%
% (c) The AQUA Project, Glasgow University, 1993-1998
%
\section[Simplify]{The main module of the simplifier}
\begin{code}
module Simplify ( simplTopBinds, simplExpr ) where
#include "HsVersions.h"
import DynFlags
import SimplMonad
import Type hiding ( substTy, extendTvSubst, substTyVar )
import SimplEnv
import SimplUtils
import FamInstEnv ( FamInstEnv )
import Id
import MkId ( seqId, realWorldPrimId )
import MkCore ( mkImpossibleExpr )
import IdInfo
import Name ( mkSystemVarName, isExternalName )
import Coercion hiding ( substCo, substTy, substCoVar, extendTvSubst )
import OptCoercion ( optCoercion )
import FamInstEnv ( topNormaliseType )
import DataCon ( DataCon, dataConWorkId, dataConRepStrictness )
import CoreMonad ( Tick(..), SimplifierMode(..) )
import CoreSyn
import Demand ( isStrictDmd )
import PprCore ( pprParendExpr, pprCoreExpr )
import CoreUnfold
import CoreUtils
import qualified CoreSubst
import CoreArity
import Rules ( lookupRule, getRules )
import BasicTypes ( isMarkedStrict, Arity )
import CostCentre ( currentCCS, pushCCisNop )
import TysPrim ( realWorldStatePrimTy )
import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
import MonadUtils ( foldlM, mapAccumLM )
import Maybes ( orElse, isNothing )
import Data.List ( mapAccumL )
import Outputable
import FastString
import Pair
\end{code}
The guts of the simplifier is in this module, but the driver loop for
the simplifier is in SimplCore.lhs.
-----------------------------------------
*** IMPORTANT NOTE ***
-----------------------------------------
The simplifier used to guarantee that the output had no shadowing, but
it does not do so any more. (Actually, it never did!) The reason is
documented with simplifyArgs.
-----------------------------------------
*** IMPORTANT NOTE ***
-----------------------------------------
Many parts of the simplifier return a bunch of "floats" as well as an
expression. This is wrapped as a datatype SimplUtils.FloatsWith.
All "floats" are let-binds, not case-binds, but some non-rec lets may
be unlifted (with RHS ok-for-speculation).
-----------------------------------------
ORGANISATION OF FUNCTIONS
-----------------------------------------
simplTopBinds
- simplify all top-level binders
- for NonRec, call simplRecOrTopPair
- for Rec, call simplRecBind
------------------------------
simplExpr (applied lambda) ==> simplNonRecBind
simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
------------------------------
simplRecBind [binders already simplfied]
- use simplRecOrTopPair on each pair in turn
simplRecOrTopPair [binder already simplified]
Used for: recursive bindings (top level and nested)
top-level non-recursive bindings
Returns:
- check for PreInlineUnconditionally
- simplLazyBind
simplNonRecBind
Used for: non-top-level non-recursive bindings
beta reductions (which amount to the same thing)
Because it can deal with strict arts, it takes a
"thing-inside" and returns an expression
- check for PreInlineUnconditionally
- simplify binder, including its IdInfo
- if strict binding
simplStrictArg
mkAtomicArgs
completeNonRecX
else
simplLazyBind
addFloats
simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
Used for: binding case-binder and constr args in a known-constructor case
- check for PreInLineUnconditionally
- simplify binder
- completeNonRecX
------------------------------
simplLazyBind: [binder already simplified, RHS not]
Used for: recursive bindings (top level and nested)
top-level non-recursive bindings
non-top-level, but *lazy* non-recursive bindings
[must not be strict or unboxed]
Returns floats + an augmented environment, not an expression
- substituteIdInfo and add result to in-scope
[so that rules are available in rec rhs]
- simplify rhs
- mkAtomicArgs
- float if exposes constructor or PAP
- completeBind
completeNonRecX: [binder and rhs both simplified]
- if the the thing needs case binding (unlifted and not ok-for-spec)
build a Case
else
completeBind
addFloats
completeBind: [given a simplified RHS]
[used for both rec and non-rec bindings, top level and not]
- try PostInlineUnconditionally
- add unfolding [this is the only place we add an unfolding]
- add arity
Right hand sides and arguments
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In many ways we want to treat
(a) the right hand side of a let(rec), and
(b) a function argument
in the same way. But not always! In particular, we would
like to leave these arguments exactly as they are, so they
will match a RULE more easily.
f (g x, h x)
g (+ x)
It's harder to make the rule match if we ANF-ise the constructor,
or eta-expand the PAP:
f (let { a = g x; b = h x } in (a,b))
g (\y. + x y)
On the other hand if we see the let-defns
p = (g x, h x)
q = + x
then we *do* want to ANF-ise and eta-expand, so that p and q
can be safely inlined.
Even floating lets out is a bit dubious. For let RHS's we float lets
out if that exposes a value, so that the value can be inlined more vigorously.
For example
r = let x = e in (x,x)
Here, if we float the let out we'll expose a nice constructor. We did experiments
that showed this to be a generally good thing. But it was a bad thing to float
lets out unconditionally, because that meant they got allocated more often.
For function arguments, there's less reason to expose a constructor (it won't
get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
So for the moment we don't float lets out of function arguments either.
Eta expansion
~~~~~~~~~~~~~~
For eta expansion, we want to catch things like
case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
If the \x was on the RHS of a let, we'd eta expand to bring the two
lambdas together. And in general that's a good thing to do. Perhaps
we should eta expand wherever we find a (value) lambda? Then the eta
expansion at a let RHS can concentrate solely on the PAP case.
%************************************************************************
%* *
\subsection{Bindings}
%* *
%************************************************************************
\begin{code}
simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
simplTopBinds env0 binds0
= do {
; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
; dflags <- getDOptsSmpl
; let dump_flag = dopt Opt_D_verbose_core2core dflags
; env2 <- simpl_binds dump_flag env1 binds0
; freeTick SimplifierDone
; return env2 }
where
simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
simpl_binds _ env [] = return env
simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
simpl_bind env bind
; simpl_binds dump env' binds }
trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
trace_bind False _ = \x -> x
simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel NonRecursive b b' r
where
(env', b') = addBndrRules env b (lookupRecBndr env b)
\end{code}
%************************************************************************
%* *
\subsection{Lazy bindings}
%* *
%************************************************************************
simplRecBind is used for
* recursive bindings only
\begin{code}
simplRecBind :: SimplEnv -> TopLevelFlag
-> [(InId, InExpr)]
-> SimplM SimplEnv
simplRecBind env0 top_lvl pairs0
= do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
; env1 <- go (zapFloats env_with_info) triples
; return (env0 `addRecFloats` env1) }
where
add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
where
(env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
go env [] = return env
go env ((old_bndr, new_bndr, rhs) : pairs)
= do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
; go env' pairs }
\end{code}
simplOrTopPair is used for
* recursive bindings (whether top level or not)
* top-level non-recursive bindings
It assumes the binder has already been simplified, but not its IdInfo.
\begin{code}
simplRecOrTopPair :: SimplEnv
-> TopLevelFlag -> RecFlag
-> InId -> OutBndr -> InExpr
-> SimplM SimplEnv
simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
| preInlineUnconditionally env top_lvl old_bndr rhs
= do { tick (PreInlineUnconditionally old_bndr)
; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
| otherwise
= simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env
\end{code}
simplLazyBind is used for
* [simplRecOrTopPair] recursive bindings (whether top level or not)
* [simplRecOrTopPair] top-level non-recursive bindings
* [simplNonRecE] non-top-level *lazy* non-recursive bindings
Nota bene:
1. It assumes that the binder is *already* simplified,
and is in scope, and its IdInfo too, except unfolding
2. It assumes that the binder type is lifted.
3. It does not check for pre-inline-unconditionallly;
that should have been done already.
