import CoreSyn
import CoreLint ( showPass, endPass )
-import CoreUtils ( exprType, tcEqExpr, mkPiTypes )
+import CoreUtils ( exprType, mkPiTypes )
import CoreFVs ( exprsFreeVars )
-import CoreSubst ( Subst, mkSubst, substExpr )
import CoreTidy ( tidyRules )
import PprCore ( pprRules )
import WwLib ( mkWorkerArgs )
-import DataCon ( dataConRepArity, isVanillaDataCon )
-import Type ( tyConAppArgs, tyVarsOfTypes )
-import Unify ( coreRefineTys )
+import DataCon ( dataConRepArity, dataConUnivTyVars )
+import Type ( Type, tyConAppArgs )
+import Coercion ( coercionKind )
+import Rules ( matchN )
import Id ( Id, idName, idType, isDataConWorkId_maybe,
- mkUserLocal, mkSysLocal )
+ mkUserLocal, mkSysLocal, idUnfolding, isLocalId )
import Var ( Var )
import VarEnv
import VarSet
import ErrUtils ( dumpIfSet_dyn )
import DynFlags ( DynFlags, DynFlag(..) )
import BasicTypes ( Activation(..) )
-import Maybes ( orElse )
-import Util ( mapAccumL, lengthAtLeast, notNull )
+import Maybes ( orElse, catMaybes, isJust )
+import Util ( zipWithEqual, lengthAtLeast, notNull )
import List ( nubBy, partition )
import UniqSupply
import Outputable
import FastString
+import UniqFM
\end{code}
-----------------------------------------------------
At the call to f, we see that the argument, n is know to be (I# n#),
and n is evaluated elsewhere in the body of f, so we can play the same
-trick as above. However we don't want to do that if the boxed version
-of n is needed (else we'd avoid the eval but pay more for re-boxing n).
-So in this case we want that the *only* uses of n are in case statements.
+trick as above.
+Note [Reboxing]
+~~~~~~~~~~~~~~~
+We must be careful not to allocate the same constructor twice. Consider
+ f p = (...(case p of (a,b) -> e)...p...,
+ ...let t = (r,s) in ...t...(f t)...)
+At the recursive call to f, we can see that t is a pair. But we do NOT want
+to make a specialised copy:
+ f' a b = let p = (a,b) in (..., ...)
+because now t is allocated by the caller, then r and s are passed to the
+recursive call, which allocates the (r,s) pair again.
+
+This happens if
+ (a) the argument p is used in other than a case-scrutinsation way.
+ (b) the argument to the call is not a 'fresh' tuple; you have to
+ look into its unfolding to see that it's a tuple
+
+Hence the "OR" part of Note [Good arguments] below.
+
+ALTERNATIVE: pass both boxed and unboxed versions. This no longer saves
+allocation, but does perhaps save evals. In the RULE we'd have
+something like
+
+ f (I# x#) = f' (I# x#) x#
+
+If at the call site the (I# x) was an unfolding, then we'd have to
+rely on CSE to eliminate the duplicate allocation.... This alternative
+doesn't look attractive enough to pursue.
+
+
+Note [Good arguments]
+~~~~~~~~~~~~~~~~~~~~~
So we look for
* A self-recursive function. Ignore mutual recursion for now,
That same parameter is scrutinised by a case somewhere in
the RHS of the function
AND
- Those are the only uses of the parameter
+ Those are the only uses of the parameter (see Note [Reboxing])
+What to abstract over
+~~~~~~~~~~~~~~~~~~~~~
There's a bit of a complication with type arguments. If the call
site looks like
* Find the free variables of the abstracted pattern
* Pass these variables, less any that are in scope at
- the fn defn.
+ the fn defn. But see Note [Shadowing] below.
NOTICE that we only abstract over variables that are not in scope,
in f_spec's RHS.
+Note [Shadowing]
+~~~~~~~~~~~~~~~~
+In this pass we gather up usage information that may mention variables
+that are bound between the usage site and the definition site; or (more
+seriously) may be bound to something different at the definition site.
+For example:
+
+ f x = letrec g y v = let x = ...
+ in ...(g (a,b) x)...
+
+Since 'x' is in scope at the call site, we may make a rewrite rule that
+looks like
+ RULE forall a,b. g (a,b) x = ...
+But this rule will never match, because it's really a different 'x' at
+the call site -- and that difference will be manifest by the time the
+simplifier gets to it. [A worry: the simplifier doesn't *guarantee*
+no-shadowing, so perhaps it may not be distinct?]
+
+Anyway, the rule isn't actually wrong, it's just not useful. One possibility
+is to run deShadowBinds before running SpecConstr, but instead we run the
+simplifier. That gives the simplest possible program for SpecConstr to
+chew on; and it virtually guarantees no shadowing.
