\begin{code}
module CoreUnfold (
- Unfolding, UnfoldingGuidance, -- types
+ Unfolding, UnfoldingGuidance, -- Abstract types
noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
mkOtherCon, otherCons,
unfoldingTemplate, maybeUnfoldingTemplate,
isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
- hasUnfolding, hasSomeUnfolding,
+ hasUnfolding, hasSomeUnfolding, neverUnfold,
couldBeSmallEnoughToInline,
certainlyWillInline,
import CmdLineOpts ( opt_UF_CreationThreshold,
opt_UF_UseThreshold,
- opt_UF_ScrutConDiscount,
opt_UF_FunAppDiscount,
- opt_UF_PrimArgDiscount,
- opt_UF_KeenessFactor,
- opt_UF_CheapOp, opt_UF_DearOp,
- opt_UnfoldCasms, opt_PprStyle_Debug,
- opt_D_dump_inlinings
+ opt_UF_KeenessFactor,
+ opt_UF_DearOp, opt_UnfoldCasms,
+ DynFlags, DynFlag(..), dopt
)
import CoreSyn
import PprCore ( pprCoreExpr )
import OccurAnal ( occurAnalyseGlobalExpr )
-import BinderInfo ( )
-import CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
-import Id ( Id, idType, idFlavour, idUnique, isId, idWorkerInfo,
+import CoreUtils ( exprIsValue, exprIsCheap, exprIsTrivial )
+import Id ( Id, idType, idFlavour, isId,
idSpecialisation, idInlinePragma, idUnfolding,
isPrimOpId_maybe
)
import VarSet
-import Name ( isLocallyDefined )
-import Literal ( isLitLitLit )
+import Literal ( isLitLitLit, litIsDupable )
import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
-import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..), insideLam, workerExists )
-import TyCon ( tyConFamilySize )
-import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
-import Unique ( Unique, buildIdKey, augmentIdKey )
-import Maybes ( maybeToBool )
+import IdInfo ( InlinePragInfo(..), OccInfo(..), IdFlavour(..),
+ isNeverInlinePrag
+ )
+import Type ( isUnLiftedType )
+import PrelNames ( hasKey, buildIdKey, augmentIdKey )
import Bag
-import List ( maximumBy )
-import Util ( isIn, lengthExceeds )
+import FastTypes
import Outputable
#if __GLASGOW_HASKELL__ >= 404
#endif
\end{code}
+
%************************************************************************
%* *
-\subsection{@Unfolding@ and @UnfoldingGuidance@ types}
+\subsection{Making unfoldings}
%* *
%************************************************************************
\begin{code}
-data Unfolding
- = NoUnfolding
-
- | OtherCon [AltCon] -- It ain't one of these
- -- (OtherCon xs) also indicates that something has been evaluated
- -- and hence there's no point in re-evaluating it.
- -- OtherCon [] is used even for non-data-type values
- -- to indicated evaluated-ness. Notably:
- -- data C = C !(Int -> Int)
- -- case x of { C f -> ... }
- -- Here, f gets an OtherCon [] unfolding.
-
- | CompulsoryUnfolding CoreExpr -- There is no "original" definition,
- -- so you'd better unfold.
-
- | CoreUnfolding -- An unfolding with redundant cached information
- CoreExpr -- Template; binder-info is correct
- Bool -- This is a top-level binding
- Bool -- exprIsCheap template (cached); it won't duplicate (much) work
- -- if you inline this in more than one place
- Bool -- exprIsValue template (cached); it is ok to discard a `seq` on
- -- this variable
- Bool -- exprIsBottom template (cached)
- UnfoldingGuidance -- Tells about the *size* of the template.
