%
-% (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}
-#include "HsVersions.h"
-
module CoreUnfold (
- SimpleUnfolding(..), Unfolding(..), UnfoldingGuidance(..), -- types
- UfExpr, RdrName, -- For closure (delete in 1.3)
-
- FormSummary(..), mkFormSummary, whnfOrBottom, exprSmallEnoughToDup, exprIsTrivial,
+ Unfolding(..), UnfoldingGuidance, -- types
- noUnfolding, mkMagicUnfolding, mkUnfolding, getUnfoldingTemplate,
+ noUnfolding, mkUnfolding, getUnfoldingTemplate,
+ isEvaldUnfolding, hasUnfolding,
- smallEnoughToInline, couldBeSmallEnoughToInline, certainlySmallEnoughToInline,
- inlineUnconditionally,
+ couldBeSmallEnoughToInline,
+ certainlySmallEnoughToInline,
+ okToUnfoldInHiFile,
calcUnfoldingGuidance,
- PragmaInfo(..) -- Re-export
+ callSiteInline, blackListed
) where
-IMP_Ubiq()
-#if defined (__GLASGOW_HASKELL__) && __GLASGOW_HASKELL__ <= 201
-IMPORT_DELOOPER(IdLoop) -- for paranoia checking;
- -- and also to get mkMagicUnfoldingFun
-IMPORT_DELOOPER(PrelLoop) -- for paranoia checking
-IMPORT_DELOOPER(SmplLoop)
-#else
-import {-# SOURCE #-} MagicUFs
-#endif
-
-import Bag ( emptyBag, unitBag, unionBags, Bag )
+#include "HsVersions.h"
-import CmdLineOpts ( opt_UnfoldingCreationThreshold,
- opt_UnfoldingUseThreshold,
- opt_UnfoldingConDiscount
+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 Constants ( uNFOLDING_CHEAP_OP_COST,
- uNFOLDING_DEAR_OP_COST,
- uNFOLDING_NOREP_LIT_COST
- )
-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 CostCentre ( ccMentionsId )
-import Id ( SYN_IE(Id), idType, getIdArity, isBottomingId, isDataCon,
- idWantsToBeINLINEd, idMustBeINLINEd, idMustNotBeINLINEd,
- SYN_IE(IdSet), GenId{-instances-} )
-import PrimOp ( primOpCanTriggerGC, fragilePrimOp, PrimOp(..) )
-import IdInfo ( ArityInfo(..), bottomIsGuaranteed )
-import Literal ( isNoRepLit, isLitLitLit )
-import Pretty
-import TyCon ( tyConFamilySize )
-import Type ( maybeAppDataTyConExpandingDicts )
-import Unique ( Unique )
-import UniqSet ( emptyUniqSet, unitUniqSet, mkUniqSet,
- addOneToUniqSet, unionUniqSets
+import BinderInfo ( )
+import CoreUtils ( coreExprType, exprIsTrivial, mkFormSummary, whnfOrBottom,
+ FormSummary(..) )
+import Id ( Id, idType, idUnique, isId,
+ getIdSpecialisation, getInlinePragma, getIdUnfolding
)
-import Usage ( SYN_IE(UVar) )
+import VarSet
+import Const ( Con(..), isLitLitLit, isWHNFCon )
+import PrimOp ( PrimOp(..), primOpIsDupable )
+import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..) )
+import TyCon ( tyConFamilySize )
+import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe )
+import Const ( isNoRepLit )
+import Unique ( Unique, buildIdKey, augmentIdKey, runSTRepIdKey )
import Maybes ( maybeToBool )
-import Util ( isIn, panic, assertPanic )
-#if __GLASGOW_HASKELL__ >= 202
+import Bag
+import Util ( isIn, lengthExceeds )
import Outputable
-
-#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.
