= tagZero xs #-}
So tagZero's RHS mentions pmap, and pmap's RULE mentions tagZero.
However, tagZero can only be inlined in phase 1 and later, while
the RULE is only active *before* phase 1. So there's no problem.
To make this work, we look for the RHS free vars only for
*active* rules. That's the reason for the is_active argument
to idRhsRuleVars, and the occ_rule_act field of the OccEnv.
* Note [Weak loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~
There is a last nasty wrinkle. Suppose we have
Rec { f = f_rhs
RULE f [] = g
h = h_rhs
g = h
...more...
}
Remember that we simplify the RULES before any RHS (see Note
[Rules are visible in their own rec group] above).
So we must *not* postInlineUnconditionally 'g', even though
its RHS turns out to be trivial. (I'm assuming that 'g' is
not choosen as a loop breaker.) Why not? Because then we
drop the binding for 'g', which leaves it out of scope in the
RULE!
We "solve" this by making g a "weak" or "rules-only" loop breaker,
with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
has IAmLoopBreaker False. So
Inline postInlineUnconditionally
IAmLoopBreaker False no no
IAmLoopBreaker True yes no
other yes yes
The **sole** reason for this kind of loop breaker is so that
postInlineUnconditionally does not fire. Ugh.
* Note [Rule dependency info]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
The VarSet in a SpecInfo is used for dependency analysis in the
occurrence analyser. We must track free vars in *both* lhs and rhs.
Hence use of idRuleVars, rather than idRuleRhsVars in occAnalBind.
Why both? Consider
x = y
RULE f x = 4
Then if we substitute y for x, we'd better do so in the
rule's LHS too, so we'd better ensure the dependency is respected
* Note [Inline rules]
~~~~~~~~~~~~~~~~~~~
None of the above stuff about RULES applies to Inline Rules,
stored in a CoreUnfolding. The unfolding, if any, is simplified
at the same time as the regular RHS of the function, so it should
be treated *exactly* like an extra RHS.
There is a danger that we'll be sub-optimal if we see this
f = ...f...
[INLINE f = ..no f...]
where f is recursive, but the INLINE is not. This can just about
happen with a sufficiently odd set of rules; eg
foo :: Int -> Int
{-# INLINE [1] foo #-}
foo x = x+1
bar :: Int -> Int
{-# INLINE [1] bar #-}
bar x = foo x + 1
{-# RULES "foo" [~1] forall x. foo x = bar x #-}
Here the RULE makes bar recursive; but it's INLINE pragma remains
non-recursive. It's tempting to then say that 'bar' should not be
a loop breaker, but an attempt to do so goes wrong in two ways:
a) We may get
$df = ...$cfoo...
$cfoo = ...$df....
[INLINE $cfoo = ...no-$df...]
But we want $cfoo to depend on $df explicitly so that we
put the bindings in the right order to inline $df in $cfoo
and perhaps break the loop altogether. (Maybe this
b)
Example [eftInt]
~~~~~~~~~~~~~~~
Example (from GHC.Enum):
eftInt :: Int# -> Int# -> [Int]
eftInt x y = ...(non-recursive)...
{-# INLINE [0] eftIntFB #-}
eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
eftIntFB c n x y = ...(non-recursive)...
{-# RULES
"eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
"eftIntList" [1] eftIntFB (:) [] = eftInt
#-}
Example [Specialisation rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this group, which is typical of what SpecConstr builds:
fs a = ....f (C a)....
f x = ....f (C a)....
{-# RULE f (C a) = fs a #-}
So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
- the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
- fs is inlined (say it's small)
- now there's another opportunity to apply the RULE
This showed up when compiling Control.Concurrent.Chan.getChanContents.
\begin{code}
occAnalBind _ env (Rec pairs) body_usage
= foldr (occAnalRec env) (body_usage, []) sccs
where
bndr_set = mkVarSet (map fst pairs)
sccs :: [SCC (Node Details)]
sccs = stronglyConnCompFromEdgedVerticesR rec_edges
rec_edges :: [Node Details]
rec_edges = map make_node pairs
make_node (bndr, rhs)
= (details, varUnique bndr, keysUFM out_edges)
where
details = ND { nd_bndr = bndr, nd_rhs = rhs'
, nd_uds = rhs_usage3, nd_inl = inl_fvs}
(rhs_usage1, rhs') = occAnalRhs env NoOccInfo rhs
rhs_usage2 = addIdOccs rhs_usage1 rule_fvs
rhs_usage3 = addIdOccs rhs_usage2 unf_fvs
unf = realIdUnfolding bndr
unf_fvs = stableUnfoldingVars unf
rule_fvs = idRuleVars bndr
inl_fvs = rhs_fvs `unionVarSet` unf_fvs
rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage1
out_edges = intersectUFM_C (\b _ -> b) bndr_set rhs_usage3
occAnalRec :: OccEnv -> SCC (Node Details)
-> (UsageDetails, [CoreBind])
-> (UsageDetails, [CoreBind])
occAnalRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs, nd_uds = rhs_usage}, _, _))
(body_usage, binds)
| not (bndr `usedIn` body_usage)
= (body_usage, binds)
| otherwise
= (body_usage' +++ rhs_usage,
NonRec tagged_bndr rhs : binds)
where
(body_usage', tagged_bndr) = tagBinder body_usage bndr
occAnalRec env (CyclicSCC nodes) (body_usage, binds)
| not (any (`usedIn` body_usage) bndrs)
