{- (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 #-} {-# LANGUAGE BangPatterns #-} {-# OPTIONS_GHC -Wno-incomplete-record-updates #-} module GHC.Core.Unfold ( Unfolding, UnfoldingGuidance, -- Abstract types UnfoldingOpts (..), defaultUnfoldingOpts, updateCreationThreshold, updateUseThreshold, updateFunAppDiscount, updateDictDiscount, updateVeryAggressive, updateCaseScaling, updateCaseThreshold, ArgSummary(..), couldBeSmallEnoughToInline, inlineBoringOk, certainlyWillInline, smallEnoughToInline, callSiteInline, CallCtxt(..), calcUnfoldingGuidance ) where #include "HsVersions.h" import GHC.Prelude import GHC.Driver.Session import GHC.Driver.Ppr import GHC.Core import GHC.Core.Utils import GHC.Types.Id import GHC.Types.Demand ( isDeadEndSig ) import GHC.Core.DataCon import GHC.Types.Literal import GHC.Builtin.PrimOps import GHC.Types.Id.Info import GHC.Types.Basic ( Arity, InlineSpec(..), inlinePragmaSpec ) import GHC.Core.Type import GHC.Builtin.Names import GHC.Builtin.Types.Prim ( realWorldStatePrimTy ) import GHC.Data.Bag import GHC.Utils.Logger import GHC.Utils.Misc import GHC.Utils.Outputable import GHC.Types.ForeignCall import GHC.Types.Name import GHC.Types.Tickish import qualified Data.ByteString as BS import Data.List (isPrefixOf) -- | Unfolding options data UnfoldingOpts = UnfoldingOpts { UnfoldingOpts -> Int unfoldingCreationThreshold :: !Int -- ^ Threshold above which unfoldings are not *created* , UnfoldingOpts -> Int unfoldingUseThreshold :: !Int -- ^ Threshold above which unfoldings are not *inlined* , UnfoldingOpts -> Int unfoldingFunAppDiscount :: !Int -- ^ Discount for lambdas that are used (applied) , UnfoldingOpts -> Int unfoldingDictDiscount :: !Int -- ^ Discount for dictionaries , UnfoldingOpts -> Bool unfoldingVeryAggressive :: !Bool -- ^ Force inlining in many more cases -- Don't consider depth up to x , UnfoldingOpts -> Int unfoldingCaseThreshold :: !Int -- Penalize depth with 1/x , UnfoldingOpts -> Int unfoldingCaseScaling :: !Int } defaultUnfoldingOpts :: UnfoldingOpts defaultUnfoldingOpts :: UnfoldingOpts defaultUnfoldingOpts = UnfoldingOpts { unfoldingCreationThreshold :: Int unfoldingCreationThreshold = Int 750 -- The unfoldingCreationThreshold threshold must be reasonably high -- to take account of possible discounts. -- E.g. 450 is not enough in 'fulsom' for Interval.sqr to -- inline into Csg.calc (The unfolding for sqr never makes it -- into the interface file.) , unfoldingUseThreshold :: Int unfoldingUseThreshold = Int 90 -- Last adjusted upwards in #18282, when I reduced -- the result discount for constructors. , unfoldingFunAppDiscount :: Int unfoldingFunAppDiscount = Int 60 -- Be fairly keen to inline a function if that means -- we'll be able to pick the right method from a dictionary , unfoldingDictDiscount :: Int unfoldingDictDiscount = Int 30 -- Be fairly keen to inline a function if that means -- we'll be able to pick the right method from a dictionary , unfoldingVeryAggressive :: Bool unfoldingVeryAggressive = Bool False -- Only apply scaling once we are deeper than threshold cases -- in an RHS. , unfoldingCaseThreshold :: Int unfoldingCaseThreshold = Int 2 -- Penalize depth with (size*depth)/scaling , unfoldingCaseScaling :: Int unfoldingCaseScaling = Int 30 } -- Helpers for "GHC.Driver.Session" updateCreationThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateCreationThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateCreationThreshold Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingCreationThreshold :: Int unfoldingCreationThreshold = Int n } updateUseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateUseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateUseThreshold Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingUseThreshold :: Int unfoldingUseThreshold = Int n } updateFunAppDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts updateFunAppDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts updateFunAppDiscount Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingFunAppDiscount :: Int unfoldingFunAppDiscount = Int n } updateDictDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts updateDictDiscount :: Int -> UnfoldingOpts -> UnfoldingOpts updateDictDiscount Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingDictDiscount :: Int unfoldingDictDiscount = Int n } updateVeryAggressive :: Bool -> UnfoldingOpts -> UnfoldingOpts updateVeryAggressive :: Bool -> UnfoldingOpts -> UnfoldingOpts updateVeryAggressive Bool n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingVeryAggressive :: Bool unfoldingVeryAggressive = Bool n } updateCaseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateCaseThreshold :: Int -> UnfoldingOpts -> UnfoldingOpts updateCaseThreshold Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingCaseThreshold :: Int unfoldingCaseThreshold = Int n } updateCaseScaling :: Int -> UnfoldingOpts -> UnfoldingOpts updateCaseScaling :: Int -> UnfoldingOpts -> UnfoldingOpts updateCaseScaling Int n UnfoldingOpts opts = UnfoldingOpts opts { unfoldingCaseScaling :: Int unfoldingCaseScaling = Int n } {- 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 actually are. 