ghc-9.12.4: GHC/Core/Opt/OccurAnal.hs
{-# LANGUAGE ViewPatterns #-}
{-# OPTIONS_GHC -cpp -Wno-incomplete-record-updates #-}
{-# OPTIONS_GHC -fmax-worker-args=12 #-}
-- The -fmax-worker-args=12 is there because the main functions
-- are strict in the OccEnv, and it turned out that with the default settting
-- some functions would unbox the OccEnv ad some would not, depending on how
-- many /other/ arguments the function has. Inconsistent unboxing is very
-- bad for performance, so I increased the limit to allow it to unbox
-- consistently.
-- AK: Seems we no longer unbox OccEnv now anyway so it might be redundant.
{-
(c) The GRASP/AQUA Project, Glasgow University, 1992-1998
************************************************************************
* *
\section[OccurAnal]{Occurrence analysis pass}
* *
************************************************************************
The occurrence analyser re-typechecks a core expression, returning a new
core expression with (hopefully) improved usage information.
-}
module GHC.Core.Opt.OccurAnal (
occurAnalysePgm,
occurAnalyseExpr,
zapLambdaBndrs, BinderSwapDecision(..), scrutOkForBinderSwap
) where
import GHC.Prelude hiding ( head, init, last, tail )
import GHC.Core
import GHC.Core.FVs
import GHC.Core.Utils ( exprIsTrivial, isDefaultAlt, isExpandableApp,
mkCastMCo, mkTicks )
import GHC.Core.Opt.Arity ( joinRhsArity, isOneShotBndr )
import GHC.Core.Coercion
import GHC.Core.Predicate ( isDictId )
import GHC.Core.Type
import GHC.Core.TyCo.FVs ( tyCoVarsOfMCo )
import GHC.Data.Maybe( orElse )
import GHC.Data.Graph.Directed ( SCC(..), Node(..)
, stronglyConnCompFromEdgedVerticesUniq
, stronglyConnCompFromEdgedVerticesUniqR )
import GHC.Types.Unique
import GHC.Types.Unique.FM
import GHC.Types.Unique.Set
import GHC.Types.Id
import GHC.Types.Id.Info
import GHC.Types.Basic
import GHC.Types.Tickish
import GHC.Types.Var.Set
import GHC.Types.Var.Env
import GHC.Types.Var
import GHC.Types.Demand ( argOneShots, argsOneShots, isDeadEndSig )
import GHC.Utils.Outputable
import GHC.Utils.Panic
import GHC.Utils.Misc
import GHC.Builtin.Names( runRWKey )
import GHC.Unit.Module( Module )
import Data.List (mapAccumL)
{-
************************************************************************
* *
occurAnalysePgm, occurAnalyseExpr
* *
************************************************************************
Here's the externally-callable interface:
-}
-- | Do occurrence analysis, and discard occurrence info returned
occurAnalyseExpr :: CoreExpr -> CoreExpr
occurAnalyseExpr expr = expr'
where
WUD _ expr' = occAnal initOccEnv expr
occurAnalysePgm :: Module -- Used only in debug output
-> (Id -> Bool) -- Active unfoldings
-> (Activation -> Bool) -- Active rules
-> [CoreRule] -- Local rules for imported Ids
-> CoreProgram -> CoreProgram
occurAnalysePgm this_mod active_unf active_rule imp_rules binds
| isEmptyDetails final_usage
= occ_anald_binds
| otherwise -- See Note [Glomming]
= warnPprTrace True "Glomming in" (hang (ppr this_mod <> colon) 2 (ppr final_usage))
occ_anald_glommed_binds
where
init_env = initOccEnv { occ_rule_act = active_rule
, occ_unf_act = active_unf }
WUD final_usage occ_anald_binds = go binds init_env
WUD _ occ_anald_glommed_binds = occAnalRecBind init_env TopLevel
imp_rule_edges
(flattenBinds binds)
initial_uds
-- It's crucial to re-analyse the glommed-together bindings
-- so that we establish the right loop breakers. Otherwise
-- we can easily create an infinite loop (#9583 is an example)
--
-- Also crucial to re-analyse the /original/ bindings
-- in case the first pass accidentally discarded as dead code
-- a binding that was actually needed (albeit before its
-- definition site). #17724 threw this up.
initial_uds = addManyOccs emptyDetails (rulesFreeVars imp_rules)
-- The RULES declarations keep things alive!
-- imp_rule_edges maps a top-level local binder 'f' to the
-- RHS free vars of any IMP-RULE, a local RULE for an imported function,
-- where 'f' appears on the LHS
-- e.g. RULE foldr f = blah
-- imp_rule_edges contains f :-> fvs(blah)
-- We treat such RULES as extra rules for 'f'
-- See Note [Preventing loops due to imported functions rules]
imp_rule_edges :: ImpRuleEdges
imp_rule_edges = foldr (plusVarEnv_C (++)) emptyVarEnv
[ mapVarEnv (const [(act,rhs_fvs)]) $ getUniqSet $
exprsFreeIds args `delVarSetList` bndrs
| Rule { ru_act = act, ru_bndrs = bndrs
, ru_args = args, ru_rhs = rhs } <- imp_rules
-- Not BuiltinRules; see Note [Plugin rules]
, let rhs_fvs = exprFreeIds rhs `delVarSetList` bndrs ]
go :: [CoreBind] -> OccEnv -> WithUsageDetails [CoreBind]
go [] _ = WUD initial_uds []
go (bind:binds) env = occAnalBind env TopLevel
imp_rule_edges bind (go binds) (++)
{- *********************************************************************
* *
IMP-RULES
Local rules for imported functions
* *
********************************************************************* -}
type ImpRuleEdges = IdEnv [(Activation, VarSet)]
-- Mapping from a local Id 'f' to info about its IMP-RULES,
-- i.e. /local/ rules for an imported Id that mention 'f' on the LHS
-- We record (a) its Activation and (b) the RHS free vars
-- See Note [IMP-RULES: local rules for imported functions]
noImpRuleEdges :: ImpRuleEdges
noImpRuleEdges = emptyVarEnv
lookupImpRules :: ImpRuleEdges -> Id -> [(Activation,VarSet)]
lookupImpRules imp_rule_edges bndr
= case lookupVarEnv imp_rule_edges bndr of
Nothing -> []
Just vs -> vs
impRulesScopeUsage :: [(Activation,VarSet)] -> UsageDetails
-- Variable mentioned in RHS of an IMP-RULE for the bndr,
-- whether active or not
impRulesScopeUsage imp_rules_info
= foldr add emptyDetails imp_rules_info
where
add (_,vs) usage = addManyOccs usage vs
impRulesActiveFvs :: (Activation -> Bool) -> VarSet
-> [(Activation,VarSet)] -> VarSet
impRulesActiveFvs is_active bndr_set vs
= foldr add emptyVarSet vs `intersectVarSet` bndr_set
where
add (act,vs) acc | is_active act = vs `unionVarSet` acc
| otherwise = acc
{- Note [IMP-RULES: local rules for imported functions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We quite often have
* A /local/ rule
* for an /imported/ function
like this:
foo x = blah
{-# RULE "map/foo" forall xs. map foo xs = xs #-}
We call them IMP-RULES. They are important in practice, and occur a
lot in the libraries.
IMP-RULES are held in mg_rules of ModGuts, and passed in to
occurAnalysePgm.
Main Invariant:
* Throughout, we treat an IMP-RULE that mentions 'f' on its LHS
just like a RULE for f.
Note [IMP-RULES: unavoidable loops]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this
f = /\a. B.g a
RULE B.g Int = 1 + f Int
Note that
* The RULE is for an imported function.
* f is non-recursive
Now we
can get
f Int --> B.g Int Inlining f
--> 1 + f Int Firing RULE
and so the simplifier goes into an infinite loop. This
would not happen if the RULE was for a local function,
because we keep track of dependencies through rules. But
that is pretty much impossible to do for imported Ids. Suppose
f's definition had been
f = /\a. C.h a
where (by some long and devious process), C.h eventually inlines to
B.g. We could only spot such loops by exhaustively following
unfoldings of C.h etc, in case we reach B.g, and hence (via the RULE)
f.
We regard this potential infinite loop as a *programmer* error.
It's up the programmer not to write silly rules like
RULE f x = f x
and the example above is just a more complicated version.
Note [Specialising imported functions] (referred to from Specialise)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
For *automatically-generated* rules, the programmer can't be
responsible for the "programmer error" in Note [IMP-RULES: unavoidable
loops]. In particular, consider specialising a recursive function
defined in another module. If we specialise a recursive function B.g,
we get
g_spec = .....(B.g Int).....
RULE B.g Int = g_spec
Here, g_spec doesn't look recursive, but when the rule fires, it
becomes so. And if B.g was mutually recursive, the loop might not be
as obvious as it is here.
To avoid this,
* When specialising a function that is a loop breaker,
give a NOINLINE pragma to the specialised function
Note [Preventing loops due to imported functions rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider:
import GHC.Base (foldr)
{-# RULES "filterList" forall p. foldr (filterFB (:) p) [] = filter p #-}
filter p xs = build (\c n -> foldr (filterFB c p) n xs)
filterFB c p = ...
f = filter p xs
Note that filter is not a loop-breaker, so what happens is:
f = filter p xs
= {inline} build (\c n -> foldr (filterFB c p) n xs)
= {inline} foldr (filterFB (:) p) [] xs
= {RULE} filter p xs
We are in an infinite loop.
A more elaborate example (that I actually saw in practice when I went to
mark GHC.List.filter as INLINABLE) is as follows. Say I have this module:
{-# LANGUAGE RankNTypes #-}
module GHCList where
import Prelude hiding (filter)
import GHC.Base (build)
{-# INLINABLE filter #-}
filter :: (a -> Bool) -> [a] -> [a]
filter p [] = []
filter p (x:xs) = if p x then x : filter p xs else filter p xs
{-# NOINLINE [0] filterFB #-}
filterFB :: (a -> b -> b) -> (a -> Bool) -> a -> b -> b
filterFB c p x r | p x = x `c` r
| otherwise = r
{-# RULES
"filter" [~1] forall p xs. filter p xs = build (\c n -> foldr
(filterFB c p) n xs)
"filterList" [1] forall p. foldr (filterFB (:) p) [] = filter p
#-}
Then (because RULES are applied inside INLINABLE unfoldings, but inlinings
are not), the unfolding given to "filter" in the interface file will be:
filter p [] = []
filter p (x:xs) = if p x then x : build (\c n -> foldr (filterFB c p) n xs)
else build (\c n -> foldr (filterFB c p) n xs
Note that because this unfolding does not mention "filter", filter is not
marked as a strong loop breaker. Therefore at a use site in another module:
filter p xs
= {inline}
case xs of [] -> []
(x:xs) -> if p x then x : build (\c n -> foldr (filterFB c p) n xs)
else build (\c n -> foldr (filterFB c p) n xs)
build (\c n -> foldr (filterFB c p) n xs)
= {inline} foldr (filterFB (:) p) [] xs
= {RULE} filter p xs
And we are in an infinite loop again, except that this time the loop is producing an
infinitely large *term* (an unrolling of filter) and so the simplifier finally
dies with "ticks exhausted"
SOLUTION: we treat the rule "filterList" as an extra rule for 'filterFB'
because it mentions 'filterFB' on the LHS. This is the Main Invariant
in Note [IMP-RULES: local rules for imported functions].
So, during loop-breaker analysis:
- for each active RULE for a local function 'f' we add an edge between
'f' and the local FVs of the rule RHS
- for each active RULE for an *imported* function we add dependency
edges between the *local* FVS of the rule LHS and the *local* FVS of
the rule RHS.
Even with this extra hack we aren't always going to get things
right. For example, it might be that the rule LHS mentions an imported
Id, and another module has a RULE that can rewrite that imported Id to
one of our local Ids.
Note [Plugin rules]
~~~~~~~~~~~~~~~~~~~
Conal Elliott (#11651) built a GHC plugin that added some
BuiltinRules (for imported Ids) to the mg_rules field of ModGuts, to
do some domain-specific transformations that could not be expressed
with an ordinary pattern-matching CoreRule. But then we can't extract
the dependencies (in imp_rule_edges) from ru_rhs etc, because a
BuiltinRule doesn't have any of that stuff.
So we simply assume that BuiltinRules have no dependencies, and filter
them out from the imp_rule_edges comprehension.
Note [Glomming]
~~~~~~~~~~~~~~~
RULES for imported Ids can make something at the top refer to
something at the bottom:
foo = ...(B.f @Int)...
$sf = blah
RULE: B.f @Int = $sf
Applying this rule makes foo refer to $sf, although foo doesn't appear to
depend on $sf. (And, as in Note [IMP-RULES: local rules for imported functions], the
dependency might be more indirect. For example, foo might mention C.t
rather than B.f, where C.t eventually inlines to B.f.)
NOTICE that this cannot happen for rules whose head is a
locally-defined function, because we accurately track dependencies
through RULES. It only happens for rules whose head is an imported
function (B.f in the example above).
Solution:
- When simplifying, bring all top level identifiers into
scope at the start, ignoring the Rec/NonRec structure, so
that when 'h' pops up in f's rhs, we find it in the in-scope set
(as the simplifier generally expects). This happens in simplTopBinds.
- In the occurrence analyser, if there are any out-of-scope
occurrences that pop out of the top, which will happen after
firing the rule: f = \x -> h x
h = \y -> 3
then just glom all the bindings into a single Rec, so that
the *next* iteration of the occurrence analyser will sort
them all out. This part happens in occurAnalysePgm.
This is a legitimate situation where the need for glomming doesn't
point to any problems. However, when GHC is compiled with -DDEBUG, we
produce a warning addressed to the GHC developers just in case we
require glomming due to an out-of-order reference that is caused by
some earlier transformation stage misbehaving.
-}
{-
************************************************************************
* *
Bindings
* *
************************************************************************
Note [Recursive bindings: the grand plan]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Loop breaking is surprisingly subtle. First read the section 4 of
"Secrets of the GHC inliner". This describes our basic plan. We
avoid infinite inlinings by choosing loop breakers, and ensuring that
a loop breaker cuts each loop.
See also Note [Inlining and hs-boot files] in GHC.Core.ToIface, which
deals with a closely related source of infinite loops.
When we come across a binding group
Rec { x1 = r1; ...; xn = rn }
we treat it like this (occAnalRecBind):
1. Note [Forming Rec groups]
Occurrence-analyse each right hand side, and build a
"Details" for each binding to capture the results.
Wrap the details in a LetrecNode, ready for SCC analysis.
All this is done by makeNode.
The edges of this graph are the "scope edges".
2. Do SCC-analysis on these Nodes:
- Each CyclicSCC will become a new Rec
- Each AcyclicSCC will become a new NonRec
The key property is that every free variable of a binding is
accounted for by the scope edges, so that when we are done
everything is still in scope.
3. For each AcyclicSCC, just make a NonRec binding.
4. For each CyclicSCC of the scope-edge SCC-analysis in (2), we
identify suitable loop-breakers to ensure that inlining terminates.
This is done by occAnalRec.
To do so, form the loop-breaker graph, do SCC analysis. For each
CyclicSCC we choose a loop breaker, delete all edges to that node,
re-analyse the SCC, and iterate. See Note [Choosing loop breakers]
for the details
Note [Dead code]
~~~~~~~~~~~~~~~~
Dropping dead code for a cyclic Strongly Connected Component is done
in a very simple way:
the entire SCC is dropped if none of its binders are mentioned
in the body; otherwise the whole thing is kept.
The key observation is that dead code elimination happens after
dependency analysis: so 'occAnalBind' processes SCCs instead of the
original term's binding groups.
Thus 'occAnalBind' does indeed drop 'f' in an example like
letrec f = ...g...
g = ...(...g...)...
in
...g...
when 'g' no longer uses 'f' at all (eg 'f' does not occur in a RULE in
'g'). 'occAnalBind' first consumes 'CyclicSCC g' and then it consumes
'AcyclicSCC f', where 'body_usage' won't contain 'f'.
Note [Forming Rec groups]
~~~~~~~~~~~~~~~~~~~~~~~~~
The key point about the "Forming Rec groups" step is that it /preserves
scoping/. If 'x' is mentioned, it had better be bound somewhere. So if
we start with
Rec { f = ...h...
; g = ...f...
; h = ...f... }
we can split into SCCs
Rec { f = ...h...
; h = ..f... }
NonRec { g = ...f... }
We put bindings {f = ef; g = eg } in a Rec group if "f uses g" and "g
uses f", no matter how indirectly. We do a SCC analysis with an edge
f -> g if "f mentions g". That is, g is free in:
a) the rhs 'ef'
b) or the RHS of a rule for f, whether active or inactive
Note [Rules are extra RHSs]
c) or the LHS or a rule for f, whether active or inactive
Note [Rule dependency info]
d) the RHS of an /active/ local IMP-RULE
Note [IMP-RULES: local rules for imported functions]
(b) and (c) apply regardless of the activation of the RULE, because even if
the rule is inactive its free variables must be bound. But (d) doesn't need
to worry about this because IMP-RULES are always notionally at the bottom
of the file.
* 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).
So in Example [eftInt], eftInt and eftIntFB will be put in the
same Rec, even though their 'main' RHSs are both non-recursive.
We must also include inactive rules, so that their free vars
remain in scope.
* Note [Rule dependency info]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
The VarSet in a RuleInfo 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 = v+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 RULE appears to mention 'x'
as well as 'v'
* Note [Rules are visible in their own rec group]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We want the rules for 'f' to be visible in f's right-hand side.
And we'd like them to be visible in other functions in f's Rec
group. E.g. in Note [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 [TailUsageDetails when forming Rec groups]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The `TailUsageDetails` stored in the `nd_uds` field of a `NodeDetails` is
computed by `occAnalLamTail` applied to the RHS, not `occAnalExpr`.
That is because the binding might still become a *non-recursive join point* in
the AcyclicSCC case of dependency analysis!
Hence we do the delayed `adjustTailUsage` in `occAnalRec`/`tagRecBinders` to get
a regular, adjusted UsageDetails.
See Note [Join points and unfoldings/rules] for more details on the contract.
Note [Stable unfoldings]
~~~~~~~~~~~~~~~~~~~~~~~~
None of the above stuff about RULES applies to a stable unfolding
stored in a CoreUnfolding. The unfolding, if any, is simplified
at the same time as the regular RHS of the function (ie *not* like
Note [Rules are visible in their own rec group]), so it should be
treated *exactly* like an extra RHS.
Or, rather, when computing loop-breaker edges,
* If f has an INLINE pragma, and it is active, we treat the
INLINE rhs as f's rhs
* If it's inactive, we treat f as having no rhs
* If it has no INLINE pragma, we look at f's actual 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
#-}
Note [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 [Rules for recursive functions] in GHC.Core.Opt.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.
Hence the transitive rule_fv_env stuff described in
Note [Rules and loop breakers].
