%
-% (c) The AQUA Project, Glasgow University, 1994-1996
+% (c) The AQUA Project, Glasgow University, 1994-1998
%
\section[CoreUnfold]{Core-syntax unfoldings}
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 @SimpleUnfolding@ unfolding, you will
+literal''). In the corner of a @CoreUnfolding@ unfolding, you will
find, unsurprisingly, a Core expression.
\begin{code}
module CoreUnfold (
- SimpleUnfolding(..), Unfolding(..), UnfoldingGuidance(..), -- types
- UfExpr, RdrName, -- For closure (delete in 1.3)
+ Unfolding, UnfoldingGuidance, -- types
- FormSummary(..), mkFormSummary, whnfOrBottom, exprSmallEnoughToDup, exprIsTrivial,
+ noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
+ mkOtherCon, otherCons,
+ unfoldingTemplate, maybeUnfoldingTemplate,
+ isEvaldUnfolding, isCheapUnfolding,
+ hasUnfolding, hasSomeUnfolding,
- noUnfolding, mkMagicUnfolding, mkUnfolding, getUnfoldingTemplate,
+ couldBeSmallEnoughToInline,
+ certainlySmallEnoughToInline,
+ okToUnfoldInHiFile,
- smallEnoughToInline, couldBeSmallEnoughToInline, certainlySmallEnoughToInline,
- inlineUnconditionally,
+ calcUnfoldingGuidance,
- calcUnfoldingGuidance,
-
- PragmaInfo(..) -- Re-export
+ callSiteInline, blackListed
) where
#include "HsVersions.h"
-import {-# SOURCE #-} MagicUFs ( MagicUnfoldingFun, mkMagicUnfoldingFun )
-
-import Bag ( emptyBag, unitBag, unionBags, Bag )
-
-import CmdLineOpts ( opt_UnfoldingCreationThreshold,
- opt_UnfoldingUseThreshold,
- opt_UnfoldingConDiscount,
- opt_UnfoldingKeenessFactor
- )
-import Constants ( uNFOLDING_CHEAP_OP_COST,
- uNFOLDING_DEAR_OP_COST,
- uNFOLDING_NOREP_LIT_COST
+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_UF_NoRepLit,
+ opt_UnfoldCasms, opt_PprStyle_Debug,
+ opt_D_dump_inlinings
)
-import BinderInfo ( BinderInfo, isOneFunOcc, isOneSafeFunOcc
- )
-import PragmaInfo ( PragmaInfo(..) )
import CoreSyn
-import CoreUtils ( unTagBinders )
-import HsCore ( UfExpr )
-import RdrHsSyn ( RdrName )
+import PprCore ( pprCoreExpr )
import OccurAnal ( occurAnalyseGlobalExpr )
-import CoreUtils ( coreExprType )
-import Id ( Id, idType, getIdArity, isBottomingId, isDataCon,
- idWantsToBeINLINEd, idMustBeINLINEd, idMustNotBeINLINEd,
- IdSet, GenId{-instances-} )
-import PrimOp ( primOpCanTriggerGC, fragilePrimOp, PrimOp(..) )
-import IdInfo ( ArityInfo(..), bottomIsGuaranteed )
-import Literal ( isNoRepLit, isLitLitLit )
-import TyCon ( tyConFamilySize )
-import Type ( splitAlgTyConApp_maybe )
-import Unique ( Unique )
-import UniqSet ( emptyUniqSet, unitUniqSet, mkUniqSet,
- addOneToUniqSet, unionUniqSets
+import BinderInfo ( )
+import CoreUtils ( coreExprType, exprIsTrivial, exprIsValue, exprIsCheap )
+import Id ( Id, idType, idUnique, isId, getIdWorkerInfo,
+ getIdSpecialisation, getInlinePragma, getIdUnfolding
)
+import VarSet
+import Name ( isLocallyDefined )
+import Const ( Con(..), isLitLitLit, isWHNFCon )
+import PrimOp ( PrimOp(..), primOpIsDupable )
+import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..), insideLam, workerExists )
+import TyCon ( tyConFamilySize )
+import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe, isUnLiftedType )
+import Const ( isNoRepLit )
+import Unique ( Unique, buildIdKey, augmentIdKey )
import Maybes ( maybeToBool )
-import Util ( isIn, panic, assertPanic )
+import Bag
+import Util ( isIn, lengthExceeds )
import Outputable
+
+#if __GLASGOW_HASKELL__ >= 404
+import GlaExts ( fromInt )
+#endif
\end{code}
%************************************************************************
data Unfolding
= NoUnfolding
- | CoreUnfolding SimpleUnfolding
+ | OtherCon [Con] -- 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
+ UnfoldingGuidance -- Tells about the *size* of the template.