= (body_usage, binds)
| otherwise
= (final_usage, Rec pairs : binds)
where
bndrs = [b | (ND { nd_bndr = b }, _, _) <- nodes]
bndr_set = mkVarSet bndrs
non_boring bndr = isId bndr &&
(isStableUnfolding (realIdUnfolding bndr) || idHasRules bndr)
total_usage = foldl add_usage body_usage nodes
add_usage usage_so_far (ND { nd_uds = rhs_usage }, _, _) = usage_so_far +++ rhs_usage
(final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
tag_node usage (details@ND { nd_bndr = bndr }, k, ks)
= (usage `delVarEnv` bndr, (details { nd_bndr = bndr2 }, k, ks))
where
bndr2 | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr1
| otherwise = bndr1
bndr1 = setBinderOcc usage bndr
all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars)
emptyVarSet bndrs
pairs | any non_boring bndrs
= foldr (reOrderRec 0) [] $
stronglyConnCompFromEdgedVerticesR loop_breaker_edges
| otherwise
= reOrderCycle 0 tagged_nodes []
loop_breaker_edges = map mk_node tagged_nodes
mk_node (details@(ND { nd_inl = inl_fvs }), k, _) = (details, k, new_ks)
where
new_ks = keysUFM (fst (extendFvs rule_fv_env inl_fvs))
rule_fv_env :: IdEnv IdSet
rule_fv_env = transClosureFV init_rule_fvs
init_rule_fvs
| Just is_active <- occ_rule_act env
= [ (b, rule_fvs)
| b <- bndrs
, isId b
, let rule_fvs = idRuleRhsVars is_active b
`intersectVarSet` bndr_set
, not (isEmptyVarSet rule_fvs)]
| otherwise
= []
\end{code}
@reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
strongly connected component (there's guaranteed to be a cycle). It returns the
same pairs, but
a) in a better order,
b) with some of the Ids having a IAmALoopBreaker pragma
The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
that the simplifier can guarantee not to loop provided it never records an inlining
for these no-inline guys.
Furthermore, the order of the binds is such that if we neglect dependencies
on the no-inline Ids then the binds are topologically sorted. This means
that the simplifier will generally do a good job if it works from top bottom,
recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
==============
[June 98: I don't understand the following paragraphs, and I've
changed the a=b case again so that it isn't a special case any more.]
Here's a case that bit me:
letrec
a = b
b = \x. BIG
in
...a...a...a....
Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
Perhaps something cleverer would suffice.
===============
\begin{code}
type Node details = (details, Unique, [Unique])
data Details
= ND { nd_bndr :: Id
, nd_rhs :: CoreExpr
, nd_uds :: UsageDetails
, nd_inl :: IdSet
}
reOrderRec :: Int -> SCC (Node Details)
-> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
reOrderRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _))
pairs = (bndr, rhs) : pairs
reOrderRec depth (CyclicSCC cycle) pairs = reOrderCycle depth cycle pairs
reOrderCycle :: Int -> [Node Details] -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
reOrderCycle _ [] _
= panic "reOrderCycle"
reOrderCycle _ [(ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)] pairs
=
(makeLoopBreaker False bndr, rhs) : pairs
reOrderCycle depth (bind : binds) pairs
=
foldr (reOrderRec new_depth)
([ (makeLoopBreaker False bndr, rhs)
| (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _) <- chosen_binds] ++ pairs)
(stronglyConnCompFromEdgedVerticesR unchosen)
where
(chosen_binds, unchosen) = choose_loop_breaker [bind] (score bind) [] binds
approximate_loop_breaker = depth >= 2
new_depth | approximate_loop_breaker = 0
| otherwise = depth+1
choose_loop_breaker loop_binds _loop_sc acc []
= (loop_binds, acc)
choose_loop_breaker loop_binds loop_sc acc (bind : binds)
| sc < loop_sc
= choose_loop_breaker [bind] sc (loop_binds ++ acc) binds
| approximate_loop_breaker && sc == loop_sc
= choose_loop_breaker (bind : loop_binds) loop_sc acc binds
| otherwise
= choose_loop_breaker loop_binds loop_sc (bind : acc) binds
where
sc = score bind
score :: Node Details -> Int
score (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)
| not (isId bndr) = 100
| isDFunId bndr = 9
| Just inl_source <- isStableCoreUnfolding_maybe (idUnfolding bndr)
= case inl_source of
InlineWrapper {} -> 10
_other -> 3
| is_con_app rhs = 5
| exprIsTrivial rhs = 10
| isOneOcc (idOccInfo bndr) = 2
| canUnfold (realIdUnfolding bndr) = 1
| otherwise = 0
is_con_app (Var v) = isConLikeId v
is_con_app (App f _) = is_con_app f
is_con_app (Lam _ e) = is_con_app e
is_con_app (Note _ e) = is_con_app e
is_con_app _ = False
makeLoopBreaker :: Bool -> Id -> Id
makeLoopBreaker weak bndr
= ASSERT2( isId bndr, ppr bndr ) setIdOccInfo bndr (IAmALoopBreaker weak)
\end{code}
Note [Complexity of loop breaking]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The loop-breaking algorithm knocks out one binder at a time, and
performs a new SCC analysis on the remaining binders. That can
behave very badly in tightly-coupled groups of bindings; in the
worst case it can be (N**2)*log N, because it does a full SCC
on N, then N-1, then N-2 and so on.
