\begin{code}
module SimplUtils (
-- Rebuilding
- mkLam, mkCase, prepareAlts, bindCaseBndr,
+ mkLam, mkCase, prepareAlts,
-- Inlining,
preInlineUnconditionally, postInlineUnconditionally,
- activeInline, activeRule,
+ activeUnfolding, activeUnfInRule, activeRule,
simplEnvForGHCi, simplEnvForRules, updModeForInlineRules,
-- The continuation type
import Name
import Id
import Var ( isCoVar )
-import NewDemand
+import Demand
import SimplMonad
import Type hiding( substTy )
import Coercion ( coercionKind )
vanilla_discounts, arg_discounts :: [Int]
vanilla_discounts = repeat 0
arg_discounts = case idUnfolding fun of
- CoreUnfolding {uf_guidance = UnfoldIfGoodArgs {ug_args = discounts}}
+ CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
-> discounts ++ vanilla_discounts
_ -> vanilla_discounts
vanilla_stricts = repeat False
arg_stricts
- = case splitStrictSig (idNewStrictness fun) of
+ = case splitStrictSig (idStrictness fun) of
(demands, result_info)
| not (demands `lengthExceeds` n_val_args)
-> -- Enough args, use the strictness given.
-> Unfolding
-> Bool
postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
- | not active = False
- | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
+ | not active = False
+ | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
-- because it might be referred to "earlier"
- | isExportedId bndr = False
- | isInlineRule unfolding = False -- Note [InlineRule and postInlineUnconditionally]
- | exprIsTrivial rhs = True
+ | isExportedId bndr = False
+ | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
+ | exprIsTrivial rhs = True
| otherwise
= case occ_info of
-- The point of examining occ_info here is that for *non-values*
-- case v of
-- True -> case x of ...
-- False -> case x of ...
- -- I'm not sure how important this is in practice
+ -- This is very important in practice; e.g. wheel-seive1 doubles
+ -- in allocation if you miss this out
OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
-> smallEnoughToInline unfolding -- Small enough to dup
-- ToDo: consider discount on smallEnoughToInline if int_cxt is true
SimplPhase n _ -> isActive n act
act = idInlineActivation bndr
-activeInline :: SimplEnv -> OutId -> Bool
-activeInline env id
- | isNonRuleLoopBreaker (idOccInfo id) -- Things with an INLINE pragma may have
- -- an unfolding *and* be a loop breaker
- = False -- (maybe the knot is not yet untied)
- | otherwise
+activeUnfolding :: SimplEnv -> IdUnfoldingFun
+activeUnfolding env
+ = case getMode env of
+ SimplGently { sm_inline = False } -> active_unfolding_minimal
+ SimplGently { sm_inline = True } -> active_unfolding_gentle
+ SimplPhase n _ -> active_unfolding n
+
+activeUnfInRule :: SimplEnv -> IdUnfoldingFun
+-- When matching in RULE, we want to "look through" an unfolding
+-- if *rules* are on, even if *inlinings* are not. A notable example
+-- is DFuns, which really we want to match in rules like (op dfun)
+-- in gentle mode.
+activeUnfInRule env
= case getMode env of
- SimplGently { sm_inline = inlining_on }
- -> inlining_on && isEarlyActive act
- -- See Note [Gentle mode]
-
- -- NB: we used to have a second exception, for data con wrappers.
- -- On the grounds that we use gentle mode for rule LHSs, and
- -- they match better when data con wrappers are inlined.
- -- But that only really applies to the trivial wrappers (like (:)),
- -- and they are now constructed as Compulsory unfoldings (in MkId)
- -- so they'll happen anyway.
-
- SimplPhase n _ -> isActive n act
+ SimplGently { sm_rules = False } -> active_unfolding_minimal
+ SimplGently { sm_rules = True } -> active_unfolding_gentle
+ SimplPhase n _ -> active_unfolding n
+
+active_unfolding_minimal :: IdUnfoldingFun
+-- Compuslory unfoldings only
+-- Ignore SimplGently, because we want to inline regardless;
+-- the Id has no top-level binding at all
+--
+-- NB: we used to have a second exception, for data con wrappers.
