{- (c) The University of Glasgow 2006 (c) The AQUA Project, Glasgow University, 1994-1998 Core-syntax unfoldings Unfoldings (which can travel across module boundaries) are in Core syntax (namely @CoreExpr@s). 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 @CoreUnfolding@ unfolding, you will find, unsurprisingly, a Core expression. -} {-# LANGUAGE CPP #-} module CoreUnfold ( Unfolding, UnfoldingGuidance, -- Abstract types noUnfolding, mkImplicitUnfolding, mkUnfolding, mkCoreUnfolding, mkTopUnfolding, mkSimpleUnfolding, mkWorkerUnfolding, mkInlineUnfolding, mkInlineUnfoldingWithArity, mkInlinableUnfolding, mkWwInlineRule, mkCompulsoryUnfolding, mkDFunUnfolding, specUnfolding, ArgSummary(..), couldBeSmallEnoughToInline, inlineBoringOk, certainlyWillInline, smallEnoughToInline, callSiteInline, CallCtxt(..), -- Reexport from CoreSubst (it only live there so it can be used -- by the Very Simple Optimiser) exprIsConApp_maybe, exprIsLiteral_maybe ) where #include "HsVersions.h" import GhcPrelude import DynFlags import CoreSyn import OccurAnal ( occurAnalyseExpr_NoBinderSwap ) import CoreOpt import CoreArity ( manifestArity ) import CoreUtils import Id import Demand ( isBottomingSig ) import DataCon import Literal import PrimOp import IdInfo import BasicTypes ( Arity, InlineSpec(..), inlinePragmaSpec ) import Type import PrelNames import TysPrim ( realWorldStatePrimTy ) import Bag import Util import Outputable import ForeignCall import Name import qualified Data.ByteString as BS import Data.List {- ************************************************************************ * * \subsection{Making unfoldings} * * ************************************************************************ -} mkTopUnfolding :: DynFlags -> Bool -> CoreExpr -> Unfolding mkTopUnfolding :: DynFlags -> Bool -> CoreExpr -> Unfolding mkTopUnfolding DynFlags dflags Bool is_bottoming CoreExpr rhs = DynFlags -> UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding mkUnfolding DynFlags dflags UnfoldingSource InlineRhs Bool True Bool is_bottoming CoreExpr rhs mkImplicitUnfolding :: DynFlags -> CoreExpr -> Unfolding -- For implicit Ids, do a tiny bit of optimising first mkImplicitUnfolding :: DynFlags -> CoreExpr -> Unfolding mkImplicitUnfolding DynFlags dflags CoreExpr expr = DynFlags -> Bool -> CoreExpr -> Unfolding mkTopUnfolding DynFlags dflags Bool False (DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags CoreExpr expr) -- Note [Top-level flag on inline rules] -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- Slight hack: note that mk_inline_rules conservatively sets the -- top-level flag to True. It gets set more accurately by the simplifier -- Simplify.simplUnfolding. mkSimpleUnfolding :: DynFlags -> CoreExpr -> Unfolding mkSimpleUnfolding :: DynFlags -> CoreExpr -> Unfolding mkSimpleUnfolding DynFlags dflags CoreExpr rhs = DynFlags -> UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding mkUnfolding DynFlags dflags UnfoldingSource InlineRhs Bool False Bool False CoreExpr rhs mkDFunUnfolding :: [Var] -> DataCon -> [CoreExpr] -> Unfolding mkDFunUnfolding :: [Var] -> DataCon -> [CoreExpr] -> Unfolding mkDFunUnfolding [Var] bndrs DataCon con [CoreExpr] ops = DFunUnfolding :: [Var] -> DataCon -> [CoreExpr] -> Unfolding DFunUnfolding { df_bndrs :: [Var] df_bndrs = [Var] bndrs , df_con :: DataCon df_con = DataCon con , df_args :: [CoreExpr] df_args = (CoreExpr -> CoreExpr) -> [CoreExpr] -> [CoreExpr] forall a b. (a -> b) -> [a] -> [b] map CoreExpr -> CoreExpr occurAnalyseExpr_NoBinderSwap [CoreExpr] ops } -- See Note [Occurrence analysis of unfoldings] mkWwInlineRule :: DynFlags -> CoreExpr -> Arity -> Unfolding mkWwInlineRule :: DynFlags -> CoreExpr -> Arity -> Unfolding mkWwInlineRule DynFlags dflags CoreExpr expr Arity arity = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource InlineStable Bool True (DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags CoreExpr expr) (UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_arity :: Arity ug_arity = Arity arity, ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk , ug_boring_ok :: Bool ug_boring_ok = Bool boringCxtNotOk }) mkCompulsoryUnfolding :: CoreExpr -> Unfolding mkCompulsoryUnfolding :: CoreExpr -> Unfolding mkCompulsoryUnfolding CoreExpr expr -- Used for things that absolutely must be unfolded = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource InlineCompulsory Bool True (DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags unsafeGlobalDynFlags CoreExpr expr) (UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_arity :: Arity ug_arity = Arity 0 -- Arity of unfolding doesn't matter , ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk, ug_boring_ok :: Bool ug_boring_ok = Bool boringCxtOk }) mkWorkerUnfolding :: DynFlags -> (CoreExpr -> CoreExpr) -> Unfolding -> Unfolding -- See Note [Worker-wrapper for INLINABLE functions] in WorkWrap mkWorkerUnfolding :: DynFlags -> (CoreExpr -> CoreExpr) -> Unfolding -> Unfolding mkWorkerUnfolding DynFlags dflags CoreExpr -> CoreExpr work_fn (CoreUnfolding { uf_src :: Unfolding -> UnfoldingSource uf_src = UnfoldingSource src, uf_tmpl :: Unfolding -> CoreExpr uf_tmpl = CoreExpr tmpl , uf_is_top :: Unfolding -> Bool uf_is_top = Bool top_lvl }) | UnfoldingSource -> Bool isStableSource UnfoldingSource src = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource src Bool top_lvl CoreExpr new_tmpl UnfoldingGuidance guidance where new_tmpl :: CoreExpr new_tmpl = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags (CoreExpr -> CoreExpr work_fn CoreExpr tmpl) guidance :: UnfoldingGuidance guidance = DynFlags -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance DynFlags dflags Bool False CoreExpr new_tmpl mkWorkerUnfolding DynFlags _ CoreExpr -> CoreExpr _ Unfolding _ = Unfolding noUnfolding -- | Make an unfolding that may be used unsaturated -- (ug_unsat_ok = unSaturatedOk) and that is reported as having its -- manifest arity (the number of outer lambdas applications will -- resolve before doing any work). mkInlineUnfolding :: CoreExpr -> Unfolding mkInlineUnfolding :: CoreExpr -> Unfolding mkInlineUnfolding CoreExpr expr = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource InlineStable Bool True -- Note [Top-level flag on inline rules] CoreExpr expr' UnfoldingGuidance guide where expr' :: CoreExpr expr' = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags unsafeGlobalDynFlags CoreExpr expr guide :: UnfoldingGuidance guide = UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_arity :: Arity ug_arity = CoreExpr -> Arity manifestArity CoreExpr expr' , ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk , ug_boring_ok :: Bool ug_boring_ok = Bool boring_ok } boring_ok :: Bool boring_ok = CoreExpr -> Bool inlineBoringOk CoreExpr expr' -- | Make an unfolding that will be used once the RHS has been saturated -- to the given arity. mkInlineUnfoldingWithArity :: Arity -> CoreExpr -> Unfolding mkInlineUnfoldingWithArity :: Arity -> CoreExpr -> Unfolding mkInlineUnfoldingWithArity Arity arity CoreExpr expr = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource InlineStable Bool True -- Note [Top-level flag on inline rules] CoreExpr expr' UnfoldingGuidance guide where expr' :: CoreExpr expr' = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags unsafeGlobalDynFlags CoreExpr expr guide :: UnfoldingGuidance guide = UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_arity :: Arity ug_arity = Arity arity , ug_unsat_ok :: Bool ug_unsat_ok = Bool needSaturated , ug_boring_ok :: Bool ug_boring_ok = Bool boring_ok } -- See Note [INLINE pragmas and boring contexts] as to why we need to look -- at the arity here. boring_ok :: Bool boring_ok | Arity arity Arity -> Arity -> Bool forall a. Eq a => a -> a -> Bool == Arity 0 = Bool True | Bool otherwise = CoreExpr -> Bool inlineBoringOk CoreExpr expr' mkInlinableUnfolding :: DynFlags -> CoreExpr -> Unfolding mkInlinableUnfolding :: DynFlags -> CoreExpr -> Unfolding mkInlinableUnfolding DynFlags dflags CoreExpr expr = DynFlags -> UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding mkUnfolding DynFlags dflags UnfoldingSource InlineStable Bool False Bool False CoreExpr expr' where expr' :: CoreExpr expr' = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags CoreExpr expr specUnfolding :: DynFlags -> [Var] -> (CoreExpr -> CoreExpr) -> [CoreArg] -- LHS arguments in the RULE -> Unfolding -> Unfolding -- See Note [Specialising unfoldings] -- specUnfolding spec_bndrs spec_args unf -- = \spec_bndrs. unf spec_args -- specUnfolding :: DynFlags -> [Var] -> (CoreExpr -> CoreExpr) -> [CoreExpr] -> Unfolding -> Unfolding specUnfolding DynFlags dflags [Var] spec_bndrs CoreExpr -> CoreExpr spec_app [CoreExpr] rule_lhs_args df :: Unfolding df@(DFunUnfolding { df_bndrs :: Unfolding -> [Var] df_bndrs = [Var] old_bndrs, df_con :: Unfolding -> DataCon df_con = DataCon con, df_args :: Unfolding -> [CoreExpr] df_args = [CoreExpr] args }) = ASSERT2( rule_lhs_args `equalLength` old_bndrs , ppr df $$ ppr rule_lhs_args ) -- For this ASSERT see Note [DFunUnfoldings] in GHC.Core.Opt.Specialise [Var] -> DataCon -> [CoreExpr] -> Unfolding mkDFunUnfolding [Var] spec_bndrs DataCon con ((CoreExpr -> CoreExpr) -> [CoreExpr] -> [CoreExpr] forall a b. (a -> b) -> [a] -> [b] map CoreExpr -> CoreExpr spec_arg [CoreExpr] args) -- For DFunUnfoldings we transform -- \obs. MkD <op1> ... <opn> -- to -- \sbs. MkD ((\obs. <op1>) spec_args) ... ditto <opn> where spec_arg :: CoreExpr -> CoreExpr spec_arg CoreExpr arg = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags (CoreExpr -> CoreExpr) -> CoreExpr -> CoreExpr forall a b. (a -> b) -> a -> b $ CoreExpr -> CoreExpr spec_app ([Var] -> CoreExpr -> CoreExpr forall b. [b] -> Expr b -> Expr b mkLams [Var] old_bndrs CoreExpr arg) -- The beta-redexes created by spec_app will be -- simplified away by simplOptExpr specUnfolding DynFlags dflags [Var] spec_bndrs CoreExpr -> CoreExpr spec_app [CoreExpr] rule_lhs_args (CoreUnfolding { uf_src :: Unfolding -> UnfoldingSource uf_src = UnfoldingSource src, uf_tmpl :: Unfolding -> CoreExpr uf_tmpl = CoreExpr tmpl , uf_is_top :: Unfolding -> Bool uf_is_top = Bool top_lvl , uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfoldingGuidance old_guidance }) | UnfoldingSource -> Bool isStableSource UnfoldingSource src -- See Note [Specialising unfoldings] , UnfWhen { ug_arity :: UnfoldingGuidance -> Arity ug_arity = Arity old_arity } <- UnfoldingGuidance old_guidance = UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource src Bool top_lvl CoreExpr new_tmpl (UnfoldingGuidance old_guidance { ug_arity :: Arity ug_arity = Arity old_arity Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity arity_decrease }) where new_tmpl :: CoreExpr new_tmpl = DynFlags -> CoreExpr -> CoreExpr simpleOptExpr DynFlags dflags (CoreExpr -> CoreExpr) -> CoreExpr -> CoreExpr forall a b. (a -> b) -> a -> b $ [Var] -> CoreExpr -> CoreExpr forall b. [b] -> Expr