To avoid this, we switch plans after 2 (or whatever) attempts:
Plan A: pick one binder with the lowest score, make it
a loop breaker, and try again
Plan B: pick *all* binders with the lowest score, make them
all loop breakers, and try again
Since there are only a small finite number of scores, this will
terminate in a constant number of iterations, rather than O(N)
iterations.
You might thing that it's very unlikely, but RULES make it much
more likely. Here's a real example from Trac #1969:
Rec { $dm = \d.\x. op d
{-# RULES forall d. $dm Int d = $s$dm1
forall d. $dm Bool d = $s$dm2 #-}
dInt = MkD .... opInt ...
dInt = MkD .... opBool ...
opInt = $dm dInt
opBool = $dm dBool
$s$dm1 = \x. op dInt
$s$dm2 = \x. op dBool }
The RULES stuff means that we can't choose $dm as a loop breaker
(Note [Choosing loop breakers]), so we must choose at least (say)
opInt *and* opBool, and so on. The number of loop breakders is
linear in the number of instance declarations.
Note [INLINE pragmas]
~~~~~~~~~~~~~~~~~~~~~
Avoid choosing a function with an INLINE pramga as the loop breaker!
If such a function is mutually-recursive with a non-INLINE thing,
then the latter should be the loop-breaker.
Usually this is just a question of optimisation. But a particularly
bad case is wrappers generated by the demand analyser: if you make
then into a loop breaker you may get an infinite inlining loop. For
example:
rec {
$wfoo x = ....foo x....
{-loop brk-} foo x = ...$wfoo x...
}
The interface file sees the unfolding for $wfoo, and sees that foo is
strict (and hence it gets an auto-generated wrapper). Result: an
infinite inlining in the importing scope. So be a bit careful if you
change this. A good example is Tree.repTree in
nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
breaker then compiling Game.hs goes into an infinite loop. This
happened when we gave is_con_app a lower score than inline candidates:
Tree.repTree
= __inline_me (/\a. \w w1 w2 ->
case Tree.$wrepTree @ a w w1 w2 of
{ (# ww1, ww2 #) -> Branch @ a ww1 ww2 })
Tree.$wrepTree
= /\a w w1 w2 ->
(# w2_smP, map a (Tree a) (Tree.repTree a w1 w) (w w2) #)
Here we do *not* want to choose 'repTree' as the loop breaker.
Note [DFuns should not be loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It's particularly bad to make a DFun into a loop breaker. See
Note [How instance declarations are translated] in TcInstDcls
We give DFuns a higher score than ordinary CONLIKE things because
if there's a choice we want the DFun to be the non-looop breker. Eg
rec { sc = /\ a \$dC. $fBWrap (T a) ($fCT @ a $dC)
$fCT :: forall a_afE. (Roman.C a_afE) => Roman.C (Roman.T a_afE)
{-# DFUN #-}
$fCT = /\a \$dC. MkD (T a) ((sc @ a $dC) |> blah) ($ctoF @ a $dC)
}
Here 'sc' (the superclass) looks CONLIKE, but we'll never get to it
if we can't unravel the DFun first.
Note [Constructor applications]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It's really really important to inline dictionaries. Real
example (the Enum Ordering instance from GHC.Base):
rec f = \ x -> case d of (p,q,r) -> p x
g = \ x -> case d of (p,q,r) -> q x
d = (v, f, g)
Here, f and g occur just once; but we can't inline them into d.
On the other hand we *could* simplify those case expressions if
we didn't stupidly choose d as the loop breaker.
But we won't because constructor args are marked "Many".
Inlining dictionaries is really essential to unravelling
the loops in static numeric dictionaries, see GHC.Float.
Note [Closure conversion]
~~~~~~~~~~~~~~~~~~~~~~~~~
We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
The immediate motivation came from the result of a closure-conversion transformation
which generated code like this:
data Clo a b = forall c. Clo (c -> a -> b) c
($:) :: Clo a b -> a -> b
Clo f env $: x = f env x
rec { plus = Clo plus1 ()
; plus1 _ n = Clo plus2 n
; plus2 Zero n = n
; plus2 (Succ m) n = Succ (plus $: m $: n) }
If we inline 'plus' and 'plus1', everything unravels nicely. But if
we choose 'plus1' as the loop breaker (which is entirely possible
otherwise), the loop does not unravel nicely.
@occAnalRhs@ deals with the question of bindings where the Id is marked
by an INLINE pragma. For these we record that anything which occurs
in its RHS occurs many times. This pessimistically assumes that ths
inlined binder also occurs many times in its scope, but if it doesn't
we'll catch it next time round. At worst this costs an extra simplifier pass.
ToDo: try using the occurrence info for the inline'd binder.
