{- (c) The GRASP/AQUA Project, Glasgow University, 1993-1998 \section[WorkWrap]{Worker/wrapper-generating back-end of strictness analyser} -} {-# LANGUAGE CPP #-} module GHC.Core.Opt.WorkWrap ( wwTopBinds ) where import GHC.Prelude import GHC.Core.Opt.Arity ( manifestArity ) import GHC.Core import GHC.Core.Unfold import GHC.Core.Unfold.Make import GHC.Core.Utils ( exprType, exprIsHNF ) import GHC.Core.FVs ( exprFreeVars ) import GHC.Types.Var import GHC.Types.Id import GHC.Types.Id.Info import GHC.Core.Type import GHC.Types.Unique.Supply import GHC.Types.Basic import GHC.Driver.Session import GHC.Driver.Ppr import GHC.Driver.Config import GHC.Types.Demand import GHC.Types.Cpr import GHC.Types.SourceText import GHC.Core.Opt.WorkWrap.Utils import GHC.Utils.Misc import GHC.Utils.Outputable import GHC.Types.Unique import GHC.Utils.Panic import GHC.Core.FamInstEnv import GHC.Utils.Monad #include "HsVersions.h" {- We take Core bindings whose binders have: \begin{enumerate} \item Strictness attached (by the front-end of the strictness analyser), and / or \item Constructed Product Result information attached by the CPR analysis pass. \end{enumerate} and we return some ``plain'' bindings which have been worker/wrapper-ified, meaning: \begin{enumerate} \item Functions have been split into workers and wrappers where appropriate. If a function has both strictness and CPR properties then only one worker/wrapper doing both transformations is produced; \item Binders' @IdInfos@ have been updated to reflect the existence of these workers/wrappers (this is where we get STRICTNESS and CPR pragma info for exported values). \end{enumerate} -} wwTopBinds :: DynFlags -> FamInstEnvs -> UniqSupply -> CoreProgram -> CoreProgram wwTopBinds :: DynFlags -> FamInstEnvs -> UniqSupply -> CoreProgram -> CoreProgram wwTopBinds DynFlags dflags FamInstEnvs fam_envs UniqSupply us CoreProgram top_binds = UniqSupply -> UniqSM CoreProgram -> CoreProgram forall a. UniqSupply -> UniqSM a -> a initUs_ UniqSupply us (UniqSM CoreProgram -> CoreProgram) -> UniqSM CoreProgram -> CoreProgram forall a b. (a -> b) -> a -> b $ do [CoreProgram] top_binds' <- (CoreBind -> UniqSM CoreProgram) -> CoreProgram -> UniqSM [CoreProgram] forall (t :: * -> *) (m :: * -> *) a b. (Traversable t, Monad m) => (a -> m b) -> t a -> m (t b) mapM (DynFlags -> FamInstEnvs -> CoreBind -> UniqSM CoreProgram wwBind DynFlags dflags FamInstEnvs fam_envs) CoreProgram top_binds CoreProgram -> UniqSM CoreProgram forall (m :: * -> *) a. Monad m => a -> m a return ([CoreProgram] -> CoreProgram forall (t :: * -> *) a. Foldable t => t [a] -> [a] concat [CoreProgram] top_binds') {- ************************************************************************ * * \subsection[wwBind-wwExpr]{@wwBind@ and @wwExpr@} * * ************************************************************************ @wwBind@ works on a binding, trying each \tr{(binder, expr)} pair in turn. Non-recursive case first, then recursive... -} wwBind :: DynFlags -> FamInstEnvs -> CoreBind -> UniqSM [CoreBind] -- returns a WwBinding intermediate form; -- the caller will convert to Expr/Binding, -- as appropriate. wwBind :: DynFlags -> FamInstEnvs -> CoreBind -> UniqSM CoreProgram wwBind DynFlags dflags FamInstEnvs fam_envs (NonRec Var binder Expr Var rhs) = do Expr Var new_rhs <- DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var rhs [(Var, Expr Var)] new_pairs <- DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] tryWW DynFlags dflags FamInstEnvs fam_envs RecFlag NonRecursive Var binder Expr Var new_rhs CoreProgram -> UniqSM CoreProgram forall (m :: * -> *) a. Monad m => a -> m a return [Var -> Expr Var -> CoreBind forall b. b -> Expr b -> Bind b NonRec Var b Expr Var e | (Var b,Expr Var e) <- [(Var, Expr Var)] new_pairs] -- Generated bindings must be non-recursive -- because the original binding was. wwBind DynFlags dflags FamInstEnvs fam_envs (Rec [(Var, Expr Var)] pairs) = CoreBind -> CoreProgram forall (m :: * -> *) a. Monad m => a -> m a return (CoreBind -> CoreProgram) -> ([(Var, Expr Var)] -> CoreBind) -> [(Var, Expr Var)] -> CoreProgram forall b c a. (b -> c) -> (a -> b) -> a -> c . [(Var, Expr Var)] -> CoreBind forall b. [(b, Expr b)] -> Bind b Rec ([(Var, Expr Var)] -> CoreProgram) -> UniqSM [(Var, Expr Var)] -> UniqSM CoreProgram forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b <$> ((Var, Expr Var) -> UniqSM [(Var, Expr Var)]) -> [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a b. Monad m => (a -> m [b]) -> [a] -> m [b] concatMapM (Var, Expr Var) -> UniqSM [(Var, Expr Var)] do_one [(Var, Expr Var)] pairs where do_one :: (Var, Expr Var) -> UniqSM [(Var, Expr Var)] do_one (Var binder, Expr Var rhs) = do Expr Var new_rhs <- DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var rhs DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] tryWW DynFlags dflags FamInstEnvs fam_envs RecFlag Recursive Var binder Expr Var new_rhs {- @wwExpr@ basically just walks the tree, looking for appropriate annotations that can be used. Remember it is @wwBind@ that does the matching by looking for strict arguments of the correct type. @wwExpr@ is a version that just returns the ``Plain'' Tree. -} wwExpr :: DynFlags -> FamInstEnvs -> CoreExpr -> UniqSM CoreExpr wwExpr :: DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags _ FamInstEnvs _ e :: Expr Var e@(Type {}) = Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return Expr Var e wwExpr DynFlags _ FamInstEnvs _ e :: Expr Var e@(Coercion {}) = Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return Expr Var e wwExpr DynFlags _ FamInstEnvs _ e :: Expr Var e@(Lit {}) = Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return Expr Var e wwExpr DynFlags _ FamInstEnvs _ e :: Expr Var e@(Var {}) = Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return Expr Var e wwExpr DynFlags dflags FamInstEnvs fam_envs (Lam Var binder Expr Var expr) = Var -> Expr Var -> Expr Var forall b. b -> Expr b -> Expr b Lam Var new_binder (Expr Var -> Expr Var) -> UniqSM (Expr Var) -> UniqSM (Expr Var) forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b <$> DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var expr where new_binder :: Var new_binder | Var -> Bool isId Var binder = Var -> Var zapIdUsedOnceInfo Var binder | Bool otherwise = Var binder -- See Note [Zapping Used Once info in WorkWrap] wwExpr DynFlags dflags FamInstEnvs fam_envs (App Expr Var f Expr Var a) = Expr Var -> Expr Var -> Expr Var forall b. Expr b -> Expr b -> Expr b App (Expr Var -> Expr Var -> Expr Var) -> UniqSM (Expr Var) -> UniqSM (Expr Var -> Expr Var) forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b <$> DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var f UniqSM (Expr Var -> Expr Var) -> UniqSM (Expr Var) -> UniqSM (Expr Var) forall (f :: * -> *) a b. Applicative f => f (a -> b) -> f a -> f b <*> DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var a wwExpr DynFlags dflags FamInstEnvs fam_envs (Tick CoreTickish note Expr Var expr) = CoreTickish -> Expr Var -> Expr Var forall b. CoreTickish -> Expr b -> Expr b Tick CoreTickish note (Expr Var -> Expr Var) -> UniqSM (Expr Var) -> UniqSM (Expr Var) forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b <$> DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var expr wwExpr DynFlags dflags FamInstEnvs fam_envs (Cast Expr Var expr CoercionR co) = do Expr Var new_expr <- DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var expr Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return (Expr Var -> CoercionR -> Expr Var forall b. Expr b -> CoercionR -> Expr b Cast Expr Var new_expr CoercionR co) wwExpr DynFlags dflags FamInstEnvs fam_envs (Let CoreBind bind Expr Var expr) = CoreProgram -> Expr Var -> Expr Var forall b. [Bind b] -> Expr b -> Expr b mkLets (CoreProgram -> Expr Var -> Expr Var) -> UniqSM CoreProgram -> UniqSM (Expr Var -> Expr Var) forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b <$> DynFlags -> FamInstEnvs -> CoreBind -> UniqSM CoreProgram wwBind DynFlags dflags FamInstEnvs fam_envs CoreBind bind UniqSM (Expr Var -> Expr Var) -> UniqSM (Expr Var) -> UniqSM (Expr Var) forall (f :: * -> *) a b. Applicative f => f (a -> b) -> f a -> f b <*> DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var expr wwExpr DynFlags dflags FamInstEnvs fam_envs (Case Expr Var expr Var binder Type ty [Alt Var] alts) = do Expr Var new_expr <- DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var expr [Alt Var] new_alts <- (Alt Var -> UniqSM (Alt Var)) -> [Alt Var] -> UniqSM [Alt Var] forall (t :: * -> *) (m :: * -> *) a b. (Traversable t, Monad m) => (a -> m b) -> t a -> m (t b) mapM Alt Var -> UniqSM (Alt Var) ww_alt [Alt Var] alts let new_binder :: Var new_binder = Var -> Var zapIdUsedOnceInfo Var binder -- See Note [Zapping Used Once info in WorkWrap] Expr Var -> UniqSM (Expr Var) forall (m :: * -> *) a. Monad m => a -> m a return (Expr Var -> Var -> Type -> [Alt Var] -> Expr Var forall b. Expr b -> b -> Type -> [Alt b] -> Expr b Case Expr Var new_expr Var new_binder Type ty [Alt Var] new_alts) where ww_alt :: Alt Var -> UniqSM (Alt Var) ww_alt (Alt AltCon con [Var] binders Expr Var rhs) = do Expr Var new_rhs <- DynFlags -> FamInstEnvs -> Expr Var -> UniqSM (Expr Var) wwExpr DynFlags dflags FamInstEnvs fam_envs Expr Var rhs let new_binders :: [Var] new_binders = [ if Var -> Bool isId Var b then Var -> Var zapIdUsedOnceInfo Var b else Var b | Var b <- [Var] binders ] -- See Note [Zapping Used Once info in WorkWrap] Alt Var -> UniqSM (Alt Var) forall (m :: * -> *) a. Monad m => a -> m a return (AltCon -> [Var] -> Expr Var -> Alt Var forall b. AltCon -> [b] -> Expr b -> Alt b Alt AltCon con [Var] new_binders Expr Var new_rhs) {- ************************************************************************ * * \subsection[tryWW]{@tryWW@: attempt a worker/wrapper pair} * * ************************************************************************ @tryWW@ just accumulates arguments, converts strictness info from the front-end into the proper form, then calls @mkWwBodies@ to do the business. The only reason this is monadised is for the unique supply. Note [Don't w/w INLINE things] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ It's very important to refrain from w/w-ing an INLINE