Note [Occurrence analysis for join points]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider these two somewhat artificial programs (#22404)
Program (P1) Program (P2)
------------------------------ -------------------------------------
let v = <small thunk> in let v = <small thunk> in
join j = case v of (a,b) -> a
in case x of in case x of
A -> case v of (a,b) -> a A -> j
B -> case v of (a,b) -> a B -> j
C -> case v of (a,b) -> b C -> case v of (a,b) -> b
D -> [] D -> []
In (P1), `v` gets allocated, as a thunk, every time this code is executed. But
notice that `v` occurs at most once in any case branch; the occurrence analyser
spots this and returns a OneOcc{ occ_n_br = 3 } for `v`. Then the code in
GHC.Core.Opt.Simplify.Utils.postInlineUnconditionally inlines `v` at its three
use sites, and discards the let-binding. That way, we avoid allocating `v` in
the A,B,C branches (though we still compute it of course), and branch D
doesn't involve <small thunk> at all. This sometimes makes a Really Big
Difference.
In (P2) we have shared the common RHS of A, B, in a join point `j`. We would
like to inline `v` in just the same way as in (P1). But the usual strategy
for let bindings is conservative and uses `andUDs` to combine usage from j's
RHS to its body; as if `j` was called on every code path (once, albeit). In
the case of (P2), we'll get ManyOccs for `v`. Important optimisation lost!
Solving this problem makes the Simplifier less fragile. For example,
the Simplifier might inline `j`, and convert (P2) into (P1)... or it might
not, depending in a perhaps-fragile way on the size of the join point.
I was motivated to implement this feature of the occurrence analyser
when trying to make optimisation join points simpler and more robust
(see e.g. #23627).
The occurrence analyser therefore has clever code that behaves just as
if you inlined `j` at all its call sites. Here is a tricky variant
to keep in mind:
Program (P3)
-------------------------------
join j = case v of (a,b) -> a
in case f v of
A -> j
B -> j
C -> []
If you mentally inline `j` you'll see that `v` is used twice on the path
through A, so it should have ManyOcc. Bear this case in mind!
* We treat /non-recursive/ join points specially. Recursive join points are
treated like any other letrec, as before. Moreover, we only give this special
treatment to /pre-existing/ non-recursive join points, not the ones that we
discover for the first time in this sweep of the occurrence analyser.
* In occ_env, the new (occ_join_points :: IdEnv OccInfoEnv) maps
each in-scope non-recursive join point, such as `j` above, to
a "zeroed form" of its RHS's usage details. The "zeroed form"
* deletes ManyOccs
* maps a OneOcc to OneOcc{ occ_n_br = 0 }
In our example, occ_join_points will be extended with
[j :-> [v :-> OneOcc{occ_n_br=0}]]
See addJoinPoint.
* At an occurrence of a join point, we do everything as normal, but add in the
UsageDetails from the occ_join_points. See mkOneOcc.
* Crucially, at the NonRec binding of the join point, in `occAnalBind`, we use
`orUDs`, not `andUDs` to combine the usage from the RHS with the usage from
the body.
Here are the consequences
* Because of the perhaps-surprising OneOcc{occ_n_br=0} idea of the zeroed
form, the occ_n_br field of a OneOcc binder still counts the number of
/actual lexical occurrences/ of the variable. In Program P2, for example,
`v` will end up with OneOcc{occ_n_br=2}, not occ_n_br=3.
There are two lexical occurrences of `v`!
(NB: `orUDs` adds occ_n_br together, so occ_n_br=1 is impossible, too.)
* In the tricky (P3) we'll get an `andUDs` of
* OneOcc{occ_n_br=0} from the occurrences of `j`)
* OneOcc{occ_n_br=1} from the (f v)
These are `andUDs` together in `addOccInfo`, and hence
`v` gets ManyOccs, just as it should. Clever!
There are a couple of tricky wrinkles
(W1) Consider this example which shadows `j`:
join j = rhs in
in case x of { K j -> ..j..; ... }
Clearly when we come to the pattern `K j` we must drop the `j`
entry in occ_join_points.
This is done by `drop_shadowed_joins` in `addInScope`.
(W2) Consider this example which shadows `v`:
join j = ...v...
in case x of { K v -> ..j..; ... }
We can't make j's occurrences in the K alternative give rise to an
occurrence of `v` (via occ_join_points), because it'll just be deleted by
the `K v` pattern. Yikes. This is rare because shadowing is rare, but
it definitely can happen. Solution: when bringing `v` into scope at
the `K v` pattern, chuck out of occ_join_points any elements whose
UsageDetails mentions `v`. Instead, just `andUDs` all that usage in
right here.
This requires work in two places.
* In `preprocess_env`, we detect if the newly-bound variables intersect
the free vars of occ_join_points. (These free vars are conveniently
simply the domain of the OccInfoEnv for that join point.) If so,
we zap the entire occ_join_points.
* In `postprcess_uds`, we add the chucked-out join points to the
returned UsageDetails, with `andUDs`.
(W3) Consider this example, which shadows `j`, but this time in an argument
join j = rhs
in f (case x of { K j -> ...; ... })
We can zap the entire occ_join_points when looking at the argument,
because `j` can't posibly occur -- it's a join point! And the smaller
occ_join_points is, the better. Smaller to look up in mkOneOcc, and
more important, less looking-up when checking (W2).
This is done in setNonTailCtxt. It's important /not/ to do this for
join-point RHS's because of course `j` can occur there!
NB: this is just about efficiency: it is always safe /not/ to zap the
occ_join_points.
(W4) What if the join point binding has a stable unfolding, or RULES?
They are just alternative right-hand sides, and at each call site we
will use only one of them. So again, we can use `orUDs` to combine
usage info from all these alternatives RHSs.
Wrinkles (W1) and (W2) are very similar to Note [Binder swap] (BS3).
Note [Finding join points]
~~~~~~~~~~~~~~~~~~~~~~~~~~
It's the occurrence analyser's job to find bindings that we can turn into join
points, but it doesn't perform that transformation right away. Rather, it marks
the eligible bindings as part of their occurrence data, leaving it to the
simplifier (or to simpleOptPgm) to actually change the binder's 'IdDetails'.
The simplifier then eta-expands the RHS if needed and then updates the
occurrence sites. Dividing the work this way means that the occurrence analyser
still only takes one pass, yet one can always tell the difference between a
function call and a jump by looking at the occurrence (because the same pass
changes the 'IdDetails' and propagates the binders to their occurrence sites).
To track potential join points, we use the 'occ_tail' field of OccInfo. A value
of `AlwaysTailCalled n` indicates that every occurrence of the variable is a
tail call with `n` arguments (counting both value and type arguments). Otherwise
'occ_tail' will be 'NoTailCallInfo'. The tail call info flows bottom-up with the
rest of 'OccInfo' until it goes on the binder.
Note [Join arity prediction based on joinRhsArity]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In general, the join arity from tail occurrences of a join point (O) may be
higher or lower than the manifest join arity of the join body (M). E.g.,
-- M > O:
let f x y = x + y -- M = 2
in if b then f 1 else f 2 -- O = 1
==> { Contify for join arity 1 }
join f x = \y -> x + y
in if b then jump f 1 else jump f 2
-- M < O
let f = id -- M = 0
in if ... then f 12 else f 13 -- O = 1
==> { Contify for join arity 1, eta-expand f }
join f x = id x
in if b then jump f 12 else jump f 13
But for *recursive* let, it is crucial that both arities match up, consider
letrec f x y = if ... then f x else True
in f 42
Here, M=2 but O=1. If we settled for a joinrec arity of 1, the recursive jump
would not happen in a tail context! Contification is invalid here.
So indeed it is crucial to demand that M=O.
(Side note: Actually, we could be more specific: Let O1 be the join arity of
occurrences from the letrec RHS and O2 the join arity from the let body. Then
we need M=O1 and M<=O2 and could simply eta-expand the RHS to match O2 later.
M=O is the specific case where we don't want to eta-expand. Neither the join
points paper nor GHC does this at the moment.)
We can capitalise on this observation and conclude that *if* f could become a
joinrec (without eta-expansion), it will have join arity M.
Now, M is just the result of 'joinRhsArity', a rather simple, local analysis.
It is also the join arity inside the 'TailUsageDetails' returned by
'occAnalLamTail', so we can predict join arity without doing any fixed-point
iteration or really doing any deep traversal of let body or RHS at all.
We check for M in the 'adjustTailUsage' call inside 'tagRecBinders'.
All this is quite apparent if you look at the contification transformation in
Fig. 5 of "Compiling without Continuations" (which does not account for
eta-expansion at all, mind you). The letrec case looks like this
letrec f = /\as.\xs. L[us] in L'[es]
... and a bunch of conditions establishing that f only occurs
in app heads of join arity (len as + len xs) inside us and es ...
The syntactic form `/\as.\xs. L[us]` forces M=O iff `f` occurs in `us`. However,
for non-recursive functions, this is the definition of contification from the
paper:
let f = /\as.\xs.u in L[es] ... conditions ...
Note that u could be a lambda itself, as we have seen. No relationship between M
and O to exploit here.
Note [Join points and unfoldings/rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
let j2 y = blah
let j x = j2 (x+x)
{-# INLINE [2] j #-}
in case e of { A -> j 1; B -> ...; C -> j 2 }
Before j is inlined, we'll have occurrences of j2 in
both j's RHS and in its stable unfolding. We want to discover
j2 as a join point. So 'occAnalUnfolding' returns an unadjusted
'TailUsageDetails', like 'occAnalLamTail'. We adjust the usage details of the
unfolding to the actual join arity using the same 'adjustTailArity' as for
the RHS, see Note [Adjusting right-hand sides].
Same with rules. Suppose we have:
let j :: Int -> Int
j y = 2 * y
let k :: Int -> Int -> Int
{-# RULES "SPEC k 0" k 0 y = j y #-}
k x y = x + 2 * y
in case e of { A -> k 1 2; B -> k 3 5; C -> blah }
We identify k as a join point, and we want j to be a join point too.
Without the RULE it would be, and we don't want the RULE to mess it
up. So provided the join-point arity of k matches the args of the
rule we can allow the tail-call info from the RHS of the rule to
propagate.
* Note that the join arity of the RHS and that of the unfolding or RULE might
mismatch:
let j x y = j2 (x+x)
{-# INLINE[2] j = \x. g #-}
{-# RULE forall x y z. j x y z = h 17 #-}
in j 1 2
So it is crucial that we adjust each TailUsageDetails individually
with the actual join arity 2 here before we combine with `andUDs`.
Here, that means losing tail call info on `g` and `h`.
* Wrinkle for Rec case: We store one TailUsageDetails in the node Details for
RHS, unfolding and RULE combined. Clearly, if they don't agree on their join
arity, we have to do some adjusting. We choose to adjust to the join arity
of the RHS, because that is likely the join arity that the join point will
have; see Note [Join arity prediction based on joinRhsArity].
If the guess is correct, then tail calls in the RHS are preserved; a necessary
condition for the whole binding becoming a joinrec.
The guess can only be incorrect in the 'AcyclicSCC' case when the binding
becomes a non-recursive join point with a different join arity. But then the
eventual call to 'adjustTailUsage' in 'tagRecBinders'/'occAnalRec' will
be with a different join arity and destroy unsound tail call info with
'markNonTail'.
* Wrinkle for RULES. Suppose the example was a bit different:
let j :: Int -> Int
j y = 2 * y
k :: Int -> Int -> Int
{-# RULES "SPEC k 0" k 0 = j #-}
k x y = x + 2 * y
in ...
If we eta-expanded the rule all would be well, but as it stands the
one arg of the rule don't match the join-point arity of 2.
Conceivably we could notice that a potential join point would have
an "undersaturated" rule and account for it. This would mean we
could make something that's been specialised a join point, for
instance. But local bindings are rarely specialised, and being
overly cautious about rules only costs us anything when, for some `j`:
* Before specialisation, `j` has non-tail calls, so it can't be a join point.
* During specialisation, `j` gets specialised and thus acquires rules.
* Sometime afterward, the non-tail calls to `j` disappear (as dead code, say),
and so now `j` *could* become a join point.
This appears to be very rare in practice. TODO Perhaps we should gather
statistics to be sure.
------------------------------------------------------------
Note [Adjusting right-hand sides]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
There's a bit of a dance we need to do after analysing a lambda expression or
a right-hand side. In particular, we need to
a) call 'markAllNonTail' *unless* the binding is for a join point, and
the TailUsageDetails from the RHS has the right join arity; e.g.
join j x y = case ... of
A -> j2 p
B -> j2 q
in j a b
Here we want the tail calls to j2 to be tail calls of the whole expression
b) call 'markAllInsideLam' *unless* the binding is for a thunk, a one-shot
lambda, or a non-recursive join point
Some examples, with how the free occurrences in e (assumed not to be a value
lambda) get marked:
inside lam non-tail-called
------------------------------------------------------------
let x = e No Yes
let f = \x -> e Yes Yes
let f = \x{OneShot} -> e No Yes
\x -> e Yes Yes
join j x = e No No
joinrec j x = e Yes No
There are a few other caveats; most importantly, if we're marking a binding as
'AlwaysTailCalled', it's *going* to be a join point, so we treat it as one so
that the effect cascades properly. Consequently, at the time the RHS is
analysed, we won't know what adjustments to make; thus 'occAnalLamTail' must
return the unadjusted 'TailUsageDetails', to be adjusted by 'adjustTailUsage'
once join-point-hood has been decided and eventual one-shot annotations have
been added through 'markNonRecJoinOneShots'.
It is not so simple to see that 'occAnalNonRecBind' and 'occAnalRecBind' indeed
perform a similar sequence of steps. Thus, here is an interleaving of events
of both functions, serving as a specification:
1. Call 'occAnalLamTail' to find usage information for the RHS.
Recursive case: 'makeNode'
Non-recursive case: 'occAnalNonRecBind'
2. (Analyse the binding's scope. Done in 'occAnalBind'/`occAnal Let{}`.
Same whether recursive or not.)
3. Call 'tagNonRecBinder' or 'tagRecBinders', which decides whether to make
the binding a join point.
Cyclic Recursive case: 'mkLoopBreakerNodes'
Acyclic Recursive case: `occAnalRec AcyclicSCC{}`
Non-recursive case: 'occAnalNonRecBind'
4. Non-recursive join point: Call 'markNonRecJoinOneShots' so that e.g.,
FloatOut sees one-shot annotations on lambdas
Acyclic Recursive case: `occAnalRec AcyclicSCC{}` calls 'adjustNonRecRhs'
Non-recursive case: 'occAnalNonRecBind' calls 'adjustNonRecRhs'
5. Call 'adjustTailUsage' accordingly.
Cyclic Recursive case: 'tagRecBinders'
Acyclic Recursive case: 'adjustNonRecRhs'
Non-recursive case: 'adjustNonRecRhs'
-}
------------------------------------------------------------------
-- occAnalBind
------------------------------------------------------------------
occAnalBind
:: OccEnv
-> TopLevelFlag
-> ImpRuleEdges
-> CoreBind
-> (OccEnv -> WithUsageDetails r) -- Scope of the bind
-> ([CoreBind] -> r -> r) -- How to combine the scope with new binds
-> WithUsageDetails r -- Of the whole let(rec)
-- AK: While not allocating any less inlining occAnalBind turns calls to the
-- passed functions into known calls with all the benefits that brings.
-- On a version of T26425 with 6k alternatives this improved compile
-- by 10-20% with -O.
{-# INLINE occAnalBind #-}
occAnalBind env lvl ire (Rec pairs) thing_inside combine
= addInScopeList env (map fst pairs) $ \env ->
let WUD body_uds body' = thing_inside env
WUD bind_uds binds' = occAnalRecBind env lvl ire pairs body_uds
in WUD bind_uds (combine binds' body')
occAnalBind !env lvl ire (NonRec bndr rhs) thing_inside combine
| isTyVar bndr -- A type let; we don't gather usage info
= let !(WUD body_uds res) = addInScopeOne env bndr thing_inside
in WUD body_uds (combine [NonRec bndr rhs] res)
-- /Existing/ non-recursive join points
-- See Note [Occurrence analysis for join points]
| mb_join@(JoinPoint {}) <- idJoinPointHood bndr
= -- Analyse the RHS and /then/ the body
let -- Analyse the rhs first, generating rhs_uds
!(rhs_uds_s, bndr', rhs') = occAnalNonRecRhs env lvl ire mb_join bndr rhs
rhs_uds = foldl1' orUDs rhs_uds_s -- NB: orUDs. See (W4) of
-- Note [Occurrence analysis for join points]
-- Now analyse the body, adding the join point
-- into the environment with addJoinPoint
!(WUD body_uds (occ, body)) = occAnalNonRecBody env bndr' $ \env ->
thing_inside (addJoinPoint env bndr' rhs_uds)
in
if isDeadOcc occ -- Drop dead code; see Note [Dead code]
then WUD body_uds body
else WUD (rhs_uds `orUDs` body_uds) -- Note `orUDs`
(combine [NonRec (fst (tagNonRecBinder lvl occ bndr')) rhs']
body)
-- The normal case, including newly-discovered join points
-- Analyse the body and /then/ the RHS
| WUD body_uds (occ,body) <- occAnalNonRecBody env bndr thing_inside
= if isDeadOcc occ -- Drop dead code; see Note [Dead code]
then WUD body_uds body
else let
-- Get the join info from the *new* decision; NB: bndr is not already a JoinId
-- See Note [Join points and unfoldings/rules]
-- => join arity O of Note [Join arity prediction based on joinRhsArity]
(tagged_bndr, mb_join) = tagNonRecBinder lvl occ bndr
!(rhs_uds_s, final_bndr, rhs') = occAnalNonRecRhs env lvl ire mb_join tagged_bndr rhs
in WUD (foldr andUDs body_uds rhs_uds_s) -- Note `andUDs`
(combine [NonRec final_bndr rhs'] body)
-----------------
occAnalNonRecBody :: OccEnv -> Id
-> (OccEnv -> WithUsageDetails r) -- Scope of the bind
-> (WithUsageDetails (OccInfo, r))
occAnalNonRecBody env bndr thing_inside
= addInScopeOne env bndr $ \env ->
let !(WUD inner_uds res) = thing_inside env
!occ = lookupLetOccInfo inner_uds bndr
in WUD inner_uds (occ, res)
-----------------
occAnalNonRecRhs :: OccEnv -> TopLevelFlag -> ImpRuleEdges
-> JoinPointHood -> Id -> CoreExpr
-> ([UsageDetails], Id, CoreExpr)
occAnalNonRecRhs !env lvl imp_rule_edges mb_join bndr rhs
| null rules, null imp_rule_infos
= -- Fast path for common case of no rules. This is only worth
-- 0.1% perf on average, but it's also only a line or two of code
( [adj_rhs_uds, adj_unf_uds], final_bndr_no_rules, final_rhs )
| otherwise
= (adj_rhs_uds : adj_unf_uds : adj_rule_uds, final_bndr_with_rules, final_rhs )
where
--------- Right hand side ---------
-- For join points, set occ_encl to OccVanilla, via setTailCtxt. If we have
-- join j = Just (f x) in ...