[March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
[June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
\begin{code}
occAnalRhs :: OccEnv
-> OccInfo -> CoreExpr
-> (UsageDetails, CoreExpr)
occAnalRhs env occ rhs
= occAnal ctxt rhs
where
ctxt | certainly_inline = env
| otherwise = rhsCtxt env
certainly_inline = case occ of
OneOcc in_lam one_br _ -> not in_lam && one_br
_ -> False
addIdOccs :: UsageDetails -> VarSet -> UsageDetails
addIdOccs usage id_set = foldVarSet add usage id_set
where
add v u | isId v = addOneOcc u v NoOccInfo
| otherwise = u
\end{code}
Expressions
~~~~~~~~~~~
\begin{code}
occAnal :: OccEnv
-> CoreExpr
-> (UsageDetails,
CoreExpr)
occAnal _ (Type t) = (emptyDetails, Type t)
occAnal env (Var v) = (mkOneOcc env v False, Var v)
\end{code}
We regard variables that occur as constructor arguments as "dangerousToDup":
\begin{verbatim}
module A where
f x = let y = expensive x in
let z = (True,y) in
(case z of {(p,q)->q}, case z of {(p,q)->q})
\end{verbatim}
We feel free to duplicate the WHNF (True,y), but that means
that y may be duplicated thereby.
If we aren't careful we duplicate the (expensive x) call!
Constructors are rather like lambdas in this way.
\begin{code}
occAnal _ expr@(Lit _) = (emptyDetails, expr)
\end{code}
\begin{code}
occAnal env (Note note@(SCC _) body)
= case occAnal env body of { (usage, body') ->
(mapVarEnv markInsideSCC usage, Note note body')
}
occAnal env (Note note body)
= case occAnal env body of { (usage, body') ->
(usage, Note note body')
}
occAnal env (Cast expr co)
= case occAnal env expr of { (usage, expr') ->
(markManyIf (isRhsEnv env) usage, Cast expr' co)
}
\end{code}
\begin{code}
occAnal env app@(App _ _)
= occAnalApp env (collectArgs app)
occAnal env (Lam x body) | isTyCoVar x
= case occAnal env body of { (body_usage, body') ->
(body_usage, Lam x body')
}
occAnal env expr@(Lam _ _)
= case occAnal env_body body of { (body_usage, body') ->
let
(final_usage, tagged_binders) = tagLamBinders body_usage binders'
really_final_usage = if linear then
final_usage
else
mapVarEnv markInsideLam final_usage
in
(really_final_usage,
mkLams tagged_binders body') }
where
env_body = vanillaCtxt (trimOccEnv env binders)
(binders, body) = collectBinders expr
binders' = oneShotGroup env binders
linear = all is_one_shot binders'
is_one_shot b = isId b && isOneShotBndr b
occAnal env (Case scrut bndr ty alts)
= case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
case mapAndUnzip occ_anal_alt alts of { (alts_usage_s, alts') ->
let
alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
(alts_usage1, tagged_bndr) = tag_case_bndr alts_usage bndr
total_usage = scrut_usage +++ alts_usage1
in
total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
where
tag_case_bndr usage bndr
= case lookupVarEnv usage bndr of
Nothing -> (usage, setIdOccInfo bndr IAmDead)
Just _ -> (usage `delVarEnv` bndr, setIdOccInfo bndr NoOccInfo)
alt_env = mkAltEnv env scrut bndr
occ_anal_alt = occAnalAlt alt_env bndr
occ_anal_scrut (Var v) (alt1 : other_alts)
| not (null other_alts) || not (isDefaultAlt alt1)
= (mkOneOcc env v True, Var v)
occ_anal_scrut scrut _alts
= occAnal (vanillaCtxt env) scrut
occAnal env (Let bind body)
= case occAnal env_body body of { (body_usage, body') ->
case occAnalBind env env_body bind body_usage of { (final_usage, new_binds) ->
(final_usage, mkLets new_binds body') }}
where
env_body = trimOccEnv env (bindersOf bind)
occAnalArgs :: OccEnv -> [CoreExpr] -> (UsageDetails, [CoreExpr])
occAnalArgs env args
= case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
(foldr (+++) emptyDetails arg_uds_s, args')}
where
arg_env = vanillaCtxt env
\end{code}
Applications are dealt with specially because we want
the "build hack" to work.
\begin{code}
occAnalApp :: OccEnv
-> (Expr CoreBndr, [Arg CoreBndr])
-> (UsageDetails, Expr CoreBndr)
occAnalApp env (Var fun, args)
= case args_stuff of { (args_uds, args') ->
let
final_args_uds = markManyIf (isRhsEnv env && is_exp) args_uds
in
(fun_uds +++ final_args_uds, mkApps (Var fun) args') }
where
fun_uniq = idUnique fun
fun_uds = mkOneOcc env fun (valArgCount args > 0)
is_exp = isExpandableApp fun (valArgCount args)
args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
| fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
| fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
| fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
| otherwise = occAnalArgs env args
occAnalApp env (fun, args)
= case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
case occAnalArgs env args of { (args_uds, args') ->
let
final_uds = fun_uds +++ args_uds
in
(final_uds, mkApps fun' args') }}
markManyIf :: Bool
-> UsageDetails
-> UsageDetails
markManyIf True uds = mapVarEnv markMany uds
markManyIf False uds = uds
appSpecial :: OccEnv
-> Int -> CtxtTy
-> [CoreExpr]
-> (UsageDetails, [CoreExpr])
appSpecial env n ctxt args
= go n args
where
arg_env = vanillaCtxt env
go _ [] = (emptyDetails, [])
go 1 (arg:args)
= case occAnal (setCtxtTy arg_env ctxt) arg of { (arg_uds, arg') ->
case occAnalArgs env args of { (args_uds, args') ->
(arg_uds +++ args_uds, arg':args') }}
go n (arg:args)
= case occAnal arg_env arg of { (arg_uds, arg') ->
case go (n1) args of { (args_uds, args') ->
(arg_uds +++ args_uds, arg':args') }}
\end{code}
Note [Binders in case alternatives]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
case x of y { (a,b) -> f y }
We treat 'a', 'b' as dead, because they don't physically occur in the
case alternative. (Indeed, a variable is dead iff it doesn't occur in
its scope in the output of OccAnal.) It really helps to know when
binders are unused. See esp the call to isDeadBinder in
Simplify.mkDupableAlt
In this example, though, the Simplifier will bring 'a' and 'b' back to
life, beause it binds 'y' to (a,b) (imagine got inlined and
scrutinised y).
