X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcoreSyn%2FCoreUnfold.lhs;h=d57f1886fc0377e7b7d91397a2c169374a9915a2;hb=28a464a75e14cece5db40f2765a29348273ff2d2;hp=6853b96b7be758264a544671df535b4fb9aea642;hpb=cbdeae8fc8a1c72d20d89241acae8a313214b51c;p=ghc-hetmet.git diff --git a/ghc/compiler/coreSyn/CoreUnfold.lhs b/ghc/compiler/coreSyn/CoreUnfold.lhs index 6853b96..d57f188 100644 --- a/ghc/compiler/coreSyn/CoreUnfold.lhs +++ b/ghc/compiler/coreSyn/CoreUnfold.lhs @@ -17,42 +17,35 @@ module CoreUnfold ( Unfolding, UnfoldingGuidance, -- Abstract types noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding, - mkOtherCon, otherCons, + evaldUnfolding, mkOtherCon, otherCons, unfoldingTemplate, maybeUnfoldingTemplate, isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding, hasUnfolding, hasSomeUnfolding, neverUnfold, couldBeSmallEnoughToInline, - certainlyWillInline, - okToUnfoldInHiFile, + certainlyWillInline, smallEnoughToInline, - callSiteInline, blackListed + callSiteInline ) where #include "HsVersions.h" -import CmdLineOpts ( opt_UF_CreationThreshold, - opt_UF_UseThreshold, - opt_UF_FunAppDiscount, - opt_UF_KeenessFactor, - opt_UF_DearOp, opt_UnfoldCasms, - DynFlags, DynFlag(..), dopt +import StaticFlags ( opt_UF_CreationThreshold, opt_UF_UseThreshold, + opt_UF_FunAppDiscount, opt_UF_KeenessFactor, + opt_UF_DearOp, ) +import DynFlags ( DynFlags, DynFlag(..), dopt ) import CoreSyn import PprCore ( pprCoreExpr ) -import OccurAnal ( occurAnalyseGlobalExpr ) -import CoreUtils ( exprIsValue, exprIsCheap, exprIsTrivial ) +import OccurAnal ( occurAnalyseExpr ) +import CoreUtils ( exprIsHNF, exprIsCheap, exprIsTrivial ) import Id ( Id, idType, isId, - idSpecialisation, idInlinePragma, idUnfolding, - isFCallId_maybe, globalIdDetails + idUnfolding, globalIdDetails ) -import VarSet -import Literal ( isLitLitLit, litSize ) +import DataCon ( isUnboxedTupleCon ) +import Literal ( litSize ) import PrimOp ( primOpIsDupable, primOpOutOfLine ) -import ForeignCall ( okToExposeFCall ) -import IdInfo ( InlinePragInfo(..), OccInfo(..), GlobalIdDetails(..), - isNeverInlinePrag - ) +import IdInfo ( OccInfo(..), GlobalIdDetails(..) ) import Type ( isUnLiftedType ) import PrelNames ( hasKey, buildIdKey, augmentIdKey ) import Bag @@ -60,7 +53,7 @@ import FastTypes import Outputable #if __GLASGOW_HASKELL__ >= 404 -import GlaExts ( fromInt ) +import GLAEXTS ( Int# ) #endif \end{code} @@ -75,9 +68,10 @@ import GlaExts ( fromInt ) mkTopUnfolding expr = mkUnfolding True {- Top level -} expr mkUnfolding top_lvl expr - = CoreUnfolding (occurAnalyseGlobalExpr expr) + = CoreUnfolding (occurAnalyseExpr expr) top_lvl - (exprIsValue expr) + + (exprIsHNF expr) -- Already evaluated (exprIsCheap expr) @@ -94,7 +88,7 @@ mkUnfolding top_lvl expr -- it gets fixed up next round mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded - = CompulsoryUnfolding (occurAnalyseGlobalExpr expr) + = CompulsoryUnfolding (occurAnalyseExpr expr) \end{code} @@ -133,13 +127,13 @@ calcUnfoldingGuidance bOMB_OUT_SIZE expr -- but no more. in - case (sizeExpr bOMB_OUT_SIZE val_binders body) of + case (sizeExpr (iUnbox bOMB_OUT_SIZE) val_binders body) of TooBig | not inline -> UnfoldNever -- A big function with an INLINE pragma must -- have an UnfoldIfGoodArgs guidance - | inline -> UnfoldIfGoodArgs n_val_binders + | otherwise -> UnfoldIfGoodArgs n_val_binders (map (const 0) val_binders) max_inline_size 0 @@ -177,7 +171,7 @@ calcUnfoldingGuidance bOMB_OUT_SIZE expr \end{code} \begin{code} -sizeExpr :: 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 @@ -221,7 +215,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr where rhs_size = foldr (addSize . size_up . snd) sizeZero pairs - size_up (Case (Var v) _ alts) + size_up (Case (Var v) _ _ alts) | v `elem` top_args -- We are scrutinising an argument variable = {- I'm nuking this special case; BUT see the comment with case alternatives. @@ -245,7 +239,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- into h; if we inline f first, while it looks small, then g's -- wrapper will get inlined