%
% (c) The GRASP/AQUA Project, Glasgow University, 19921998
%
%************************************************************************
%* *
\section[OccurAnal]{Occurrence analysis pass}
%* *
%************************************************************************
The occurrence analyser retypechecks a core expression, returning a new
core expression with (hopefully) improved usage information.
\begin{code}
module OccurAnal (
occurAnalysePgm, occurAnalyseExpr
) where
#include "HsVersions.h"
import CoreSyn
import CoreFVs
import CoreUtils ( exprIsTrivial, isDefaultAlt )
import Coercion ( mkSymCoercion )
import Id
import Name ( localiseName )
import IdInfo
import BasicTypes
import VarSet
import VarEnv
import Maybes ( orElse )
import Digraph ( SCC(..), stronglyConnCompFromEdgedVerticesR )
import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
import Unique ( Unique )
import UniqFM ( keysUFM, intersectUFM_C, foldUFM_Directly )
import Util ( mapAndUnzip )
import Outputable
import Data.List
\end{code}
%************************************************************************
%* *
\subsection[OccurAnalmain]{Counting occurrences: main function}
%* *
%************************************************************************
Here's the externallycallable interface:
\begin{code}
occurAnalysePgm :: [CoreBind] -> [CoreBind]
occurAnalysePgm binds
= snd (go initOccEnv binds)
where
go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
go _ []
= (emptyDetails, [])
go env (bind:binds)
= (final_usage, bind' ++ binds')
where
(bs_usage, binds') = go env binds
(final_usage, bind') = occAnalBind env bind bs_usage
occurAnalyseExpr :: CoreExpr -> CoreExpr
occurAnalyseExpr expr = snd (occAnal initOccEnv expr)
\end{code}
%************************************************************************
%* *
\subsection[OccurAnalmain]{Counting occurrences: main function}
%* *
%************************************************************************
Bindings
~~~~~~~~
\begin{code}
occAnalBind :: OccEnv
-> CoreBind
-> UsageDetails
-> (UsageDetails,
[CoreBind])
occAnalBind env (NonRec binder rhs) body_usage
| isTyVar binder
= (body_usage, [NonRec binder rhs])
| not (binder `usedIn` body_usage)
= (body_usage, [])
| otherwise
= (body_usage' +++ addRuleUsage rhs_usage binder,
[NonRec tagged_binder rhs'])
where
(body_usage', tagged_binder) = tagBinder body_usage binder
(rhs_usage, rhs') = occAnalRhs env tagged_binder rhs
\end{code}
Note [Dead code]
~~~~~~~~~~~~~~~~
Dropping dead code for recursive bindings is done in a very simple way:
the entire set of bindings is dropped if none of its binders are
mentioned in its body; otherwise none are.
This seems to miss an obvious improvement.
letrec f = ...g...
g = ...f...
in
...g...
===>
letrec f = ...g...
g = ...(...g...)...
in
...g...
Now 'f' is unused! But it's OK! Dependency analysis will sort this
out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
dropped. It isn't easy to do a perfect job in one blow. Consider
letrec f = ...g...
g = ...h...
h = ...k...
k = ...m...
m = ...m...
in
...m...
Note [Loop breaking and RULES]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Loop breaking is surprisingly subtle. First read the section 4 of
"Secrets of the GHC inliner". This describes our basic plan.
However things are made quite a bit more complicated by RULES. Remember
* Note [Rules are extra RHSs]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
keeps the specialised "children" alive. If the parent dies
(because it isn't referenced any more), then the children will die
too (unless they are already referenced directly).
To that end, we build a Rec group for each cyclic strongly
connected component,
*treating f's rules as extra RHSs for 'f'*.
When we make the Rec groups we include variables free in *either*
LHS *or* RHS of the rule. The former might seems silly, but see
Note [Rule dependency info].
So in Example [eftInt], eftInt and eftIntFB will be put in the
same Rec, even though their 'main' RHSs are both nonrecursive.
* Note [Rules are visible in their own rec group]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We want the rules for 'f' to be visible in f's righthand side.
And we'd like them to be visible in other functions in f's Rec
group. E.g. in Example [Specialisation rules] we want f' rule
to be visible in both f's RHS, and fs's RHS.
This means that we must simplify the RULEs first, before looking
at any of the definitions. This is done by Simplify.simplRecBind,
when it calls addLetIdInfo.