\begin{code}
occAnalAlt :: OccEnv
-> CoreBndr
-> CoreAlt
-> (UsageDetails, Alt IdWithOccInfo)
occAnalAlt env case_bndr (con, bndrs, rhs)
= let
env' = trimOccEnv env bndrs
in
case occAnal env' rhs of { (rhs_usage1, rhs1) ->
let
proxies = getProxies env' case_bndr
(rhs_usage2, rhs2) = foldrBag wrapProxy (rhs_usage1, rhs1) proxies
(alt_usg, tagged_bndrs) = tagLamBinders rhs_usage2 bndrs
bndrs' = tagged_bndrs
in
(alt_usg, (con, bndrs', rhs2)) }
wrapProxy :: ProxyBind -> (UsageDetails, CoreExpr) -> (UsageDetails, CoreExpr)
wrapProxy (bndr, rhs_var, co) (body_usg, body)
| not (bndr `usedIn` body_usg)
= (body_usg, body)
| otherwise
= (body_usg' +++ rhs_usg, Let (NonRec tagged_bndr rhs) body)
where
(body_usg', tagged_bndr) = tagBinder body_usg bndr
rhs_usg = unitVarEnv rhs_var NoOccInfo
rhs = mkCoerceI co (Var rhs_var)
\end{code}
%************************************************************************
%* *
OccEnv
%* *
%************************************************************************
\begin{code}
data OccEnv
= OccEnv { occ_encl :: !OccEncl
, occ_ctxt :: !CtxtTy
, occ_proxy :: ProxyEnv
, occ_rule_fvs :: ImpRuleUsage
, occ_rule_act :: Maybe (Activation -> Bool)
}
data OccEncl
= OccRhs
| OccVanilla
instance Outputable OccEncl where
ppr OccRhs = ptext (sLit "occRhs")
ppr OccVanilla = ptext (sLit "occVanilla")
type CtxtTy = [Bool]
initOccEnv :: Maybe (Activation -> Bool) -> [CoreRule]
-> OccEnv
initOccEnv active_rule imp_rules
= OccEnv { occ_encl = OccVanilla
, occ_ctxt = []
, occ_proxy = PE emptyVarEnv emptyVarSet
, occ_rule_fvs = findImpRuleUsage active_rule imp_rules
, occ_rule_act = active_rule }
vanillaCtxt :: OccEnv -> OccEnv
vanillaCtxt env = env { occ_encl = OccVanilla, occ_ctxt = [] }
rhsCtxt :: OccEnv -> OccEnv
rhsCtxt env = env { occ_encl = OccRhs, occ_ctxt = [] }
setCtxtTy :: OccEnv -> CtxtTy -> OccEnv
setCtxtTy env ctxt = env { occ_ctxt = ctxt }
isRhsEnv :: OccEnv -> Bool
isRhsEnv (OccEnv { occ_encl = OccRhs }) = True
isRhsEnv (OccEnv { occ_encl = OccVanilla }) = False
oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
oneShotGroup (OccEnv { occ_ctxt = ctxt }) bndrs
= go ctxt bndrs []
where
go _ [] rev_bndrs = reverse rev_bndrs
go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
| isId bndr = go ctxt bndrs (bndr':rev_bndrs)
where
bndr' | lin_ctxt = setOneShotLambda bndr
| otherwise = bndr
go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
addAppCtxt env@(OccEnv { occ_ctxt = ctxt }) args
= env { occ_ctxt = replicate (valArgCount args) True ++ ctxt }
\end{code}
%************************************************************************
%* *
ImpRuleUsage
%* *
%************************************************************************
\begin{code}
type ImpRuleUsage = NameEnv UsageDetails
\end{code}
Note [ImpRuleUsage]
~~~~~~~~~~~~~~~~
Consider this, where A.g is an imported Id
f x = A.g x
{-# RULE "foo" forall x. A.g x = f x #-}
Obviously there's a loop, but the danger is that the occurrence analyser
will say that 'f' is not a loop breaker. Then the simplifier will
optimise 'f' to
f x = f x
and then gaily inline 'f'. Result infinite loop. More realistically,
these kind of rules are generated when specialising imported INLINABLE Ids.