-- we do not want to float the (f x) to
-- let y = f x in join j = Just y in ...
-- That's that OccRhs would do; but there's no point because
-- j will never be scrutinised.
rhs_env = mkRhsOccEnv env NonRecursive rhs_ctxt mb_join bndr rhs
rhs_ctxt = mkNonRecRhsCtxt lvl bndr unf
-- See Note [Join arity prediction based on joinRhsArity]
-- Match join arity O from mb_join_arity with manifest join arity M as
-- returned by of occAnalLamTail. It's totally OK for them to mismatch;
-- hence adjust the UDs from the RHS
WUD adj_rhs_uds final_rhs = adjustNonRecRhs mb_join $
occAnalLamTail rhs_env rhs
final_bndr_with_rules
| noBinderSwaps env = bndr -- See Note [Unfoldings and rules]
| otherwise = bndr `setIdSpecialisation` mkRuleInfo rules'
`setIdUnfolding` unf1
final_bndr_no_rules
| noBinderSwaps env = bndr -- See Note [Unfoldings and rules]
| otherwise = bndr `setIdUnfolding` unf1
--------- Unfolding ---------
-- See Note [Join points and unfoldings/rules]
unf = idUnfolding bndr
WTUD unf_tuds unf1 = occAnalUnfolding rhs_env unf
adj_unf_uds = adjustTailArity mb_join unf_tuds
--------- Rules ---------
-- See Note [Rules are extra RHSs] and Note [Rule dependency info]
-- and Note [Join points and unfoldings/rules]
rules = idCoreRules bndr
rules_w_uds = map (occAnalRule rhs_env) rules
rules' = map fstOf3 rules_w_uds
imp_rule_infos = lookupImpRules imp_rule_edges bndr
imp_rule_uds = [impRulesScopeUsage imp_rule_infos]
-- imp_rule_uds: consider
-- h = ...
-- g = ...
-- RULE map g = h
-- Then we want to ensure that h is in scope everywhere
-- that g is (since the RULE might turn g into h), so
-- we make g mention h.
adj_rule_uds :: [UsageDetails]
adj_rule_uds = imp_rule_uds ++
[ l `andUDs` adjustTailArity mb_join r
| (_,l,r) <- rules_w_uds ]
mkNonRecRhsCtxt :: TopLevelFlag -> Id -> Unfolding -> OccEncl
-- Precondition: Id is not a join point
mkNonRecRhsCtxt lvl bndr unf
| certainly_inline = OccVanilla -- See Note [Cascading inlines]
| otherwise = OccRhs
where
certainly_inline -- See Note [Cascading inlines]
= -- mkNonRecRhsCtxt is only used for non-join points, so occAnalBind
-- has set the OccInfo for this binder before calling occAnalNonRecRhs
case idOccInfo bndr of
OneOcc { occ_in_lam = NotInsideLam, occ_n_br = 1 }
-> active && not stable_unf && not top_bottoming
_ -> False
active = isAlwaysActive (idInlineActivation bndr)
stable_unf = isStableUnfolding unf
top_bottoming = isTopLevel lvl && isDeadEndId bndr
-----------------
occAnalRecBind :: OccEnv -> TopLevelFlag -> ImpRuleEdges -> [(Var,CoreExpr)]
-> UsageDetails -> WithUsageDetails [CoreBind]
-- For a recursive group, we
-- * occ-analyse all the RHSs
-- * compute strongly-connected components
-- * feed those components to occAnalRec
-- See Note [Recursive bindings: the grand plan]
occAnalRecBind !rhs_env lvl imp_rule_edges pairs body_usage
= foldr (occAnalRec rhs_env lvl) (WUD body_usage []) sccs
where
sccs :: [SCC NodeDetails]
sccs = stronglyConnCompFromEdgedVerticesUniq nodes
nodes :: [LetrecNode]
nodes = map (makeNode rhs_env imp_rule_edges bndr_set) pairs
bndrs = map fst pairs
bndr_set = mkVarSet bndrs
-----------------------------
occAnalRec :: OccEnv -> TopLevelFlag
-> SCC NodeDetails
-> WithUsageDetails [CoreBind]
-> WithUsageDetails [CoreBind]
-- The NonRec case is just like a Let (NonRec ...) above
occAnalRec !_ lvl
(AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = wtuds }))
(WUD body_uds binds)
| isDeadOcc occ -- Check for dead code: see Note [Dead code]
= WUD body_uds binds
| otherwise
= let (bndr', mb_join) = tagNonRecBinder lvl occ bndr
!(WUD rhs_uds' rhs') = adjustNonRecRhs mb_join wtuds
in WUD (body_uds `andUDs` rhs_uds')
(NonRec bndr' rhs' : binds)
where
occ = lookupLetOccInfo body_uds bndr
-- The Rec case is the interesting one
-- See Note [Recursive bindings: the grand plan]
-- See Note [Loop breaking]
occAnalRec env lvl (CyclicSCC details_s) (WUD body_uds binds)
| not (any needed details_s)
= -- Check for dead code: see Note [Dead code]
-- NB: Only look at body_uds, ignoring uses in the SCC
WUD body_uds binds
| otherwise
= WUD final_uds (Rec pairs : binds)
where
all_simple = all nd_simple details_s
needed :: NodeDetails -> Bool
needed (ND { nd_bndr = bndr }) = isExportedId bndr || bndr `elemVarEnv` body_env
body_env = ud_env body_uds
------------------------------
-- Make the nodes for the loop-breaker analysis
-- See Note [Choosing loop breakers] for loop_breaker_nodes
final_uds :: UsageDetails
loop_breaker_nodes :: [LoopBreakerNode]
WUD final_uds loop_breaker_nodes = mkLoopBreakerNodes env lvl body_uds details_s
------------------------------
weak_fvs :: VarSet
weak_fvs = mapUnionVarSet nd_weak_fvs details_s
---------------------------
-- Now reconstruct the cycle
pairs :: [(Id,CoreExpr)]
pairs | all_simple = reOrderNodes 0 weak_fvs loop_breaker_nodes []
| otherwise = loopBreakNodes 0 weak_fvs loop_breaker_nodes []
-- In the common case when all are "simple" (no rules at all)
-- the loop_breaker_nodes will include all the scope edges
-- so a SCC computation would yield a single CyclicSCC result;
-- and reOrderNodes deals with exactly that case.
-- Saves a SCC analysis in a common case
{- *********************************************************************
* *
Loop breaking
* *
********************************************************************* -}
{- Note [Choosing loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In Step 4 in Note [Recursive bindings: the grand plan]), occAnalRec does
loop-breaking on each CyclicSCC of the original program:
* mkLoopBreakerNodes: Form the loop-breaker graph for that CyclicSCC
* loopBreakNodes: Do SCC analysis on it
* reOrderNodes: For each CyclicSCC, pick a loop breaker
* Delete edges to that loop breaker
* Do another SCC analysis on that reduced SCC
* Repeat
To form the loop-breaker graph, we construct a new set of Nodes, the
"loop-breaker nodes", with the same details but different edges, the
"loop-breaker edges". The loop-breaker nodes have both more and fewer
dependencies than the scope edges:
More edges:
If f calls g, and g has an active rule that mentions h then
we add an edge from f -> h. See Note [Rules and loop breakers].
Fewer edges: we only include dependencies
* only on /active/ rules,
* on rule /RHSs/ (not LHSs)
The scope edges, by contrast, must be much more inclusive.
The nd_simple flag tracks the common case when a binding has no RULES
at all, in which case the loop-breaker edges will be identical to the
scope edges.
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 [inl_fvs]
~~~~~~~~~~~~~~
Note that the loop-breaker graph includes edges for occurrences in
/both/ the RHS /and/ the stable unfolding. Consider this, which actually
occurred when compiling BooleanFormula.hs in GHC:
Rec { lvl1 = go
; lvl2[StableUnf = go] = lvl1
; go = ...go...lvl2... }
From the point of view of infinite inlining, we need only these edges:
lvl1 :-> go
lvl2 :-> go -- The RHS lvl1 will never be used for inlining
go :-> go, lvl2
But the danger is that, lacking any edge to lvl1, we'll put it at the
end thus
Rec { lvl2[ StableUnf = go] = lvl1
; go[LoopBreaker] = ...go...lvl2... }
; lvl1[Occ=Once] = go }
And now the Simplifer will try to use PreInlineUnconditionally on lvl1
(which occurs just once), but because it is last we won't actually
substitute in lvl2. Sigh.
To avoid this possibility, we include edges from lvl2 to /both/ its
stable unfolding /and/ its RHS. Hence the defn of inl_fvs in
makeNode. Maybe we could be more clever, but it's very much a corner
case.
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 chosen as a loop breaker.) Why not? Because then we
drop the binding for 'g', which leaves it out of scope in the
RULE!
Here's a somewhat different example of the same thing
Rec { q = r
; r = ...p...
; p = p_rhs
RULE p [] = q }
Here the RULE is "below" q, but we *still* can't postInlineUnconditionally
q, because the RULE for p is active throughout. So the RHS of r
might rewrite to r = ...q...
So q must remain in scope in the output program!
We "solve" this by:
Make q a "weak" loop breaker (OccInfo = IAmLoopBreaker True)
iff q is a mentioned in the RHS of any RULE (active on not)
in the Rec group
Note the "active or not" comment; even if a RULE is inactive, we
want its RHS free vars to stay alive (#20820)!
A normal "strong" loop breaker has IAmLoopBreaker False. So:
Inline postInlineUnconditionally
strong IAmLoopBreaker False no no
weak IAmLoopBreaker True yes no
other yes yes
The **sole** reason for this kind of loop breaker is so that
postInlineUnconditionally does not fire. Ugh.
Annoyingly, since we simplify the rules *first* we'll never inline
q into p's RULE. That trivial binding for q will hang around until
we discard the rule. Yuk. But it's rare.
Note [Rules and loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When we form the loop-breaker graph (Step 4 in Note [Recursive
bindings: the grand plan]), we must be careful about RULEs.
For a start, we want a loop breaker to cut every cycle, so inactive
rules play no part; we need only consider /active/ rules.
See Note [Finding rule RHS free vars]
The second point is more subtle. A RULE is like an equation for
'f' that is *always* inlined if it is applicable. We do *not* disable
rules for loop-breakers. It's up to whoever makes the rules to make
sure that the rules themselves always terminate. See Note [Rules for
recursive functions] in GHC.Core.Opt.Simplify
Hence, if
f's RHS (or its stable unfolding if it has one) 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. Example: see Note
[Specialisation rules]. So our plan is this:
Take the free variables of f's RHS, and augment it with all the
variables reachable by a transitive sequence RULES from those
starting points.
That is the whole reason for computing rule_fv_env in mkLoopBreakerNodes.
Wrinkles:
* We only consider /active/ rules. See Note [Finding rule RHS free vars]
* We need only consider free vars that are also binders in this Rec
group. See also Note [Finding rule RHS free vars]
* 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])
* Why "transitive sequence of rules"? Because active rules apply
unconditionally, without checking loop-breaker-ness.
See Note [Loop breaker dependencies].
Note [Finding rule RHS free vars]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this real example from Data Parallel Haskell
tagZero :: Array Int -> Array Tag
{-# INLINE [1] tagZeroes #-}
tagZero xs = pmap (\x -> fromBool (x==0)) xs
{-# RULES "tagZero" [~1] forall xs n.
pmap fromBool <blah blah> = 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 occ_rule_act field
of the OccEnv.
Note [loopBreakNodes]
~~~~~~~~~~~~~~~~~~~~~
loopBreakNodes is applied to the list of nodes for a cyclic strongly
connected component (there's guaranteed to be a cycle). It returns
the same nodes, 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.
-}
type Binding = (Id,CoreExpr)
-- See Note [loopBreakNodes]
loopBreakNodes :: Int
-> VarSet -- Binders whose dependencies may be "missing"
-- See Note [Weak loop breakers]
-> [LoopBreakerNode]
-> [Binding] -- Append these to the end
-> [Binding]
-- Return the bindings sorted into a plausible order, and marked with loop breakers.
-- See Note [loopBreakNodes]
loopBreakNodes depth weak_fvs nodes binds
= -- pprTrace "loopBreakNodes" (ppr nodes) $
go (stronglyConnCompFromEdgedVerticesUniqR nodes)
where
go [] = binds
go (scc:sccs) = loop_break_scc scc (go sccs)
loop_break_scc scc binds
= case scc of
AcyclicSCC node -> nodeBinding (mk_non_loop_breaker weak_fvs) node : binds
CyclicSCC nodes -> reOrderNodes depth weak_fvs nodes binds
----------------------------------
reOrderNodes :: Int -> VarSet -> [LoopBreakerNode] -> [Binding] -> [Binding]
-- Choose a loop breaker, mark it no-inline,
-- and call loopBreakNodes on the rest
reOrderNodes _ _ [] _ = panic "reOrderNodes"
reOrderNodes _ _ [node] binds = nodeBinding mk_loop_breaker node : binds
reOrderNodes depth weak_fvs (node : nodes) binds
= -- pprTrace "reOrderNodes" (vcat [ text "unchosen" <+> ppr unchosen
-- , text "chosen" <+> ppr chosen_nodes ]) $
loopBreakNodes new_depth weak_fvs unchosen $
(map (nodeBinding mk_loop_breaker) chosen_nodes ++ binds)
where
(chosen_nodes, unchosen) = chooseLoopBreaker approximate_lb
(snd_score (node_payload node))
[node] [] nodes
approximate_lb = depth >= 2
new_depth | approximate_lb = 0
| otherwise = depth+1
-- After two iterations (d=0, d=1) give up
-- and approximate, returning to d=0
nodeBinding :: (Id -> Id) -> LoopBreakerNode -> Binding
nodeBinding set_id_occ (node_payload -> SND { snd_bndr = bndr, snd_rhs = rhs})
= (set_id_occ bndr, rhs)
mk_loop_breaker :: Id -> Id
mk_loop_breaker bndr
= bndr `setIdOccInfo` occ'
where
occ' = strongLoopBreaker { occ_tail = tail_info }
tail_info = tailCallInfo (idOccInfo bndr)
mk_non_loop_breaker :: VarSet -> Id -> Id
-- See Note [Weak loop breakers]
mk_non_loop_breaker weak_fvs bndr
| bndr `elemVarSet` weak_fvs = setIdOccInfo bndr occ'
| otherwise = bndr
where
occ' = weakLoopBreaker { occ_tail = tail_info }
tail_info = tailCallInfo (idOccInfo bndr)
----------------------------------
chooseLoopBreaker :: Bool -- True <=> Too many iterations,
-- so approximate
-> NodeScore -- Best score so far
-> [LoopBreakerNode] -- Nodes with this score
-> [LoopBreakerNode] -- Nodes with higher scores
-> [LoopBreakerNode] -- Unprocessed nodes
-> ([LoopBreakerNode], [LoopBreakerNode])
-- This loop looks for the bind with the lowest score
-- to pick as the loop breaker. The rest accumulate in
chooseLoopBreaker _ _ loop_nodes acc []
= (loop_nodes, acc) -- Done
-- If approximate_loop_breaker is True, we pick *all*
-- nodes with lowest score, else just one
-- See Note [Complexity of loop breaking]
chooseLoopBreaker approx_lb loop_sc loop_nodes acc (node : nodes)
| approx_lb
, rank sc == rank loop_sc
= chooseLoopBreaker approx_lb loop_sc (node : loop_nodes) acc nodes
| sc `betterLB` loop_sc -- Better score so pick this new one
= chooseLoopBreaker approx_lb sc [node] (loop_nodes ++ acc) nodes
| otherwise -- Worse score so don't pick it
= chooseLoopBreaker approx_lb loop_sc loop_nodes (node : acc) nodes
where
sc = snd_score (node_payload node)
{-
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 #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 breakers is
linear in the number of instance declarations.
Note [Loop breakers and INLINE/INLINABLE 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.
It's vital to distinguish between INLINE and INLINABLE (the
Bool returned by hasStableCoreUnfolding_maybe). If we start with
Rec { {-# INLINABLE f #-}
f x = ...f... }
and then worker/wrapper it through strictness analysis, we'll get
Rec { {-# INLINABLE $wf #-}
$wf p q = let x = (p,q) in ...f...
{-# INLINE f #-}
f x = case x of (p,q) -> $wf p q }
Now it is vital that we choose $wf as the loop breaker, so we can
inline 'f' in '$wf'.
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 GHC.Tc.TyCl.Instance
We give DFuns a higher score than ordinary CONLIKE things because
if there's a choice we want the DFun to be the non-loop breaker. 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.
@occAnalUnfolding@ 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 this
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 loopBreakSCC.
[June 98, SLPJ] I've undone this change; I don't understand it. See notes with loopBreakSCC.
************************************************************************
* *
Making nodes
* *
************************************************************************
-}
-- | Digraph node as constructed by 'makeNode' and consumed by 'occAnalRec'.
-- The Unique key is gotten from the Id.
type LetrecNode = Node Unique NodeDetails
-- | Node details as consumed by 'occAnalRec'.
data NodeDetails
= ND { nd_bndr :: Id -- Binder
, nd_rhs :: !(WithTailUsageDetails CoreExpr)
-- ^ RHS, already occ-analysed
-- With TailUsageDetails from RHS, and RULES, and stable unfoldings,
-- ignoring phase (ie assuming all are active).
-- NB: Unadjusted TailUsageDetails, as if this Node becomes a
-- non-recursive join point!
-- See Note [TailUsageDetails when forming Rec groups]
, nd_inl :: IdSet -- Free variables of the stable unfolding and the RHS
-- but excluding any RULES
-- This is the IdSet that may be used if the Id is inlined
, nd_simple :: Bool -- True iff this binding has no local RULES
-- If all nodes are simple we don't need a loop-breaker
-- dep-anal before reconstructing.