- | MagicUnfolding
- Unique -- Unique of the Id whose magic unfolding this is
- MagicUnfoldingFun
-
-
-data SimpleUnfolding
- = SimpleUnfolding -- An unfolding with redundant cached information
+ | CoreUnfolding -- 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
-
+ CoreExpr -- Template; binder-info is correct
+\end{code}
+\begin{code}
noUnfolding = NoUnfolding
-mkUnfolding inline_prag expr
+mkUnfolding expr
= let
-- strictness mangling (depends on there being no CSE)
- ufg = calcUnfoldingGuidance inline_prag opt_UnfoldingCreationThreshold expr
+ ufg = calcUnfoldingGuidance opt_UF_CreationThreshold 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 }
-
-mkMagicUnfolding :: Unique -> Unfolding
-mkMagicUnfolding tag = MagicUnfolding tag (mkMagicUnfoldingFun tag)
+ CoreUnfolding (mkFormSummary expr) ufg occ
getUnfoldingTemplate :: Unfolding -> CoreExpr
-getUnfoldingTemplate (CoreUnfolding (SimpleUnfolding _ _ expr))
- = unTagBinders expr
+getUnfoldingTemplate (CoreUnfolding _ _ expr) = expr
getUnfoldingTemplate other = panic "getUnfoldingTemplate"
+isEvaldUnfolding :: Unfolding -> Bool
+isEvaldUnfolding (OtherCon _) = True
+isEvaldUnfolding (CoreUnfolding ValueForm _ expr) = True
+isEvaldUnfolding other = False
+
+hasUnfolding :: Unfolding -> Bool
+hasUnfolding NoUnfolding = False
+hasUnfolding other = True
data UnfoldingGuidance
= UnfoldNever
-- 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
\begin{code}
instance Outputable UnfoldingGuidance where
- ppr sty UnfoldAlways = ptext SLIT("_ALWAYS_")
- ppr sty (UnfoldIfGoodArgs t v cs size discount)
- = hsep [ptext SLIT("_IF_ARGS_"), int t, int v,
+ ppr UnfoldAlways = ptext SLIT("ALWAYS")
+ ppr UnfoldNever = ptext SLIT("NEVER")
+ ppr (UnfoldIfGoodArgs v cs size discount)
+ = hsep [ptext SLIT("IF_ARGS"), int v,
if null cs -- always print *something*
then char 'X'
else hcat (map (text . show) cs),
%************************************************************************
%* *
-\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 sty VarForm = ptext SLIT("Var")
- ppr sty ValueForm = ptext SLIT("Value")
- ppr sty BottomForm = ptext SLIT("Bot")
- ppr sty OtherForm = ptext SLIT("Other")
-
-mkFormSummary ::GenCoreExpr bndr Id tyvar uvar -> 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 { (use_binders, ty_binders, val_binders, body) ->
+calcUnfoldingGuidance bOMB_OUT_SIZE expr
+ | exprIsTrivial expr -- Often trivial expressions are never bound
+ -- to an expression, but it can happen. For
+ -- example, the Id for a nullary constructor has
+ -- a trivial expression as its unfolding, and
+ -- we want to make sure that we always unfold it.