Solution: treat an occurrence of A.g as an occurrence of all the local Ids
that occur on the RULE's RHS. This mapping from imported Id to local Ids
is held in occ_rule_fvs.
\begin{code}
findImpRuleUsage :: Maybe (Activation -> Bool) -> [CoreRule] -> ImpRuleUsage
findImpRuleUsage Nothing _ = emptyNameEnv
findImpRuleUsage (Just is_active) rules
= mkNameEnv [ (f, mapUFM (\_ -> NoOccInfo) ls)
| f <- rule_names
, let ls = find_lcl_deps f
, not (isEmptyVarSet ls) ]
where
rule_names = map ru_fn rules
rule_name_set = mkNameSet rule_names
imp_deps :: NameEnv VarSet
imp_deps = foldr add_imp emptyNameEnv rules
add_imp rule acc
| is_active (ruleActivation rule)
= extendNameEnv_C unionVarSet acc (ru_fn rule)
(exprSomeFreeVars keep_imp (ru_rhs rule))
| otherwise = acc
keep_imp v = isId v && (idName v `elemNameSet` rule_name_set)
full_imp_deps = transClosureFV (ufmToList imp_deps)
lcl_deps :: NameEnv VarSet
lcl_deps = foldr add_lcl emptyNameEnv rules
add_lcl rule acc = extendNameEnv_C unionVarSet acc (ru_fn rule)
(exprFreeIds (ru_rhs rule))
find_lcl_deps :: Name -> VarSet
find_lcl_deps f
= foldVarSet (unionVarSet . lookup_lcl . idName) (lookup_lcl f)
(lookupNameEnv full_imp_deps f `orElse` emptyVarSet)
lookup_lcl :: Name -> VarSet
lookup_lcl g = lookupNameEnv lcl_deps g `orElse` emptyVarSet
transClosureFV :: Uniquable a => [(a, VarSet)] -> UniqFM VarSet
transClosureFV fv_list
| no_change = env
| otherwise = transClosureFV new_fv_list
where
env = listToUFM fv_list
(no_change, new_fv_list) = mapAccumL bump True fv_list
bump no_change (b,fvs)
| no_change_here = (no_change, (b,fvs))
| otherwise = (False, (b,new_fvs))
where
(new_fvs, no_change_here) = extendFvs env fvs
extendFvs :: UniqFM VarSet -> VarSet -> (VarSet, Bool)
extendFvs env s
= foldVarSet add (s, True) s
where
add v (vs, no_change_so_far)
= case lookupUFM env v of
Just fvs | not (fvs `subVarSet` s)
-> (vs `unionVarSet` fvs, False)
_ -> (vs, no_change_so_far)
\end{code}
%************************************************************************
%* *
ProxyEnv
%* *
%************************************************************************
\begin{code}
data ProxyEnv
= PE (IdEnv (Id, [(Id,CoercionI)])) VarSet
\end{code}
Note [ProxyEnv]
~~~~~~~~~~~~~~~
The ProxyEnv keeps track of the connection between case binders and
scrutinee. Specifically, if
sc |-> (sc, [...(cb, co)...])
is a binding in the ProxyEnv, then
cb = sc |> coi
Typically we add such a binding when encountering the case expression
case (sc |> coi) of cb { ... }
Things to note:
* The domain of the ProxyEnv is the variable (or casted variable)
scrutinees of enclosing cases. This is additionally used
to ensure we gather occurrence info even for GlobalId scrutinees;
see Note [Binder swap for GlobalId scrutinee]
* The ProxyEnv is just an optimisation; you can throw away any
element without losing correctness. And we do so when pushing
it inside a binding (see trimProxyEnv).
* One scrutinee might map to many case binders: Eg
case sc of cb1 { DEFAULT -> ....case sc of cb2 { ... } .. }
INVARIANTS
* If sc1 |-> (sc2, [...(cb, co)...]), then sc1==sc2
It's a UniqFM and we sometimes need the domain Id
* Any particular case binder 'cb' occurs only once in entire range
* No loops
The Main Reason for having a ProxyEnv is so that when we encounter
case e of cb { pi -> ri }
we can find all the in-scope variables derivable from 'cb',
and effectively add let-bindings for them (or at least for the
ones *mentioned* in ri) thus:
case e of cb { pi -> let { x = ..cb..; y = ...cb.. }
in ri }
In this way we'll replace occurrences of 'x', 'y' with 'cb',
which implements the Binder-swap idea (see Note [Binder swap])
The function getProxies finds these bindings; then we
add just the necessary ones, using wrapProxy.
Note [Binder swap]
~~~~~~~~~~~~~~~~~~
We do these two transformations right here:
(1) case x of b { pi -> ri }
==>
case x of b { pi -> let x=b in ri }
(2) case (x |> co) of b { pi -> ri }
==>
case (x |> co) of b { pi -> let x = b |> sym co in ri }
Why (2)? See Note [Case of cast]
In both cases, in a particular alternative (pi -> ri), we only
add the binding if
(a) x occurs free in (pi -> ri)
(ie it occurs in ri, but is not bound in pi)
(b) the pi does not bind b (or the free vars of co)
We need (a) and (b) for the inserted binding to be correct.
For the alternatives where we inject the binding, we can transfer
all x's OccInfo to b. And that is the point.
Notice that
* The deliberate shadowing of 'x'.
* That (a) rapidly becomes false, so no bindings are injected.
The reason for doing these transformations here is because it allows
us to adjust the OccInfo for 'x' and 'b' as we go.
* Suppose the only occurrences of 'x' are the scrutinee and in the
ri; then this transformation makes it occur just once, and hence
get inlined right away.
* If we do this in the Simplifier, we don't know whether 'x' is used
in ri, so we are forced to pessimistically zap b's OccInfo even
though it is typically dead (ie neither it nor x appear in the
ri). There's nothing actually wrong with zapping it, except that
it's kind of nice to know which variables are dead. My nose
tells me to keep this information as robustly as possible.