b -> Expr b mkLams [Var] spec_bndrs (CoreExpr -> CoreExpr) -> CoreExpr -> CoreExpr forall a b. (a -> b) -> a -> b $ CoreExpr -> CoreExpr spec_app CoreExpr tmpl -- The beta-redexes created by spec_app -- will besimplified away by simplOptExpr arity_decrease :: Arity arity_decrease = (CoreExpr -> Bool) -> [CoreExpr] -> Arity forall a. (a -> Bool) -> [a] -> Arity count CoreExpr -> Bool forall b. Expr b -> Bool isValArg [CoreExpr] rule_lhs_args Arity -> Arity -> Arity forall a. Num a => a -> a -> a - (Var -> Bool) -> [Var] -> Arity forall a. (a -> Bool) -> [a] -> Arity count Var -> Bool isId [Var] spec_bndrs specUnfolding DynFlags _ [Var] _ CoreExpr -> CoreExpr _ [CoreExpr] _ Unfolding _ = Unfolding noUnfolding {- Note [Specialising unfoldings] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ When we specialise a function for some given type-class arguments, we use specUnfolding to specialise its unfolding. Some important points: * If the original function has a DFunUnfolding, the specialised one must do so too! Otherwise we lose the magic rules that make it interact with ClassOps * There is a bit of hack for INLINABLE functions: f :: Ord a => .... f = <big-rhs> {- INLINABLE f #-} Now if we specialise f, should the specialised version still have an INLINABLE pragma? If it does, we'll capture a specialised copy of <big-rhs> as its unfolding, and that probaby won't inline. But if we don't, the specialised version of <big-rhs> might be small enough to inline at a call site. This happens with Control.Monad.liftM3, and can cause a lot more allocation as a result (nofib n-body shows this). Moreover, keeping the INLINABLE thing isn't much help, because the specialised function (probaby) isn't overloaded any more. Conclusion: drop the INLINEALE pragma. In practice what this means is: if a stable unfolding has UnfoldingGuidance of UnfWhen, we keep it (so the specialised thing too will always inline) if a stable unfolding has UnfoldingGuidance of UnfIfGoodArgs (which arises from INLINABLE), we discard it Note [Honour INLINE on 0-ary bindings] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Consider x = <expensive> {-# INLINE x #-} f y = ...x... The semantics of an INLINE pragma is inline x at every call site, provided it is saturated; that is, applied to at least as many arguments as appear on the LHS of the Haskell source definition. (This soure-code-derived arity is stored in the `ug_arity` field of the `UnfoldingGuidance`.) In the example, x's ug_arity is 0, so we should inline it at every use site. It's rare to have such an INLINE pragma (usually INLINE Is on functions), but it's occasionally very important (#15578, #15519). In #15519 we had something like x = case (g a b) of I# r -> T r {-# INLINE x #-} f y = ...(h x).... where h is strict. So we got f y = ...(case g a b of I# r -> h (T r))... and that in turn allowed SpecConstr to ramp up performance. How do we deliver on this? By adjusting the ug_boring_ok flag in mkInlineUnfoldingWithArity; see Note [INLINE pragmas and boring contexts] NB: there is a real risk that full laziness will float it right back out again. Consider again x = factorial 200 {-# INLINE x #-} f y = ...x... After inlining we get f y = ...(factorial 200)... but it's entirely possible that full laziness will do lvl23 = factorial 200 f y = ...lvl23... That's a problem for another day. Note [INLINE pragmas and boring contexts] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ An INLINE pragma uses mkInlineUnfoldingWithArity to build the unfolding. That sets the ug_boring_ok flag to False if the function is not tiny (inlineBoringOK), so that even INLINE functions are not inlined in an utterly boring context. E.g. \x y. Just (f y x) Nothing is gained by inlining f here, even if it has an INLINE pragma. But for 0-ary bindings, we want to inline regardless; see Note [Honour INLINE on 0-ary bindings]. I'm a bit worried that it's possible for the same kind of problem to arise for non-0-ary functions too, but let's wait and see. -} mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding -- Occurrence-analyses the expression before capturing it mkCoreUnfolding :: UnfoldingSource -> Bool -> CoreExpr -> UnfoldingGuidance -> Unfolding mkCoreUnfolding UnfoldingSource src Bool top_lvl CoreExpr expr UnfoldingGuidance guidance = CoreUnfolding :: CoreExpr -> UnfoldingSource -> Bool -> Bool -> Bool -> Bool -> Bool -> UnfoldingGuidance -> Unfolding CoreUnfolding { uf_tmpl :: CoreExpr uf_tmpl = CoreExpr -> CoreExpr occurAnalyseExpr_NoBinderSwap CoreExpr expr, -- See Note [Occurrence analysis of unfoldings] uf_src :: UnfoldingSource uf_src = UnfoldingSource src, uf_is_top :: Bool uf_is_top = Bool top_lvl, uf_is_value :: Bool uf_is_value = CoreExpr -> Bool exprIsHNF CoreExpr expr, uf_is_conlike :: Bool uf_is_conlike = CoreExpr -> Bool exprIsConLike CoreExpr expr, uf_is_work_free :: Bool uf_is_work_free = CoreExpr -> Bool exprIsWorkFree CoreExpr expr, uf_expandable :: Bool uf_expandable = CoreExpr -> Bool exprIsExpandable CoreExpr expr, uf_guidance :: UnfoldingGuidance uf_guidance = UnfoldingGuidance guidance } mkUnfolding :: DynFlags -> UnfoldingSource -> Bool -- Is top-level -> Bool -- Definitely a bottoming binding -- (only relevant for top-level bindings) -> CoreExpr -> Unfolding -- Calculates unfolding guidance -- Occurrence-analyses the expression before capturing it mkUnfolding :: DynFlags -> UnfoldingSource -> Bool -> Bool -> CoreExpr -> Unfolding mkUnfolding DynFlags dflags UnfoldingSource src Bool is_top_lvl Bool is_bottoming CoreExpr expr = CoreUnfolding :: CoreExpr -> UnfoldingSource -> Bool -> Bool -> Bool -> Bool -> Bool -> UnfoldingGuidance -> Unfolding CoreUnfolding { uf_tmpl :: CoreExpr uf_tmpl = CoreExpr -> CoreExpr occurAnalyseExpr_NoBinderSwap CoreExpr expr, -- See Note [Occurrence analysis of unfoldings] uf_src :: UnfoldingSource uf_src = UnfoldingSource src, uf_is_top :: Bool uf_is_top = Bool is_top_lvl, uf_is_value :: Bool uf_is_value = CoreExpr -> Bool exprIsHNF CoreExpr expr, uf_is_conlike :: Bool uf_is_conlike = CoreExpr -> Bool exprIsConLike CoreExpr expr, uf_expandable :: Bool uf_expandable = CoreExpr -> Bool exprIsExpandable CoreExpr expr, uf_is_work_free :: Bool uf_is_work_free = CoreExpr -> Bool exprIsWorkFree CoreExpr expr, uf_guidance :: UnfoldingGuidance uf_guidance = UnfoldingGuidance guidance } where is_top_bottoming :: Bool is_top_bottoming = Bool is_top_lvl Bool -> Bool -> Bool && Bool is_bottoming guidance :: UnfoldingGuidance guidance = DynFlags -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance DynFlags dflags Bool is_top_bottoming CoreExpr expr -- NB: *not* (calcUnfoldingGuidance (occurAnalyseExpr_NoBinderSwap expr))! -- See Note [Calculate unfolding guidance on the non-occ-anal'd expression] {- Note [Occurrence analysis of unfoldings] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ We do occurrence-analysis of unfoldings once and for all, when the unfolding is built, rather than each time we inline them. But given this decision it's vital that we do *always* do it. Consider this unfolding \x -> letrec { f = ...g...; g* = f } in body where g* is (for some strange reason) the loop breaker. If we don't occ-anal it when reading it in, we won't mark g as a loop breaker, and we may inline g entirely in body, dropping its binding, and leaving the occurrence in f out of scope. This happened in #8892, where the unfolding in question was a DFun unfolding. But more generally, the simplifier is designed on the basis that it is looking at occurrence-analysed expressions, so better ensure that they acutally are. We use occurAnalyseExpr_NoBinderSwap instead of occurAnalyseExpr; see Note [No binder swap in unfoldings]. Note [No binder swap in unfoldings] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The binder swap can temporarily violate Core Lint, by assinging a LocalId binding to a GlobalId. For example, if A.foo{r872} is a GlobalId with unique r872, then case A.foo{r872} of bar { K x -> ...(A.foo{r872})... } gets transformed to case A.foo{r872} of bar { K x -> let foo{r872} = bar in ...(A.foo{r872})... This is usually not a problem, because the simplifier will transform this to: case A.foo{r872} of bar { K x -> ...(bar)... However, after occurrence analysis but before simplification, this extra 'let' violates the Core Lint invariant that we do not have local 'let' bindings for GlobalIds. That seems (just) tolerable for the occurrence analysis that happens just before the Simplifier, but not for unfoldings, which are Linted independently. As a quick workaround, we disable binder swap in this module. See #16288 and #16296 for further plans. Note [Calculate unfolding guidance on the non-occ-anal'd expression] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Notice that we give the non-occur-analysed expression to calcUnfoldingGuidance. In some ways it'd be better to occur-analyse first; for example, sometimes during simplification, there's a large let-bound thing which has been substituted, and so is now dead; so 'expr' contains two copies of the thing while the occurrence-analysed expression doesn't. Nevertheless, we *don't* and *must not* occ-analyse before computing the size because a) The size computation bales out after a while, whereas occurrence analysis does not. b) Residency increases sharply if you occ-anal first. I'm not 100% sure why, but it's a large effect. Compiling Cabal went from residency of 534M to over 800M with this one change. This can occasionally mean that the guidance is very pessimistic; it gets fixed up next round. And it should be rare, because large let-bound things that are dead are usually caught by preInlineUnconditionally ************************************************************************ * * \subsection{The UnfoldingGuidance type} * * ************************************************************************ -} inlineBoringOk :: CoreExpr -> Bool -- See Note [INLINE for small functions] -- True => the result of inlining the expression is -- no bigger than the expression itself -- eg (\x y -> f y x) -- This is a quick and dirty version. It doesn't attempt -- to deal with (\x y z -> x (y z)) -- The really important one is (x `cast` c) inlineBoringOk :: CoreExpr -> Bool inlineBoringOk CoreExpr e = Arity -> CoreExpr -> Bool go Arity 0 CoreExpr e where go :: Int -> CoreExpr -> Bool go :: Arity -> CoreExpr -> Bool go Arity credit (Lam Var x CoreExpr e) | Var -> Bool isId Var x = Arity -> CoreExpr -> Bool go (Arity creditArity -> Arity -> Arity forall a. Num a => a -> a -> a +Arity 1) CoreExpr e | Bool otherwise = Arity -> CoreExpr -> Bool go Arity credit CoreExpr e -- See Note [Count coercion arguments in boring contexts] go Arity credit (App CoreExpr f (Type {})) = Arity -> CoreExpr -> Bool go Arity credit CoreExpr f go Arity credit (App CoreExpr f CoreExpr a) | Arity credit Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 , CoreExpr -> Bool exprIsTrivial CoreExpr