%
-% (c) The AQUA Project, Glasgow University, 1994-1996
+% (c) The AQUA Project, Glasgow University, 1994-1998
%
\section[CoreUnfold]{Core-syntax unfoldings}
Unfoldings (which can travel across module boundaries) are in Core
syntax (namely @CoreExpr@s).
-The type @UnfoldingDetails@ sits ``above'' simply-Core-expressions
+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 @GenForm@ unfolding, you will
+literal''). In the corner of a @CoreUnfolding@ unfolding, you will
find, unsurprisingly, a Core expression.
\begin{code}
-#include "HsVersions.h"
-
module CoreUnfold (
- UnfoldingDetails(..), UnfoldingGuidance(..), -- types
- FormSummary(..),
-
- mkFormSummary,
- mkGenForm, mkLitForm, mkConForm,
- whnfDetails,
- mkMagicUnfolding,
- calcUnfoldingGuidance,
- mentionedInUnfolding
+ Unfolding, UnfoldingGuidance, -- Abstract types
+
+ noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
+ evaldUnfolding, mkOtherCon, otherCons,
+ unfoldingTemplate, maybeUnfoldingTemplate,
+ isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
+ hasUnfolding, hasSomeUnfolding, neverUnfold,
+
+ couldBeSmallEnoughToInline,
+ certainlyWillInline, smallEnoughToInline,
+
+ callSiteInline
) where
-IMP_Ubiq()
-IMPORT_DELOOPER(IdLoop) -- for paranoia checking;
- -- and also to get mkMagicUnfoldingFun
-IMPORT_DELOOPER(PrelLoop) -- for paranoia checking
+#include "HsVersions.h"
-import Bag ( emptyBag, unitBag, unionBags, Bag )
-import BinderInfo ( oneTextualOcc, oneSafeOcc )
-import CgCompInfo ( uNFOLDING_CHEAP_OP_COST,
- uNFOLDING_DEAR_OP_COST,
- uNFOLDING_NOREP_LIT_COST
+import StaticFlags ( opt_UF_CreationThreshold, opt_UF_UseThreshold,
+ opt_UF_FunAppDiscount, opt_UF_KeenessFactor,
+ opt_UF_DearOp,
)
+import DynFlags ( DynFlags, DynFlag(..), dopt )
import CoreSyn
-import CoreUtils ( coreExprType, manifestlyWHNF )
-import CostCentre ( ccMentionsId )
-import Id ( SYN_IE(IdSet), GenId{-instances-} )
-import IdInfo ( bottomIsGuaranteed )
-import Literal ( isNoRepLit, isLitLitLit )
-import Pretty
-import PrimOp ( primOpCanTriggerGC, PrimOp(..) )
-import TyCon ( tyConFamilySize )
-import Type ( getAppDataTyConExpandingDicts )
-import UniqSet ( emptyUniqSet, unitUniqSet, mkUniqSet,
- addOneToUniqSet, unionUniqSets
+import PprCore ( pprCoreExpr )
+import OccurAnal ( occurAnalyseExpr )
+import CoreUtils ( exprIsHNF, exprIsCheap, exprIsTrivial )
+import Id ( Id, idType, isId,
+ idUnfolding, globalIdDetails
)
-import Usage ( SYN_IE(UVar) )
-import Util ( isIn, panic )
-
-whatsMentionedInId = panic "whatsMentionedInId (CoreUnfold)"
-getMentionedTyConsAndClassesFromType = panic "getMentionedTyConsAndClassesFromType (CoreUnfold)"
+import DataCon ( isUnboxedTupleCon )
+import Literal ( litSize )
+import PrimOp ( primOpIsDupable, primOpOutOfLine )
+import IdInfo ( OccInfo(..), GlobalIdDetails(..) )
+import Type ( isUnLiftedType )
+import PrelNames ( hasKey, buildIdKey, augmentIdKey )
+import Bag
+import FastTypes
+import Outputable
+
+#if __GLASGOW_HASKELL__ >= 404
+import GLAEXTS ( Int# )
+#endif
\end{code}
+
%************************************************************************
%* *
-\subsection{@UnfoldingDetails@ and @UnfoldingGuidance@ types}
+\subsection{Making unfoldings}
%* *
%************************************************************************
-(And @FormSummary@, too.)
-
\begin{code}
-data UnfoldingDetails
- = NoUnfoldingDetails
-
- | OtherLitForm
- [Literal] -- It is a literal, but definitely not one of these
-
- | OtherConForm
- [Id] -- It definitely isn't one of these constructors
- -- This captures the situation in the default branch of
- -- a case: case x of
- -- c1 ... -> ...
- -- c2 ... -> ...
- -- v -> default-rhs
- -- Then in default-rhs we know that v isn't c1 or c2.
- --
- -- NB. In the degenerate: case x of {v -> default-rhs}
- -- x will be bound to
- -- OtherConForm []
- -- which captures the idea that x is eval'd but we don't
- -- know which constructor.
-
-
- | GenForm
- FormSummary -- Tells whether the template is a WHNF or bottom
- TemplateOutExpr -- The template
- UnfoldingGuidance -- Tells about the *size* of the template.
-
- | MagicForm
- Unique -- of the Id whose magic unfolding this is
- MagicUnfoldingFun
-
-type TemplateOutExpr = GenCoreExpr (Id, BinderInfo) Id TyVar UVar
- -- An OutExpr with occurrence info attached. This is used as
- -- a template in GeneralForms.
