import CoreSyn -- All of it
import OccurAnal ( occurAnalyseExpr )
import CoreFVs ( exprFreeVars, exprsFreeVars, bindFreeVars, rulesFreeVars )
-import CoreUtils ( exprType )
+import CoreUtils ( exprType, eqExprX )
import PprCore ( pprRules )
-import Type ( Type, TvSubstEnv, tcEqTypeX )
+import Type ( Type, TvSubstEnv )
import TcType ( tcSplitTyConApp_maybe )
import CoreTidy ( tidyRules )
import Id
import Outputable
import FastString
import Maybes
-import OrdList
import Bag
import Util
import Data.List
\end{code}
+Note [Overall plumbing for rules]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* After the desugarer:
+ - The ModGuts initially contains mg_rules :: [CoreRule] of
+ locally-declared rules for imported Ids.
+ - Locally-declared rules for locally-declared Ids are attached to
+ the IdInfo for that Id. See Note [Attach rules to local ids] in
+ DsBinds
+
+* TidyPgm strips off all the rules from local Ids and adds them to
+ mg_rules, so that the ModGuts has *all* the locally-declared rules.
+
+* The HomePackageTable contains a ModDetails for each home package
+ module. Each contains md_rules :: [CoreRule] of rules declared in
+ that module. The HomePackageTable grows as ghc --make does its
+ up-sweep. In batch mode (ghc -c), the HPT is empty; all imported modules
+ are treated by the "external" route, discussed next, regardless of
+ which package they come from.
+
+* The ExternalPackageState has a single eps_rule_base :: RuleBase for
+ Ids in other packages. This RuleBase simply grow monotonically, as
+ ghc --make compiles one module after another.
+
+ During simplification, interface files may get demand-loaded,
+ as the simplifier explores the unfoldings for Ids it has in
+ its hand. (Via an unsafePerformIO; the EPS is really a cache.)
+ That in turn may make the EPS rule-base grow. In contrast, the
+ HPT never grows in this way.
+
+* The result of all this is that during Core-to-Core optimisation
+ there are four sources of rules:
+
+ (a) Rules in the IdInfo of the Id they are a rule for. These are
+ easy: fast to look up, and if you apply a substitution then
+ it'll be applied to the IdInfo as a matter of course.
+
+ (b) Rules declared in this module for imported Ids, kept in the
+ ModGuts. If you do a substitution, you'd better apply the
+ substitution to these. There are seldom many of these.
+
+ (c) Rules declared in the HomePackageTable. These never change.
+
+ (d) Rules in the ExternalPackageTable. These can grow in response
+ to lazy demand-loading of interfaces.
+
+* At the moment (c) is carried in a reader-monad way by the CoreMonad.
+ The HomePackageTable doesn't have a single RuleBase because technically
+ we should only be able to "see" rules "below" this module; so we
+ generate a RuleBase for (c) by combing rules from all the modules
+ "below" us. That's whye we can't just select the home-package RuleBase
+ from HscEnv.
+
+ [NB: we are inconsistent here. We should do the same for external
+ pacakges, but we don't. Same for type-class instances.]
+
+* So in the outer simplifier loop, we combine (b-d) into a single
+ RuleBase, reading
+ (b) from the ModGuts,
+ (c) from the CoreMonad, and
+ (d) from its mutable variable
+ [Of coures this means that we won't see new EPS rules that come in
+ during a single simplifier iteration, but that probably does not
+ matter.]