+-- On the grounds that we use gentle mode for rule LHSs, and
+-- they match better when data con wrappers are inlined.
+-- But that only really applies to the trivial wrappers (like (:)),
+-- and they are now constructed as Compulsory unfoldings (in MkId)
+-- so they'll happen anyway.
+active_unfolding_minimal id
+ | isCompulsoryUnfolding unf = unf
+ | otherwise = NoUnfolding
where
- act = idInlineActivation id
+ unf = realIdUnfolding id -- Never a loop breaker
+
+active_unfolding_gentle :: IdUnfoldingFun
+-- Anything that is early-active
+-- See Note [Gentle mode]
+active_unfolding_gentle id
+ | isEarlyActive (idInlineActivation id) = idUnfolding id
+ | otherwise = NoUnfolding
+ -- idUnfolding checks for loop-breakers
+ -- Things with an INLINE pragma may have
+ -- an unfolding *and* be a loop breaker
+ -- (maybe the knot is not yet untied)
+
+active_unfolding :: CompilerPhase -> IdUnfoldingFun
+active_unfolding n id
+ | isActive n (idInlineActivation id) = idUnfolding id
+ | otherwise = NoUnfolding
activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
-- Nothing => No rules at all
prepareAlts tries these things:
-1. If several alternatives are identical, merge them into
- a single DEFAULT alternative. I've occasionally seen this
- making a big difference:
+1. Eliminate alternatives that cannot match, including the
+ DEFAULT alternative.
- case e of =====> case e of
- C _ -> f x D v -> ....v....
- D v -> ....v.... DEFAULT -> f x
- DEFAULT -> f x
+2. If the DEFAULT alternative can match only one possible constructor,
+ then make that constructor explicit.
+ e.g.
+ case e of x { DEFAULT -> rhs }
+ ===>
+ case e of x { (a,b) -> rhs }
+ where the type is a single constructor type. This gives better code
+ when rhs also scrutinises x or e.
- The point is that we merge common RHSs, at least for the DEFAULT case.
- [One could do something more elaborate but I've never seen it needed.]
- To avoid an expensive test, we just merge branches equal to the *first*
- alternative; this picks up the common cases
- a) all branches equal
- b) some branches equal to the DEFAULT (which occurs first)
+3. Returns a list of the constructors that cannot holds in the
+ DEFAULT alternative (if there is one)
-2. Case merging:
- case e of b { ==> case e of b {
- p1 -> rhs1 p1 -> rhs1
- ... ...
- pm -> rhsm pm -> rhsm
- _ -> case b of b' { pn -> let b'=b in rhsn
- pn -> rhsn ...
- ... po -> let b'=b in rhso
- po -> rhso _ -> let b'=b in rhsd
- _ -> rhsd
- }
-
- which merges two cases in one case when -- the default alternative of
- the outer case scrutises the same variable as the outer case This
- transformation is called Case Merging. It avoids that the same
- variable is scrutinised multiple times.
+Here "cannot match" includes knowledge from GADTs
+It's a good idea do do this stuff before simplifying the alternatives, to
+avoid simplifying alternatives we know can't happen, and to come up with
+the list of constructors that are handled, to put into the IdInfo of the
+case binder, for use when simplifying the alternatives.
-The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
+Eliminating the default alternative in (1) isn't so obvious, but it can
+happen:
- x | p `is` 1 -> e1
- | p `is` 2 -> e2
- ...etc...
+data Colour = Red | Green | Blue
-where @is@ was something like
-
- p `is` n = p /= (-1) && p == n
+f x = case x of
+ Red -> ..
+ Green -> ..
+ DEFAULT -> h x
-This gave rise to a horrible sequence of cases
+h y = case y of
+ Blue -> ..
