X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2FcoreSyn%2FCoreUnfold.lhs;h=767006076d60c7ed9d5e07bd9965442236865cac;hp=ad2a391ec27fedbb89a7f649cfed945b7dea460a;hb=fa1c8a7e7013b1e9a37326b80abadec737c9347e;hpb=b6cc5851fa49720b31d989d210c8e43dc27cbfe6 diff --git a/compiler/coreSyn/CoreUnfold.lhs b/compiler/coreSyn/CoreUnfold.lhs index ad2a391..7670060 100644 --- a/compiler/coreSyn/CoreUnfold.lhs +++ b/compiler/coreSyn/CoreUnfold.lhs @@ -1,7 +1,9 @@ % +% (c) The University of Glasgow 2006 % (c) The AQUA Project, Glasgow University, 1994-1998 % -\section[CoreUnfold]{Core-syntax unfoldings} + +Core-syntax unfoldings Unfoldings (which can travel across module boundaries) are in Core syntax (namely @CoreExpr@s). @@ -25,36 +27,29 @@ module CoreUnfold ( couldBeSmallEnoughToInline, certainlyWillInline, smallEnoughToInline, - callSiteInline + callSiteInline, CallCtxt(..) + ) where #include "HsVersions.h" -import StaticFlags ( opt_UF_CreationThreshold, opt_UF_UseThreshold, - opt_UF_FunAppDiscount, opt_UF_KeenessFactor, - opt_UF_DearOp, - ) -import DynFlags ( DynFlags, DynFlag(..), dopt ) +import StaticFlags +import DynFlags import CoreSyn import PprCore () -- Instances -import OccurAnal ( occurAnalyseExpr ) -import CoreUtils ( exprIsHNF, exprIsCheap, exprIsTrivial ) -import Id ( Id, idType, isId, - idUnfolding, globalIdDetails - ) -import DataCon ( isUnboxedTupleCon ) -import Literal ( litSize ) -import PrimOp ( primOpIsDupable, primOpOutOfLine ) -import IdInfo ( GlobalIdDetails(..) ) -import Type ( isUnLiftedType ) -import PrelNames ( hasKey, buildIdKey, augmentIdKey ) +import OccurAnal +import CoreUtils +import Id +import DataCon +import Literal +import PrimOp +import IdInfo +import Type +import PrelNames import Bag import FastTypes import Outputable -#if __GLASGOW_HASKELL__ >= 404 -import GLAEXTS ( Int# ) -#endif \end{code} @@ -65,8 +60,10 @@ import GLAEXTS ( Int# ) %************************************************************************ \begin{code} +mkTopUnfolding :: CoreExpr -> Unfolding mkTopUnfolding expr = mkUnfolding True {- Top level -} expr +mkUnfolding :: Bool -> CoreExpr -> Unfolding mkUnfolding top_lvl expr = CoreUnfolding (occurAnalyseExpr expr) top_lvl @@ -95,6 +92,7 @@ instance Outputable Unfolding where = ptext SLIT("Unf") <+> sep [ppr top <+> ppr hnf <+> ppr cheap <+> ppr g, ppr e] +mkCompulsoryUnfolding :: CoreExpr -> Unfolding mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded = CompulsoryUnfolding (occurAnalyseExpr expr) \end{code} @@ -172,14 +170,14 @@ calcUnfoldingGuidance bOMB_OUT_SIZE expr -- We want to say "2 value binders". Why? So that -- we take account of information given for the arguments - go inline rev_vbs (Note InlineMe e) = go True rev_vbs e + go _ rev_vbs (Note InlineMe e) = go True rev_vbs e go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e | otherwise = go inline rev_vbs e go inline rev_vbs e = (inline, reverse rev_vbs, e) \end{code} \begin{code} -sizeExpr :: Int# -- Bomb out if it gets bigger than this +sizeExpr :: FastInt -- Bomb out if it gets bigger than this -> [Id] -- Arguments; we're interested in which of these -- get case'd -> CoreExpr @@ -188,10 +186,10 @@ sizeExpr :: Int# -- Bomb out if it gets bigger than this sizeExpr bOMB_OUT_SIZE top_args expr = size_up expr where - size_up (Type t) = sizeZero -- Types cost nothing - size_up (Var v) = sizeOne + size_up (Type _) = sizeZero -- Types cost nothing + size_up (Var _) = sizeOne - size_up (Note InlineMe body) = sizeOne -- Inline notes make it look very small + size_up (Note InlineMe _) = sizeOne -- Inline notes make it look very small -- This can be important. If you have an instance decl like this: -- instance Foo a => Foo [a] where -- {-# INLINE op1, op2 #-} @@ -199,11 +197,11 