small as the call itself, the
inline unconditionally, regardless of how boring the context is.

Things to note:

(1) We inline *unconditionally* if inlined thing is smaller (using sizeExpr)
    than the thing it's replacing.  Notice that
      (f x) --> (g 3)             -- YES, unconditionally
      (f x) --> x : []            -- YES, *even though* there are two
                                  --      arguments to the cons
      x     --> g 3               -- NO
      x     --> Just v            -- NO

    It's very important not to unconditionally replace a variable by
    a non-atomic term.

(2) We do this even if the thing isn't saturated, else we end up with the
    silly situation that
       f x y = x
       ...map (f 3)...
    doesn't inline.  Even in a boring context, inlining without being
    saturated will give a lambda instead of a PAP, and will be more
    efficient at runtime.

(3) However, when the function's arity > 0, we do insist that it
    has at least one value argument at the call site.  (This check is
    made in the UnfWhen case of callSiteInline.) Otherwise we find this:
         f = /\a \x:a. x
         d = /\b. MkD (f b)
    If we inline f here we get
         d = /\b. MkD (\x:b. x)
    and then prepareRhs floats out the argument, abstracting the type
    variables, so we end up with the original again!

(4) We must be much more cautious about arity-zero things. Consider
       let x = y +# z in ...
    In *size* terms primops look very small, because the generate a
    single instruction, but we do not want to unconditionally replace
    every occurrence of x with (y +# z).  So we only do the
    unconditional-inline thing for *trivial* expressions.

    NB: you might think that PostInlineUnconditionally would do this
    but it doesn't fire for top-level things; see GHC.Core.Opt.Simplify.Utils
    Note [Top level and postInlineUnconditionally]

Note [Count coercion arguments in boring contexts]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In inlineBoringOK, we ignore type arguments when deciding whether an
expression is okay to inline into boring contexts. This is good, since
if we have a definition like

  let y = x @Int in f y y

there’s no reason not to inline y at both use sites — no work is
actually duplicated. It may seem like the same reasoning applies to
coercion arguments, and indeed, in #17182 we changed inlineBoringOK to
treat coercions the same way.

However, this isn’t a good idea: unlike type arguments, which have
no runtime representation, coercion arguments *do* have a runtime
representation (albeit the zero-width VoidRep, see Note [Coercion tokens]
in "GHC.CoreToStg"). This caused trouble in #17787 for DataCon wrappers for
nullary GADT constructors: the wrappers would be inlined and each use of
the constructor would lead to a separate allocation instead of just
sharing the wrapper closure.

The solution: don’t ignore coercion arguments after all.
-}

uncondInline :: CoreExpr -> Arity -> Int -> Bool
-- Inline unconditionally if there no size increase
-- Size of call is arity (+1 for the function)
-- See Note [INLINE for small functions]
uncondInline :: CoreExpr -> Int -> Int -> Bool
uncondInline CoreExpr
rhs Int
arity Int
size
  | Int
arity Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
0 = Int
size Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
<= Int
10 Int -> Int -> Int
forall a. Num a => a -> a -> a
* (Int
arity Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
1) -- See Note [INLINE for small functions] (1)
  | Bool
otherwise = CoreExpr -> Bool
exprIsTrivial CoreExpr
rhs        -- See Note [INLINE for small functions] (4)

sizeExpr :: UnfoldingOpts
         -> Int             -- Bomb out if it gets bigger than this
         -> [Id]            -- Arguments; we're interested in which of these
                            -- get case'd
         -> CoreExpr
         -> ExprSize

-- Note [Computing the size of an expression]

-- Forcing bOMB_OUT_SIZE early prevents repeated
-- unboxing of the Int argument.
sizeExpr :: UnfoldingOpts -> Int -> [Id] -> CoreExpr -> ExprSize
sizeExpr UnfoldingOpts
opts !Int
bOMB_OUT_SIZE [Id]
top_args CoreExpr
expr
  = CoreExpr -> ExprSize
size_up CoreExpr
expr
  where
    size_up :: CoreExpr -> ExprSize
size_up (Cast CoreExpr
e CoercionR
_) = CoreExpr -> ExprSize
size_up CoreExpr
e
    size_up (Tick CoreTickish
_ CoreExpr
e) = CoreExpr -> ExprSize
size_up CoreExpr
e
    size_up (Type Type
_)   = ExprSize
sizeZero           -- Types cost nothing
    size_up (Coercion CoercionR
_) = ExprSize
sizeZero
    size_up (Lit Literal
lit)  = Int -> ExprSize
sizeN (Literal -> Int
litSize Literal
lit)
    size_up (Var Id
f) | Id -> Bool
isZeroBitId Id
f = ExprSize
sizeZero
                      -- Make sure we get constructor discounts even
                      -- on nullary constructors
                    | Bool
otherwise       = Id -> [CoreExpr] -> Int -> ExprSize
size_up_call Id
f [] Int
0