+ = UnfoldAlways
+
+ | otherwise
+ = case collectBinders expr of { (binders, body) ->
+ let
+ val_binders = filter isId binders
+ in
case (sizeExpr bOMB_OUT_SIZE val_binders body) of
TooBig -> UnfoldNever
SizeIs size cased_args scrut_discount
-> UnfoldIfGoodArgs
- (length ty_binders)
(length val_binders)
(map discount_for val_binders)
(I# size)
(I# scrut_discount)
where
- discount_for b
- | is_data && b `is_elem` cased_args = tyConFamilySize tycon
- | otherwise = 0
- where
- (is_data, tycon)
- = case (maybeAppDataTyConExpandingDicts (idType b)) of
- Nothing -> (False, panic "discount")
- Just (tc,_,_) -> (True, tc)
-
- is_elem = isIn "calcUnfoldingGuidance" }
+ discount_for b
+ | num_cases == 0 = 0
+ | is_fun_ty = num_cases * opt_UF_FunAppDiscount
+ | is_data_ty = num_cases * tyConFamilySize tycon * 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)
+ }
\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 InlineMe _) = sizeTwo -- The idea is that this is one more
+ -- than the size of the "call" (i.e. 1)
+ -- We want to reply "no" to noSizeIncrease
+ -- for a bare reference (i.e. applied to no args)
+ -- to an INLINE thing
- 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 (Note _ body) = size_up body -- Notes cost nothing
- 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
- (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)
- = nukeScrutDiscount (size_up rhs)
- `addSize`
- size_up body
+ size_up (Con con args) = foldr (addSize . size_up)
+ (size_up_con con args)
+ args
- 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"
+ size_up (Lam b e) | isId b = size_up e `addSizeN` 1
+ | otherwise = size_up e
- ------------
- -- 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 (Let (NonRec binder rhs) body)
+ = nukeScrutDiscount (size_up rhs) `addSize`
+ size_up body `addSizeN`
+ 1 -- For the allocation
- ------------
- size_up_alts scrut_ty (AlgAlts alts deflt)
- = (foldr (addSize . size_alg_alt) (size_up_deflt deflt) alts)
- `addSizeN`
- alt_cost
+ size_up (Let (Rec pairs) body)
+ = 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`
+ 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 . size_up) (fun_discount 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
+ fun_discount (Var fun) | idUnique fun == buildIdKey = buildSize
+ | fun `is_elem` args = scrutArg fun
+ fun_discount other = sizeZero
+
+ ------------
+ 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 (maybeAppDataTyConExpandingDicts 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.
+
+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).
+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}
-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 &&
- discounted_size <= opt_UnfoldingUseThreshold
- where
+couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
+couldBeSmallEnoughToInline UnfoldNever = False
+couldBeSmallEnoughToInline other = True
- enough_args n [] | n > 0 = False -- A function with no value args => don't unfold
- enough_args _ _ = True -- Otherwise it's ok to try
+certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
+certainlySmallEnoughToInline UnfoldNever = False
+certainlySmallEnoughToInline UnfoldAlways = True
+certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold
+\end{code}
- discounted_size = (size - args_discount) - result_discount
+@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.
- args_discount = sum (zipWith arg_discount discount_vec arg_is_evald_s)
- result_discount | result_is_scruted = scrut_discount
- | otherwise = 0
+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.
- arg_discount no_of_constrs is_evald
- | is_evald = 1 + no_of_constrs * opt_UnfoldingConDiscount
- | otherwise = 1
+\begin{code}
+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}
-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.
+
+%************************************************************************
+%* *
+\subsection{callSiteInline}
+%* *
+%************************************************************************
+
+This is the key function. It decides whether to inline a variable at a call site
+
+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
+
+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}
---UNUSED?
-couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool
-couldBeSmallEnoughToInline guidance = smallEnoughToInline (repeat True) True guidance
+callSiteInline :: Bool -- True <=> the Id is black listed
+ -> Bool -- 'inline' note at call site
+ -> Id -- The Id
+ -> [CoreExpr] -- Arguments
+ -> Bool -- True <=> continuation is interesting
+ -> Maybe CoreExpr -- Unfolding, if any
+
+
+callSiteInline black_listed inline_call id args interesting_cont
+ = case getIdUnfolding id of {
+ NoUnfolding -> Nothing ;
+ OtherCon _ -> Nothing ;
+ CoreUnfolding form guidance unf_template ->
+
+ let
+ result | yes_or_no = Just unf_template
+ | otherwise = Nothing
+
+ inline_prag = getInlinePragma id
+ arg_infos = map interestingArg val_args
+ val_args = filter isValArg args
+ whnf = whnfOrBottom form
+
+ yes_or_no =
+ case inline_prag of
+ IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
+ IMustNotBeINLINEd -> False
+ IAmALoopBreaker -> False
+ IMustBeINLINEd -> True -- Overrides absolutely everything, including the black list
+ ICanSafelyBeINLINEd in_lam one_br -> consider in_lam one_br
+ NoInlinePragInfo -> consider InsideLam False
+
+ consider in_lam one_branch
+ | black_listed = False
+ | inline_call = True
+ | 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.