The Maybe (Id,CoreExpr) passed to occAnalAlt is the extra let-binding
{x=b}; it's Nothing if the binder-swap doesn't happen.
There is a danger though. Consider
let v = x +# y
in case (f v) of w -> ...v...v...
And suppose that (f v) expands to just v. Then we'd like to
use 'w' instead of 'v' in the alternative. But it may be too
late; we may have substituted the (cheap) x+#y for v in the
same simplifier pass that reduced (f v) to v.
I think this is just too bad. CSE will recover some of it.
Note [Case of cast]
~~~~~~~~~~~~~~~~~~~
Consider case (x `cast` co) of b { I# ->
... (case (x `cast` co) of {...}) ...
We'd like to eliminate the inner case. That is the motivation for
equation (2) in Note [Binder swap]. When we get to the inner case, we
inline x, cancel the casts, and away we go.
Note [Binder swap on GlobalId scrutinees]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When the scrutinee is a GlobalId we must take care in two ways
i) In order to *know* whether 'x' occurs free in the RHS, we need its
occurrence info. BUT, we don't gather occurrence info for
GlobalIds. That's one use for the (small) occ_proxy env in OccEnv is
for: it says "gather occurrence info for these.
ii) We must call localiseId on 'x' first, in case it's a GlobalId, or
has an External Name. See, for example, SimplEnv Note [Global Ids in
the substitution].
Note [getProxies is subtle]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
The code for getProxies isn't all that obvious. Consider
case v |> cov of x { DEFAULT ->
case x |> cox1 of y { DEFAULT ->
case x |> cox2 of z { DEFAULT -> r
These will give us a ProxyEnv looking like:
x |-> (x, [(y, cox1), (z, cox2)])
v |-> (v, [(x, cov)])
From this we want to extract the bindings
x = z |> sym cox2
v = x |> sym cov
y = x |> cox1
Notice that later bindings may mention earlier ones, and that
we need to go "both ways".
Historical note [no-case-of-case]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We *used* to suppress the binder-swap in case expressions when
-fno-case-of-case is on. Old remarks:
"This happens in the first simplifier pass,
and enhances full laziness. Here's the bad case:
f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
If we eliminate the inner case, we trap it inside the I# v -> arm,
which might prevent some full laziness happening. I've seen this
in action in spectral/cichelli/Prog.hs:
[(m,n) | m <- [1..max], n <- [1..max]]
Hence the check for NoCaseOfCase."
However, now the full-laziness pass itself reverses the binder-swap, so this
check is no longer necessary.
Historical note [Suppressing the case binder-swap]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This old note describes a problem that is also fixed by doing the
binder-swap in OccAnal:
There is another situation when it might make sense to suppress the
case-expression binde-swap. If we have
case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
...other cases .... }
We'll perform the binder-swap for the outer case, giving
case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
...other cases .... }
But there is no point in doing it for the inner case, because w1 can't
be inlined anyway. Furthermore, doing the case-swapping involves
zapping w2's occurrence info (see paragraphs that follow), and that
forces us to bind w2 when doing case merging. So we get
case x of w1 { A -> let w2 = w1 in e1
B -> let w2 = w1 in e2
...other cases .... }
This is plain silly in the common case where w2 is dead.
Even so, I can't see a good way to implement this idea. I tried
not doing the binder-swap if the scrutinee was already evaluated
but that failed big-time:
data T = MkT !Int
case v of w { MkT x ->
case x of x1 { I# y1 ->
case x of x2 { I# y2 -> ...
Notice that because MkT is strict, x is marked "evaluated". But to
eliminate the last case, we must either make sure that x (as well as
x1) has unfolding MkT y1. THe straightforward thing to do is to do
the binder-swap. So this whole note is a no-op.
It's fixed by doing the binder-swap in OccAnal because we can do the
binder-swap unconditionally and still get occurrence analysis
information right.