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 = Int -> CoreExpr -> Bool go Int 0 CoreExpr e where go :: Int -> CoreExpr -> Bool go :: Int -> CoreExpr -> Bool go Int credit (Lam Id x CoreExpr e) | Id -> Bool isId Id x = Int -> CoreExpr -> Bool go (Int creditforall a. Num a => a -> a -> a +Int 1) CoreExpr e | Bool otherwise = Int -> CoreExpr -> Bool go Int credit CoreExpr e -- See Note [Count coercion arguments in boring contexts] go Int credit (App CoreExpr f (Type {})) = Int -> CoreExpr -> Bool go Int credit CoreExpr f go Int credit (App CoreExpr f CoreExpr a) | Int credit forall a. Ord a => a -> a -> Bool > Int 0 , CoreExpr -> Bool exprIsTrivial CoreExpr a = Int -> CoreExpr -> Bool go (Int creditforall a. Num a => a -> a -> a -Int 1) CoreExpr f go Int credit (Tick CoreTickish _ CoreExpr e) = Int -> CoreExpr -> Bool go Int credit CoreExpr e -- dubious go Int credit (Cast CoreExpr e CoercionR _) = Int -> CoreExpr -> Bool go Int credit CoreExpr e go Int credit (Case CoreExpr scrut Id _ Type _ [Alt AltCon _ [Id] _ CoreExpr rhs]) -- See Note [Inline unsafeCoerce] | CoreExpr -> Bool isUnsafeEqualityProof CoreExpr scrut = Int -> CoreExpr -> Bool go Int credit CoreExpr rhs go Int _ (Var {}) = Bool boringCxtOk go Int _ CoreExpr _ = Bool boringCxtNotOk calcUnfoldingGuidance :: UnfoldingOpts -> Bool -- Definitely a top-level, bottoming binding -> CoreExpr -- Expression to look at -> UnfoldingGuidance calcUnfoldingGuidance :: UnfoldingOpts -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance UnfoldingOpts opts Bool is_top_bottoming (Tick CoreTickish t CoreExpr expr) | Bool -> Bool not (forall (pass :: TickishPass). GenTickish pass -> Bool tickishIsCode CoreTickish t) -- non-code ticks don't matter for unfolding = UnfoldingOpts -> Bool -> CoreExpr -> UnfoldingGuidance calcUnfoldingGuidance UnfoldingOpts opts Bool is_top_bottoming CoreExpr expr calcUnfoldingGuidance UnfoldingOpts opts Bool is_top_bottoming CoreExpr expr = case UnfoldingOpts -> Int -> [Id] -> CoreExpr -> ExprSize sizeExpr UnfoldingOpts opts Int bOMB_OUT_SIZE [Id] val_bndrs CoreExpr body of ExprSize TooBig -> UnfoldingGuidance UnfNever SizeIs Int size Bag (Id, Int) cased_bndrs Int scrut_discount | CoreExpr -> Int -> Int -> Bool uncondInline CoreExpr expr Int n_val_bndrs Int size -> UnfWhen { ug_unsat_ok :: Bool ug_unsat_ok = Bool unSaturatedOk , ug_boring_ok :: Bool ug_boring_ok = Bool boringCxtOk , ug_arity :: Int ug_arity = Int 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 -> -- If you don't force this then we retain all the Ids let !discounts :: [Int] discounts = forall a b. (a -> b) -> [a] -> [b] strictMap (Bag (Id, Int) -> Id -> Int mk_discount Bag (Id, Int) cased_bndrs) [Id] val_bndrs in UnfIfGoodArgs { ug_args :: [Int] ug_args = [Int] discounts , ug_size :: Int ug_size = Int size , ug_res :: Int ug_res = Int scrut_discount } where ([Id] bndrs, CoreExpr body) = forall b. Expr b -> ([b], Expr b) collectBinders CoreExpr expr bOMB_OUT_SIZE :: Int bOMB_OUT_SIZE = UnfoldingOpts -> Int unfoldingCreationThreshold UnfoldingOpts opts -- Bomb out if size gets bigger than this val_bndrs :: [Id] val_bndrs = forall a. (a -> Bool) -> [a] -> [a] filter Id -> Bool isId [Id] bndrs n_val_bndrs :: Int n_val_bndrs = forall (t :: * -> *) a. Foldable t => t a -> Int length [Id] val_bndrs mk_discount :: Bag (Id,Int) -> Id -> Int mk_discount :: Bag (Id, Int) -> Id -> Int mk_discount Bag (Id, Int) cbs Id bndr = forall (t :: * -> *) b a. Foldable t => (b -> a -> b) -> b -> t a -> b foldl' Int -> (Id, Int) -> Int combine Int 0 Bag (Id, Int) cbs where combine :: Int -> (Id, Int) -> Int combine Int acc (Id bndr', Int disc) | Id bndr forall a. Eq a => a -> a -> Bool == Id bndr' = Int acc Int -> Int -> Int `plus_disc` Int disc | Bool otherwise = Int acc plus_disc :: Int -> Int -> Int plus_disc :: Int -> Int -> Int plus_disc | Type -> Bool isFunTy (Id -> Type idType Id bndr) = forall a. Ord a => a -> a -> a max | Bool otherwise = forall a. Num a => a -> a -> a (+) -- See Note [Function and non-function discounts] {- Note [Inline unsafeCoerce] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ We really want to inline unsafeCoerce, even when applied to boring arguments. It doesn't look as if its RHS is smaller than the call unsafeCoerce x = case unsafeEqualityProof @a @b of UnsafeRefl -> x but that case is discarded -- see Note [Implementing unsafeCoerce] in base:Unsafe.Coerce. Moreover, if we /don't/ inline it, we may be left with f (unsafeCoerce x) which will build a thunk -- bad, bad, bad. Conclusion: we really want inlineBoringOk to be True of the RHS of unsafeCoerce. This is (U4) in Note [Implementing unsafeCoerce]. 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 universally 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 GHC.Core.Opt.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 GHC.Core.Opt.Simplify.Utils 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 "GHC.CoreToStg"). 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 -> Int -> Int -> Bool uncondInline CoreExpr rhs Int arity Int size | Int arity forall a. Ord a => a -> a -> Bool > Int 0 = Int size forall a. Ord a => a -> a -> Bool <= Int 10 forall a. Num a => a -> a -> a * (Int arity forall a. Num a => a -> a -> a + Int 1) -- See Note [INLINE for small functions] (1) | Bool otherwise = CoreExpr -> Bool exprIsTrivial CoreExpr rhs -- See Note [INLINE for small functions] (4) sizeExpr :: UnfoldingOpts -> 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] -- Forcing bOMB_OUT_SIZE early prevents repeated -- unboxing of the Int argument. sizeExpr :: UnfoldingOpts -> Int -> [Id] -> CoreExpr -> ExprSize sizeExpr UnfoldingOpts opts !Int bOMB_OUT_SIZE [Id] top_args CoreExpr expr = CoreExpr -> ExprSize size_up CoreExpr expr where size_up :: CoreExpr -> ExprSize size_up (Cast CoreExpr e CoercionR _) = CoreExpr -> ExprSize size_up CoreExpr e size_up (Tick CoreTickish _ CoreExpr e) = CoreExpr -> ExprSize size_up CoreExpr e size_up (Type Type _) = ExprSize sizeZero -- Types cost nothing size_up (Coercion CoercionR _) = ExprSize sizeZero size_up (Lit Literal lit) = Int -> ExprSize sizeN (Literal -> Int litSize Literal lit) size_up (Var Id f) | Id -> Bool isRealWorldId Id f = ExprSize sizeZero -- Make sure we get constructor discounts even -- on nullary constructors | Bool otherwise = Id -> [CoreExpr] -> Int -> ExprSize size_up_call Id f [] Int 0 size_up (App