\begin{code}
simplLazyBind :: SimplEnv
-> TopLevelFlag -> RecFlag
-> InId -> OutId
-> InExpr -> SimplEnv
-> SimplM SimplEnv
simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
=
do { let rhs_env = rhs_se `setInScope` env
(tvs, body) = case collectTyBinders rhs of
(tvs, body) | not_lam body -> (tvs,body)
| otherwise -> ([], rhs)
not_lam (Lam _ _) = False
not_lam _ = True
; (body_env, tvs') <- simplBinders rhs_env tvs
; (body_env1, body1) <- simplExprF body_env body mkRhsStop
; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
; (env', rhs')
<- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
then
do { rhs' <- mkLam env tvs' (wrapFloats body_env1 body1)
; return (env, rhs') }
else if null tvs then
do { tick LetFloatFromLet
; return (addFloats env body_env2, body2) }
else
do { tick LetFloatFromLet
; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
; rhs' <- mkLam env tvs' body3
; env' <- foldlM (addPolyBind top_lvl) env poly_binds
; return (env', rhs') }
; completeBind env' top_lvl bndr bndr1 rhs' }
\end{code}
A specialised variant of simplNonRec used when the RHS is already simplified,
notably in knownCon. It uses case-binding where necessary.
\begin{code}
simplNonRecX :: SimplEnv
-> InId
-> OutExpr
-> SimplM SimplEnv
simplNonRecX env bndr new_rhs
| isDeadBinder bndr
= return env
| Coercion co <- new_rhs
= return (extendCvSubst env bndr co)
| otherwise
= do { (env', bndr') <- simplBinder env bndr
; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
completeNonRecX :: TopLevelFlag -> SimplEnv
-> Bool
-> InId
-> OutId
-> OutExpr
-> SimplM SimplEnv
completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
= do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
; (env2, rhs2) <-
if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
then do { tick LetFloatFromLet
; return (addFloats env env1, rhs1) }
else return (env, wrapFloats env1 rhs1)
; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
\end{code}
{- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
Doing so risks exponential behaviour, because new_rhs has been simplified once already
In the cases described by the folowing commment, postInlineUnconditionally will
catch many of the relevant cases.
-- This happens; for example, the case_bndr during case of
-- known constructor: case (a,b) of x { (p,q) -> ... }
-- Here x isn't mentioned in the RHS, so we don't want to
-- create the (dead) let-binding let x = (a,b) in ...
--
-- Similarly, single occurrences can be inlined vigourously
-- e.g. case (f x, g y) of (a,b) -> ....
-- If a,b occur once we can avoid constructing the let binding for them.
Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
-- Consider case I# (quotInt# x y) of
-- I# v -> let w = J# v in ...
-- If we gaily inline (quotInt# x y) for v, we end up building an
-- extra thunk:
-- let w = J# (quotInt# x y) in ...
-- because quotInt# can fail.
| preInlineUnconditionally env NotTopLevel bndr new_rhs
= thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
-}
----------------------------------
prepareRhs takes a putative RHS, checks whether it's a PAP or
constructor application and, if so, converts it to ANF, so that the
resulting thing can be inlined more easily. Thus
x = (f a, g b)
becomes
t1 = f a
t2 = g b
x = (t1,t2)
We also want to deal well cases like this
v = (f e1 `cast` co) e2
Here we want to make e1,e2 trivial and get
x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
That's what the 'go' loop in prepareRhs does
\begin{code}
prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
prepareRhs top_lvl env id (Cast rhs co)
| Pair ty1 _ty2 <- coercionKind co
, not (isUnLiftedType ty1)
= do { (env', rhs') <- makeTrivialWithInfo top_lvl env sanitised_info rhs
; return (env', Cast rhs' co) }
where
sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
`setDemandInfo` demandInfo info
info = idInfo id
prepareRhs top_lvl env0 _ rhs0
= do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
; return (env1, rhs1) }
where
go n_val_args env (Cast rhs co)
= do { (is_exp, env', rhs') <- go n_val_args env rhs
; return (is_exp, env', Cast rhs' co) }
go n_val_args env (App fun (Type ty))
= do { (is_exp, env', rhs') <- go n_val_args env fun
; return (is_exp, env', App rhs' (Type ty)) }
go n_val_args env (App fun arg)
= do { (is_exp, env', fun') <- go (n_val_args+1) env fun
; case is_exp of
True -> do { (env'', arg') <- makeTrivial top_lvl env' arg
; return (True, env'', App fun' arg') }
False -> return (False, env, App fun arg) }
go n_val_args env (Var fun)
= return (is_exp, env, Var fun)
where
is_exp = isExpandableApp fun n_val_args
go _ env other
= return (False, env, other)
\end{code}
Note [Float coercions]
~~~~~~~~~~~~~~~~~~~~~~
When we find the binding
x = e `cast` co
we'd like to transform it to
x' = e
x = x `cast` co -- A trivial binding
There's a chance that e will be a constructor application or function, or something
like that, so moving the coerion to the usage site may well cancel the coersions
and lead to further optimisation. Example:
data family T a :: *
data instance T Int = T Int
foo :: Int -> Int -> Int
foo m n = ...
where
x = T m
go 0 = 0
go n = case x of { T m -> go (n-m) }
-- This case should optimise
Note [Preserve strictness when floating coercions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In the Note [Float coercions] transformation, keep the strictness info.
Eg
f = e `cast` co -- f has strictness SSL
When we transform to
f' = e -- f' also has strictness SSL
f = f' `cast` co -- f still has strictness SSL
Its not wrong to drop it on the floor, but better to keep it.
Note [Float coercions (unlifted)]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
BUT don't do [Float coercions] if 'e' has an unlifted type.
This *can* happen:
foo :: Int = (error (# Int,Int #) "urk")
`cast` CoUnsafe (# Int,Int #) Int
If do the makeTrivial thing to the error call, we'll get
foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
But 'v' isn't in scope!
These strange casts can happen as a result of case-of-case
bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
(# p,q #) -> p+q
\begin{code}
makeTrivial :: TopLevelFlag -> SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
makeTrivial top_lvl env expr = makeTrivialWithInfo top_lvl env vanillaIdInfo expr
makeTrivialWithInfo :: TopLevelFlag -> SimplEnv -> IdInfo
-> OutExpr -> SimplM (SimplEnv, OutExpr)
makeTrivialWithInfo top_lvl env info expr
| exprIsTrivial expr
|| not (bindingOk top_lvl expr expr_ty)
= return (env, expr)
| otherwise
= do { uniq <- getUniqueM
; let name = mkSystemVarName uniq (fsLit "a")
var = mkLocalIdWithInfo name expr_ty info
; env' <- completeNonRecX top_lvl env False var var expr
; expr' <- simplVar env' var
; return (env', expr') }
where
expr_ty = exprType expr
bindingOk :: TopLevelFlag -> CoreExpr -> Type -> Bool
bindingOk top_lvl _ expr_ty
| isTopLevel top_lvl = not (isUnLiftedType expr_ty)
| otherwise = True
\end{code}
Note [Cannot trivialise]
~~~~~~~~~~~~~~~~~~~~~~~~
Consider tih
f :: Int -> Addr#
foo :: Bar
foo = Bar (f 3)
Then we can't ANF-ise foo, even though we'd like to, because
we can't make a top-level binding for the Addr# (f 3). And if
so we don't want to turn it into
foo = let x = f 3 in Bar x
because we'll just end up inlining x back, and that makes the
simplifier loop. Better not to ANF-ise it at all.