+ = WARN( case in_lam of { NotInsideLam -> True; other -> False },
+ text "callSiteInline:oneOcc" <+> ppr id )
+ -- If it has one occurrence, not inside a lambda, PreInlineUnconditionally
+ -- should have zapped it already
+ whnf && (not (null args) || interesting_cont)
+
+ | otherwise -- Occurs (textually) more than once, so look at its size
+ = case guidance of
+ UnfoldAlways -> True
+ 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)
+ -> case in_lam of
+ NotInsideLam -> True
+ InsideLam -> whnf
+
+ | not (or arg_infos || really_interesting_cont)
+ -- If it occurs more than once, there must be something interesting
+ -- about some argument, or the result, to make it worth inlining
+ -> False
+
+ | otherwise
+ -> case in_lam of
+ NotInsideLam -> small_enough
+ InsideLam -> whnf && small_enough
+
+ where
+ n_args = length arg_infos
+ enough_args = n_args >= n_vals_wanted
+ really_interesting_cont | n_args < n_vals_wanted = False -- Too few args
+ | n_args == n_vals_wanted = interesting_cont
+ | otherwise = True -- Extra args
+ -- This rather elaborate defn for really_interesting_cont is important
+ -- Consider an I# = INLINE (\x -> I# {x})
+ -- The unfolding guidance deems it to have size 2, and no arguments.
+ -- So in an application (I# y) we must take the extra arg 'y' as
+ -- evidence of 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 "inline prag:" <+> ppr inline_prag,
+ text "arg infos" <+> ppr arg_infos,
+ text "interesting continuation" <+> ppr interesting_cont,
+ text "whnf" <+> ppr whnf,
+ 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
+ }
-certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool
-certainlySmallEnoughToInline guidance = smallEnoughToInline (repeat False) False guidance
+-- An argument is interesting if it has *some* structure
+-- We are here trying to avoid unfolding a function that
+-- is applied only to variables that have no unfolding
+-- (i.e. they are probably lambda bound): f x y z
+-- There is little point in inlining f here.
+interestingArg (Type _) = False
+interestingArg (App fn (Type _)) = interestingArg fn
+interestingArg (Var v) = hasUnfolding (getIdUnfolding v)
+interestingArg other = True
+
+
+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).
+ = 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}
-Predicates
-~~~~~~~~~~
-@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
+%************************************************************************
+%* *
+\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.
\begin{code}
-inlineUnconditionally :: Bool -> Id -> BinderInfo -> Bool
+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.
+
+-- Phase 0: used for 'no inlinings please'
+blackListed rule_vars (Just 0)
+ = \v -> True
+
+-- Phase 1: don't inline any rule-y things or things with specialisations
+blackListed rule_vars (Just 1)
+ = \v -> let v_uniq = idUnique v
+ in v `elemVarSet` rule_vars
+ || not (isEmptyCoreRules (getIdSpecialisation v))
+ || v_uniq == runSTRepIdKey
+
+-- Phase 2: allow build/augment to inline, and specialisations
+blackListed rule_vars (Just 2)
+ = \v -> let v_uniq = idUnique v
+ in (v `elemVarSet` rule_vars && not (v_uniq == buildIdKey ||
+ v_uniq == augmentIdKey))
+ || v_uniq == runSTRepIdKey
+
+-- Otherwise just go for it
+blackListed rule_vars phase
+ = \v -> False
+\end{code}
-inlineUnconditionally ok_to_dup id occ_info
- | idMustNotBeINLINEd id = False
- | isOneFunOcc occ_info
- && idMustBeINLINEd id = True
+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!
- | isOneSafeFunOcc (ok_to_dup || idWantsToBeINLINEd id) occ_info
- = True
+Yet we do want to inline runST sometime, so we can avoid
+needless code. Solution: black list it until the last moment.
- | otherwise
- = False
-\end{code}