later anyway. To avoid this nasty -- ordering difference, we make (case a of (x,y) -> ...), - -- *where a is one of the arguments* look free. + -- *where a is one of the arguments* look free. other -> -} @@ -269,9 +263,9 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- The 1+ is a little discount for reduced allocation in the caller alts_size tot_size _ = tot_size - - size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize` - foldr (addSize . size_up_alt) sizeZero alts +-- gaw 2004 + size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize` + foldr (addSize . size_up_alt) sizeZero alts -- We don't charge for the case itself -- It's a strict thing, and the price of the call -- is paid by scrut. Also consider @@ -298,7 +292,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr | fun `hasKey` augmentIdKey = augmentSize | otherwise = case globalIdDetails fun of - DataConId dc -> conSizeN (valArgCount args) + DataConWorkId dc -> conSizeN dc (valArgCount args) FCallId fc -> sizeN opt_UF_DearOp PrimOpId op -> primOpSize op (valArgCount args) @@ -336,58 +330,66 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- These addSize things have to be here because -- I don't want to give them bOMB_OUT_SIZE as an argument - addSizeN TooBig _ = TooBig - addSizeN (SizeIs n xs d) m - | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig - | otherwise = SizeIs n_tot xs d - where - n_tot = n +# iUnbox m + addSizeN TooBig _ = TooBig + addSizeN (SizeIs n xs d) m = mkSizeIs bOMB_OUT_SIZE (n +# iUnbox m) xs d - addSize TooBig _ = TooBig - addSize _ TooBig = TooBig - addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2) - | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig - | otherwise = SizeIs n_tot xys d_tot - where - n_tot = n1 +# n2 - d_tot = d1 +# d2 - xys = xs `unionBags` ys + addSize TooBig _ = TooBig + addSize _ TooBig = TooBig + addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2) + = mkSizeIs bOMB_OUT_SIZE (n1 +# n2) (xs `unionBags` ys) (d1 +# d2) \end{code} Code for manipulating sizes \begin{code} - data ExprSize = TooBig | SizeIs FastInt -- Size found (Bag (Id,Int)) -- Arguments cased herein, and discount for each such FastInt -- Size to subtract if result is scrutinised -- by a case expression - +-- subtract the discount before deciding whether to bale out. eg. we +-- want to inline a large constructor application into a selector: +-- tup = (a_1, ..., a_99) +-- x = case tup of ... +-- +mkSizeIs max n xs d | (n -# d) ># max = TooBig + | otherwise = SizeIs n xs d + maxSize TooBig _ = TooBig maxSize _ TooBig = TooBig maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1 | otherwise = s2 -sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0) -sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0) +sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0) +sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0) sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0) -conSizeN n = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1) +conSizeN dc n + | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1) + | otherwise = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1) -- Treat constructors as size 1; we are keen to expose them -- (and we charge separately for their args). We can't treat -- them as size zero, else we find that (iBox x) has size 1, -- which is the same as a lone variable; and hence 'v' will -- always be replaced by (iBox x), where v is bound to iBox x. + -- + -- However, unboxed tuples count as size zero + -- I found occasions where we had + -- f x y z = case op# x y z of { s -> (# s, () #) } + -- and f wasn't getting inlined primOpSize op n_args | not (primOpIsDupable op) = sizeN opt_UF_DearOp - | not (primOpOutOfLine op) = sizeN (1 - n_args) + | not (primOpOutOfLine op) = sizeN (2 - n_args) -- Be very keen to inline simple primops. - -- We give a discount of 1 for each arg so that (op# x y z) costs 1. - -- I found occasions where we had - -- f x y z = case op# x y z of { s -> (# s, () #) } - -- and f wasn't getting inlined + -- We give a discount