\begin{code}
extendProxyEnv :: ProxyEnv -> Id -> CoercionI -> Id -> ProxyEnv
extendProxyEnv pe scrut co case_bndr
| scrut == case_bndr = PE env1 fvs1
| otherwise = PE env2 fvs2
where
PE env1 fvs1 = trimProxyEnv pe [case_bndr]
env2 = extendVarEnv_Acc add single env1 scrut1 (case_bndr,co)
single cb_co = (scrut1, [cb_co])
add cb_co (x, cb_cos) = (x, cb_co:cb_cos)
fvs2 = fvs1 `unionVarSet` freeVarsCoI co
`extendVarSet` case_bndr
`extendVarSet` scrut1
scrut1 = mkLocalId (localiseName (idName scrut)) (idType scrut)
type ProxyBind = (Id, Id, CoercionI)
getProxies :: OccEnv -> Id -> Bag ProxyBind
getProxies (OccEnv { occ_proxy = PE pe _ }) case_bndr
=
go_fwd case_bndr
where
fwd_pe :: IdEnv (Id, CoercionI)
fwd_pe = foldVarEnv add1 emptyVarEnv pe
where
add1 (x,ycos) env = foldr (add2 x) env ycos
add2 x (y,co) env = extendVarEnv env y (x,co)
go_fwd :: Id -> Bag ProxyBind
go_fwd case_bndr =
go_fwd' case_bndr
go_fwd' case_bndr
| Just (scrut, co) <- lookupVarEnv fwd_pe case_bndr
= unitBag (scrut, case_bndr, mkSymCoI co)
`unionBags` go_fwd scrut
`unionBags` go_bwd scrut [pr | pr@(cb,_) <- lookup_bwd scrut
, cb /= case_bndr]
| otherwise
= emptyBag
lookup_bwd :: Id -> [(Id, CoercionI)]
lookup_bwd scrut = case lookupVarEnv pe scrut of
Nothing -> []
Just (_, cb_cos) -> cb_cos
go_bwd :: Id -> [(Id, CoercionI)] -> Bag ProxyBind
go_bwd scrut cb_cos = foldr (unionBags . go_bwd1 scrut) emptyBag cb_cos
go_bwd1 :: Id -> (Id, CoercionI) -> Bag ProxyBind
go_bwd1 scrut (case_bndr, co)
=
unitBag (case_bndr, scrut, co)
`unionBags` go_bwd case_bndr (lookup_bwd case_bndr)
mkAltEnv :: OccEnv -> CoreExpr -> Id -> OccEnv
mkAltEnv env scrut cb
= env { occ_encl = OccVanilla, occ_proxy = pe' }
where
pe = occ_proxy env
pe' = case scrut of
Var v -> extendProxyEnv pe v (IdCo (idType v)) cb
Cast (Var v) co -> extendProxyEnv pe v (ACo co) cb
_other -> trimProxyEnv pe [cb]
trimOccEnv :: OccEnv -> [CoreBndr] -> OccEnv
trimOccEnv env bndrs = env { occ_proxy = trimProxyEnv (occ_proxy env) bndrs }
trimProxyEnv :: ProxyEnv -> [CoreBndr] -> ProxyEnv
trimProxyEnv (PE pe fvs) bndrs
| not (bndr_set `intersectsVarSet` fvs)
= PE pe fvs
| otherwise
= PE pe' (fvs `minusVarSet` bndr_set)
where
pe' = mapVarEnv trim pe
bndr_set = mkVarSet bndrs
trim (scrut, cb_cos) | scrut `elemVarSet` bndr_set = (scrut, [])
| otherwise = (scrut, filterOut discard cb_cos)
discard (cb,co) = bndr_set `intersectsVarSet`
extendVarSet (freeVarsCoI co) cb
freeVarsCoI :: CoercionI -> VarSet
freeVarsCoI (IdCo t) = tyVarsOfType t
freeVarsCoI (ACo co) = tyVarsOfType co
\end{code}
%************************************************************************
%* *
\subsection[OccurAnal-types]{OccEnv}
%* *
%************************************************************************
\begin{code}
type UsageDetails = IdEnv OccInfo
(+++), combineAltsUsageDetails
:: UsageDetails -> UsageDetails -> UsageDetails
(+++) usage1 usage2
= plusVarEnv_C addOccInfo usage1 usage2
combineAltsUsageDetails usage1 usage2
= plusVarEnv_C orOccInfo usage1 usage2
addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
addOneOcc usage id info
= plusVarEnv_C addOccInfo usage (unitVarEnv id info)
emptyDetails :: UsageDetails
emptyDetails = (emptyVarEnv :: UsageDetails)
usedIn :: Id -> UsageDetails -> Bool
v `usedIn` details = isExportedId v || v `elemVarEnv` details
type IdWithOccInfo = Id
tagLamBinders :: UsageDetails
-> [Id]
-> (UsageDetails,
[IdWithOccInfo])
tagLamBinders usage binders = usage' `seq` (usage', bndrs')
where
(usage', bndrs') = mapAccumR tag_lam usage binders
tag_lam usage bndr = (usage2, setBinderOcc usage bndr)
where
usage1 = usage `delVarEnv` bndr
usage2 | isId bndr = addIdOccs usage1 (idUnfoldingVars bndr)
| otherwise = usage1
tagBinder :: UsageDetails
-> Id
-> (UsageDetails,
IdWithOccInfo)
tagBinder usage binder
= let
usage' = usage `delVarEnv` binder
binder' = setBinderOcc usage binder
in
usage' `seq` (usage', binder')
setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
setBinderOcc usage bndr
| isTyCoVar bndr = bndr
| isExportedId bndr = case idOccInfo bndr of
NoOccInfo -> bndr
_ -> setIdOccInfo bndr NoOccInfo
| otherwise = setIdOccInfo bndr occ_info
where
occ_info = lookupVarEnv usage bndr `orElse` IAmDead
\end{code}
%************************************************************************
%* *
\subsection{Operations over OccInfo}
%* *
%************************************************************************
\begin{code}
mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
mkOneOcc env id int_cxt
| isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
| PE env _ <- occ_proxy env
, id `elemVarEnv` env = unitVarEnv id NoOccInfo
| Just uds <- lookupNameEnv (occ_rule_fvs env) (idName id)
= uds
| otherwise = emptyDetails
markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
markMany _ = NoOccInfo
markInsideSCC occ = markMany occ
markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
markInsideLam occ = occ
addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
addOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )
NoOccInfo
orOccInfo (OneOcc in_lam1 _ int_cxt1)
(OneOcc in_lam2 _ int_cxt2)
= OneOcc (in_lam1 || in_lam2)
False
(int_cxt1 && int_cxt2)
orOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )
NoOccInfo
\end{code}