+ DEFAULT -> [ case y of ... ]
- case p of
- (-1) -> $j p
- 1 -> e1
- DEFAULT -> $j p
-
-and similarly in cascade for all the join points!
-
-Note [Dead binders]
-~~~~~~~~~~~~~~~~~~~~
-We do this *here*, looking at un-simplified alternatives, because we
-have to check that r doesn't mention the variables bound by the
-pattern in each alternative, so the binder-info is rather useful.
+If we inline h into f, the default case of the inlined h can't happen.
+If we don't notice this, we may end up filtering out *all* the cases
+of the inner case y, which give us nowhere to go!
\begin{code}
-prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
-prepareAlts env scrut case_bndr' alts
- = do { dflags <- getDOptsSmpl
- ; alts <- combineIdenticalAlts case_bndr' alts
-
- ; let (alts_wo_default, maybe_deflt) = findDefault alts
+prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
+prepareAlts scrut case_bndr' alts
+ = do { let (alts_wo_default, maybe_deflt) = findDefault alts
alt_cons = [con | (con,_,_) <- alts_wo_default]
imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
-- "imposs_deflt_cons" are handled
-- EITHER by the context,
-- OR by a non-DEFAULT branch in this case expression.
- ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
+ ; default_alts <- prepareDefault case_bndr' mb_tc_app
imposs_deflt_cons maybe_deflt
; let trimmed_alts = filterOut impossible_alt alts_wo_default
- merged_alts = mergeAlts trimmed_alts default_alts
+ merged_alts = mergeAlts trimmed_alts default_alts
-- We need the mergeAlts in case the new default_alt
-- has turned into a constructor alternative.
-- The merge keeps the inner DEFAULT at the front, if there is one
impossible_alt _ = False
---------------------------------------------------
--- 1. Merge identical branches
---------------------------------------------------
-combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
-
-combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
- | all isDeadBinder bndrs1, -- Remember the default
- length filtered_alts < length con_alts -- alternative comes first
- -- Also Note [Dead binders]
- = do { tick (AltMerge case_bndr)
- ; return ((DEFAULT, [], rhs1) : filtered_alts) }
- where
- filtered_alts = filter keep con_alts
- keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
-
-combineIdenticalAlts _ alts = return alts
-
--------------------------------------------------------------------------
--- Prepare the default alternative
--------------------------------------------------------------------------
-prepareDefault :: DynFlags
- -> SimplEnv
- -> OutId -- Case binder; need just for its type. Note that as an
+prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
-- OutId, it has maximum information; this is important.
-- Test simpl013 is an example
-> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
-> Maybe InExpr -- Rhs
-> SimplM [InAlt] -- Still unsimplified
-- We use a list because it's what mergeAlts expects,
- -- And becuase case-merging can cause many to show up
-
-------- Merge nested cases ----------
-prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
- | dopt Opt_CaseMerge dflags
- , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
- , DoneId inner_scrut_var' <- substId env inner_scrut_var
- -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
- , inner_scrut_var' == outer_bndr
- -- NB: the substId means that if the outer scrutinee was a
- -- variable, and inner scrutinee is the same variable,
- -- then inner_scrut_var' will be outer_bndr
- -- via the magic of simplCaseBinder
- = do { tick (CaseMerge outer_bndr)
-
- ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
- ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
- not (con `elem` imposs_cons) ]
- -- NB: filter out any imposs_cons. Example:
- -- case x of
- -- A -> e1
- -- DEFAULT -> case x of
- -- A -> e2
- -- B -> e3
- -- When we merge, we must ensure that e1 takes
- -- precedence over e2 as the value for A!
- }
- -- Warning: don't call prepareAlts recursively!
- -- Firstly, there's no point, because inner alts have already had
- -- mkCase applied to them, so they won't have a case in their default
- -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
- -- in munge_rhs may put a case into the DEFAULT branch!
-
--------- Fill in known constructor -----------
-prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
+prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
| -- This branch handles the case where we are
-- scrutinisng an algebraic data type
isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
-- which would be quite legitmate. But it's a really obscure corner, and
-- not worth wasting code on.