+
+Note [Specialising for constant parameters]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+This one is about specialising on a *constant* (but not necessarily
+constructor) argument
+
+ foo :: Int -> (Int -> Int) -> Int
+ foo 0 f = 0
+ foo m f = foo (f m) (+1)
+
+It produces
+
+ lvl_rmV :: GHC.Base.Int -> GHC.Base.Int
+ lvl_rmV =
+ \ (ds_dlk :: GHC.Base.Int) ->
+ case ds_dlk of wild_alH { GHC.Base.I# x_alG ->
+ GHC.Base.I# (GHC.Prim.+# x_alG 1)
+
+ T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
+ GHC.Prim.Int#
+ T.$wfoo =
+ \ (ww_sme :: GHC.Prim.Int#) (w_smg :: GHC.Base.Int -> GHC.Base.Int) ->
+ case ww_sme of ds_Xlw {
+ __DEFAULT ->
+ case w_smg (GHC.Base.I# ds_Xlw) of w1_Xmo { GHC.Base.I# ww1_Xmz ->
+ T.$wfoo ww1_Xmz lvl_rmV
+ };
+ 0 -> 0
+ }
+
+The recursive call has lvl_rmV as its argument, so we could create a specialised copy
+with that argument baked in; that is, not passed at all. Now it can perhaps be inlined.
+
+When is this worth it? Call the constant 'lvl'
+- If 'lvl' has an unfolding that is a constructor, see if the corresponding
+ parameter is scrutinised anywhere in the body.
+
+- If 'lvl' has an unfolding that is a inlinable function, see if the corresponding
+ parameter is applied (...to enough arguments...?)
+
+ Also do this is if the function has RULES?
+
+Also
+
+Note [Specialising for lambda parameters]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ foo :: Int -> (Int -> Int) -> Int
+ foo 0 f = 0
+ foo m f = foo (f m) (\n -> n-m)
+
+This is subtly different from the previous one in that we get an
+explicit lambda as the argument:
+
+ T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
+ GHC.Prim.Int#
+ T.$wfoo =
+ \ (ww_sm8 :: GHC.Prim.Int#) (w_sma :: GHC.Base.Int -> GHC.Base.Int) ->
+ case ww_sm8 of ds_Xlr {
+ __DEFAULT ->
+ case w_sma (GHC.Base.I# ds_Xlr) of w1_Xmf { GHC.Base.I# ww1_Xmq ->
+ T.$wfoo
+ ww1_Xmq
+ (\ (n_ad3 :: GHC.Base.Int) ->
+ case n_ad3 of wild_alB { GHC.Base.I# x_alA ->
+ GHC.Base.I# (GHC.Prim.-# x_alA ds_Xlr)
+ })
+ };
+ 0 -> 0
+ }
+
+I wonder if SpecConstr couldn't be extended to handle this? After all,
+lambda is a sort of constructor for functions and perhaps it already
+has most of the necessary machinery?
+
+Furthermore, there's an immediate win, because you don't need to allocate the lamda
+at the call site; and if perchance it's called in the recursive call, then you
+may avoid allocating it altogether. Just like for constructors.
+
+Looks cool, but probably rare...but it might be easy to implement.
+
+
+Note [SpecConstr for casts]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Consider
+ data family T a :: *
+ data instance T Int = T Int
+
+ foo n = ...
+ where
+ go (T 0) = 0
+ go (T n) = go (T (n-1))
+
+The recursive call ends up looking like
+ go (T (I# ...) `cast` g)
+So we want to spot the construtor application inside the cast.
+That's why we have the Cast case in argToPat
+
+
+-----------------------------------------------------
+ Stuff not yet handled
+-----------------------------------------------------
+
+Here are notes arising from Roman's work that I don't want to lose.
+
+Example 1
+~~~~~~~~~
+ data T a = T !a
+
+ foo :: Int -> T Int -> Int
+ foo 0 t = 0
+ foo x t | even x = case t of { T n -> foo (x-n) t }
+ | otherwise = foo (x-1) t
+
+SpecConstr does no specialisation, because the second recursive call
+looks like a boxed use of the argument. A pity.