A case in point is literal strings (a MachStr is not regarded as
trivial):
foo = Ptr "blob"#
We don't want to ANF-ise this.
%************************************************************************
%* *
\subsection{Completing a lazy binding}
%* *
%************************************************************************
completeBind
* deals only with Ids, not TyVars
* takes an already-simplified binder and RHS
* is used for both recursive and non-recursive bindings
* is used for both top-level and non-top-level bindings
It does the following:
- tries discarding a dead binding
- tries PostInlineUnconditionally
- add unfolding [this is the only place we add an unfolding]
- add arity
It does *not* attempt to do let-to-case. Why? Because it is used for
- top-level bindings (when let-to-case is impossible)
- many situations where the "rhs" is known to be a WHNF
(so let-to-case is inappropriate).
Nor does it do the atomic-argument thing
\begin{code}
completeBind :: SimplEnv
-> TopLevelFlag
-> InId
-> OutId -> OutExpr
-> SimplM SimplEnv
completeBind env top_lvl old_bndr new_bndr new_rhs
| isCoVar old_bndr
= case new_rhs of
Coercion co -> return (extendCvSubst env old_bndr co)
_ -> return (addNonRec env new_bndr new_rhs)
| otherwise
= ASSERT( isId new_bndr )
do { let old_info = idInfo old_bndr
old_unf = unfoldingInfo old_info
occ_info = occInfo old_info
; (new_arity, final_rhs) <- tryEtaExpand env new_bndr new_rhs
; new_unfolding <- simplUnfolding env top_lvl old_bndr final_rhs old_unf
; if postInlineUnconditionally env top_lvl new_bndr occ_info final_rhs new_unfolding
then do { tick (PostInlineUnconditionally old_bndr)
; return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
else
do { let info1 = idInfo new_bndr `setArityInfo` new_arity
info2 = info1 `setUnfoldingInfo` new_unfolding
info3 | isEvaldUnfolding new_unfolding = zapDemandInfo info2 `orElse` info2
| otherwise = info2
final_id = new_bndr `setIdInfo` info3
;
return (addNonRec env final_id final_rhs) } }
addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
addPolyBind top_lvl env (NonRec poly_id rhs)
= do { unfolding <- simplUnfolding env top_lvl poly_id rhs noUnfolding
; let final_id = setIdInfo poly_id $
idInfo poly_id `setUnfoldingInfo` unfolding
`setArityInfo` exprArity rhs
; return (addNonRec env final_id rhs) }
addPolyBind _ env bind@(Rec _)
= return (extendFloats env bind)
simplUnfolding :: SimplEnv-> TopLevelFlag
-> InId
-> OutExpr
-> Unfolding -> SimplM Unfolding
simplUnfolding env _ _ _ (DFunUnfolding ar con ops)
= return (DFunUnfolding ar con ops')
where
ops' = map (substExpr (text "simplUnfolding") env) ops
simplUnfolding env top_lvl id _
(CoreUnfolding { uf_tmpl = expr, uf_arity = arity
, uf_src = src, uf_guidance = guide })
| isStableSource src
= do { expr' <- simplExpr rule_env expr
; let src' = CoreSubst.substUnfoldingSource (mkCoreSubst (text "inline-unf") env) src
is_top_lvl = isTopLevel top_lvl
; case guide of
UnfWhen sat_ok _
-> let guide' = UnfWhen sat_ok (inlineBoringOk expr')
in return (mkCoreUnfolding src' is_top_lvl expr' arity guide')
_other
-> let bottoming = isBottomingId id
in bottoming `seq`
return (mkUnfolding src' is_top_lvl bottoming expr')
}
where
act = idInlineActivation id
rule_env = updMode (updModeForInlineRules act) env
simplUnfolding _ top_lvl id new_rhs _
= let bottoming = isBottomingId id
in bottoming `seq`
return (mkUnfolding InlineRhs (isTopLevel top_lvl) bottoming new_rhs)
\end{code}
Note [Force bottoming field]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We need to force bottoming, or the new unfolding holds
on to the old unfolding (which is part of the id).
Note [Arity decrease]
~~~~~~~~~~~~~~~~~~~~~
Generally speaking the arity of a binding should not decrease. But it *can*
legitimately happen becuase of RULES. Eg
f = g Int
where g has arity 2, will have arity 2. But if there's a rewrite rule
g Int --> h
where h has arity 1, then f's arity will decrease. Here's a real-life example,
which is in the output of Specialise:
Rec {
$dm {Arity 2} = \d.\x. op d
{-# RULES forall d. $dm Int d = $s$dm #-}
dInt = MkD .... opInt ...
opInt {Arity 1} = $dm dInt
$s$dm {Arity 0} = \x. op dInt }
Here opInt has arity 1; but when we apply the rule its arity drops to 0.
That's why Specialise goes to a little trouble to pin the right arity
on specialised functions too.
Note [Setting the new unfolding]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* If there's an INLINE pragma, we simplify the RHS gently. Maybe we
should do nothing at all, but simplifying gently might get rid of
more crap.
* If not, we make an unfolding from the new RHS. But *only* for
non-loop-breakers. Making loop breakers not have an unfolding at all
means that we can avoid tests in exprIsConApp, for example. This is
important: if exprIsConApp says 'yes' for a recursive thing, then we
can get into an infinite loop
If there's an InlineRule on a loop breaker, we hang on to the inlining.
It's pretty dodgy, but the user did say 'INLINE'. May need to revisit
this choice.
Note [Setting the demand info]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
If the unfolding is a value, the demand info may
go pear-shaped, so we nuke it. Example:
let x = (a,b) in
case x of (p,q) -> h p q x
Here x is certainly demanded. But after we've nuked
the case, we'll get just
let x = (a,b) in h a b x
and now x is not demanded (I'm assuming h is lazy)
This really happens. Similarly
let f = \x -> e in ...f..f...
After inlining f at some of its call sites the original binding may
(for example) be no longer strictly demanded.
The solution here is a bit ad hoc...
%************************************************************************
%* *
\subsection[Simplify-simplExpr]{The main function: simplExpr}
%* *
%************************************************************************
The reason for this OutExprStuff stuff is that we want to float *after*
simplifying a RHS, not before. If we do so naively we get quadratic
behaviour as things float out.
To see why it's important to do it after, consider this (real) example:
let t = f x
in fst t
==>
let t = let a = e1
b = e2
in (a,b)
in fst t
==>
let a = e1
b = e2
t = (a,b)
in
a -- Can't inline a this round, cos it appears twice
==>
e1
Each of the ==> steps is a round of simplification. We'd save a
whole round if we float first. This can cascade. Consider
let f = g d
in \x -> ...f...
==>
let f = let d1 = ..d.. in \y -> e
in \x -> ...f...
==>
let d1 = ..d..
in \x -> ...(\y ->e)...