, nd_weak_fvs :: IdSet -- Variables bound in this Rec group that are free
-- in the RHS of any rule (active or not) for this bndr
-- See Note [Weak loop breakers]
, nd_active_rule_fvs :: IdSet -- Variables bound in this Rec group that are free
-- in the RHS of an active rule for this bndr
-- See Note [Rules and loop breakers]
}
instance Outputable NodeDetails where
ppr nd = text "ND" <> braces
(sep [ text "bndr =" <+> ppr (nd_bndr nd)
, text "uds =" <+> ppr uds
, text "inl =" <+> ppr (nd_inl nd)
, text "simple =" <+> ppr (nd_simple nd)
, text "active_rule_fvs =" <+> ppr (nd_active_rule_fvs nd)
])
where
WTUD uds _ = nd_rhs nd
-- | Digraph with simplified and completely occurrence analysed
-- 'SimpleNodeDetails', retaining just the info we need for breaking loops.
type LoopBreakerNode = Node Unique SimpleNodeDetails
-- | Condensed variant of 'NodeDetails' needed during loop breaking.
data SimpleNodeDetails
= SND { snd_bndr :: IdWithOccInfo -- OccInfo accurate
, snd_rhs :: CoreExpr -- properly occur-analysed
, snd_score :: NodeScore
}
instance Outputable SimpleNodeDetails where
ppr nd = text "SND" <> braces
(sep [ text "bndr =" <+> ppr (snd_bndr nd)
, text "score =" <+> ppr (snd_score nd)
])
-- The NodeScore is compared lexicographically;
-- e.g. lower rank wins regardless of size
type NodeScore = ( Int -- Rank: lower => more likely to be picked as loop breaker
, Int -- Size of rhs: higher => more likely to be picked as LB
-- Maxes out at maxExprSize; we just use it to prioritise
-- small functions
, Bool ) -- Was it a loop breaker before?
-- True => more likely to be picked
-- Note [Loop breakers, node scoring, and stability]
rank :: NodeScore -> Int
rank (r, _, _) = r
makeNode :: OccEnv -> ImpRuleEdges -> VarSet
-> (Var, CoreExpr) -> LetrecNode
-- See Note [Recursive bindings: the grand plan]
makeNode !env imp_rule_edges bndr_set (bndr, rhs)
= -- pprTrace "makeNode" (ppr bndr <+> ppr (sizeVarSet bndr_set)) $
DigraphNode { node_payload = details
, node_key = varUnique bndr
, node_dependencies = nonDetKeysUniqSet scope_fvs }
-- It's OK to use nonDetKeysUniqSet here as stronglyConnCompFromEdgedVerticesR
-- is still deterministic with edges in nondeterministic order as
-- explained in Note [Deterministic SCC] in GHC.Data.Graph.Directed.
where
details = ND { nd_bndr = bndr'
, nd_rhs = WTUD (TUD rhs_ja unadj_scope_uds) rhs'
, nd_inl = inl_fvs
, nd_simple = null rules_w_uds && null imp_rule_info
, nd_weak_fvs = weak_fvs
, nd_active_rule_fvs = active_rule_fvs }
bndr' | noBinderSwaps env = bndr -- See Note [Unfoldings and rules]
| otherwise = bndr `setIdUnfolding` unf'
`setIdSpecialisation` mkRuleInfo rules'
-- NB: Both adj_unf_uds and adj_rule_uds have been adjusted to match the
-- JoinArity rhs_ja of unadj_rhs_uds.
unadj_inl_uds = unadj_rhs_uds `andUDs` adj_unf_uds
unadj_scope_uds = unadj_inl_uds `andUDs` adj_rule_uds
-- Note [Rules are extra RHSs]
-- Note [Rule dependency info]
scope_fvs = udFreeVars bndr_set unadj_scope_uds
-- scope_fvs: all occurrences from this binder: RHS, unfolding,
-- and RULES, both LHS and RHS thereof, active or inactive
inl_fvs = udFreeVars bndr_set unadj_inl_uds
-- inl_fvs: vars that would become free if the function was inlined.
-- We conservatively approximate that by the free vars from the RHS
-- and the unfolding together.
-- See Note [inl_fvs]
--------- Right hand side ---------
-- Constructing the edges for the main Rec computation
-- See Note [Forming Rec groups]
-- and Note [TailUsageDetails when forming Rec groups]
-- Compared to occAnalNonRecBind, we can't yet adjust the RHS because
-- (a) we don't yet know the final joinpointhood. It might not become a
-- join point after all!
-- (b) we don't even know whether it stays a recursive RHS after the SCC
-- analysis we are about to seed! So we can't markAllInsideLam in
-- advance, because if it ends up as a non-recursive join point we'll
-- consider it as one-shot and don't need to markAllInsideLam.
-- Instead, do the occAnalLamTail call here and postpone adjustTailUsage
-- until occAnalRec. In effect, we pretend that the RHS becomes a
-- non-recursive join point and fix up later with adjustTailUsage.
rhs_env = mkRhsOccEnv env Recursive OccRhs (idJoinPointHood bndr) bndr rhs
-- If bndr isn't an /existing/ join point (so idJoinPointHood = NotJoinPoint),
-- it's safe to zap the occ_join_points, because they can't occur in RHS.
WTUD (TUD rhs_ja unadj_rhs_uds) rhs' = occAnalLamTail rhs_env rhs
-- The corresponding call to adjustTailUsage is in occAnalRec and tagRecBinders
--------- Unfolding ---------
-- See Note [Join points and unfoldings/rules]
unf = realIdUnfolding bndr -- realIdUnfolding: Ignore loop-breaker-ness
-- here because that is what we are setting!
WTUD unf_tuds unf' = occAnalUnfolding rhs_env unf
adj_unf_uds = adjustTailArity (JoinPoint rhs_ja) unf_tuds
-- `rhs_ja` is `joinRhsArity rhs` and is the prediction for source M
-- of Note [Join arity prediction based on joinRhsArity]
--------- IMP-RULES --------
is_active = occ_rule_act env :: Activation -> Bool
imp_rule_info = lookupImpRules imp_rule_edges bndr
imp_rule_uds = impRulesScopeUsage imp_rule_info
imp_rule_fvs = impRulesActiveFvs is_active bndr_set imp_rule_info
--------- All rules --------
-- See Note [Join points and unfoldings/rules]
-- `rhs_ja` is `joinRhsArity rhs'` and is the prediction for source M
-- of Note [Join arity prediction based on joinRhsArity]
rules_w_uds :: [(CoreRule, UsageDetails, UsageDetails)]
rules_w_uds = [ (r,l,adjustTailArity (JoinPoint rhs_ja) rhs_wuds)
| rule <- idCoreRules bndr
, let (r,l,rhs_wuds) = occAnalRule rhs_env rule ]
rules' = map fstOf3 rules_w_uds
adj_rule_uds = foldr add_rule_uds imp_rule_uds rules_w_uds
add_rule_uds (_, l, r) uds = l `andUDs` r `andUDs` uds
-------- active_rule_fvs ------------
active_rule_fvs = foldr add_active_rule imp_rule_fvs rules_w_uds
add_active_rule (rule, _, rhs_uds) fvs
| is_active (ruleActivation rule)
= udFreeVars bndr_set rhs_uds `unionVarSet` fvs
| otherwise
= fvs
-------- weak_fvs ------------
-- See Note [Weak loop breakers]
weak_fvs = foldr add_rule emptyVarSet rules_w_uds
add_rule (_, _, rhs_uds) fvs = udFreeVars bndr_set rhs_uds `unionVarSet` fvs
mkLoopBreakerNodes :: OccEnv -> TopLevelFlag
-> UsageDetails -- for BODY of let
-> [NodeDetails]
-> WithUsageDetails [LoopBreakerNode] -- with OccInfo up-to-date
-- See Note [Choosing loop breakers]
-- This function primarily creates the Nodes for the
-- loop-breaker SCC analysis. More specifically:
-- a) tag each binder with its occurrence info
-- b) add a NodeScore to each node
-- c) make a Node with the right dependency edges for
-- the loop-breaker SCC analysis
-- d) adjust each RHS's usage details according to
-- the binder's (new) shotness and join-point-hood
mkLoopBreakerNodes !env lvl body_uds details_s
= WUD final_uds (zipWithEqual "mkLoopBreakerNodes" mk_lb_node details_s bndrs')
where
WUD final_uds bndrs' = tagRecBinders lvl body_uds details_s
mk_lb_node nd@(ND { nd_bndr = old_bndr, nd_inl = inl_fvs
, nd_rhs = WTUD _ rhs }) new_bndr
= DigraphNode { node_payload = simple_nd
, node_key = varUnique old_bndr
, node_dependencies = nonDetKeysUniqSet lb_deps }
-- It's OK to use nonDetKeysUniqSet here as
-- stronglyConnCompFromEdgedVerticesR is still deterministic with edges
-- in nondeterministic order as explained in
-- Note [Deterministic SCC] in GHC.Data.Graph.Directed.
where
simple_nd = SND { snd_bndr = new_bndr, snd_rhs = rhs, snd_score = score }
score = nodeScore env new_bndr lb_deps nd
lb_deps = extendFvs_ rule_fv_env inl_fvs
-- See Note [Loop breaker dependencies]
rule_fv_env :: IdEnv IdSet
-- Maps a variable f to the variables from this group
-- reachable by a sequence of RULES starting with f
-- Domain is *subset* of bound vars (others have no rule fvs)
-- See Note [Finding rule RHS free vars]
-- Why transClosureFV? See Note [Loop breaker dependencies]
rule_fv_env = transClosureFV $ mkVarEnv $
[ (b, rule_fvs)
| ND { nd_bndr = b, nd_active_rule_fvs = rule_fvs } <- details_s
, not (isEmptyVarSet rule_fvs) ]
{- Note [Loop breaker dependencies]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The loop breaker dependencies of x in a recursive
group { f1 = e1; ...; fn = en } are:
- The "inline free variables" of f: the fi free in
f's stable unfolding and RHS; see Note [inl_fvs]
- Any fi reachable from those inline free variables by a sequence
of RULE rewrites. Remember, rule rewriting is not affected
by fi being a loop breaker, so we have to take the transitive
closure in case f is the only possible loop breaker in the loop.
Hence rule_fv_env. We need only account for /active/ rules.
-}
------------------------------------------
nodeScore :: OccEnv
-> Id -- Binder with new occ-info
-> VarSet -- Loop-breaker dependencies
-> NodeDetails
-> NodeScore
nodeScore !env new_bndr lb_deps
(ND { nd_bndr = old_bndr, nd_rhs = WTUD _ bind_rhs })
| not (isId old_bndr) -- A type or coercion variable is never a loop breaker
= (100, 0, False)
| old_bndr `elemVarSet` lb_deps -- Self-recursive things are great loop breakers
= (0, 0, True) -- See Note [Self-recursion and loop breakers]
| not (occ_unf_act env old_bndr) -- A binder whose inlining is inactive (e.g. has
= (0, 0, True) -- a NOINLINE pragma) makes a great loop breaker
| exprIsTrivial rhs
= mk_score 10 -- Practically certain to be inlined
-- Used to have also: && not (isExportedId bndr)
-- But I found this sometimes cost an extra iteration when we have
-- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
-- where df is the exported dictionary. Then df makes a really
-- bad choice for loop breaker
| DFunUnfolding { df_args = args } <- old_unf
-- Never choose a DFun as a loop breaker
-- Note [DFuns should not be loop breakers]
= (9, length args, is_lb)
-- Data structures are more important than INLINE pragmas
-- so that dictionary/method recursion unravels
| CoreUnfolding { uf_guidance = UnfWhen {} } <- old_unf
= mk_score 6
| is_con_app rhs -- Data types help with cases:
= mk_score 5 -- Note [Constructor applications]
| isStableUnfolding old_unf
, can_unfold
= mk_score 3
| isOneOcc (idOccInfo new_bndr)
= mk_score 2 -- Likely to be inlined
| can_unfold -- The Id has some kind of unfolding
= mk_score 1
| otherwise
= (0, 0, is_lb)
where
mk_score :: Int -> NodeScore
mk_score rank = (rank, rhs_size, is_lb)
-- is_lb: see Note [Loop breakers, node scoring, and stability]
is_lb = isStrongLoopBreaker (idOccInfo old_bndr)
old_unf = realIdUnfolding old_bndr
can_unfold = canUnfold old_unf
rhs = case old_unf of
CoreUnfolding { uf_src = src, uf_tmpl = unf_rhs }
| isStableSource src
-> unf_rhs
_ -> bind_rhs
-- 'bind_rhs' is irrelevant for inlining things with a stable unfolding
rhs_size = case old_unf of
CoreUnfolding { uf_guidance = guidance }
| UnfIfGoodArgs { ug_size = size } <- guidance
-> size
_ -> cheapExprSize rhs
-- Checking for a constructor application
-- Cheap and cheerful; the simplifier moves casts out of the way
-- The lambda case is important to spot x = /\a. C (f a)
-- which comes up when C is a dictionary constructor and
-- f is a default method.
-- Example: the instance for Show (ST s a) in GHC.ST
--
-- However we *also* treat (\x. C p q) as a con-app-like thing,
-- Note [Closure conversion]
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 (Tick _ e) = is_con_app e
is_con_app (Let _ e) = is_con_app e -- let x = let y = blah in (a,b)
is_con_app _ = False -- We will float the y out, so treat
-- the x-binding as a con-app (#20941)
maxExprSize :: Int
maxExprSize = 20 -- Rather arbitrary
cheapExprSize :: CoreExpr -> Int
-- Maxes out at maxExprSize
cheapExprSize e
= go 0 e
where
go n e | n >= maxExprSize = n
| otherwise = go1 n e
go1 n (Var {}) = n+1
go1 n (Lit {}) = n+1
go1 n (Type {}) = n
go1 n (Coercion {}) = n
go1 n (Tick _ e) = go1 n e
go1 n (Cast e _) = go1 n e
go1 n (App f a) = go (go1 n f) a
go1 n (Lam b e)
| isTyVar b = go1 n e
| otherwise = go (n+1) e
go1 n (Let b e) = gos (go1 n e) (rhssOfBind b)
go1 n (Case e _ _ as) = gos (go1 n e) (rhssOfAlts as)
gos n [] = n
gos n (e:es) | n >= maxExprSize = n
| otherwise = gos (go1 n e) es
betterLB :: NodeScore -> NodeScore -> Bool
-- If n1 `betterLB` n2 then choose n1 as the loop breaker
betterLB (rank1, size1, lb1) (rank2, size2, _)
| rank1 < rank2 = True
| rank1 > rank2 = False
| size1 < size2 = False -- Make the bigger n2 into the loop breaker
| size1 > size2 = True
| lb1 = True -- Tie-break: if n1 was a loop breaker before, choose it
| otherwise = False -- See Note [Loop breakers, node scoring, and stability]
{- Note [Self-recursion and loop breakers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
If we have
rec { f = ...f...g...
; g = .....f... }
then 'f' has to be a loop breaker anyway, so we may as well choose it
right away, so that g can inline freely.
This is really just a cheap hack. Consider
rec { f = ...g...
; g = ..f..h...
; h = ...f....}
Here f or g are better loop breakers than h; but we might accidentally
choose h. Finding the minimal set of loop breakers is hard.
Note [Loop breakers, node scoring, and stability]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
To choose a loop breaker, we give a NodeScore to each node in the SCC,
and pick the one with the best score (according to 'betterLB').
We need to be jolly careful (#12425, #12234) about the stability
of this choice. Suppose we have
let rec { f = ...g...g...
; g = ...f...f... }
in
case x of
True -> ...f..
False -> ..f...
In each iteration of the simplifier the occurrence analyser OccAnal
chooses a loop breaker. Suppose in iteration 1 it choose g as the loop
breaker. That means it is free to inline f.
Suppose that GHC decides to inline f in the branches of the case, but
(for some reason; eg it is not saturated) in the rhs of g. So we get
let rec { f = ...g...g...
; g = ...f...f... }
in
case x of
True -> ...g...g.....
False -> ..g..g....
Now suppose that, for some reason, in the next iteration the occurrence
analyser chooses f as the loop breaker, so it can freely inline g. And
again for some reason the simplifier inlines g at its calls in the case
branches, but not in the RHS of f. Then we get
let rec { f = ...g...g...
; g = ...f...f... }
in
case x of
True -> ...(...f...f...)...(...f..f..).....
False -> ..(...f...f...)...(..f..f...)....
You can see where this is going! Each iteration of the simplifier
doubles the number of calls to f or g. No wonder GHC is slow!
(In the particular example in comment:3 of #12425, f and g are the two
mutually recursive fmap instances for CondT and Result. They are both
marked INLINE which, oddly, is why they don't inline in each other's
RHS, because the call there is not saturated.)
The root cause is that we flip-flop on our choice of loop breaker. I
always thought it didn't matter, and indeed for any single iteration
to terminate, it doesn't matter. But when we iterate, it matters a
lot!!
So The Plan is this:
If there is a tie, choose the node that
was a loop breaker last time round
Hence the is_lb field of NodeScore
Note [Strictness in the occurrence analyser]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
By carefully making the occurrence analyser strict in some places, we can
dramatically reduce its memory residency. Among other things we:
* Evaluate the result of `tagLamBinder` and friends, so that the binder (or its
OccInfo) does not retain the entire `UsageDetails`. Also use `strictMap` in `tagLamBinders`.
* In `combineUsageDetailsWith`, the fields of the data constructor are strict, and we use
`strictPlusVarEnv` on the maps that are bound to be needed later on to avoid thunks being
stored in the values.
These measures reduced residency for test T26425 by a factor of at least 5x.
-}
{- *********************************************************************
* *
Lambda groups
* *
********************************************************************* -}
{- Note [Occurrence analysis for lambda binders]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
For value lambdas we do a special hack. Consider
(\x. \y. ...x...)
If we did nothing, x is used inside the \y, so would be marked
as dangerous to dup. But in the common case where the abstraction
is applied to two arguments this is over-pessimistic, which delays
inlining x, which forces more simplifier iterations.
So the occurrence analyser collaborates with the simplifier to treat
a /lambda-group/ specially. A lambda-group is a contiguous run of
lambda and casts, e.g.
Lam x (Lam y (Cast (Lam z body) co))
* Occurrence analyser: we just mark each binder in the lambda-group
(here: x,y,z) with its occurrence info in the *body* of the
lambda-group. See occAnalLamTail.
* Simplifier. The simplifier is careful when partially applying
lambda-groups. See the call to zapLambdaBndrs in
GHC.Core.Opt.Simplify.simplExprF1
GHC.Core.SimpleOpt.simple_app
* Why do we take care to account for intervening casts? Answer:
currently we don't do eta-expansion and cast-swizzling in a stable
unfolding (see Historical-note [Eta-expansion in stable unfoldings]).
So we can get
f = \x. ((\y. ...x...y...) |> co)
Now, since the lambdas aren't together, the occurrence analyser will
say that x is OnceInLam. Now if we have a call
(f e1 |> co) e2
we'll end up with
let x = e1 in ...x..e2...
and it'll take an extra iteration of the Simplifier to substitute for x.
A thought: a lambda-group is pretty much what GHC.Core.Opt.Arity.manifestArity
recognises except that the latter looks through (some) ticks. Maybe a lambda
group should also look through (some) ticks?