CoreExpr fun CoreExpr arg) | 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] -> Int -> ExprSize size_up_app CoreExpr fun [CoreExpr arg] (if forall b. Expr b -> Bool isRealWorldExpr CoreExpr arg then Int 1 else Int 0) size_up (Lam Id b CoreExpr e) | Id -> Bool isId Id b Bool -> Bool -> Bool && Bool -> Bool not (Id -> Bool isRealWorldId Id b) = UnfoldingOpts -> ExprSize -> ExprSize lamScrutDiscount UnfoldingOpts opts (CoreExpr -> ExprSize size_up CoreExpr e ExprSize -> Int -> ExprSize `addSizeN` Int 10) | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr e size_up (Let (NonRec Id binder CoreExpr rhs) CoreExpr body) = (Id, CoreExpr) -> ExprSize size_up_rhs (Id binder, CoreExpr rhs) ExprSize -> ExprSize -> ExprSize `addSizeNSD` CoreExpr -> ExprSize size_up CoreExpr body ExprSize -> Int -> ExprSize `addSizeN` forall {a}. Num a => Id -> a size_up_alloc Id binder size_up (Let (Rec [(Id, CoreExpr)] pairs) CoreExpr body) = forall (t :: * -> *) a b. Foldable t => (a -> b -> b) -> b -> t a -> b foldr (ExprSize -> ExprSize -> ExprSize addSizeNSD forall b c a. (b -> c) -> (a -> b) -> a -> c . (Id, CoreExpr) -> ExprSize size_up_rhs) (CoreExpr -> ExprSize size_up CoreExpr body ExprSize -> Int -> ExprSize `addSizeN` forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a sum (forall a b. (a -> b) -> [a] -> [b] map (forall {a}. Num a => Id -> a size_up_alloc forall b c a. (b -> c) -> (a -> b) -> a -> c . forall a b. (a, b) -> a fst) [(Id, CoreExpr)] pairs)) [(Id, CoreExpr)] pairs size_up (Case CoreExpr e Id _ Type _ [Alt Id] alts) | forall (t :: * -> *) a. Foldable t => t a -> Bool null [Alt Id] alts = CoreExpr -> ExprSize size_up CoreExpr e -- case e of {} never returns, so take size of scrutinee size_up (Case CoreExpr e Id _ Type _ [Alt Id] alts) -- Now alts is non-empty | Just Id v <- CoreExpr -> Maybe Id is_top_arg CoreExpr e -- We are scrutinising an argument variable = let alt_sizes :: [ExprSize] alt_sizes = forall a b. (a -> b) -> [a] -> [b] map Alt Id -> ExprSize size_up_alt [Alt Id] 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 Int tot Bag (Id, Int) tot_disc Int tot_scrut) -- Size of all alternatives (SizeIs Int max Bag (Id, Int) _ Int _) -- Size of biggest alternative = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int tot (forall a. a -> Bag a unitBag (Id v, Int 20 forall a. Num a => a -> a -> a + Int tot forall a. Num a => a -> a -> a - Int max) forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Id, Int) tot_disc) Int 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 (forall (t :: * -> *) a. Foldable t => (a -> a -> a) -> t a -> a foldr1 ExprSize -> ExprSize -> ExprSize addAltSize [ExprSize] alt_sizes) -- alts is non-empty (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 :: CoreExpr -> Maybe Id is_top_arg (Var Id v) | Id v forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Id] top_args = forall a. a -> Maybe a Just Id v is_top_arg (Cast CoreExpr e CoercionR _) = CoreExpr -> Maybe Id is_top_arg CoreExpr e is_top_arg CoreExpr _ = forall a. Maybe a Nothing size_up (Case CoreExpr e Id _ Type _ [Alt Id] alts) = CoreExpr -> ExprSize size_up CoreExpr e ExprSize -> ExprSize -> ExprSize `addSizeNSD` forall (t :: * -> *) a b. Foldable t => (a -> b -> b) -> b -> t a -> b foldr (ExprSize -> ExprSize -> ExprSize addAltSize forall b c a. (b -> c) -> (a -> b) -> a -> c . Alt Id -> ExprSize size_up_alt) ExprSize case_size [Alt Id] alts where case_size :: ExprSize case_size | forall b. Expr b -> Bool is_inline_scrut CoreExpr e, forall a. [a] -> Int -> Bool lengthAtMost [Alt Id] alts Int 1 = Int -> ExprSize sizeN (-Int 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 Id v) = HasDebugCallStack => Type -> Bool isUnliftedType (Id -> Type idType Id v) is_inline_scrut Expr b scrut | (Var Id f, [Expr b] _) <- forall b. Expr b -> (Expr b, [Expr b]) collectArgs Expr b scrut = case Id -> IdDetails idDetails Id 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 :: (Id, CoreExpr) -> ExprSize size_up_rhs (Id bndr, CoreExpr rhs) | Just Int join_arity <- Id -> Maybe Int isJoinId_maybe Id bndr -- Skip arguments to join point , ([Id] _bndrs, CoreExpr body) <- forall b. Int -> Expr b -> ([b], Expr b) collectNBinders Int 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] -> Int -> ExprSize size_up_app (App CoreExpr fun CoreExpr arg) [CoreExpr] args Int voids | forall b. Expr b -> Bool isTyCoArg CoreExpr arg = CoreExpr -> [CoreExpr] -> Int -> ExprSize size_up_app CoreExpr fun [CoreExpr] args Int voids | forall b. Expr b -> Bool isRealWorldExpr CoreExpr arg = CoreExpr -> [CoreExpr] -> Int -> ExprSize size_up_app CoreExpr fun (CoreExpr argforall a. a -> [a] -> [a] :[CoreExpr] args) (Int voids forall a. Num a => a -> a -> a + Int 1) | Bool otherwise = CoreExpr -> ExprSize size_up CoreExpr arg ExprSize -> ExprSize -> ExprSize `addSizeNSD` CoreExpr -> [CoreExpr] -> Int -> ExprSize size_up_app CoreExpr fun (CoreExpr argforall a. a -> [a] -> [a] :[CoreExpr] args) Int voids size_up_app (Var Id fun) [CoreExpr] args Int voids = Id -> [CoreExpr] -> Int -> ExprSize size_up_call Id fun [CoreExpr] args Int voids size_up_app (Tick CoreTickish _ CoreExpr expr) [CoreExpr] args Int voids = CoreExpr -> [CoreExpr] -> Int -> ExprSize size_up_app CoreExpr expr [CoreExpr] args Int voids size_up_app (Cast CoreExpr expr CoercionR _) [CoreExpr] args Int voids = CoreExpr -> [CoreExpr] -> Int -> ExprSize size_up_app CoreExpr expr [CoreExpr] args Int voids size_up_app CoreExpr other [CoreExpr] args Int voids = CoreExpr -> ExprSize size_up CoreExpr other ExprSize -> Int -> ExprSize `addSizeN` Int -> Int -> Int callSize (forall (t :: * -> *) a. Foldable t => t a -> Int length [CoreExpr] args) Int 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 :: Id -> [CoreExpr] -> Int -> ExprSize