Only in this second round can the \y be applied, and it
might do the same again.
\begin{code}
simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
simplExpr env expr = simplExprC env expr mkBoringStop
simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
simplExprC env expr cont
=
do { (env', expr') <- simplExprF (zapFloats env) expr cont
;
return (wrapFloats env' expr') }
simplExprF :: SimplEnv -> InExpr -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplExprF env e cont
=
simplExprF1 env e cont
simplExprF1 :: SimplEnv -> InExpr -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplExprF1 env (Var v) cont = simplIdF env v cont
simplExprF1 env (Lit lit) cont = rebuild env (Lit lit) cont
simplExprF1 env (Note n expr) cont = simplNote env n expr cont
simplExprF1 env (Cast body co) cont = simplCast env body co cont
simplExprF1 env (Coercion co) cont = simplCoercionF env co cont
simplExprF1 env (Type ty) cont = ASSERT( contIsRhsOrArg cont )
rebuild env (Type (substTy env ty)) cont
simplExprF1 env (App fun arg) cont = simplExprF env fun $
ApplyTo NoDup arg env cont
simplExprF1 env expr@(Lam {}) cont
= simplLam env zapped_bndrs body cont
where
(bndrs, body) = collectBinders expr
zapped_bndrs | need_to_zap = map zap bndrs
| otherwise = bndrs
need_to_zap = any zappable_bndr (drop n_args bndrs)
n_args = countArgs cont
zappable_bndr b = isId b && not (isOneShotBndr b)
zap b | isTyVar b = b
| otherwise = zapLamIdInfo b
simplExprF1 env (Case scrut bndr _ alts) cont
| sm_case_case (getMode env)
=
simplExprF env scrut (Select NoDup bndr alts env cont)
| otherwise
=
do { case_expr' <- simplExprC env scrut
(Select NoDup bndr alts env mkBoringStop)
; rebuild env case_expr' cont }
simplExprF1 env (Let (Rec pairs) body) cont
= do { env' <- simplRecBndrs env (map fst pairs)
; env'' <- simplRecBind env' NotTopLevel pairs
; simplExprF env'' body cont }
simplExprF1 env (Let (NonRec bndr rhs) body) cont
= simplNonRecE env bndr (rhs, env) ([], body) cont
simplType :: SimplEnv -> InType -> SimplM OutType
simplType env ty
=
seqType new_ty `seq` return new_ty
where
new_ty = substTy env ty
simplCoercionF :: SimplEnv -> InCoercion -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplCoercionF env co cont
= do { co' <- simplCoercion env co
; simpl_co co' cont }
where
simpl_co co (CoerceIt g cont)
= simpl_co new_co cont
where
new_co = mkSymCo g0 `mkTransCo` co `mkTransCo` g1
[g0, g1] = decomposeCo 2 g
simpl_co co cont
= seqCo co `seq` rebuild env (Coercion co) cont
simplCoercion :: SimplEnv -> InCoercion -> SimplM OutCoercion
simplCoercion env co
= let opt_co = optCoercion (getCvSubst env) co
in opt_co `seq` return opt_co
\end{code}
%************************************************************************
%* *
\subsection{The main rebuilder}
%* *
%************************************************************************
\begin{code}
rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
rebuild env expr cont
= case cont of
Stop {} -> return (env, expr)
CoerceIt co cont -> rebuild env (Cast expr co) cont
Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
StrictArg info _ cont -> rebuildCall env (info `addArgTo` expr) cont
StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
; simplLam env' bs body cont }
ApplyTo dup_flag arg se cont
| isSimplified dup_flag -> rebuild env (App expr arg) cont
| otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
; rebuild env (App expr arg') cont }
\end{code}
%************************************************************************
%* *
\subsection{Lambdas}
%* *
%************************************************************************
\begin{code}
simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplCast env body co0 cont0
= do { co1 <- simplCoercion env co0
;
simplExprF env body (addCoerce co1 cont0) }
where
addCoerce co cont = add_coerce co (coercionKind co) cont
add_coerce _co (Pair s1 k1) cont
| s1 `eqType` k1 = cont
add_coerce co1 (Pair s1 _k2) (CoerceIt co2 cont)
| (Pair _l1 t1) <- coercionKind co2
, s1 `eqType` t1 = cont
| otherwise = CoerceIt (mkTransCo co1 co2) cont
add_coerce co (Pair s1s2 _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
| Just (tyvar,_) <- splitForAllTy_maybe s1s2
= ASSERT( isTyVar tyvar )
ApplyTo Simplified (Type arg_ty') (zapSubstEnv arg_se) (addCoerce new_cast cont)
where
new_cast = mkInstCo co arg_ty'
arg_ty' | isSimplified dup = arg_ty
| otherwise = substTy (arg_se `setInScope` env) arg_ty
add_coerce co (Pair s1s2 t1t2) (ApplyTo dup arg arg_se cont)
| isFunTy s1s2
, isFunTy t1t2
= ApplyTo dup new_arg (zapSubstEnv arg_se) (addCoerce co2 cont)
where
[co1, co2] = decomposeCo 2 co
new_arg = mkCoerce (mkSymCo co1) arg'
arg' = substExpr (text "move-cast") arg_se' arg
arg_se' = arg_se `setInScope` env
add_coerce co _ cont = CoerceIt co cont
\end{code}
%************************************************************************
%* *
\subsection{Lambdas}
%* *
%************************************************************************
Note [Zap unfolding when beta-reducing]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Lambda-bound variables can have stable unfoldings, such as
$j = \x. \b{Unf=Just x}. e
See Note [Case binders and join points] below; the unfolding for lets
us optimise e better. However when we beta-reduce it we want to
revert to using the actual value, otherwise we can end up in the
stupid situation of
let x = blah in
let b{Unf=Just x} = y
in ...b...
Here it'd be far better to drop the unfolding and use the actual RHS.
\begin{code}
simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplLam env [] body cont = simplExprF env body cont
simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
= do { tick (BetaReduction bndr)
; simplNonRecE env (zap_unfolding bndr) (arg, arg_se) (bndrs, body) cont }
where
zap_unfolding bndr
| isId bndr, isStableUnfolding (realIdUnfolding bndr)
= setIdUnfolding bndr NoUnfolding
| otherwise = bndr
simplLam env bndrs body cont
= do { (env', bndrs') <- simplLamBndrs env bndrs
; body' <- simplExpr env' body
; new_lam <- mkLam env' bndrs' body'
; rebuild env' new_lam cont }
simplNonRecE :: SimplEnv
-> InBndr
-> (InExpr, SimplEnv)
-> ([InBndr], InExpr)
-> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
= ASSERT( isTyVar bndr )
do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
| preInlineUnconditionally env NotTopLevel bndr rhs
= do { tick (PreInlineUnconditionally bndr)
;
simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
| isStrictId bndr
= do { simplExprF (rhs_se `setFloats` env) rhs
(StrictBind bndr bndrs body env cont) }
| otherwise
= ASSERT( not (isTyVar bndr) )
do { (env1, bndr1) <- simplNonRecBndr env bndr
; let (env2, bndr2) = addBndrRules env1 bndr bndr1
; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
; simplLam env3 bndrs body cont }
\end{code}
%************************************************************************
%* *
\subsection{Notes}
%* *
%************************************************************************
\begin{code}
simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
-> SimplM (SimplEnv, OutExpr)
simplNote env (SCC cc) e cont
| pushCCisNop cc (getEnclosingCC env)
= simplExprF env e cont
| otherwise
= do { e' <- simplExpr (setEnclosingCC env currentCCS) e
; rebuild env (mkSCC cc e') cont }
simplNote env (CoreNote s) e cont
= do { e' <- simplExpr env e
; rebuild env (Note (CoreNote s) e') cont }
\end{code}
%************************************************************************
%* *
Variables
%* *
%************************************************************************
\begin{code}
simplVar :: SimplEnv -> InVar -> SimplM OutExpr
simplVar env var
| isTyVar var = return (Type (substTyVar env var))
| isCoVar var = return (Coercion (substCoVar env var))
| otherwise
= case substId env var of
DoneId var1 -> return (Var var1)
DoneEx e -> return e
ContEx tvs cvs ids e -> simplExpr (setSubstEnv env tvs cvs ids) e