-}
isOneShotFun :: CoreExpr -> Bool
-- The top level lambdas, ignoring casts, of the expression
-- are all one-shot. If there aren't any lambdas at all, this is True
isOneShotFun (Lam b e) = isOneShotBndr b && isOneShotFun e
isOneShotFun (Cast e _) = isOneShotFun e
isOneShotFun _ = True
zapLambdaBndrs :: CoreExpr -> FullArgCount -> CoreExpr
-- If (\xyz. t) appears under-applied to only two arguments,
-- we must zap the occ-info on x,y, because they appear under the \z
-- See Note [Occurrence analysis for lambda binders] in GHc.Core.Opt.OccurAnal
--
-- NB: `arg_count` includes both type and value args
zapLambdaBndrs fun arg_count
= -- If the lambda is fully applied, leave it alone; if not
-- zap the OccInfo on the lambdas that do have arguments,
-- so they beta-reduce to use-many Lets rather than used-once ones.
zap arg_count fun `orElse` fun
where
zap :: FullArgCount -> CoreExpr -> Maybe CoreExpr
-- Nothing => No need to change the occ-info
-- Just e => Had to change
zap 0 e | isOneShotFun e = Nothing -- All remaining lambdas are one-shot
| otherwise = Just e -- in which case no need to zap
zap n (Cast e co) = do { e' <- zap n e; return (Cast e' co) }
zap n (Lam b e) = do { e' <- zap (n-1) e
; return (Lam (zap_bndr b) e') }
zap _ _ = Nothing -- More arguments than lambdas
zap_bndr b | isTyVar b = b
| otherwise = zapLamIdInfo b
occAnalLamTail :: OccEnv -> CoreExpr -> WithTailUsageDetails CoreExpr
-- ^ See Note [Occurrence analysis for lambda binders].
-- It does the following:
-- * Sets one-shot info on the lambda binder from the OccEnv, and
-- removes that one-shot info from the OccEnv
-- * Sets the OccEnv to OccVanilla when going under a value lambda
-- * Tags each lambda with its occurrence information
-- * Walks through casts
-- * Package up the analysed lambda with its manifest join arity
--
-- This function does /not/ do
-- markAllInsideLam or
-- markAllNonTail
-- The caller does that, via adjustTailUsage (mostly calls go through
-- adjustNonRecRhs). Every call to occAnalLamTail must ultimately call
-- adjustTailUsage to discharge the assumed join arity.
--
-- In effect, the analysis result is for a non-recursive join point with
-- manifest arity and adjustTailUsage does the fixup.
-- See Note [Adjusting right-hand sides]
occAnalLamTail env expr
= let !(WUD usage expr') = occ_anal_lam_tail env expr
in WTUD (TUD (joinRhsArity expr) usage) expr'
occ_anal_lam_tail :: OccEnv -> CoreExpr -> WithUsageDetails CoreExpr
-- Does not markInsideLam etc for the outmost batch of lambdas
occ_anal_lam_tail env expr@(Lam {})
= go env [] expr
where
go :: OccEnv -> [Var] -> CoreExpr -> WithUsageDetails CoreExpr
go env rev_bndrs (Lam bndr body)
| isTyVar bndr
= go env (bndr:rev_bndrs) body
-- Important: Unlike a value binder, do not modify occ_encl
-- to OccVanilla, so that with a RHS like
-- \(@ x) -> K @x (f @x)
-- we'll see that (K @x (f @x)) is in a OccRhs, and hence refrain
-- from inlining f. See the beginning of Note [Cascading inlines].
| otherwise
= let (env_one_shots', bndr')
= case occ_one_shots env of
[] -> ([], bndr)
(os : oss) -> (oss, updOneShotInfo bndr os)
-- Use updOneShotInfo, not setOneShotInfo, as pre-existing
-- one-shot info might be better than what we can infer, e.g.
-- due to explicit use of the magic 'oneShot' function.
-- See Note [oneShot magic]
env' = env { occ_encl = OccVanilla, occ_one_shots = env_one_shots' }
in go env' (bndr':rev_bndrs) body
go env rev_bndrs body
= addInScope env rev_bndrs $ \env ->
let !(WUD usage body') = occ_anal_lam_tail env body
-- See Note [Strictness in the occurrence analyser]
wrap_lam !body !bndr = let !bndr' = tagLamBinder usage bndr
in Lam bndr' body
in WUD (usage `addLamCoVarOccs` rev_bndrs)
(foldl' wrap_lam body' rev_bndrs)
-- For casts, keep going in the same lambda-group
-- See Note [Occurrence analysis for lambda binders]
occ_anal_lam_tail env (Cast expr co)
= let WUD usage expr' = occ_anal_lam_tail env expr
-- usage1: see Note [Gather occurrences of coercion variables]
usage1 = addManyOccs usage (coVarsOfCo co)
-- usage2: see Note [Occ-anal and cast worker/wrapper]
usage2 = case expr of
Var {} | isRhsEnv env -> markAllMany usage1
_ -> usage1
-- usage3: you might think this was not necessary, because of
-- the markAllNonTail in adjustTailUsage; but not so! For a
-- join point, adjustTailUsage doesn't do this; yet if there is
-- a cast, we must! Also: why markAllNonTail? See
-- GHC.Core.Lint: Note Note [Join points and casts]
usage3 = markAllNonTail usage2
in WUD usage3 (Cast expr' co)
occ_anal_lam_tail env expr -- Not Lam, not Cast
= occAnal env expr
{- Note [Occ-anal and cast worker/wrapper]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider y = e; x = y |> co
If we mark y as used-once, we'll inline y into x, and the Cast
worker/wrapper transform will float it straight back out again. See
Note [Cast worker/wrapper] in GHC.Core.Opt.Simplify.
So in this particular case we want to mark 'y' as Many. It's very
ad-hoc, but it's also simple. It's also what would happen if we gave
the binding for x a stable unfolding (as we usually do for wrappers, thus
y = e
{-# INLINE x #-}
x = y |> co
Now y appears twice -- once in x's stable unfolding, and once in x's
RHS. So it'll get a Many occ-info. (Maybe Cast w/w should create a stable
unfolding, which would obviate this Note; but that seems a bit of a
heavyweight solution.)
We only need to this in occAnalLamTail, not occAnal, because the top leve
of a right hand side is handled by occAnalLamTail.
-}
{- *********************************************************************
* *
Right hand sides
* *
********************************************************************* -}
occAnalUnfolding :: OccEnv
-> Unfolding
-> WithTailUsageDetails Unfolding
-- Occurrence-analyse a stable unfolding;
-- discard a non-stable one altogether and return empty usage details.
occAnalUnfolding !env unf
= case unf of
unf@(CoreUnfolding { uf_tmpl = rhs, uf_src = src })
| isStableSource src ->
let
WTUD (TUD rhs_ja uds) rhs' = occAnalLamTail env rhs
unf' = unf { uf_tmpl = rhs' }
in WTUD (TUD rhs_ja (markAllMany uds)) unf'
-- markAllMany: see Note [Occurrences in stable unfoldings]
| otherwise -> WTUD (TUD 0 emptyDetails) unf
-- For non-Stable unfoldings we leave them undisturbed, but
-- don't count their usage because the simplifier will discard them.
-- We leave them undisturbed because nodeScore uses their size info
-- to guide its decisions. It's ok to leave un-substituted
-- expressions in the tree because all the variables that were in
-- scope remain in scope; there is no cloning etc.
unf@(DFunUnfolding { df_bndrs = bndrs, df_args = args })
-> let WUD uds args' = addInScopeList env bndrs $ \ env ->
occAnalList env args
in WTUD (TUD 0 uds) (unf { df_args = args' })
-- No need to use tagLamBinders because we
-- never inline DFuns so the occ-info on binders doesn't matter
unf -> WTUD (TUD 0 emptyDetails) unf
occAnalRule :: OccEnv
-> CoreRule
-> (CoreRule, -- Each (non-built-in) rule
UsageDetails, -- Usage details for LHS
TailUsageDetails) -- Usage details for RHS
occAnalRule env rule@(Rule { ru_bndrs = bndrs, ru_args = args, ru_rhs = rhs })
= (rule', lhs_uds', TUD rhs_ja rhs_uds')
where
rule' = rule { ru_args = args', ru_rhs = rhs' }
WUD lhs_uds args' = addInScopeList env bndrs $ \env ->
occAnalList env args
lhs_uds' = markAllManyNonTail lhs_uds
WUD rhs_uds rhs' = addInScopeList env bndrs $ \env ->
occAnal env rhs
-- Note [Rules are extra RHSs]
-- Note [Rule dependency info]
rhs_uds' = markAllMany rhs_uds
rhs_ja = length args -- See Note [Join points and unfoldings/rules]
occAnalRule _ other_rule = (other_rule, emptyDetails, TUD 0 emptyDetails)
{- Note [Occurrences in stable unfoldings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
f p = BIG
{-# INLINE g #-}
g y = not (f y)
where this is the /only/ occurrence of 'f'. So 'g' will get a stable
unfolding. Now suppose that g's RHS gets optimised (perhaps by a rule
or inlining f) so that it doesn't mention 'f' any more. Now the last
remaining call to f is in g's Stable unfolding. But, even though there
is only one syntactic occurrence of f, we do /not/ want to do
preinlineUnconditionally here!
The INLINE pragma says "inline exactly this RHS"; perhaps the
programmer wants to expose that 'not', say. If we inline f that will make
the Stable unfoldign big, and that wasn't what the programmer wanted.
Another way to think about it: if we inlined g as-is into multiple
call sites, now there's be multiple calls to f.
Bottom line: treat all occurrences in a stable unfolding as "Many".
We still leave tail call information intact, though, as to not spoil
potential join points.
Note [Unfoldings and rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Generally unfoldings and rules are already occurrence-analysed, so we
don't want to reconstruct their trees; we just want to analyse them to
find how they use their free variables.
EXCEPT if there is a binder-swap going on, in which case we do want to
produce a new tree.
So we have a fast-path that keeps the old tree if the occ_bs_env is
empty. This just saves a bit of allocation and reconstruction; not
a big deal.
Two tricky corners:
* Dead bindings (#22761). Supose we have
Unfolding = \x. let y = foo in x+1
which includes a dead binding for `y`. In occAnalUnfolding we occ-anal
the unfolding and produce /no/ occurrences of `foo` (since `y` is
dead). But if we discard the occ-analysed syntax tree (which we do on
our fast path), and use the old one, we still /have/ an occurrence of
`foo` -- and that can lead to out-of-scope variables (#22761).
Solution: always keep occ-analysed trees in unfoldings and rules, so they
have no dead code. See Note [OccInfo in unfoldings and rules] in GHC.Core.
* One-shot binders. Consider
{- f has Stable unfolding \p q -> blah
Demand on f is LC(L,C(1,!P(L)); that is, one-shot in its second ar -}
f = \x y. blah
Now we `mkRhsOccEnv` will build an OccEnv for f's RHS that has
occ_one_shots = [NoOneShortInfo, OneShotLam]
This will put OneShotLam on the \y. And it'll put it on the \q. But the
noBinderSwap check will mean that we discard this new occ-anal'd unfolding
and keep the old one, with no OneShotInfo.
This looks a little inconsistent, but the Stable unfolding is just used for
inlinings; OneShotInfo isn't a lot of use here.
Note [Cascading inlines]
~~~~~~~~~~~~~~~~~~~~~~~~
By default we use an OccRhs for the RHS of a binding. This tells the
occ anal n that it's looking at an RHS, which has an effect in
occAnalApp. In particular, for constructor applications, it makes
the arguments appear to have NoOccInfo, so that we don't inline into
them. Thus x = f y
k = Just x
we do not want to inline x.
But there's a problem. Consider
x1 = a0 : []
x2 = a1 : x1
x3 = a2 : x2
g = f x3
First time round, it looks as if x1 and x2 occur as an arg of a
let-bound constructor ==> give them a many-occurrence.
But then x3 is inlined (unconditionally as it happens) and
next time round, x2 will be, and the next time round x1 will be
Result: multiple simplifier iterations. Sigh.
So, when analysing the RHS of x3 we notice that x3 will itself
definitely inline the next time round, and so we analyse x3's rhs in
an OccVanilla context, not OccRhs. Hence the "certainly_inline" stuff.
Annoyingly, we have to approximate GHC.Core.Opt.Simplify.Utils.preInlineUnconditionally.
If (a) the RHS is expandable (see isExpandableApp in occAnalApp), and
(b) certainly_inline says "yes" when preInlineUnconditionally says "no"
then the simplifier iterates indefinitely:
x = f y
k = Just x -- We decide that k is 'certainly_inline'
v = ...k... -- but preInlineUnconditionally doesn't inline it
inline ==>
k = Just (f y)
v = ...k...
float ==>
x1 = f y
k = Just x1
v = ...k...
This is worse than the slow cascade, so we only want to say "certainly_inline"
if it really is certain. Look at the note with preInlineUnconditionally
for the various clauses. See #24582 for an example of the two getting out of sync.
************************************************************************
* *
Expressions
* *
************************************************************************
-}
occAnalList :: OccEnv -> [CoreExpr] -> WithUsageDetails [CoreExpr]
occAnalList !_ [] = WUD emptyDetails []
occAnalList env (e:es) = let
(WUD uds1 e') = occAnal env e
(WUD uds2 es') = occAnalList env es
in WUD (uds1 `andUDs` uds2) (e' : es')
occAnal :: OccEnv
-> CoreExpr
-> WithUsageDetails CoreExpr -- Gives info only about the "interesting" Ids
occAnal !_ expr@(Lit _) = WUD emptyDetails expr
occAnal env expr@(Var _) = occAnalApp env (expr, [], [])
-- At one stage, I gathered the idRuleVars for the variable here too,
-- which in a way is the right thing to do.
-- But that went wrong right after specialisation, when
-- the *occurrences* of the overloaded function didn't have any
-- rules in them, so the *specialised* versions looked as if they
-- weren't used at all.
occAnal _ expr@(Type ty)
= WUD (addManyOccs emptyDetails (coVarsOfType ty)) expr
occAnal _ expr@(Coercion co)
= WUD (addManyOccs emptyDetails (coVarsOfCo co)) expr
-- See Note [Gather occurrences of coercion variables]
{- Note [Gather occurrences of coercion variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We need to gather info about what coercion variables appear, for two reasons:
1. So that we can sort them into the right place when doing dependency analysis.
2. So that we know when they are surely dead.
It is useful to know when they a coercion variable is surely dead,
when we want to discard a case-expression, in GHC.Core.Opt.Simplify.rebuildCase.
For example (#20143):
case unsafeEqualityProof @blah of
UnsafeRefl cv -> ...no use of cv...
Here we can discard the case, since unsafeEqualityProof always terminates.
But only if the coercion variable 'cv' is unused.
Another example from #15696: we had something like
case eq_sel d of co -> ...(typeError @(...co...) "urk")...
Then 'd' was substituted by a dictionary, so the expression
simpified to
case (Coercion <blah>) of cv -> ...(typeError @(...cv...) "urk")...
We can only drop the case altogether if 'cv' is unused, which is not
the case here.
Conclusion: we need accurate dead-ness info for CoVars.
We gather CoVar occurrences from:
* The (Type ty) and (Coercion co) cases of occAnal
* The type 'ty' of a lambda-binder (\(x:ty). blah)
See addCoVarOccs
But it is not necessary to gather CoVars from the types of other binders.
* For let-binders, if the type mentions a CoVar, so will the RHS (since
it has the same type)
* For case-alt binders, if the type mentions a CoVar, so will the scrutinee
(since it has the same type)
-}
occAnal env (Tick tickish body)
= WUD usage' (Tick tickish body')
where
WUD usage body' = occAnal env body
usage'
| tickish `tickishScopesLike` SoftScope
= usage -- For soft-scoped ticks (including SourceNotes) we don't want
-- to lose join-point-hood, so we don't mess with `usage` (#24078)
-- For a non-soft tick scope, we can inline lambdas only, so we
-- abandon tail calls, and do markAllInsideLam too: usage_lam
| Breakpoint _ _ ids _ <- tickish
= -- Never substitute for any of the Ids in a Breakpoint
addManyOccs usage_lam (mkVarSet ids)
| otherwise
= usage_lam
usage_lam = markAllNonTail (markAllInsideLam usage)
-- TODO There may be ways to make ticks and join points play
-- nicer together, but right now there are problems:
-- let j x = ... in tick<t> (j 1)
-- Making j a join point may cause the simplifier to drop t
-- (if the tick is put into the continuation). So we don't
-- count j 1 as a tail call.
-- See #14242.
occAnal env (Cast expr co)
= let (WUD usage expr') = occAnal env expr
usage1 = addManyOccs usage (coVarsOfCo co)
-- usage2: see Note [Gather occurrences of coercion variables]
usage2 = markAllNonTail usage1
-- usage3: calls inside expr aren't tail calls any more
in WUD usage2 (Cast expr' co)
occAnal env app@(App _ _)
= occAnalApp env (collectArgsTicks tickishFloatable app)
occAnal env expr@(Lam {})
= adjustNonRecRhs NotJoinPoint $ -- NotJoinPoint <=> markAllManyNonTail
occAnalLamTail env expr
occAnal env (Case scrut bndr ty alts)
= let
WUD scrut_usage scrut' = occAnal (setScrutCtxt env alts) scrut
WUD alts_usage (tagged_bndr, alts')
= addInScopeOne env bndr $ \env ->
let alt_env = addBndrSwap scrut' bndr $
setTailCtxt env -- Kill off OccRhs
WUD alts_usage alts' = do_alts alt_env alts
!tagged_bndr = tagLamBinder alts_usage bndr
-- See Note [Strictness in the occurrence analyser]
in WUD alts_usage (tagged_bndr, alts')
total_usage = markAllNonTail scrut_usage `andUDs` alts_usage
-- Alts can have tail calls, but the scrutinee can't
in WUD total_usage (Case scrut' tagged_bndr ty alts')
where
do_alts :: OccEnv -> [CoreAlt] -> WithUsageDetails [CoreAlt]
do_alts _ [] = WUD emptyDetails []
do_alts env (alt:alts) = WUD (uds1 `orUDs` uds2) (alt':alts')
where
WUD uds1 alt' = do_alt env alt
WUD uds2 alts' = do_alts env alts
do_alt !env (Alt con bndrs rhs)
= addInScopeList env bndrs $ \ env ->
let WUD rhs_usage rhs' = occAnal env rhs
!tagged_bndrs = tagLamBinders rhs_usage bndrs
-- See Note [Strictness in the occurrence analyser]
in -- See Note [Binders in case alternatives]
WUD rhs_usage (Alt con tagged_bndrs rhs')
occAnal env (Let bind body)
-- TODO: Would be nice to use a strict version of mkLets here
= occAnalBind env NotTopLevel noImpRuleEdges bind
(\env -> occAnal env body) mkLets
occAnalArgs :: OccEnv -> CoreExpr -> [CoreExpr]
-> [OneShots] -- Very commonly empty, notably prior to dmd anal
-> WithUsageDetails CoreExpr
-- The `fun` argument is just an accumulating parameter,
-- the base for building the application we return
occAnalArgs !env fun args !one_shots
= go emptyDetails fun args one_shots
where
env_args = setNonTailCtxt encl env
-- Make bottoming functions interesting
-- See Note [Bottoming function calls]
encl | Var f <- fun, isDeadEndSig (idDmdSig f) = OccScrut
| otherwise = OccVanilla
go uds fun [] _ = WUD uds fun
go uds fun (arg:args) one_shots
= go (uds `andUDs` arg_uds) (fun `App` arg') args one_shots'
where
!(WUD arg_uds arg') = occAnal arg_env arg
!(arg_env, one_shots')
| isTypeArg arg
= (env_args, one_shots)
| otherwise
= case one_shots of
[] -> (env_args, []) -- Fast path; one_shots is often empty
(os : one_shots') -> (setOneShots os env_args, one_shots')
{-
Applications are dealt with specially because we want
the "build hack" to work.