a = Arity -> CoreExpr -> Bool go (Arity creditArity -> Arity -> Arity forall a. Num a => a -> a -> a -Arity 1) CoreExpr f go Arity credit (Tick Tickish Var _ CoreExpr e) = Arity -> CoreExpr -> Bool go Arity credit CoreExpr e -- dubious go Arity credit (Cast CoreExpr e Coercion _) = Arity -> CoreExpr -> Bool go Arity credit CoreExpr e go Arity _ (Var {}) = Bool boringCxtOk go Arity _ CoreExpr _ = Bool boringCxtNotOk calcUnfoldingGuidance :: DynFlags -> Bool -- Definitely a top-level, bottoming binding -> CoreExpr -- Expression to look at -> UnfoldingGuidance calcUnfoldingGuidance :: DynFlags -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance DynFlags dflags Bool is_top_bottoming (Tick Tickish Var t CoreExpr expr) | Bool -> Bool not (Tickish Var -> Bool forall id. Tickish id -> Bool tickishIsCode Tickish Var t) -- non-code ticks don't matter for unfolding = DynFlags -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance DynFlags dflags Bool is_top_bottoming CoreExpr expr calcUnfoldingGuidance DynFlags dflags Bool is_top_bottoming CoreExpr expr = case DynFlags -> Arity -> [Var] -> CoreExpr -> ExprSize sizeExpr DynFlags dflags Arity bOMB_OUT_SIZE [Var] val_bndrs CoreExpr body of ExprSize TooBig -> UnfoldingGuidance UnfNever SizeIs Arity size Bag (Var, Arity) cased_bndrs Arity scrut_discount | CoreExpr -> Arity -> Arity -> Bool uncondInline CoreExpr expr Arity n_val_bndrs Arity size -> UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk , ug_boring_ok :: Bool ug_boring_ok = Bool boringCxtOk , ug_arity :: Arity ug_arity = Arity n_val_bndrs } -- Note [INLINE for small functions] | Bool is_top_bottoming -> UnfoldingGuidance UnfNever -- See Note [Do not inline top-level bottoming functions] | Bool otherwise -> UnfIfGoodArgs :: [Arity] -> Arity -> Arity -> UnfoldingGuidance UnfIfGoodArgs { ug_args :: [Arity] ug_args = (Var -> Arity) -> [Var] -> [Arity] forall a b. (a -> b) -> [a] -> [b] map (Bag (Var, Arity) -> Var -> Arity mk_discount Bag (Var, Arity) cased_bndrs) [Var] val_bndrs , ug_size :: Arity ug_size = Arity size , ug_res :: Arity ug_res = Arity scrut_discount } where ([Var] bndrs, CoreExpr body) = CoreExpr -> ([Var], CoreExpr) forall b. Expr b -> ([b], Expr b) collectBinders CoreExpr expr bOMB_OUT_SIZE :: Arity bOMB_OUT_SIZE = DynFlags -> Arity ufCreationThreshold DynFlags dflags -- Bomb out if size gets bigger than this val_bndrs :: [Var] val_bndrs = (Var -> Bool) -> [Var] -> [Var] forall a. (a -> Bool) -> [a] -> [a] filter Var -> Bool isId [Var] bndrs n_val_bndrs :: Arity n_val_bndrs = [Var] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [Var] val_bndrs mk_discount :: Bag (Id,Int) -> Id -> Int mk_discount :: Bag (Var, Arity) -> Var -> Arity mk_discount Bag (Var, Arity) cbs Var bndr = (Arity -> (Var, Arity) -> Arity) -> Arity -> Bag (Var, Arity) -> Arity forall (t :: * -> *) b a. Foldable t => (b -> a -> b) -> b -> t a -> b foldl' Arity -> (Var, Arity) -> Arity combine Arity 0 Bag (Var, Arity) cbs where combine :: Arity -> (Var, Arity) -> Arity combine Arity acc (Var bndr', Arity disc) | Var bndr Var -> Var -> Bool forall a. Eq a => a -> a -> Bool == Var bndr' = Arity acc Arity -> Arity -> Arity `plus_disc` Arity disc | Bool otherwise = Arity acc plus_disc :: Int -> Int -> Int plus_disc :: Arity -> Arity -> Arity plus_disc | Type -> Bool isFunTy (Var -> Type idType Var bndr) = Arity -> Arity -> Arity forall a. Ord a => a -> a -> a max | Bool otherwise = Arity -> Arity -> Arity forall a. Num a => a -> a -> a (+) -- See Note [Function and non-function discounts] {- Note [Computing the size of an expression] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The basic idea of sizeExpr is obvious enough: count nodes. But getting the heuristics right has taken a long time. Here's the basic strategy: * Variables, literals: 0 (Exception for string literals, see litSize.) * Function applications (f e1 .. en): 1 + #value args * Constructor applications: 1, regardless of #args * Let(rec): 1 + size of components * Note, cast: 0 Examples Size Term -------------- 0 42# 0 x 0 True 2 f x 1 Just x 4 f (g x) Notice that 'x' counts 0, while (f x) counts 2. That's deliberate: there's a function call to account for. Notice also that constructor applications are very cheap, because exposing them to a caller is so valuable. [25/5/11] All sizes are now multiplied by 10, except for primops (which have sizes like 1 or 4. This makes primops look fantastically cheap, and seems to be almost unversally beneficial. Done partly as a result of #4978. Note [Do not inline top-level bottoming functions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The FloatOut pass has gone to some trouble to float out calls to 'error' and similar friends. See Note [Bottoming floats] in SetLevels. Do not re-inline them! But we *do* still inline if they are very small (the uncondInline stuff). Note [INLINE for small functions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Consider {-# INLINE f #-} f x = Just x g y = f y Then f's RHS is no larger than its LHS, so we should inline it into even the most boring context. In general, f the function is sufficiently small that its body is as small as the call itself, the inline unconditionally, regardless of how boring the context is. Things to note: (1) We inline *unconditionally* if inlined thing is smaller (using sizeExpr) than the thing it's replacing. Notice that (f x) --> (g 3) -- YES, unconditionally (f x) --> x : [] -- YES, *even though* there are two -- arguments to the cons x --> g 3 -- NO x --> Just v -- NO It's very important not to unconditionally replace a variable by a non-atomic term. (2) We do this even if the thing isn't saturated, else we end up with the silly situation that f x y = x ...map (f 3)... doesn't inline. Even in a boring context, inlining without being saturated will give a lambda instead of a PAP, and will be more efficient at runtime. (3) However, when the function's arity > 0, we do insist that it has at least one value argument at the call site. (This check is made in the UnfWhen case of callSiteInline.) Otherwise we find this: f = /\a \x:a. x d = /\b. MkD (f b) If we inline f here we get d = /\b. MkD (\x:b. x) and then prepareRhs floats out the argument, abstracting the type variables, so we end up with the original again! (4) We must be much more cautious about arity-zero things. Consider let x = y +# z in ... In *size* terms primops look very small, because the generate a single instruction, but we do not want to unconditionally replace every occurrence of x with (y +# z). So we only do the unconditional-inline thing for *trivial* expressions. NB: you might think that PostInlineUnconditionally would do this but it doesn't fire for top-level things; see SimplUtils Note [Top level and postInlineUnconditionally] Note [Count coercion arguments in boring contexts] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In inlineBoringOK, we ignore type arguments when deciding whether an expression is okay to inline into boring contexts. This is good, since if we have a definition like let y = x @Int in f y y there’s no reason not to inline y at both use sites — no work is actually duplicated. It may seem like the same reasoning applies to coercion arguments, and indeed, in #17182 we changed inlineBoringOK to treat coercions the same way. However, this isn’t a good idea: unlike type arguments, which have no runtime representation, coercion arguments *do* have a runtime representation (albeit the zero-width VoidRep, see Note [Coercion tokens] in CoreToStg.hs). This caused trouble in #17787 for DataCon wrappers for nullary GADT constructors: the wrappers would be inlined and each use of the constructor would lead to a separate allocation instead of just sharing the wrapper closure. The solution: don’t ignore coercion arguments after all. -} uncondInline :: CoreExpr -> Arity -> Int -> Bool -- Inline unconditionally if there no size increase -- Size of call is arity (+1 for the function) -- See Note [INLINE for small functions] uncondInline :: CoreExpr -> Arity -> Arity -> Bool uncondInline CoreExpr rhs Arity arity Arity size | Arity arity Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 = Arity size Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool <= Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity arity Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 1) -- See Note [INLINE for small functions] (1) | Bool otherwise = CoreExpr -> Bool exprIsTrivial CoreExpr rhs -- See Note [INLINE for small functions] (4) sizeExpr :: DynFlags -> Int -- Bomb out if it gets bigger than this -> [Id] -- Arguments; we're interested in which of these -- get case'd -> CoreExpr -> ExprSize -- Note [Computing the size of an expression] sizeExpr :: DynFlags -> Arity -> [Var] -> CoreExpr -> ExprSize sizeExpr DynFlags dflags Arity bOMB_OUT_SIZE [Var] top_args CoreExpr expr = CoreExpr -> ExprSize size_up CoreExpr expr where size_up :: CoreExpr -> ExprSize size_up (Cast CoreExpr e Coercion _) = CoreExpr -> ExprSize size_up CoreExpr e size_up (Tick Tickish Var _ CoreExpr e) = CoreExpr -> ExprSize size_up CoreExpr e size_up (Type Type _) = ExprSize sizeZero -- Types cost nothing size_up (Coercion Coercion _) = ExprSize sizeZero size_up (Lit Literal lit) = Arity -> ExprSize sizeN (Literal -> Arity litSize Literal lit) size_up (Var Var f) | Var -> Bool isRealWorldId Var f = ExprSize sizeZero -- Make sure we get constructor discounts even -- on nullary constructors | Bool otherwise = Var -> [CoreExpr] -> Arity -> ExprSize size_up_call Var f [] Arity 0 size_up (App CoreExpr fun CoreExpr arg) | CoreExpr -> Bool forall b. Expr b -> Bool isTyCoArg CoreExpr arg = CoreExpr -> ExprSize size_up CoreExpr fun | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr arg ExprSize -> ExprSize -> ExprSize `addSizeNSD` CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr fun [CoreExpr arg] (if CoreExpr -> Bool forall b. Expr b -> Bool isRealWorldExpr CoreExpr arg then Arity 1 else Arity 0) size_up (Lam Var b CoreExpr e) | Var -> Bool isId Var b Bool -> Bool -> Bool && Bool -> Bool not (Var -> Bool isRealWorldId Var b) = DynFlags -> ExprSize -> ExprSize lamScrutDiscount DynFlags dflags (CoreExpr -> ExprSize size_up CoreExpr e ExprSize -> Arity -> ExprSize `addSizeN` Arity 10) | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr e size_up (Let (NonRec Var binder CoreExpr rhs) CoreExpr body) = (Var, CoreExpr) -> ExprSize size_up_rhs (Var binder, CoreExpr rhs) ExprSize -> ExprSize -> ExprSize `addSizeNSD` CoreExpr -> ExprSize size_up CoreExpr body ExprSize -> Arity -> ExprSize `addSizeN` Var -> Arity forall p. Num p => Var -> p size_up_alloc Var binder size_up (Let (Rec [(Var, CoreExpr)] pairs) CoreExpr body) = ((Var, CoreExpr) -> ExprSize -> ExprSize) -> ExprSize -> [(Var, CoreExpr)] -> ExprSize forall (t :: * -> *) a b. Foldable t => (a -> b -> b) -> b -> t a -> b foldr (ExprSize -> ExprSize -> ExprSize addSizeNSD (ExprSize -> ExprSize -> ExprSize) -> ((Var, CoreExpr) -> ExprSize) -> (Var, CoreExpr) -> ExprSize -> ExprSize forall b c a. (b -> c) -> (a -> b) -> a -> c . (Var, CoreExpr) -> ExprSize size_up_rhs) (CoreExpr -> ExprSize size_up CoreExpr body ExprSize -> Arity -> ExprSize `addSizeN` [Arity] -> Arity forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a sum (((Var, CoreExpr) -> Arity) -> [(Var, CoreExpr)] -> [Arity] forall a b. (a -> b) -> [a] -> [b] map (Var -> Arity forall p. Num p => Var -> p size_up_alloc (Var -> Arity) -> ((Var, CoreExpr) -> Var) -> (Var, CoreExpr) -> Arity forall b c a. (b -> c) -> (a -> b) -> a -> c . (Var, CoreExpr) -> Var forall a b. (a, b) -> a fst) [(Var, CoreExpr)] pairs)) [(Var, CoreExpr)] pairs size_up (Case CoreExpr e Var _ Type _ [Alt Var] alts) | [Alt Var] -> Bool forall (t :: * -> *) a. Foldable t => t a -> Bool null [Alt Var] alts = CoreExpr -> ExprSize size_up CoreExpr e -- case e of {} never returns, so take size of scrutinee size_up (Case CoreExpr e Var _ Type _ [Alt Var] alts) -- Now alts is non-empty | Just Var v <- CoreExpr -> Maybe Var forall b. Expr b -> Maybe Var is_top_arg CoreExpr e -- We are scrutinising an argument variable = let alt_sizes :: [ExprSize] alt_sizes = (Alt Var -> ExprSize) -> [Alt Var] -> [ExprSize] forall a b. (a -> b) -> [a] -> [b] map Alt Var -> ExprSize size_up_alt [Alt Var] alts -- alts_size tries to compute a good discount for -- the case when we are scrutinising an argument variable alts_size :: ExprSize -> ExprSize -> ExprSize alts_size (SizeIs Arity tot Bag (Var, Arity) tot_disc Arity tot_scrut) -- Size of all alternatives (SizeIs Arity max Bag (Var, Arity) _ Arity _) -- Size of biggest alternative = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity tot ((Var, Arity) -> Bag (Var, Arity) forall a. a -> Bag a unitBag (Var v, Arity 20 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity tot Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity max) Bag (Var, Arity) -> Bag (Var, Arity) -> Bag (Var, Arity) forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Var, Arity) tot_disc) Arity tot_scrut -- If the variable is known, we produce a -- discount that will take us back to 'max', -- the size of the largest alternative The -- 1+ is a little discount for reduced -- allocation in the caller -- -- Notice though, that we return tot_disc, -- the total discount from all branches. I -- think that's right. alts_size ExprSize tot_size ExprSize _ = ExprSize tot_size in ExprSize -> ExprSize -> ExprSize alts_size ((ExprSize -> ExprSize -> ExprSize) -> [ExprSize] -> ExprSize forall (t :: * -> *) a. Foldable t => (a -> a -> a) -> t a -> a foldr1 ExprSize -> ExprSize -> ExprSize addAltSize [ExprSize] alt_sizes) -- alts is non-empty ((ExprSize -> ExprSize -> ExprSize) -> [ExprSize] -> ExprSize forall (t :: * -> *) a. Foldable t => (a -> a -> a) -> t a -> a foldr1 ExprSize -> ExprSize -> ExprSize maxSize [ExprSize] alt_sizes) -- Good to inline if an arg is scrutinised, because -- that may eliminate allocation in the caller -- And it eliminates the case itself where is_top_arg :: Expr b -> Maybe Var is_top_arg (Var Var v) | Var v Var -> [Var] -> Bool forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Var] top_args = Var -> Maybe Var forall a. a -> Maybe a Just Var v is_top_arg (Cast Expr b e Coercion _) = Expr b -> Maybe Var is_top_arg Expr b e is_top_arg Expr b _ = Maybe Var forall a. Maybe a Nothing size_up (Case CoreExpr e Var _ Type _ [Alt Var] alts) = CoreExpr -> ExprSize size_up CoreExpr e ExprSize -> ExprSize -> ExprSize `addSizeNSD` (Alt Var -> ExprSize -> ExprSize) -> ExprSize -> [Alt Var] -> ExprSize forall (t :: * -> *) a b. Foldable t => (a -> b -> b) -> b -> t a -> b foldr (ExprSize -> ExprSize -> ExprSize addAltSize (ExprSize -> ExprSize -> ExprSize) -> (Alt Var -> ExprSize) -> Alt Var -> ExprSize -> ExprSize forall b c a. (b -> c) -> (a -> b) -> a -> c . Alt Var -> ExprSize size_up_alt) ExprSize case_size [Alt Var] alts where case_size :: ExprSize case_size | CoreExpr -> Bool forall b. Expr b -> Bool is_inline_scrut CoreExpr e, [Alt Var] -> Arity -> Bool forall a. [a] -> Arity -> Bool lengthAtMost [Alt Var] alts Arity 1 = Arity -> ExprSize sizeN (-Arity 10) | Bool otherwise = ExprSize sizeZero -- Normally we don't charge for the case itself, but -- we charge one per alternative (see size_up_alt, -- below) to account for the cost of the info table -- and comparisons. -- -- However, in certain cases (see is_inline_scrut -- below), no code is generated for the case unless -- there are multiple alts. In these cases we -- subtract one, making the first alt free. -- e.g. case x# +# y# of _ -> ... should cost 1 -- case touch# x# of _ -> ... should cost 0 -- (see #4978) -- -- I would like to not have the "lengthAtMost alts 1" -- condition above, but without that some programs got worse -- (spectral/hartel/event and spectral/para). I don't fully -- understand why. (SDM 24/5/11) -- unboxed variables, inline primops and unsafe foreign calls -- are all "inline" things: is_inline_scrut :: Expr b -> Bool is_inline_scrut (Var Var v) = HasDebugCallStack => Type -> Bool Type -> Bool isUnliftedType (Var -> Type idType Var v) is_inline_scrut Expr b scrut | (Var Var f, [Expr b] _) <- Expr b -> (Expr b, [Expr b]) forall b. Expr b -> (Expr b, [Expr b]) collectArgs Expr b scrut = case Var -> IdDetails idDetails Var f of FCallId ForeignCall fc -> Bool -> Bool not (ForeignCall -> Bool isSafeForeignCall ForeignCall fc) PrimOpId PrimOp op -> Bool -> Bool not (PrimOp -> Bool primOpOutOfLine PrimOp op) IdDetails _other -> Bool False | Bool otherwise = Bool False size_up_rhs :: (Var, CoreExpr) -> ExprSize size_up_rhs (Var bndr, CoreExpr rhs) | Just Arity join_arity <- Var -> Maybe Arity isJoinId_maybe Var bndr -- Skip arguments to join point , ([Var] _bndrs, CoreExpr body) <- Arity -> CoreExpr -> ([Var], CoreExpr) forall b. Arity -> Expr b -> ([b], Expr b) collectNBinders Arity join_arity CoreExpr rhs = CoreExpr -> ExprSize size_up CoreExpr body | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr rhs ------------ -- size_up_app is used when there's ONE OR MORE value args size_up_app :: CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app (App CoreExpr fun CoreExpr arg) [CoreExpr] args Arity voids | CoreExpr -> Bool forall b. Expr b -> Bool isTyCoArg CoreExpr arg = CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr fun [CoreExpr] args Arity voids | CoreExpr -> Bool forall b. Expr b -> Bool isRealWorldExpr CoreExpr arg = CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr fun (CoreExpr argCoreExpr -> [CoreExpr] -> [CoreExpr] forall a. a -> [a] -> [a] :[CoreExpr] args) (Arity voids Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 1) | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr arg ExprSize -> ExprSize -> ExprSize `addSizeNSD` CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr fun (CoreExpr argCoreExpr -> [CoreExpr] -> [CoreExpr] forall a. a -> [a] -> [a] :[CoreExpr] args) Arity voids size_up_app (Var Var fun) [CoreExpr] args Arity voids = Var -> [CoreExpr] -> Arity -> ExprSize size_up_call Var fun [CoreExpr] args Arity voids size_up_app (Tick Tickish Var _ CoreExpr expr) [CoreExpr] args Arity voids = CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr expr [CoreExpr] args Arity voids size_up_app (Cast CoreExpr expr Coercion _) [CoreExpr] args Arity voids = CoreExpr -> [CoreExpr] -> Arity -> ExprSize size_up_app CoreExpr expr [CoreExpr] args Arity voids size_up_app CoreExpr other [CoreExpr] args Arity voids = CoreExpr -> ExprSize size_up CoreExpr other ExprSize -> Arity -> ExprSize `addSizeN` Arity -> Arity -> Arity callSize ([CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] args) Arity voids -- if the lhs is not an App or a Var, or an invisible thing like a -- Tick or Cast, then we should charge for a complete call plus the -- size of the lhs itself. ------------ size_up_call :: Id -> [CoreExpr] -> Int -> ExprSize size_up_call :: Var -> [CoreExpr] -> Arity -> ExprSize size_up_call Var fun [CoreExpr] val_args Arity voids = case Var -> IdDetails idDetails Var fun of FCallId ForeignCall _ -> Arity -> ExprSize sizeN (Arity -> Arity -> Arity callSize ([CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] val_args) Arity voids) DataConWorkId DataCon dc -> DataCon -> Arity -> ExprSize conSize DataCon dc ([CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] val_args) PrimOpId PrimOp op -> PrimOp -> Arity -> ExprSize primOpSize PrimOp op ([CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] val_args) ClassOpId Class _ -> DynFlags -> [Var] -> [CoreExpr] -> ExprSize classOpSize DynFlags dflags [Var] top_args [CoreExpr] val_args IdDetails _ -> DynFlags -> [Var] -> Var -> Arity -> Arity -> ExprSize funSize DynFlags dflags [Var] top_args Var fun ([CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] val_args) Arity voids ------------ size_up_alt :: Alt Var -> ExprSize size_up_alt (AltCon _con, [Var] _bndrs, CoreExpr rhs) = CoreExpr -> ExprSize size_up CoreExpr rhs ExprSize -> Arity -> ExprSize `addSizeN` Arity 10 -- Don't charge for args, so that wrappers look cheap -- (See comments about wrappers with Case) -- -- IMPORTANT: *do* charge 1 for the alternative, else we -- find that giant case nests are treated as practically free -- A good example is Foreign.C.Error.errnoToIOError ------------ -- Cost to allocate binding with given binder size_up_alloc :: Var -> p size_up_alloc Var bndr | Var -> Bool isTyVar Var bndr -- Doesn't exist at runtime Bool -> Bool -> Bool || Var -> Bool isJoinId Var bndr -- Not allocated at all Bool -> Bool -> Bool || HasDebugCallStack => Type -> Bool Type -> Bool isUnliftedType (Var -> Type idType Var bndr) -- Doesn't live in heap = p 0 | Bool otherwise = p 10 ------------ -- These addSize things have to be here because -- I don't want to give them bOMB_OUT_SIZE as an argument addSizeN :: ExprSize -> Arity -> ExprSize addSizeN ExprSize TooBig Arity _ = ExprSize TooBig addSizeN (SizeIs Arity n Bag (Var, Arity) xs Arity d) Arity m = Arity -> Arity -> Bag (Var, Arity) -> Arity -> ExprSize mkSizeIs Arity bOMB_OUT_SIZE (Arity n Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity m) Bag (Var, Arity) xs Arity d -- addAltSize is used to add the sizes of case alternatives addAltSize :: ExprSize -> ExprSize -> ExprSize addAltSize ExprSize TooBig ExprSize _ = ExprSize TooBig addAltSize ExprSize _ ExprSize TooBig = ExprSize TooBig addAltSize (SizeIs Arity n1 Bag (Var, Arity) xs Arity d1) (SizeIs Arity n2 Bag (Var, Arity) ys Arity d2) = Arity -> Arity -> Bag (Var, Arity) -> Arity -> ExprSize mkSizeIs Arity bOMB_OUT_SIZE (Arity n1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n2) (Bag (Var, Arity) xs Bag (Var, Arity) -> Bag (Var, Arity) -> Bag (Var, Arity) forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Var, Arity) ys) (Arity d1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity d2) -- Note [addAltSize result discounts] -- This variant