size_up_call Id fun [CoreExpr] val_args Int voids = case Id -> IdDetails idDetails Id fun of FCallId ForeignCall _ -> Int -> ExprSize sizeN (Int -> Int -> Int callSize (forall (t :: * -> *) a. Foldable t => t a -> Int length [CoreExpr] val_args) Int voids) DataConWorkId DataCon dc -> DataCon -> Int -> ExprSize conSize DataCon dc (forall (t :: * -> *) a. Foldable t => t a -> Int length [CoreExpr] val_args) PrimOpId PrimOp op -> PrimOp -> Int -> ExprSize primOpSize PrimOp op (forall (t :: * -> *) a. Foldable t => t a -> Int length [CoreExpr] val_args) ClassOpId Class _ -> UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize classOpSize UnfoldingOpts opts [Id] top_args [CoreExpr] val_args IdDetails _ -> UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize funSize UnfoldingOpts opts [Id] top_args Id fun (forall (t :: * -> *) a. Foldable t => t a -> Int length [CoreExpr] val_args) Int voids ------------ size_up_alt :: Alt Id -> ExprSize size_up_alt (Alt AltCon _con [Id] _bndrs CoreExpr rhs) = CoreExpr -> ExprSize size_up CoreExpr rhs ExprSize -> Int -> ExprSize `addSizeN` Int 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 :: Id -> a size_up_alloc Id bndr | Id -> Bool isTyVar Id bndr -- Doesn't exist at runtime Bool -> Bool -> Bool || Id -> Bool isJoinId Id bndr -- Not allocated at all Bool -> Bool -> Bool || HasDebugCallStack => Type -> Bool isUnliftedType (Id -> Type idType Id bndr) -- Doesn't live in heap = a 0 | Bool otherwise = a 10 ------------ -- These addSize things have to be here because -- I don't want to give them bOMB_OUT_SIZE as an argument addSizeN :: ExprSize -> Int -> ExprSize addSizeN ExprSize TooBig Int _ = ExprSize TooBig addSizeN (SizeIs Int n Bag (Id, Int) xs Int d) Int m = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize mkSizeIs Int bOMB_OUT_SIZE (Int n forall a. Num a => a -> a -> a + Int m) Bag (Id, Int) xs Int 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 Int n1 Bag (Id, Int) xs Int d1) (SizeIs Int n2 Bag (Id, Int) ys Int d2) = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize mkSizeIs Int bOMB_OUT_SIZE (Int n1 forall a. Num a => a -> a -> a + Int n2) (Bag (Id, Int) xs forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Id, Int) ys) (Int d1 forall a. Num a => a -> a -> a + Int 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 Int n1 Bag (Id, Int) xs Int _) (SizeIs Int n2 Bag (Id, Int) ys Int d2) = Int -> Int -> Bag (Id, Int) -> Int -> ExprSize mkSizeIs Int bOMB_OUT_SIZE (Int n1 forall a. Num a => a -> a -> a + Int n2) (Bag (Id, Int) xs forall a. Bag a -> Bag a -> Bag a `unionBags` Bag (Id, Int) ys) Int d2 -- Ignore d1 isRealWorldId :: Id -> Bool isRealWorldId Id id = Id -> Type idType Id id Type -> Type -> Bool `eqType` Type realWorldStatePrimTy -- an expression of type State# RealWorld must be a variable isRealWorldExpr :: Expr b -> Bool isRealWorldExpr (Var Id id) = Id -> Bool isRealWorldId Id id isRealWorldExpr (Tick CoreTickish _ 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 GHC.Core.Unfold.sizeExpr litSize :: Literal -> Int litSize (LitNumber LitNumType LitNumInteger Integer _) = Int 100 -- Note [Size of literal integers] litSize (LitNumber LitNumType LitNumNatural Integer _) = Int 100 litSize (LitString ByteString str) = Int 10 forall a. Num a => a -> a -> a + Int 10 forall a. Num a => a -> a -> a * ((ByteString -> Int BS.length ByteString str forall a. Num a => a -> a -> a + Int 3) forall a. Integral a => a -> a -> a `div` Int 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 = Int 0 -- Must match size of nullary constructors -- Key point: if x |-> 4, then x must inline unconditionally -- (eg via case binding) classOpSize :: UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize -- See Note [Conlike is interesting] classOpSize :: UnfoldingOpts -> [Id] -> [CoreExpr] -> ExprSize classOpSize UnfoldingOpts _ [Id] _ [] = ExprSize sizeZero classOpSize UnfoldingOpts opts [Id] top_args (CoreExpr arg1 : [CoreExpr] other_args) = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int size Bag (Id, Int) arg_discount Int 0 where size :: Int size = Int 20 forall a. Num a => a -> a -> a + (Int 10 forall a. Num a => a -> a -> a * forall (t :: * -> *) a. Foldable t => t a -> Int 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 (Id, Int) arg_discount = case CoreExpr arg1 of Var Id dict | Id dict forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Id] top_args -> forall a. a -> Bag a unitBag (Id dict, UnfoldingOpts -> Int unfoldingDictDiscount UnfoldingOpts opts) CoreExpr _other -> 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 :: Int -> Int -> Int callSize Int n_val_args Int voids = Int 10 forall a. Num a => a -> a -> a * (Int 1 forall a. Num a => a -> a -> a + Int n_val_args forall a. Num a => a -> a -> a - Int 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 :: Int -> Int -> Int jumpSize Int n_val_args Int voids = Int 2 forall a. Num a => a -> a -> a * (Int 1 forall a. Num a => a -> a -> a + Int n_val_args forall a. Num a => a -> a -> a - Int 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 :: UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize -- Size for functions that are not constructors or primops -- Note [Function applications] funSize :: UnfoldingOpts -> [Id] -> Id -> Int -> Int -> ExprSize funSize UnfoldingOpts opts [Id] top_args Id fun Int n_val_args Int voids | Id fun forall a. Uniquable a => a -> Unique -> Bool `hasKey` Unique buildIdKey = ExprSize buildSize | Id fun forall a. Uniquable a => a -> Unique -> Bool `hasKey` Unique augmentIdKey = ExprSize augmentSize | Bool otherwise = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int size Bag (Id, Int) arg_discount Int res_discount where some_val_args :: Bool some_val_args = Int n_val_args forall a. Ord a => a -> a -> Bool > Int 0 is_join :: Bool is_join = Id -> Bool isJoinId Id fun size :: Int size | Bool is_join = Int -> Int -> Int jumpSize Int n_val_args Int voids | Bool -> Bool not Bool some_val_args = Int 0 | Bool otherwise = Int -> Int -> Int callSize Int n_val_args Int voids -- DISCOUNTS -- See Note [Function and non-function discounts] arg_discount :: Bag (Id, Int) arg_discount | Bool some_val_args Bool -> Bool -> Bool && Id fun forall (t :: * -> *) a. (Foldable t, Eq a) => a -> t a -> Bool `elem` [Id] top_args = forall a. a -> Bag a unitBag (Id fun, UnfoldingOpts -> Int unfoldingFunAppDiscount UnfoldingOpts opts) | Bool otherwise = forall a. Bag a emptyBag -- If the function is an argument and is applied -- to some values, give it an arg-discount res_discount :: Int res_discount | Id -> Int idArity Id fun forall a. Ord a => a -> a -> Bool > Int n_val_args = UnfoldingOpts -> Int unfoldingFunAppDiscount UnfoldingOpts opts | Bool otherwise = Int 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 -> Int -> ExprSize conSize DataCon dc Int n_val_args | Int n_val_args forall a. Eq a => a -> a -> Bool == Int 0 = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 0 forall a. Bag a emptyBag Int 10 -- Like variables -- See Note [Unboxed tuple size and result discount] | DataCon -> Bool isUnboxedTupleDataCon DataCon dc = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 0 forall a. Bag a emptyBag Int 10 -- See Note [Constructor size and result discount] | Bool otherwise = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 10 forall a. Bag a emptyBag Int 10 {- 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. Historical note 1: Until Jun 2020 we gave it a "bit of extra incentive" via a discount of 10*(1 + n_val_args), but that was FAR too much (#18282). In particular, consider a huge case tree like let r = case y1 of Nothing -> B1 a b c Just v1 -> case y2 of Nothing -> B1 c b a Just v2 -> ... If conSize gives a cost of 10 (regardless of n_val_args) and a discount of 10, that'll make each alternative RHS cost zero. We charge 10 for each case alternative (see size_up_alt). If we give a bigger discount (say 20) in conSize, we'll make the case expression cost *nothing*, and that can make a huge case tree cost nothing. This leads to massive, sometimes exponenial inlinings (#18282). In short, don't give a discount that give a negative size to a sub-expression! Historical note 2: Much longer ago, 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. When fixing #18282 (see Note [Constructor size and result discount]) I changed the result discount to be just 10, not 10*(1+n_val_args). 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 (IS n). We just use an arbitrary big-ish constant here so that, in particular, we don't inline top-level defns like n = IS 5 There's no point in doing so -- any optimisations will see the IS 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 -> Int -> ExprSize primOpSize PrimOp op Int n_val_args = if PrimOp -> Bool primOpOutOfLine PrimOp op then Int -> ExprSize sizeN (Int op_size forall a. Num a => a -> a -> a + Int n_val_args) else Int -> ExprSize sizeN Int op_size where op_size :: Int op_size = PrimOp -> Int primOpCodeSize PrimOp op buildSize :: ExprSize buildSize :: ExprSize buildSize = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 0 forall a. Bag a emptyBag Int 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 = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 0 forall a. Bag a emptyBag Int 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 :: UnfoldingOpts -> ExprSize -> ExprSize lamScrutDiscount :: UnfoldingOpts -> ExprSize -> ExprSize lamScrutDiscount UnfoldingOpts opts (SizeIs Int n Bag (Id, Int) vs Int _) = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int n Bag (Id, Int) vs (UnfoldingOpts -> Int unfoldingFunAppDiscount UnfoldingOpts opts) lamScrutDiscount UnfoldingOpts _ 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 thresholds are defined in 'UnfoldingOpts'. They are: unfoldingCreationThreshold At a definition site, if the unfolding is bigger than this, we may discard it altogether unfoldingUseThreshold At a call site, if the unfolding, less discounts, is smaller than this, then it's small enough inline unfoldingDictDiscount 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 unfoldingFunAppDiscount Discount for a function argument that is applied. Quite large, because if we inline we avoid the higher-order call. unfoldingVeryAggressive If True, the compiler ignores all the thresholds and inlines very aggressively. It still adheres to arity, simplifier phase control and loop breakers. Historical Note: Before April 2020 we had another factor, ufKeenessFactor, which would scale the discounts before they were subtracted from the size. This was justified with the following comment: -- We multiply 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. However, this is highly suspect since it means that we subtract a *scaled* size from an absolute size, resulting in crazy (e.g. negative) scores in some cases (#15304). We consequently killed off ufKeenessFactor and bumped up the ufUseThreshold to compensate. 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 -> Int _es_size_is :: {-# UNPACK #-} !Int -- ^ Size found , ExprSize -> Bag (Id, Int) _es_args :: !