X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcoreSyn%2FCoreUnfold.lhs;h=44fe5a7799f8473edd26fd3264775c8cd2b2d0e0;hb=207fd30f24ba4175bb532d862c87b560e171b651;hp=7aec06e516423af4fa2ee86e3cb8cc8f5ae6d5bc;hpb=0596517a9b4b2b32e5d375a986351102ac4540fc;p=ghc-hetmet.git diff --git a/ghc/compiler/coreSyn/CoreUnfold.lhs b/ghc/compiler/coreSyn/CoreUnfold.lhs index 7aec06e..44fe5a7 100644 --- a/ghc/compiler/coreSyn/CoreUnfold.lhs +++ b/ghc/compiler/coreSyn/CoreUnfold.lhs @@ -1,235 +1,150 @@ % -% (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(..), + Unfolding(..), UnfoldingGuidance, -- types + + noUnfolding, mkUnfolding, getUnfoldingTemplate, + isEvaldUnfolding, hasUnfolding, + + couldBeSmallEnoughToInline, + certainlySmallEnoughToInline, + okToUnfoldInHiFile, - mkFormSummary, - mkGenForm, - mkMagicUnfolding, - modifyUnfoldingDetails, calcUnfoldingGuidance, - mentionedInUnfolding + + callSiteInline, blackListed ) where -import Ubiq -import IdLoop -- for paranoia checking -import 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 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 CoreSyn -import CoreUtils ( coreExprType ) -import CostCentre ( ccMentionsId ) -import Id ( IdSet(..), GenId{-instances-} ) -import IdInfo ( bottomIsGuaranteed ) -import Literal ( isNoRepLit, isLitLitLit ) -import MagicUFs ( mkMagicUnfoldingFun, MagicUnfoldingFun ) -import Pretty -import PrimOp ( PrimOp(..) ) -import Type ( getAppDataTyCon ) -import UniqSet ( emptyUniqSet, singletonUniqSet, mkUniqSet, - unionUniqSets +import PprCore ( pprCoreExpr ) +import CoreUtils ( whnfOrBottom ) +import OccurAnal ( occurAnalyseGlobalExpr ) +import BinderInfo ( ) +import CoreUtils ( coreExprType, exprIsTrivial, mkFormSummary, + FormSummary(..) ) +import Id ( Id, idType, idUnique, isId, + getIdSpecialisation, getInlinePragma, getIdUnfolding ) -import Usage ( UVar(..) ) -import Util ( isIn, panic ) - -manifestlyWHNF = panic "manifestlyWHNF (CoreUnfold)" -primOpCanTriggerGC = panic "primOpCanTriggerGC (CoreUnfold)" -getTyConFamilySize = panic "getTyConFamilySize (CoreUnfold)" -whatsMentionedInId = panic "whatsMentionedInId (CoreUnfold)" -getMentionedTyConsAndClassesFromType = panic "getMentionedTyConsAndClassesFromType (CoreUnfold)" +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 Bag +import Util ( isIn, lengthExceeds ) +import Outputable \end{code} %************************************************************************ %* * -\subsection{@UnfoldingDetails@ and @UnfoldingGuidance@ types} +\subsection{@Unfolding@ and @UnfoldingGuidance@ types} %* * %************************************************************************ -(And @FormSummary@, too.) - \begin{code} -data UnfoldingDetails - = NoUnfoldingDetails - - | LitForm - Literal - - | OtherLitForm - [Literal] -- It is a literal, but definitely not one of these - - | ConForm - Id -- The constructor - [CoreArg] -- Value arguments; NB OutArgs, already cloned - - | 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 - Bool -- True <=> At most one textual occurrence of the - -- binder in its scope, *or* - -- if we are happy to duplicate this - -- binding. - 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 +data Unfolding + = NoUnfolding + + | 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. + + | 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. + CoreExpr -- Template; binder-info is correct \end{code} \begin{code} -data UnfoldingGuidance - = UnfoldNever -- Don't do it! +noUnfolding = NoUnfolding + +mkUnfolding expr + = let + -- strictness mangling (depends on there being no CSE) + ufg = calcUnfoldingGuidance opt_UF_CreationThreshold expr + occ = occurAnalyseGlobalExpr expr + in + CoreUnfolding (mkFormSummary expr) ufg occ + +getUnfoldingTemplate :: Unfolding -> CoreExpr +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 | 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 -- and "n" value args - | 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 + [Int] -- Discount if the argument is evaluated. -- (i.e., a simplification will definitely - -- be possible). + -- be possible). One elt of the list per *value* arg. + 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 + 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.) \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, + 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 ppChar 'X' - else ppBesides (map pp_c cs), - ppInt size ] - where - pp_c False = ppChar 'X' - pp_c True = ppChar 'C' -\end{code} - - -%************************************************************************ -%* * -\subsection{@mkGenForm@ and @modifyUnfoldingDetails@} -%* * -%************************************************************************ - -\begin{code} -mkGenForm :: Bool -- Ok to Dup code down different case branches, - -- because of either a flag saying so, - -- or alternatively the object is *SMALL* - -> BinderInfo -- - -> FormSummary - -> TemplateOutExpr -- Template - -> UnfoldingGuidance -- Tells about the *size* of the template. - -> UnfoldingDetails - -mkGenForm safe_to_dup occ_info