Note [Bottoming function calls]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
let x = (a,b) in
case p of
A -> ...(error x)..
B -> ...(ertor x)...
postInlineUnconditionally may duplicate x's binding, but sometimes it
does so only if the use site IsInteresting. Pushing allocation into error
branches is good, so we try to make bottoming calls look interesting, by
setting occ_encl = OccScrut for such calls.
The slightly-artificial test T21128 is a good example. It's probably
not a huge deal.
Note [Arguments of let-bound constructors]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
f x = let y = expensive x in
let z = (True,y) in
(case z of {(p,q)->q}, case z of {(p,q)->q})
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.
-}
occAnalApp :: OccEnv
-> (Expr CoreBndr, [Arg CoreBndr], [CoreTickish])
-> WithUsageDetails (Expr CoreBndr)
-- Naked variables (not applied) end up here too
occAnalApp !env (Var fun, args, ticks)
-- Account for join arity of runRW# continuation
-- See Note [Simplification of runRW#]
--
-- NB: Do not be tempted to make the next (Var fun, args, tick)
-- equation into an 'otherwise' clause for this equation
-- The former has a bang-pattern to occ-anal the args, and
-- we don't want to occ-anal them twice in the runRW# case!
-- This caused #18296
| fun `hasKey` runRWKey
, [t1, t2, arg] <- args
, WUD usage arg' <- adjustNonRecRhs (JoinPoint 1) $ occAnalLamTail env arg
= let app_out = mkTicks ticks $ mkApps (Var fun) [t1, t2, arg']
in WUD usage app_out
occAnalApp env (Var fun_id, args, ticks)
= let app_out = mkTicks ticks app'
in WUD all_uds app_out
where
-- Lots of banged bindings: this is a very heavily bit of code,
-- so it pays not to make lots of thunks here, all of which
-- will ultimately be forced.
!(fun', fun_id') = lookupBndrSwap env fun_id
!(WUD args_uds app') = occAnalArgs env fun' args one_shots
fun_uds = mkOneOcc env fun_id' int_cxt n_args
-- NB: fun_uds is computed for fun_id', not fun_id
-- See (BS1) in Note [The binder-swap substitution]
all_uds = fun_uds `andUDs` final_args_uds
!final_args_uds = markAllNonTail $
markAllInsideLamIf (isRhsEnv env && is_exp) $
-- isRhsEnv: see Note [OccEncl]
args_uds
-- We mark the free vars of the argument of a constructor or PAP
-- as "inside-lambda", if it is the RHS of a let(rec).
-- This means that nothing gets inlined into a constructor or PAP
-- argument position, which is what we want. Typically those
-- constructor arguments are just variables, or trivial expressions.
-- We use inside-lam because it's like eta-expanding the PAP.
--
-- This is the *whole point* of the isRhsEnv predicate
-- See Note [Arguments of let-bound constructors]
!n_val_args = valArgCount args
!n_args = length args
!int_cxt = case occ_encl env of
OccScrut -> IsInteresting
_other | n_val_args > 0 -> IsInteresting
| otherwise -> NotInteresting
!is_exp = isExpandableApp fun_id n_val_args
-- See Note [CONLIKE pragma] in GHC.Types.Basic
-- The definition of is_exp should match that in GHC.Core.Opt.Simplify.prepareRhs
one_shots = argsOneShots (idDmdSig fun_id) guaranteed_val_args
guaranteed_val_args = n_val_args + length (takeWhile isOneShotInfo
(occ_one_shots env))
-- See Note [Sources of one-shot information], bullet point A']
occAnalApp env (fun, args, ticks)
= let app_out = mkTicks ticks app'
in WUD (markAllNonTail (fun_uds `andUDs` args_uds)) app_out
where
!(WUD args_uds app') = occAnalArgs env fun' args []
!(WUD fun_uds fun') = occAnal (addAppCtxt env args) fun
-- The addAppCtxt is a bit cunning. One iteration of the simplifier
-- often leaves behind beta redexes like
-- (\x y -> e) a1 a2
-- Here we would like to mark x,y as one-shot, and treat the whole
-- thing much like a let. We do this by pushing some OneShotLam items
-- onto the context stack.
addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
addAppCtxt env@(OccEnv { occ_one_shots = ctxt }) args
| n_val_args > 0
= env { occ_one_shots = replicate n_val_args OneShotLam ++ ctxt
, occ_encl = OccVanilla }
-- OccVanilla: the function part of the application
-- is no longer on OccRhs or OccScrut
| otherwise
= env
where
n_val_args = valArgCount args
{-
Note [Sources of one-shot information]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The occurrence analyser obtains one-shot-lambda information from two sources:
A: Saturated applications: eg f e1 .. en
In general, given a call (f e1 .. en) we can propagate one-shot info from
f's strictness signature into e1 .. en, but /only/ if n is enough to
saturate the strictness signature. A strictness signature like
f :: C(1,C(1,L))LS
means that *if f is applied to three arguments* then it will guarantee to
call its first argument at most once, and to call the result of that at
most once. But if f has fewer than three arguments, all bets are off; e.g.
map (f (\x y. expensive) e2) xs
Here the \x y abstraction may be called many times (once for each element of
xs) so we should not mark x and y as one-shot. But if it was
map (f (\x y. expensive) 3 2) xs
then the first argument of f will be called at most once.
The one-shot info, derived from f's strictness signature, is
computed by 'argsOneShots', called in occAnalApp.
A': Non-obviously saturated applications: eg build (f (\x y -> expensive))
where f is as above.
In this case, f is only manifestly applied to one argument, so it does not
look saturated. So by the previous point, we should not use its strictness
signature to learn about the one-shotness of \x y. But in this case we can:
build is fully applied, so we may use its strictness signature; and from
that we learn that build calls its argument with two arguments *at most once*.
So there is really only one call to f, and it will have three arguments. In
that sense, f is saturated, and we may proceed as described above.
Hence the computation of 'guaranteed_val_args' in occAnalApp, using
'(occ_one_shots env)'. See also #13227, comment:9
B: Let-bindings: eg let f = \c. let ... in \n -> blah
in (build f, build f)
Propagate one-shot info from the demand-info on 'f' to the
lambdas in its RHS (which may not be syntactically at the top)
This information must have come from a previous run of the demand
analyser.
Previously, the demand analyser would *also* set the one-shot information, but
that code was buggy (see #11770), so doing it only in on place, namely here, is
saner.
Note [OneShots]
~~~~~~~~~~~~~~~
When analysing an expression, the occ_one_shots argument contains information
about how the function is being used. The length of the list indicates
how many arguments will eventually be passed to the analysed expression,
and the OneShotInfo indicates whether this application is once or multiple times.
Example:
Context of f occ_one_shots when analysing f
f 1 2 [OneShot, OneShot]
map (f 1) [OneShot, NoOneShotInfo]
build f [OneShot, OneShot]
f 1 2 `seq` f 2 1 [NoOneShotInfo, OneShot]
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, because it binds 'y' to (a,b) (imagine got inlined and
scrutinised y).
-}
{-
************************************************************************
* *
OccEnv
* *
************************************************************************
-}
data OccEnv
= OccEnv { occ_encl :: !OccEncl -- Enclosing context information
, occ_one_shots :: !OneShots -- See Note [OneShots]
, occ_unf_act :: Id -> Bool -- Which Id unfoldings are active
, occ_rule_act :: Activation -> Bool -- Which rules are active
-- See Note [Finding rule RHS free vars]
-- See Note [The binder-swap substitution]
-- If x :-> (y, co) is in the env,
-- then please replace x by (y |> mco)
-- Invariant of course: idType x = exprType (y |> mco)
, occ_bs_env :: !(IdEnv (OutId, MCoercion))
-- Domain is Global and Local Ids
-- Range is just Local Ids
, occ_bs_rng :: !VarSet
-- Vars (TyVars and Ids) free in the range of occ_bs_env
-- Usage details of the RHS of in-scope non-recursive join points
-- Invariant: no Id maps to an empty OccInfoEnv
-- See Note [Occurrence analysis for join points]
, occ_join_points :: !JoinPointInfo
}
type JoinPointInfo = IdEnv OccInfoEnv
-----------------------------
{- Note [OccEncl]
~~~~~~~~~~~~~~~~~
OccEncl is used to control whether to inline into constructor arguments.
* OccRhs: consider
let p = <blah> in
let x = Just p
in ...case p of ...
Here `p` occurs syntactically once, but we want to mark it as InsideLam
to stop `p` inlining. We want to leave the x-binding as a constructor
applied to variables, so that the Simplifier can simplify that inner `case`.
The OccRhs just tells occAnalApp to mark occurrences in constructor args
* OccScrut: consider (case x of ...). Here we want to give `x` OneOcc
with "interesting context" field int_cxt = True. The OccScrut tells
occAnalApp (which deals with lone variables too) when to set this field
to True.
-}
data OccEncl -- See Note [OccEncl]
= OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
| OccScrut -- Scrutintee of a case
| OccVanilla -- Everything else
instance Outputable OccEncl where
ppr OccRhs = text "occRhs"
ppr OccScrut = text "occScrut"
ppr OccVanilla = text "occVanilla"
-- See Note [OneShots]
type OneShots = [OneShotInfo]
initOccEnv :: OccEnv
initOccEnv
= OccEnv { occ_encl = OccVanilla
, occ_one_shots = []
-- To be conservative, we say that all
-- inlines and rules are active
, occ_unf_act = \_ -> True
, occ_rule_act = \_ -> True
, occ_join_points = emptyVarEnv
, occ_bs_env = emptyVarEnv
, occ_bs_rng = emptyVarSet }
noBinderSwaps :: OccEnv -> Bool
noBinderSwaps (OccEnv { occ_bs_env = bs_env }) = isEmptyVarEnv bs_env
setScrutCtxt :: OccEnv -> [CoreAlt] -> OccEnv
setScrutCtxt !env alts
= setNonTailCtxt encl env
where
encl | interesting_alts = OccScrut
| otherwise = OccVanilla
interesting_alts = case alts of
[] -> False
[alt] -> not (isDefaultAlt alt)
_ -> True
-- 'interesting_alts' is True if the case has at least one
-- non-default alternative. That in turn influences
-- pre/postInlineUnconditionally. Grep for "occ_int_cxt"!
{- Note [The OccEnv for a right hand side]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
How do we create the OccEnv for a RHS (in mkRhsOccEnv)?
For a non-join point binding, x = rhs
* occ_encl: set to OccRhs; but see `mkNonRecRhsCtxt` for wrinkles
* occ_join_points: zap them!
* occ_one_shots: initialise from the idDemandInfo;
see Note [Sources of one-shot information]
For a join point binding, j x = rhs
* occ_encl: Consider
x = e
join j = Just x
We want to inline x into j right away, so we don't want to give the join point
a OccRhs (#14137); we want OccVanilla. It's not a huge deal, because the
FloatIn pass knows to float into join point RHSs; and the simplifier does not
float things out of join point RHSs. But it's a simple, cheap thing to do.
* occ_join_points: no need to zap.
* occ_one_shots: we start with one-shot-info from the context, which indeed
applies to the /body/ of the join point, after walking past the binders.
So we add to the front a OneShotInfo for each value-binder of the join
point: see `extendOneShotsForJoinPoint`. (Failing to account for the join-point
binders caused #25096.)
For the join point binders themselves, of a /non-recursive/ join point,
we make the binder a OneShotLam. Again see `extendOneShotsForJoinPoint`.
These one-shot infos then get attached to the binder by `occAnalLamTail`.
-}
setNonTailCtxt :: OccEncl -> OccEnv -> OccEnv
setNonTailCtxt ctxt !env
= env { occ_encl = ctxt
, occ_one_shots = []
, occ_join_points = zapJoinPointInfo (occ_join_points env) }
setTailCtxt :: OccEnv -> OccEnv
setTailCtxt !env = env { occ_encl = OccVanilla }
-- Preserve occ_one_shots, occ_join points
-- Do not use OccRhs for the RHS of a join point (which is a tail ctxt):
mkRhsOccEnv :: OccEnv -> RecFlag -> OccEncl -> JoinPointHood -> Id -> CoreExpr -> OccEnv
-- See Note [The OccEnv for a right hand side]
-- For a join point:
-- - Keep occ_one_shots, occ_joinPoints from the context
-- - But push enough OneShotInfo onto occ_one_shots to account
-- for the join-point value binders
-- - Set occ_encl to OccVanilla
-- For non-join points
-- - Zap occ_one_shots and occ_join_points
-- - Set occ_encl to specified OccEncl
mkRhsOccEnv env@(OccEnv { occ_one_shots = ctxt_one_shots, occ_join_points = ctxt_join_points })
is_rec encl jp_hood bndr rhs
| JoinPoint join_arity <- jp_hood
= env { occ_encl = OccVanilla
, occ_one_shots = extendOneShotsForJoinPoint is_rec join_arity rhs ctxt_one_shots
, occ_join_points = ctxt_join_points }
| otherwise
= env { occ_encl = encl
, occ_one_shots = argOneShots (idDemandInfo bndr)
-- argOneShots: see Note [Sources of one-shot information]
, occ_join_points = zapJoinPointInfo ctxt_join_points }
zapJoinPointInfo :: JoinPointInfo -> JoinPointInfo
-- (zapJoinPointInfo jp_info) basically just returns emptyVarEnv (hence zapped).
-- See (W3) of Note [Occurrence analysis for join points]
--
-- Zapping improves efficiency, slightly, if you accidentally introduce a bug,
-- in which you zap [jx :-> uds] and then find an occurrence of jx anyway, you
-- might lose those uds, and that might mean we don't record all occurrencs, and
-- that means we duplicate a redex.... a very nasty bug (which I encountered!).
-- Hence this DEBUG code which doesn't remove jx from the envt; it just gives it
-- emptyDetails, which in turn causes a panic in mkOneOcc. That will catch this
-- bug before it does any damage.
#ifdef DEBUG
zapJoinPointInfo jp_info = mapVarEnv (\ _ -> emptyVarEnv) jp_info
#else
zapJoinPointInfo _ = emptyVarEnv
#endif
extendOneShotsForJoinPoint
:: RecFlag -> JoinArity -> CoreExpr
-> [OneShotInfo] -> [OneShotInfo]
-- Push enough OneShortInfos on the front of ctxt_one_shots
-- to account for the value lambdas of the join point
extendOneShotsForJoinPoint is_rec join_arity rhs ctxt_one_shots
= go join_arity rhs
where
-- For a /non-recursive/ join point we can mark all
-- its join-lambda as one-shot; and it's a good idea to do so
-- But not so for recursive ones
os = case is_rec of
NonRecursive -> OneShotLam
Recursive -> NoOneShotInfo
go 0 _ = ctxt_one_shots
go n (Lam b rhs)
| isId b = os : go (n-1) rhs
| otherwise = go (n-1) rhs
go _ _ = [] -- Not enough lambdas. This can legitimately happen.
-- e.g. let j = case ... in j True
-- This will become an arity-1 join point after the
-- simplifier has eta-expanded it; but it may not have
-- enough lambdas /yet/. (Lint checks that JoinIds do
-- have enough lambdas.)