+
+ $wfoo_sFw :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
+ $wfoo_sFw =
+ \ (ww_sFo [Just L] :: GHC.Prim.Int#) (w_sFq [Just L] :: T.T GHC.Base.Int) ->
+ case ww_sFo of ds_Xw6 [Just L] {
+ __DEFAULT ->
+ case GHC.Prim.remInt# ds_Xw6 2 of wild1_aEF [Dead Just A] {
+ __DEFAULT -> $wfoo_sFw (GHC.Prim.-# ds_Xw6 1) w_sFq;
+ 0 ->
+ case w_sFq of wild_Xy [Just L] { T.T n_ad5 [Just U(L)] ->
+ case n_ad5 of wild1_aET [Just A] { GHC.Base.I# y_aES [Just L] ->
+ $wfoo_sFw (GHC.Prim.-# ds_Xw6 y_aES) wild_Xy
+ } } };
+ 0 -> 0
+
+Example 2
+~~~~~~~~~
+ data a :*: b = !a :*: !b
+ data T a = T !a
+
+ foo :: (Int :*: T Int) -> Int
+ foo (0 :*: t) = 0
+ foo (x :*: t) | even x = case t of { T n -> foo ((x-n) :*: t) }
+ | otherwise = foo ((x-1) :*: t)
+
+Very similar to the previous one, except that the parameters are now in
+a strict tuple. Before SpecConstr, we have
+
+ $wfoo_sG3 :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
+ $wfoo_sG3 =
+ \ (ww_sFU [Just L] :: GHC.Prim.Int#) (ww_sFW [Just L] :: T.T
+ GHC.Base.Int) ->
+ case ww_sFU of ds_Xws [Just L] {
+ __DEFAULT ->
+ case GHC.Prim.remInt# ds_Xws 2 of wild1_aEZ [Dead Just A] {
+ __DEFAULT ->
+ case ww_sFW of tpl_B2 [Just L] { T.T a_sFo [Just A] ->
+ $wfoo_sG3 (GHC.Prim.-# ds_Xws 1) tpl_B2 -- $wfoo1
+ };
+ 0 ->
+ case ww_sFW of wild_XB [Just A] { T.T n_ad7 [Just S(L)] ->
+ case n_ad7 of wild1_aFd [Just L] { GHC.Base.I# y_aFc [Just L] ->
+ $wfoo_sG3 (GHC.Prim.-# ds_Xws y_aFc) wild_XB -- $wfoo2
+ } } };
+ 0 -> 0 }
+
+We get two specialisations:
+"SC:$wfoo1" [0] __forall {a_sFB :: GHC.Base.Int sc_sGC :: GHC.Prim.Int#}
+ Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int a_sFB)
+ = Foo.$s$wfoo1 a_sFB sc_sGC ;
+"SC:$wfoo2" [0] __forall {y_aFp :: GHC.Prim.Int# sc_sGC :: GHC.Prim.Int#}
+ Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int (GHC.Base.I# y_aFp))
+ = Foo.$s$wfoo y_aFp sc_sGC ;
+
+But perhaps the first one isn't good. After all, we know that tpl_B2 is
+a T (I# x) really, because T is strict and Int has one constructor. (We can't
+unbox the strict fields, becuase T is polymorphic!)
+
+
+
%************************************************************************
%* *
\subsection{Top level wrapper stuff}
%************************************************************************
\begin{code}
-data ScEnv = SCE { scope :: VarEnv HowBound,
+data ScEnv = SCE { scope :: InScopeEnv,
-- Binds all non-top-level variables in scope
cons :: ConstrEnv
}
+type InScopeEnv = VarEnv HowBound
+
type ConstrEnv = IdEnv ConValue
data ConValue = CV AltCon [CoreArg]
-- Variables known to be bound to a constructor
instance Outputable ConValue where
ppr (CV con args) = ppr con <+> interpp'SP args
-refineConstrEnv :: Subst -> ConstrEnv -> ConstrEnv
--- The substitution is a type substitution only
-refineConstrEnv subst env = mapVarEnv refine_con_value env
- where
- refine_con_value (CV con args) = CV con (map (substExpr subst) args)
-
emptyScEnv = SCE { scope = emptyVarEnv, cons = emptyVarEnv }
-data HowBound = RecFun -- These are the recursive functions for which
- -- we seek interesting call patterns
+data HowBound = RecFun -- These are the recursive functions for which
+ -- we seek interesting call patterns
- | RecArg -- These are those functions' arguments; we are
- -- interested to see if those arguments are scrutinised
+ | RecArg -- These are those functions' arguments, or their sub-components;
+ -- we gather occurrence information for these
- | Other -- We track all others so we know what's in scope
- -- This is used in spec_one to check what needs to be
- -- passed as a parameter and what is in scope at the
- -- function definition site
+ | Other -- We track all others so we know what's in scope
+ -- This is used in spec_one to check what needs to be
+ -- passed as a parameter and what is in scope at the
+ -- function definition site
instance Outputable HowBound where
ppr RecFun = text "RecFun"
-- C x y -> ...