WhnfForm template guidance - = GenForm (oneTextualOcc safe_to_dup occ_info) WhnfForm template guidance - -mkGenForm safe_to_dup occ_info form_summary template guidance - | oneSafeOcc safe_to_dup occ_info -- Non-WHNF with only safe occurrences - = GenForm True form_summary template guidance - - | otherwise -- Not a WHNF, many occurrences - = NoUnfoldingDetails -\end{code} - -\begin{code} -modifyUnfoldingDetails - :: Bool -- OK to dup - -> BinderInfo -- New occurrence info for the thing - -> UnfoldingDetails - -> UnfoldingDetails - -modifyUnfoldingDetails ok_to_dup occ_info - (GenForm only_one form_summary template guidance) - | only_one = mkGenForm ok_to_dup occ_info form_summary template guidance - -modifyUnfoldingDetails ok_to_dup occ_info other = other + then char 'X' + else hcat (map (text . show) cs), + int size, + int discount ] \end{code} @@ -241,516 +156,493 @@ modifyUnfoldingDetails ok_to_dup occ_info other = other \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 scc_s_OK bOMB_OUT_SIZE expr - = let - (use_binders, ty_binders, val_binders, body) = collectBinders expr +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 scc_s_OK bOMB_OUT_SIZE val_binders body) of + case (sizeExpr bOMB_OUT_SIZE val_binders body) of - Nothing -> UnfoldNever + TooBig -> UnfoldNever - Just (size, cased_args) - -> let - uf = UnfoldIfGoodArgs - (length ty_binders) + SizeIs size cased_args scrut_discount + -> UnfoldIfGoodArgs (length val_binders) - [ b `is_elem` cased_args | b <- val_binders ] - size - in - -- pprTrace "calcUnfold:" (ppAbove (ppr PprDebug uf) (ppr PprDebug expr)) - uf - where - is_elem = isIn "calcUnfoldingGuidance" + (map discount_for val_binders) + (I# size) + (I# scrut_discount) + where + 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 :: 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 (I# bOMB_OUT_SIZE) 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 (SCC _ (Con _ _)) = Nothing -- **** HACK ***** - size_up (SCC lbl body) - = if scc_s_OK then size_up body else 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 (Type t) = sizeZero -- Types cost nothing + size_up (Var v) = sizeOne - size_up (Let (NonRec binder rhs) body) - = size_up rhs - `addSize` - size_up body - `addSizeN` - 1 + 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 (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 (Note _ body) = size_up body -- Notes cost nothing - ------------ - size_up_arg arg = if isValArg arg then sizeOne else sizeZero{-it's free-} + size_up (App fun (Type t)) = size_up fun + size_up (App fun arg) = size_up_app fun `addSize` size_up arg - ------------ - size_up_alts scrut_ty (AlgAlts alts deflt) - = foldr (addSize . size_alg_alt) (size_up_deflt deflt) alts - `addSizeN` - (case (getTyConFamilySize tycon) of { Just n -> n }) - -- 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 + size_up (Con con args) = foldr (addSize . size_up) + (size_up_con con args) + args - (tycon, _, _) = getAppDataTyCon scrut_ty + size_up (Lam b e) | isId b = size_up e `addSizeN` 1 + | otherwise = size_up e - 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 (NonRec binder rhs) body) + = nukeScrutDiscount (size_up rhs) `addSize` + size_up body `addSizeN` + 1 -- For the allocation + + 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 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 + + ------------ + -- A function application with at least one value argument + -- so if the function is an argument give it an arg-discount + size_up_app (App fun arg) = size_up_app fun `addSize` size_up arg + size_up_app fun = arg_discount fun `addSize` size_up fun + + ------------ + size_up_alt (con, bndrs, rhs) = size_up rhs + -- Don't charge for args, so that wrappers look cheap ------------ - 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 + + 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 ------------ - -- 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 + -- 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" ------------ - 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 + -- These addSize things have to be here because + -- I don't want to give them bOMB_OUT_SIZE as an argument - addSize Nothing _ = Nothing - addSize _ Nothing = Nothing - addSize (Just (n, xs)) (Just (m, ys)) - | tot < bOMB_OUT_SIZE = Just (tot, xys) - | otherwise = Nothing + addSizeN TooBig _ = TooBig + addSizeN (SizeIs n xs d) (I# m) + | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d + | otherwise = TooBig where - tot = n+m - xys = xs ++ ys + n_tot = n +# m + + addSize TooBig _ = TooBig + addSize _ TooBig = TooBig + addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2) + | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot + | otherwise = TooBig + where + n_tot = n1 +# n2 + d_tot = d1 +# d2 + xys = xs `unionBags` ys \end{code} -%************************************************************************ -%* * -\subsection[unfoldings-for-ifaces]{Processing unfoldings for interfaces} -%* * -%************************************************************************ +Code for