-
-mkMagicUnfolding :: Unique -> UnfoldingDetails
-mkMagicUnfolding tag = MagicForm tag (mkMagicUnfoldingFun tag)
-
-data FormSummary
- = WhnfForm -- Expression is WHNF
- | BottomForm -- Expression is guaranteed to be bottom. We're more gung
- -- ho about inlining such things, because it can't waste work
- | OtherForm -- Anything else
-
-instance Outputable FormSummary where
- ppr sty WhnfForm = ppStr "WHNF"
- ppr sty BottomForm = ppStr "Bot"
- ppr sty OtherForm = ppStr "Other"
-
---???mkFormSummary :: StrictnessInfo -> GenCoreExpr bndr Id -> FormSummary
-mkFormSummary si expr
- | manifestlyWHNF expr = WhnfForm
- | bottomIsGuaranteed si = BottomForm
-
- -- Chances are that the Id will be decorated with strictness info
- -- telling that the RHS is definitely bottom. This *might* not be the
- -- case, if it's been a while since strictness analysis, but leaving out
- -- the test for manifestlyBottom makes things a little more efficient.
- -- We can always put it back...
- -- | manifestlyBottom expr = BottomForm
-
- | otherwise = OtherForm
-
-whnfDetails :: UnfoldingDetails -> Bool -- True => thing is evaluated
-whnfDetails (GenForm WhnfForm _ _) = True
-whnfDetails (OtherLitForm _) = True
-whnfDetails (OtherConForm _) = True
-whnfDetails other = False
-\end{code}
-
-\begin{code}
-data UnfoldingGuidance
- = UnfoldNever -- Don't do it!
-
- | UnfoldAlways -- There is no "original" definition,
- -- so you'd better unfold. Or: something
- -- so cheap to unfold (e.g., 1#) that
- -- you should do it absolutely always.
-
- | EssentialUnfolding -- Like UnfoldAlways, but you *must* do
- -- it absolutely always.
- -- This is what we use for data constructors
- -- and PrimOps, because we don't feel like
- -- generating curried versions "just in case".
-
- | UnfoldIfGoodArgs Int -- if "m" type args and "n" value args; and
- Int -- those val args are manifestly data constructors
- [Bool] -- the val-arg positions marked True
- -- (i.e., a simplification will definitely
- -- be possible).
- Int -- The "size" of the unfolding; to be elaborated
- -- later. ToDo
-
- | BadUnfolding -- This is used by TcPragmas if the *lazy*
- -- lintUnfolding test fails
- -- It will never escape from the IdInfo as
- -- it is caught by getInfo_UF and converted
- -- to NoUnfoldingDetails
-\end{code}
-
-\begin{code}
-instance Outputable UnfoldingGuidance where
- ppr sty UnfoldNever = ppStr "_N_"
- ppr sty UnfoldAlways = ppStr "_ALWAYS_"
- ppr sty EssentialUnfolding = ppStr "_ESSENTIAL_" -- shouldn't appear in an iface
- ppr sty (UnfoldIfGoodArgs t v cs size)
- = ppCat [ppStr "_IF_ARGS_", ppInt t, ppInt v,
- if null cs -- always print *something*
- then ppChar 'X'
- else ppBesides (map pp_c cs),
- ppInt size ]
- where
- pp_c False = ppChar 'X'
- pp_c True = ppChar 'C'
+mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
+
+mkUnfolding top_lvl expr
+ = CoreUnfolding (occurAnalyseExpr expr)
+ top_lvl
+
+ (exprIsHNF 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
+ -- Nevertheless, we don't occ-analyse before computing the size because the
+ -- size computation bales out after a while, whereas occurrence analysis does not.
+ --
+ -- This can occasionally mean that the guidance is very pessimistic;
+ -- it gets fixed up next round
+
+mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
+ = CompulsoryUnfolding (occurAnalyseExpr expr)
\end{code}
%************************************************************************
%* *
-\subsection{@mkGenForm@ and friends}
+\subsection{The UnfoldingGuidance type}
%* *
%************************************************************************
\begin{code}
-mkGenForm :: FormSummary
- -> TemplateOutExpr -- Template
- -> UnfoldingGuidance -- Tells about the *size* of the template.
- -> UnfoldingDetails
-
-mkGenForm = GenForm
-
--- two shorthand variants:
-mkLitForm lit = mk_go_for_it (Lit lit)
-mkConForm con args = mk_go_for_it (Con con args)
-
-mk_go_for_it expr = mkGenForm WhnfForm expr UnfoldAlways
+instance Outputable UnfoldingGuidance where
+ ppr UnfoldNever = ptext SLIT("NEVER")
+ ppr (UnfoldIfGoodArgs v cs size discount)
+ = hsep [ ptext SLIT("IF_ARGS"), int v,
+ brackets (hsep (map int cs)),
+ int size,
+ int discount ]
\end{code}
-%************************************************************************
-%* *
-\subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
-%* *
-%************************************************************************
\begin{code}
calcUnfoldingGuidance
- :: Bool -- True <=> OK if _scc_s appear in expr
- -> Int -- bomb out if size gets bigger than this
- -> CoreExpr -- expression to look at
+ :: Int -- bomb out if size gets bigger than this
+ -> CoreExpr -- expression to look at
-> UnfoldingGuidance
+calcUnfoldingGuidance bOMB_OUT_SIZE expr
+ = case collect_val_bndrs expr of { (inline, val_binders, body) ->
+ let
+ n_val_binders = length val_binders
+
+ max_inline_size = n_val_binders+2
+ -- The idea is that if there is an INLINE pragma (inline is True)
+ -- and there's a big body, we give a size of n_val_binders+2. This
+ -- This is just enough to fail the no-size-increase test in callSiteInline,
+ -- so that INLINE things don't get inlined into entirely boring contexts,
+ -- but no more.