- | MagicUnfolding
- Unique -- Unique of the Id whose magic unfolding this is
- MagicUnfoldingFun
+seqUnfolding :: Unfolding -> ()
+seqUnfolding (CoreUnfolding e top b1 b2 g)
+ = seqExpr e `seq` top `seq` b1 `seq` b2 `seq` seqGuidance g
+seqUnfolding other = ()
+\end{code}
+\begin{code}
+noUnfolding = NoUnfolding
+mkOtherCon = OtherCon
-data SimpleUnfolding
- = SimpleUnfolding -- An unfolding with redundant cached information
- FormSummary -- Tells whether the template is a WHNF or bottom
- UnfoldingGuidance -- Tells about the *size* of the template.
- SimplifiableCoreExpr -- Template
+mkTopUnfolding expr = mkUnfolding True expr
+mkUnfolding top_lvl expr
+ = CoreUnfolding (occurAnalyseGlobalExpr expr)
+ top_lvl
+ (exprIsCheap expr)
+ (exprIsValue expr)
+ (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
-noUnfolding = NoUnfolding
+mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
+ = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
-mkUnfolding inline_prag expr
- = let
- -- strictness mangling (depends on there being no CSE)
- ufg = calcUnfoldingGuidance inline_prag opt_UnfoldingCreationThreshold expr
- occ = occurAnalyseGlobalExpr expr
- cuf = CoreUnfolding (SimpleUnfolding (mkFormSummary expr) ufg occ)
-
- cont = case occ of { Var _ -> cuf; _ -> cuf }
- in
- case ufg of { UnfoldAlways -> cont; _ -> cont }
+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 = []
-mkMagicUnfolding :: Unique -> Unfolding
-mkMagicUnfolding tag = MagicUnfolding tag (mkMagicUnfoldingFun tag)
+isEvaldUnfolding :: Unfolding -> Bool
+isEvaldUnfolding (OtherCon _) = True
+isEvaldUnfolding (CoreUnfolding _ _ _ is_evald _) = is_evald
+isEvaldUnfolding other = False
-getUnfoldingTemplate :: Unfolding -> CoreExpr
-getUnfoldingTemplate (CoreUnfolding (SimpleUnfolding _ _ expr))
- = unTagBinders expr
-getUnfoldingTemplate other = panic "getUnfoldingTemplate"
+isCheapUnfolding :: Unfolding -> Bool
+isCheapUnfolding (CoreUnfolding _ _ is_cheap _ _) = is_cheap
+isCheapUnfolding 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
- | 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.
-
- | UnfoldIfGoodArgs Int -- if "m" type args
- Int -- and "n" value args
+ | UnfoldIfGoodArgs Int -- and "n" value args
[Int] -- Discount if the argument is evaluated.
-- (i.e., a simplification will definitely
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}
\begin{code}
instance Outputable UnfoldingGuidance where
- ppr UnfoldAlways = ptext SLIT("_ALWAYS_")
- ppr (UnfoldIfGoodArgs t v cs size discount)
- = hsep [ptext SLIT("_IF_ARGS_"), int t, int v,
- if null cs -- always print *something*
- then char 'X'
- else hcat (map (text . show) cs),
+ 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{Figuring out things about expressions}
-%* *
-%************************************************************************
-
-\begin{code}
-data FormSummary
- = VarForm -- Expression is a variable (or scc var, etc)
- | ValueForm -- Expression is a value: i.e. a value-lambda,constructor, or literal
- | 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 VarForm = ptext SLIT("Var")
- ppr ValueForm = ptext SLIT("Value")
- ppr BottomForm = ptext SLIT("Bot")
- ppr OtherForm = ptext SLIT("Other")
-
-mkFormSummary ::GenCoreExpr bndr Id flexi -> FormSummary
-
-mkFormSummary expr
- = go (0::Int) expr -- The "n" is the number of (value) arguments so far
- where
- go n (Lit _) = ASSERT(n==0) ValueForm
- go n (Con _ _) = ASSERT(n==0) ValueForm
- go n (Prim _ _) = OtherForm
- go n (SCC _ e) = go n e
- go n (Coerce _ _ e) = go n e
-
- go n (Let (NonRec b r) e) | exprIsTrivial r = go n e -- let f = f' alpha in (f,g)
- -- should be treated as a value
- go n (Let _ e) = OtherForm
- go n (Case _ _) = OtherForm
-
- go 0 (Lam (ValBinder x) e) = ValueForm -- NB: \x.bottom /= bottom!