ignores the result discount from its LEFT argument -- It's used when the second argument isn't part of the result addSizeNSD :: ExprSize -> ExprSize -> ExprSize addSizeNSD ExprSize TooBig ExprSize _ = ExprSize TooBig addSizeNSD ExprSize _ ExprSize TooBig = ExprSize TooBig addSizeNSD (SizeIs Arity n1 Bag (Var, Arity) xs Arity _) (SizeIs Arity n2 Bag (Var, Arity) ys Arity d2) = Arity -> Arity -> Bag (Var, Arity) -> Arity -> ExprSize mkSizeIs Arity bOMB_OUT_SIZE (Arity n1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n2) (Bag (Var, Arity) xs Bag (Var, Arity) -> Bag (Var, Arity) -> Bag (Var, Arity) forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Var, Arity) ys) Arity d2 -- Ignore d1 isRealWorldId :: Var -> Bool isRealWorldId Var id = Var -> Type idType Var id Type -> Type -> Bool `eqType` Type realWorldStatePrimTy -- an expression of type State# RealWorld must be a variable isRealWorldExpr :: Expr b -> Bool isRealWorldExpr (Var Var id) = Var -> Bool isRealWorldId Var id isRealWorldExpr (Tick Tickish Var _ Expr b e) = Expr b -> Bool isRealWorldExpr Expr b e isRealWorldExpr Expr b _ = Bool False -- | Finds a nominal size of a string literal. litSize :: Literal -> Int -- Used by CoreUnfold.sizeExpr litSize :: Literal -> Arity litSize (LitNumber LitNumType LitNumInteger Integer _ Type _) = Arity 100 -- Note [Size of literal integers] litSize (LitNumber LitNumType LitNumNatural Integer _ Type _) = Arity 100 litSize (LitString ByteString str) = Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * ((ByteString -> Arity BS.length ByteString str Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 3) Arity -> Arity -> Arity forall a. Integral a => a -> a -> a `div` Arity 4) -- If size could be 0 then @f "x"@ might be too small -- [Sept03: make literal strings a bit bigger to avoid fruitless -- duplication of little strings] litSize Literal _other = Arity 0 -- Must match size of nullary constructors -- Key point: if x |-> 4, then x must inline unconditionally -- (eg via case binding) classOpSize :: DynFlags -> [Id] -> [CoreExpr] -> ExprSize -- See Note [Conlike is interesting] classOpSize :: DynFlags -> [Var] -> [CoreExpr] -> ExprSize classOpSize DynFlags _ [Var] _ [] = ExprSize sizeZero classOpSize DynFlags dflags [Var] top_args (CoreExpr arg1 : [CoreExpr] other_args) = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity size Bag (Var, Arity) arg_discount Arity 0 where size :: Arity size = Arity 20 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + (Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * [CoreExpr] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [CoreExpr] other_args) -- If the class op is scrutinising a lambda bound dictionary then -- give it a discount, to encourage the inlining of this function -- The actual discount is rather arbitrarily chosen arg_discount :: Bag (Var, Arity) arg_discount = case CoreExpr arg1 of Var Var dict | Var dict Var -> [Var] -> Bool forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Var] top_args -> (Var, Arity) -> Bag (Var, Arity) forall a. a -> Bag a unitBag (Var dict, DynFlags -> Arity ufDictDiscount DynFlags dflags) CoreExpr _other -> Bag (Var, Arity) forall a. Bag a emptyBag -- | The size of a function call callSize :: Int -- ^ number of value args -> Int -- ^ number of value args that are void -> Int callSize :: Arity -> Arity -> Arity callSize Arity n_val_args Arity voids = Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity 1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n_val_args Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity voids) -- The 1+ is for the function itself -- Add 1 for each non-trivial arg; -- the allocation cost, as in let(rec) -- | The size of a jump to a join point jumpSize :: Int -- ^ number of value args -> Int -- ^ number of value args that are void -> Int jumpSize :: Arity -> Arity -> Arity jumpSize Arity n_val_args Arity voids = Arity 2 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity 1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n_val_args Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity voids) -- A jump is 20% the size of a function call. Making jumps free reopens -- bug #6048, but making them any more expensive loses a 21% improvement in -- spectral/puzzle. TODO Perhaps adjusting the default threshold would be a -- better solution? funSize :: DynFlags -> [Id] -> Id -> Int -> Int -> ExprSize -- Size for functions that are not constructors or primops -- Note [Function applications] funSize :: DynFlags -> [Var] -> Var -> Arity -> Arity -> ExprSize funSize DynFlags dflags [Var] top_args Var fun Arity n_val_args Arity voids | Var fun Var -> Unique -> Bool forall a. Uniquable a => a -> Unique -> Bool `hasKey` Unique buildIdKey = ExprSize buildSize | Var fun Var -> Unique -> Bool forall a. Uniquable a => a -> Unique -> Bool `hasKey` Unique augmentIdKey = ExprSize augmentSize | Bool otherwise = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity size Bag (Var, Arity) arg_discount Arity res_discount where some_val_args :: Bool some_val_args = Arity n_val_args Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 is_join :: Bool is_join = Var -> Bool isJoinId Var fun size :: Arity size | Bool is_join = Arity -> Arity -> Arity jumpSize Arity n_val_args Arity voids | Bool -> Bool not Bool some_val_args = Arity 0 | Bool otherwise = Arity -> Arity -> Arity callSize Arity n_val_args Arity voids -- DISCOUNTS -- See Note [Function and non-function discounts] arg_discount :: Bag (Var, Arity) arg_discount | Bool some_val_args Bool -> Bool -> Bool && Var fun Var -> [Var] -> Bool forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Var] top_args = (Var, Arity) -> Bag (Var, Arity) forall a. a -> Bag a unitBag (Var fun, DynFlags -> Arity ufFunAppDiscount DynFlags dflags) | Bool otherwise = Bag (Var, Arity) forall a. Bag a emptyBag -- If the function is an argument and is applied -- to some values, give it an arg-discount res_discount :: Arity res_discount | Var -> Arity idArity Var fun Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity n_val_args = DynFlags -> Arity ufFunAppDiscount DynFlags dflags | Bool otherwise = Arity 0 -- If the function is partially applied, show a result discount -- XXX maybe behave like ConSize for eval'd variable conSize :: DataCon -> Int -> ExprSize conSize :: DataCon -> Arity -> ExprSize conSize DataCon dc Arity n_val_args | Arity n_val_args Arity -> Arity -> Bool forall a. Eq a => a -> a -> Bool == Arity 0 = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 0 Bag (Var, Arity) forall a. Bag a emptyBag Arity 10 -- Like variables -- See Note [Unboxed tuple size and result discount] | DataCon -> Bool isUnboxedTupleCon DataCon dc = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 0 Bag (Var, Arity) forall a. Bag a emptyBag (Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity 1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n_val_args)) -- See Note [Constructor size and result discount] | Bool otherwise = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 10 Bag (Var, Arity) forall a. Bag a emptyBag (Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity 1 Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n_val_args)) -- XXX still looks to large to me {- Note [Constructor size and result discount] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Treat a constructors application as size 10, regardless of how many arguments it has; 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 (Just x) has size 0, which is the same as a lone variable; and hence 'v' will always be replaced by (Just x), where v is bound to Just x. The "result discount" is applied if the result of the call is scrutinised (say by a case). For a constructor application that will mean the constructor application will disappear, so we don't need to charge it to the function. So the discount should at least match the cost of the constructor application, namely 10. But to give a bit of extra incentive we give a discount of 10*(1 + n_val_args). Simon M tried a MUCH bigger discount: (10 * (10 + n_val_args)), and said it was an "unambiguous win", but its terribly dangerous because a function with many many case branches, each finishing with a constructor, can have an arbitrarily large discount. This led to terrible code bloat: see #6099. Note [Unboxed tuple size and result discount] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 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. I tried giving unboxed tuples a *result discount* of zero (see the commented-out line). Why? When returned as a result they do not allocate, so maybe we don't want to charge so much for them If you have a non-zero discount here, we find that workers often get inlined back into wrappers, because it look like f x = case $wf x of (# a,b #) -> (a,b) and we are keener because of the case. However while this change shrank binary sizes by 0.5% it also made spectral/boyer allocate 5% more. All other changes were very small. So it's not a big deal but I didn't adopt the idea. Note [Function and non-function discounts] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ We want a discount if the function is applied. A good example is monadic combinators with continuation arguments, where inlining is quite important. But we don't want a big discount when a function is called many times (see the detailed comments with #6048) because if the function is big it won't be inlined at its many call sites and no benefit results. Indeed, we can get exponentially big inlinings this way; that is what #6048 is about. On the other hand, for data-valued arguments, if there are lots of case expressions in the body, each one will get smaller if we apply the function to a constructor application, so we *want* a big discount if the argument is scrutinised by many case expressions. Conclusion: - For functions, take the max of the discounts - For data values, take the sum of the discounts Note [Literal integer size] ~~~~~~~~~~~~~~~~~~~~~~~~~~~ Literal integers *can* be big (mkInteger [...coefficients...]), but need not be (S# n). We just use an arbitrary big-ish constant here so that, in particular, we don't inline top-level defns like n = S# 5 There's no point in doing so -- any optimisations will see the S# through n's unfolding. Nor will a big size inhibit unfoldings functions that mention a literal Integer, because the float-out pass will float all those constants to top level. -} primOpSize :: PrimOp -> Int -> ExprSize primOpSize :: PrimOp -> Arity -> ExprSize primOpSize PrimOp op Arity n_val_args = if PrimOp -> Bool primOpOutOfLine PrimOp op then Arity -> ExprSize sizeN (Arity op_size Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity n_val_args) else Arity -> ExprSize sizeN Arity op_size where op_size :: Arity op_size = PrimOp -> Arity primOpCodeSize PrimOp op buildSize :: ExprSize buildSize :: ExprSize buildSize = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 