+
+
%************************************************************************
%* *
\subsection[specialisation-IdInfo]{Specialisation info about an @Id@}
%************************************************************************
%* *
-\subsection{Matching}
+ Matching
%* *
%************************************************************************
-Note [Extra args in rule matching]
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-If we find a matching rule, we return (Just (rule, rhs)),
-but the rule firing has only consumed as many of the input args
-as the ruleArity says. It's up to the caller to keep track
-of any left-over args. E.g. if you call
- lookupRule ... f [e1, e2, e3]
-and it returns Just (r, rhs), where r has ruleArity 2
-then the real rewrite is
- f e1 e2 e3 ==> rhs e3
-
-You might think it'd be cleaner for lookupRule to deal with the
-leftover arguments, by applying 'rhs' to them, but the main call
-in the Simplifier works better as it is. Reason: the 'args' passed
-to lookupRule are the result of a lazy substitution
-
\begin{code}
-- | The main rule matching function. Attempts to apply all (active)
-- supplied rules to this instance of an application in a given
go ms (r:rs) = case (matchRule is_active id_unf in_scope args rough_args r) of
Just e -> go ((r,e):ms) rs
Nothing -> -- pprTrace "match failed" (ppr r $$ ppr args $$
- -- ppr [(arg_id, unfoldingTemplate unf) | Var arg_id <- args, let unf = idUnfolding arg_id, isCheapUnfolding unf] )
- go ms rs
+ -- ppr [ (arg_id, unfoldingTemplate unf)
+ -- | Var arg_id <- args
+ -- , let unf = idUnfolding arg_id
+ -- , isCheapUnfolding unf] )
+ go ms rs
findBest :: (Id, [CoreExpr])
-> (CoreRule,CoreExpr) -> [(CoreRule,CoreExpr)] -> (CoreRule,CoreExpr)
noBlackList :: Activation -> Bool
noBlackList _ = False -- Nothing is black listed
+\end{code}
+
+Note [Extra args in rule matching]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+If we find a matching rule, we return (Just (rule, rhs)),
+but the rule firing has only consumed as many of the input args
+as the ruleArity says. It's up to the caller to keep track
+of any left-over args. E.g. if you call
+ lookupRule ... f [e1, e2, e3]
+and it returns Just (r, rhs), where r has ruleArity 2
+then the real rewrite is
+ f e1 e2 e3 ==> rhs e3
+
+You might think it'd be cleaner for lookupRule to deal with the
+leftover arguments, by applying 'rhs' to them, but the main call
+in the Simplifier works better as it is. Reason: the 'args' passed
+to lookupRule are the result of a lazy substitution
+\begin{code}
+------------------------------------
matchRule :: (Activation -> Bool) -> IdUnfoldingFun
-> InScopeSet
-> [CoreExpr] -> [Maybe Name]
| ruleCantMatch tpl_tops rough_args = Nothing
| otherwise
= case matchN id_unf in_scope tpl_vars tpl_args args of
- Nothing -> Nothing
- Just (binds, tpl_vals) -> Just (mkLets binds $
- rule_fn `mkApps` tpl_vals)
+ Nothing -> Nothing
+ Just (bind_wrapper, tpl_vals) -> Just (bind_wrapper $
+ rule_fn `mkApps` tpl_vals)
where
rule_fn = occurAnalyseExpr (mkLams tpl_vars rhs)
-- We could do this when putting things into the rulebase, I guess
-\end{code}
-\begin{code}
--- For a given match template and context, find bindings to wrap around
--- the entire result and what should be substituted for each template variable.
--- Fail if there are two few actual arguments from the target to match the template
+---------------------------------------
matchN :: IdUnfoldingFun
-> InScopeSet -- ^ In-scope variables
-> [Var] -- ^ Match template type variables
-> [CoreExpr] -- ^ Match template
-> [CoreExpr] -- ^ Target; can have more elements than the template
- -> Maybe ([CoreBind],
+ -> Maybe (BindWrapper, -- Floated bindings; see Note [Matching lets]
[CoreExpr])
+-- For a given match template and context, find bindings to wrap around
+-- the entire result and what should be substituted for each template variable.
+-- Fail if there are two few actual arguments from the target to match the template
matchN id_unf in_scope tmpl_vars tmpl_es target_es
= do { (tv_subst, id_subst, binds)
<- go init_menv emptySubstEnv tmpl_es target_es
- ; return (fromOL binds,
+ ; return (binds,
map (lookup_tmpl tv_subst id_subst) tmpl_vars') }
where
(init_rn_env, tmpl_vars') = mapAccumL rnBndrL (mkRnEnv2 in_scope) tmpl_vars
lookup_tmpl :: TvSubstEnv -> IdSubstEnv -> Var -> CoreExpr
lookup_tmpl tv_subst id_subst tmpl_var'
- | isTyVar tmpl_var' = case lookupVarEnv tv_subst tmpl_var' of
+ | isTyCoVar tmpl_var' = case lookupVarEnv tv_subst tmpl_var' of
Just ty -> Type ty
Nothing -> unbound tmpl_var'
| otherwise = case lookupVarEnv id_subst tmpl_var' of
-- * The domain of the TvSubstEnv and IdSubstEnv are the template
-- variables passed into the match.
--
--- * The (OrdList CoreBind) in a SubstEnv are the bindings floated out
+-- * The BindWrapper in a SubstEnv are the bindings floated out
-- from nested matches; see the Let case of match, below
--
-type SubstEnv = (TvSubstEnv, IdSubstEnv, OrdList CoreBind)
+type SubstEnv = (TvSubstEnv, IdSubstEnv, BindWrapper)
+
+type BindWrapper = CoreExpr -> CoreExpr
+ -- See Notes [Matching lets] and [Matching cases]
+ -- we represent the floated bindings as a core-to-core function
+
type IdSubstEnv = IdEnv CoreExpr
emptySubstEnv :: SubstEnv
-emptySubstEnv = (emptyVarEnv, emptyVarEnv, nilOL)
-
+emptySubstEnv = (emptyVarEnv, emptyVarEnv, \e -> e)
-- At one stage I tried to match even if there are more
-- template args than real args.