setOneShots :: OneShots -> OccEnv -> OccEnv
setOneShots os !env
| null os = env -- Fast path for common case
| otherwise = env { occ_one_shots = os }
isRhsEnv :: OccEnv -> Bool
isRhsEnv (OccEnv { occ_encl = cxt }) = case cxt of
OccRhs -> True
_ -> False
addInScopeList :: OccEnv -> [Var]
-> (OccEnv -> WithUsageDetails a) -> WithUsageDetails a
{-# INLINE addInScopeList #-}
addInScopeList env bndrs thing_inside
| null bndrs = thing_inside env -- E.g. nullary constructors in a `case`
| otherwise = addInScope env bndrs thing_inside
addInScopeOne :: OccEnv -> Id
-> (OccEnv -> WithUsageDetails a) -> WithUsageDetails a
{-# INLINE addInScopeOne #-}
addInScopeOne env bndr = addInScope env [bndr]
addInScope :: OccEnv -> [Var]
-> (OccEnv -> WithUsageDetails a) -> WithUsageDetails a
{-# INLINE addInScope #-}
-- This function is called a lot, so we want to inline the fast path
-- so we don't have to allocate thing_inside and call it
-- The bndrs must include TyVars as well as Ids, because of
-- (BS3) in Note [Binder swap]
-- We do not assume that the bndrs are in scope order; in fact the
-- call in occ_anal_lam_tail gives them to addInScope in /reverse/ order
-- Fast path when the is no environment-munging to do
-- This is rather common: notably at top level, but nested too
addInScope env bndrs thing_inside
| isEmptyVarEnv (occ_bs_env env)
, isEmptyVarEnv (occ_join_points env)
, WUD uds res <- thing_inside env
= WUD (delBndrsFromUDs bndrs uds) res
addInScope env bndrs thing_inside
= WUD uds' res
where
bndr_set = mkVarSet bndrs
!(env', bad_joins) = preprocess_env env bndr_set
!(WUD uds res) = thing_inside env'
uds' = postprocess_uds bndrs bad_joins uds
preprocess_env :: OccEnv -> VarSet -> (OccEnv, JoinPointInfo)
preprocess_env env@(OccEnv { occ_join_points = join_points
, occ_bs_rng = bs_rng_vars })
bndr_set
| bad_joins = (drop_shadowed_swaps (drop_shadowed_joins env), join_points)
| otherwise = (drop_shadowed_swaps env, emptyVarEnv)
where
drop_shadowed_swaps :: OccEnv -> OccEnv
-- See Note [The binder-swap substitution] (BS3)
drop_shadowed_swaps env@(OccEnv { occ_bs_env = swap_env })
| isEmptyVarEnv swap_env
= env
| bs_rng_vars `intersectsVarSet` bndr_set
= env { occ_bs_env = emptyVarEnv, occ_bs_rng = emptyVarSet }
| otherwise
= env { occ_bs_env = swap_env `minusUFM` bndr_fm }
drop_shadowed_joins :: OccEnv -> OccEnv
-- See Note [Occurrence analysis for join points] wrinkle2 (W1) and (W2)
drop_shadowed_joins env = env { occ_join_points = emptyVarEnv }
-- bad_joins is true if it would be wrong to push occ_join_points inwards
-- (a) `bndrs` includes any of the occ_join_points
-- (b) `bndrs` includes any variables free in the RHSs of occ_join_points
bad_joins :: Bool
bad_joins = nonDetStrictFoldVarEnv_Directly is_bad False join_points
bndr_fm :: UniqFM Var Var
bndr_fm = getUniqSet bndr_set
is_bad :: Unique -> OccInfoEnv -> Bool -> Bool
is_bad uniq join_uds rest
= uniq `elemUniqSet_Directly` bndr_set ||
not (bndr_fm `disjointUFM` join_uds) ||
rest
postprocess_uds :: [Var] -> JoinPointInfo -> UsageDetails -> UsageDetails
postprocess_uds bndrs bad_joins uds
= add_bad_joins (delBndrsFromUDs bndrs uds)
where
add_bad_joins :: UsageDetails -> UsageDetails
-- Add usage info for occ_join_points that we cannot push inwards
-- because of shadowing
-- See Note [Occurrence analysis for join points] wrinkle (W2)
add_bad_joins uds
| isEmptyVarEnv bad_joins = uds
| otherwise = modifyUDEnv extend_with_bad_joins uds
extend_with_bad_joins :: OccInfoEnv -> OccInfoEnv
extend_with_bad_joins env
= nonDetStrictFoldUFM_Directly add_bad_join env bad_joins
add_bad_join :: Unique -> OccInfoEnv -> OccInfoEnv -> OccInfoEnv
-- Behave like `andUDs` when adding in the bad_joins
add_bad_join uniq join_env env
| uniq `elemVarEnvByKey` env = plusVarEnv_C andLocalOcc env join_env
| otherwise = env
addJoinPoint :: OccEnv -> Id -> UsageDetails -> OccEnv
addJoinPoint env bndr rhs_uds
| isEmptyVarEnv zeroed_form
= env
| otherwise
= env { occ_join_points = extendVarEnv (occ_join_points env) bndr zeroed_form }
where
zeroed_form = mkZeroedForm rhs_uds
mkZeroedForm :: UsageDetails -> OccInfoEnv
-- See Note [Occurrence analysis for join points] for "zeroed form"
mkZeroedForm (UD { ud_env = rhs_occs })
= mapMaybeUFM do_one rhs_occs
where
do_one :: LocalOcc -> Maybe LocalOcc
do_one (ManyOccL {}) = Nothing
do_one occ@(OneOccL {}) = Just (occ { lo_n_br = 0 })
--------------------
transClosureFV :: VarEnv VarSet -> VarEnv VarSet
-- If (f,g), (g,h) are in the input, then (f,h) is in the output
-- as well as (f,g), (g,h)
transClosureFV env
| no_change = env
| otherwise = transClosureFV (listToUFM_Directly new_fv_list)
where
(no_change, new_fv_list) = mapAccumL bump True (nonDetUFMToList env)
-- It's OK to use nonDetUFMToList here because we'll forget the
-- ordering by creating a new set with listToUFM
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_ :: VarEnv VarSet -> VarSet -> VarSet
extendFvs_ env s = fst (extendFvs env s) -- Discard the Bool flag
extendFvs :: VarEnv VarSet -> VarSet -> (VarSet, Bool)
-- (extendFVs env s) returns
-- (s `union` env(s), env(s) `subset` s)
extendFvs env s
| isNullUFM env
= (s, True)
| otherwise
= (s `unionVarSet` extras, extras `subVarSet` s)
where
extras :: VarSet -- env(s)
extras = nonDetStrictFoldUFM unionVarSet emptyVarSet $
-- It's OK to use nonDetStrictFoldUFM here because unionVarSet commutes
intersectUFM_C (\x _ -> x) env (getUniqSet s)
{-
************************************************************************
* *
Binder swap
* *
************************************************************************
Note [Binder swap]
~~~~~~~~~~~~~~~~~~
The "binder swap" transformation swaps occurrence of the
scrutinee of a case for occurrences of the case-binder:
(1) case x of b { pi -> ri }
==>
case x of b { pi -> ri[b/x] }
(2) case (x |> co) of b { pi -> ri }
==>
case (x |> co) of b { pi -> ri[b |> sym co/x] }
The substitution ri[b/x] etc is done by the occurrence analyser.
See Note [The binder-swap substitution].
There are two reasons for making this swap:
(A) It reduces the number of occurrences of the scrutinee, x.
That in turn might reduce its occurrences to one, so we
can inline it and save an allocation. E.g.
let x = factorial y in case x of b { I# v -> ...x... }
If we replace 'x' by 'b' in the alternative we get
let x = factorial y in case x of b { I# v -> ...b... }
and now we can inline 'x', thus
case (factorial y) of b { I# v -> ...b... }
(B) The case-binder b has unfolding information; in the
example above we know that b = I# v. That in turn allows
nested cases to simplify. Consider
case x of b { I# v ->
...(case x of b2 { I# v2 -> rhs })...
If we replace 'x' by 'b' in the alternative we get
case x of b { I# v ->
...(case b of b2 { I# v2 -> rhs })...
and now it is trivial to simplify the inner case:
case x of b { I# v ->
...(let b2 = b in rhs)...
The same can happen even if the scrutinee is a variable
with a cast: see Note [Case of cast]
The reason for doing these transformations /here in the occurrence
analyser/ 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 instead the Simplifier replaces occurrences of x with
occurrences of b, that will mess up b's occurrence info. That in
turn might have consequences.
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 [The binder-swap substitution]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The binder-swap is implemented by the occ_bs_env field of OccEnv.
There are two main pieces:
* Given case x |> co of b { alts }
we add [x :-> (b, sym co)] to the occ_bs_env environment; this is
done by addBndrSwap.
* Then, at an occurrence of a variable, we look up in the occ_bs_env
to perform the swap. This is done by lookupBndrSwap.
Some tricky corners:
(BS1) We do the substitution before gathering occurrence info. So in
the above example, an occurrence of x turns into an occurrence
of b, and that's what we gather in the UsageDetails. It's as
if the binder-swap occurred before occurrence analysis. See
the computation of fun_uds in occAnalApp.
(BS2) When doing a lookup in occ_bs_env, we may need to iterate,
as you can see implemented in lookupBndrSwap. Why?
Consider case x of a { 1# -> e1; DEFAULT ->
case x of b { 2# -> e2; DEFAULT ->
case x of c { 3# -> e3; DEFAULT -> ..x..a..b.. }}}
At the first case addBndrSwap will extend occ_bs_env with
[x :-> a]
At the second case we occ-anal the scrutinee 'x', which looks up
'x in occ_bs_env, returning 'a', as it should.
Then addBndrSwap will add [a :-> b] to occ_bs_env, yielding
occ_bs_env = [x :-> a, a :-> b]
At the third case we'll again look up 'x' which returns 'a'.
But we don't want to stop the lookup there, else we'll end up with
case x of a { 1# -> e1; DEFAULT ->
case a of b { 2# -> e2; DEFAULT ->
case a of c { 3# -> e3; DEFAULT -> ..a..b..c.. }}}
Instead, we want iterate the lookup in addBndrSwap, to give
case x of a { 1# -> e1; DEFAULT ->
case a of b { 2# -> e2; DEFAULT ->
case b of c { 3# -> e3; DEFAULT -> ..c..c..c.. }}}
This makes a particular difference for case-merge, which works
only if the scrutinee is the case-binder of the immediately enclosing
case (Note [Merge Nested Cases] in GHC.Core.Opt.Simplify.Utils
See #19581 for the bug report that showed this up.
(BS3) We need care when shadowing. Suppose [x :-> b] is in occ_bs_env,
and we encounter:
(i) \x. blah
Here we want to delete the x-binding from occ_bs_env
(ii) \b. blah
This is harder: we really want to delete all bindings that
have 'b' free in the range. That is a bit tiresome to implement,
so we compromise. We keep occ_bs_rng, which is the set of
free vars of rng(occc_bs_env). If a binder shadows any of these
variables, we discard all of occ_bs_env. Safe, if a bit
brutal. NB, however: the simplifer de-shadows the code, so the
next time around this won't happen.
These checks are implemented in addInScope.
(i) is needed only for Ids, but (ii) is needed for tyvars too (#22623)
because if occ_bs_env has [x :-> ...a...] where `a` is a tyvar, we
must not replace `x` by `...a...` under /\a. ...x..., or similarly
under a case pattern match that binds `a`.
An alternative would be for the occurrence analyser to do cloning as
it goes. In principle it could do so, but it'd make it a bit more
complicated and there is no great benefit. The simplifer uses
cloning to get a no-shadowing situation, the care-when-shadowing
behaviour above isn't needed for long.
(BS4) The domain of occ_bs_env can include GlobaIds. Eg
case M.foo of b { alts }
We extend occ_bs_env with [M.foo :-> b]. That's fine.
(BS5) We have to apply the occ_bs_env substitution uniformly,
including to (local) rules and unfoldings.
(BS6) We must be very careful with dictionaries.
See Note [Care with binder-swap on dictionaries]
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 [Care with binder-swap on dictionaries]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This Note explains why we need isDictId in scrutOkForBinderSwap.
Consider this tricky example (#21229, #21470):
class Sing (b :: Bool) where sing :: Bool
instance Sing 'True where sing = True
instance Sing 'False where sing = False
f :: forall a. Sing a => blah
h = \ @(a :: Bool) ($dSing :: Sing a)
let the_co = Main.N:Sing[0] <a> :: Sing a ~R# Bool
case ($dSing |> the_co) of wild
True -> f @'True (True |> sym the_co)
False -> f @a dSing
Now do a binder-swap on the case-expression:
h = \ @(a :: Bool) ($dSing :: Sing a)
let the_co = Main.N:Sing[0] <a> :: Sing a ~R# Bool
case ($dSing |> the_co) of wild
True -> f @'True (True |> sym the_co)
False -> f @a (wild |> sym the_co)
And now substitute `False` for `wild` (since wild=False in the False branch):
h = \ @(a :: Bool) ($dSing :: Sing a)
let the_co = Main.N:Sing[0] <a> :: Sing a ~R# Bool
case ($dSing |> the_co) of wild
True -> f @'True (True |> sym the_co)
False -> f @a (False |> sym the_co)
And now we have a problem. The specialiser will specialise (f @a d)a (for all
vtypes a and dictionaries d!!) with the dictionary (False |> sym the_co), using
Note [Specialising polymorphic dictionaries] in GHC.Core.Opt.Specialise.
The real problem is the binder-swap. It swaps a dictionary variable $dSing
(of kind Constraint) for a term variable wild (of kind Type). And that is
dangerous: a dictionary is a /singleton/ type whereas a general term variable is
not. In this particular example, Bool is most certainly not a singleton type!
Conclusion:
for a /dictionary variable/ do not perform
the clever cast version of the binder-swap
Hence the subtle isDictId in scrutOkForBinderSwap.
Note [Zap case binders in proxy bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From the original
case x of cb(dead) { p -> ...x... }
we will get
case x of cb(live) { p -> ...cb... }
Core Lint never expects to find an *occurrence* of an Id marked
as Dead, so we must zap the OccInfo on cb before making the
binding x = cb. See #5028.
NB: the OccInfo on /occurrences/ really doesn't matter much; the simplifier
doesn't use it. So this is only to satisfy the perhaps-over-picky Lint.
-}
addBndrSwap :: OutExpr -> Id -> OccEnv -> OccEnv
-- See Note [The binder-swap substitution]
addBndrSwap scrut case_bndr
env@(OccEnv { occ_bs_env = swap_env, occ_bs_rng = rng_vars })
| DoBinderSwap scrut_var mco <- scrutOkForBinderSwap scrut
, scrut_var /= case_bndr
-- Consider: case x of x { ... }
-- Do not add [x :-> x] to occ_bs_env, else lookupBndrSwap will loop
= env { occ_bs_env = extendVarEnv swap_env scrut_var (case_bndr', mco)
, occ_bs_rng = rng_vars `extendVarSet` case_bndr'
`unionVarSet` tyCoVarsOfMCo mco }
| otherwise
= env
where
case_bndr' = zapIdOccInfo case_bndr
-- See Note [Zap case binders in proxy bindings]
-- | See bBinderSwaOk.
data BinderSwapDecision
= NoBinderSwap
| DoBinderSwap OutVar MCoercion
scrutOkForBinderSwap :: OutExpr -> BinderSwapDecision
-- If (scrutOkForBinderSwap e = DoBinderSwap v mco, then
-- v = e |> mco
-- See Note [Case of cast]
-- See Note [Care with binder-swap on dictionaries]
--
-- We use this same function in SpecConstr, and Simplify.Iteration,
-- when something binder-swap-like is happening
scrutOkForBinderSwap (Var v) = DoBinderSwap v MRefl
scrutOkForBinderSwap (Cast (Var v) co)
| not (isDictId v) = DoBinderSwap v (MCo (mkSymCo co))
-- Cast: see Note [Case of cast]
-- isDictId: see Note [Care with binder-swap on dictionaries]
-- The isDictId rejects a Constraint/Constraint binder-swap, perhaps
-- over-conservatively. But I have never seen one, so I'm leaving
-- the code as simple as possible. Losing the binder-swap in a
-- rare case probably has very low impact.
scrutOkForBinderSwap (Tick _ e) = scrutOkForBinderSwap e -- Drop ticks
scrutOkForBinderSwap _ = NoBinderSwap
lookupBndrSwap :: OccEnv -> Id -> (CoreExpr, Id)
-- See Note [The binder-swap substitution]
-- Returns an expression of the same type as Id
lookupBndrSwap env@(OccEnv { occ_bs_env = bs_env }) bndr
= case lookupVarEnv bs_env bndr of {
Nothing -> (Var bndr, bndr) ;
Just (bndr1, mco) ->
-- Why do we iterate here?
-- See (BS2) in Note [The binder-swap substitution]
case lookupBndrSwap env bndr1 of
(fun, fun_id) -> (mkCastMCo fun mco, fun_id) }
{- Historical note [Proxy let-bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We used to do the binder-swap transformation by introducing
a proxy let-binding, thus;
case x of b { pi -> ri }
==>
case x of b { pi -> let x = b in ri }
But that had two problems:
1. If 'x' is an imported GlobalId, we'd end up with a GlobalId
on the LHS of a let-binding which isn't allowed. We worked
around this for a while by "localising" x, but it turned
out to be very painful #16296,
2. In CorePrep we use the occurrence analyser to do dead-code
elimination (see Note [Dead code in CorePrep]). But that
occasionally led to an unlifted let-binding
case x of b { DEFAULT -> let x::Int# = b in ... }
which disobeys one of CorePrep's output invariants (no unlifted
let-bindings) -- see #5433.
Doing a substitution (via occ_bs_env) is much better.
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.
************************************************************************
* *
\subsection[OccurAnal-types]{OccEnv}
* *
************************************************************************
Note [UsageDetails and zapping]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
On many occasions, we must modify all gathered occurrence data at once. For
instance, all occurrences underneath a (non-one-shot) lambda set the
'occ_in_lam' flag to become 'True'. We could use 'mapVarEnv' to do this, but
that takes O(n) time and we will do this often---in particular, there are many
places where tail calls are not allowed, and each of these causes all variables
to get marked with 'NoTailCallInfo'.
Instead of relying on `mapVarEnv`, then, we carry three 'IdEnv's around along
with the 'OccInfoEnv'. Each of these extra environments is a "zapped set"
recording which variables have been zapped in some way. Zapping all occurrence
info then simply means setting the corresponding zapped set to the whole
'OccInfoEnv', a fast O(1) operation.
Note [LocalOcc]
~~~~~~~~~~~~~~~
LocalOcc is used purely internally, in the occurrence analyser. It differs from
GHC.Types.Basic.OccInfo because it has only OneOcc and ManyOcc; it does not need
IAmDead or IAmALoopBreaker.
Note that `OneOccL` doesn't meant that it occurs /syntactially/ only once; it
means that it is /used/ only once. It might occur syntactically many times.
For example, in (case x of A -> y; B -> y; C -> True),
* `y` is used only once
* but it occurs syntactically twice
-}
type OccInfoEnv = IdEnv LocalOcc -- A finite map from an expression's
-- free variables to their usage
data LocalOcc -- See Note [LocalOcc]
= OneOccL { lo_n_br :: {-# UNPACK #-} !BranchCount -- Number of syntactic occurrences
, lo_tail :: !TailCallInfo
-- Combining (AlwaysTailCalled 2) and (AlwaysTailCalled 3)
-- gives NoTailCallInfo
, lo_int_cxt :: !InterestingCxt }
| ManyOccL !TailCallInfo
instance Outputable LocalOcc where
ppr (OneOccL { lo_n_br = n, lo_tail = tci })
= text "OneOccL" <> braces (ppr n <> comma <> ppr tci)
ppr (ManyOccL tci) = text "ManyOccL" <> braces (ppr tci)
localTailCallInfo :: LocalOcc -> TailCallInfo
localTailCallInfo (OneOccL { lo_tail = tci }) = tci
localTailCallInfo (ManyOccL tci) = tci
type ZappedSet = OccInfoEnv -- Values are ignored
data UsageDetails
= UD { ud_env :: !OccInfoEnv
, ud_z_many :: !ZappedSet -- apply 'markMany' to these
, ud_z_in_lam :: !ZappedSet -- apply 'markInsideLam' to these
, ud_z_tail :: !ZappedSet -- zap tail-call info for these
}
-- INVARIANT: All three zapped sets are subsets of ud_env
instance Outputable UsageDetails where
ppr ud@(UD { ud_env = env, ud_z_tail = z_tail })
= text "UD" <+> (braces $ fsep $ punctuate comma $
[ ppr uq <+> text ":->" <+> ppr (lookupOccInfoByUnique ud uq)
| (uq, _) <- nonDetStrictFoldVarEnv_Directly do_one [] env ])
$$ nest 2 (text "ud_z_tail" <+> ppr z_tail)
where
do_one :: Unique -> LocalOcc -> [(Unique,LocalOcc)] -> [(Unique,LocalOcc)]
do_one uniq occ occs = (uniq, occ) : occs
---------------------
-- | TailUsageDetails captures the result of applying 'occAnalLamTail'
-- to a function `\xyz.body`. The TailUsageDetails pairs together
-- * the number of lambdas (including type lambdas: a JoinArity)
-- * UsageDetails for the `body` of the lambda, unadjusted by `adjustTailUsage`.