simplIdF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
simplIdF env var cont
= case substId env var of
DoneEx e -> simplExprF (zapSubstEnv env) e cont
ContEx tvs cvs ids e -> simplExprF (setSubstEnv env tvs cvs ids) e cont
DoneId var1 -> completeCall env var1 cont
completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
completeCall env var cont
= do {
dflags <- getDOptsSmpl
; let (lone_variable, arg_infos, call_cont) = contArgs cont
n_val_args = length arg_infos
interesting_cont = interestingCallContext call_cont
unfolding = activeUnfolding env var
maybe_inline = callSiteInline dflags var unfolding
lone_variable arg_infos interesting_cont
; case maybe_inline of {
Just expr
-> do { tick (UnfoldingDone var)
; trace_inline dflags expr cont $
simplExprF (zapSubstEnv env) expr cont }
; Nothing -> do
{ rule_base <- getSimplRules
; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
; rebuildCall env info cont
}}}
where
trace_inline dflags unfolding cont stuff
| not (dopt Opt_D_dump_inlinings dflags) = stuff
| not (dopt Opt_D_verbose_core2core dflags)
= if isExternalName (idName var) then
pprDefiniteTrace "Inlining done:" (ppr var) stuff
else stuff
| otherwise
= pprDefiniteTrace ("Inlining done: " ++ showSDoc (ppr var))
(vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
text "Cont: " <+> ppr cont])
stuff
rebuildCall :: SimplEnv
-> ArgInfo
-> SimplCont
-> SimplM (SimplEnv, OutExpr)
rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
| not (contIsTrivial cont)
= return (env, mk_coerce res)
where
res = mkApps (Var fun) (reverse rev_args)
res_ty = exprType res
cont_ty = contResultType env res_ty cont
co = mkUnsafeCo res_ty cont_ty
mk_coerce expr | cont_ty `eqType` res_ty = expr
| otherwise = mkCoerce co expr
rebuildCall env info (ApplyTo dup_flag (Type arg_ty) se cont)
= do { arg_ty' <- if isSimplified dup_flag then return arg_ty
else simplType (se `setInScope` env) arg_ty
; rebuildCall env (info `addArgTo` Type arg_ty') cont }
rebuildCall env info@(ArgInfo { ai_encl = encl_rules
, ai_strs = str:strs, ai_discs = disc:discs })
(ApplyTo dup_flag arg arg_se cont)
| isSimplified dup_flag
= rebuildCall env (addArgTo info' arg) cont
| str
=
simplExprF (arg_se `setFloats` env) arg
(StrictArg info' cci cont)
| otherwise
= do { arg' <- simplExprC (arg_se `setInScope` env) arg
(mkLazyArgStop cci)
; rebuildCall env (addArgTo info' arg') cont }
where
info' = info { ai_strs = strs, ai_discs = discs }
cci | encl_rules || disc > 0 = ArgCtxt encl_rules
| otherwise = BoringCtxt
rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
= do {
; let args = reverse rev_args
env' = zapSubstEnv env
; mb_rule <- tryRules env rules fun args cont
; case mb_rule of {
Just (n_args, rule_rhs) -> simplExprF env' rule_rhs $
pushSimplifiedArgs env' (drop n_args args) cont ;
; Nothing -> rebuild env (mkApps (Var fun) args) cont
} }
\end{code}
Note [RULES apply to simplified arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It's very desirable to try RULES once the arguments have been simplified, because
doing so ensures that rule cascades work in one pass. Consider
{-# RULES g (h x) = k x
f (k x) = x #-}
...f (g (h x))...
Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
we match f's rules against the un-simplified RHS, it won't match. This
makes a particularly big difference when superclass selectors are involved:
op ($p1 ($p2 (df d)))
We want all this to unravel in one sweeep.
Note [Avoid redundant simplification]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Because RULES apply to simplified arguments, there's a danger of repeatedly
simplifying already-simplified arguments. An important example is that of
(>>=) d e1 e2
Here e1, e2 are simplified before the rule is applied, but don't really
participate in the rule firing. So we mark them as Simplified to avoid
re-simplifying them.
Note [Shadowing]
~~~~~~~~~~~~~~~~
This part of the simplifier may break the no-shadowing invariant
Consider
f (...(\a -> e)...) (case y of (a,b) -> e')
where f is strict in its second arg
If we simplify the innermost one first we get (...(\a -> e)...)
Simplifying the second arg makes us float the case out, so we end up with
case y of (a,b) -> f (...(\a -> e)...) e'
So the output does not have the no-shadowing invariant. However, there is
no danger of getting name-capture, because when the first arg was simplified
we used an in-scope set that at least mentioned all the variables free in its
static environment, and that is enough.
We can't just do innermost first, or we'd end up with a dual problem:
case x of (a,b) -> f e (...(\a -> e')...)
I spent hours trying to recover the no-shadowing invariant, but I just could
not think of an elegant way to do it. The simplifier is already knee-deep in
continuations. We have to keep the right in-scope set around; AND we have
to get the effect that finding (error "foo") in a strict arg position will
discard the entire application and replace it with (error "foo"). Getting
all this at once is TOO HARD!
%************************************************************************
%* *
Rewrite rules
%* *
%************************************************************************
\begin{code}
tryRules :: SimplEnv -> [CoreRule]
-> Id -> [OutExpr] -> SimplCont
-> SimplM (Maybe (Arity, CoreExpr))
tryRules env rules fn args call_cont
| null rules
= return Nothing
| otherwise
= do { case lookupRule (activeRule env) (getUnfoldingInRuleMatch env)
(getInScope env) fn args rules of {
Nothing -> return Nothing ;
Just (rule, rule_rhs) ->
do { tick (RuleFired (ru_name rule))
; dflags <- getDOptsSmpl
; trace_dump dflags rule rule_rhs $
return (Just (ruleArity rule, rule_rhs)) }}}
where
trace_dump dflags rule rule_rhs stuff
| not (dopt Opt_D_dump_rule_firings dflags)
, not (dopt Opt_D_dump_rule_rewrites dflags) = stuff
| not (dopt Opt_D_dump_rule_rewrites dflags)
= pprDefiniteTrace "Rule fired:" (ftext (ru_name rule)) stuff
| otherwise
= pprDefiniteTrace "Rule fired"
(vcat [text "Rule:" <+> ftext (ru_name rule),
text "Before:" <+> hang (ppr fn) 2 (sep (map pprParendExpr args)),
text "After: " <+> pprCoreExpr rule_rhs,
text "Cont: " <+> ppr call_cont])
stuff
\end{code}
Note [Rules for recursive functions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
You might think that we shouldn't apply rules for a loop breaker:
doing so might give rise to an infinite loop, because a RULE is
rather like an extra equation for the function:
RULE: f (g x) y = x+y
Eqn: f a y = a-y
But it's too drastic to disable rules for loop breakers.
Even the foldr/build rule would be disabled, because foldr
is recursive, and hence a loop breaker:
foldr k z (build g) = g k z
So it's up to the programmer: rules can cause divergence
%************************************************************************
%* *
Rebuilding a case expression
%* *
%************************************************************************
Note [Case elimination]
~~~~~~~~~~~~~~~~~~~~~~~
The case-elimination transformation discards redundant case expressions.
Start with a simple situation:
case x# of ===> let y# = x# in e
y# -> e
(when x#, y# are of primitive type, of course). We can't (in general)
do this for algebraic cases, because we might turn bottom into
non-bottom!
The code in SimplUtils.prepareAlts has the effect of generalise this
idea to look for a case where we're scrutinising a variable, and we
know that only the default case can match. For example:
case x of
0# -> ...
DEFAULT -> ...(case x of
0# -> ...
DEFAULT -> ...) ...
Here the inner case is first trimmed to have only one alternative, the
DEFAULT, after which it's an instance of the previous case. This
really only shows up in eliminating error-checking code.
Note that SimplUtils.mkCase combines identical RHSs. So
case e of ===> case e of DEFAULT -> r
True -> r
False -> r
Now again the case may be elminated by the CaseElim transformation.