-calcUnfoldingGuidance scc_s_OK bOMB_OUT_SIZE expr
- = let
- (use_binders, ty_binders, val_binders, body) = collectBinders expr
in
- case (sizeExpr scc_s_OK bOMB_OUT_SIZE val_binders body) of
-
- Nothing -> UnfoldNever
-
- Just (size, cased_args)
- -> let
- uf = UnfoldIfGoodArgs
- (length ty_binders)
- (length val_binders)
- [ b `is_elem` cased_args | b <- val_binders ]
- size
- in
- -- pprTrace "calcUnfold:" (ppAbove (ppr PprDebug uf) (ppr PprDebug expr))
- uf
+ case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of
+
+ TooBig
+ | not inline -> UnfoldNever
+ -- A big function with an INLINE pragma must
+ -- have an UnfoldIfGoodArgs guidance
+ | otherwise -> UnfoldIfGoodArgs n_val_binders
+ (map (const 0) val_binders)
+ max_inline_size 0
+
+ SizeIs size cased_args scrut_discount
+ -> UnfoldIfGoodArgs
+ n_val_binders
+ (map discount_for val_binders)
+ final_size
+ (iBox scrut_discount)
+ where
+ boxed_size = iBox size
+
+ final_size | inline = boxed_size `min` max_inline_size
+ | otherwise = boxed_size
+
+ -- Sometimes an INLINE thing is smaller than n_val_binders+2.
+ -- A particular case in point is a constructor, which has size 1.
+ -- We want to inline this regardless, hence the `min`
+
+ discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
+ 0 cased_args
+ }
where
- is_elem = isIn "calcUnfoldingGuidance"
+ collect_val_bndrs e = go False [] e
+ -- We need to be a bit careful about how we collect the
+ -- value binders. In ptic, if we see
+ -- __inline_me (\x y -> e)
+ -- We want to say "2 value binders". Why? So that
+ -- we take account of information given for the arguments
+
+ go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
+ go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
+ | otherwise = go inline rev_vbs e
+ go inline rev_vbs e = (inline, reverse rev_vbs, e)
\end{code}
\begin{code}
-sizeExpr :: Bool -- True <=> _scc_s OK
- -> Int -- Bomb out if it gets bigger than this
+sizeExpr :: Int# -- Bomb out if it gets bigger than this
-> [Id] -- Arguments; we're interested in which of these
-- get case'd
-> CoreExpr
- -> Maybe (Int, -- Size
- [Id] -- Subset of args which are cased
- )
+ -> ExprSize
-sizeExpr scc_s_OK bOMB_OUT_SIZE args expr
+sizeExpr bOMB_OUT_SIZE top_args expr
= size_up expr
where
- size_up (Var v) = sizeOne
- size_up (App fun arg) = size_up fun `addSize` size_up_arg arg
- size_up (Lit lit) = if isNoRepLit lit
- then sizeN uNFOLDING_NOREP_LIT_COST
- else sizeOne
+ size_up (Type t) = sizeZero -- Types cost nothing
+ size_up (Var v) = sizeOne
- size_up (SCC _ (Con _ _)) = Nothing -- **** HACK *****
- size_up (SCC lbl body)
- = if scc_s_OK then size_up body else Nothing
+ size_up (Note InlineMe body) = sizeOne -- Inline notes make it look very small
+ -- This can be important. If you have an instance decl like this:
+ -- instance Foo a => Foo [a] where
+ -- {-# INLINE op1, op2 #-}
+ -- op1 = ...
+ -- op2 = ...
+ -- then we'll get a dfun which is a pair of two INLINE lambdas
- size_up (Coerce _ _ body) = size_up body -- Coercions cost nothing
+ size_up (Note _ body) = size_up body -- Other notes cost nothing
- size_up (Con con args) = -- 1 + # of val args
- sizeN (1 + numValArgs args)
- size_up (Prim op args) = sizeN op_cost -- NB: no charge for PrimOp args
- where
- op_cost = if primOpCanTriggerGC op
- then uNFOLDING_DEAR_OP_COST
- -- these *tend* to be more expensive;
- -- number chosen to avoid unfolding (HACK)
- else uNFOLDING_CHEAP_OP_COST
-
- size_up expr@(Lam _ _)
- = let
- (uvars, tyvars, args, body) = collectBinders expr
- in
- size_up body `addSizeN` length args
+ size_up (App fun (Type t)) = size_up fun
+ size_up (App fun arg) = size_up_app fun [arg]
- size_up (Let (NonRec binder rhs) body)
- = size_up rhs
- `addSize`
- size_up body
- `addSizeN`
- 1
+ size_up (Lit lit) = sizeN (litSize lit)
- size_up (Let (Rec pairs) body)
- = foldr addSize sizeZero [size_up rhs | (_,rhs) <- pairs]
- `addSize`
- size_up body
- `addSizeN`
- length pairs
-
- size_up (Case scrut alts)
- = size_up_scrut scrut
- `addSize`
- size_up_alts (coreExprType scrut) alts
- -- We charge for the "case" itself in "size_up_alts"
+ size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
+ | otherwise = size_up e
- ------------
- size_up_arg arg = if isValArg arg then sizeOne else sizeZero{-it's free-}
-
- ------------
- size_up_alts scrut_ty (AlgAlts alts deflt)
- = foldr (addSize . size_alg_alt) (size_up_deflt deflt) alts
- `addSizeN` (tyConFamilySize tycon)
- -- NB: we charge N for an alg. "case", where N is
- -- the number of constructors in the thing being eval'd.
- -- (You'll eventually get a "discount" of N if you
- -- think the "case" is likely to go away.)
- where
- size_alg_alt (con,args,rhs) = size_up rhs
- -- Don't charge for args, so that wrappers look cheap
-
- (tycon, _, _) = --trace "CoreUnfold.getAppDataTyConExpandingDicts" $
- getAppDataTyConExpandingDicts scrut_ty
+ size_up (Let (NonRec binder rhs) body)
+ = nukeScrutDiscount (size_up rhs) `addSize`
+ size_up body `addSizeN`
+ (if isUnLiftedType (idType binder) then 0 else 1)
+ -- For the allocation
+ -- If the binder has an unlifted type there is no allocation
- size_up_alts _ (PrimAlts alts deflt)
- = foldr (addSize . size_prim_alt) (size_up_deflt deflt) alts
- -- *no charge* for a primitive "case"!