-- See Note [Notes in RULE matching]
match id_unfolding_fun menv subst e1 (Var v2) -- Note [Expanding variables]
- | not (locallyBoundR rn_env v2) -- Note [Do not expand locally-bound variables]
- , Just e2' <- expandUnfolding (id_unfolding_fun v2')
+ | not (inRnEnvR rn_env v2) -- Note [Do not expand locally-bound variables]
+ , Just e2' <- expandUnfolding_maybe (id_unfolding_fun v2')
= match id_unfolding_fun (menv { me_env = nukeRnEnvR rn_env }) subst e1 e2'
where
v2' = lookupRnInScope rn_env v2
-- Notice that we look up v2 in the in-scope set
-- See Note [Lookup in-scope]
-- No need to apply any renaming first (hence no rnOccR)
- -- becuase of the not-locallyBoundR
+ -- because of the not-inRnEnvR
match idu menv (tv_subst, id_subst, binds) e1 (Let bind e2)
- | all freshly_bound bndrs -- See Note [Matching lets]
- , not (any (locallyBoundR rn_env) bind_fvs)
+ | okToFloat rn_env bndrs (bindFreeVars bind) -- See Note [Matching lets]
= match idu (menv { me_env = rn_env' })
- (tv_subst, id_subst, binds `snocOL` bind')
- e1 e2'
+ (tv_subst, id_subst, binds . Let bind)
+ e1 e2
where
rn_env = me_env menv
- bndrs = bindersOf bind
- bind_fvs = varSetElems (bindFreeVars bind)
- freshly_bound x = not (x `rnInScope` rn_env)
- bind' = bind
- e2' = e2
- rn_env' = extendRnInScopeList rn_env bndrs
+ rn_env' = extendRnInScopeList rn_env bndrs
+ bndrs = bindersOf bind
+
+{- Disabled: see Note [Matching cases] below
+match idu menv (tv_subst, id_subst, binds) e1
+ (Case scrut case_bndr ty [(con, alt_bndrs, rhs)])
+ | exprOkForSpeculation scrut -- See Note [Matching cases]
+ , okToFloat rn_env bndrs (exprFreeVars scrut)
+ = match idu (menv { me_env = rn_env' })
+ (tv_subst, id_subst, binds . case_wrap)
+ e1 rhs
+ where
+ rn_env = me_env menv
+ rn_env' = extendRnInScopeList rn_env bndrs
+ bndrs = case_bndr : alt_bndrs
+ case_wrap rhs' = Case scrut case_bndr ty [(con, alt_bndrs, rhs')]
+-}
match _ _ subst (Lit lit1) (Lit lit2)
| lit1 == lit2
Nothing
------------------------------------------
+okToFloat :: RnEnv2 -> [Var] -> VarSet -> Bool
+okToFloat rn_env bndrs bind_fvs
+ = all freshly_bound bndrs
+ && foldVarSet ((&&) . not_captured) True bind_fvs
+ where
+ freshly_bound x = not (x `rnInScope` rn_env)
+ not_captured fv = not (inRnEnvR rn_env fv)
+
+------------------------------------------
match_var :: IdUnfoldingFun
-> MatchEnv
-> SubstEnv
-- c.f. match_ty below
; return (tv_subst', extendVarEnv id_subst v1' e2, binds) }
- Just e1' | eqExpr idu (nukeRnEnvL rn_env) e1' e2
+ Just e1' | eqExprX idu (nukeRnEnvL rn_env) e1' e2
-> Just subst
| otherwise
bindings in the SubstEnv
We can only do this if
- (a) Widening the scope of w does not capture any variables
- We use a conservative test: w is not already in scope
- If not, we clone the binders, and substitute
- (b) The free variables of R are not bound by the part of the
- target expression outside the let binding; e.g.
- f (\v. let w = v+1 in g E)
- Here we obviously cannot float the let-binding for w.
+ (a) Widening the scope of w does not capture any variables
+ We use a conservative test: w is not already in scope
+ If not, we clone the binders, and substitute
+ (b) The free variables of R are not bound by the part of the
+ target expression outside the let binding; e.g.
+ f (\v. let w = v+1 in g E)
+ Here we obviously cannot float the let-binding for w.
You may think rule (a) would never apply, because rule matching is
mostly invoked from the simplifier, when we have just run substExpr
(let x=y+1 in (x,x), let x=y-1 in (x,x))
--> let x=y+1 in let x1=y-1 in ((x,x),(x1,x1))
-
+Note [Matching cases]
+~~~~~~~~~~~~~~~~~~~~~
+{- NOTE: This idea is currently disabled. It really only works if
+ the primops involved are OkForSpeculation, and, since
+ they have side effects readIntOfAddr and touch are not.