- go n (Lam (ValBinder x) e) = go (n-1) e -- Applied lambda
- go n (Lam other_binder e) = go n e
-
- go n (App fun arg) | isValArg arg = go (n+1) fun
- go n (App fun other_arg) = go n fun
-
- go n (Var f) | isBottomingId f = BottomForm
- | isDataCon f = ValueForm -- Can happen inside imported unfoldings
- go 0 (Var f) = VarForm
- go n (Var f) = case getIdArity f of
- ArityExactly a | n < a -> ValueForm
- ArityAtLeast a | n < a -> ValueForm
- other -> OtherForm
-
-whnfOrBottom :: FormSummary -> Bool
-whnfOrBottom VarForm = True
-whnfOrBottom ValueForm = True
-whnfOrBottom BottomForm = True
-whnfOrBottom OtherForm = False
-\end{code}
-
-@exprIsTrivial@ is true of expressions we are unconditionally happy to duplicate;
-simple variables and constants, and type applications.
-
-\begin{code}
-exprIsTrivial (Var v) = True
-exprIsTrivial (Lit lit) = not (isNoRepLit lit)
-exprIsTrivial (App e (TyArg _)) = exprIsTrivial e
-exprIsTrivial (Coerce _ _ e) = exprIsTrivial e
-exprIsTrivial other = False
-\end{code}
-
-\begin{code}
-exprSmallEnoughToDup (Con _ _) = True -- Could check # of args
-exprSmallEnoughToDup (Prim op _) = not (fragilePrimOp op) -- Could check # of args
-exprSmallEnoughToDup (Lit lit) = not (isNoRepLit lit)
-exprSmallEnoughToDup (Coerce _ _ e) = exprSmallEnoughToDup e
-exprSmallEnoughToDup expr
- = case (collectArgs expr) of { (fun, _, vargs) ->
- case fun of
- Var v | length vargs <= 4 -> True
- _ -> False
- }
-
-\end{code}
-
-
-%************************************************************************
-%* *
\subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}
%* *
%************************************************************************
\begin{code}
calcUnfoldingGuidance
- :: PragmaInfo -- INLINE pragma stuff
- -> Int -- bomb out if size gets bigger than this
+ :: Int -- bomb out if size gets bigger than this
-> CoreExpr -- expression to look at
-> UnfoldingGuidance
-
-calcUnfoldingGuidance IMustBeINLINEd bOMB_OUT_SIZE expr = UnfoldAlways -- Always inline if the INLINE pragma says so
-calcUnfoldingGuidance IWantToBeINLINEd bOMB_OUT_SIZE expr = UnfoldAlways -- Always inline if the INLINE pragma says so
-calcUnfoldingGuidance IMustNotBeINLINEd bOMB_OUT_SIZE expr = UnfoldNever -- ...and vice versa...
-
-calcUnfoldingGuidance NoPragmaInfo bOMB_OUT_SIZE expr
- = case collectBinders expr of { (ty_binders, val_binders, body) ->
+calcUnfoldingGuidance bOMB_OUT_SIZE expr
+ = case collect_val_bndrs expr of { (inline, val_binders, body) ->
+ let
+ n_val_binders = length val_binders
+ in
case (sizeExpr bOMB_OUT_SIZE val_binders body) of
- TooBig -> UnfoldNever
+ TooBig
+ | not inline -> UnfoldNever
+ -- A big function with an INLINE pragma must
+ -- have an UnfoldIfGoodArgs guidance
+ | inline -> UnfoldIfGoodArgs n_val_binders
+ (map (const 0) val_binders)
+ (n_val_binders + 2) 0
+ -- See comments with final_size below
SizeIs size cased_args scrut_discount
-> UnfoldIfGoodArgs
- (length ty_binders)
- (length val_binders)
+ n_val_binders
(map discount_for val_binders)
- (I# size)
+ final_size
(I# scrut_discount)
where
- discount_for b
- | is_data && b `is_elem` cased_args = tyConFamilySize tycon
- | otherwise = 0
- where
- (is_data, tycon)
- = case (splitAlgTyConApp_maybe (idType b)) of
- Nothing -> (False, panic "discount")
- Just (tc,_,_) -> (True, tc)
-
- is_elem = isIn "calcUnfoldingGuidance" }
+ boxed_size = I# size
+
+ final_size | inline = 0 -- Trying very agresssive inlining of INLINE things.