+ size_up (Let (Rec pairs) body)
+ = nukeScrutDiscount rhs_size `addSize`
+ size_up body `addSizeN`
+ length pairs -- For the allocation
where
- size_prim_alt (lit,rhs) = size_up rhs
+ rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
+
+ 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
+
+ where
+ 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, 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
+
+-- gaw 2004
+ size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize`
+ foldr (addSize . size_up_alt) sizeZero alts
+ -- We don't charge for the case itself
+ -- It's a strict thing, and the price of the call
+ -- is paid by scrut. Also consider
+ -- case f x of DEFAULT -> e
+ -- This is just ';'! Don't charge for it.
+
+ ------------
+ size_up_app (App fun arg) args
+ | isTypeArg arg = size_up_app fun args
+ | otherwise = size_up_app fun (arg:args)
+ size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
+ (size_up_fun fun args)
+ args
+
+ -- A function application with at least one value argument
+ -- so if the function is an argument give it an arg-discount
+ --
+ -- Also behave specially if the function is a build
+ --
+ -- Also if the function is a constant Id (constr or primop)
+ -- compute discounts specially
+ size_up_fun (Var fun) args
+ | fun `hasKey` buildIdKey = buildSize
+ | fun `hasKey` augmentIdKey = augmentSize
+ | otherwise
+ = case globalIdDetails fun of
+ DataConWorkId dc -> conSizeN dc (valArgCount args)
+
+ FCallId fc -> sizeN opt_UF_DearOp
+ PrimOpId op -> primOpSize op (valArgCount args)
+ -- foldr addSize (primOpSize op) (map arg_discount args)
+ -- At one time I tried giving an arg-discount if a primop
+ -- is applied to one of the function's arguments, but it's
+ -- not good. At the moment, any unlifted-type arg gets a
+ -- 'True' for 'yes I'm evald', so we collect the discount even
+ -- if we know nothing about it. And just having it in a primop
+ -- doesn't help at all if we don't know something more.
+
+ other -> fun_discount fun `addSizeN`
+ (1 + length (filter (not . exprIsTrivial) args))
+ -- The 1+ is for the function itself
+ -- Add 1 for each non-trivial arg;
+ -- the allocation cost, as in let(rec)
+ -- Slight hack here: for constructors the args are almost always
+ -- trivial; and for primops they are almost always prim typed
+ -- We should really only count for non-prim-typed args in the
+ -- general case, but that seems too much like hard work
+
+ size_up_fun other args = size_up other
+
+ ------------
+ size_up_alt (con, bndrs, rhs) = size_up rhs
+ -- Don't charge for args, so that wrappers look cheap
+ -- (See comments about wrappers with Case)
------------
- size_up_deflt NoDefault = sizeZero
- size_up_deflt (BindDefault binder rhs) = size_up rhs
+ -- 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
------------
- -- Scrutinees. There are two things going on here.
- -- First, we want to record if we're case'ing an argument
- -- Second, we want to charge nothing for the srutinee if it's just
- -- a variable. That way wrapper-like things look cheap.
- size_up_scrut (Var v) | v `is_elem` args = Just (0, [v])
- | otherwise = Just (0, [])
- size_up_scrut other = size_up other
-
- is_elem :: Id -> [Id] -> Bool
- is_elem = isIn "size_up_scrut"
-
- ------------
- sizeZero = Just (0, [])
- sizeOne = Just (1, [])
- sizeN n = Just (n, [])
- sizeVar v = Just (0, [v])
-
- addSizeN Nothing _ = Nothing
- addSizeN (Just (n, xs)) m
- | tot < bOMB_OUT_SIZE = Just (tot, xs)
- | otherwise = Nothing
- where
- tot = n+m
-
- addSize Nothing _ = Nothing
- addSize _ Nothing = Nothing
- addSize (Just (n, xs)) (Just (m, ys))
- | tot < bOMB_OUT_SIZE = Just (tot, xys)
- | otherwise = Nothing
- where
- tot = n+m
- xys = xs ++ ys
+ -- 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) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d
+
+ addSize TooBig _ = TooBig
+ addSize _ TooBig = TooBig
+ addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
+ = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2)
\end{code}
-%************************************************************************
-%* *
-\subsection[unfoldings-for-ifaces]{Processing unfoldings for interfaces}
-%* *
-%************************************************************************
-
-Of course, the main thing we do to unfoldings-for-interfaces is {\em
-print} them. But, while we're at it, we collect info about
-``mentioned'' Ids, etc., etc.---we're going to need this stuff anyway.
-
-%************************************************************************
-%* *
-\subsubsection{Monad stuff for the unfolding-generation game}
-%* *
-%************************************************************************
+Code for manipulating sizes
\begin{code}
-type UnfoldM bndr thing
- = IdSet -- in-scope Ids (passed downwards only)
- -> (bndr -> Id) -- to extract an Id from a binder (down only)
-
- -> (Bag Id, -- mentioned global vars (ditto)
- Bag TyCon, -- ditto, tycons
- Bag Class, -- ditto, classes
- Bool) -- True <=> mentions something litlit-ish
-
- -> (thing, (Bag Id, Bag TyCon, Bag Class, Bool)) -- accumulated...
-\end{code}
-
-A little stuff for in-scopery:
-\begin{code}
-no_in_scopes :: IdSet
-add1 :: IdSet -> Id -> IdSet
-add_some :: IdSet -> [Id] -> IdSet
-
-no_in_scopes = emptyUniqSet
-in_scopes `add1` x = addOneToUniqSet in_scopes x
-in_scopes `add_some` xs = in_scopes `unionUniqSets` mkUniqSet xs
-\end{code}
-
-The can-see-inside-monad functions are the usual sorts of things.