* Note [Choosing loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We avoid infinite inlinings by choosing loop breakers, and
ensuring that a loop breaker cuts each loop. But what is a
"loop"? In particular, a RULE is like an equation for 'f' that
is *always* inlined if it is applicable. We do *not* disable
rules for loopbreakers. It's up to whoever makes the rules to
make sure that the rules themselves alwasys terminate. See Note
[Rules for recursive functions] in Simplify.lhs
Hence, if
f's RHS mentions g, and
g has a RULE that mentions h, and
h has a RULE that mentions f
then we *must* choose f to be a loop breaker. In general, take the
free variables of f's RHS, and augment it with all the variables
reachable by RULES from those starting points. That is the whole
reason for computing rule_fv_env in occAnalBind. (Of course we
only consider free vars that are also binders in this Rec group.)
Note that when we compute this rule_fv_env, we only consider variables
free in the *RHS* of the rule, in contrast to the way we build the
Rec group in the first place (Note [Rule dependency info])
Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
chosen as a loop breaker, because their RHSs don't mention each other.
And indeed both can be inlined safely.
Note that the edges of the graph we use for computing loop breakers
are not the same as the edges we use for computing the Rec blocks.
That's why we compute
rec_edges for the Rec block analysis
loop_breaker_edges for the loop breaker analysis
* 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...
}
Remmber 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.)
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 postInlineUnconditinoally
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 addRuleUsage.
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
Example [eftInt]
~~~~~~~~~~~~~~~
Example (from GHC.Enum):
eftInt :: Int# -> Int# -> [Int]
eftInt x y = ...(nonrecursive)...
eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
eftIntFB c n x y = ...(nonrecursive)...
Example [Specialisation rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this group, which is typical of what SpecConstr builds:
fs a = ....f (C a)....
f x = ....f (C 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 [Selfrecursive 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 (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)
= (ND bndr rhs' rhs_usage rhs_fvs, idUnique bndr, out_edges)
where
(rhs_usage, rhs') = occAnalRhs env bndr rhs
rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage
out_edges = keysUFM (rhs_fvs `unionVarSet` idRuleVars bndr)
occAnalRec :: SCC (Node Details) -> (UsageDetails, [CoreBind])
-> (UsageDetails, [CoreBind])
occAnalRec (AcyclicSCC (ND bndr rhs rhs_usage _, _, _)) (body_usage, binds)
| not (bndr `usedIn` body_usage)
= (body_usage, binds)
| otherwise
= (body_usage' +++ addRuleUsage rhs_usage bndr,
NonRec tagged_bndr rhs : binds)
where
(body_usage', tagged_bndr) = tagBinder body_usage bndr
occAnalRec (CyclicSCC nodes) (body_usage, binds)
| not (any (`usedIn` body_usage) bndrs)
= (body_usage, binds)
| otherwise
= (final_usage, Rec pairs : binds)
where
bndrs = [b | (ND b _ _ _, _, _) <- nodes]
bndr_set = mkVarSet bndrs
total_usage = foldl add_usage body_usage nodes
add_usage body_usage (ND bndr _ rhs_usage _, _, _)
= body_usage +++ addRuleUsage rhs_usage bndr
(final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
tag_node usage (ND bndr rhs rhs_usage rhs_fvs, k, ks)
= (usage `delVarEnv` bndr, (ND bndr2 rhs rhs_usage rhs_fvs, 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 | no_rules = reOrderCycle 0 tagged_nodes []
| otherwise = foldr (reOrderRec 0) [] $
stronglyConnCompFromEdgedVerticesR loop_breaker_edges
loop_breaker_edges = map mk_node tagged_nodes
mk_node (details@(ND _ _ _ rhs_fvs), k, _) = (details, k, new_ks)
where
new_ks = keysUFM (extendFvs rule_fv_env rhs_fvs rhs_fvs)
rule_fv_env :: IdEnv IdSet
rule_fv_env = rule_loop init_rule_fvs
no_rules = null init_rule_fvs
init_rule_fvs = [(b, rule_fvs)
| b <- bndrs
, let rule_fvs = idRuleRhsVars b `intersectVarSet` bndr_set
, not (isEmptyVarSet rule_fvs)]
rule_loop :: [(Id,IdSet)] -> IdEnv IdSet
rule_loop fv_list
| no_change = env
| otherwise = rule_loop new_fv_list
where
env = mkVarEnv init_rule_fvs
(no_change, new_fv_list) = mapAccumL bump True fv_list
bump no_change (b,fvs)
| new_fvs `subVarSet` fvs = (no_change, (b,fvs))
| otherwise = (False, (b,new_fvs `unionVarSet` fvs))
where
new_fvs = extendFvs env emptyVarSet fvs
idRuleRhsVars :: Id -> VarSet
idRuleRhsVars id = foldr (unionVarSet . ruleRhsFreeVars) emptyVarSet (idCoreRules id)
extendFvs :: IdEnv IdSet -> IdSet -> IdSet -> IdSet
extendFvs env fvs id_set
= foldUFM_Directly add fvs id_set
where
add uniq _ fvs
= case lookupVarEnv_Directly env uniq of
Just fvs' -> fvs' `unionVarSet` fvs
Nothing -> fvs
\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 noinline guys.