-- we want to bind b, and perhaps scrut too, to (C x y)
extendCaseBndrs :: ScEnv -> Id -> CoreExpr -> AltCon -> [Var] -> ScEnv
-extendCaseBndrs env case_bndr scrut DEFAULT alt_bndrs
- = extendBndrs env (case_bndr : alt_bndrs)
-
-extendCaseBndrs env case_bndr scrut con@(LitAlt lit) alt_bndrs
- = ASSERT( null alt_bndrs ) extendAlt env case_bndr scrut (CV con []) []
-
-extendCaseBndrs env case_bndr scrut con@(DataAlt data_con) alt_bndrs
- | isVanillaDataCon data_con
- = extendAlt env case_bndr scrut (CV con vanilla_args) alt_bndrs
-
- | otherwise -- GADT
- = extendAlt env1 case_bndr scrut (CV con gadt_args) alt_bndrs
+extendCaseBndrs env case_bndr scrut con alt_bndrs
+ = case con of
+ DEFAULT -> env1
+ LitAlt lit -> extendCons env1 scrut case_bndr (CV con [])
+ DataAlt dc -> extend_data_con dc
where
- vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
- map varToCoreExpr alt_bndrs
-
- gadt_args = map (substExpr subst . varToCoreExpr) alt_bndrs
- -- This call generates some bogus warnings from substExpr,
- -- because it's inconvenient to put all the Ids in scope
- -- Will be fixed when we move to FC
-
- (alt_tvs, _) = span isTyVar alt_bndrs
- Just (tv_subst, is_local) = coreRefineTys data_con alt_tvs (idType case_bndr)
- subst = mkSubst in_scope tv_subst emptyVarEnv -- No Id substitition
- in_scope = mkInScopeSet (tyVarsOfTypes (varEnvElts tv_subst))
-
- env1 | is_local = env
- | otherwise = env { cons = refineConstrEnv subst (cons env) }
-
-
-
-extendAlt :: ScEnv -> Id -> CoreExpr -> ConValue -> [Var] -> ScEnv
-extendAlt env case_bndr scrut val alt_bndrs
- = let
- env1 = SCE { scope = extendVarEnvList (scope env) [(b,Other) | b <- case_bndr : alt_bndrs],
- cons = extendVarEnv (cons env) case_bndr val }
- in
- case scrut of
- Var v -> -- Bind the scrutinee in the ConstrEnv if it's a variable
- -- Also forget if the scrutinee is a RecArg, because we're
- -- now in the branch of a case, and we don't want to
- -- record a non-scrutinee use of v if we have
- -- case v of { (a,b) -> ...(f v)... }
- SCE { scope = extendVarEnv (scope env1) v Other,
- cons = extendVarEnv (cons env1) v val }
- other -> env1
+ cur_scope = scope env
+ env1 = env { scope = extendVarEnvList cur_scope
+ [(b,how_bound) | b <- case_bndr:alt_bndrs] }
+
+ -- Record RecArg for the components iff the scrutinee is RecArg
+ -- I think the only reason for this is to keep the usage envt small
+ -- so is it worth it at all?
+ -- [This comment looks plain wrong to me, so I'm ignoring it
+ -- "Also forget if the scrutinee is a RecArg, because we're
+ -- now in the branch of a case, and we don't want to
+ -- record a non-scrutinee use of v if we have
+ -- case v of { (a,b) -> ...(f v)... }" ]
+ how_bound = get_how scrut
+ where
+ get_how (Var v) = lookupVarEnv cur_scope v `orElse` Other
+ get_how (Cast e _) = get_how e
+ get_how (Note _ e) = get_how e
+ get_how other = Other
+
+ extend_data_con data_con =
+ extendCons env1 scrut case_bndr (CV con vanilla_args)
+ where
+ vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
+ varsToCoreExprs alt_bndrs
+
+extendCons :: ScEnv -> CoreExpr -> Id -> ConValue -> ScEnv
+extendCons env scrut case_bndr val
+ = case scrut of
+ Var v -> env { cons = extendVarEnv cons1 v val }
+ other -> env { cons = cons1 }
+ where
+ cons1 = extendVarEnv (cons env) case_bndr val
-- When we encounter a recursive function binding
-- f = \x y -> ...
combineUsages [] = nullUsage
combineUsages us = foldr1 combineUsage us
-data ArgOcc = CaseScrut
- | OtherOcc
- | Both
+lookupOcc :: ScUsage -> Var -> (ScUsage, ArgOcc)
+lookupOcc (SCU { calls = sc_calls, occs = sc_occs }) bndr
+ = (SCU {calls = sc_calls, occs = delVarEnv sc_occs bndr},
+ lookupVarEnv sc_occs bndr `orElse` NoOcc)
+
+lookupOccs :: ScUsage -> [Var] -> (ScUsage, [ArgOcc])
+lookupOccs (SCU { calls = sc_calls, occs = sc_occs }) bndrs
+ = (SCU {calls = sc_calls, occs = delVarEnvList sc_occs bndrs},
+ [lookupVarEnv sc_occs b `orElse` NoOcc | b <- bndrs])
+
+data ArgOcc = NoOcc -- Doesn't occur at all; or a type argument
+ | UnkOcc -- Used in some unknown way
+
+ | ScrutOcc (UniqFM [ArgOcc]) -- See Note [ScrutOcc]
+
+ | BothOcc -- Definitely taken apart, *and* perhaps used in some other way
+
+{- Note [ScrutOcc]
+
+An occurrence of ScrutOcc indicates that the thing, or a `cast` version of the thing,
+is *only* taken apart or applied.