-
-seqUnfolding :: Unfolding -> ()
-seqUnfolding (CoreUnfolding e top b1 b2 b3 g)
- = seqExpr e `seq` top `seq` b1 `seq` b2 `seq` b3 `seq` seqGuidance g
-seqUnfolding other = ()
-\end{code}
-
-\begin{code}
-noUnfolding = NoUnfolding
-mkOtherCon = OtherCon
+mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
-mkTopUnfolding cpr_info expr = mkUnfolding True {- Top level -} cpr_info expr
-
-mkUnfolding top_lvl cpr_info expr
+mkUnfolding top_lvl expr
= CoreUnfolding (occurAnalyseGlobalExpr expr)
top_lvl
- (exprIsCheap expr)
(exprIsValue expr)
- (exprIsBottom expr)
- (calcUnfoldingGuidance opt_UF_CreationThreshold cpr_info expr)
+ -- Already evaluated
+
+ (exprIsCheap expr)
+ -- OK to inline inside a lambda
+
+ (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
-- 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
mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
= CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
+\end{code}
-unfoldingTemplate :: Unfolding -> CoreExpr
-unfoldingTemplate (CoreUnfolding expr _ _ _ _ _) = expr
-unfoldingTemplate (CompulsoryUnfolding expr) = expr
-unfoldingTemplate other = panic "getUnfoldingTemplate"
-
-maybeUnfoldingTemplate :: Unfolding -> Maybe CoreExpr
-maybeUnfoldingTemplate (CoreUnfolding expr _ _ _ _ _) = Just expr
-maybeUnfoldingTemplate (CompulsoryUnfolding expr) = Just expr
-maybeUnfoldingTemplate other = Nothing
-
-otherCons (OtherCon cons) = cons
-otherCons other = []
-
-isValueUnfolding :: Unfolding -> Bool
- -- Returns False for OtherCon
-isValueUnfolding (CoreUnfolding _ _ _ is_evald _ _) = is_evald
-isValueUnfolding other = False
-
-isEvaldUnfolding :: Unfolding -> Bool
- -- Returns True for OtherCon
-isEvaldUnfolding (OtherCon _) = True
-isEvaldUnfolding (CoreUnfolding _ _ _ is_evald _ _) = is_evald
-isEvaldUnfolding other = False
-
-isCheapUnfolding :: Unfolding -> Bool
-isCheapUnfolding (CoreUnfolding _ _ is_cheap _ _ _) = is_cheap
-isCheapUnfolding other = False
-
-isCompulsoryUnfolding :: Unfolding -> Bool
-isCompulsoryUnfolding (CompulsoryUnfolding _) = True
-isCompulsoryUnfolding other = False
-
-hasUnfolding :: Unfolding -> Bool
-hasUnfolding (CoreUnfolding _ _ _ _ _ _) = True
-hasUnfolding (CompulsoryUnfolding _) = True
-hasUnfolding other = False
-
-hasSomeUnfolding :: Unfolding -> Bool
-hasSomeUnfolding NoUnfolding = False
-hasSomeUnfolding other = True
-
-data UnfoldingGuidance
- = UnfoldNever
- | UnfoldIfGoodArgs Int -- and "n" value args
-
- [Int] -- Discount if the argument is evaluated.
- -- (i.e., a simplification will definitely
- -- be possible). One elt of the list per *value* arg.
-
- Int -- The "size" of the unfolding; to be elaborated
- -- later. ToDo
-
- Int -- Scrutinee discount: the discount to substract if the thing is in
- -- a context (case (thing args) of ...),
- -- (where there are the right number of arguments.)
-seqGuidance (UnfoldIfGoodArgs n ns a b) = n `seq` sum ns `seq` a `seq` b `seq` ()
-seqGuidance other = ()
-\end{code}
+%************************************************************************
+%* *
+\subsection{The UnfoldingGuidance type}
+%* *
+%************************************************************************
\begin{code}
instance Outputable UnfoldingGuidance where
\end{code}
-%************************************************************************
-%* *
-\subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
-%* *
-%************************************************************************
-
\begin{code}
calcUnfoldingGuidance
:: Int -- bomb out if size gets bigger than this
- -> CprInfo -- CPR info for this RHS
-> CoreExpr -- expression to look at
-> UnfoldingGuidance
-calcUnfoldingGuidance bOMB_OUT_SIZE cpr_info expr
+calcUnfoldingGuidance bOMB_OUT_SIZE expr
= case collect_val_bndrs expr of { (inline, val_binders, body) ->
let
n_val_binders = length val_binders
-- so that INLINE things don't get inlined into entirely boring contexts,
-- but no more.