Furthermore, the order of the binds is such that if we neglect dependencies
on the noinline 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....
Reordering doesn't change the order of bindings, but there was no loopbreaker.
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 Id
CoreExpr
UsageDetails
IdSet
reOrderRec :: Int -> SCC (Node Details)
-> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
reOrderRec _ (AcyclicSCC (ND bndr 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 _ [bind] pairs
= (makeLoopBreaker False bndr, rhs) : pairs
where
(ND bndr rhs _ _, _, _) = bind
reOrderCycle depth (bind : binds) pairs
=
foldr (reOrderRec new_depth)
([ (makeLoopBreaker False bndr, rhs)
| (ND bndr 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 bndr rhs _ _, _, _)
| workerExists (idWorkerInfo bndr) = 10
| exprIsTrivial rhs = 5
| is_con_app rhs = 3
| inlineCandidate bndr rhs = 2
| not (neverUnfold (idUnfolding bndr)) = 1
| otherwise = 0
inlineCandidate :: Id -> CoreExpr -> Bool
inlineCandidate _ (Note InlineMe _) = True
inlineCandidate id _ = isOneOcc (idOccInfo id)
is_con_app (Var v) = isDataConWorkId 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 = setIdOccInfo bndr (IAmALoopBreaker weak)
\end{code}
Note [Complexity of loop breaking]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The loopbreaking algorithm knocks out one binder at a time, and
performs a new SCC analysis on the remaining binders. That can
behave very badly in tightlycoupled groups of bindings; in the
worst case it can be (N**2)*log N, because it does a full SCC
on N, then N1, then N2 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
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]
~~~~~~~~~~~~~~~~~~~~~
Never choose a function with an INLINE pramga as the loop breaker!
If such a function is mutuallyrecursive with a nonINLINE thing,
then the latter should be the loopbreaker.
A particular 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 {
$wfoo x = ....foo x....
foo x = ...$wfoo x...
}
The interface file sees the unfolding for $wfoo, and sees that foo is
strict (and hence it gets an autogenerated wrapper). Result: an
infinite inlining in the importing scope. So be a bit careful if you
change this. A good example is Tree.repTree in
nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
breaker then compiling Game.hs goes into an infinite loop (this
happened when we gave is_con_app a lower score than inline candidates).
Note [Constructor applications]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It's really really important to inline dictionaries. Real
example (the Enum Ordering instance from GHC.Base):
rec f = \ x -> case d of (p,q,r) -> p x
g = \ x -> case d of (p,q,r) -> q x
d = (v, f, g)
Here, f and g occur just once; but we can't inline them into d.
On the other hand we *could* simplify those case expressions if
we didn't stupidly choose d as the loop breaker.
But we won't because constructor args are marked "Many".
Inlining dictionaries is really essential to unravelling
the loops in static numeric dictionaries, see GHC.Float.
Note [Closure conversion]
~~~~~~~~~~~~~~~~~~~~~~~~~
We treat (\x. C p q) as a highscore candidate in the letrec scoring algorithm.
The immediate motivation came from the result of a closureconversion transformation
which generated code like this:
data Clo a b = forall c. Clo (c -> a -> b) c
($:) :: 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
-> Id -> CoreExpr
-> (UsageDetails, CoreExpr)
occAnalRhs env id rhs
= occAnal ctxt rhs
where
ctxt | certainly_inline id = env
| otherwise = rhsCtxt env
certainly_inline id = case idOccInfo id of
OneOcc in_lam one_br _ -> not in_lam && one_br
_ -> False
\end{code}
\begin{code}
addRuleUsage :: UsageDetails -> Id -> UsageDetails
addRuleUsage usage id
= foldVarSet add usage (idRuleVars id)
where
add v u = addOneOcc u v NoOccInfo
\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 InlineMe body)
= case occAnal env body of { (usage, body') ->
(mapVarEnv markMany usage, Note InlineMe body')
}
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') ->
(markRhsUds env True usage, Cast expr' co)
}
\end{code}
\begin{code}
occAnal env app@(App _ _)
= occAnalApp env (collectArgs app)
occAnal env (Lam x body) | isTyVar 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) = tagBinders 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 env
(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_usage' = addCaseBndrUsage alts_usage
(alts_usage1, tagged_bndr) = tagBinder alts_usage' bndr
total_usage = scrut_usage +++ alts_usage1
in
total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
where
addCaseBndrUsage usage = case lookupVarEnv usage bndr of
Nothing -> usage
Just _ -> extendVarEnv usage bndr NoOccInfo
alt_env = mkAltEnv env bndr_swap
bndr_swap = case scrut of
Var v -> Just (v, Var bndr)
Cast (Var v) co -> Just (v, Cast (Var bndr) (mkSymCoercion co))
_other -> Nothing
occ_anal_alt = occAnalAlt alt_env bndr bndr_swap
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 of { (body_usage, body') ->
case occAnalBind env bind body_usage of { (final_usage, new_binds) ->
(final_usage, mkLets new_binds body') }}
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 = markRhsUds env is_pap 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_pap = isConLikeId fun || valArgCount args < idArity fun
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') }}
markRhsUds :: OccEnv
-> Bool
-> UsageDetails
-> UsageDetails
markRhsUds env is_pap arg_uds
| isRhsEnv env && is_pap = mapVarEnv markMany arg_uds
| otherwise = arg_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 [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 letbinding
{x=b}; it's Nothing if the binderswap 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 [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 what the (small) occ_scrut_ids set 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].