+
+ Functions, literal: ScrutOcc emptyUFM
+ Data constructors: ScrutOcc subs,
+
+where (subs :: UniqFM [ArgOcc]) gives usage of the *pattern-bound* components,
+The domain of the UniqFM is the Unique of the data constructor
+
+The [ArgOcc] is the occurrences of the *pattern-bound* components
+of the data structure. E.g.
+ data T a = forall b. MkT a b (b->a)
+A pattern binds b, x::a, y::b, z::b->a, but not 'a'!
+
+-}
instance Outputable ArgOcc where
- ppr CaseScrut = ptext SLIT("case-scrut")
- ppr OtherOcc = ptext SLIT("other-occ")
- ppr Both = ptext SLIT("case-scrut and other")
+ ppr (ScrutOcc xs) = ptext SLIT("scrut-occ") <> ppr xs
+ ppr UnkOcc = ptext SLIT("unk-occ")
+ ppr BothOcc = ptext SLIT("both-occ")
+ ppr NoOcc = ptext SLIT("no-occ")
+
+combineOcc NoOcc occ = occ
+combineOcc occ NoOcc = occ
+combineOcc (ScrutOcc xs) (ScrutOcc ys) = ScrutOcc (plusUFM_C combineOccs xs ys)
+combineOcc UnkOcc UnkOcc = UnkOcc
+combineOcc _ _ = BothOcc
+
+combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc]
+combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys
+
+conArgOccs :: ArgOcc -> AltCon -> [ArgOcc]
+-- Find usage of components of data con; returns [UnkOcc...] if unknown
+-- See Note [ScrutOcc] for the extra UnkOccs in the vanilla datacon case
-combineOcc CaseScrut CaseScrut = CaseScrut
-combineOcc OtherOcc OtherOcc = OtherOcc
-combineOcc _ _ = Both
+conArgOccs (ScrutOcc fm) (DataAlt dc)
+ | Just pat_arg_occs <- lookupUFM fm dc
+ = [UnkOcc | tv <- dataConUnivTyVars dc] ++ pat_arg_occs
+
+conArgOccs other con = repeat UnkOcc
\end{code}
scExpr env e@(Type t) = returnUs (nullUsage, e)
scExpr env e@(Lit l) = returnUs (nullUsage, e)
-scExpr env e@(Var v) = returnUs (varUsage env v OtherOcc, e)
+scExpr env e@(Var v) = returnUs (varUsage env v UnkOcc, e)
scExpr env (Note n e) = scExpr env e `thenUs` \ (usg,e') ->
returnUs (usg, Note n e')
+scExpr env (Cast e co)= scExpr env e `thenUs` \ (usg,e') ->
+ returnUs (usg, Cast e' co)
scExpr env (Lam b e) = scExpr (extendBndr env b) e `thenUs` \ (usg,e') ->
returnUs (usg, Lam b e')
scExpr env (Case scrut b ty alts)
- = sc_scrut scrut `thenUs` \ (scrut_usg, scrut') ->
- mapAndUnzipUs sc_alt alts `thenUs` \ (alts_usgs, alts') ->
- returnUs (combineUsages alts_usgs `combineUsage` scrut_usg,
- Case scrut' b ty alts')
+ = do { (alt_usgs, alt_occs, alts') <- mapAndUnzip3Us sc_alt alts
+ ; let (alt_usg, b_occ) = lookupOcc (combineUsages alt_usgs) b
+ scrut_occ = foldr combineOcc b_occ alt_occs
+ -- The combined usage of the scrutinee is given
+ -- by scrut_occ, which is passed to scScrut, which
+ -- in turn treats a bare-variable scrutinee specially
+ ; (scrut_usg, scrut') <- scScrut env scrut scrut_occ
+ ; return (alt_usg `combineUsage` scrut_usg,
+ Case scrut' b ty alts') }
where
- sc_scrut e@(Var v) = returnUs (varUsage env v CaseScrut, e)
- sc_scrut e = scExpr env e
-
- sc_alt (con,bs,rhs) = scExpr env1 rhs `thenUs` \ (usg,rhs') ->
- returnUs (usg, (con,bs,rhs'))
- where
- env1 = extendCaseBndrs env b scrut con bs
+ sc_alt (con,bs,rhs)