-
-\begin{code}
-thenUf :: UnfoldM bndr a -> (a -> UnfoldM bndr b) -> UnfoldM bndr b
-thenUf m k in_scopes get_id mentioneds
- = case m in_scopes get_id mentioneds of { (v, mentioneds1) ->
- k v in_scopes get_id mentioneds1 }
-
-thenUf_ :: UnfoldM bndr a -> UnfoldM bndr b -> UnfoldM bndr b
-thenUf_ m k in_scopes get_id mentioneds
- = case m in_scopes get_id mentioneds of { (_, mentioneds1) ->
- k in_scopes get_id mentioneds1 }
-
-mapUf :: (a -> UnfoldM bndr b) -> [a] -> UnfoldM bndr [b]
-mapUf f [] = returnUf []
-mapUf f (x:xs)
- = f x `thenUf` \ r ->
- mapUf f xs `thenUf` \ rs ->
- returnUf (r:rs)
-
-returnUf :: a -> UnfoldM bndr a
-returnUf v in_scopes get_id mentioneds = (v, mentioneds)
-
-addInScopesUf :: [Id] -> UnfoldM bndr a -> UnfoldM bndr a
-addInScopesUf more_in_scopes m in_scopes get_id mentioneds
- = m (in_scopes `add_some` more_in_scopes) get_id mentioneds
-
-getInScopesUf :: UnfoldM bndr IdSet
-getInScopesUf in_scopes get_id mentioneds = (in_scopes, mentioneds)
-
-extractIdsUf :: [bndr] -> UnfoldM bndr [Id]
-extractIdsUf binders in_scopes get_id mentioneds
- = (map get_id binders, mentioneds)
-
-consider_Id :: Id -> UnfoldM bndr ()
-consider_Id var in_scopes get_id (ids, tcs, clss, has_litlit)
- = let
- (ids2, tcs2, clss2) = whatsMentionedInId in_scopes var
- in
- ((), (ids `unionBags` ids2,
- tcs `unionBags` tcs2,
- clss `unionBags`clss2,
- has_litlit))
-\end{code}
-
-\begin{code}
-addToMentionedIdsUf :: Id -> UnfoldM bndr ()
-addToMentionedTyConsUf :: Bag TyCon -> UnfoldM bndr ()
-addToMentionedClassesUf :: Bag Class -> UnfoldM bndr ()
-litlit_oops :: UnfoldM bndr ()
-
-addToMentionedIdsUf add_me in_scopes get_id (ids, tcs, clss, has_litlit)
- = ((), (ids `unionBags` unitBag add_me, tcs, clss, has_litlit))
-
-addToMentionedTyConsUf add_mes in_scopes get_id (ids, tcs, clss, has_litlit)
- = ((), (ids, tcs `unionBags` add_mes, clss, has_litlit))
-
-addToMentionedClassesUf add_mes in_scopes get_id (ids, tcs, clss, has_litlit)
- = ((), (ids, tcs, clss `unionBags` add_mes, has_litlit))
-
-litlit_oops in_scopes get_id (ids, tcs, clss, _)
- = ((), (ids, tcs, clss, True))
+data ExprSize = TooBig
+ | SizeIs FastInt -- Size found
+ (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
+ FastInt -- Size to subtract if result is scrutinised
+ -- by a case expression
+
+-- 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 max n xs d | (n -# d) ># max = TooBig
+ | otherwise = SizeIs n xs d
+
+maxSize TooBig _ = TooBig
+maxSize _ TooBig = TooBig
+maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
+ | otherwise = s2
+
+sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
+sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
+sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
+conSizeN dc n
+ | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1)
+ | otherwise = 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 (iBox x) has size 1,
+ -- which is the same as a lone variable; and hence 'v' will
+ -- always be replaced by (iBox x), where v is bound to iBox x.
+ --
+ -- 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
+
+primOpSize op n_args
+ | not (primOpIsDupable op) = sizeN opt_UF_DearOp
+ | not (primOpOutOfLine op) = sizeN (2 - 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 2.
+ -- We can't make it cost 1, else we'll inline let v = (op# x y z)
+ -- at every use of v, which is excessive.
+ --
+ -- A good example is:
+ -- let x = +# p q in C {x}
+ -- Even though x get's an occurrence of 'many', its RHS looks cheap,
+ -- and there's a good chance it'll get inlined back into C's RHS. Urgh!
+ | otherwise = sizeOne
+
+buildSize = SizeIs (-2#) emptyBag 4#
+ -- 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 becuause 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 = SizeIs (-2#) emptyBag 4#
+ -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
+ -- e plus ys. The -2 accounts for the \cn
+
+nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
+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 { d -> SizeIs n vs (iUnbox d) }
+lamScrutDiscount TooBig = TooBig
\end{code}
%************************************************************************
%* *
-\subsubsection{Gathering up info for an interface-unfolding}
+\subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
%* *
%************************************************************************
-\begin{code}
-{-
-mentionedInUnfolding
- :: (bndr -> Id) -- so we can get Ids out of binders
- -> GenCoreExpr bndr Id -- input expression
- -> (Bag Id, Bag TyCon, Bag Class,
- -- what we found mentioned in the expr
- Bool -- True <=> mentions a ``litlit''-ish thing
- -- (the guy on the other side of an interface
- -- may not be able to handle it)
- )
--}
-
-mentionedInUnfolding get_id expr
- = case (ment_expr expr no_in_scopes get_id (emptyBag, emptyBag, emptyBag, False)) of
- (_, (ids_bag, tcs_bag, clss_bag, has_litlit)) ->
- (ids_bag, tcs_bag, clss_bag, has_litlit)
-\end{code}
+We have very limited information about an unfolding expression: (1)~so
+many type arguments and so many value arguments expected---for our
+purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
+a single integer. (3)~An ``argument info'' vector. For this, what we
+have at the moment is a Boolean per argument position that says, ``I
+will look with great favour on an explicit constructor in this
+position.'' (4)~The ``discount'' to subtract if the expression
+is being scrutinised.