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- op2 = ... -- then we'll get a dfun which is a pair of two INLINE lambdas - size_up (Note _ body) = size_up body -- Other notes cost nothing + size_up (Note _ body) = size_up body -- Other notes cost nothing - size_up (Cast e _) = size_up e + size_up (Cast e _) = size_up e - size_up (App fun (Type t)) = size_up fun + size_up (App fun (Type _)) = size_up fun size_up (App fun arg) = size_up_app fun [arg] size_up (Lit lit) = sizeN (litSize lit) @@ -239,7 +237,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr case alts of - [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0# + [alt] -> size_up_alt alt `addSize` SizeIs (_ILIT(0)) (unitBag (v, 1)) (_ILIT(0)) -- We want to make wrapper-style evaluation look cheap, so that -- when we inline a wrapper it doesn't make call site (much) bigger -- Otherwise we get nasty phase ordering stuff: @@ -265,15 +263,14 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- alts_size tries to compute a good discount for -- the case when we are scrutinising an argument variable - alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives - (SizeIs max max_disc max_scrut) -- Size of biggest alternative - = SizeIs tot (unitBag (v, iBox (_ILIT 1 +# tot -# max)) `unionBags` max_disc) max_scrut + alts_size (SizeIs tot _tot_disc _tot_scrut) -- Size of all alternatives + (SizeIs max max_disc max_scrut) -- Size of biggest alternative + = SizeIs tot (unitBag (v, iBox (_ILIT(1) +# tot -# max)) `unionBags` max_disc) max_scrut -- If the variable is known, we produce a discount that -- will take us back to 'max', the size of rh largest alternative -- The 1+ is a little discount for reduced allocation in the caller alts_size tot_size _ = tot_size --- gaw 2004 size_up (Case e _ _ alts) = nukeScrutDiscount (size_up e) `addSize` foldr (addSize . size_up_alt) sizeZero alts -- We don't charge for the case itself @@ -304,7 +301,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr = case globalIdDetails fun of DataConWorkId dc -> conSizeN dc (valArgCount args) - FCallId fc -> sizeN opt_UF_DearOp + FCallId _ -> sizeN opt_UF_DearOp PrimOpId op -> primOpSize op (valArgCount args) -- foldr addSize (primOpSize op) (map arg_discount args) -- At one time I tried giving an arg-discount if a primop @@ -314,7 +311,7 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- if we know nothing about it. And just having it in a primop -- doesn't help at all if we don't know something more. - other -> fun_discount fun `addSizeN` + _ -> fun_discount fun `addSizeN` (1 + length (filter (not . exprIsTrivial) args)) -- The 1+ is for the function itself -- Add 1 for each non-trivial arg; @@ -324,17 +321,17 @@ sizeExpr bOMB_OUT_SIZE top_args expr -- We should really only count for non-prim-typed args in the -- general case, but that seems too much like hard work - size_up_fun other args = size_up other + size_up_fun other _ = size_up other ------------ - size_up_alt (con, bndrs, rhs) = size_up rhs + size_up_alt (_con, _bndrs, rhs) = size_up rhs -- Don't charge for args, so that wrappers look cheap -- (See comments about wrappers with Case) ------------ -- We want to record if we're case'ing, or applying, an argument - fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0# - fun_discount other = sizeZero + fun_discount v | v `elem` top_args = SizeIs (_ILIT(0)) (unitBag (v, opt_UF_FunAppDiscount)) (_ILIT(0)) + fun_discount _ = sizeZero ------------ -- These addSize things have to be here because @@ -363,20 +360,26 @@ data ExprSize = TooBig -- tup = (a_1, ..., a_99) -- x = case tup of ... -- +mkSizeIs :: FastInt -> FastInt -> Bag (Id, Int) -> FastInt -> ExprSize mkSizeIs max n xs d | (n -# d) ># max = TooBig | otherwise = SizeIs n xs d +maxSize :: ExprSize -> ExprSize -> ExprSize maxSize TooBig _ = TooBig maxSize _ TooBig = TooBig maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1 | otherwise = s2 -sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0) -sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0) -sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0) +sizeZero, sizeOne :: ExprSize +sizeN :: Int -> ExprSize +conSizeN :: DataCon ->Int -> ExprSize + +sizeZero = SizeIs (_ILIT(0)) emptyBag (_ILIT(0)) +sizeOne = SizeIs (_ILIT(1)) emptyBag (_ILIT(0)) +sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT(0)) conSizeN dc n - | isUnboxedTupleCon dc = SizeIs (_ILIT 0) emptyBag (iUnbox n +# _ILIT 1) - | otherwise = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1) + | 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, @@ -388,6 +391,7 @@ conSizeN dc n -- f x y z = case op# x y z of { s -> (# s, () #) } -- and f wasn't getting inlined +primOpSize :: PrimOp -> Int -> ExprSize primOpSize op n_args | not (primOpIsDupable op) = sizeN opt_UF_DearOp | not (primOpOutOfLine op) = sizeN (2 - n_args) @@ -402,7 +406,8 @@ primOpSize op n_args -- and there's a good chance it'll get inlined back into C's RHS. Urgh! | otherwise = sizeOne -buildSize = SizeIs (-2#) emptyBag 4# +buildSize :: ExprSize +buildSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(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 @@ -410,16 +415,19 @@ buildSize = SizeIs (-2#) emptyBag 4# -- build is saturated (it usually is). The "-2" discounts for the \c n, -- The "4" is rather arbitrary. -augmentSize = SizeIs (-2#) emptyBag 4# +augmentSize :: ExprSize +augmentSize = SizeIs (_ILIT(-2)) emptyBag (_ILIT(4)) -- Ditto (augment t (\cn -> e) ys) should cost only the cost of -- e plus ys. The -2 accounts for the \cn - -nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0# -nukeScrutDiscount TooBig = TooBig + +nukeScrutDiscount :: ExprSize -> ExprSize +nukeScrutDiscount (SizeIs n vs _) = SizeIs n vs (_ILIT(0)) +nukeScrutDiscount TooBig = TooBig -- When we return a lambda, give a discount if it's used (applied) -lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) } -lamScrutDiscount TooBig = TooBig +lamScrutDiscount :: ExprSize -> ExprSize +lamScrutDiscount (SizeIs n vs _) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) } +lamScrutDiscount TooBig = TooBig \end{code} @@ -460,20 +468,20 @@ Just the same as smallEnoughToInline, except that it has no actual arguments. \begin{code} couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of - UnfoldNever -> False - other -> True + UnfoldNever -> False + _ -> True certainlyWillInline :: Unfolding -> Bool -- Sees if the unfolding is pretty certain to inline certainlyWillInline (CoreUnfolding _ _ _ is_cheap (UnfoldIfGoodArgs n_vals _ size _)) = is_cheap && size - (n_vals +1) <= opt_UF_UseThreshold -certainlyWillInline other +certainlyWillInline _ = False smallEnoughToInline :: Unfolding -> Bool smallEnoughToInline (CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ _ size _)) = size <= opt_UF_UseThreshold -smallEnoughToInline other +smallEnoughToInline _ = False \end{code} @@ -503,15 +511,32 @@ StrictAnal.addStrictnessInfoToTopId callSiteInline :: DynFlags -> Bool -- True <=> the Id can be inlined -> Id -- The Id + -> Bool -- True if there are are no arguments at all (incl type args) -> [Bool] -- One for each value arg; True if it is interesting - -> Bool -- True <=> continuation is interesting + -> CallCtxt -- True <=> continuation is interesting -> Maybe CoreExpr -- Unfolding, if any -callSiteInline dflags active_inline id arg_infos interesting_cont +data CallCtxt = BoringCtxt + + | ArgCtxt Bool -- We're somewhere in the RHS of function with rules + -- => be keener to inline + Int -- We *are* the argument of a function with this