+ = do { let env1 = extendCaseBndrs env b scrut con bs
+ ; (usg,rhs') <- scExpr env1 rhs
+ ; let (usg', arg_occs) = lookupOccs usg bs
+ scrut_occ = case con of
+ DataAlt dc -> ScrutOcc (unitUFM dc arg_occs)
+ other -> ScrutOcc emptyUFM
+ ; return (usg', scrut_occ, (con,bs,rhs')) }
scExpr env (Let bind body)
= scBind env bind `thenUs` \ (env', bind_usg, bind') ->
returnUs (bind_usg `combineUsage` body_usg, Let bind' body')
scExpr env e@(App _ _)
- = let
- (fn, args) = collectArgs e
- in
- mapAndUnzipUs (scExpr env) (fn:args) `thenUs` \ (usgs, (fn':args')) ->
+ = do { let (fn, args) = collectArgs e
+ ; (fn_usg, fn') <- scScrut env fn (ScrutOcc emptyUFM)
-- Process the function too. It's almost always a variable,
-- but not always. In particular, if this pass follows float-in,
-- which it may, we can get
-- (let f = ...f... in f) arg1 arg2
- let
- call_usg = case fn of
- Var f | Just RecFun <- lookupScopeEnv env f
- -> SCU { calls = unitVarEnv f [(cons env, args)],
- occs = emptyVarEnv }
- other -> nullUsage
- in
- returnUs (combineUsages usgs `combineUsage` call_usg, mkApps fn' args')
+ -- We use scScrut to record the fact that the function is called
+ -- Perhpas we should check that it has at least one value arg,
+ -- but currently we don't bother
+
+ ; (arg_usgs, args') <- mapAndUnzipUs (scExpr env) args
+ ; let call_usg = case fn of
+ Var f | Just RecFun <- lookupScopeEnv env f
+ -> SCU { calls = unitVarEnv f [(cons env, args)],
+ occs = emptyVarEnv }
+ other -> nullUsage
+ ; return (combineUsages arg_usgs `combineUsage` fn_usg
+ `combineUsage` call_usg,
+ mkApps fn' args') }
+
+
+----------------------
+scScrut :: ScEnv -> CoreExpr -> ArgOcc -> UniqSM (ScUsage, CoreExpr)
+-- Used for the scrutinee of a case,
+-- or the function of an application.
+-- Remember to look through casts
+scScrut env e@(Var v) occ = returnUs (varUsage env v occ, e)
+scScrut env (Cast e co) occ = do { (usg, e') <- scScrut env e occ
+ ; returnUs (usg, Cast e' co) }
+scScrut env e occ = scExpr env e
----------------------
scBind env (Rec [(fn,rhs)])
| notNull val_bndrs
= scExpr env_fn_body body `thenUs` \ (usg, body') ->
- specialise env fn bndrs body usg `thenUs` \ (rules, spec_prs) ->
+ specialise env fn bndrs body' usg `thenUs` \ (rules, spec_prs) ->
+ -- Note body': the specialised copies should be based on the
+ -- optimised version of the body, in case there were
+ -- nested functions inside.
let
SCU { calls = calls, occs = occs } = usg
in
-> UniqSM ([CoreRule], -- Rules
[(Id,CoreExpr)]) -- Bindings
-specialise env fn bndrs body (SCU {calls=calls, occs=occs})
- = getUs `thenUs` \ us ->
- let
- all_calls = lookupVarEnv calls fn `orElse` []
-
- good_calls :: [[CoreArg]]
- good_calls = [ pats
- | (con_env, call_args) <- all_calls,
- call_args `lengthAtLeast` n_bndrs, -- App is saturated
- let call = bndrs `zip` call_args,
- any (good_arg con_env occs) call, -- At least one arg is a constr app
- let (_, pats) = argsToPats con_env us call_args
- ]
- in
- mapAndUnzipUs (spec_one env fn (mkLams bndrs body))
- (nubBy same_call good_calls `zip` [1..])