This includes things like (==# a# b#)::Bool so that we simplify
case ==# a# b# of { True -> x; False -> x }
to just
x
This particular example shows up in default methods for
comparision operations (e.g. in (>=) for Int.Int32)
Note [CaseElimination: lifted case]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We also make sure that we deal with this very common case,
where x has a lifted type:
case e of
x -> ...x...
Here we are using the case as a strict let; if x is used only once
then we want to inline it. We have to be careful that this doesn't
make the program terminate when it would have diverged before, so we
check that
(a) 'e' is already evaluated (it may so if e is a variable)
Specifically we check (exprIsHNF e)
or
(b) the scrutinee is a variable and 'x' is used strictly
or
(c) 'x' is not used at all and e is ok-for-speculation
For the (c), consider
case (case a ># b of { True -> (p,q); False -> (q,p) }) of
r -> blah
The scrutinee is ok-for-speculation (it looks inside cases), but we do
not want to transform to
let r = case a ># b of { True -> (p,q); False -> (q,p) }
in blah
because that builds an unnecessary thunk.
Further notes about case elimination
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider: test :: Integer -> IO ()
test = print
Turns out that this compiles to:
Print.test
= \ eta :: Integer
eta1 :: State# RealWorld ->
case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
case hPutStr stdout
(PrelNum.jtos eta ($w[] @ Char))
eta1
of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
Notice the strange '<' which has no effect at all. This is a funny one.
It started like this:
f x y = if x < 0 then jtos x
else if y==0 then "" else jtos x
At a particular call site we have (f v 1). So we inline to get
if v < 0 then jtos x
else if 1==0 then "" else jtos x
Now simplify the 1==0 conditional:
if v<0 then jtos v else jtos v
Now common-up the two branches of the case:
case (v<0) of DEFAULT -> jtos v
Why don't we drop the case? Because it's strict in v. It's technically
wrong to drop even unnecessary evaluations, and in practice they
may be a result of 'seq' so we *definitely* don't want to drop those.
I don't really know how to improve this situation.
\begin{code}
rebuildCase, reallyRebuildCase
:: SimplEnv
-> OutExpr
-> InId
-> [InAlt]
-> SimplCont
-> SimplM (SimplEnv, OutExpr)
rebuildCase env scrut case_bndr alts cont
| Lit lit <- scrut
= do { tick (KnownBranch case_bndr)
; case findAlt (LitAlt lit) alts of
Nothing -> missingAlt env case_bndr alts cont
Just (_, bs, rhs) -> simple_rhs bs rhs }
| Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
= do { tick (KnownBranch case_bndr)
; case findAlt (DataAlt con) alts of
Nothing -> missingAlt env case_bndr alts cont
Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
case_bndr bs rhs cont
}
where
simple_rhs bs rhs = ASSERT( null bs )
do { env' <- simplNonRecX env case_bndr scrut
; simplExprF env' rhs cont }
rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
| all isDeadBinder bndrs
, if isUnLiftedType (idType case_bndr)
then ok_for_spec
else elim_lifted
= do { tick (CaseElim case_bndr)
; env' <- simplNonRecX env case_bndr scrut
; simplExprF env' rhs cont }
where
elim_lifted
= exprIsHNF scrut
|| (strict_case_bndr && scrut_is_var scrut)
|| (is_plain_seq && ok_for_spec)
ok_for_spec = exprOkForSpeculation scrut
is_plain_seq = isDeadBinder case_bndr
strict_case_bndr = isStrictDmd (idDemandInfo case_bndr)
scrut_is_var (Cast s _) = scrut_is_var s
scrut_is_var (Var v) = not (isTickBoxOp v)
scrut_is_var _ = False
rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
| all isDeadBinder (case_bndr : bndrs)
= do { let rhs' = substExpr (text "rebuild-case") env rhs
out_args = [Type (substTy env (idType case_bndr)),
Type (exprType rhs'), scrut, rhs']
; rule_base <- getSimplRules
; mb_rule <- tryRules env (getRules rule_base seqId) seqId out_args cont
; case mb_rule of
Just (n_args, res) -> simplExprF (zapSubstEnv env)
(mkApps res (drop n_args out_args))
cont
Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
rebuildCase env scrut case_bndr alts cont
= reallyRebuildCase env scrut case_bndr alts cont
reallyRebuildCase env scrut case_bndr alts cont
= do {
(env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
; if null alts' then missingAlt env case_bndr alts cont
else do
{ dflags <- getDOptsSmpl
; case_expr <- mkCase dflags scrut' case_bndr' alts'
; rebuild env' case_expr nodup_cont } }
\end{code}
simplCaseBinder checks whether the scrutinee is a variable, v. If so,
try to eliminate uses of v in the RHSs in favour of case_bndr; that
way, there's a chance that v will now only be used once, and hence
inlined.
Historical note: we use to do the "case binder swap" in the Simplifier
so there were additional complications if the scrutinee was a variable.
Now the binder-swap stuff is done in the occurrence analyer; see
OccurAnal Note [Binder swap].
Note [zapOccInfo]
~~~~~~~~~~~~~~~~~
If the case binder is not dead, then neither are the pattern bound
variables:
case of x { (a,b) ->
case x of { (p,q) -> p } }
Here (a,b) both look dead, but come alive after the inner case is eliminated.
The point is that we bring into the envt a binding
let x = (a,b)
after the outer case, and that makes (a,b) alive. At least we do unless
the case binder is guaranteed dead.
In practice, the scrutinee is almost always a variable, so we pretty
much always zap the OccInfo of the binders. It doesn't matter much though.
Note [Improving seq]
~~~~~~~~~~~~~~~~~~~
Consider
type family F :: * -> *
type instance F Int = Int
... case e of x { DEFAULT -> rhs } ...
where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
case e `cast` co of x'::Int
I# x# -> let x = x' `cast` sym co
in rhs
so that 'rhs' can take advantage of the form of x'.
Notice that Note [Case of cast] (in OccurAnal) may then apply to the result.
Nota Bene: We only do the [Improving seq] transformation if the
case binder 'x' is actually used in the rhs; that is, if the case
is *not* a *pure* seq.
a) There is no point in adding the cast to a pure seq.
b) There is a good reason not to: doing so would interfere
with seq rules (Note [Built-in RULES for seq] in MkId).
In particular, this [Improving seq] thing *adds* a cast
while [Built-in RULES for seq] *removes* one, so they
just flip-flop.
You might worry about
case v of x { __DEFAULT ->
... case (v `cast` co) of y { I# -> ... }}
This is a pure seq (since x is unused), so [Improving seq] won't happen.
But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
case v of x { __DEFAULT ->
... case (x `cast` co) of y { I# -> ... }}
Now the outer case is not a pure seq, so [Improving seq] will happen,
and then the inner case will disappear.
The need for [Improving seq] showed up in Roman's experiments. Example:
foo :: F Int -> Int -> Int
foo t n = t `seq` bar n
where
bar 0 = 0
bar n = bar (n - case t of TI i -> i)
Here we'd like to avoid repeated evaluating t inside the loop, by
taking advantage of the `seq`.
