import CmdLineOpts ( opt_UF_CreationThreshold,
opt_UF_UseThreshold,
- opt_UF_ScrutConDiscount,
opt_UF_FunAppDiscount,
- opt_UF_PrimArgDiscount,
- opt_UF_KeenessFactor,
+ opt_UF_KeenessFactor,
opt_UF_CheapOp, opt_UF_DearOp,
opt_UnfoldCasms, opt_PprStyle_Debug,
opt_D_dump_inlinings
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 Id ( Id, idType, idFlavour, isId, idWorkerInfo,
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 IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), IdFlavour(..), CprInfo(..),
+ insideLam, workerExists, isNeverInlinePrag
+ )
import Type ( splitFunTy_maybe, isUnLiftedType )
-import Unique ( Unique, buildIdKey, augmentIdKey )
-import Maybes ( maybeToBool )
+import PrelNames ( hasKey, buildIdKey, augmentIdKey )
import Bag
-import List ( maximumBy )
-import Util ( isIn, lengthExceeds )
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
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) o 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
-- 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
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
callSiteInline 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,
#ifdef DEBUG
if opt_D_dump_inlinings 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))