, let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
- impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
+ impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
= case filterOut impossible all_cons of
[] -> return [] -- Eliminate the default alternative
-- altogether if it can't match
_ -> return [(DEFAULT, [], deflt_rhs)]
| debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
- -- This can legitimately happen for type families, so don't report that
+ -- Check for no data constructors
+ -- This can legitimately happen for type families, so don't report that
= pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
$ return [(DEFAULT, [], deflt_rhs)]
--------- Catch-all cases -----------
-prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
+prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
= return [(DEFAULT, [], deflt_rhs)]
-prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
+prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
= return [] -- No default branch
\end{code}
-=================================================================================
+%************************************************************************
+%* *
+ mkCase
+%* *
+%************************************************************************
mkCase tries these things
-1. Eliminate the case altogether if possible
+1. Merge Nested Cases
+
+ case e of b { ==> case e of b {
+ p1 -> rhs1 p1 -> rhs1
+ ... ...
+ pm -> rhsm pm -> rhsm
+ _ -> case b of b' { pn -> let b'=b in rhsn
+ pn -> rhsn ...
+ ... po -> let b'=b in rhso
+ po -> rhso _ -> let b'=b in rhsd
+ _ -> rhsd
+ }
+
+ which merges two cases in one case when -- the default alternative of
+ the outer case scrutises the same variable as the outer case. This
+ transformation is called Case Merging. It avoids that the same
+ variable is scrutinised multiple times.
-2. Case-identity:
+2. Eliminate Identity Case
case e of ===> e
True -> True;
and similar friends.
+3. Merge identical alternatives.
+ If several alternatives are identical, merge them into
+ a single DEFAULT alternative. I've occasionally seen this
+ making a big difference:
+
+ case e of =====> case e of
+ C _ -> f x D v -> ....v....
+ D v -> ....v.... DEFAULT -> f x
+ DEFAULT -> f x
+
+ The point is that we merge common RHSs, at least for the DEFAULT case.
+ [One could do something more elaborate but I've never seen it needed.]
+ To avoid an expensive test, we just merge branches equal to the *first*
+ alternative; this picks up the common cases
+ a) all branches equal
+ b) some branches equal to the DEFAULT (which occurs first)
+
+The case where Merge Identical Alternatives transformation showed up
+was like this (base/Foreign/C/Err/Error.lhs):
+
+ x | p `is` 1 -> e1
+ | p `is` 2 -> e2
+ ...etc...
+
+where @is@ was something like
+
+ p `is` n = p /= (-1) && p == n
+
+This gave rise to a horrible sequence of cases
+
+ case p of
+ (-1) -> $j p
+ 1 -> e1
+ DEFAULT -> $j p
+
+and similarly in cascade for all the join points!
+
\begin{code}
-mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
- -> SimplM OutExpr
+mkCase, mkCase1, mkCase2
+ :: DynFlags
+ -> OutExpr -> OutId
+ -> [OutAlt] -- Alternatives in standard (increasing) order
+ -> SimplM OutExpr
+
+--------------------------------------------------
+-- 1. Merge Nested Cases
+--------------------------------------------------
+
+mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
+ | dopt Opt_CaseMerge dflags
+ , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
+ , inner_scrut_var == outer_bndr
+ = do { tick (CaseMerge outer_bndr)
+
+ ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
+ (con, args, wrap_rhs rhs)
+ -- Simplifier's no-shadowing invariant should ensure
+ -- that outer_bndr is not shadowed by the inner patterns
+ wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
+ -- The let is OK even for unboxed binders,
+
+ wrapped_alts | isDeadBinder inner_bndr = inner_alts
+ | otherwise = map wrap_alt inner_alts
+
+ merged_alts = mergeAlts outer_alts wrapped_alts
+ -- NB: mergeAlts gives priority to the left
+ -- case x of
+ -- A -> e1
+ -- DEFAULT -> case x of
+ -- A -> e2
+ -- B -> e3
+ -- When we merge, we must ensure that e1 takes
+ -- precedence over e2 as the value for A!