    size_up (App CoreExpr
fun CoreExpr
arg)
      | CoreExpr -> Bool
forall b. Expr b -> Bool
isTyCoArg CoreExpr
arg = CoreExpr -> ExprSize
size_up CoreExpr
fun
      | Bool
otherwise     = CoreExpr -> ExprSize
size_up CoreExpr
arg  ExprSize -> ExprSize -> ExprSize
`addSizeNSD`
                        CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
fun [CoreExpr
arg] (if CoreExpr -> Bool
forall b. Expr b -> Bool
isZeroBitExpr CoreExpr
arg then Int
1 else Int
0)

    size_up (Lam Id
b CoreExpr
e)
      | Id -> Bool
isId Id
b Bool -> Bool -> Bool
&& Bool -> Bool
not (Id -> Bool
isZeroBitId Id
b) = UnfoldingOpts -> ExprSize -> ExprSize
lamScrutDiscount UnfoldingOpts
opts (CoreExpr -> ExprSize
size_up CoreExpr
e ExprSize -> Int -> ExprSize
`addSizeN` Int
10)
      | Bool
otherwise = CoreExpr -> ExprSize
size_up CoreExpr
e

    size_up (Let (NonRec Id
binder CoreExpr
rhs) CoreExpr
body)
      = (Id, CoreExpr) -> ExprSize
size_up_rhs (Id
binder, CoreExpr
rhs) ExprSize -> ExprSize -> ExprSize
`addSizeNSD`
        CoreExpr -> ExprSize
size_up CoreExpr
body              ExprSize -> Int -> ExprSize
`addSizeN`
        Id -> Int
forall {a}. Num a => Id -> a
size_up_alloc Id
binder

    size_up (Let (Rec [(Id, CoreExpr)]
pairs) CoreExpr
body)
      = ((Id, CoreExpr) -> ExprSize -> ExprSize)
-> ExprSize -> [(Id, CoreExpr)] -> ExprSize
forall a b. (a -> b -> b) -> b -> [a] -> b
forall (t :: * -> *) a b.
Foldable t =>
(a -> b -> b) -> b -> t a -> b
foldr (ExprSize -> ExprSize -> ExprSize
addSizeNSD (ExprSize -> ExprSize -> ExprSize)
-> ((Id, CoreExpr) -> ExprSize)
-> (Id, CoreExpr)
-> ExprSize
-> ExprSize
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (Id, CoreExpr) -> ExprSize
size_up_rhs)
              (CoreExpr -> ExprSize
size_up CoreExpr
body ExprSize -> Int -> ExprSize
`addSizeN` [Int] -> Int
forall a. Num a => [a] -> a
forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a
sum (((Id, CoreExpr) -> Int) -> [(Id, CoreExpr)] -> [Int]
forall a b. (a -> b) -> [a] -> [b]
map (Id -> Int
forall {a}. Num a => Id -> a
size_up_alloc (Id -> Int) -> ((Id, CoreExpr) -> Id) -> (Id, CoreExpr) -> Int
forall b c a. (b -> c) -> (a -> b) -> a -> c
. (Id, CoreExpr) -> Id
forall a b. (a, b) -> a
fst) [(Id, CoreExpr)]
pairs))
              [(Id, CoreExpr)]
pairs

    size_up (Case CoreExpr
e Id
_ Type
_ [Alt Id]
alts)
        | [Alt Id] -> Bool
forall a. [a] -> Bool
forall (t :: * -> *) a. Foldable t => t a -> Bool
null [Alt Id]
alts
        = CoreExpr -> ExprSize
size_up CoreExpr
e    -- case e of {} never returns, so take size of scrutinee

    size_up (Case CoreExpr
e Id
_ Type
_ [Alt Id]
alts)
        -- Now alts is non-empty
        | Just Id
v <- CoreExpr -> Maybe Id
is_top_arg CoreExpr
e -- We are scrutinising an argument variable
        = let
            alt_sizes :: [ExprSize]
alt_sizes = (Alt Id -> ExprSize) -> [Alt Id] -> [ExprSize]
forall a b. (a -> b) -> [a] -> [b]
map Alt Id -> ExprSize
size_up_alt [Alt Id]
alts

                  -- alts_size tries to compute a good discount for
                  -- the case when we are scrutinising an argument variable
            alts_size :: ExprSize -> ExprSize -> ExprSize
alts_size (SizeIs Int
tot Bag (Id, Int)
tot_disc Int
tot_scrut)
                          -- Size of all alternatives
                      (SizeIs Int
max Bag (Id, Int)
_        Int
_)
                          -- Size of biggest alternative
                  = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
tot ((Id, Int) -> Bag (Id, Int)
forall a. a -> Bag a
unitBag (Id
v, Int
20 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
tot Int -> Int -> Int
forall a. Num a => a -> a -> a
- Int
max)
                      Bag (Id, Int) -> Bag (Id, Int) -> Bag (Id, Int)
forall a. Bag a -> Bag a -> Bag a
`unionBags` Bag (Id, Int)
tot_disc) Int
tot_scrut
                          -- If the variable is known, we produce a
                          -- discount that will take us back to 'max',
                          -- the size of the largest alternative The
                          -- 1+ is a little discount for reduced
                          -- allocation in the caller
                          --
                          -- Notice though, that we return tot_disc,
                          -- the total discount from all branches.  I
                          -- think that's right.