+
+Assuming we have enough type- and value arguments (if not, we give up
+immediately), then we see if the ``discounted size'' is below some
+(semi-arbitrary) threshold. It works like this: for every argument
+position where we're looking for a constructor AND WE HAVE ONE in our
+hands, we get a (again, semi-arbitrary) discount [proportion to the
+number of constructors in the type being scrutinized].
+
+If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
+and the expression in question will evaluate to a constructor, we use
+the computed discount size *for the result only* rather than
+computing the argument discounts. Since we know the result of
+the expression is going to be taken apart, discounting its size
+is more accurate (see @sizeExpr@ above for how this discount size
+is computed).
+
+We use this one to avoid exporting inlinings that we ``couldn't possibly
+use'' on the other side. Can be overridden w/ flaggery.
+Just the same as smallEnoughToInline, except that it has no actual arguments.
\begin{code}
---ment_expr :: GenCoreExpr bndr Id -> UnfoldM bndr ()
-
-ment_expr (Var v) = consider_Id v
-ment_expr (Lit l) = consider_lit l
-
-ment_expr expr@(Lam _ _)
- = let
- (uvars, tyvars, args, body) = collectBinders expr
- in
- extractIdsUf args `thenUf` \ bs_ids ->
- addInScopesUf bs_ids (
- -- this considering is just to extract any mentioned types/classes
- mapUf consider_Id bs_ids `thenUf_`
- ment_expr body
- )
-
-ment_expr (App fun arg)
- = ment_expr fun `thenUf_`
- ment_arg arg
-
-ment_expr (Con c as)
- = consider_Id c `thenUf_`
- mapUf ment_arg as `thenUf_`
- returnUf ()
-
-ment_expr (Prim op as)
- = ment_op op `thenUf_`
- mapUf ment_arg as `thenUf_`
- returnUf ()
- where
- ment_op (CCallOp str is_asm may_gc arg_tys res_ty)
- = mapUf ment_ty arg_tys `thenUf_`
- ment_ty res_ty
- ment_op other_op = returnUf ()
-
-ment_expr (Case scrutinee alts)
- = ment_expr scrutinee `thenUf_`
- ment_alts alts
-
-ment_expr (Let (NonRec bind rhs) body)
- = ment_expr rhs `thenUf_`
- extractIdsUf [bind] `thenUf` \ bi@[bind_id] ->
- addInScopesUf bi (
- ment_expr body `thenUf_`
- consider_Id bind_id )
-
-ment_expr (Let (Rec pairs) body)
- = let
- binders = map fst pairs
- rhss = map snd pairs
- in
- extractIdsUf binders `thenUf` \ binder_ids ->
- addInScopesUf binder_ids (
- mapUf ment_expr rhss `thenUf_`
- mapUf consider_Id binder_ids `thenUf_`
- ment_expr body )
-
-ment_expr (SCC cc expr)
- = (case (ccMentionsId cc) of
- Just id -> consider_Id id
- Nothing -> returnUf ()
- )
- `thenUf_` ment_expr expr
-
-ment_expr (Coerce _ _ _) = panic "ment_expr:Coerce"
-
--------------
-ment_ty ty
- = let
- (tycons, clss) = getMentionedTyConsAndClassesFromType ty
- in
- addToMentionedTyConsUf tycons `thenUf_`
- addToMentionedClassesUf clss
-
--------------
-
-ment_alts alg_alts@(AlgAlts alts deflt)
- = mapUf ment_alt alts `thenUf_`
- ment_deflt deflt
- where
- ment_alt alt@(con, params, rhs)
- = consider_Id con `thenUf_`
- extractIdsUf params `thenUf` \ param_ids ->
- addInScopesUf param_ids (
- -- "consider" them so we can chk out their types...
- mapUf consider_Id param_ids `thenUf_`
- ment_expr rhs )
-
-ment_alts (PrimAlts alts deflt)
- = mapUf ment_alt alts `thenUf_`
- ment_deflt deflt
- where
- ment_alt alt@(lit, rhs) = ment_expr rhs
-
-----------------
-ment_deflt NoDefault
- = returnUf ()
-
-ment_deflt d@(BindDefault b rhs)
- = extractIdsUf [b] `thenUf` \ bi@[b_id] ->
- addInScopesUf bi (
- consider_Id b_id `thenUf_`
- ment_expr rhs )
-
------------
-ment_arg (VarArg v) = consider_Id v
-ment_arg (LitArg l) = consider_lit l
-ment_arg (TyArg ty) = ment_ty ty
-ment_arg (UsageArg _) = returnUf ()
-
------------
-consider_lit lit
- | isLitLitLit lit = litlit_oops `thenUf_` returnUf ()
- | otherwise = returnUf ()
+couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
+couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
+ UnfoldNever -> False
+ other -> True
+
+certainlyWillInline :: Unfolding -> Bool
+ -- Sees if the unfolding is pretty certain to inline
+certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _))
+ = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold
+certainlyWillInline other
+ = False
+
+smallEnoughToInline :: Unfolding -> Bool
+smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _))
+ = size <= opt_UF_UseThreshold
+smallEnoughToInline other
+ = False
\end{code}
%************************************************************************
%* *
-\subsubsection{Printing unfoldings in interfaces}
+\subsection{callSiteInline}
%* *
%************************************************************************
-Printing Core-expression unfoldings is sufficiently delicate that we
-give it its own function.
-\begin{code}
-{- OLD:
-pprCoreUnfolding
- :: CoreExpr
- -> Pretty
-
-pprCoreUnfolding expr
- = let
- (_, renamed) = instCoreExpr uniqSupply_u expr
- -- We rename every unfolding with a "steady" unique supply,
- -- so that the names won't constantly change.
- -- One place we *MUST NOT* use a splittable UniqueSupply!