arg discount + -- => be keener to inline + -- INVARIANT: ArgCtxt False 0 ==> BoringCtxt + + | CaseCtxt -- We're the scrutinee of a case + -- that decomposes its scrutinee + +instance Outputable CallCtxt where + ppr BoringCtxt = ptext SLIT("BoringCtxt") + ppr (ArgCtxt _ _) = ptext SLIT("ArgCtxt") + ppr CaseCtxt = ptext SLIT("CaseCtxt") + +callSiteInline dflags active_inline id lone_variable arg_infos cont_info = case idUnfolding id of { NoUnfolding -> Nothing ; - OtherCon cs -> Nothing ; + OtherCon _ -> Nothing ; CompulsoryUnfolding unf_template -> Just unf_template ; -- CompulsoryUnfolding => there is no top-level binding @@ -528,9 +553,7 @@ callSiteInline dflags active_inline id arg_infos interesting_cont n_val_args = length arg_infos - yes_or_no - | not active_inline = False - | otherwise = is_cheap && consider_safe False + yes_or_no = active_inline && is_cheap && consider_safe -- We consider even the once-in-one-branch -- occurrences, because they won't all have been -- caught by preInlineUnconditionally. In particular, @@ -539,14 +562,13 @@ callSiteInline dflags active_inline id arg_infos interesting_cont -- pre-inline will not have inlined it for fear of -- invalidating the occurrence info in the rhs. - consider_safe once + consider_safe -- consider_safe decides whether it's a good idea to -- inline something, given that there's no -- work-duplication issue (the caller checks that). = case guidance of UnfoldNever -> False UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount - | enough_args && size <= (n_vals_wanted + 1) -- Inline unconditionally if there no size increase -- Size of call is n_vals_wanted (+1 for the function) @@ -556,43 +578,47 @@ callSiteInline dflags active_inline id arg_infos interesting_cont -> some_benefit && small_enough where - some_benefit = or arg_infos || really_interesting_cont || - (not is_top && ({- once || -} (n_vals_wanted > 0 && enough_args))) - -- [was (once && not in_lam)] - -- If it occurs more than once, there must be - -- something interesting about some argument, or the - -- result context, to make it worth inlining - -- - -- If a function has a nested defn we also record - -- some-benefit, on the grounds that we are often able - -- to eliminate the binding, and hence the allocation, - -- for the function altogether; this is good for join - -- points. But this only makes sense for *functions*; - -- inlining a constructor doesn't help allocation - -- unless the result is scrutinised. UNLESS the - -- constructor occurs just once, albeit possibly in - -- multiple case branches. Then inlining it doesn't - -- increase allocation, but it does increase the - -- chance that the constructor won't be allocated at - -- all in the branches that don't use it. - - enough_args = n_val_args >= n_vals_wanted - really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args - | n_val_args == n_vals_wanted = interesting_cont - | otherwise = True -- Extra args + enough_args = n_val_args >= n_vals_wanted + + some_benefit = or arg_infos || really_interesting_cont + -- There must be something interesting + -- about some argument, or the result + -- context, to make it worth inlining + + really_interesting_cont + | n_val_args < n_vals_wanted = False -- Too few args + | n_val_args == n_vals_wanted = interesting_saturated_call + | otherwise = True -- Extra args -- really_interesting_cont tells if the result of the -- call is in an interesting context. + interesting_saturated_call + = case cont_info of + BoringCtxt -> not is_top && n_vals_wanted > 0 -- Note [Nested functions] + CaseCtxt -> not lone_variable || not is_value -- Note [Lone variables] + ArgCtxt {} -> True + -- Was: n_vals_wanted > 0; but see test eyeball/inline1.hs + small_enough = (size - discount) <= opt_UF_UseThreshold - discount = computeDiscount n_vals_wanted arg_discounts res_discount - arg_infos really_interesting_cont + discount = computeDiscount n_vals_wanted arg_discounts + res_discount' arg_infos + res_discount' = case cont_info of + BoringCtxt -> 0 + CaseCtxt -> res_discount + ArgCtxt _ _ -> 4 `min` res_discount + -- res_discount can be very large when a function returns + -- construtors; but we only want to invoke that large discount + -- when there's a case continuation. + -- Otherwise we, rather arbitrarily, threshold it. Yuk. + -- But we want to aovid inlining large functions that return + -- constructors into contexts that are simply "interesting" in if dopt Opt_D_dump_inlinings dflags then pprTrace "Considering inlining" (ppr id <+> vcat [text "active:" <+> ppr active_inline, text "arg infos" <+> ppr arg_infos, - text "interesting continuation" <+> ppr interesting_cont, + text "interesting continuation" <+> ppr cont_info, text "is value:" <+> ppr is_value, text "is cheap:" <+> ppr is_cheap, text "guidance" <+> ppr guidance, @@ -601,9 +627,78 @@ callSiteInline dflags active_inline id arg_infos interesting_cont else result } +\end{code} + +Note [Nested functions] +~~~~~~~~~~~~~~~~~~~~~~~ +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. + +Note [Lone variables] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +The "lone-variable" case is important. I spent ages messing about +with unsatisfactory varaints, but this is nice. The idea is that if a +variable appears all alone + as an arg of lazy fn, or rhs Stop + as scrutinee of a case Select + as arg of a strict fn ArgOf +AND + it is bound to a value +then we should not inline it (unless there is some other reason, +e.g. is is the sole occurrence). That is what is happening at +the use of 'lone_variable' in 'interesting_saturated_call'. + +Why? At least in the case-scrutinee situation, turning + let x = (a,b) in case x of y -> ... +into + let x = (a,b) in case (a,b) of y -> ... +and thence to + let x = (a,b) in let y = (a,b) in ... +is bad if the binding for x will remain. + +Another example: I discovered that strings +were getting inlined straight back into applications of 'error' +because the latter is strict. + s = "foo" + f = \x -> ...(error s)... + +Fundamentally such contexts should not encourage inlining because the +context can ``see'' the unfolding of the variable (e.g. case or a +RULE) so there's no gain. If the thing is bound to a value. + +However, watch out: + + * Consider this: + foo = _inline_ (\n. [n]) + bar = _inline_ (foo 20) + baz = \n. case bar of { (m:_) -> m + n } + Here we really want to inline 'bar' so that we can inline 'foo' + and the whole thing unravels as it should obviously do. This is + important: in the NDP project, 'bar' generates a closure data + structure rather than a list. + + * Even a type application or coercion isn't a lone variable. + Consider + case $fMonadST @ RealWorld of { :DMonad a b c -> c } + We had better inline that sucker! The case won't see through it. + + For now, I'm treating treating a variable applied to types + in a *lazy* context "lone". The motivating example was + f = /\a. \x. BIG + g = /\a. \y. h (f a) + There's no advantage in inlining f here, and perhaps + a significant disadvantage. Hence some_val_args in the Stop case -computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int -computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used +\begin{code} +computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Int +computeDiscount n_vals_wanted arg_discounts result_discount arg_infos -- 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 @@ -625,8 +720,4 @@ computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used 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}