            alts_size ExprSize
tot_size ExprSize
_ = ExprSize
tot_size
          in
          ExprSize -> ExprSize -> ExprSize
alts_size ((ExprSize -> ExprSize -> ExprSize) -> [ExprSize] -> ExprSize
forall a. (a -> a -> a) -> [a] -> a
forall (t :: * -> *) a. Foldable t => (a -> a -> a) -> t a -> a
foldr1 ExprSize -> ExprSize -> ExprSize
addAltSize [ExprSize]
alt_sizes)  -- alts is non-empty
                    ((ExprSize -> ExprSize -> ExprSize) -> [ExprSize] -> ExprSize
forall a. (a -> a -> a) -> [a] -> a
forall (t :: * -> *) a. Foldable t => (a -> a -> a) -> t a -> a
foldr1 ExprSize -> ExprSize -> ExprSize
maxSize    [ExprSize]
alt_sizes)
                -- Good to inline if an arg is scrutinised, because
                -- that may eliminate allocation in the caller
                -- And it eliminates the case itself
        where
          is_top_arg :: CoreExpr -> Maybe Id
is_top_arg (Var Id
v) | Id
v Id -> [Id] -> Bool
forall a. Eq a => a -> [a] -> Bool
forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool
`elem` [Id]
top_args = Id -> Maybe Id
forall a. a -> Maybe a
Just Id
v
          is_top_arg (Cast CoreExpr
e CoercionR
_) = CoreExpr -> Maybe Id
is_top_arg CoreExpr
e
          is_top_arg CoreExpr
_ = Maybe Id
forall a. Maybe a
Nothing


    size_up (Case CoreExpr
e Id
_ Type
_ [Alt Id]
alts) = CoreExpr -> ExprSize
size_up CoreExpr
e  ExprSize -> ExprSize -> ExprSize
`addSizeNSD`
                                (Alt Id -> ExprSize -> ExprSize)
-> ExprSize -> [Alt Id] -> ExprSize
forall a b. (a -> b -> b) -> b -> [a] -> b
forall (t :: * -> *) a b.
Foldable t =>
(a -> b -> b) -> b -> t a -> b
foldr (ExprSize -> ExprSize -> ExprSize
addAltSize (ExprSize -> ExprSize -> ExprSize)
-> (Alt Id -> ExprSize) -> Alt Id -> ExprSize -> ExprSize
forall b c a. (b -> c) -> (a -> b) -> a -> c
. Alt Id -> ExprSize
size_up_alt) ExprSize
case_size [Alt Id]
alts
      where
          case_size :: ExprSize
case_size
           | CoreExpr -> Bool
forall b. Expr b -> Bool
is_inline_scrut CoreExpr
e, [Alt Id] -> Int -> Bool
forall a. [a] -> Int -> Bool
lengthAtMost [Alt Id]
alts Int
1 = Int -> ExprSize
sizeN (-Int
10)
           | Bool
otherwise = ExprSize
sizeZero
                -- Normally we don't charge for the case itself, but
                -- we charge one per alternative (see size_up_alt,
                -- below) to account for the cost of the info table
                -- and comparisons.
                --
                -- However, in certain cases (see is_inline_scrut
                -- below), no code is generated for the case unless
                -- there are multiple alts.  In these cases we
                -- subtract one, making the first alt free.
                -- e.g. case x# +# y# of _ -> ...   should cost 1
                --      case touch# x# of _ -> ...  should cost 0
                -- (see #4978)
                --
                -- I would like to not have the "lengthAtMost alts 1"
                -- condition above, but without that some programs got worse
                -- (spectral/hartel/event and spectral/para).  I don't fully
                -- understand why. (SDM 24/5/11)

                -- unboxed variables, inline primops and unsafe foreign calls
                -- are all "inline" things:
          is_inline_scrut :: Expr b -> Bool
is_inline_scrut (Var Id
v) =
            HasDebugCallStack => Type -> Bool
Type -> Bool
isUnliftedType (Id -> Type
idType Id
v)
              -- isUnliftedType is OK here: scrutinees have a fixed RuntimeRep (search for FRRCase)
          is_inline_scrut Expr b
scrut
              | (Var Id
f, [Expr b]
_) <- Expr b -> (Expr b, [Expr b])
forall b. Expr b -> (Expr b, [Expr b])
collectArgs Expr b
scrut
                = case Id -> IdDetails
idDetails Id
f of
                    FCallId ForeignCall
fc    -> Bool -> Bool
not (ForeignCall -> Bool
isSafeForeignCall ForeignCall
fc)
                    PrimOpId PrimOp
op ConcreteTyVars
_ -> Bool -> Bool
not (PrimOp -> Bool
primOpOutOfLine PrimOp
op)
                    IdDetails
_other        -> Bool
False
              | Bool
otherwise
                = Bool
False

    size_up_rhs :: (Id, CoreExpr) -> ExprSize
size_up_rhs (Id
bndr, CoreExpr
rhs)
      | JoinPoint Int
join_arity <- Id -> JoinPointHood
idJoinPointHood Id
bndr
        -- Skip arguments to join point
      , ([Id]
_bndrs, CoreExpr
body) <- Int -> CoreExpr -> ([Id], CoreExpr)
forall b. Int -> Expr b -> ([b], Expr b)
collectNBinders Int
join_arity CoreExpr
rhs
      = CoreExpr -> ExprSize
size_up CoreExpr
body
      | Bool
otherwise
      = CoreExpr -> ExprSize
size_up CoreExpr
rhs