At one point I did transformation in LiberateCase, but it's more
robust here. (Otherwise, there's a danger that we'll simply drop the
'seq' altogether, before LiberateCase gets to see it.)
\begin{code}
simplAlts :: SimplEnv
-> OutExpr
-> InId
-> [InAlt]
-> SimplCont
-> SimplM (OutExpr, OutId, [OutAlt])
simplAlts env scrut case_bndr alts cont'
=
do { let env0 = zapFloats env
; (env1, case_bndr1) <- simplBinder env0 case_bndr
; fam_envs <- getFamEnvs
; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
case_bndr case_bndr1 alts
; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
; let mb_var_scrut = case scrut' of { Var v -> Just v; _ -> Nothing }
; alts' <- mapM (simplAlt alt_env' mb_var_scrut
imposs_deflt_cons case_bndr' cont') in_alts
; return (scrut', case_bndr', alts') }
improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
-> OutExpr -> InId -> OutId -> [InAlt]
-> SimplM (SimplEnv, OutExpr, OutId)
improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
| not (isDeadBinder case_bndr)
, Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
= do { case_bndr2 <- newId (fsLit "nt") ty2
; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCo co)
env2 = extendIdSubst env case_bndr rhs
; return (env2, scrut `Cast` co, case_bndr2) }
improveSeq _ env scrut _ case_bndr1 _
= return (env, scrut, case_bndr1)
simplAlt :: SimplEnv
-> Maybe OutId
-> [AltCon]
-> OutId
-> SimplCont
-> InAlt
-> SimplM OutAlt
simplAlt env scrut imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
= ASSERT( null bndrs )
do { let env' = addBinderUnfolding env scrut case_bndr'
(mkOtherCon imposs_deflt_cons)
; rhs' <- simplExprC env' rhs cont'
; return (DEFAULT, [], rhs') }
simplAlt env scrut _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
= ASSERT( null bndrs )
do { let env' = addBinderUnfolding env scrut case_bndr'
(mkSimpleUnfolding (Lit lit))
; rhs' <- simplExprC env' rhs cont'
; return (LitAlt lit, [], rhs') }
simplAlt env scrut _ case_bndr' cont' (DataAlt con, vs, rhs)
= do {
let vs_with_evals = add_evals (dataConRepStrictness con)
; (env', vs') <- simplLamBndrs env vs_with_evals
; let inst_tys' = tyConAppArgs (idType case_bndr')
con_args = map Type inst_tys' ++ varsToCoreExprs vs'
unf = mkSimpleUnfolding (mkConApp con con_args)
env'' = addBinderUnfolding env' scrut case_bndr' unf
; rhs' <- simplExprC env'' rhs cont'
; return (DataAlt con, vs', rhs') }
where
add_evals the_strs
= go vs the_strs
where
go [] [] = []
go (v:vs') strs | isTyVar v = v : go vs' strs
go (v:vs') (str:strs)
| isMarkedStrict str = evald_v : go vs' strs
| otherwise = zapped_v : go vs' strs
where
zapped_v = zapBndrOccInfo keep_occ_info v
evald_v = zapped_v `setIdUnfolding` evaldUnfolding
go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
keep_occ_info = isDeadBinder case_bndr' && isNothing scrut
addBinderUnfolding :: SimplEnv -> Maybe OutId -> Id -> Unfolding -> SimplEnv
addBinderUnfolding env scrut bndr unf
= case scrut of
Just v -> modifyInScope env1 (v `setIdUnfolding` unf)
_ -> env1
where
env1 = modifyInScope env bndr_w_unf
bndr_w_unf = bndr `setIdUnfolding` unf
zapBndrOccInfo :: Bool -> Id -> Id
zapBndrOccInfo keep_occ_info pat_id
| keep_occ_info = pat_id
| otherwise = zapIdOccInfo pat_id
\end{code}
Note [Add unfolding for scrutinee]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In general it's unlikely that a variable scrutinee will appear
in the case alternatives case x of { ...x unlikely to appear... }
because the binder-swap in OccAnal has got rid of all such occcurrences
See Note [Binder swap] in OccAnal.
BUT it is still VERY IMPORTANT to add a suitable unfolding for a
variable scrutinee, in simplAlt. Here's why
case x of y
(a,b) -> case b of c
I# v -> ...(f y)...
There is no occurrence of 'b' in the (...(f y)...). But y gets
the unfolding (a,b), and *that* mentions b. If f has a RULE
RULE f (p, I# q) = ...
we want that rule to match, so we must extend the in-scope env with a
suitable unfolding for 'y'. It's *essential* for rule matching; but
it's also good for case-elimintation -- suppose that 'f' was inlined
and did multi-level case analysis, then we'd solve it in one
simplifier sweep instead of two.
Exactly the same issue arises in SpecConstr;
see Note [Add scrutinee to ValueEnv too] in SpecConstr
%************************************************************************
%* *
\subsection{Known constructor}
%* *
%************************************************************************
We are a bit careful with occurrence info. Here's an example
(\x* -> case x of (a*, b) -> f a) (h v, e)
where the * means "occurs once". This effectively becomes
case (h v, e) of (a*, b) -> f a)
and then
let a* = h v; b = e in f a
and then
f (h v)
All this should happen in one sweep.
\begin{code}
knownCon :: SimplEnv
-> OutExpr
-> DataCon -> [OutType] -> [OutExpr]
-> InId -> [InBndr] -> InExpr
-> SimplCont
-> SimplM (SimplEnv, OutExpr)
knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
= do { env' <- bind_args env bs dc_args
; env'' <- bind_case_bndr env'
; simplExprF env'' rhs cont }
where
zap_occ = zapBndrOccInfo (isDeadBinder bndr)
bind_args env' [] _ = return env'
bind_args env' (b:bs') (Type ty : args)
= ASSERT( isTyVar b )
bind_args (extendTvSubst env' b ty) bs' args
bind_args env' (b:bs') (arg : args)
= ASSERT( isId b )
do { let b' = zap_occ b
; env'' <- simplNonRecX env' b' arg
; bind_args env'' bs' args }
bind_args _ _ _ =
pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
text "scrut:" <+> ppr scrut
bind_case_bndr env
| isDeadBinder bndr = return env
| exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
| otherwise = do { dc_args <- mapM (simplVar env) bs
; let con_app = Var (dataConWorkId dc)
`mkTyApps` dc_ty_args
`mkApps` dc_args
; simplNonRecX env bndr con_app }
missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
missingAlt env case_bndr alts cont
= WARN( True, ptext (sLit "missingAlt") <+> ppr case_bndr )
return (env, mkImpossibleExpr res_ty)
where
res_ty = contResultType env (substTy env (coreAltsType alts)) cont
\end{code}
%************************************************************************
%* *
\subsection{Duplicating continuations}
%* *
%************************************************************************
\begin{code}
prepareCaseCont :: SimplEnv
-> [InAlt] -> SimplCont
-> SimplM (SimplEnv, SimplCont, SimplCont)
prepareCaseCont env alts cont
| many_alts alts = mkDupableCont env cont
| otherwise = return (env, cont, mkBoringStop)
where
many_alts :: [InAlt] -> Bool
many_alts [] = False
many_alts [_] = False
many_alts (alt:alts)
| is_bot_alt alt = many_alts alts
| otherwise = not (all is_bot_alt alts)
is_bot_alt (_,_,rhs) = exprIsBottom rhs
\end{code}
Note [Bottom alternatives]
~~~~~~~~~~~~~~~~~~~~~~~~~~
When we have
case (case x of { A -> error .. ; B -> e; C -> error ..)
of alts
then we can just duplicate those alts because the A and C cases
will disappear immediately. This is more direct than creating
join points and inlining them away; and in some cases we would
not even create the join points (see Note [Single-alternative case])
and we would keep the case-of-case which is silly. See Trac #4930.