(Bag (Id,Int)) -- ^ Arguments cased herein, and discount for each such , ExprSize -> Int _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 Int a Bag (Id, Int) _ Int c) = SDoc -> SDoc brackets (Int -> SDoc int Int a SDoc -> SDoc -> SDoc <+> Int -> SDoc int Int 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 :: Int -> Int -> Bag (Id, Int) -> Int -> ExprSize mkSizeIs Int max Int n Bag (Id, Int) xs Int d | (Int n forall a. Num a => a -> a -> a - Int d) forall a. Ord a => a -> a -> Bool > Int max = ExprSize TooBig | Bool otherwise = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int n Bag (Id, Int) xs Int 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 Int n1 Bag (Id, Int) _ Int _) s2 :: ExprSize s2@(SizeIs Int n2 Bag (Id, Int) _ Int _) | Int n1 forall a. Ord a => a -> a -> Bool > Int n2 = ExprSize s1 | Bool otherwise = ExprSize s2 sizeZero :: ExprSize sizeN :: Int -> ExprSize sizeZero :: ExprSize sizeZero = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int 0 forall a. Bag a emptyBag Int 0 sizeN :: Int -> ExprSize sizeN Int n = Int -> Bag (Id, Int) -> Int -> ExprSize SizeIs Int n forall a. Bag a emptyBag Int 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 :: UnfoldingOpts -> Int -> CoreExpr -> Bool couldBeSmallEnoughToInline :: UnfoldingOpts -> Int -> CoreExpr -> Bool couldBeSmallEnoughToInline UnfoldingOpts opts Int threshold CoreExpr rhs = case UnfoldingOpts -> Int -> [Id] -> CoreExpr -> ExprSize sizeExpr UnfoldingOpts opts Int threshold [] CoreExpr body of ExprSize TooBig -> Bool False ExprSize _ -> Bool True where ([Id] _, CoreExpr body) = forall b. Expr b -> ([b], Expr b) collectBinders CoreExpr rhs ---------------- smallEnoughToInline :: UnfoldingOpts -> Unfolding -> Bool smallEnoughToInline :: UnfoldingOpts -> Unfolding -> Bool smallEnoughToInline UnfoldingOpts opts (CoreUnfolding {uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfIfGoodArgs {ug_size :: UnfoldingGuidance -> Int ug_size = Int size}}) = Int size forall a. Ord a => a -> a -> Bool <= UnfoldingOpts -> Int unfoldingUseThreshold UnfoldingOpts opts smallEnoughToInline UnfoldingOpts _ Unfolding _ = Bool False ---------------- certainlyWillInline :: UnfoldingOpts -> 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 :: UnfoldingOpts -> IdInfo -> Maybe Unfolding certainlyWillInline UnfoldingOpts opts IdInfo fn_info = case Unfolding fn_unf of CoreUnfolding { uf_tmpl :: Unfolding -> CoreExpr uf_tmpl = CoreExpr expr, uf_guidance :: Unfolding -> UnfoldingGuidance uf_guidance = UnfoldingGuidance guidance, uf_src :: Unfolding -> UnfoldingSource uf_src = UnfoldingSource src } | Bool loop_breaker -> forall a. Maybe a Nothing -- Won't inline, so try w/w | Bool noinline -> forall a. Maybe a Nothing -- See Note [Worker-wrapper for NOINLINE functions] | Bool otherwise -> case UnfoldingGuidance guidance of UnfoldingGuidance UnfNever -> forall a. Maybe a Nothing UnfWhen {} -> forall a. a -> Maybe a Just (Unfolding fn_unf { uf_src :: UnfoldingSource uf_src = UnfoldingSource src' }) -- INLINE functions have UnfWhen UnfIfGoodArgs { ug_size :: UnfoldingGuidance -> Int ug_size = Int size, ug_args :: UnfoldingGuidance -> [Int] ug_args = [Int] args } -> CoreExpr -> Int -> [Int] -> UnfoldingSource -> Maybe Unfolding do_cunf CoreExpr expr Int size [Int] args UnfoldingSource src' where src' :: UnfoldingSource src' = case UnfoldingSource src of UnfoldingSource InlineRhs -> UnfoldingSource InlineStable UnfoldingSource _ -> UnfoldingSource src -- Do not change InlineCompulsory! DFunUnfolding {} -> 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 -> 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) forall a. Eq a => a -> a -> Bool == InlineSpec NoInline fn_unf :: Unfolding fn_unf = IdInfo -> Unfolding unfoldingInfo IdInfo fn_info -- 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 -> Int -> [Int] -> UnfoldingSource -> Maybe Unfolding do_cunf CoreExpr expr Int size [Int] args UnfoldingSource src' | IdInfo -> Int arityInfo IdInfo fn_info forall a. Ord a => a -> a -> Bool > Int 0 -- See Note [certainlyWillInline: be careful of thunks] , Bool -> Bool not (StrictSig -> Bool isDeadEndSig (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 :: Int unf_arity = forall (t :: * -> *) a. Foldable t => t a -> Int length [Int] args , Int size forall a. Num a => a -> a -> a - (Int 10 forall a. Num a => a -> a -> a * (Int unf_arity forall a. Num a => a -> a -> a + Int 1)) forall a. Ord a => a -> a -> Bool <= UnfoldingOpts -> Int unfoldingUseThreshold UnfoldingOpts opts = forall a. a -> Maybe a Just (Unfolding fn_unf { uf_src :: UnfoldingSource uf_src = UnfoldingSource src' , uf_guidance :: UnfoldingGuidance uf_guidance = UnfWhen { ug_arity :: Int ug_arity = Int 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 = 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 :: Logger -> DynFlags -> Int -- Case depth -> 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 :: Logger -> DynFlags -> Int -> Id -> Bool -> Bool -> [ArgSummary] -> CallCtxt -> Maybe CoreExpr callSiteInline Logger logger DynFlags dflags !Int case_depth Id id Bool active_unfolding Bool lone_variable [ArgSummary] arg_infos CallCtxt cont_info = case Id -> Unfolding idUnfolding Id 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 -> Logger -> DynFlags -> Int -> Id -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding Logger logger DynFlags dflags Int case_depth Id id Bool lone_variable [ArgSummary] arg_infos CallCtxt cont_info CoreExpr unf_template Bool is_wf Bool is_exp UnfoldingGuidance guidance | Bool otherwise -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String "Inactive unfolding:" (forall a. Outputable a => a -> SDoc ppr Id id) forall a. Maybe a Nothing Unfolding NoUnfolding -> forall a. Maybe a Nothing Unfolding BootUnfolding -> forall a. Maybe a Nothing OtherCon {} -> forall a. Maybe a Nothing DFunUnfolding {} -> forall a. Maybe a Nothing -- Never unfold a DFun -- | Report the inlining of an identifier's RHS to the user, if requested. traceInline :: Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline :: forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id inline_id String str SDoc doc a result -- We take care to ensure that doc is used in only one branch, ensuring