    ------------
    -- size_up_app is used when there's ONE OR MORE value args
    size_up_app :: CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app (App CoreExpr
fun CoreExpr
arg) [CoreExpr]
args Int
voids
        | CoreExpr -> Bool
forall b. Expr b -> Bool
isTyCoArg CoreExpr
arg                  = CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
fun [CoreExpr]
args Int
voids
        | CoreExpr -> Bool
forall b. Expr b -> Bool
isZeroBitExpr CoreExpr
arg              = CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
fun (CoreExpr
argCoreExpr -> [CoreExpr] -> [CoreExpr]
forall a. a -> [a] -> [a]
:[CoreExpr]
args) (Int
voids Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
1)
        | Bool
otherwise                      = CoreExpr -> ExprSize
size_up CoreExpr
arg  ExprSize -> ExprSize -> ExprSize
`addSizeNSD`
                                           CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
fun (CoreExpr
argCoreExpr -> [CoreExpr] -> [CoreExpr]
forall a. a -> [a] -> [a]
:[CoreExpr]
args) Int
voids
    size_up_app (Var Id
fun)     [CoreExpr]
args Int
voids = Id -> [CoreExpr] -> Int -> ExprSize
size_up_call Id
fun [CoreExpr]
args Int
voids
    size_up_app (Tick CoreTickish
_ CoreExpr
expr) [CoreExpr]
args Int
voids = CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
expr [CoreExpr]
args Int
voids
    size_up_app (Cast CoreExpr
expr CoercionR
_) [CoreExpr]
args Int
voids = CoreExpr -> [CoreExpr] -> Int -> ExprSize
size_up_app CoreExpr
expr [CoreExpr]
args Int
voids
    size_up_app CoreExpr
other         [CoreExpr]
args Int
voids = CoreExpr -> ExprSize
size_up CoreExpr
other ExprSize -> Int -> ExprSize
`addSizeN`
                                           Int -> Int -> Int
callSize ([CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
args) Int
voids
       -- if the lhs is not an App or a Var, or an invisible thing like a
       -- Tick or Cast, then we should charge for a complete call plus the
       -- size of the lhs itself.

    ------------
    size_up_call :: Id -> [CoreExpr] -> Int -> ExprSize
    size_up_call :: Id -> [CoreExpr] -> Int -> ExprSize
size_up_call Id
fun [CoreExpr]
val_args Int
voids
       = case Id -> IdDetails
idDetails Id
fun of
           FCallId ForeignCall
_        -> Int -> ExprSize
sizeN (Int -> Int -> Int
callSize ([CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
val_args) Int
voids)
           DataConWorkId DataCon
dc -> DataCon -> Int -> ExprSize
conSize    DataCon
dc ([CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
val_args)
           PrimOpId PrimOp
op ConcreteTyVars
_    -> PrimOp -> Int -> ExprSize
primOpSize PrimOp
op ([CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
val_args)
           ClassOpId {}     -> UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize
classOpSize UnfoldingOpts
opts [Id]
top_args [CoreExpr]
val_args
           IdDetails
_                -> UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize
funSize UnfoldingOpts
opts [Id]
top_args Id
fun ([CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
val_args) Int
voids

    ------------
    size_up_alt :: Alt Id -> ExprSize
size_up_alt (Alt AltCon
_con [Id]
_bndrs CoreExpr
rhs) = CoreExpr -> ExprSize
size_up CoreExpr
rhs ExprSize -> Int -> ExprSize
`addSizeN` Int
10
        -- Don't charge for args, so that wrappers look cheap
        -- (See comments about wrappers with Case)
        --
        -- IMPORTANT: *do* charge 1 for the alternative, else we
        -- find that giant case nests are treated as practically free
        -- A good example is Foreign.C.Error.errnoToIOError

    ------------
    -- Cost to allocate binding with given binder
    size_up_alloc :: Id -> a
size_up_alloc Id
bndr
      |  Id -> Bool
isTyVar Id
bndr                    -- Doesn't exist at runtime
      Bool -> Bool -> Bool
|| Id -> Bool
isJoinId Id
bndr                   -- Not allocated at all
      Bool -> Bool -> Bool
|| Bool -> Bool
not (Type -> Bool
isBoxedType (Id -> Type
idType Id
bndr)) -- Doesn't live in heap
      = a
0
      | Bool
otherwise
      = a
10