+ -- Reason: we don't want to call the un-inlined version,
+ -- because its body is awful
+ -- boxed_size `min` (n_val_binders + 2) -- Trying "+2" again...
+ | otherwise = boxed_size
+ -- 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+1. This
+ -- This is enough to pass the no-size-increase test in callSiteInline,
+ -- but no more.
+ -- I tried n_val_binders+2, to just defeat the test, on the grounds that
+ -- we don't want to inline an INLINE thing into a totally boring context,
+ -- but I found that some wrappers (notably one for a join point) weren't
+ -- getting inlined, and that was terrible. In that particular case, the
+ -- call site applied the wrapper to realWorld#, so if we made that an
+ -- "interesting" value the inlining would have happened... but it was
+ -- simpler to inline wrappers a little more eagerly instead.
+ --
+ -- Sometimes, though, 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
+ | num_cases == 0 = 0
+ | is_fun_ty = num_cases * opt_UF_FunAppDiscount
+ | is_data_ty = num_cases * opt_UF_ScrutConDiscount
+ | otherwise = num_cases * opt_UF_PrimArgDiscount
+ where
+ num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args
+ -- Count occurrences of b in cased_args
+ arg_ty = idType b
+ is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)
+ (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of
+ Nothing -> (False, panic "discount")
+ Just (tc,_,_) -> (True, tc)
+ }
+ where
+
+ 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 (I# bOMB_OUT_SIZE) args expr
= size_up expr
where
- size_up (Var v) = sizeZero
- size_up (Lit lit) | isNoRepLit lit = sizeN uNFOLDING_NOREP_LIT_COST
- | otherwise = sizeZero
+ size_up (Type t) = sizeZero -- Types cost nothing
+ size_up (Var v) = sizeOne
- size_up (SCC lbl body) = size_up body -- SCCs cost nothing
- size_up (Coerce _ _ body) = size_up body -- Coercions cost nothing
+ size_up (Note _ body) = size_up body -- Notes cost nothing
- size_up (App fun arg) = size_up fun `addSize` size_up_arg arg
- -- NB Zero cost for for type applications;
- -- others cost 1 or more
+ size_up (App fun (Type t)) = size_up fun
+ size_up (App fun arg) = size_up_app fun [arg]
- size_up (Con con args) = conSizeN (numValArgs args)
- -- We don't count 1 for the constructor because we're
- -- quite keen to get constructors into the open
-
- 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
- (tyvars, args, body) = collectBinders expr
- in
- size_up body `addSizeN` length args
+ size_up (Con con args) = foldr (addSize . size_up)
+ (size_up_con con args)
+ args
+
+ size_up (Lam b e) | isId b = size_up e `addSizeN` 1
+ | otherwise = size_up e
size_up (Let (NonRec binder rhs) body)
- = nukeScrutDiscount (size_up rhs)
- `addSize`
- size_up 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 (Let (Rec pairs) body)
- = nukeScrutDiscount (foldr addSize sizeZero [size_up rhs | (_,rhs) <- pairs])
- `addSize`
- size_up body
-
- size_up (Case scrut alts)
- = nukeScrutDiscount (size_up scrut)
- `addSize`
- arg_discount scrut
- `addSize`
- size_up_alts (coreExprType scrut) alts
- -- We charge for the "case" itself in "size_up_alts"
-
- ------------
- -- In an application we charge 0 for type application
- -- 1 for most anything else
- -- N for norep_lits
- size_up_arg (LitArg lit) | isNoRepLit lit = sizeN uNFOLDING_NOREP_LIT_COST
- size_up_arg (TyArg _) = sizeZero
- size_up_arg other = sizeOne
-
- ------------
- size_up_alts scrut_ty (AlgAlts alts deflt)
- = (foldr (addSize . size_alg_alt) (size_up_deflt deflt) alts)
- `addSizeN`
- alt_cost
+ = nukeScrutDiscount rhs_size `addSize`
+ size_up body `addSizeN`
+ length pairs -- For the allocation
where
- size_alg_alt (con,args,rhs) = size_up rhs
+ rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
+
+ size_up (Case scrut _ alts)
+ = nukeScrutDiscount (size_up scrut) `addSize`
+ arg_discount scrut `addSize`
+ foldr (addSize . size_up_alt) sizeZero alts
+ `addSizeN` 1 -- charge one for the case itself.