that -- the simplifier can push its allocation into the branch. See Note [INLINE -- conditional tracing utilities]. | Bool enable = forall a. Logger -> TraceAction a putTraceMsg Logger logger DynFlags dflags String str SDoc doc a result | Bool otherwise = a result where enable :: Bool enable | 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 = Bool True | Just String prefix <- DynFlags -> Maybe String inlineCheck DynFlags dflags = String prefix forall a. Eq a => [a] -> [a] -> Bool `isPrefixOf` OccName -> String occNameString (forall a. NamedThing a => a -> OccName getOccName Id inline_id) | Bool otherwise = Bool False {-# INLINE traceInline #-} -- see Note [INLINE conditional tracing utilities] {- Note [Avoid inlining into deeply nested cases] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Consider a function f like this: f arg1 arg2 = case ... ... -> g arg1 ... -> g arg2 This function is small. So should be safe to inline. However sometimes this doesn't quite work out like that. Consider this code: f1 arg1 arg2 ... = ... case _foo of alt1 -> ... f2 arg1 ... alt2 -> ... f2 arg2 ... f2 arg1 arg2 ... = ... case _foo of alt1 -> ... f3 arg1 ... alt2 -> ... f3 arg2 ... f3 arg1 arg2 ... = ... ... repeats up to n times. And then f1 is applied to some arguments: foo = ... f1 <interestingArgs> ... Initially f2..fn are not interesting to inline so we don't. However we see that f1 is applied to interesting args. So it's an obvious choice to inline those: foo = ... case _foo of alt1 -> ... f2 <interestingArg> ... alt2 -> ... f2 <interestingArg> ... As a result we go and inline f2 both mentions of f2 in turn are now applied to interesting arguments and f2 is small: foo = ... case _foo of alt1 -> ... case _foo of alt1 -> ... f3 <interestingArg> ... alt2 -> ... f3 <interestingArg> ... alt2 -> ... case _foo of alt1 -> ... f3 <interestingArg> ... alt2 -> ... f3 <interestingArg> ... The same thing happens for each binding up to f_n, duplicating the amount of inlining done in each step. Until at some point we are either done or run out of simplifier ticks/RAM. This pattern happened #18730. To combat this we introduce one more heuristic when weighing inlining decision. We keep track of a "case-depth". Which increases each time we look inside a case expression with more than one alternative. We then apply a penalty to inlinings based on the case-depth at which they would be inlined. Bounding the number of inlinings in such a scenario. The heuristic can be tuned in two ways: * We can ignore the first n levels of case nestings for inlining decisions using -funfolding-case-threshold. * The penalty grows linear with the depth. It's computed as size*(depth-threshold)/scaling. Scaling can be set with -funfolding-case-scaling. Some guidance on setting these defaults: * A low treshold (<= 2) is needed to prevent exponential cases from spiraling out of control. We picked 2 for no particular reason. * Scaling the penalty by any more than 30 means the reproducer from T18730 won't compile even with reasonably small values of n. Instead it will run out of runs/ticks. This means to positively affect the reproducer a scaling <= 30 is required. * A scaling of >= 15 still causes a few very large regressions on some nofib benchmarks. (+80% for gc/fulsom, +90% for real/ben-raytrace, +20% for spectral/fibheaps) * A scaling of >= 25 showed no regressions on nofib. However it showed a number of (small) regression for compiler perf benchmarks. The end result is that we are settling for a scaling of 30, with a threshold of 2. This gives us minimal compiler perf regressions. No nofib runtime regressions and will still avoid this pattern sometimes. This is a "safe" default, where we err on the side of compiler blowup instead of risking runtime regressions. For cases where the default falls short the flag can be changed to allow more/less inlining as needed on a per-module basis. -} tryUnfolding :: Logger -> DynFlags -> Int -> Id -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding :: Logger -> DynFlags -> Int -> Id -> Bool -> [ArgSummary] -> CallCtxt -> CoreExpr -> Bool -> Bool -> UnfoldingGuidance -> Maybe CoreExpr tryUnfolding Logger logger DynFlags dflags !Int case_depth Id 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 -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (String -> SDoc text String "UnfNever") forall a. Maybe a Nothing UnfWhen { ug_arity :: UnfoldingGuidance -> Int ug_arity = Int 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 || UnfoldingOpts -> Bool unfoldingVeryAggressive UnfoldingOpts uf_opts) -- See Note [INLINE for small functions (3)] -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc empty Bool True) (forall a. a -> Maybe a Just CoreExpr unf_template) | Bool otherwise -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc empty Bool False) forall a. Maybe a Nothing where some_benefit :: Bool some_benefit = Int -> Bool calc_some_benefit Int uf_arity enough_args :: Bool enough_args = (Int n_val_args forall a. Ord a => a -> a -> Bool >= Int uf_arity) Bool -> Bool -> Bool || (Bool unsat_ok Bool -> Bool -> Bool && Int n_val_args forall a. Ord a => a -> a -> Bool > Int 0) UnfIfGoodArgs { ug_args :: UnfoldingGuidance -> [Int] ug_args = [Int] arg_discounts, ug_res :: UnfoldingGuidance -> Int ug_res = Int res_discount, ug_size :: UnfoldingGuidance -> Int ug_size = Int size } | UnfoldingOpts -> Bool unfoldingVeryAggressive UnfoldingOpts uf_opts -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool True) (forall a. a -> Maybe a Just CoreExpr unf_template) | Bool is_wf Bool -> Bool -> Bool && Bool some_benefit Bool -> Bool -> Bool && Bool small_enough -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool True) (forall a. a -> Maybe a Just CoreExpr unf_template) | Bool