--- Experimental thing commented in for now
--- max_inline_size = case cpr_info of
--- NoCPRInfo -> n_val_binders + 2
--- ReturnsCPR -> n_val_binders + 1
-
- -- However, the wrapper for a CPR'd function is particularly good to inline,
- -- even in a boring context, because we may get to do update in place:
- -- let x = case y of { I# y# -> I# (y# +# 1#) }
- -- Hence the case on cpr_info
-
in
case (sizeExpr bOMB_OUT_SIZE val_binders body) of
n_val_binders
(map discount_for val_binders)
final_size
- (I# scrut_discount)
+ (iBox scrut_discount)
where
- boxed_size = I# size
+ boxed_size = iBox size
final_size | inline = boxed_size `min` max_inline_size
| otherwise = boxed_size
-> CoreExpr
-> ExprSize
-sizeExpr (I# bOMB_OUT_SIZE) top_args expr
+sizeExpr bOMB_OUT_SIZE top_args expr
= size_up expr
where
size_up (Type t) = sizeZero -- Types cost nothing
size_up (Note _ body) = size_up body -- Notes cost nothing
- size_up (App fun (Type t)) = size_up fun
- size_up (App fun arg) = size_up_app fun [arg]
+ size_up (App fun (Type t)) = size_up fun
+ size_up (App fun arg) = size_up_app fun [arg]
- size_up (Lit lit) = sizeOne
+ size_up (Lit lit) | litIsDupable lit = sizeOne
+ | otherwise = sizeN opt_UF_DearOp -- For lack of anything better
size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
| otherwise = size_up e
where
rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
- -- We want to make wrapper-style evaluation look cheap, so that
- -- when we inline a wrapper it doesn't make call site (much) bigger
- -- Otherwise we get nasty phase ordering stuff:
- -- f x = g x x
- -- h y = ...(f e)...
- -- If we inline g's wrapper, f looks big, and doesn't get inlined
- -- into h; if we inline f first, while it looks small, then g's
- -- wrapper will get inlined later anyway. To avoid this nasty
- -- ordering difference, we make (case a of (x,y) -> ...) look free.
- size_up (Case (Var v) _ [alt])
- | v `elem` top_args
- = size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
+ size_up (Case (Var v) _ alts)
+ | v `elem` top_args -- We are scrutinising an argument variable
+ =
+{- I'm nuking this special case; BUT see the comment with case alternatives.
+
+ (a) It's too eager. We don't want to inline a wrapper into a
+ context with no benefit.
+ E.g. \ x. f (x+x) no point in inlining (+) here!
+
+ (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
+ aren't scrutinising arguments any more
+
+ case alts of
+
+ [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
+ -- We want to make wrapper-style evaluation look cheap, so that
+ -- when we inline a wrapper it doesn't make call site (much) bigger
+ -- Otherwise we get nasty phase ordering stuff:
+ -- f x = g x x
+ -- h y = ...(f e)...
+ -- If we inline g's wrapper, f looks big, and doesn't get inlined
+ -- into h; if we inline f first, while it looks small, then g's
+ -- wrapper will get inlined later anyway. To avoid this nasty
+ -- ordering difference, we make (case a of (x,y) -> ...),
+ -- *where a is one of the arguments* look free.