+
+-- Just charge for the alts that exist, not the ones that might exist
+-- `addSizeN`
+-- case (splitAlgTyConApp_maybe (coreExprType scrut)) of
+-- Nothing -> 1
+-- Just (tc,_,_) -> tyConFamilySize tc
+
+ ------------
+ size_up_app (App fun arg) args = size_up_app fun (arg:args)
+ size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
+ (size_up_fun fun)
+ 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
+ size_up_fun (Var fun) | idUnique fun == buildIdKey = buildSize
+ | idUnique fun == augmentIdKey = augmentSize
+ | fun `is_elem` args = scrutArg fun `addSize` sizeOne
+ size_up_fun other = size_up other
+
+ ------------
+ size_up_alt (con, bndrs, rhs) = size_up rhs
-- Don't charge for args, so that wrappers look cheap
- -- 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.)
- -- It's important to charge for alternatives. If you don't then you
- -- get size 1 for things like:
- -- case x of { A -> 1#; B -> 2#; ... lots }
-
- alt_cost :: Int
- alt_cost
- = case (splitAlgTyConApp_maybe scrut_ty) of
- Nothing -> 1
- Just (tc,_,_) -> tyConFamilySize tc
-
- size_up_alts _ (PrimAlts alts deflt)
- = foldr (addSize . size_prim_alt) (size_up_deflt deflt) alts
- -- *no charge* for a primitive "case"!
- where
- size_prim_alt (lit,rhs) = size_up rhs
-
------------
- size_up_deflt NoDefault = sizeZero
- size_up_deflt (BindDefault binder rhs) = size_up rhs
+ size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit
+ | otherwise = sizeOne
- ------------
- -- We want to record if we're case'ing an argument
+ size_up_con (DataCon dc) args = conSizeN (valArgCount args)
+
+ size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)
+ -- Give an arg-discount if a primop is applies to
+ -- one of the function's arguments
+ where
+ op_cost | primOpIsDupable op = opt_UF_CheapOp
+ | otherwise = opt_UF_DearOp
+
+ -- We want to record if we're case'ing, or applying, an argument
arg_discount (Var v) | v `is_elem` args = scrutArg v
arg_discount other = sizeZero
+ ------------
is_elem :: Id -> [Id] -> Bool
is_elem = isIn "size_up_scrut"
where
n_tot = n1 +# n2
d_tot = d1 +# d2
- xys = xs ++ ys
-
-
+ xys = xs `unionBags` ys
\end{code}
Code for manipulating sizes
data ExprSize = TooBig
| SizeIs Int# -- Size found
- [Id] -- Arguments cased herein
+ (Bag Id) -- Arguments cased herein
Int# -- Size to subtract if result is scrutinised
-- by a case expression
-sizeZero = SizeIs 0# [] 0#
-sizeOne = SizeIs 1# [] 0#
-sizeN (I# n) = SizeIs n [] 0#
-conSizeN (I# n) = SizeIs n [] n
-scrutArg v = SizeIs 0# [v] 0#
+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#)
+ -- Treat constructors as size 1, that unfoldAlways responsds 'False'
+ -- when asked about 'x' when x is bound to (C 3#).
+ -- This avoids gratuitous 'ticks' when x itself appears as an
+ -- atomic constructor argument.
+
+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
+
+scrutArg v = SizeIs 0# (unitBag v) 0#
nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
nukeScrutDiscount TooBig = TooBig
\end{code}
+
%************************************************************************
%* *
\subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
is more accurate (see @sizeExpr@ above for how this discount size
is computed).
-\begin{code}
-smallEnoughToInline :: [Bool] -- Evaluated-ness of value arguments
- -> Bool -- Result is scrutinised
- -> UnfoldingGuidance
- -> Bool -- True => unfold it
-
-smallEnoughToInline _ _ UnfoldAlways = True
-smallEnoughToInline _ _ UnfoldNever = False
-smallEnoughToInline arg_is_evald_s result_is_scruted
- (UnfoldIfGoodArgs m_tys_wanted n_vals_wanted discount_vec size scrut_discount)
- = enough_args n_vals_wanted arg_is_evald_s &&
- size - discount <= opt_UnfoldingUseThreshold
- where
-
- enough_args n [] | n > 0 = False -- A function with no value args => don't unfold
- enough_args _ _ = True -- Otherwise it's ok to try
-
- -- 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.
