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 CoreUtils ( exprIsValue, exprIsCheap, exprIsBottom, exprIsTrivial )
-import Id ( Id, idType, idFlavour, isId, idWorkerInfo,
+import CoreUtils ( exprIsValue, exprIsCheap, exprIsTrivial )
+import Id ( Id, idType, idFlavour, isId,
idSpecialisation, idInlinePragma, idUnfolding,
isPrimOpId_maybe
)
import VarSet
import Literal ( isLitLitLit, litIsDupable )
import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
-import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..),
- insideLam, workerExists, isNeverInlinePrag
+import IdInfo ( InlinePragInfo(..), OccInfo(..), IdFlavour(..),
+ isNeverInlinePrag
)
-import Type ( splitFunTy_maybe, isUnLiftedType )
-import Unique ( Unique, buildIdKey, augmentIdKey, hasKey )
+import Type ( isUnLiftedType )
+import PrelNames ( hasKey, buildIdKey, augmentIdKey )
import Bag
+import FastTypes
import Outputable
#if __GLASGOW_HASKELL__ >= 404
mkUnfolding top_lvl expr
= CoreUnfolding (occurAnalyseGlobalExpr expr)
top_lvl
- (exprIsCheap expr)
(exprIsValue expr)
- (exprIsBottom 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
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) | litIsDupable lit = sizeOne
| otherwise = sizeN opt_UF_DearOp -- For lack of anything better
size_up (Case (Var v) _ alts)
| v `elem` top_args -- We are scrutinising an argument variable
- = case alts of
+ =
+{- 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
-- 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
+ 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
-- 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
------------
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
-- 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}
certainlyWillInline v
= case idUnfolding v of
- CoreUnfolding _ _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
+ CoreUnfolding _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
-> is_value
&& size - (n_vals +1) <= opt_UF_UseThreshold
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 cs -> Nothing ;
-- 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
| from_INLINE -> has_rules -- Black list until final phase
| otherwise -> True -- Always blacklisted
- IMustNotBeINLINEd from_inline (Just threshold)
- | from_inline -> (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