+specialise env fn bndrs body body_usg
+ = do { let (_, bndr_occs) = lookupOccs body_usg bndrs
+
+ ; mb_calls <- -- pprTrace "specialise" (ppr fn <+> ppr bndrs <+> ppr bndr_occs) $
+ mapM (callToPats (scope env) bndr_occs)
+ (lookupVarEnv (calls body_usg) fn `orElse` [])
+
+ ; let good_calls :: [([Var], [CoreArg])]
+ good_calls = catMaybes mb_calls
+ in_scope = mkInScopeSet $ unionVarSets $
+ [ exprsFreeVars pats `delVarSetList` vs
+ | (vs,pats) <- good_calls ]
+ uniq_calls = nubBy (same_call in_scope) good_calls
+ ; mapAndUnzipUs (spec_one env fn (mkLams bndrs body))
+ (uniq_calls `zip` [1..]) }
where
- n_bndrs = length bndrs
- same_call as1 as2 = and (zipWith tcEqExpr as1 as2)
-
----------------------
-good_arg :: ConstrEnv -> IdEnv ArgOcc -> (CoreBndr, CoreArg) -> Bool
-good_arg con_env arg_occs (bndr, arg)
- = case is_con_app_maybe con_env arg of
- Just _ -> bndr_usg_ok arg_occs bndr arg
- other -> False
-
-bndr_usg_ok :: IdEnv ArgOcc -> Var -> CoreArg -> Bool
-bndr_usg_ok arg_occs bndr arg
- = case lookupVarEnv arg_occs bndr of
- Just CaseScrut -> True -- Used only by case scrutiny
- Just Both -> case arg of -- Used by case and elsewhere
- App _ _ -> True -- so the arg should be an explicit con app
- other -> False
- other -> False -- Not used, or used wonkily
-
+ -- Two calls are the same if they match both ways
+ same_call in_scope (vs1,as1)(vs2,as2)
+ = isJust (matchN in_scope vs1 as1 as2)
+ && isJust (matchN in_scope vs2 as2 as1)
+
+callToPats :: InScopeEnv -> [ArgOcc] -> Call
+ -> UniqSM (Maybe ([Var], [CoreExpr]))
+ -- The VarSet is the variables to quantify over in the rule
+ -- The [CoreExpr] are the argument patterns for the rule
+callToPats in_scope bndr_occs (con_env, args)
+ | length args < length bndr_occs -- Check saturated
+ = return Nothing
+ | otherwise
+ = do { prs <- argsToPats in_scope con_env (args `zip` bndr_occs)
+ ; let (good_pats, pats) = unzip prs
+ pat_fvs = varSetElems (exprsFreeVars pats)
+ qvars = filter (not . (`elemVarEnv` in_scope)) pat_fvs
+ -- Quantify over variables that are not in sccpe
+ -- See Note [Shadowing] at the top
+
+ ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $
+ if or good_pats
+ then return (Just (qvars, pats))
+ else return Nothing }
---------------------
spec_one :: ScEnv
-> Id -- Function
-> CoreExpr -- Rhs of the original function
- -> ([CoreArg], Int)
+ -> (([Var], [CoreArg]), Int)
-> UniqSM (CoreRule, (Id,CoreExpr)) -- Rule and binding
-- spec_one creates a specialised copy of the function, together
f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
-}
-spec_one env fn rhs (pats, rule_number)
+spec_one env fn rhs ((vars_to_bind, pats), rule_number)
= getUniqueUs `thenUs` \ spec_uniq ->
let
fn_name = idName fn
fn_loc = nameSrcLoc fn_name
spec_occ = mkSpecOcc (nameOccName fn_name)
- pat_fvs = varSetElems (exprsFreeVars pats)
- vars_to_bind = filter not_avail pat_fvs
- not_avail v = not (v `elemVarEnv` scope env)
+
-- Put the type variables first; the type of a term
-- variable may mention a type variable
(tvs, ids) = partition isTyVar vars_to_bind
This code deals with analysing call-site arguments to see whether
they are constructor applications.
+
\begin{code}
-- argToPat takes an actual argument, and returns an abstracted
-- version, consisting of just the "constructor skeleton" of the
-- placeholder variables. For example:
-- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
-argToPat :: ConstrEnv -> UniqSupply -> CoreArg -> (UniqSupply, CoreExpr)
-argToPat env us (Type ty)
- = (us, Type ty)
-
-argToPat env us arg
- | Just (CV dc args) <- is_con_app_maybe env arg
- = let
- (us',args') = argsToPats env us args
- in
- (us', mk_con_app dc args')
-
-argToPat env us (Var v) -- Don't uniqify existing vars,
- = (us, Var v) -- so that we can spot when we pass them twice
-
-argToPat env us arg
- = (us1, Var (mkSysLocal FSLIT("sc") (uniqFromSupply us2) (exprType arg)))
+argToPat :: InScopeEnv -- What's in scope at the fn defn site
+ -> ConstrEnv -- ConstrEnv at the call site
+ -> CoreArg -- A call arg (or component thereof)
+ -> ArgOcc
+ -> UniqSM (Bool, CoreArg)
+-- Returns (interesting, pat),
+-- where pat is the pattern derived from the argument
+-- intersting=True if the pattern is non-trivial (not a variable or type)
+-- E.g. x:xs --> (True, x:xs)
+-- f xs --> (False, w) where w is a fresh wildcard
+-- (f xs, 'c') --> (True, (w, 'c')) where w is a fresh wildcard
+-- \x. x+y --> (True, \x. x+y)
+-- lvl7 --> (True, lvl7) if lvl7 is bound
+-- somewhere further out
+
+argToPat in_scope con_env arg@(Type ty) arg_occ
+ = return (False, arg)
+
+argToPat in_scope con_env (Var v) arg_occ
+ | not (isLocalId v) || v `elemVarEnv` in_scope
+ = -- The recursive call passes a variable that
+ -- is in scope at the function definition site
+ -- It's worth specialising on this if
+ -- (a) it's used in an interesting way in the body
+ -- (b) we know what its value is
+ if (case arg_occ of { UnkOcc -> False; other -> True }) -- (a)
+ && isValueUnfolding (idUnfolding v) -- (b)
+ then return (True, Var v)
+ else wildCardPat (idType v)
+
+argToPat in_scope con_env (Let _ arg) arg_occ
+ = argToPat in_scope con_env arg arg_occ
+ -- Look through let expressions
+ -- e.g. f (let v = rhs in \y -> ...v...)