manipulating sizes + +\begin{code} + +data ExprSize = TooBig + | SizeIs Int# -- Size found + (Bag Id) -- Arguments cased herein + Int# -- Size to subtract if result is scrutinised + -- by a case expression + +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. + +scrutArg v = SizeIs 0# (unitBag v) 0# + +nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0# +nukeScrutDiscount TooBig = TooBig +\end{code} -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} +\subsection[considerUnfolding]{Given all the info, do (not) do the unfolding} %* * %************************************************************************ -\begin{code} -type UnfoldM bndr thing - = IdSet -- in-scope Ids (passed downwards only) - -> (bndr -> Id) -- to extract an Id from a binder (down only) +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 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. - -> (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 = in_scopes `unionUniqSets` singletonUniqSet x -in_scopes `add_some` xs = in_scopes `unionUniqSets` mkUniqSet xs +couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool +couldBeSmallEnoughToInline UnfoldNever = False +couldBeSmallEnoughToInline other = True + +certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool +certainlySmallEnoughToInline UnfoldNever = False +certainlySmallEnoughToInline UnfoldAlways = True +certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold \end{code} -The can-see-inside-monad functions are the usual sorts of things. +@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. -\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} +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} -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)) +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} %************************************************************************ %* * -\subsubsection{Gathering up info for an interface-unfolding} +\subsection{callSiteInline} %* * %************************************************************************ -\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} +This is the key function. It decides whether to inline a variable at a call site -\begin{code} ---ment_expr :: GenCoreExpr bndr Id -> UnfoldM bndr () +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 -ment_expr (Var v) = consider_Id v -ment_expr (Lit l) = consider_lit l +If the thing is in WHNF, there's no danger of duplicating work, +so we can inline if it occurs once, or is small -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 () +\begin{code} +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. + = case in_lam of + NotInsideLam -> True + InsideLam -> 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 + -- evidene 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 + } + +-- 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 - 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_ty ty - = let - (tycons, clss) = getMentionedTyConsAndClassesFromType ty - in - addToMentionedTyConsUf tycons `thenUf_` - addToMentionedClassesUf clss + arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos) -------------- + mk_arg_discount discount is_evald | is_evald = discount + | otherwise = 0 -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 () + -- 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} + %************************************************************************ %* * -\subsubsection{Printing unfoldings in interfaces} +\subsection{Black-listing} %* * %************************************************************************ -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 +Inlining is controlled by the "Inline phase" number, which is set +by the per-simplification-pass '-finline-phase' flag. -ppr_Unfolding = PprUnfolding (panic "CoreUnfold:ppr_Unfolding") -\end{code} +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} -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 +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} - collect_tyvars (CoTyLam tyv e) = ( tyv:tyvs, e_after ) - where (tyvs, e_after) = collect_tyvars e - collect_tyvars other_e = ( [], other_e ) -ppr_uf_Expr in_scopes expr@(App fun_expr atom) - = let - (fun, args) = collect_args expr [] +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 - 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) - -ppr_uf_Expr in_scopes (CoTyApp expr ty) - = ppCat [ppPStr SLIT("_TYAPP_ "), ppr_uf_Expr in_scopes expr, - ppChar '{', pprParendUniType ppr_Unfolding ty, ppChar '}'] - -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] - 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] + \ 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! -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] -\end{code} +Yet we do want to inline runST sometime, so we can avoid +needless code. Solution: black list it until the last moment. -\begin{code} -ppr_uf_Binder :: Id -> Pretty -ppr_uf_Binder v - = ppBesides [ppLparen, pprIdInUnfolding (singletonUniqSet 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 -} -\end{code}