0 Bag (Var, Arity) forall a. Bag a emptyBag Arity 40 -- 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 because build is -- very like a constructor. We don't bother to check that the -- build is saturated (it usually is). The "-2" discounts for the \c n, -- The "4" is rather arbitrary. augmentSize :: ExprSize augmentSize :: ExprSize augmentSize = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 0 Bag (Var, Arity) forall a. Bag a emptyBag Arity 40 -- Ditto (augment t (\cn -> e) ys) should cost only the cost of -- e plus ys. The -2 accounts for the \cn -- When we return a lambda, give a discount if it's used (applied) lamScrutDiscount :: DynFlags -> ExprSize -> ExprSize lamScrutDiscount :: DynFlags -> ExprSize -> ExprSize lamScrutDiscount DynFlags dflags (SizeIs Arity n Bag (Var, Arity) vs Arity _) = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity n Bag (Var, Arity) vs (DynFlags -> Arity ufFunAppDiscount DynFlags dflags) lamScrutDiscount DynFlags _ ExprSize TooBig = ExprSize TooBig {- Note [addAltSize result discounts] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ When adding the size of alternatives, we *add* the result discounts too, rather than take the *maximum*. For a multi-branch case, this gives a discount for each branch that returns a constructor, making us keener to inline. I did try using 'max' instead, but it makes nofib 'rewrite' and 'puzzle' allocate significantly more, and didn't make binary sizes shrink significantly either. Note [Discounts and thresholds] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Constants for discounts and thesholds are defined in main/DynFlags, all of form ufXxxx. They are: ufCreationThreshold At a definition site, if the unfolding is bigger than this, we may discard it altogether ufUseThreshold At a call site, if the unfolding, less discounts, is smaller than this, then it's small enough inline ufKeenessFactor Factor by which the discounts are multiplied before subtracting from size ufDictDiscount The discount for each occurrence of a dictionary argument as an argument of a class method. Should be pretty small else big functions may get inlined ufFunAppDiscount Discount for a function argument that is applied. Quite large, because if we inline we avoid the higher-order call. ufDearOp The size of a foreign call or not-dupable PrimOp ufVeryAggressive If True, the compiler ignores all the thresholds and inlines very aggressively. It still adheres to arity, simplifier phase control and loop breakers. Note [Function applications] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In a function application (f a b) - If 'f' is an argument to the function being analysed, and there's at least one value arg, record a FunAppDiscount for f - If the application if a PAP (arity > 2 in this example) record a *result* discount (because inlining with "extra" args in the call may mean that we now get a saturated application) Code for manipulating sizes -} -- | The size of a candidate expression for unfolding data ExprSize = TooBig | SizeIs { ExprSize -> Arity _es_size_is :: {-# UNPACK #-} !Int -- ^ Size found , ExprSize -> Bag (Var, Arity) _es_args :: !(Bag (Id,Int)) -- ^ Arguments cased herein, and discount for each such , ExprSize -> Arity _es_discount :: {-# UNPACK #-} !Int -- ^ Size to subtract if result is scrutinised by a case -- expression } instance Outputable ExprSize where ppr :: ExprSize -> SDoc ppr ExprSize TooBig = String -> SDoc text String "TooBig" ppr (SizeIs Arity a Bag (Var, Arity) _ Arity c) = SDoc -> SDoc brackets (Arity -> SDoc int Arity a SDoc -> SDoc -> SDoc <+> Arity -> SDoc int Arity c) -- 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 :: Int -> Int -> Bag (Id, Int) -> Int -> ExprSize mkSizeIs :: Arity -> Arity -> Bag (Var, Arity) -> Arity -> ExprSize mkSizeIs Arity max Arity n Bag (Var, Arity) xs Arity d | (Arity n Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity d) Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity max = ExprSize TooBig | Bool otherwise = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity n Bag (Var, Arity) xs Arity d maxSize :: ExprSize -> ExprSize -> ExprSize maxSize :: ExprSize -> ExprSize -> ExprSize maxSize ExprSize TooBig ExprSize _ = ExprSize TooBig maxSize ExprSize _ ExprSize TooBig = ExprSize TooBig maxSize s1 :: ExprSize s1@(SizeIs Arity n1 Bag (Var, Arity) _ Arity _) s2 :: ExprSize s2@(SizeIs Arity n2 Bag (Var, Arity) _ Arity _) | Arity n1 Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity n2 = ExprSize s1 | Bool otherwise = ExprSize s2 sizeZero :: ExprSize sizeN :: Int -> ExprSize sizeZero :: ExprSize sizeZero = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity 0 Bag (Var, Arity) forall a. Bag a emptyBag Arity 0 sizeN :: Arity -> ExprSize sizeN Arity n = Arity -> Bag (Var, Arity) -> Arity -> ExprSize SizeIs Arity n Bag (Var, Arity) forall a. Bag a emptyBag Arity 0 {- ************************************************************************ * * \subsection[considerUnfolding]{Given all the info, do (not) do the unfolding} * * ************************************************************************ We use 'couldBeSmallEnoughToInline' 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. -} couldBeSmallEnoughToInline :: DynFlags -> Int -> CoreExpr -> Bool couldBeSmallEnoughToInline :: DynFlags -> Arity -> CoreExpr -> Bool couldBeSmallEnoughToInline DynFlags dflags Arity threshold CoreExpr rhs = case DynFlags -> Arity -> [Var] -> CoreExpr -> ExprSize sizeExpr DynFlags dflags Arity threshold [] CoreExpr body of ExprSize TooBig -> Bool False ExprSize _ -> Bool True where ([Var] _, CoreExpr body) = CoreExpr -> ([Var], CoreExpr) forall b. Expr b -> ([b], Expr b) collectBinders CoreExpr rhs ---------------- smallEnoughToInline :: DynFlags -> Unfolding -> Bool smallEnoughToInline :: DynFlags -> Unfolding -> Bool smallEnoughToInline DynFlags dflags (CoreUnfolding {uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfIfGoodArgs {ug_size :: UnfoldingGuidance -> Arity ug_size = Arity size}}) = Arity size Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool <= DynFlags -> Arity ufUseThreshold DynFlags dflags smallEnoughToInline DynFlags _ Unfolding _ = Bool False ---------------- certainlyWillInline :: DynFlags -> IdInfo -> Maybe Unfolding -- ^ Sees if the unfolding is pretty certain to inline. -- If so, return a *stable* unfolding for it, that will always inline. certainlyWillInline :: DynFlags -> IdInfo -> Maybe Unfolding certainlyWillInline DynFlags dflags IdInfo fn_info = case IdInfo -> Unfolding unfoldingInfo IdInfo fn_info of CoreUnfolding { uf_tmpl :: Unfolding -> CoreExpr uf_tmpl = CoreExpr e, uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfoldingGuidance g } | Bool loop_breaker -> Maybe Unfolding forall a. Maybe a Nothing -- Won't inline, so try w/w | Bool noinline -> Maybe Unfolding forall a. Maybe a Nothing -- See Note [Worker-wrapper for NOINLINE functions] | Bool otherwise -> CoreExpr -> UnfoldingGuidance -> Maybe Unfolding do_cunf CoreExpr e UnfoldingGuidance g -- Depends on size, so look at that DFunUnfolding {} -> Unfolding -> Maybe Unfolding forall a. a -> Maybe a Just Unfolding fn_unf -- Don't w/w DFuns; it never makes sense -- to do so, and even if it is currently a -- loop breaker, it may not be later Unfolding _other_unf -> Maybe Unfolding forall a. Maybe a Nothing where loop_breaker :: Bool loop_breaker = OccInfo -> Bool isStrongLoopBreaker (IdInfo -> OccInfo occInfo IdInfo fn_info) noinline :: Bool noinline = InlinePragma -> InlineSpec inlinePragmaSpec (IdInfo -> InlinePragma inlinePragInfo IdInfo fn_info) InlineSpec -> InlineSpec -> Bool forall a. Eq a => a -> a -> Bool == InlineSpec NoInline fn_unf :: Unfolding fn_unf = IdInfo -> Unfolding unfoldingInfo IdInfo fn_info do_cunf :: CoreExpr -> UnfoldingGuidance -> Maybe Unfolding do_cunf :: CoreExpr -> UnfoldingGuidance -> Maybe Unfolding do_cunf CoreExpr _ UnfoldingGuidance UnfNever = Maybe Unfolding forall a. Maybe a Nothing do_cunf CoreExpr _ (UnfWhen {}) = Unfolding -> Maybe Unfolding forall a. a -> Maybe a Just (Unfolding fn_unf { uf_src :: UnfoldingSource uf_src = UnfoldingSource InlineStable }) -- INLINE functions have UnfWhen -- The UnfIfGoodArgs case seems important. If we w/w small functions -- binary sizes go up by 10%! (This is with SplitObjs.) -- I'm not totally sure why. -- INLINABLE functions come via this path -- See Note [certainlyWillInline: INLINABLE] do_cunf CoreExpr expr (UnfIfGoodArgs { ug_size :: UnfoldingGuidance -> Arity ug_size = Arity size, ug_args :: UnfoldingGuidance -> [Arity] ug_args = [Arity] args }) | IdInfo -> Arity arityInfo IdInfo fn_info Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 -- See Note [certainlyWillInline: be careful of thunks] , Bool -> Bool not (StrictSig -> Bool isBottomingSig (IdInfo -> StrictSig strictnessInfo IdInfo fn_info)) -- Do not unconditionally inline a bottoming functions even if -- it seems smallish. We've carefully lifted it out to top level, -- so we don't want to re-inline it. , let unf_arity :: Arity unf_arity = [Arity] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [Arity] args , Arity size Arity -> Arity -> Arity forall a. Num a => a -> a -> a - (Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * (Arity unf_arity Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 1)) Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool <= DynFlags -> Arity ufUseThreshold DynFlags dflags = Unfolding -> Maybe Unfolding forall a. a -> Maybe a Just (Unfolding fn_unf { uf_src :: UnfoldingSource uf_src = UnfoldingSource InlineStable , uf_guidance :: UnfoldingGuidance uf_guidance = UnfWhen :: Arity -> Bool -> Bool -> UnfoldingGuidance UnfWhen { ug_arity :: Arity ug_arity = Arity unf_arity , ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk , ug_boring_ok :: Bool ug_boring_ok = CoreExpr -> Bool inlineBoringOk CoreExpr expr } }) -- Note the "unsaturatedOk". A function like f = \ab. a -- will certainly inline, even if partially applied (f e), so we'd -- better make sure that the transformed inlining has the same property | Bool otherwise = Maybe Unfolding forall a. Maybe a Nothing {- Note [certainlyWillInline: be careful of thunks] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Don't claim that thunks will certainly inline, because that risks work duplication. Even if the work duplication is not great (eg is_cheap holds), it can make a big difference in an inner loop In #5623 we found that the WorkWrap phase thought that y = case x of F# v -> F# (v +# v) was certainlyWillInline, so the addition got duplicated. Note that we check arityInfo instead of the arity of the unfolding to detect this case. This is so that we don't accidentally fail to inline small partial applications, like `f = g 42` (where `g` recurses into `f`) where g has arity 2 (say). Here there is no risk of work duplication, and the RHS is tiny, so certainlyWillInline should return True. But `unf_arity` is zero! However f's arity, gotten from `arityInfo fn_info`, is 1. Failing to say that `f` will inline forces W/W to generate a potentially huge worker for f that will immediately cancel with `g`'s wrapper anyway, causing unnecessary churn in the Simplifier while arriving at the same result. Note [certainlyWillInline: INLINABLE] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ certainlyWillInline /must/ return Nothing for a large INLINABLE thing, even though we have a stable inlining, so that strictness w/w takes place. It makes a big difference to efficiency, and the w/w pass knows how to transfer the INLINABLE info to the worker; see WorkWrap Note [Worker-wrapper for INLINABLE functions] ************************************************************************ * * \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 NOTE: we don't want to inline top-level functions that always diverge. It just makes the code bigger. Tt turns out that the convenient way to prevent them inlining is to give them a NOINLINE pragma, which we do in StrictAnal.addStrictnessInfoToTopId -} callSiteInline :: DynFlags -> Id -- The Id -> Bool -- True <=> unfolding is active -> Bool -- True if there are no arguments at all (incl type args) -> [ArgSummary] -- One for each value arg; True if it is interesting -> CallCtxt -- True <=> continuation is interesting -> Maybe CoreExpr -- Unfolding, if any data ArgSummary = TrivArg -- Nothing interesting | NonTrivArg -- Arg has structure | ValueArg -- Arg is a con-app or PAP -- ..or con-like. Note [Conlike is interesting] instance Outputable ArgSummary where ppr :: ArgSummary -> SDoc ppr ArgSummary TrivArg = String -> SDoc text String "TrivArg" ppr ArgSummary NonTrivArg = String -> SDoc text String "NonTrivArg" ppr ArgSummary ValueArg = String -> SDoc text String "ValueArg" nonTriv :: ArgSummary -> Bool nonTriv :: ArgSummary -> Bool nonTriv ArgSummary TrivArg = Bool False nonTriv ArgSummary _ = Bool True data CallCtxt = BoringCtxt | RhsCtxt -- Rhs of a let-binding; see Note [RHS of lets] | DiscArgCtxt -- Argument of a function with non-zero arg discount | RuleArgCtxt -- We are somewhere in the argument of a function with rules | ValAppCtxt -- We're applied to at least one value arg -- This arises when we have ((f x |> co) y) -- Then the (f x) has argument 'x' but in a ValAppCtxt | CaseCtxt -- We're the scrutinee of a case -- that decomposes its scrutinee instance Outputable CallCtxt where ppr :: CallCtxt -> SDoc ppr CallCtxt CaseCtxt = String -> SDoc text String "CaseCtxt" ppr CallCtxt ValAppCtxt = String -> SDoc text String "ValAppCtxt" ppr CallCtxt BoringCtxt = String -> SDoc text String "BoringCtxt" ppr CallCtxt RhsCtxt = String -> SDoc text String "RhsCtxt" ppr CallCtxt DiscArgCtxt = String -> SDoc text String "DiscArgCtxt" ppr CallCtxt RuleArgCtxt = String -> SDoc text String "RuleArgCtxt" callSiteInline :: DynFlags -> Var -> Bool -> Bool -> [ArgSummary] -> CallCtxt -> Maybe CoreExpr callSiteInline DynFlags dflags Var id Bool active_unfolding Bool lone_variable [ArgSummary] arg_infos CallCtxt cont_info = case Var -> Unfolding idUnfolding Var id of -- idUnfolding checks for loop-breakers, returning NoUnfolding -- Things with an INLINE pragma may have an unfolding *and* -- be a loop breaker (maybe the knot is not yet untied) CoreUnfolding { uf_tmpl :: Unfolding -> CoreExpr uf_tmpl = CoreExpr unf_template , uf_is_work_free :: Unfolding -> Bool uf_is_work_free = Bool is_wf , uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfoldingGuidance guidance, uf_expandable :: Unfolding -> Bool uf_expandable = Bool is_exp } | Bool active_unfolding -> DynFlags -> Var -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding DynFlags dflags Var id Bool lone_variable [ArgSummary] arg_infos CallCtxt cont_info CoreExpr unf_template Bool is_wf Bool is_exp UnfoldingGuidance guidance | Bool otherwise -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String "Inactive unfolding:" (Var -> SDoc forall a. Outputable a => a -> SDoc ppr Var id) Maybe CoreExpr forall a. Maybe a Nothing Unfolding NoUnfolding -> Maybe CoreExpr forall a. Maybe a Nothing Unfolding BootUnfolding -> Maybe CoreExpr forall a. Maybe a Nothing OtherCon {} -> Maybe CoreExpr forall a. Maybe a Nothing DFunUnfolding {} -> Maybe CoreExpr forall a. Maybe a Nothing -- Never unfold a DFun traceInline :: DynFlags -> Id -> String -> SDoc -> a -> a traceInline :: DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var inline_id String str SDoc doc a result | Just String prefix <- DynFlags -> Maybe String inlineCheck DynFlags dflags = if String prefix String -> String -> Bool forall a. Eq a => [a] -> [a] -> Bool `isPrefixOf` OccName -> String occNameString (Var -> OccName forall a. NamedThing a => a -> OccName getOccName Var inline_id) then String -> SDoc -> a -> a forall a. String -> SDoc -> a -> a pprTrace String str SDoc doc a result else a result | DumpFlag -> DynFlags -> Bool dopt DumpFlag Opt_D_dump_inlinings DynFlags dflags Bool -> Bool -> Bool && DumpFlag -> DynFlags -> Bool dopt DumpFlag Opt_D_verbose_core2core DynFlags dflags = String -> SDoc -> a -> a forall a. String -> SDoc -> a -> a pprTrace String str SDoc doc a result | Bool otherwise = a result -- | This is an awful but temporary workaround for #17615, where the -- case analysis from the 'ufVeryAggressive' selector causes the entire -- 'DynFlags' to be unpacked into local bindings (due to binder swap). This -- results in a tremendous amount of stack spillage, severely bloating the code -- generated for 'callSiteInline'. -- -- The right solution here is likely to fix binder swap to avoid this terrible -- behavior (since there are likely other instances of this as well) but this -- case was serious enough that it showed up in a CPU profile and consequently -- I wanted to fix it for 8.10. very_aggressive :: DynFlags -> Bool very_aggressive :: DynFlags -> Bool very_aggressive = DynFlags -> Bool ufVeryAggressive {-# NOINLINE very_aggressive #-} tryUnfolding :: DynFlags -> Id -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding :: DynFlags -> Var -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding DynFlags dflags Var id Bool lone_variable [ArgSummary] arg_infos CallCtxt cont_info CoreExpr unf_template Bool is_wf Bool is_exp UnfoldingGuidance guidance = case UnfoldingGuidance guidance of UnfoldingGuidance UnfNever -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (String -> SDoc text String "UnfNever") Maybe CoreExpr forall a. Maybe a Nothing UnfWhen { ug_arity :: UnfoldingGuidance -> Arity ug_arity = Arity uf_arity, ug_unsat_ok :: UnfoldingGuidance -> Bool ug_unsat_ok = Bool unsat_ok, ug_boring_ok :: UnfoldingGuidance -> Bool ug_boring_ok = Bool boring_ok } | Bool enough_args Bool -> Bool -> Bool && (Bool boring_ok Bool -> Bool -> Bool || Bool some_benefit Bool -> Bool -> Bool || DynFlags -> Bool very_aggressive DynFlags dflags) -- See Note [INLINE for small functions (3)] -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (Bool -> SDoc -> Bool -> SDoc forall a. Outputable a => a -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc empty Bool True) (CoreExpr -> Maybe CoreExpr forall a. a -> Maybe a Just CoreExpr unf_template) | Bool otherwise -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (Bool -> SDoc -> Bool -> SDoc forall a. Outputable a => a -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc empty Bool False) Maybe CoreExpr forall a. Maybe a Nothing where some_benefit :: Bool some_benefit = Arity -> Bool calc_some_benefit Arity uf_arity enough_args :: Bool enough_args = (Arity n_val_args Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool >= Arity uf_arity) Bool -> Bool -> Bool || (Bool unsat_ok Bool -> Bool -> Bool && Arity n_val_args Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0) UnfIfGoodArgs { ug_args :: UnfoldingGuidance -> [Arity] ug_args = [Arity] arg_discounts, ug_res :: UnfoldingGuidance -> Arity ug_res = Arity res_discount, ug_size :: UnfoldingGuidance -> Arity ug_size = Arity size } | DynFlags -> Bool very_aggressive DynFlags dflags -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (Bool -> SDoc -> Bool -> SDoc forall a. Outputable a => a -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool True) (CoreExpr -> Maybe CoreExpr forall a. a -> Maybe a Just CoreExpr unf_template) | Bool is_wf Bool -> Bool -> Bool && Bool some_benefit Bool -> Bool -> Bool && Bool small_enough -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (Bool -> SDoc -> Bool -> SDoc forall a. Outputable a => a -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool True) (CoreExpr -> Maybe CoreExpr forall a. a -> Maybe a Just CoreExpr unf_template) | Bool otherwise -> DynFlags -> Var -> String -> SDoc -> Maybe CoreExpr -> Maybe CoreExpr forall a. DynFlags -> Var -> String -> SDoc -> a -> a traceInline DynFlags dflags Var id String str (Bool -> SDoc -> Bool -> SDoc forall a. Outputable a => a -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool False) Maybe CoreExpr forall a. Maybe a Nothing where some_benefit :: Bool some_benefit = Arity -> Bool calc_some_benefit ([Arity] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [Arity] arg_discounts) extra_doc :: SDoc extra_doc = String -> SDoc text String "discounted size =" SDoc -> SDoc -> SDoc <+> Arity -> SDoc int Arity discounted_size discounted_size :: Arity discounted_size = Arity size Arity -> Arity -> Arity forall a. Num a => a -> a -> a - Arity discount small_enough :: Bool small_enough = Arity discounted_size Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool <= DynFlags -> Arity ufUseThreshold DynFlags dflags discount :: Arity discount = DynFlags -> [Arity] -> Arity -> [ArgSummary] -> CallCtxt -> Arity computeDiscount DynFlags dflags [Arity] arg_discounts Arity res_discount [ArgSummary] arg_infos CallCtxt cont_info where mk_doc :: a -> SDoc -> Bool -> SDoc mk_doc a some_benefit SDoc extra_doc Bool yes_or_no = [SDoc] -> SDoc vcat [ String -> SDoc text String "arg infos" SDoc -> SDoc -> SDoc <+> [ArgSummary] -> SDoc forall a. Outputable a => a -> SDoc ppr [ArgSummary] arg_infos , String -> SDoc text String "interesting continuation" SDoc -> SDoc -> SDoc <+> CallCtxt -> SDoc forall a. Outputable a => a -> SDoc ppr CallCtxt cont_info , String -> SDoc text String "some_benefit" SDoc -> SDoc -> SDoc <+> a -> SDoc forall a. Outputable a => a -> SDoc ppr a some_benefit , String -> SDoc text String "is exp:" SDoc -> SDoc -> SDoc <+> Bool -> SDoc forall a. Outputable a => a -> SDoc ppr Bool is_exp , String -> SDoc text String "is work-free:" SDoc -> SDoc -> SDoc <+> Bool -> SDoc forall a. Outputable a => a -> SDoc ppr Bool is_wf , String -> SDoc text String "guidance" SDoc -> SDoc -> SDoc <+> UnfoldingGuidance -> SDoc forall a. Outputable a => a -> SDoc ppr UnfoldingGuidance guidance , SDoc extra_doc , String -> SDoc text String "ANSWER =" SDoc -> SDoc -> SDoc <+> if Bool yes_or_no then String -> SDoc text String "YES" else