function (ie one with a stable unfolding) because the wrapper will then overwrite the old stable unfolding with the wrapper code. Furthermore, if the programmer has marked something as INLINE, we may lose by w/w'ing it. If the strictness analyser is run twice, this test also prevents wrappers (which are INLINEd) from being re-done. (You can end up with several liked-named Ids bouncing around at the same time---absolute mischief.) Notice that we refrain from w/w'ing an INLINE function even if it is in a recursive group. It might not be the loop breaker. (We could test for loop-breaker-hood, but I'm not sure that ever matters.) Note [Worker-wrapper for INLINABLE functions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ If we have {-# INLINABLE f #-} f :: Ord a => [a] -> Int -> a f x y = ....f.... where f is strict in y, we might get a more efficient loop by w/w'ing f. But that would make a new unfolding which would overwrite the old one! So the function would no longer be INLNABLE, and in particular will not be specialised at call sites in other modules. This comes in practice (#6056). Solution: do the w/w for strictness analysis, but transfer the Stable unfolding to the *worker*. So we will get something like this: {-# INLINE[0] f #-} f :: Ord a => [a] -> Int -> a f d x y = case y of I# y' -> fw d x y' {-# INLINABLE[0] fw #-} fw :: Ord a => [a] -> Int# -> a fw d x y' = let y = I# y' in ...f... How do we "transfer the unfolding"? Easy: by using the old one, wrapped in work_fn! See GHC.Core.Unfold.mkWorkerUnfolding. Note [No worker-wrapper for record selectors] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ We sometimes generate a lot of record selectors, and generally the don't benefit from worker/wrapper. Yes, mkWwBodies would find a w/w split, but it is then suppressed by the certainlyWillInline test in splitFun. The wasted effort in mkWwBodies makes a measurable difference in compile time (see MR !2873), so although it's a terribly ad-hoc test, we just check here for record selectors, and do a no-op in that case. I did look for a generalisation, so that it's not just record selectors that benefit. But you'd need a cheap test for "this function will definitely get a w/w split" and that's hard to predict in advance...the logic in mkWwBodies is complex. So I've left the super-simple test, with this Note to explain. Note [Worker-wrapper for NOINLINE functions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ We used to disable worker/wrapper for NOINLINE things, but it turns out this can cause unnecessary reboxing of values. Consider {-# NOINLINE f #-} f :: Int -> a f x = error (show x) g :: Bool -> Bool -> Int -> Int g True True p = f p g False True p = p + 1 g b False p = g b True p the strictness analysis will discover f and g are strict, but because f has no wrapper, the worker for g will rebox p. So we get $wg x y p# = let p = I# p# in -- Yikes! Reboxing! case x of False -> case y of False -> $wg False True p# True -> +# p# 1# True -> case y of False -> $wg True True p# True -> case f p of { } g x y p = case p of (I# p#) -> $wg x y p# Now, in this case the reboxing will float into the True branch, and so the allocation will only happen on the error path. But it won't float inwards if there are multiple branches that call (f p), so the reboxing will happen on every call of g. Disaster. Solution: do worker/wrapper even on NOINLINE things; but move the NOINLINE pragma to the worker. (See #13143 for a real-world example.) It is crucial that we do this for *all* NOINLINE functions. #10069 demonstrates what happens when we promise to w/w a (NOINLINE) leaf function, but fail to deliver: data C = C Int# Int# {-# NOINLINE c1 #-} c1 :: C -> Int# c1 (C _ n) = n {-# NOINLINE fc #-} fc :: C -> Int# fc c = 2 *# c1 c Failing to w/w `c1`, but still w/wing `fc` leads to the following code: c1 :: C -> Int# c1 (C _ n) = n $wfc :: Int# -> Int# $wfc n = let c = C 0# n in 2 #* c1 c fc :: C -> Int# fc (C _ n) = $wfc n Yikes! The reboxed `C` in `$wfc` can't cancel out, so we are in a bad place. This generalises to any function that derives its strictness signature from its callees, so we have to make sure that when a function announces particular strictness properties, we have to w/w them accordingly, even if it means splitting a NOINLINE function. Note [Worker activation] ~~~~~~~~~~~~~~~~~~~~~~~~ Follows on from Note [Worker-wrapper for INLINABLE functions] It is *vital* that if the worker gets an INLINABLE pragma (from the original function), then the worker has the same phase activation as the wrapper (or later). That is necessary to allow the wrapper to inline into the worker's unfolding: see GHC.Core.Opt.Simplify.Utils Note [Simplifying inside stable unfoldings]. If the original is NOINLINE, it's important that the work inherit the original activation. Consider {-# NOINLINE expensive #-} expensive x = x + 1 f y = let z = expensive y in ... If expensive's worker inherits the wrapper's activation, we'll get this (because of the compromise in point (2) of Note [Wrapper activation]) {-# NOINLINE[0] $wexpensive #-} $wexpensive x = x + 1 {-# INLINE[0] expensive #-} expensive x = $wexpensive x f y = let z = expensive y in ... and $wexpensive will be immediately inlined into expensive, followed by expensive into f. This effectively removes the original NOINLINE! Otherwise, nothing is lost by giving the worker the same activation as the wrapper, because the worker won't have any chance of inlining until the wrapper does; there's no point in giving it an earlier activation. Note [Don't w/w inline small non-loop-breaker things] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In general, we refrain from w/w-ing *small* functions, which are not loop breakers, because they'll inline anyway. But we must take care: it may look small now, but get to be big later after other inlining has happened. So we take the precaution of adding a StableUnfolding for any such functions. I made this change when I observed a big function at the end of compilation with a useful strictness signature but no w-w. (It was small during demand analysis, we refrained from w/w, and then got big when something was inlined in its rhs.) When I measured it on nofib, it didn't make much difference; just a few percent improved allocation on one benchmark (bspt/Euclid.space). But nothing got worse. There is an infelicity though. We may get something like f = g val ==> g x = case gw x of r -> I# r f {- InlineStable, Template = g val -} f = case gw x of r -> I# r The code for f duplicates that for g, without any real benefit. It won't really be executed, because calls to f will go via the inlining. Note [Don't w/w join points for CPR] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ There's no point in exploiting CPR info on a join point. If the whole function is getting CPR'd, then the case expression around the worker function will get pushed into the join point by the simplifier, which will have the same effect that w/w'ing for CPR would have - the result will be returned in an unboxed tuple. f z = let join j x y = (x+1, y+1) in case z of A -> j 1 2 B -> j 2 3 => f z = case $wf z of (# a, b #) -> (a, b) $wf z = case (let join j x y = (x+1, y+1) in case z of A -> j 1 2 B -> j 2 3) of (a, b) -> (# a, b #) => f z = case $wf z of (# a, b #) -> (a, b) $wf z = let join j x y = (# x+1, y+1 #) in case z of A -> j 1 2 B -> j 2 3 Note that we still want to give @j@ the CPR property, so that @f@ has it. So CPR *analyse* join points as regular functions, but don't *transform* them. Doing W/W for returned products on a join point would be tricky anyway, as the worker could not be a join point because it would not be tail-called. However, doing the *argument* part of W/W still works for join points, since the wrapper body will make a tail call: f z = let join j x y = x + y in ... => f z = let join $wj x# y# = x# +# y# j x y = case x of I# x# -> case y of I# y# -> $wj x# y# in ... Note [Wrapper activation] ~~~~~~~~~~~~~~~~~~~~~~~~~ When should the wrapper inlining be active? 1. It must not be active earlier than the current Activation of the Id 2. It should be active at some point, despite (1) because of Note [Worker-wrapper for NOINLINE functions] 3. For ordinary functions with no pragmas we want to inline the wrapper as early as possible (#15056). Suppose another module defines f x = g x x and suppose there is some RULE for (g True True). Then if we have a call (f True), we'd expect to inline 'f' and the RULE will fire. But if f is w/w'd (which it might be), we want the inlining to occur just as if it hadn't been. (This only matters if f's RHS is big enough to w/w, but small enough to inline given the call site, but that can happen.) 4. We do not want to inline the wrapper before specialisation. module Foo where f :: Num a => a -> Int -> a f n 0 = n -- Strict in the Int, hence wrapper f n x = f (n+n) (x-1) g :: Int -> Int g x = f x x -- Provokes a specialisation for f module Bar where import Foo h :: Int -> Int h x = f 3 x In module Bar we want to give specialisations a chance to fire before inlining f's wrapper. Historical note: At one stage I tried making the wrapper inlining always-active, and that had a very bad effect on nofib/imaginary/x2n1; a wrapper was inlined before the specialisation fired. Reminder: Note [Don't w/w INLINE things], so we don't need to worry about INLINE things here. Conclusion: - If the user said NOINLINE[n], respect that - If the user said NOINLINE, inline the wrapper only after phase 0, the last user-visible phase. That means that all rules will have had a chance to fire. What phase is after phase 0? Answer: FinalPhase, that's the reason it exists. NB: Similar to InitialPhase, users can't write INLINE[Final] f; it's syntactically illegal. - Otherwise inline wrapper in phase 2. That allows the 'gentle' simplification pass to apply specialisation rules Note [Wrapper NoUserInlinePrag] ~~~~~~~~~~~~~~~~~~~~~~~~~~~ We use NoUserInlinePrag on the wrapper, to say that there is no user-specified inline pragma. (The worker inherits that; see Note [Worker-wrapper for INLINABLE functions].) The wrapper has no pragma given by the user. (Historical note: we used to give the wrapper an INLINE pragma, but CSE will not