+
+ other ->
+-}
+ alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
+ (foldr1 maxSize alt_sizes)
+
-- Good to inline if an arg is scrutinised, because
-- that may eliminate allocation in the caller
-- And it eliminates the case itself
- | otherwise
- = size_up_alt alt
-
- -- Scrutinising one of the argument variables,
- -- with more than one alternative
- size_up (Case (Var v) _ alts)
- | v `elem` top_args
- = alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
- (foldr1 maxSize alt_sizes)
+
where
- v_in_args = v `elem` top_args
alt_sizes = map size_up_alt alts
+ -- alts_size tries to compute a good discount for
+ -- the case when we are scrutinising an argument variable
alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
(SizeIs max max_disc max_scrut) -- Size of biggest alternative
- = SizeIs tot (unitBag (v, I# (1# +# tot -# max)) `unionBags` max_disc) max_scrut
+ = SizeIs tot (unitBag (v, iBox (_ILIT 1 +# tot -# max)) `unionBags` max_disc) max_scrut
-- If the variable is known, we produce a discount that
-- will take us back to 'max', the size of rh largest alternative
-- The 1+ is a little discount for reduced allocation in the caller
-
alts_size tot_size _ = tot_size
-- Also if the function is a constant Id (constr or primop)
-- compute discounts specially
size_up_fun (Var fun) args
- | idUnique fun == buildIdKey = buildSize
- | idUnique fun == augmentIdKey = augmentSize
+ | fun `hasKey` buildIdKey = buildSize
+ | fun `hasKey` augmentIdKey = augmentSize
| otherwise
= case idFlavour fun of
DataConId dc -> conSizeN (valArgCount args)
------------
size_up_alt (con, bndrs, rhs) = size_up rhs
- -- Don't charge for args, so that wrappers look cheap
+ -- Don't charge for args, so that wrappers look cheap
+ -- (See comments about wrappers with Case)
------------
-- We want to record if we're case'ing, or applying, an argument
fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
- fun_discount other = sizeZero
+ fun_discount other = sizeZero
------------
-- These addSize things have to be here because
-- I don't want to give them bOMB_OUT_SIZE as an argument
addSizeN TooBig _ = TooBig
- addSizeN (SizeIs n xs d) (I# m)
- | n_tot ># bOMB_OUT_SIZE = TooBig
+ addSizeN (SizeIs n xs d) m
+ | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig
| otherwise = SizeIs n_tot xs d
where
- n_tot = n +# m
+ n_tot = n +# iUnbox m
addSize TooBig _ = TooBig
addSize _ TooBig = TooBig
addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
- | n_tot ># bOMB_OUT_SIZE = TooBig
+ | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig
| otherwise = SizeIs n_tot xys d_tot
where
n_tot = n1 +# n2
\begin{code}
data ExprSize = TooBig
- | SizeIs Int# -- Size found
+ | SizeIs FastInt -- Size found
(Bag (Id,Int)) -- Arguments cased herein, and discount for each such
- Int# -- Size to subtract if result is scrutinised
+ FastInt -- Size to subtract if result is scrutinised
-- by a case expression
-isTooBig TooBig = True
-isTooBig _ = False
maxSize TooBig _ = TooBig
maxSize _ TooBig = TooBig
maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
| otherwise = s2
-sizeZero = SizeIs 0# emptyBag 0#
-sizeOne = SizeIs 1# emptyBag 0#
-sizeTwo = SizeIs 2# emptyBag 0#
-sizeN (I# n) = SizeIs n emptyBag 0#
-conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)
+sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
+sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
+sizeTwo = SizeIs (_ILIT 2) emptyBag (_ILIT 0)
+sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
+conSizeN n = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1)
-- Treat constructors as size 1; 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 (I# x) has size 1,
+ -- them as size zero, else we find that (iBox x) has size 1,
-- which is the same as a lone variable; and hence 'v' will
- -- always be replaced by (I# x), where v is bound to I# x.
+ -- always be replaced by (iBox x), where v is bound to iBox x.
primOpSize op n_args
| not (primOpIsDupable op) = sizeN opt_UF_DearOp
- | not (primOpOutOfLine op) = sizeZero -- These are good to inline
+ | not (primOpOutOfLine op) = sizeN (1 - n_args)
+ -- Be very keen to inline simple primops.
+ -- We give a discount of 1 for each arg so that (op# x y z) costs 1.