+ -- Here we can specialise for f (\y -> ...)
+ -- because the rule-matcher will look through the let.
+
+argToPat in_scope con_env (Cast arg co) arg_occ
+ = do { (interesting, arg') <- argToPat in_scope con_env arg arg_occ
+ ; if interesting then
+ return (interesting, Cast arg' co)
+ else
+ wildCardPat (snd (coercionKind co)) }
+
+argToPat in_scope con_env arg arg_occ
+ | is_value_lam arg
+ = return (True, arg)
where
- (us1,us2) = splitUniqSupply us
-
-argsToPats :: ConstrEnv -> UniqSupply -> [CoreArg] -> (UniqSupply, [CoreExpr])
-argsToPats env us args = mapAccumL (argToPat env) us args
+ is_value_lam (Lam v e) -- Spot a value lambda, even if
+ | isId v = True -- it is inside a type lambda
+ | otherwise = is_value_lam e
+ is_value_lam other = False
+
+argToPat in_scope con_env arg arg_occ
+ | Just (CV dc args) <- is_con_app_maybe con_env arg
+ , case arg_occ of
+ ScrutOcc _ -> True -- Used only by case scrutinee
+ BothOcc -> case arg of -- Used by case scrut
+ App {} -> True -- ...and elsewhere...
+ other -> False
+ other -> False -- No point; the arg is not decomposed
+ = do { args' <- argsToPats in_scope con_env (args `zip` conArgOccs arg_occ dc)
+ ; return (True, mk_con_app dc (map snd args')) }
+
+argToPat in_scope con_env (Var v) arg_occ
+ = -- A variable bound inside the function.
+ -- Don't make a wild-card, because we may usefully share
+ -- e.g. f a = let x = ... in f (x,x)
+ -- NB: this case follows the lambda and con-app cases!!
+ return (False, Var v)
+
+-- The default case: make a wild-card
+argToPat in_scope con_env arg arg_occ = wildCardPat (exprType arg)
+
+wildCardPat :: Type -> UniqSM (Bool, CoreArg)
+wildCardPat ty = do { uniq <- getUniqueUs
+ ; let id = mkSysLocal FSLIT("sc") uniq ty
+ ; return (False, Var id) }
+
+argsToPats :: InScopeEnv -> ConstrEnv
+ -> [(CoreArg, ArgOcc)]
+ -> UniqSM [(Bool, CoreArg)]
+argsToPats in_scope con_env args
+ = mapUs do_one args
+ where
+ do_one (arg,occ) = argToPat in_scope con_env arg occ
\end{code}
\begin{code}
is_con_app_maybe :: ConstrEnv -> CoreExpr -> Maybe ConValue
is_con_app_maybe env (Var v)
- = lookupVarEnv env v
- -- You might think we could look in the idUnfolding here
- -- but that doesn't take account of which branch of a
- -- case we are in, which is the whole point
+ = case lookupVarEnv env v of
+ Just stuff -> Just stuff
+ -- You might think we could look in the idUnfolding here
+ -- but that doesn't take account of which branch of a
+ -- case we are in, which is the whole point
+
+ Nothing | isCheapUnfolding unf
+ -> is_con_app_maybe env (unfoldingTemplate unf)
+ where
+ unf = idUnfolding v
+ -- However we do want to consult the unfolding
+ -- as well, for let-bound constructors!
+
+ other -> Nothing
is_con_app_maybe env (Lit lit)
= Just (CV (LitAlt lit) [])
mk_con_app :: AltCon -> [CoreArg] -> CoreExpr
mk_con_app (LitAlt lit) [] = Lit lit
mk_con_app (DataAlt con) args = mkConApp con args
+mk_con_app other args = panic "SpecConstr.mk_con_app"
\end{code}