otherwise -> forall a. Logger -> DynFlags -> Id -> String -> SDoc -> a -> a traceInline Logger logger DynFlags dflags Id id String str (Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool False) forall a. Maybe a Nothing where some_benefit :: Bool some_benefit = Int -> Bool calc_some_benefit (forall (t :: * -> *) a. Foldable t => t a -> Int length [Int] arg_discounts) extra_doc :: SDoc extra_doc = [SDoc] -> SDoc vcat [ String -> SDoc text String "case depth =" SDoc -> SDoc -> SDoc <+> Int -> SDoc int Int case_depth , String -> SDoc text String "depth based penalty =" SDoc -> SDoc -> SDoc <+> Int -> SDoc int Int depth_penalty , String -> SDoc text String "discounted size =" SDoc -> SDoc -> SDoc <+> Int -> SDoc int Int adjusted_size ] -- See Note [Avoid inlining into deeply nested cases] depth_treshold :: Int depth_treshold = UnfoldingOpts -> Int unfoldingCaseThreshold UnfoldingOpts uf_opts depth_scaling :: Int depth_scaling = UnfoldingOpts -> Int unfoldingCaseScaling UnfoldingOpts uf_opts depth_penalty :: Int depth_penalty | Int case_depth forall a. Ord a => a -> a -> Bool <= Int depth_treshold = Int 0 | Bool otherwise = (Int size forall a. Num a => a -> a -> a * (Int case_depth forall a. Num a => a -> a -> a - Int depth_treshold)) forall a. Integral a => a -> a -> a `div` Int depth_scaling adjusted_size :: Int adjusted_size = Int size forall a. Num a => a -> a -> a + Int depth_penalty forall a. Num a => a -> a -> a - Int discount small_enough :: Bool small_enough = Int adjusted_size forall a. Ord a => a -> a -> Bool <= UnfoldingOpts -> Int unfoldingUseThreshold UnfoldingOpts uf_opts discount :: Int discount = [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int computeDiscount [Int] arg_discounts Int res_discount [ArgSummary] arg_infos CallCtxt cont_info where uf_opts :: UnfoldingOpts uf_opts = DynFlags -> UnfoldingOpts unfoldingOpts DynFlags dflags mk_doc :: Bool -> SDoc -> Bool -> SDoc mk_doc Bool some_benefit SDoc extra_doc Bool yes_or_no = [SDoc] -> SDoc vcat [ String -> SDoc text String "arg infos" SDoc -> SDoc -> SDoc <+> forall a. Outputable a => a -> SDoc ppr [ArgSummary] arg_infos , String -> SDoc text String "interesting continuation" SDoc -> SDoc -> SDoc <+> forall a. Outputable a => a -> SDoc ppr CallCtxt cont_info , String -> SDoc text String "some_benefit" SDoc -> SDoc -> SDoc <+> forall a. Outputable a => a -> SDoc ppr Bool some_benefit , String -> SDoc text String "is exp:" SDoc -> SDoc -> SDoc <+> forall a. Outputable a => a -> SDoc ppr Bool is_exp , String -> SDoc text String "is work-free:" SDoc -> SDoc -> SDoc <+> forall a. Outputable a => a -> SDoc ppr Bool is_wf , String -> SDoc text String "guidance" SDoc -> SDoc -> 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"] ctx :: SDocContext ctx = DynFlags -> PprStyle -> SDocContext initSDocContext DynFlags dflags PprStyle defaultDumpStyle str :: String str = String "Considering inlining: " forall a. [a] -> [a] -> [a] ++ SDocContext -> SDoc -> String showSDocDump SDocContext ctx (forall a. Outputable a => a -> SDoc ppr Id id) n_val_args :: Int n_val_args = forall (t :: * -> *) a. Foldable t => t a -> Int 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 :: Int -> Bool calc_some_benefit Int 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 = Int n_val_args forall a. Ord a => a -> a -> Bool >= Int uf_arity over_saturated :: Bool over_saturated = Int n_val_args forall a. Ord a => a -> a -> Bool > Int uf_arity interesting_args :: Bool interesting_args = 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 -> Int uf_arity forall a. Ord a => a -> a -> Bool > Int 0 -- See Note [Unfold info lazy contexts] CallCtxt DiscArgCtxt -> Int uf_arity forall a. Ord a => a -> a -> Bool > Int 0 -- Note [Inlining in ArgCtxt] CallCtxt RhsCtxt -> Int uf_arity forall a. Ord a => a -> a -> Bool > Int 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 GHC.Core.Utils) 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 :: [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int computeDiscount :: [Int] -> Int -> [ArgSummary] -> CallCtxt -> Int computeDiscount [Int] arg_discounts Int res_discount [ArgSummary] arg_infos CallCtxt cont_info = Int 10 -- Discount of 10 because the result replaces the call -- so we count 10 for the function itself forall a. Num a => a -> a -> a + Int 10 forall a. Num a => a -> a -> a * forall (t :: * -> *) a. Foldable t => t a -> Int length [Int] actual_arg_discounts -- Discount of 10 for each arg supplied, -- because the result replaces the call forall a. Num a => a -> a -> a + Int total_arg_discount forall a. Num a => a -> a -> a + Int res_discount' where actual_arg_discounts :: [Int] actual_arg_discounts = forall a b c. (a -> b -> c) -> [a] -> [b] -> [c] zipWith forall {a}. Num a => a -> ArgSummary -> a mk_arg_discount [Int] arg_discounts [ArgSummary] arg_infos total_arg_discount :: Int total_arg_discount = forall (t :: * -> *) a. (Foldable t, Num a) => t a -> a sum [Int] actual_arg_discounts mk_arg_discount :: a -> ArgSummary -> a mk_arg_discount a _ ArgSummary TrivArg = a 0 mk_arg_discount a _ ArgSummary NonTrivArg = a 10 mk_arg_discount a discount ArgSummary ValueArg = a discount res_discount' :: Int res_discount' | Ordering LT <- [Int] arg_discounts forall a b. [a] -> [b] -> Ordering `compareLength` [ArgSummary] arg_infos = Int res_discount -- Over-saturated | Bool otherwise = case CallCtxt cont_info of CallCtxt BoringCtxt -> Int 0 CallCtxt CaseCtxt -> Int res_discount -- Presumably a constructor CallCtxt ValAppCtxt -> Int res_discount -- Presumably a function CallCtxt _ -> Int 40 forall a. Ord a => a -> a -> a `min` Int 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 avoid inlining large functions that return -- constructors into contexts that are simply "interesting"