of 1 for each arg so that (op# x y z) costs 2. + -- We can't make it cost 1, else we'll inline let v = (op# x y z) + -- at every use of v, which is excessive. + -- + -- A good example is: + -- let x = +# p q in C {x} + -- Even though x get's an occurrence of 'many', its RHS looks cheap, + -- and there's a good chance it'll get inlined back into C's RHS. Urgh! | otherwise = sizeOne buildSize = SizeIs (-2#) emptyBag 4# @@ -451,48 +453,20 @@ couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold UnfoldNever -> False other -> True -certainlyWillInline :: Id -> Bool - -- Sees if the Id is pretty certain to inline -certainlyWillInline v - = case idUnfolding v of - - CoreUnfolding _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _) - -> is_value - && size - (n_vals +1) <= opt_UF_UseThreshold - - other -> False +certainlyWillInline :: Unfolding -> Bool + -- Sees if the unfolding is pretty certain to inline +certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _)) + = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold +certainlyWillInline other + = False + +smallEnoughToInline :: Unfolding -> Bool +smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _)) + = size <= opt_UF_UseThreshold +smallEnoughToInline other + = False \end{code} -@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. - -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} -okToUnfoldInHiFile :: CoreExpr -> Bool -okToUnfoldInHiFile e = opt_UnfoldCasms || go e - where - -- Race over an expression looking for CCalls.. - go (Var v) = case isFCallId_maybe v of - Just fcall -> okToExposeFCall fcall - Nothing -> True - go (Lit lit) = not (isLitLitLit lit) - 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)) && - not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ]) - go (Note _ body) = go body - go (Type _) = True -\end{code} - - %************************************************************************ %* * \subsection{callSiteInline} @@ -517,7 +491,7 @@ StrictAnal.addStrictnessInfoToTopId \begin{code} callSiteInline :: DynFlags - -> Bool -- True <=> the Id is black listed + -> Bool -- True <=> the Id can be inlined -> Bool -- 'inline' note at call site -> OccInfo -> Id -- The Id @@ -526,15 +500,18 @@ callSiteInline :: DynFlags -> Maybe CoreExpr -- Unfolding, if any -callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont +callSiteInline dflags active_inline inline_call occ id arg_infos interesting_cont = case idUnfolding id of { NoUnfolding -> Nothing ; OtherCon cs -> Nothing ; - CompulsoryUnfolding unf_template | black_listed -> Nothing - | otherwise -> Just unf_template ; - -- Constructors have compulsory unfoldings, but - -- may have rules, in which case they are - -- black listed till later + + CompulsoryUnfolding unf_template -> Just unf_template ; + -- CompulsoryUnfolding => there is no top-level binding + -- for these things, so we must inline it. + -- Only a couple of primop-like things have + -- compulsory unfoldings (see MkId.lhs). + -- We don't allow them to be inactive + CoreUnfolding unf_template is_top is_value is_cheap guidance -> let @@ -544,49 +521,33 @@ callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont n_val_args = length arg_infos yes_or_no - | black_listed = False - | otherwise = case occ of + | not active_inline = 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 + --OneOcc in_lam _ _ -> (not in_lam || is_cheap) && consider_safe True + other -> is_cheap && consider_safe False + -- we consider even the once-in-one-branch + -- occurrences, because they won't all have been + -- caught by preInlineUnconditionally. In particular, + -- if the occurrence is once inside a lambda, and the + -- rhs is cheap but not a manifest lambda, then + -- pre-inline will not have inlined it for fear of + -- invalidating the occurrence info in the rhs. + + consider_safe once + -- consider_safe decides whether it's a good idea to + -- inline something, given that there's no + -- work-duplication issue (the caller checks that). | inline_call = True - | once_in_one_branch - -- Be very keen