Historical note [no-case-of-case]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We *used* to suppress the binder-swap in case expressoins 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 binderswap. So this whole note is a noop.
It's fixed by doing the binderswap in OccAnal because we can do the
binderswap unconditionally and still get occurrence analysis
information right.
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 [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.) This invariant is It really
helpe 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
-> Maybe (Id, CoreExpr)
-> CoreAlt
-> (UsageDetails, Alt IdWithOccInfo)
occAnalAlt env case_bndr mb_scrut_var (con, bndrs, rhs)
= case occAnal env rhs of { (rhs_usage, rhs') ->
let
(alt_usg, tagged_bndrs) = tagBinders rhs_usage bndrs
bndrs' = tagged_bndrs
in
case mb_scrut_var of
Just (scrut_var, scrut_rhs)
| scrut_var `localUsedIn` alt_usg
, not (any shadowing bndrs)
-> (addOneOcc usg_wo_scrut case_bndr NoOccInfo,
(con, bndrs', Let (NonRec scrut_var2 scrut_rhs) rhs'))
where
scrut_var1 = mkLocalId (localiseName (idName scrut_var)) (idType scrut_var)
(usg_wo_scrut, scrut_var2) = tagBinder alt_usg scrut_var1
shadowing bndr = bndr `elemVarSet` rhs_fvs
rhs_fvs = exprFreeVars scrut_rhs
_other -> (alt_usg, (con, bndrs', rhs')) }
\end{code}
%************************************************************************
%* *
\subsection[OccurAnaltypes]{OccEnv}
%* *
%************************************************************************
\begin{code}
data OccEnv
= OccEnv { occ_encl :: !OccEncl
, occ_ctxt :: !CtxtTy
, occ_scrut_ids :: !GblScrutIds }
type GblScrutIds = IdSet
data OccEncl
= OccRhs
| OccVanilla
type CtxtTy = [Bool]
initOccEnv :: OccEnv
initOccEnv = OccEnv { occ_encl = OccRhs
, occ_ctxt = []
, occ_scrut_ids = emptyVarSet }
vanillaCtxt :: OccEnv -> OccEnv
vanillaCtxt env = OccEnv { occ_encl = OccVanilla, occ_ctxt = []
, occ_scrut_ids = occ_scrut_ids env }
rhsCtxt :: OccEnv -> OccEnv
rhsCtxt env = OccEnv { occ_encl = OccRhs, occ_ctxt = []
, occ_scrut_ids = occ_scrut_ids env }
mkAltEnv :: OccEnv -> Maybe (Id, CoreExpr) -> OccEnv
mkAltEnv env (Just (scrut_id, _))
| not (isLocalId scrut_id)
= OccEnv { occ_encl = OccVanilla
, occ_scrut_ids = extendVarSet (occ_scrut_ids env) scrut_id
, occ_ctxt = occ_ctxt env }
mkAltEnv env _
| isRhsEnv env = env { occ_encl = OccVanilla }
| otherwise = env
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}
%************************************************************************
%* *
\subsection[OccurAnaltypes]{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)
localUsedIn, usedIn :: Id -> UsageDetails -> Bool
v `localUsedIn` details = v `elemVarEnv` details
v `usedIn` details = isExportedId v || v `localUsedIn` details
type IdWithOccInfo = Id
tagBinders :: UsageDetails
-> [Id]
-> (UsageDetails,
[IdWithOccInfo])
tagBinders usage binders
= let
usage' = usage `delVarEnvList` binders
uss = map (setBinderOcc usage) binders
in
usage' `seq` (usage', uss)
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
| isTyVar 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)
| id `elemVarSet` occ_scrut_ids env = unitVarEnv id NoOccInfo
| 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}