+
+ ; mkCase1 dflags scrut outer_bndr merged_alts
+ }
+ -- Warning: don't call mkCase recursively!
+ -- Firstly, there's no point, because inner alts have already had
+ -- mkCase applied to them, so they won't have a case in their default
+ -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
+ -- in munge_rhs may put a case into the DEFAULT branch!
+
+mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
--------------------------------------------------
--- 2. Identity case
+-- 2. Eliminate Identity Case
--------------------------------------------------
-mkCase scrut case_bndr alts -- Identity case
+mkCase1 _dflags scrut case_bndr alts -- Identity case
| all identity_alt alts
- = do tick (CaseIdentity case_bndr)
- return (re_cast scrut)
+ = do { tick (CaseIdentity case_bndr)
+ ; return (re_cast scrut) }
where
identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
(_,_,Cast _ co) -> Cast scrut co
_ -> scrut
+--------------------------------------------------
+-- 3. Merge Identical Alternatives
+--------------------------------------------------
+mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
+ | all isDeadBinder bndrs1 -- Remember the default
+ , length filtered_alts < length con_alts -- alternative comes first
+ -- Also Note [Dead binders]
+ = do { tick (AltMerge case_bndr)
+ ; mkCase2 dflags scrut case_bndr alts' }
+ where
+ alts' = (DEFAULT, [], rhs1) : filtered_alts
+ filtered_alts = filter keep con_alts
+ keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
+mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
--------------------------------------------------
-- Catch-all
--------------------------------------------------
-mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
+mkCase2 _dflags scrut bndr alts
+ = return (Case scrut bndr (coreAltsType alts) alts)
\end{code}
-
-When adding auxiliary bindings for the case binder, it's worth checking if
-its dead, because it often is, and occasionally these mkCase transformations
-cascade rather nicely.
-
-\begin{code}
-bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
-bindCaseBndr bndr rhs body
- | isDeadBinder bndr = body
- | otherwise = bindNonRec bndr rhs body
-\end{code}
+Note [Dead binders]
+~~~~~~~~~~~~~~~~~~~~
+Note that dead-ness is maintained by the simplifier, so that it is
+accurate after simplification as well as before.
+
+
+Note [Cascading case merge]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Case merging should cascade in one sweep, because it
+happens bottom-up
+
+ case e of a {
+ DEFAULT -> case a of b
+ DEFAULT -> case b of c {
+ DEFAULT -> e
+ A -> ea
+ B -> eb
+ C -> ec
+==>
+ case e of a {
+ DEFAULT -> case a of b
+ DEFAULT -> let c = b in e
+ A -> let c = b in ea
+ B -> eb
+ C -> ec
+==>
+ case e of a {
+ DEFAULT -> let b = a in let c = b in e
+ A -> let b = a in let c = b in ea
+ B -> let b = a in eb
+ C -> ec
+
+
+However here's a tricky case that we still don't catch, and I don't
+see how to catch it in one pass:
+
+ case x of c1 { I# a1 ->
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case x of c3 { I# a2 ->
+ case a2 of ...
+
+After occurrence analysis (and its binder-swap) we get this
+
+ case x of c1 { I# a1 ->
+ let x = c1 in -- Binder-swap addition
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case x of c3 { I# a2 ->
+ case a2 of ...
+
+When we simplify the inner case x, we'll see that
+x=c1=I# a1. So we'll bind a2 to a1, and get
+
+ case x of c1 { I# a1 ->
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case a1 of ...
+
+This is corect, but we can't do a case merge in this sweep
+because c2 /= a1. Reason: the binding c1=I# a1 went inwards
+without getting changed to c1=I# c2.
+
+I don't think this is worth fixing, even if I knew how. It'll
+all come out in the next pass anyway.
+
+
\ No newline at end of file