    ------------
        -- These addSize things have to be here because
        -- I don't want to give them bOMB_OUT_SIZE as an argument
    addSizeN :: ExprSize -> Int -> ExprSize
addSizeN ExprSize
TooBig          Int
_  = ExprSize
TooBig
    addSizeN (SizeIs Int
n Bag (Id, Int)
xs Int
d) Int
m  = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize
mkSizeIs Int
bOMB_OUT_SIZE (Int
n Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
m) Bag (Id, Int)
xs Int
d

        -- addAltSize is used to add the sizes of case alternatives
    addAltSize :: ExprSize -> ExprSize -> ExprSize
addAltSize ExprSize
TooBig            ExprSize
_      = ExprSize
TooBig
    addAltSize ExprSize
_                 ExprSize
TooBig = ExprSize
TooBig
    addAltSize (SizeIs Int
n1 Bag (Id, Int)
xs Int
d1) (SizeIs Int
n2 Bag (Id, Int)
ys Int
d2)
        = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize
mkSizeIs Int
bOMB_OUT_SIZE (Int
n1 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
n2)
                                 (Bag (Id, Int)
xs Bag (Id, Int) -> Bag (Id, Int) -> Bag (Id, Int)
forall a. Bag a -> Bag a -> Bag a
`unionBags` Bag (Id, Int)
ys)
                                 (Int
d1 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
d2) -- Note [addAltSize result discounts]

        -- This variant ignores the result discount from its LEFT argument
        -- It's used when the second argument isn't part of the result
    addSizeNSD :: ExprSize -> ExprSize -> ExprSize
addSizeNSD ExprSize
TooBig            ExprSize
_      = ExprSize
TooBig
    addSizeNSD ExprSize
_                 ExprSize
TooBig = ExprSize
TooBig
    addSizeNSD (SizeIs Int
n1 Bag (Id, Int)
xs Int
_) (SizeIs Int
n2 Bag (Id, Int)
ys Int
d2)
        = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize
mkSizeIs Int
bOMB_OUT_SIZE (Int
n1 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
n2)
                                 (Bag (Id, Int)
xs Bag (Id, Int) -> Bag (Id, Int) -> Bag (Id, Int)
forall a. Bag a -> Bag a -> Bag a
`unionBags` Bag (Id, Int)
ys)
                                 Int
d2  -- Ignore d1

    -- don't count expressions such as State# RealWorld
    -- exclude join points, because they can be rep-polymorphic
    -- and typePrimRep will crash
    isZeroBitId :: Id -> Bool
isZeroBitId Id
id = Bool -> Bool
not (Id -> Bool
isJoinId Id
id) Bool -> Bool -> Bool
&& HasDebugCallStack => Type -> Bool
Type -> Bool
isZeroBitTy (Id -> Type
idType Id
id)

    isZeroBitExpr :: Expr b -> Bool
isZeroBitExpr (Var Id
id)   = Id -> Bool
isZeroBitId Id
id
    isZeroBitExpr (Tick CoreTickish
_ Expr b
e) = Expr b -> Bool
isZeroBitExpr Expr b
e
    isZeroBitExpr Expr b
_          = Bool
False

-- | Finds a nominal size of a string literal.
litSize :: Literal -> Int
-- Used by GHC.Core.Unfold.sizeExpr
litSize :: Literal -> Int
litSize (LitNumber LitNumType
LitNumBigNat Integer
_)  = Int
100
litSize (LitString ByteString
str) = Int
10 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
10 Int -> Int -> Int
forall a. Num a => a -> a -> a
* ((ByteString -> Int
BS.length ByteString
str Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
3) Int -> Int -> Int
forall a. Integral a => a -> a -> a
`div` Int
4)
        -- If size could be 0 then @f "x"@ might be too small
        -- [Sept03: make literal strings a bit bigger to avoid fruitless
        --  duplication of little strings]
litSize Literal
_other = Int
0    -- Must match size of nullary constructors
                      -- Key point: if  x |-> 4, then x must inline unconditionally
                      --            (eg via case binding)

classOpSize :: UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize
-- See Note [Conlike is interesting]
classOpSize :: UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize
classOpSize UnfoldingOpts
_ [Id]
_ []
  = ExprSize
sizeZero
classOpSize UnfoldingOpts
opts [Id]
top_args (CoreExpr
arg1 : [CoreExpr]
other_args)
  = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
size Bag (Id, Int)
arg_discount Int
0
  where
    size :: Int
size = Int
20 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ (Int
10 Int -> Int -> Int
forall a. Num a => a -> a -> a
* [CoreExpr] -> Int
forall a. [a] -> Int
forall (t :: * -> *) a. Foldable t => t a -> Int
length [CoreExpr]
other_args)
    -- If the class op is scrutinising a lambda bound dictionary then
    -- give it a discount, to encourage the inlining of this function
    -- The actual discount is rather arbitrarily chosen
    arg_discount :: Bag (Id, Int)
arg_discount = case CoreExpr
arg1 of
                     Var Id
dict | Id
dict Id -> [Id] -> Bool
forall a. Eq a => a -> [a] -> Bool
forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool
`elem` [Id]
top_args
                              -> (Id, Int) -> Bag (Id, Int)
forall a. a -> Bag a
unitBag (Id
dict, UnfoldingOpts -> Int
unfoldingDictDiscount UnfoldingOpts
opts)
                     CoreExpr
_other   -> Bag (Id, Int)
forall a. Bag a
emptyBag