happen if there is a user-specified pragma, but should happen for w/w’ed things (#14186). We don't need a pragma, because everything we needs is expressed by (a) the stable unfolding and (b) the inl_act activation.) -} tryWW :: DynFlags -> FamInstEnvs -> RecFlag -> Id -- The fn binder -> CoreExpr -- The bound rhs; its innards -- are already ww'd -> UniqSM [(Id, CoreExpr)] -- either *one* or *two* pairs; -- if one, then no worker (only -- the orig "wrapper" lives on); -- if two, then a worker and a -- wrapper. tryWW :: DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] tryWW DynFlags dflags FamInstEnvs fam_envs RecFlag is_rec Var fn_id Expr Var rhs -- See Note [Worker-wrapper for NOINLINE functions] | Just Unfolding stable_unf <- UnfoldingOpts -> IdInfo -> Maybe Unfolding certainlyWillInline UnfoldingOpts uf_opts IdInfo fn_info = [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [ (Var fn_id Var -> Unfolding -> Var `setIdUnfolding` Unfolding stable_unf, Expr Var rhs) ] -- See Note [Don't w/w INLINE things] -- See Note [Don't w/w inline small non-loop-breaker things] | Bool is_fun Bool -> Bool -> Bool && Bool is_eta_exp = DynFlags -> FamInstEnvs -> Var -> IdInfo -> [Demand] -> Divergence -> Cpr -> Expr Var -> UniqSM [(Var, Expr Var)] splitFun DynFlags dflags FamInstEnvs fam_envs Var new_fn_id IdInfo fn_info [Demand] wrap_dmds Divergence div Cpr cpr Expr Var rhs | RecFlag -> Bool isNonRec RecFlag is_rec, Bool is_thunk -- See Note [Thunk splitting] = DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] splitThunk DynFlags dflags FamInstEnvs fam_envs RecFlag is_rec Var new_fn_id Expr Var rhs | Bool otherwise = [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [ (Var new_fn_id, Expr Var rhs) ] where uf_opts :: UnfoldingOpts uf_opts = DynFlags -> UnfoldingOpts unfoldingOpts DynFlags dflags fn_info :: IdInfo fn_info = HasDebugCallStack => Var -> IdInfo Var -> IdInfo idInfo Var fn_id ([Demand] wrap_dmds, Divergence div) = StrictSig -> ([Demand], Divergence) splitStrictSig (IdInfo -> StrictSig strictnessInfo IdInfo fn_info) cpr_ty :: CprType cpr_ty = CprSig -> CprType getCprSig (IdInfo -> CprSig cprInfo IdInfo fn_info) -- Arity of the CPR sig should match idArity when it's not a join point. -- See Note [Arity trimming for CPR signatures] in GHC.Core.Opt.CprAnal cpr :: Cpr cpr = ASSERT2( isJoinId fn_id || cpr_ty == topCprType || ct_arty cpr_ty == arityInfo fn_info , ppr fn_id <> colon <+> text "ct_arty:" <+> int (ct_arty cpr_ty) <+> text "arityInfo:" <+> ppr (arityInfo fn_info)) CprType -> Cpr ct_cpr CprType cpr_ty new_fn_id :: Var new_fn_id = Var -> Var zapIdUsedOnceInfo (Var -> Var zapIdUsageEnvInfo Var fn_id) -- See Note [Zapping DmdEnv after Demand Analyzer] and -- See Note [Zapping Used Once info WorkWrap] is_fun :: Bool is_fun = [Demand] -> Bool forall (f :: * -> *) a. Foldable f => f a -> Bool notNull [Demand] wrap_dmds Bool -> Bool -> Bool || Var -> Bool isJoinId Var fn_id -- See Note [Don't eta expand in w/w] is_eta_exp :: Bool is_eta_exp = [Demand] -> ArityInfo forall (t :: * -> *) a. Foldable t => t a -> ArityInfo length [Demand] wrap_dmds ArityInfo -> ArityInfo -> Bool forall a. Eq a => a -> a -> Bool == Expr Var -> ArityInfo manifestArity Expr Var rhs is_thunk :: Bool is_thunk = Bool -> Bool not Bool is_fun Bool -> Bool -> Bool && Bool -> Bool not (Expr Var -> Bool exprIsHNF Expr Var rhs) Bool -> Bool -> Bool && Bool -> Bool not (Var -> Bool isJoinId Var fn_id) Bool -> Bool -> Bool && Bool -> Bool not (HasDebugCallStack => Type -> Bool Type -> Bool isUnliftedType (Var -> Type idType Var fn_id)) {- Note [Zapping DmdEnv after Demand Analyzer] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In the worker-wrapper pass we zap the DmdEnv. Why? (a) it is never used again (b) it wastes space (c) it becomes incorrect as things are cloned, because we don't push the substitution into it Why here? * Because we don’t want to do it in the Demand Analyzer, as we never know there when we are doing the last pass. * We want them to be still there at the end of DmdAnal, so that -ddump-str-anal contains them. * We don’t want a second pass just for that. * WorkWrap looks at all bindings anyway. We also need to do it in TidyCore.tidyLetBndr to clean up after the final, worker/wrapper-less run of the demand analyser (see Note [Final Demand Analyser run] in GHC.Core.Opt.DmdAnal). Note [Zapping Used Once info in WorkWrap] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In the worker-wrapper pass we zap the used once info in demands and in strictness signatures. Why? * The simplifier may happen to transform code in a way that invalidates the data (see #11731 for an example). * It is not used in later passes, up to code generation. So as the data is useless and possibly wrong, we want to remove it. The most convenient place to do that is the worker wrapper phase, as it runs after every run of the demand analyser besides the very last one (which is the one where we want to _keep_ the info for the code generator). We do not do it in the demand analyser for the same reasons outlined in Note [Zapping DmdEnv after Demand Analyzer] above. Note [Don't