+ Maybe we'll get back to this later . -}
+
+Consider
+ f (case readIntOffAddr# p# i# realWorld# of { (# s#, n# #) ->
+ case touch# fp s# of { _ ->
+ I# n# } } )
+This happened in a tight loop generated by stream fusion that
+Roman encountered. We'd like to treat this just like the let
+case, because the primops concerned are ok-for-speculation.
+That is, we'd like to behave as if it had been
+ case readIntOffAddr# p# i# realWorld# of { (# s#, n# #) ->
+ case touch# fp s# of { _ ->
+ f (I# n# } } )
+
Note [Lookup in-scope]
~~~~~~~~~~~~~~~~~~~~~~
Consider this example
That is why the 'lookupRnInScope' call in the (Var v2) case of 'match'
is so important.
-\begin{code}
-eqExpr :: IdUnfoldingFun -> RnEnv2 -> CoreExpr -> CoreExpr -> Bool
--- ^ A kind of shallow equality used in rule matching, so does
--- /not/ look through newtypes or predicate types
-
-eqExpr _ env (Var v1) (Var v2)
- | rnOccL env v1 == rnOccR env v2
- = True
-
--- The next two rules expand non-local variables
--- C.f. Note [Expanding variables]
--- and Note [Do not expand locally-bound variables]
-eqExpr id_unfolding_fun env (Var v1) e2
- | not (locallyBoundL env v1)
- , Just e1' <- expandUnfolding (id_unfolding_fun (lookupRnInScope env v1))
- = eqExpr id_unfolding_fun (nukeRnEnvL env) e1' e2
-
-eqExpr id_unfolding_fun env e1 (Var v2)
- | not (locallyBoundR env v2)
- , Just e2' <- expandUnfolding (id_unfolding_fun (lookupRnInScope env v2))
- = eqExpr id_unfolding_fun (nukeRnEnvR env) e1 e2'
-
-eqExpr _ _ (Lit lit1) (Lit lit2) = lit1 == lit2
-eqExpr idu env (App f1 a1) (App f2 a2) = eqExpr idu env f1 f2 && eqExpr idu env a1 a2
-eqExpr idu env (Lam v1 e1) (Lam v2 e2) = eqExpr idu (rnBndr2 env v1 v2) e1 e2
-eqExpr idu env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && eqExpr idu env e1 e2
-eqExpr idu env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && eqExpr idu env e1 e2
-eqExpr _ env (Type t1) (Type t2) = tcEqTypeX env t1 t2
-
-eqExpr idu env (Let (NonRec v1 r1) e1)
- (Let (NonRec v2 r2) e2) = eqExpr idu env r1 r2
- && eqExpr idu (rnBndr2 env v1 v2) e1 e2
-eqExpr idu env (Let (Rec ps1) e1)
- (Let (Rec ps2) e2) = equalLength ps1 ps2
- && and (zipWith eq_rhs ps1 ps2)
- && eqExpr idu env' e1 e2
- where
- env' = foldl2 rn_bndr2 env ps2 ps2
- rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
- eq_rhs (_,r1) (_,r2) = eqExpr idu env' r1 r2
-eqExpr idu env (Case e1 v1 t1 a1)
- (Case e2 v2 t2 a2) = eqExpr idu env e1 e2
- && tcEqTypeX env t1 t2
- && equalLength a1 a2
- && and (zipWith eq_alt a1 a2)
- where
- env' = rnBndr2 env v1 v2
- eq_alt (c1,vs1,r1) (c2,vs2,r2)
- = c1==c2 && eqExpr idu (rnBndrs2 env' vs1 vs2) r1 r2
-eqExpr _ _ _ _ = False
-
-eq_note :: RnEnv2 -> Note -> Note -> Bool
-eq_note _ (SCC cc1) (SCC cc2) = cc1 == cc2
-eq_note _ (CoreNote s1) (CoreNote s2) = s1 == s2
-eq_note _ _ _ = False
-\end{code}
-
-Auxiliary functions
-
-\begin{code}
-locallyBoundL, locallyBoundR :: RnEnv2 -> Var -> Bool
-locallyBoundL rn_env v = inRnEnvL rn_env v
-locallyBoundR rn_env v = inRnEnvR rn_env v
-
-
-expandUnfolding :: Unfolding -> Maybe CoreExpr
-expandUnfolding unfolding
- | isExpandableUnfolding unfolding = Just (unfoldingTemplate unfolding)
- | otherwise = Nothing
-\end{code}
-
%************************************************************************
%* *
Rule-check the program