+ -- I found occasions where we had
+ -- f x y z = case op# x y z of { s -> (# s, () #) }
+ -- and f wasn't getting inlined
| otherwise = sizeOne
buildSize = SizeIs (-2#) emptyBag 4#
nukeScrutDiscount TooBig = TooBig
-- When we return a lambda, give a discount if it's used (applied)
-lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { I# d -> SizeIs n vs d }
+lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
lamScrutDiscount TooBig = TooBig
\end{code}
\begin{code}
couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
-couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold NoCPRInfo rhs of
+couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
UnfoldNever -> False
other -> True
certainlyWillInline v
= case idUnfolding v of
- CoreUnfolding _ _ _ is_value _ (UnfoldIfGoodArgs n_vals _ size _)
+ CoreUnfolding _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
-> is_value
&& size - (n_vals +1) <= opt_UF_UseThreshold
- && not never_inline
other -> False
- where
- never_inline = case idInlinePragma v of
- IMustNotBeINLINEd False Nothing -> True
- other -> False
\end{code}
@okToUnfoldInHifile@ is used when emitting unfolding info into an interface
go (App fun arg) = go fun && go arg
go (Lam _ body) = go body
go (Let binds body) = and (map go (body :rhssOfBind binds))
- go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))
+ go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
+ not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ])
go (Note _ body) = go body
go (Type _) = True
StrictAnal.addStrictnessInfoToTopId
\begin{code}
-callSiteInline :: Bool -- True <=> the Id is black listed
+callSiteInline :: DynFlags
+ -> Bool -- True <=> the Id is black listed
-> Bool -- 'inline' note at call site
-> OccInfo
-> Id -- The Id
-> Maybe CoreExpr -- Unfolding, if any
-callSiteInline black_listed inline_call occ id arg_infos interesting_cont
+callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont
= case idUnfolding id of {
NoUnfolding -> Nothing ;
- OtherCon _ -> Nothing ;
+ OtherCon cs -> Nothing ;
CompulsoryUnfolding unf_template | black_listed -> Nothing
| otherwise -> Just unf_template ;
-- Constructors have compulsory unfoldings, but
-- may have rules, in which case they are
-- black listed till later
- CoreUnfolding unf_template is_top is_cheap is_value is_bot guidance ->
+ CoreUnfolding unf_template is_top is_value is_cheap guidance ->
let
result | yes_or_no = Just unf_template
n_val_args = length arg_infos
- ok_inside_lam = is_value || is_bot || (is_cheap && not is_top)
- -- I'm experimenting with is_cheap && not is_top
-
yes_or_no
| black_listed = False
| otherwise = case occ of
IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
IAmALoopBreaker -> False
- OneOcc in_lam one_br -> (not in_lam || ok_inside_lam) && consider_safe in_lam True one_br
- NoOccInfo -> ok_inside_lam && consider_safe True False False
+ OneOcc in_lam one_br -> (not in_lam || is_cheap) && consider_safe in_lam True one_br
+ NoOccInfo -> is_cheap && consider_safe True False False
consider_safe in_lam once once_in_one_branch
-- consider_safe decides whether it's a good idea to inline something,
in
#ifdef DEBUG
- if opt_D_dump_inlinings then
+ if dopt Opt_D_dump_inlinings dflags then
pprTrace "Considering inlining"
- (ppr id <+> vcat [text "black listed" <+> ppr black_listed,
+ (ppr id <+> vcat [text "black listed:" <+> ppr black_listed,
text "occ info:" <+> ppr occ,
text "arg infos" <+> ppr arg_infos,
text "interesting continuation" <+> ppr interesting_cont,
text "is value:" <+> ppr is_value,
text "is cheap:" <+> ppr is_cheap,
- text "is bottom:" <+> ppr is_bot,
- text "is top-level:" <+> ppr is_top,
text "guidance" <+> ppr guidance,
text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
if yes_or_no then
in that order. The meanings of these are determined by the @blackListed@ function
here.
-The final simplification doesn't have a phase number
+The final simplification doesn't have a phase number.
Pragmas
~~~~~~~
-- place that the inline phase number is looked at.
blackListed rule_vars Nothing -- Last phase
- = \v -> case idInlinePragma v of
- IMustNotBeINLINEd False Nothing -> True -- An unconditional NOINLINE pragma
- other -> False
+ = \v -> isNeverInlinePrag (idInlinePragma v)
blackListed rule_vars (Just phase)
= \v -> normal_case rule_vars phase v
| from_INLINE -> has_rules -- Black list until final phase
| otherwise -> True -- Always blacklisted
- IMustNotBeINLINEd from_inline (Just threshold)
- | from_inline -> phase < threshold && has_rules
- | otherwise -> phase < threshold || has_rules
+ IMustNotBeINLINEd from_INLINE (Just threshold)
+ | from_INLINE -> (phase < threshold && has_rules)
+ | otherwise -> (phase < threshold || has_rules)
where
has_rules = v `elemVarSet` rule_vars
|| not (isEmptyCoreRules (idSpecialisation v))