to inline something if this is its unique occurrence: - -- - -- a) Inlining gives a good chance of eliminating the original - -- binding (and hence the allocation) for the thing. - -- (Provided it's not a top level binding, in which case the - -- allocation costs nothing.) - -- - -- b) Inlining a function that is called only once exposes the - -- body function to the call site. - -- - -- 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, - -- by allocating the function itself many times - -- - -- Note: there used to be a '&& not top_level' in the guard above, - -- but that stopped us inlining top-level functions used only once, - -- which is stupid - = WARN( not in_lam, ppr id ) -- If (not in_lam) && one_br then PreInlineUnconditionally - -- should have caught it, shouldn't it? - not (null arg_infos) || interesting_cont - | otherwise = case guidance of - UnfoldNever -> False ; + UnfoldNever -> False UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount | enough_args && size <= (n_vals_wanted + 1) - -- No size increase + -- Inline unconditionally if there no size increase -- Size of call is n_vals_wanted (+1 for the function) -> True @@ -595,20 +556,25 @@ callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont 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. - + (not is_top && ({- once || -} (n_vals_wanted > 0 && enough_args))) + -- [was (once && not in_lam)] + -- If it occurs more than once, there must be + -- something interesting about some argument, or the + -- result context, to make it worth inlining + -- + -- If a function has a nested defn we also record + -- some-benefit, on the grounds that we are often able + -- to eliminate the binding, and hence the allocation, + -- for the function altogether; this is good for join + -- points. But this only makes sense for *functions*; + -- inlining a constructor doesn't help allocation + -- unless the result is scrutinised. UNLESS the + -- constructor occurs just once, albeit possibly in + -- multiple case branches. Then inlining it doesn't + -- increase allocation, but it does increase the + -- chance that the constructor won't be allocated at + -- all in the branches that don't use it. + enough_args = n_val_args >= n_vals_wanted really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args | n_val_args == n_vals_wanted = interesting_cont @@ -623,17 +589,14 @@ callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont in if dopt Opt_D_dump_inlinings dflags then pprTrace "Considering inlining" - (ppr id <+> vcat [text "black listed:" <+> ppr black_listed, + (ppr id <+> vcat [text "active:" <+> ppr active_inline, text "occ info:" <+> ppr occ, text "arg infos" <+> ppr arg_infos, text "interesting continuation" <+> ppr interesting_cont, text "is value:" <+> ppr is_value, text "is cheap:" <+> ppr is_cheap, text "guidance" <+> ppr guidance, - text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO", - if yes_or_no then - text "Unfolding =" <+> pprCoreExpr unf_template - else empty]) + text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO"]) result else result @@ -643,8 +606,8 @@ 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) + -- *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 @@ -656,7 +619,7 @@ computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used -- Discount of 1 for each arg supplied, because the -- result replaces the call round (opt_UF_KeenessFactor * - fromInt (arg_discount + result_discount)) + fromIntegral (arg_discount + result_discount)) where arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos) @@ -667,95 +630,3 @@ computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used 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 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 -> isNeverInlinePrag (idInlinePragma v) - -blackListed rule_vars (Just phase) - = \v -> normal_case rule_vars phase v - -normal_case rule_vars phase v - = case idInlinePragma 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 (idSpecialisation 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. -