String -> SDoc text String "NO"] str :: String str = String "Considering inlining: " String -> String -> String forall a. [a] -> [a] -> [a] ++ DynFlags -> SDoc -> String showSDocDump DynFlags dflags (Var -> SDoc forall a. Outputable a => a -> SDoc ppr Var id) n_val_args :: Arity n_val_args = [ArgSummary] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [ArgSummary] arg_infos -- some_benefit is used when the RHS is small enough -- and the call has enough (or too many) value -- arguments (ie n_val_args >= arity). But there must -- be *something* interesting about some argument, or the -- result context, to make it worth inlining calc_some_benefit :: Arity -> Bool -- The Arity is the number of args -- expected by the unfolding calc_some_benefit :: Arity -> Bool calc_some_benefit Arity uf_arity | Bool -> Bool not Bool saturated = Bool interesting_args -- Under-saturated -- Note [Unsaturated applications] | Bool otherwise = Bool interesting_args -- Saturated or over-saturated Bool -> Bool -> Bool || Bool interesting_call where saturated :: Bool saturated = Arity n_val_args Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool >= Arity uf_arity over_saturated :: Bool over_saturated = Arity n_val_args Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity uf_arity interesting_args :: Bool interesting_args = (ArgSummary -> Bool) -> [ArgSummary] -> Bool forall (t :: * -> *) a. Foldable t => (a -> Bool) -> t a -> Bool any ArgSummary -> Bool nonTriv [ArgSummary] arg_infos -- NB: (any nonTriv arg_infos) looks at the -- over-saturated args too which is "wrong"; -- but if over-saturated we inline anyway. interesting_call :: Bool interesting_call | Bool over_saturated = Bool True | Bool otherwise = case CallCtxt cont_info of CallCtxt CaseCtxt -> Bool -> Bool not (Bool lone_variable Bool -> Bool -> Bool && Bool is_exp) -- Note [Lone variables] CallCtxt ValAppCtxt -> Bool True -- Note [Cast then apply] CallCtxt RuleArgCtxt -> Arity uf_arity Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 -- See Note [Unfold info lazy contexts] CallCtxt DiscArgCtxt -> Arity uf_arity Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 -- Note [Inlining in ArgCtxt] CallCtxt RhsCtxt -> Arity uf_arity Arity -> Arity -> Bool forall a. Ord a => a -> a -> Bool > Arity 0 -- CallCtxt _other -> Bool False -- See Note [Nested functions] {- Note [Unfold into lazy contexts], Note [RHS of lets] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ When the call is the argument of a function with a RULE, or the RHS of a let, we are a little bit keener to inline. For example f y = (y,y,y) g y = let x = f y in ...(case x of (a,b,c) -> ...) ... We'd inline 'f' if the call was in a case context, and it kind-of-is, only we can't see it. Also x = f v could be expensive whereas x = case v of (a,b) -> a is patently cheap and may allow more eta expansion. So we treat the RHS of a let as not-totally-boring. Note [Unsaturated applications] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ When a call is not saturated, we *still* inline if one of the arguments has interesting structure. That's sometimes very important. A good example is the Ord instance for Bool in Base: Rec { $fOrdBool =GHC.Classes.D:Ord @ Bool ... $cmin_ajX $cmin_ajX [Occ=LoopBreaker] :: Bool -> Bool -> Bool $cmin_ajX = GHC.Classes.$dmmin @ Bool $fOrdBool } But the defn of GHC.Classes.$dmmin is: $dmmin :: forall a. GHC.Classes.Ord a => a -> a -> a {- Arity: 3, HasNoCafRefs, Strictness: SLL, Unfolding: (\ @ a $dOrd :: GHC.Classes.Ord a x :: a y :: a -> case @ a GHC.Classes.<= @ a $dOrd x y of wild { GHC.Types.False -> y GHC.Types.True -> x }) -} We *really* want to inline $dmmin, even though it has arity 3, in order to unravel the recursion. Note [Things to watch] ~~~~~~~~~~~~~~~~~~~~~~ * { y = I# 3; x = y `cast` co; ...case (x `cast` co) of ... } Assume x is exported, so not inlined unconditionally. Then we want x to inline unconditionally; no reason for it not to, and doing so avoids an indirection. * { x = I# 3; ....f x.... } Make sure that x does not inline unconditionally! Lest we get extra allocation. Note [Inlining an InlineRule] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ An InlineRules is used for (a) programmer INLINE pragmas (b) inlinings from worker/wrapper For (a) the RHS may be large, and our contract is that we *only* inline when the function is applied to all the arguments on the LHS of the source-code defn. (The uf_arity in the rule.) However for worker/wrapper it may be worth inlining even if the arity is not satisfied (as we do in the CoreUnfolding case) so we don't require saturation. Note [Nested functions] ~~~~~~~~~~~~~~~~~~~~~~~ At one time we treated a call of a non-top-level function as "interesting" (regardless of how boring the context) in the hope that inlining it would eliminate the binding, and its allocation. Specifically, in the default case of interesting_call we had _other -> not is_top && uf_arity > 0 But actually postInlineUnconditionally does some of this and overall it makes virtually no difference to nofib. So I simplified away this special case Note [Cast then apply] ~~~~~~~~~~~~~~~~~~~~~~ Consider myIndex = __inline_me ( (/\a. <blah>) |> co ) co :: (forall a. a -> a) ~ (forall a. T a) ... /\a.\x. case ((myIndex a) |> sym co) x of { ... } ... We need to inline myIndex to unravel this; but the actual call (myIndex a) has no value arguments. The ValAppCtxt gives it enough incentive to inline. Note [Inlining in ArgCtxt] ~~~~~~~~~~~~~~~~~~~~~~~~~~ The condition (arity > 0) here is very important, because otherwise we end up inlining top-level stuff into useless places; eg x = I# 3# f = \y. g x This can make a very big difference: it adds 16% to nofib 'integer' allocs, and 20% to 'power'. At one stage I replaced this condition by 'True' (leading to the above slow-down). The motivation was test eyeball/inline1.hs; but that seems to work ok now. NOTE: arguably, we should inline in ArgCtxt only if the result of the call is at least CONLIKE. At least for the cases where we use ArgCtxt for the RHS of a 'let', we only profit from the inlining if we get a CONLIKE thing (modulo lets). Note [Lone variables] See also Note [Interaction of exprIsWorkFree and lone variables] ~~~~~~~~~~~~~~~~~~~~~ which appears below The "lone-variable" case is important. I spent ages messing about with unsatisfactory variants, but this is nice. The idea is that if a variable appears all alone as an arg of lazy fn, or rhs BoringCtxt as scrutinee of a case CaseCtxt as arg of a fn ArgCtxt AND it is bound to a cheap expression then we should not inline it (unless there is some other reason, e.g. it is the sole occurrence). That is what is happening at the use of 'lone_variable' in 'interesting_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, provided the RHS is "expandable" (see Note [exprIsExpandable] in CoreUtils) the context can ``see'' the unfolding of the variable (e.g. case or a RULE) so there's no gain. 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. So the non-inlining of lone_variables should only apply if the unfolding is regarded as cheap; because that is when exprIsConApp_maybe looks through the unfolding. Hence the "&& is_wf" in the InlineRule branch. * 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 Note [Interaction of exprIsWorkFree and lone variables] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The lone-variable test says "don't inline if a case expression scrutinises a lone variable whose unfolding is cheap". It's very important that, under these circumstances, exprIsConApp_maybe can spot a constructor application. So, for example, we don't consider let x = e in (x,x) to be cheap, and that's good because exprIsConApp_maybe doesn't think that expression is a constructor application. In the 'not (lone_variable && is_wf)' test, I used to test is_value rather than is_wf, which was utterly wrong, because the above expression responds True to exprIsHNF, which is what sets is_value. This kind of thing can occur if you have {-# INLINE foo #-} foo = let x = e in (x,x) which Roman did. -} computeDiscount :: DynFlags -> [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int computeDiscount :: DynFlags -> [Arity] -> Arity -> [ArgSummary] -> CallCtxt -> Arity computeDiscount DynFlags dflags [Arity] arg_discounts Arity res_discount [ArgSummary] arg_infos CallCtxt cont_info -- 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. = Arity 10 -- Discount of 10 because the result replaces the call -- so we count 10 for the function itself Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity 10 Arity -> Arity -> Arity forall a. Num a => a -> a -> a * [Arity] -> Arity forall (t :: * -> *) a. Foldable t => t a -> Arity length [Arity] actual_arg_discounts -- Discount of 10 for each arg supplied, -- because the result replaces the call Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Float -> Arity forall a b. (RealFrac a, Integral b) => a -> b round (DynFlags -> Float ufKeenessFactor DynFlags dflags Float -> Float -> Float forall a. Num a => a -> a -> a * Arity -> Float forall a b. (Integral a, Num b) => a -> b fromIntegral (Arity total_arg_discount Arity -> Arity -> Arity forall a. Num a => a -> a -> a + Arity res_discount')) where actual_arg_discounts :: [Arity] actual_arg_discounts = (Arity -> ArgSummary -> Arity) -> [Arity] -> [ArgSummary] -> [Arity] forall a b c. (a -> b -> c) -> [a] -> [b] -> [c] zipWith Arity -> ArgSummary -> Arity forall p. Num p => p -> ArgSummary -> p mk_arg_discount [Arity] arg_discounts [ArgSummary] arg_infos total_arg_discount :: Arity total_arg_discount = [Arity] -> Arity forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a sum [Arity] actual_arg_discounts mk_arg_discount :: p -> ArgSummary -> p mk_arg_discount p _ ArgSummary TrivArg = p 0 mk_arg_discount p _ ArgSummary NonTrivArg = p 10 mk_arg_discount p discount ArgSummary ValueArg = p discount res_discount' :: Arity res_discount' | Ordering LT <- [Arity] arg_discounts [Arity] -> [ArgSummary] -> Ordering forall a b. [a] -> [b] -> Ordering `compareLength` [ArgSummary] arg_infos = Arity res_discount -- Over-saturated | Bool otherwise = case CallCtxt cont_info of CallCtxt BoringCtxt -> Arity 0 CallCtxt CaseCtxt -> Arity res_discount -- Presumably a constructor CallCtxt ValAppCtxt -> Arity res_discount -- Presumably a function CallCtxt _ -> Arity 40 Arity -> Arity -> Arity forall a. Ord a => a -> a -> a `min` Arity res_discount -- ToDo: this 40 `min` res_discount doesn't seem right -- for DiscArgCtxt it shouldn't matter because the function will -- get the arg discount for any non-triv arg -- for RuleArgCtxt we do want to be keener to inline; but not only -- constructor results -- for RhsCtxt I suppose that exposing a data con is good in general -- And 40 seems very arbitrary -- -- res_discount can be very large when a function returns -- constructors; 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"