-- | The size of a function call
callSize
 :: Int  -- ^ number of value args
 -> Int  -- ^ number of value args that are void
 -> Int
callSize :: Int -> Int -> Int
callSize Int
n_val_args Int
voids = Int
10 Int -> Int -> Int
forall a. Num a => a -> a -> a
* (Int
1 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
n_val_args Int -> Int -> Int
forall a. Num a => a -> a -> a
- Int
voids)
        -- The 1+ is for the function itself
        -- Add 1 for each non-trivial arg;
        -- the allocation cost, as in let(rec)

-- | The size of a jump to a join point
jumpSize
 :: Int  -- ^ number of value args
 -> Int  -- ^ number of value args that are void
 -> Int
jumpSize :: Int -> Int -> Int
jumpSize Int
n_val_args Int
voids = Int
2 Int -> Int -> Int
forall a. Num a => a -> a -> a
* (Int
1 Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
n_val_args Int -> Int -> Int
forall a. Num a => a -> a -> a
- Int
voids)
  -- A jump is 20% the size of a function call. Making jumps free reopens
  -- bug #6048, but making them any more expensive loses a 21% improvement in
  -- spectral/puzzle. TODO Perhaps adjusting the default threshold would be a
  -- better solution?

funSize :: UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize
-- Size for functions that are not constructors or primops
-- Note [Function applications]
funSize :: UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize
funSize UnfoldingOpts
opts [Id]
top_args Id
fun Int
n_val_args Int
voids
  | Id
fun Id -> Unique -> Bool
forall a. Uniquable a => a -> Unique -> Bool
`hasKey` Unique
buildIdKey   = ExprSize
buildSize
  | Id
fun Id -> Unique -> Bool
forall a. Uniquable a => a -> Unique -> Bool
`hasKey` Unique
augmentIdKey = ExprSize
augmentSize
  | Bool
otherwise = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
size Bag (Id, Int)
arg_discount Int
res_discount
  where
    some_val_args :: Bool
some_val_args = Int
n_val_args Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
0
    is_join :: Bool
is_join = Id -> Bool
isJoinId Id
fun

    size :: Int
size | Bool
is_join              = Int -> Int -> Int
jumpSize Int
n_val_args Int
voids
         | Bool -> Bool
not Bool
some_val_args    = Int
0
         | Bool
otherwise            = Int -> Int -> Int
callSize Int
n_val_args Int
voids

        --                  DISCOUNTS
        --  See Note [Function and non-function discounts]
    arg_discount :: Bag (Id, Int)
arg_discount | Bool
some_val_args Bool -> Bool -> Bool
&& Id
fun Id -> [Id] -> Bool
forall a. Eq a => a -> [a] -> Bool
forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool
`elem` [Id]
top_args
                 = (Id, Int) -> Bag (Id, Int)
forall a. a -> Bag a
unitBag (Id
fun, UnfoldingOpts -> Int
unfoldingFunAppDiscount UnfoldingOpts
opts)
                 | Bool
otherwise = Bag (Id, Int)
forall a. Bag a
emptyBag
        -- If the function is an argument and is applied
        -- to some values, give it an arg-discount

    res_discount :: Int
res_discount | Id -> Int
idArity Id
fun Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
n_val_args = UnfoldingOpts -> Int
unfoldingFunAppDiscount UnfoldingOpts
opts
                 | Bool
otherwise                = Int
0
        -- If the function is partially applied, show a result discount
-- XXX maybe behave like ConSize for eval'd variable

conSize :: DataCon -> Int -> ExprSize
conSize :: DataCon -> Int -> ExprSize
conSize DataCon
dc Int
n_val_args
  | Int
n_val_args Int -> Int -> Bool
forall a. Eq a => a -> a -> Bool
== Int
0 = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
0 Bag (Id, Int)
forall a. Bag a
emptyBag Int
10    -- Like variables

-- See Note [Unboxed tuple size and result discount]
  | DataCon -> Bool
isUnboxedTupleDataCon DataCon
dc = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
0 Bag (Id, Int)
forall a. Bag a
emptyBag Int
10

-- See Note [Constructor size and result discount]
  | Bool
otherwise = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
10 Bag (Id, Int)
forall a. Bag a
emptyBag Int
10

{- Note [Constructor size and result discount]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Treat a constructors application as size 10, regardless of how many
arguments it has; we are keen to expose them (and we charge separately
for their args).  We can't treat them as size zero, else we find that
(Just x) has size 0, which is the same as a lone variable; and hence
'v' will always be replaced by (Just x), where v is bound to Just x.

The "result discount" is applied if the result of the call is
scrutinised (say by a case).  For a constructor application that will
mean the constructor application will disappear, so we don't need to
charge it to the function.  So the discount should at least match the
cost of the constructor application, namely 10.