-- If the binding turns out to be a join point with the indicated join
-- arity, this unadjusted usage details is just what we need; otherwise we
-- need to discard tail calls. That's what `adjustTailUsage` does.
data TailUsageDetails = TUD !JoinArity !UsageDetails
instance Outputable TailUsageDetails where
ppr (TUD ja uds) = lambda <> ppr ja <> ppr uds
---------------------
data WithUsageDetails a = WUD !UsageDetails !a
data WithTailUsageDetails a = WTUD !TailUsageDetails !a
-------------------
-- UsageDetails API
andUDs:: UsageDetails -> UsageDetails -> UsageDetails
orUDs :: UsageDetails -> UsageDetails -> UsageDetails
andUDs = combineUsageDetailsWith andLocalOcc
orUDs = combineUsageDetailsWith orLocalOcc
mkOneOcc :: OccEnv -> Id -> InterestingCxt -> JoinArity -> UsageDetails
mkOneOcc !env id int_cxt arity
| not (isLocalId id)
= emptyDetails
| Just join_uds <- lookupVarEnv (occ_join_points env) id
= -- See Note [Occurrence analysis for join points]
assertPpr (not (isEmptyVarEnv join_uds)) (ppr id) $
-- We only put non-empty join-points into occ_join_points
mkSimpleDetails (extendVarEnv join_uds id occ)
| otherwise
= mkSimpleDetails (unitVarEnv id occ)
where
occ = OneOccL { lo_n_br = 1, lo_int_cxt = int_cxt
, lo_tail = AlwaysTailCalled arity }
-- Add several occurrences, assumed not to be tail calls
add_many_occ :: Var -> OccInfoEnv -> OccInfoEnv
add_many_occ v env | isId v = extendVarEnv env v (ManyOccL NoTailCallInfo)
| otherwise = env
-- Give a non-committal binder info (i.e noOccInfo) because
-- a) Many copies of the specialised thing can appear
-- b) We don't want to substitute a BIG expression inside a RULE
-- even if that's the only occurrence of the thing
-- (Same goes for INLINE.)
addManyOccs :: UsageDetails -> VarSet -> UsageDetails
addManyOccs uds var_set
| isEmptyVarSet var_set = uds
| otherwise = uds { ud_env = add_to (ud_env uds) }
where
add_to env = nonDetStrictFoldUniqSet add_many_occ env var_set
-- It's OK to use nonDetStrictFoldUniqSet here because add_many_occ commutes
addLamCoVarOccs :: UsageDetails -> [Var] -> UsageDetails
-- Add any CoVars free in the type of a lambda-binder
-- See Note [Gather occurrences of coercion variables]
addLamCoVarOccs uds bndrs
= foldr add uds bndrs
where
add bndr uds = uds `addManyOccs` coVarsOfType (varType bndr)
emptyDetails :: UsageDetails
emptyDetails = mkSimpleDetails emptyVarEnv
isEmptyDetails :: UsageDetails -> Bool
isEmptyDetails (UD { ud_env = env }) = isEmptyVarEnv env
mkSimpleDetails :: OccInfoEnv -> UsageDetails
mkSimpleDetails env = UD { ud_env = env
, ud_z_many = emptyVarEnv
, ud_z_in_lam = emptyVarEnv
, ud_z_tail = emptyVarEnv }
modifyUDEnv :: (OccInfoEnv -> OccInfoEnv) -> UsageDetails -> UsageDetails
modifyUDEnv f uds@(UD { ud_env = env }) = uds { ud_env = f env }
delBndrsFromUDs :: [Var] -> UsageDetails -> UsageDetails
-- Delete these binders from the UsageDetails
delBndrsFromUDs bndrs (UD { ud_env = env, ud_z_many = z_many
, ud_z_in_lam = z_in_lam, ud_z_tail = z_tail })
= UD { ud_env = env `delVarEnvList` bndrs
, ud_z_many = z_many `delVarEnvList` bndrs
, ud_z_in_lam = z_in_lam `delVarEnvList` bndrs
, ud_z_tail = z_tail `delVarEnvList` bndrs }
markAllMany, markAllInsideLam, markAllNonTail, markAllManyNonTail
:: UsageDetails -> UsageDetails
markAllMany ud@(UD { ud_env = env }) = ud { ud_z_many = env }
markAllInsideLam ud@(UD { ud_env = env }) = ud { ud_z_in_lam = env }
markAllNonTail ud@(UD { ud_env = env }) = ud { ud_z_tail = env }
markAllManyNonTail = markAllMany . markAllNonTail -- effectively sets to noOccInfo
markAllInsideLamIf, markAllNonTailIf :: Bool -> UsageDetails -> UsageDetails
markAllInsideLamIf True ud = markAllInsideLam ud
markAllInsideLamIf False ud = ud
markAllNonTailIf True ud = markAllNonTail ud
markAllNonTailIf False ud = ud
lookupTailCallInfo :: UsageDetails -> Id -> TailCallInfo
lookupTailCallInfo uds id
| UD { ud_z_tail = z_tail, ud_env = env } <- uds
, not (id `elemVarEnv` z_tail)
, Just occ <- lookupVarEnv env id
= localTailCallInfo occ
| otherwise
= NoTailCallInfo
udFreeVars :: VarSet -> UsageDetails -> VarSet
-- Find the subset of bndrs that are mentioned in uds
udFreeVars bndrs (UD { ud_env = env }) = restrictFreeVars bndrs env
restrictFreeVars :: VarSet -> OccInfoEnv -> VarSet
restrictFreeVars bndrs fvs = restrictUniqSetToUFM bndrs fvs
-------------------
-- Auxiliary functions for UsageDetails implementation
combineUsageDetailsWith :: (LocalOcc -> LocalOcc -> LocalOcc)
-> UsageDetails -> UsageDetails -> UsageDetails
{-# INLINE combineUsageDetailsWith #-}
combineUsageDetailsWith plus_occ_info
uds1@(UD { ud_env = env1, ud_z_many = z_many1, ud_z_in_lam = z_in_lam1, ud_z_tail = z_tail1 })
uds2@(UD { ud_env = env2, ud_z_many = z_many2, ud_z_in_lam = z_in_lam2, ud_z_tail = z_tail2 })
| isEmptyVarEnv env1 = uds2
| isEmptyVarEnv env2 = uds1
| otherwise
-- See Note [Strictness in the occurrence analyser]
-- Using strictPlusVarEnv here speeds up the test T26425 by about 10% by avoiding
-- intermediate thunks.
= UD { ud_env = strictPlusVarEnv_C plus_occ_info env1 env2
, ud_z_many = strictPlusVarEnv z_many1 z_many2
, ud_z_in_lam = plusVarEnv z_in_lam1 z_in_lam2
, ud_z_tail = strictPlusVarEnv z_tail1 z_tail2 }
lookupLetOccInfo :: UsageDetails -> Id -> OccInfo
-- Don't use locally-generated occ_info for exported (visible-elsewhere)
-- things. Instead just give noOccInfo.
-- NB: setBinderOcc will (rightly) erase any LoopBreaker info;
-- we are about to re-generate it and it shouldn't be "sticky"
lookupLetOccInfo ud id
| isExportedId id = noOccInfo
| otherwise = lookupOccInfoByUnique ud (idUnique id)
lookupOccInfo :: UsageDetails -> Id -> OccInfo
lookupOccInfo ud id = lookupOccInfoByUnique ud (idUnique id)
lookupOccInfoByUnique :: UsageDetails -> Unique -> OccInfo
lookupOccInfoByUnique (UD { ud_env = env
, ud_z_many = z_many
, ud_z_in_lam = z_in_lam
, ud_z_tail = z_tail })
uniq
= case lookupVarEnv_Directly env uniq of
Nothing -> IAmDead
Just (OneOccL { lo_n_br = n_br, lo_int_cxt = int_cxt
, lo_tail = tail_info })
| uniq `elemVarEnvByKey`z_many
-> ManyOccs { occ_tail = mk_tail_info tail_info }
| otherwise
-> OneOcc { occ_in_lam = in_lam
, occ_n_br = n_br
, occ_int_cxt = int_cxt
, occ_tail = mk_tail_info tail_info }
where
in_lam | uniq `elemVarEnvByKey` z_in_lam = IsInsideLam
| otherwise = NotInsideLam
Just (ManyOccL tail_info) -> ManyOccs { occ_tail = mk_tail_info tail_info }
where
mk_tail_info ti
| uniq `elemVarEnvByKey` z_tail = NoTailCallInfo
| otherwise = ti
-------------------
-- See Note [Adjusting right-hand sides]
adjustNonRecRhs :: JoinPointHood
-> WithTailUsageDetails CoreExpr
-> WithUsageDetails CoreExpr
-- ^ This function concentrates shared logic between occAnalNonRecBind and the
-- AcyclicSCC case of occAnalRec.
-- It returns the adjusted rhs UsageDetails combined with the body usage
adjustNonRecRhs mb_join_arity rhs_wuds@(WTUD _ rhs)
= WUD (adjustTailUsage mb_join_arity rhs_wuds) rhs
adjustTailUsage :: JoinPointHood
-> WithTailUsageDetails CoreExpr -- Rhs usage, AFTER occAnalLamTail
-> UsageDetails
adjustTailUsage mb_join_arity (WTUD (TUD rhs_ja uds) rhs)
= -- c.f. occAnal (Lam {})
markAllInsideLamIf (not one_shot) $
markAllNonTailIf (not exact_join) $
uds
where
one_shot = isOneShotFun rhs
exact_join = mb_join_arity == JoinPoint rhs_ja
adjustTailArity :: JoinPointHood -> TailUsageDetails -> UsageDetails
adjustTailArity mb_rhs_ja (TUD ja usage)
= markAllNonTailIf (mb_rhs_ja /= JoinPoint ja) usage
type IdWithOccInfo = Id
tagLamBinders :: UsageDetails -- Of scope
-> [Id] -- Binders
-> [IdWithOccInfo] -- Tagged binders
tagLamBinders usage binders
-- See Note [Strictness in the occurrence analyser]
= strictMap (tagLamBinder usage) binders
tagLamBinder :: UsageDetails -- Of scope
-> Id -- Binder
-> IdWithOccInfo -- Tagged binders
-- Used for lambda and case binders
-- No-op on TyVars
-- A lambda binder never has an unfolding, so no need to look for that
tagLamBinder usage bndr
-- See Note [Strictness in the occurrence analyser]
= setBinderOcc (markNonTail occ) bndr
-- markNonTail: don't try to make an argument into a join point
where
occ = lookupOccInfo usage bndr
tagNonRecBinder :: TopLevelFlag -- At top level?
-> OccInfo -- Of scope
-> CoreBndr -- Binder
-> (IdWithOccInfo, JoinPointHood) -- Tagged binder
-- No-op on TyVars
-- Precondition: OccInfo is not IAmDead
tagNonRecBinder lvl occ bndr
| okForJoinPoint lvl bndr tail_call_info
, AlwaysTailCalled ar <- tail_call_info
= (setBinderOcc occ bndr, JoinPoint ar)
| otherwise
= (setBinderOcc zapped_occ bndr, NotJoinPoint)
where
tail_call_info = tailCallInfo occ
zapped_occ = markNonTail occ
tagRecBinders :: TopLevelFlag -- At top level?
-> UsageDetails -- Of body of let ONLY
-> [NodeDetails]
-> WithUsageDetails -- Adjusted details for whole scope,
-- with binders removed
[IdWithOccInfo] -- Tagged binders
-- Substantially more complicated than non-recursive case. Need to adjust RHS
-- details *before* tagging binders (because the tags depend on the RHSes).
tagRecBinders lvl body_uds details_s
= let
bndrs = map nd_bndr details_s
-- 1. See Note [Join arity prediction based on joinRhsArity]
-- Determine possible join-point-hood of whole group, by testing for
-- manifest join arity M.
-- This (re-)asserts that makeNode had made tuds for that same arity M!
unadj_uds = foldr (andUDs . test_manifest_arity) body_uds details_s
test_manifest_arity ND{nd_rhs = WTUD tuds rhs}
= adjustTailArity (JoinPoint (joinRhsArity rhs)) tuds
will_be_joins = decideRecJoinPointHood lvl unadj_uds bndrs
mb_join_arity :: Id -> JoinPointHood
-- mb_join_arity: See Note [Join arity prediction based on joinRhsArity]
-- This is the source O
mb_join_arity bndr
-- Can't use willBeJoinId_maybe here because we haven't tagged
-- the binder yet (the tag depends on these adjustments!)
| will_be_joins
, AlwaysTailCalled arity <- lookupTailCallInfo unadj_uds bndr
= JoinPoint arity
| otherwise
= assert (not will_be_joins) -- Should be AlwaysTailCalled if
NotJoinPoint -- we are making join points!
-- 2. Adjust usage details of each RHS, taking into account the
-- join-point-hood decision
rhs_udss' = [ adjustTailUsage (mb_join_arity bndr) rhs_wuds
-- Matching occAnalLamTail in makeNode
| ND { nd_bndr = bndr, nd_rhs = rhs_wuds } <- details_s ]
-- 3. Compute final usage details from adjusted RHS details
adj_uds = foldr andUDs body_uds rhs_udss'
-- 4. Tag each binder with its adjusted details
bndrs' = [ setBinderOcc (lookupLetOccInfo adj_uds bndr) bndr
| bndr <- bndrs ]
in
WUD adj_uds bndrs'
setBinderOcc :: OccInfo -> CoreBndr -> CoreBndr
setBinderOcc occ_info bndr
| isTyVar bndr = bndr
| occ_info == idOccInfo bndr = bndr
| otherwise = setIdOccInfo bndr occ_info
-- | Decide whether some bindings should be made into join points or not, based
-- on its occurrences. This is
-- Returns `False` if they can't be join points. Note that it's an
-- all-or-nothing decision, as if multiple binders are given, they're
-- assumed to be mutually recursive.
--
-- It must, however, be a final decision. If we say `True` for 'f',
-- and then subsequently decide /not/ make 'f' into a join point, then
-- the decision about another binding 'g' might be invalidated if (say)
-- 'f' tail-calls 'g'.
--
-- See Note [Invariants on join points] in "GHC.Core".
decideRecJoinPointHood :: TopLevelFlag -> UsageDetails
-> [CoreBndr] -> Bool
decideRecJoinPointHood lvl usage bndrs
= all ok bndrs -- Invariant 3: Either all are join points or none are
where
ok bndr = okForJoinPoint lvl bndr (lookupTailCallInfo usage bndr)
okForJoinPoint :: TopLevelFlag -> Id -> TailCallInfo -> Bool
-- See Note [Invariants on join points]; invariants cited by number below.
-- Invariant 2 is always satisfiable by the simplifier by eta expansion.
okForJoinPoint lvl bndr tail_call_info
| isJoinId bndr -- A current join point should still be one!
= warnPprTrace lost_join "Lost join point" lost_join_doc $
True
| valid_join
= True
| otherwise
= False
where
valid_join | NotTopLevel <- lvl
, AlwaysTailCalled arity <- tail_call_info
, -- Invariant 1 as applied to LHSes of rules
all (ok_rule arity) (idCoreRules bndr)
-- Invariant 2a: stable unfoldings
-- See Note [Join points and INLINE pragmas]
, ok_unfolding arity (realIdUnfolding bndr)
-- Invariant 4: Satisfies polymorphism rule
, isValidJoinPointType arity (idType bndr)
= True
| otherwise
= False
lost_join | JoinPoint ja <- idJoinPointHood bndr
= not valid_join ||
(case tail_call_info of -- Valid join but arity differs
AlwaysTailCalled ja' -> ja /= ja'
_ -> False)
| otherwise = False
ok_rule _ BuiltinRule{} = False -- only possible with plugin shenanigans
ok_rule join_arity (Rule { ru_args = args })
= args `lengthIs` join_arity
-- Invariant 1 as applied to LHSes of rules
-- ok_unfolding returns False if we should /not/ convert a non-join-id
-- into a join-id, even though it is AlwaysTailCalled
ok_unfolding join_arity (CoreUnfolding { uf_src = src, uf_tmpl = rhs })
= not (isStableSource src && join_arity > joinRhsArity rhs)
ok_unfolding _ (DFunUnfolding {})
= False
ok_unfolding _ _
= True
lost_join_doc
= vcat [ text "bndr:" <+> ppr bndr
, text "tc:" <+> ppr tail_call_info
, text "rules:" <+> ppr (idCoreRules bndr)
, case tail_call_info of
AlwaysTailCalled arity ->
vcat [ text "ok_unf:" <+> ppr (ok_unfolding arity (realIdUnfolding bndr))
, text "ok_type:" <+> ppr (isValidJoinPointType arity (idType bndr)) ]
_ -> empty ]
{- Note [Join points and INLINE pragmas]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
f x = let g = \x. not -- Arity 1
{-# INLINE g #-}
in case x of
A -> g True True
B -> g True False
C -> blah2
Here 'g' is always tail-called applied to 2 args, but the stable
unfolding captured by the INLINE pragma has arity 1. If we try to
convert g to be a join point, its unfolding will still have arity 1
(since it is stable, and we don't meddle with stable unfoldings), and
Lint will complain (see Note [Invariants on join points], (2a), in
GHC.Core. #13413.
Moreover, since g is going to be inlined anyway, there is no benefit
from making it a join point.
If it is recursive, and uselessly marked INLINE, this will stop us
making it a join point, which is annoying. But occasionally
(notably in class methods; see Note [Instances and loop breakers] in
GHC.Tc.TyCl.Instance) we mark recursive things as INLINE but the recursion
unravels; so ignoring INLINE pragmas on recursive things isn't good
either.
See Invariant 2a of Note [Invariants on join points] in GHC.Core
************************************************************************
* *
\subsection{Operations over OccInfo}
* *
************************************************************************
-}
markNonTail :: OccInfo -> OccInfo
markNonTail IAmDead = IAmDead
markNonTail occ = occ { occ_tail = NoTailCallInfo }
andLocalOcc :: LocalOcc -> LocalOcc -> LocalOcc
andLocalOcc occ1 occ2 = ManyOccL (tci1 `andTailCallInfo` tci2)
where
!tci1 = localTailCallInfo occ1
!tci2 = localTailCallInfo occ2
orLocalOcc :: LocalOcc -> LocalOcc -> LocalOcc
-- (orLocalOcc occ1 occ2) is used
-- when combining occurrence info from branches of a case
orLocalOcc (OneOccL { lo_n_br = nbr1, lo_int_cxt = int_cxt1, lo_tail = tci1 })
(OneOccL { lo_n_br = nbr2, lo_int_cxt = int_cxt2, lo_tail = tci2 })
= OneOccL { lo_n_br = nbr1 + nbr2
, lo_int_cxt = int_cxt1 `mappend` int_cxt2
, lo_tail = tci1 `andTailCallInfo` tci2 }
orLocalOcc occ1 occ2 = andLocalOcc occ1 occ2
andTailCallInfo :: TailCallInfo -> TailCallInfo -> TailCallInfo
andTailCallInfo info@(AlwaysTailCalled arity1) (AlwaysTailCalled arity2)
| arity1 == arity2 = info
andTailCallInfo _ _ = NoTailCallInfo