- in
- ppr_uf_Expr emptyUniqSet renamed
-
-ppr_Unfolding = PprUnfolding (panic "CoreUnfold:ppr_Unfolding")
-\end{code}
-
-\begin{code}
-ppr_uf_Expr in_scopes (Var v) = pprIdInUnfolding in_scopes v
-ppr_uf_Expr in_scopes (Lit l) = ppr ppr_Unfolding l
-
-ppr_uf_Expr in_scopes (Con c as)
- = ppBesides [ppPStr SLIT("_!_ "), pprIdInUnfolding no_in_scopes c, ppSP,
- ppLbrack, ppIntersperse pp'SP{-'-} (map (pprParendUniType ppr_Unfolding) ts), ppRbrack,
- ppSP, ppLbrack, ppIntersperse pp'SP{-'-} (map (ppr_uf_Atom in_scopes) as), ppRbrack]
-ppr_uf_Expr in_scopes (Prim op as)
- = ppBesides [ppPStr SLIT("_#_ "), ppr ppr_Unfolding op, ppSP,
- ppLbrack, ppIntersperse pp'SP{-'-} (map (pprParendUniType ppr_Unfolding) ts), ppRbrack,
- ppSP, ppLbrack, ppIntersperse pp'SP{-'-} (map (ppr_uf_Atom in_scopes) as), ppRbrack]
-
-ppr_uf_Expr in_scopes (Lam binder body)
- = ppCat [ppChar '\\', ppr_uf_Binder binder,
- ppPStr SLIT("->"), ppr_uf_Expr (in_scopes `add1` binder) body]
-
-ppr_uf_Expr in_scopes (CoTyLam tyvar expr)
- = ppCat [ppPStr SLIT("_/\\_"), interppSP ppr_Unfolding (tyvar:tyvars), ppStr "->",
- ppr_uf_Expr in_scopes body]
- where
- (tyvars, body) = collect_tyvars expr
+This is the key function. It decides whether to inline a variable at a call site
- collect_tyvars (CoTyLam tyv e) = ( tyv:tyvs, e_after )
- where (tyvs, e_after) = collect_tyvars e
- collect_tyvars other_e = ( [], other_e )
+callSiteInline is used at call sites, so it is a bit more generous.
+It's a very important function that embodies lots of heuristics.
+A non-WHNF can be inlined if it doesn't occur inside a lambda,
+and occurs exactly once or
+ occurs once in each branch of a case and is small
-ppr_uf_Expr in_scopes expr@(App fun_expr atom)
- = let
- (fun, args) = collect_args expr []
- in
- ppCat [ppPStr SLIT("_APP_ "), ppr_uf_Expr in_scopes fun, ppLbrack,
- ppIntersperse pp'SP{-'-} (map (ppr_uf_Atom in_scopes) args), ppRbrack]
- where
- collect_args (App fun arg) args = collect_args fun (arg:args)
- collect_args fun args = (fun, args)
+If the thing is in WHNF, there's no danger of duplicating work,
+so we can inline if it occurs once, or is small
-ppr_uf_Expr in_scopes (CoTyApp expr ty)
- = ppCat [ppPStr SLIT("_TYAPP_ "), ppr_uf_Expr in_scopes expr,
- ppChar '{', pprParendUniType ppr_Unfolding ty, ppChar '}']
+NOTE: we don't want to inline top-level functions that always diverge.
+It just makes the code bigger. Tt turns out that the convenient way to prevent
+them inlining is to give them a NOINLINE pragma, which we do in
+StrictAnal.addStrictnessInfoToTopId
-ppr_uf_Expr in_scopes (Case scrutinee alts)
- = ppCat [ppPStr SLIT("case"), ppr_uf_Expr in_scopes scrutinee, ppStr "of {",
- pp_alts alts, ppChar '}']
- where
- pp_alts (AlgAlts alts deflt)
- = ppCat [ppPStr SLIT("_ALG_"), ppCat (map pp_alg alts), pp_deflt deflt]
- pp_alts (PrimAlts alts deflt)
- = ppCat [ppPStr SLIT("_PRIM_"), ppCat (map pp_prim alts), pp_deflt deflt]
-
- pp_alg (con, params, rhs)
- = ppBesides [pprIdInUnfolding no_in_scopes con, ppSP,
- ppIntersperse ppSP (map ppr_uf_Binder params),
- ppPStr SLIT(" -> "), ppr_uf_Expr (in_scopes `add_some` params) rhs, ppSemi]
-
- pp_prim (lit, rhs)
- = ppBesides [ppr ppr_Unfolding lit,
- ppPStr SLIT(" -> "), ppr_uf_Expr in_scopes rhs, ppSemi]
-
- pp_deflt NoDefault = ppPStr SLIT("_NO_DEFLT_")
- pp_deflt (BindDefault binder rhs)
- = ppBesides [ppr_uf_Binder binder, ppPStr SLIT(" -> "),
- ppr_uf_Expr (in_scopes `add1` binder) rhs]
-
-ppr_uf_Expr in_scopes (Let (NonRec binder rhs) body)
- = ppBesides [ppStr "let {", ppr_uf_Binder binder, ppPStr SLIT(" = "), ppr_uf_Expr in_scopes rhs,
- ppStr "} in ", ppr_uf_Expr (in_scopes `add1` binder) body]
-
-ppr_uf_Expr in_scopes (Let (Rec pairs) body)
- = ppBesides [ppStr "_LETREC_ {", ppIntersperse sep (map pp_pair pairs),
- ppStr "} in ", ppr_uf_Expr new_in_scopes body]
+\begin{code}
+callSiteInline :: DynFlags
+ -> Bool -- True <=> the Id can be inlined
+ -> Bool -- 'inline' note at call site
+ -> OccInfo
+ -> Id -- The Id
+ -> [Bool] -- One for each value arg; True if it is interesting
+ -> Bool -- True <=> continuation is interesting
+ -> Maybe CoreExpr -- Unfolding, if any
+
+
+callSiteInline dflags active_inline inline_call occ id arg_infos interesting_cont
+ = case idUnfolding id of {
+ NoUnfolding -> Nothing ;
+ OtherCon cs -> Nothing ;
+
+ CompulsoryUnfolding unf_template -> Just unf_template ;
+ -- CompulsoryUnfolding => there is no top-level binding
+ -- for these things, so we must inline it.