eta expand in w/w] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ A binding where the manifestArity of the RHS is less than idArity of the binder means GHC.Core.Opt.Arity didn't eta expand that binding. When this happens, it does so for a reason (see Note [exprArity invariant] in GHC.Core.Opt.Arity) and we probably have a PAP, cast or trivial expression as RHS. Performing the worker/wrapper split will implicitly eta-expand the binding to idArity, overriding GHC.Core.Opt.Arity's decision. Other than playing fast and loose with divergence, it's also broken for newtypes: f = (\xy.blah) |> co where co :: (Int -> Int -> Char) ~ T Then idArity is 2 (despite the type T), and it can have a StrictSig based on a threshold of 2. But we can't w/w it without a type error. The situation is less grave for PAPs, but the implicit eta expansion caused a compiler allocation regression in T15164, where huge recursive instance method groups, mostly consisting of PAPs, got w/w'd. This caused great churn in the simplifier, when simply waiting for the PAPs to inline arrived at the same output program. Note there is the worry here that such PAPs and trivial RHSs might not *always* be inlined. That would lead to reboxing, because the analysis tacitly assumes that we W/W'd for idArity and will propagate analysis information under that assumption. So far, this doesn't seem to matter in practice. See https://gitlab.haskell.org/ghc/ghc/merge_requests/312#note_192064. -} --------------------- splitFun :: DynFlags -> FamInstEnvs -> Id -> IdInfo -> [Demand] -> Divergence -> Cpr -> CoreExpr -> UniqSM [(Id, CoreExpr)] splitFun :: DynFlags -> FamInstEnvs -> Var -> IdInfo -> [Demand] -> Divergence -> Cpr -> Expr Var -> UniqSM [(Var, Expr Var)] splitFun DynFlags dflags FamInstEnvs fam_envs Var fn_id IdInfo fn_info [Demand] wrap_dmds Divergence div Cpr cpr Expr Var rhs | Var -> Bool isRecordSelector Var fn_id -- See Note [No worker/wrapper for record selectors] = [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [ (Var fn_id, Expr Var rhs ) ] | Bool otherwise = WARN( not (wrap_dmds `lengthIs` arity), ppr fn_id <+> (ppr arity $$ ppr wrap_dmds $$ ppr cpr) ) -- The arity should match the signature do { Maybe WwResult mb_stuff <- DynFlags -> FamInstEnvs -> VarSet -> Var -> [Demand] -> Cpr -> UniqSM (Maybe WwResult) mkWwBodies DynFlags dflags FamInstEnvs fam_envs VarSet rhs_fvs Var fn_id [Demand] wrap_dmds Cpr use_cpr_info ; case Maybe WwResult mb_stuff of Maybe WwResult Nothing -> [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [(Var fn_id, Expr Var rhs)] Just WwResult stuff | Just Unfolding stable_unf <- UnfoldingOpts -> IdInfo -> Maybe Unfolding certainlyWillInline (DynFlags -> UnfoldingOpts unfoldingOpts DynFlags dflags) IdInfo fn_info -> [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [ (Var fn_id Var -> Unfolding -> Var `setIdUnfolding` Unfolding stable_unf, Expr Var rhs) ] -- See Note [Don't w/w INLINE things] -- See Note [Don't w/w inline small non-loop-breaker things] | Bool otherwise -> do { Unique work_uniq <- UniqSM Unique forall (m :: * -> *). MonadUnique m => m Unique getUniqueM ; [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return (DynFlags -> Var -> IdInfo -> ArityInfo -> Expr Var -> Unique -> Divergence -> Cpr -> WwResult -> [(Var, Expr Var)] mkWWBindPair DynFlags dflags Var fn_id IdInfo fn_info ArityInfo arity Expr Var rhs Unique work_uniq Divergence div Cpr cpr WwResult stuff) } } where rhs_fvs :: VarSet rhs_fvs = Expr Var -> VarSet exprFreeVars Expr Var rhs arity :: ArityInfo arity = IdInfo -> ArityInfo arityInfo IdInfo fn_info -- The arity is set by the simplifier using exprEtaExpandArity -- So it may be more than the number of top-level-visible lambdas -- use_cpr_info is the CPR we w/w for. Note that we kill it for join points, -- see Note [Don't w/w join points for CPR]. use_cpr_info :: Cpr use_cpr_info | Var -> Bool isJoinId Var fn_id = Cpr topCpr | Bool otherwise = Cpr cpr mkWWBindPair :: DynFlags -> Id -> IdInfo -> Arity -> CoreExpr -> Unique -> Divergence -> Cpr -> ([Demand], JoinArity, Id -> CoreExpr, Expr CoreBndr -> CoreExpr) -> [(Id, CoreExpr)] mkWWBindPair :: DynFlags -> Var -> IdInfo -> ArityInfo -> Expr Var -> Unique -> Divergence -> Cpr -> WwResult -> [(Var, Expr Var)] mkWWBindPair DynFlags dflags Var fn_id IdInfo fn_info ArityInfo arity Expr Var rhs Unique work_uniq Divergence div Cpr cpr ([Demand] work_demands, ArityInfo join_arity, Var -> Expr Var wrap_fn, Expr Var -> Expr Var work_fn) = [(Var work_id, Expr Var work_rhs), (Var wrap_id, Expr Var wrap_rhs)] -- Worker first, because wrapper mentions it where simpl_opts :: SimpleOpts simpl_opts = DynFlags -> SimpleOpts initSimpleOpts DynFlags dflags work_rhs :: Expr Var work_rhs = Expr Var -> Expr Var work_fn Expr Var rhs work_act :: Activation work_act = case InlineSpec fn_inline_spec of -- See Note [Worker activation] InlineSpec NoInline -> InlinePragma -> Activation inl_act InlinePragma fn_inl_prag InlineSpec _ -> InlinePragma -> Activation inl_act InlinePragma wrap_prag work_prag :: InlinePragma work_prag = InlinePragma { inl_src :: SourceText inl_src = String -> SourceText SourceText String "{-# INLINE" , inl_inline :: InlineSpec inl_inline = InlineSpec fn_inline_spec , inl_sat :: Maybe ArityInfo inl_sat = Maybe ArityInfo forall a. Maybe a Nothing , inl_act :: Activation inl_act = Activation work_act , inl_rule :: RuleMatchInfo inl_rule = RuleMatchInfo FunLike } -- inl_inline: copy from fn_id; see Note [Worker-wrapper for INLINABLE functions] -- inl_act: see Note [Worker activation] -- inl_rule: it does not make sense for workers to be constructorlike. work_join_arity :: Maybe ArityInfo work_join_arity | Var -> Bool isJoinId Var fn_id = ArityInfo -> Maybe ArityInfo forall a. a -> Maybe a Just ArityInfo join_arity | Bool otherwise = Maybe ArityInfo forall a. Maybe a Nothing -- worker is join point iff wrapper is join point -- (see Note [Don't w/w join points for CPR]) work_id :: Var work_id = Unique -> Var -> Type -> Var mkWorkerId Unique work_uniq Var fn_id (Expr Var -> Type exprType Expr Var work_rhs) Var -> OccInfo -> Var `setIdOccInfo` IdInfo -> OccInfo occInfo IdInfo fn_info -- Copy over occurrence info from parent -- Notably whether it's a loop breaker -- Doesn't matter much, since we will simplify next, but -- seems right-er to do so Var -> InlinePragma -> Var `setInlinePragma` InlinePragma work_prag Var -> Unfolding -> Var `setIdUnfolding` SimpleOpts -> (Expr Var -> Expr Var) -> Unfolding -> Unfolding mkWorkerUnfolding SimpleOpts simpl_opts Expr Var -> Expr Var work_fn Unfolding fn_unfolding -- See Note [Worker-wrapper for INLINABLE functions] Var -> StrictSig -> Var `setIdStrictness` [Demand] -> Divergence -> StrictSig mkClosedStrictSig [Demand] work_demands Divergence div -- Even though we may not be at top level, -- it's ok to give it an empty DmdEnv Var -> CprSig -> Var `setIdCprInfo` ArityInfo -> Cpr -> CprSig mkCprSig ArityInfo work_arity Cpr work_cpr_info Var -> Demand -> Var `setIdDemandInfo` Demand worker_demand Var -> ArityInfo -> Var `setIdArity` ArityInfo work_arity -- Set the arity so that the Core Lint check that the -- arity is consistent with the demand type goes -- through Var -> Maybe ArityInfo -> Var `asJoinId_maybe` Maybe ArityInfo work_join_arity work_arity :: ArityInfo work_arity = [Demand] -> ArityInfo forall (t :: * -> *) a. Foldable t => t a -> ArityInfo length [Demand] work_demands -- See Note [Demand on the Worker] single_call :: Bool single_call = ArityInfo -> Demand -> Bool saturatedByOneShots ArityInfo arity (IdInfo -> Demand demandInfo IdInfo fn_info) worker_demand :: Demand worker_demand | Bool single_call = ArityInfo -> Demand mkWorkerDemand ArityInfo work_arity | Bool otherwise = Demand topDmd wrap_rhs :: Expr Var wrap_rhs = Var -> Expr Var wrap_fn Var work_id wrap_prag :: InlinePragma wrap_prag = InlinePragma -> InlinePragma mkStrWrapperInlinePrag InlinePragma fn_inl_prag wrap_id :: Var wrap_id = Var fn_id Var -> Unfolding -> Var `setIdUnfolding` SimpleOpts -> Expr Var -> ArityInfo -> Unfolding mkWwInlineRule SimpleOpts simpl_opts Expr Var wrap_rhs ArityInfo arity Var -> InlinePragma -> Var `setInlinePragma` InlinePragma wrap_prag Var -> OccInfo -> Var `setIdOccInfo` OccInfo noOccInfo -- Zap any loop-breaker-ness, to avoid bleating from Lint -- about a loop breaker with an INLINE rule fn_inl_prag :: InlinePragma fn_inl_prag = IdInfo -> InlinePragma inlinePragInfo IdInfo fn_info fn_inline_spec :: InlineSpec fn_inline_spec = InlinePragma -> InlineSpec inl_inline InlinePragma fn_inl_prag fn_unfolding :: Unfolding fn_unfolding = IdInfo -> Unfolding unfoldingInfo IdInfo fn_info -- Even if we don't w/w join points for CPR, we might still do so for -- strictness. In which case a join point worker keeps its original CPR -- property; see Note [Don't w/w join points for CPR]. Otherwise, the worker -- doesn't have the CPR property anymore. work_cpr_info :: Cpr work_cpr_info | Var -> Bool isJoinId Var fn_id = Cpr cpr | Bool otherwise = Cpr topCpr mkStrWrapperInlinePrag :: InlinePragma -> InlinePragma mkStrWrapperInlinePrag :: InlinePragma -> InlinePragma mkStrWrapperInlinePrag (InlinePragma { inl_act :: InlinePragma -> Activation inl_act = Activation act, inl_rule :: InlinePragma -> RuleMatchInfo inl_rule = RuleMatchInfo rule_info }) = InlinePragma { inl_src :: SourceText inl_src = String -> SourceText SourceText String "{-# INLINE" , inl_inline :: InlineSpec inl_inline = InlineSpec NoUserInlinePrag -- See Note [Wrapper NoUserInline] , inl_sat :: Maybe ArityInfo inl_sat = Maybe ArityInfo forall a. Maybe a Nothing , inl_act :: Activation inl_act = Activation wrap_act , inl_rule :: RuleMatchInfo inl_rule = RuleMatchInfo rule_info } -- RuleMatchInfo is (and must be) unaffected where wrap_act :: Activation wrap_act = case Activation act of -- See Note [Wrapper activation] Activation NeverActive -> Activation activateDuringFinal Activation FinalActive -> Activation act ActiveAfter {} -> Activation act ActiveBefore {} -> Activation activateAfterInitial Activation AlwaysActive -> Activation activateAfterInitial -- For the last two cases, see (4) in Note [Wrapper activation] -- NB: the (ActiveBefore n) isn't quite right. We really want -- it to be active *after* Initial but *before* n. We don't have -- a way to say that, alas. {- Note [Demand on the worker] ~~~~~~~~~~~~~~~~~~~~~~~~~~~ If the original function is called once, according to its demand info, then so is the worker. This is important so that the occurrence analyser can attach OneShot annotations to the worker’s lambda binders. Example: -- Original function f [Demand=<L,1*C1(U)>] :: (a,a) -> a f = \p -> ... -- Wrapper f [Demand=<L,1*C1(U)>] :: a -> a -> a f = \p -> case p of (a,b) -> $wf a b -- Worker $wf [Demand=<L,1*C1(C1(U))>] :: Int -> Int $wf = \a b -> ... We need to check whether the original function is called once, with sufficiently many arguments. This is done using saturatedByOneShots, which takes the arity of the original function (resp. the wrapper) and the demand on the original function. The demand on the worker is then calculated using mkWorkerDemand, and always of the form [Demand=<L,1*(C1(...(C1(U))))>] Note [Do not split void functions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Consider this rather common form of binding: $j = \x:Void# -> ...no use of x... Since x is not used it'll be marked as absent. But there is no point in w/w-ing because we'll simply add (\y:Void#), see GHC.Core.Opt.WorkWrap.Utils.mkWorerArgs. If x has a more interesting type (eg Int, or Int#), there *is* a point in w/w so that we don't pass the argument at all. Note [Thunk splitting] ~~~~~~~~~~~~~~~~~~~~~~ Suppose x is used strictly (never mind whether it has the CPR property). let x* = x-rhs in body splitThunk transforms like this: let x* = case x-rhs of { I# a -> I# a } in body Now simplifier will transform to case x-rhs of I# a -> let x* = I# a in body which is what we want. Now suppose x-rhs is itself a case: x-rhs = case e of { T -> I# a; F -> I# b } The join point will abstract over a, rather than over (which is what would have happened before) which is fine. Notice that x certainly has the CPR property now! In fact, splitThunk uses the function argument w/w splitting function, so that if x's demand is deeper (say U(U(L,L),L)) then the splitting will go deeper too. NB: For recursive thunks, the Simplifier is unable to float `x-rhs` out of `x*`'s RHS, because `x*` occurs freely in `x-rhs`, and will just change it back to the original definition, so we just split non-recursive thunks. Note [Thunk splitting for top-level binders] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Top-level bindings are never strict. Yet they can be absent, as T14270 shows: module T14270 (mkTrApp) where mkTrApp x y | Just ... <- ... typeRepKind x ... = undefined | otherwise = undefined typeRepKind = Tick scc undefined (T19180 is a profiling-free test case for this) Note that `typeRepKind` is not exported and its only use site in `mkTrApp` guards a bottoming expression. Thus, demand analysis figures out that `typeRepKind` is absent and splits the thunk to typeRepKind = let typeRepKind = Tick scc undefined in let typeRepKind = absentError in typeRepKind But now we have a local binding with an External Name (See Note [About the NameSorts]). That will trigger a CoreLint error, which we get around by localising the Id for the auxiliary bindings in 'splitThunk'. -} -- | See Note [Thunk splitting]. -- -- splitThunk converts the *non-recursive* binding -- x = e -- into -- x = let x' = e in -- case x' of I# y -> let x' = I# y in x' -- See comments above. Is it not beautifully short? -- Moreover, it works just as well when there are -- several binders, and if the binders are lifted -- E.g. x = e -- --> x = let x' = e in -- case x' of (a,b) -> let x' = (a,b) in x' -- Here, x' is a localised version of x, in case x is a -- top-level Id with an External Name, because Lint rejects local binders with -- External Names; see Note [About the NameSorts] in GHC.Types.Name. -- -- How can we do thunk-splitting on a top-level binder? See -- Note [Thunk splitting for top-level binders]. splitThunk :: DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] splitThunk :: DynFlags -> FamInstEnvs -> RecFlag -> Var -> Expr Var -> UniqSM [(Var, Expr Var)] splitThunk DynFlags dflags FamInstEnvs fam_envs RecFlag is_rec Var x Expr Var rhs = ASSERT(not (isJoinId x)) do { let x' :: Var x' = Var -> Var localiseId Var x -- See comment above ; (Bool useful,[Var] _, Expr Var -> Expr Var wrap_fn, Expr Var -> Expr Var work_fn) <- DynFlags -> FamInstEnvs -> Bool -> [Var] -> UniqSM (Bool, [Var], Expr Var -> Expr Var, Expr Var -> Expr Var) mkWWstr DynFlags dflags FamInstEnvs fam_envs Bool False [Var x'] ; let res :: [(Var, Expr Var)] res = [ (Var x, CoreBind -> Expr Var -> Expr Var forall b. Bind b -> Expr b -> Expr b Let (Var -> Expr Var -> CoreBind forall b. b -> Expr b -> Bind b NonRec Var x' Expr Var rhs) (Expr Var -> Expr Var wrap_fn (Expr Var -> Expr Var work_fn (Var -> Expr Var forall b. Var -> Expr b Var Var x')))) ] ; if Bool useful then ASSERT2( isNonRec is_rec, ppr x ) -- The thunk must be non-recursive [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [(Var, Expr Var)] res else [(Var, Expr Var)] -> UniqSM [(Var, Expr Var)] forall (m :: * -> *) a. Monad m => a -> m a return [(Var x, Expr Var rhs)] }