- discount :: Int
- discount = round (
- opt_UnfoldingKeenessFactor *
- fromInt (args_discount + result_discount)
- )
-
- args_discount = sum (zipWith arg_discount discount_vec arg_is_evald_s)
- result_discount | result_is_scruted = scrut_discount
- | otherwise = 0
-
- arg_discount no_of_constrs is_evald
- | is_evald = 1 + no_of_constrs * opt_UnfoldingConDiscount
- | otherwise = 1
-\end{code}
-
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}
---UNUSED?
couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
-couldBeSmallEnoughToInline guidance = smallEnoughToInline (repeat True) True guidance
+couldBeSmallEnoughToInline UnfoldNever = False
+couldBeSmallEnoughToInline other = True
certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
-certainlySmallEnoughToInline guidance = smallEnoughToInline (repeat False) False guidance
+certainlySmallEnoughToInline UnfoldNever = False
+certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
\end{code}
-Predicates
-~~~~~~~~~~
+@okToUnfoldInHifile@ is used when emitting unfolding info into an interface
+file to determine whether an unfolding candidate really should be unfolded.
+The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
+into interface files.
-@inlineUnconditionally@ decides whether a let-bound thing can
-*definitely* be inlined at each of its call sites. If so, then
-we can drop the binding right away. But remember, you have to be
-certain that every use can be inlined. So, notably, any ArgOccs
-rule this out. Since ManyOcc doesn't record FunOcc/ArgOcc
+The reason for inlining expressions containing _casm_s into interface files
+is that these fragments of C are likely to mention functions/#defines that
+will be out-of-scope when inlined into another module. This is not an
+unfixable problem for the user (just need to -#include the approp. header
+file), but turning it off seems to the simplest thing to do.
\begin{code}
-inlineUnconditionally :: Bool -> Id -> BinderInfo -> Bool
+okToUnfoldInHiFile :: CoreExpr -> Bool
+okToUnfoldInHiFile e = opt_UnfoldCasms || go e
+ where
+ -- Race over an expression looking for CCalls..
+ go (Var _) = True
+ go (Con (Literal lit) _) = not (isLitLitLit lit)
+ go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args
+ go (Con con args) = True -- con args are always atomic
+ 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 (Note _ body) = go body
+ go (Type _) = True
+
+ -- ok to unfold a PrimOp as long as it's not a _casm_
+ okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm
+ okToUnfoldPrimOp _ = True
+\end{code}
-inlineUnconditionally ok_to_dup id occ_info
- | idMustNotBeINLINEd id = False
- | isOneFunOcc occ_info
- && idMustBeINLINEd id = True
+%************************************************************************
+%* *
+\subsection{callSiteInline}
+%* *
+%************************************************************************
+
+This is the key function. It decides whether to inline a variable at a call site
- | isOneSafeFunOcc (ok_to_dup || idWantsToBeINLINEd id) occ_info
- = True
+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
- | otherwise
- = False
+If the thing is in WHNF, there's no danger of duplicating work,
+so we can inline if it occurs once, or is small
+
+\begin{code}
+callSiteInline :: Bool -- True <=> the Id is black listed
+ -> 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 black_listed inline_call occ id arg_infos interesting_cont
+ = case getIdUnfolding id of {
+ NoUnfolding -> Nothing ;
+ OtherCon _ -> Nothing ;
+ CompulsoryUnfolding unf_template -> Just unf_template ;
+ CoreUnfolding unf_template is_top is_cheap _ guidance ->
+
+ let
+ result | yes_or_no = Just unf_template
+ | otherwise = Nothing
+
+ n_val_args = length arg_infos
+
+ 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 || 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,
+ -- given that there's no work-duplication issue (the caller checks that).
+ -- once_in_one_branch = True means there's a unique textual occurrence
+ | inline_call = True
+
+ | once_in_one_branch -- Be very keen to inline something if this is its unique occurrence; that
+ -- gives a good chance of eliminating the original binding for the thing.
+ -- The only time we hold back is when substituting inside a lambda;
+ -- then if the context is totally uninteresting (not applied, not scrutinised)
+ -- there is no point in substituting because it might just increase allocation.