+ -- Only a couple of primop-like things have
+ -- compulsory unfoldings (see MkId.lhs).
+ -- We don't allow them to be inactive
+
+ CoreUnfolding unf_template is_top is_value is_cheap guidance ->
+
+ let
+ result | yes_or_no = Just unf_template
+ | otherwise = Nothing
+
+ n_val_args = length arg_infos
+
+ yes_or_no
+ | not active_inline = False
+ | otherwise = case occ of
+ IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
+ IAmALoopBreaker -> False
+ --OneOcc in_lam _ _ -> (not in_lam || is_cheap) && consider_safe True
+ other -> is_cheap && consider_safe False
+ -- we consider even the once-in-one-branch
+ -- occurrences, because they won't all have been
+ -- caught by preInlineUnconditionally. In particular,
+ -- if the occurrence is once inside a lambda, and the
+ -- rhs is cheap but not a manifest lambda, then
+ -- pre-inline will not have inlined it for fear of
+ -- invalidating the occurrence info in the rhs.
+
+ consider_safe once
+ -- consider_safe decides whether it's a good idea to
+ -- inline something, given that there's no
+ -- work-duplication issue (the caller checks that).
+ | inline_call = True
+
+ | otherwise
+ = case guidance of
+ UnfoldNever -> False
+ UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
+
+ | enough_args && size <= (n_vals_wanted + 1)
+ -- Inline unconditionally if there no size increase
+ -- Size of call is n_vals_wanted (+1 for the function)
+ -> True
+
+ | otherwise
+ -> some_benefit && small_enough
+
+ where
+ some_benefit = or arg_infos || really_interesting_cont ||
+ (not is_top && ({- once || -} (n_vals_wanted > 0 && enough_args)))
+ -- [was (once && not in_lam)]
+ -- If it occurs more than once, there must be
+ -- something interesting about some argument, or the
+ -- result context, to make it worth inlining
+ --
+ -- If a function has a nested defn we also record
+ -- some-benefit, on the grounds that we are often able
+ -- to eliminate the binding, and hence the allocation,
+ -- for the function altogether; this is good for join
+ -- points. But this only makes sense for *functions*;
+ -- inlining a constructor doesn't help allocation
+ -- unless the result is scrutinised. UNLESS the
+ -- constructor occurs just once, albeit possibly in
+ -- multiple case branches. Then inlining it doesn't
+ -- increase allocation, but it does increase the
+ -- chance that the constructor won't be allocated at
+ -- all in the branches that don't use it.
+
+ enough_args = n_val_args >= n_vals_wanted
+ really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
+ | n_val_args == n_vals_wanted = interesting_cont
+ | otherwise = True -- Extra args
+ -- really_interesting_cont tells if the result of the
+ -- call is in an interesting context.
+
+ small_enough = (size - discount) <= opt_UF_UseThreshold
+ discount = computeDiscount n_vals_wanted arg_discounts res_discount
+ arg_infos really_interesting_cont
+
+ in
+ if dopt Opt_D_dump_inlinings dflags then
+ pprTrace "Considering inlining"
+ (ppr id <+> vcat [text "active:" <+> ppr active_inline,
+ 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 "guidance" <+> ppr guidance,
+ text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"])
+ result
+ else
+ result
+ }
+
+computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
+computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
+ -- We multiple 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.
+
+ -- we also discount 1 for each argument passed, because these will
+ -- reduce with the lambdas in the function (we count 1 for a lambda
+ -- in size_up).
+ = 1 + -- Discount of 1 because the result replaces the call
+ -- so we count 1 for the function itself
+ length (take n_vals_wanted arg_infos) +
+ -- Discount of 1 for each arg supplied, because the
+ -- result replaces the call
+ round (opt_UF_KeenessFactor *
+ fromIntegral (arg_discount + result_discount))
where
- sep = ppBeside ppSemi ppSP
- new_in_scopes = in_scopes `add_some` map fst pairs
-
- pp_pair (b, rhs) = ppCat [ppr_uf_Binder b, ppEquals, ppr_uf_Expr new_in_scopes rhs]
+ arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
-ppr_uf_Expr in_scopes (SCC cc body)
- = ASSERT(not (noCostCentreAttached cc))
- ASSERT(not (currentOrSubsumedCosts cc))
- ppBesides [ppStr "_scc_ { ", ppStr (showCostCentre ppr_Unfolding False{-not as string-} cc), ppStr " } ", ppr_uf_Expr in_scopes body]
+ mk_arg_discount discount is_evald | is_evald = discount
+ | otherwise = 0
-ppr_uf_Expr in_scopes (Coerce _ _ _) = panic "ppr_uf_Expr:Coerce"
-\end{code}
-
-\begin{code}
-ppr_uf_Binder :: Id -> Pretty
-ppr_uf_Binder v
- = ppBesides [ppLparen, pprIdInUnfolding (unitUniqSet v) v, ppPStr SLIT(" :: "),
- ppr ppr_Unfolding (idType v), ppRparen]
-
-ppr_uf_Atom in_scopes (LitArg l) = ppr ppr_Unfolding l
-ppr_uf_Atom in_scopes (VarArg v) = pprIdInUnfolding in_scopes v
-END OLD -}
+ -- Don't give a result discount unless there are enough args
+ result_discount | result_used = res_discount -- Over-applied, or case scrut
+ | otherwise = 0
\end{code}
projects / ghc-hetmet.git / blobdiff