Historical note 1: Until Jun 2020 we gave it a "bit of extra
incentive" via a discount of 10*(1 + n_val_args), but that was FAR too
much (#18282).  In particular, consider a huge case tree like

   let r = case y1 of
          Nothing -> B1 a b c
          Just v1 -> case y2 of
                      Nothing -> B1 c b a
                      Just v2 -> ...

If conSize gives a cost of 10 (regardless of n_val_args) and a
discount of 10, that'll make each alternative RHS cost zero.  We
charge 10 for each case alternative (see size_up_alt).  If we give a
bigger discount (say 20) in conSize, we'll make the case expression
cost *nothing*, and that can make a huge case tree cost nothing. This
leads to massive, sometimes exponential inlinings (#18282).  In short,
don't give a discount that give a negative size to a sub-expression!

Historical note 2: Much longer ago, Simon M tried a MUCH bigger
discount: (10 * (10 + n_val_args)), and said it was an "unambiguous
win", but its terribly dangerous because a function with many many
case branches, each finishing with a constructor, can have an
arbitrarily large discount.  This led to terrible code bloat: see #6099.

Note [Unboxed tuple size and result discount]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
However, unboxed tuples count as size zero. I found occasions where we had
        f x y z = case op# x y z of { s -> (# s, () #) }
and f wasn't getting inlined.

I tried giving unboxed tuples a *result discount* of zero (see the
commented-out line).  Why?  When returned as a result they do not
allocate, so maybe we don't want to charge so much for them. If you
have a non-zero discount here, we find that workers often get inlined
back into wrappers, because it look like
    f x = case $wf x of (# a,b #) -> (a,b)
and we are keener because of the case.  However while this change
shrank binary sizes by 0.5% it also made spectral/boyer allocate 5%
more. All other changes were very small. So it's not a big deal but I
didn't adopt the idea.

When fixing #18282 (see Note [Constructor size and result discount])
I changed the result discount to be just 10, not 10*(1+n_val_args).

Note [Function and non-function discounts]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We want a discount if the function is applied. A good example is
monadic combinators with continuation arguments, where inlining is
quite important.

But we don't want a big discount when a function is called many times
(see the detailed comments with #6048) because if the function is
big it won't be inlined at its many call sites and no benefit results.
Indeed, we can get exponentially big inlinings this way; that is what
#6048 is about.

On the other hand, for data-valued arguments, if there are lots of
case expressions in the body, each one will get smaller if we apply
the function to a constructor application, so we *want* a big discount
if the argument is scrutinised by many case expressions.

Conclusion:
  - For functions, take the max of the discounts
  - For data values, take the sum of the discounts


Note [Literal integer size]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Literal integers *can* be big (mkInteger [...coefficients...]), but
need not be (IS n).  We just use an arbitrary big-ish constant here
so that, in particular, we don't inline top-level defns like
   n = IS 5
There's no point in doing so -- any optimisations will see the IS
through n's unfolding.  Nor will a big size inhibit unfoldings functions
that mention a literal Integer, because the float-out pass will float
all those constants to top level.
-}

primOpSize :: PrimOp -> Int -> ExprSize
primOpSize :: PrimOp -> Int -> ExprSize
primOpSize PrimOp
op Int
n_val_args
 = if PrimOp -> Bool
primOpOutOfLine PrimOp
op
      then Int -> ExprSize
sizeN (Int
op_size Int -> Int -> Int
forall a. Num a => a -> a -> a
+ Int
n_val_args)
      else Int -> ExprSize
sizeN Int
op_size
 where
   op_size :: Int
op_size = PrimOp -> Int
primOpCodeSize PrimOp
op


buildSize :: ExprSize
buildSize :: ExprSize
buildSize = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
0 Bag (Id, Int)
forall a. Bag a
emptyBag Int
40
        -- We really want to inline applications of build
        -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
        -- Indeed, we should add a result_discount because build is
        -- very like a constructor.  We don't bother to check that the
        -- build is saturated (it usually is).  The "-2" discounts for the \c n,
        -- The "4" is rather arbitrary.

augmentSize :: ExprSize
augmentSize :: ExprSize
augmentSize = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
0 Bag (Id, Int)
forall a. Bag a
emptyBag Int
40
        -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
        -- e plus ys. The -2 accounts for the \cn

-- When we return a lambda, give a discount if it's used (applied)
lamScrutDiscount :: UnfoldingOpts -> ExprSize -> ExprSize
lamScrutDiscount :: UnfoldingOpts -> ExprSize -> ExprSize
lamScrutDiscount UnfoldingOpts
opts (SizeIs Int
n Bag (Id, Int)
vs Int
_) = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
n Bag (Id, Int)
vs (UnfoldingOpts -> Int
unfoldingFunAppDiscount UnfoldingOpts
opts)
lamScrutDiscount UnfoldingOpts
_      ExprSize
TooBig          = ExprSize
TooBig

{-
Note [addAltSize result discounts]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When adding the size of alternatives, we *add* the result discounts
too, rather than take the *maximum*.  For a multi-branch case, this
gives a discount for each branch that returns a constructor, making us
keener to inline.  I did try using 'max' instead, but it makes nofib
'rewrite' and 'puzzle' allocate significantly more, and didn't make
binary sizes shrink significantly either.