+ = not in_lam || not (null arg_infos) || interesting_cont
+
+ | otherwise
+ = case guidance of
+ UnfoldNever -> False ;
+ UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
+
+ | enough_args && size <= (n_vals_wanted + 1)
+ -- 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)))
+ -- 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
+#ifdef DEBUG
+ if opt_D_dump_inlinings then
+ pprTrace "Considering inlining"
+ (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 cheap" <+> ppr is_cheap,
+ text "guidance" <+> ppr guidance,
+ text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
+ if yes_or_no then
+ text "Unfolding =" <+> pprCoreExpr unf_template
+ else empty])
+ result
+ else
+#endif
+ 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 *
+ fromInt (arg_discount + result_discount))
+ where
+ arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
+
+ mk_arg_discount discount is_evald | is_evald = discount
+ | otherwise = 0
+
+ -- 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}
+
+
+%************************************************************************
+%* *
+\subsection{Black-listing}
+%* *
+%************************************************************************
+
+Inlining is controlled by the "Inline phase" number, which is set
+by the per-simplification-pass '-finline-phase' flag.
+
+For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
+in that order. The meanings of these are determined by the @blackListed@ function
+here.
+
+The final simplification doesn't have a phase number
+
+Pragmas
+~~~~~~~
+ Pragma Black list if
+
+(least black listing, most inlining)
+ INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
+ INLINE foo phase is Just p *and* foo appears on LHS of rule
+ NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
+ NOINLINE foo always
+(most black listing, least inlining)
+
+\begin{code}
+blackListed :: IdSet -- Used in transformation rules
+ -> Maybe Int -- Inline phase
+ -> Id -> Bool -- True <=> blacklisted
+
+-- The blackListed function sees whether a variable should *not* be
+-- inlined because of the inline phase we are in. This is the sole
+-- place that the inline phase number is looked at.
+
+blackListed rule_vars Nothing -- Last phase
+ = \v -> case getInlinePragma v of
+ IMustNotBeINLINEd False Nothing -> True -- An unconditional NOINLINE pragma
+ other -> False
+
+blackListed rule_vars (Just 0)
+-- Phase 0: used for 'no imported inlinings please'
+-- This prevents wrappers getting inlined which in turn is bad for full laziness
+-- NEW: try using 'not a wrapper' rather than 'not imported' in this phase.
+-- This allows a little more inlining, which seems to be important, sometimes.
+-- For example PrelArr.newIntArr gets better.
+ = \v -> -- workerExists (getIdWorkerInfo v) || normal_case rule_vars 0 v
+ -- True -- Try going back to no inlinings at all
+ -- BUT: I found that there is some advantage in doing
+ -- local inlinings first. For example in fish/Main.hs
+ -- it's advantageous to inline scale_vec2 before inlining
+ -- wrappers from PrelNum that make it look big.
+ not (isLocallyDefined v) -- This seems best at the moment
+
+blackListed rule_vars (Just phase)
+ = \v -> normal_case rule_vars phase v
+
+normal_case rule_vars phase v
+ = case getInlinePragma v of
+ NoInlinePragInfo -> has_rules
+
+ IMustNotBeINLINEd from_INLINE Nothing
+ | 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
+ where
+ has_rules = v `elemVarSet` rule_vars
+ || not (isEmptyCoreRules (getIdSpecialisation v))
+\end{code}
+
+
+SLPJ 95/04: Why @runST@ must be inlined very late:
+\begin{verbatim}
+f x =
+ runST ( \ s -> let
+ (a, s') = newArray# 100 [] s
+ (_, s'') = fill_in_array_or_something a x s'
+ in
+ freezeArray# a s'' )
+\end{verbatim}
+If we inline @runST@, we'll get:
+\begin{verbatim}
+f x = let
+ (a, s') = newArray# 100 [] realWorld#{-NB-}
+ (_, s'') = fill_in_array_or_something a x s'
+ in
+ freezeArray# a s''
+\end{verbatim}
+And now the @newArray#@ binding can be floated to become a CAF, which
+is totally and utterly wrong:
+\begin{verbatim}
+f = let
+ (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
+ in
+ \ x ->
+ let (_, s'') = fill_in_array_or_something a x s' in
+ freezeArray# a s''
+\end{verbatim}
+All calls to @f@ will share a {\em single} array!
+
+Yet we do want to inline runST sometime, so we can avoid
+needless code. Solution: black list it until the last moment.
+