Note [Discounts and thresholds]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Constants for discounts and thresholds are defined in 'UnfoldingOpts'. They are:

unfoldingCreationThreshold
     At a definition site, if the unfolding is bigger than this, we
     may discard it altogether

unfoldingUseThreshold
     At a call site, if the unfolding, less discounts, is smaller than
     this, then it's small enough inline

unfoldingDictDiscount
     The discount for each occurrence of a dictionary argument
     as an argument of a class method.  Should be pretty small
     else big functions may get inlined

unfoldingFunAppDiscount
     Discount for a function argument that is applied.  Quite
     large, because if we inline we avoid the higher-order call.

unfoldingVeryAggressive
     If True, the compiler ignores all the thresholds and inlines very
     aggressively. It still adheres to arity, simplifier phase control and
     loop breakers.


Historical Note: Before April 2020 we had another factor,
ufKeenessFactor, which would scale the discounts before they were subtracted
from the size. This was justified with the following comment:

  -- We multiply the raw discounts (args_discount and result_discount)
  -- ty opt_UnfoldingKeenessFactor because the former have to do with
  --  *size* whereas the discounts imply that there's some extra
  --  *efficiency* to be gained (e.g. beta reductions, case reductions)
  -- by inlining.

However, this is highly suspect since it means that we subtract a *scaled* size
from an absolute size, resulting in crazy (e.g. negative) scores in some cases
(#15304). We consequently killed off ufKeenessFactor and bumped up the
ufUseThreshold to compensate.


Note [Function applications]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In a function application (f a b)

  - If 'f' is an argument to the function being analysed,
    and there's at least one value arg, record a FunAppDiscount for f

  - If the application if a PAP (arity > 2 in this example)
    record a *result* discount (because inlining
    with "extra" args in the call may mean that we now
    get a saturated application)

Code for manipulating sizes
-}

-- | The size of a candidate expression for unfolding
data ExprSize
    = TooBig
    | SizeIs { ExprSize -> Int
_es_size_is  :: {-# UNPACK #-} !Int -- ^ Size found
             , ExprSize -> Bag (Id, Int)
_es_args     :: !(Bag (Id,Int))
               -- ^ Arguments cased herein, and discount for each such
             , ExprSize -> Int
_es_discount :: {-# UNPACK #-} !Int
               -- ^ Size to subtract if result is scrutinised by a case
               -- expression
             }

instance Outputable ExprSize where
  ppr :: ExprSize -> SDoc
ppr ExprSize
TooBig         = String -> SDoc
forall doc. IsLine doc => String -> doc
text String
"TooBig"
  ppr (SizeIs Int
a Bag (Id, Int)
_ Int
c) = SDoc -> SDoc
forall doc. IsLine doc => doc -> doc
brackets (Int -> SDoc
forall doc. IsLine doc => Int -> doc
int Int
a SDoc -> SDoc -> SDoc
forall doc. IsLine doc => doc -> doc -> doc
<+> Int -> SDoc
forall doc. IsLine doc => Int -> doc
int Int
c)

-- subtract the discount before deciding whether to bale out. eg. we
-- want to inline a large constructor application into a selector:
--      tup = (a_1, ..., a_99)
--      x = case tup of ...
--
mkSizeIs :: Int -> Int -> Bag (Id, Int) -> Int -> ExprSize
mkSizeIs :: Int -> Int -> Bag (Id, Int) -> Int -> ExprSize
mkSizeIs Int
max Int
n Bag (Id, Int)
xs Int
d | (Int
n Int -> Int -> Int
forall a. Num a => a -> a -> a
- Int
d) Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
max = ExprSize
TooBig
                    | Bool
otherwise     = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
n Bag (Id, Int)
xs Int
d

maxSize :: ExprSize -> ExprSize -> ExprSize
maxSize :: ExprSize -> ExprSize -> ExprSize
maxSize ExprSize
TooBig         ExprSize
_                                  = ExprSize
TooBig
maxSize ExprSize
_              ExprSize
TooBig                             = ExprSize
TooBig
maxSize s1 :: ExprSize
s1@(SizeIs Int
n1 Bag (Id, Int)
_ Int
_) s2 :: ExprSize
s2@(SizeIs Int
n2 Bag (Id, Int)
_ Int
_) | Int
n1 Int -> Int -> Bool
forall a. Ord a => a -> a -> Bool
> Int
n2   = ExprSize
s1
                                              | Bool
otherwise = ExprSize
s2

sizeZero :: ExprSize
sizeN :: Int -> ExprSize

sizeZero :: ExprSize
sizeZero = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
0 Bag (Id, Int)
forall a. Bag a
emptyBag Int
0
sizeN :: Int -> ExprSize
sizeN Int
n  = Int -> Bag (Id, Int) -> Int -> ExprSize
SizeIs Int
n Bag (Id, Int)
forall a. Bag a
emptyBag Int
0