\begin{code}
mkDupableCont :: SimplEnv -> SimplCont
-> SimplM (SimplEnv, SimplCont, SimplCont)
mkDupableCont env cont
| contIsDupable cont
= return (env, cont, mkBoringStop)
mkDupableCont _ (Stop {}) = panic "mkDupableCont"
mkDupableCont env (CoerceIt ty cont)
= do { (env', dup, nodup) <- mkDupableCont env cont
; return (env', CoerceIt ty dup, nodup) }
mkDupableCont env cont@(StrictBind {})
= return (env, mkBoringStop, cont)
mkDupableCont env (StrictArg info cci cont)
= do { (env', dup, nodup) <- mkDupableCont env cont
; (env'', args') <- mapAccumLM (makeTrivial NotTopLevel) env' (ai_args info)
; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
mkDupableCont env (ApplyTo _ arg se cont)
=
do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
; arg' <- simplExpr (se `setInScope` env') arg
; (env'', arg'') <- makeTrivial NotTopLevel env' arg'
; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env'') dup_cont
; return (env'', app_cont, nodup_cont) }
mkDupableCont env cont@(Select _ case_bndr [(_, bs, _rhs)] _ _)
| all isDeadBinder bs
&& not (isUnLiftedType (idType case_bndr))
= return (env, mkBoringStop, cont)
mkDupableCont env (Select _ case_bndr alts se cont)
=
do { tick (CaseOfCase case_bndr)
; (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
; let alt_env = se `setInScope` env'
; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
; alts' <- mapM (simplAlt alt_env' Nothing [] case_bndr' dup_cont) alts
; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
; return (env'',
Select OkToDup case_bndr' alts'' (zapSubstEnv env'') mkBoringStop,
nodup_cont) }
mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
-> SimplM (SimplEnv, [InAlt])
mkDupableAlts env case_bndr' the_alts
= go env the_alts
where
go env0 [] = return (env0, [])
go env0 (alt:alts)
= do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
; (env2, alts') <- go env1 alts
; return (env2, alt' : alts' ) }
mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
-> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
mkDupableAlt env case_bndr (con, bndrs', rhs')
| exprIsDupable rhs'
= return (env, (con, bndrs', rhs'))
| otherwise
= do { let rhs_ty' = exprType rhs'
scrut_ty = idType case_bndr
case_bndr_w_unf
= case con of
DEFAULT -> case_bndr
DataAlt dc -> setIdUnfolding case_bndr unf
where
unf = mkInlineUnfolding Nothing rhs
rhs = mkConApp dc (map Type (tyConAppArgs scrut_ty)
++ varsToCoreExprs bndrs')
LitAlt {} -> WARN( True, ptext (sLit "mkDupableAlt")
<+> ppr case_bndr <+> ppr con )
case_bndr
used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
| otherwise = bndrs' ++ [case_bndr_w_unf]
abstract_over bndr
| isTyVar bndr = True
| otherwise = not (isDeadBinder bndr)
; (final_bndrs', final_args)
<- if (any isId used_bndrs')
then return (used_bndrs', varsToCoreExprs used_bndrs')
else do { rw_id <- newId (fsLit "w") realWorldStatePrimTy
; return ([rw_id], [Var realWorldPrimId]) }
; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
; let
really_final_bndrs = map one_shot final_bndrs'
one_shot v | isId v = setOneShotLambda v
| otherwise = v
join_rhs = mkLams really_final_bndrs rhs'
join_call = mkApps (Var join_bndr) final_args
; env' <- addPolyBind NotTopLevel env (NonRec join_bndr join_rhs)
; return (env', (con, bndrs', join_call)) }
\end{code}
Note [Fusing case continuations]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It's important to fuse two successive case continuations when the
first has one alternative. That's why we call prepareCaseCont here.
Consider this, which arises from thunk splitting (see Note [Thunk
splitting] in WorkWrap):
let
x* = case (case v of {pn -> rn}) of
I# a -> I# a
in body
The simplifier will find
(Var v) with continuation
Select (pn -> rn) (
Select [I# a -> I# a] (
StrictBind body Stop
So we'll call mkDupableCont on
Select [I# a -> I# a] (StrictBind body Stop)
There is just one alternative in the first Select, so we want to
simplify the rhs (I# a) with continuation (StricgtBind body Stop)
Supposing that body is big, we end up with
let $j a =
in case v of { pn -> case rn of
I# a -> $j a }
This is just what we want because the rn produces a box that
the case rn cancels with.
See Trac #4957 a fuller example.
Note [Case binders and join points]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this
case (case .. ) of c {
I# c# -> ....c....
If we make a join point with c but not c# we get
$j = \c -> ....c....
But if later inlining scrutines the c, thus
$j = \c -> ... case c of { I# y -> ... } ...
we won't see that 'c' has already been scrutinised. This actually
happens in the 'tabulate' function in wave4main, and makes a significant
difference to allocation.
An alternative plan is this:
$j = \c# -> let c = I# c# in ...c....
but that is bad if 'c' is *not* later scrutinised.
So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
(an InlineRule) that it's really I# c#, thus
$j = \c# -> \c[=I# c#] -> ...c....
Absence analysis may later discard 'c'.
NB: take great care when doing strictness analysis;
see Note [Lamba-bound unfoldings] in DmdAnal.
Also note that we can still end up passing stuff that isn't used. Before
strictness analysis we have
let $j x y c{=(x,y)} = (h c, ...)
in ...
After strictness analysis we see that h is strict, we end up with
let $j x y c{=(x,y)} = ($wh x y, ...)
and c is unused.
Note [Duplicated env]
~~~~~~~~~~~~~~~~~~~~~
Some of the alternatives are simplified, but have not been turned into a join point
So they *must* have an zapped subst-env. So we can't use completeNonRecX to
bind the join point, because it might to do PostInlineUnconditionally, and
we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
but zapping it (as we do in mkDupableCont, the Select case) is safe, and
at worst delays the join-point inlining.
Note [Small alternative rhs]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It is worth checking for a small RHS because otherwise we
get extra let bindings that may cause an extra iteration of the simplifier to
inline back in place. Quite often the rhs is just a variable or constructor.
The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
iterations because the version with the let bindings looked big, and so wasn't
inlined, but after the join points had been inlined it looked smaller, and so
was inlined.
NB: we have to check the size of rhs', not rhs.
Duplicating a small InAlt might invalidate occurrence information
However, if it *is* dupable, we return the *un* simplified alternative,
because otherwise we'd need to pair it up with an empty subst-env....
but we only have one env shared between all the alts.
(Remember we must zap the subst-env before re-simplifying something).
Rather than do this we simply agree to re-simplify the original (small) thing later.
Note [Funky mkPiTypes]
~~~~~~~~~~~~~~~~~~~~~~
Notice the funky mkPiTypes. If the contructor has existentials
it's possible that the join point will be abstracted over
type varaibles as well as term variables.
Example: Suppose we have
data T = forall t. C [t]
Then faced with
case (case e of ...) of
C t xs::[t] -> rhs
We get the join point
let j :: forall t. [t] -> ...
j = /\t \xs::[t] -> rhs
in
case (case e of ...) of
C t xs::[t] -> j t xs
Note [Join point abstaction]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
If we try to lift a primitive-typed something out
for let-binding-purposes, we will *caseify* it (!),
with potentially-disastrous strictness results. So
instead we turn it into a function: \v -> e
where v::State# RealWorld#. The value passed to this function
is realworld#, which generates (almost) no code.
There's a slight infelicity here: we pass the overall
case_bndr to all the join points if it's used in *any* RHS,
because we don't know its usage in each RHS separately
We used to say "&& isUnLiftedType rhs_ty'" here, but now
we make the join point into a function whenever used_bndrs'
is empty. This makes the join-point more CPR friendly.
Consider: let j = if .. then I# 3 else I# 4
in case .. of { A -> j; B -> j; C -> ... }
Now CPR doesn't w/w j because it's a thunk, so
that means that the enclosing function can't w/w either,
which is a lose. Here's the example that happened in practice:
kgmod :: Int -> Int -> Int
kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
then 78
else 5
I have seen a case alternative like this:
True -> \v -> ...
It's a bit silly to add the realWorld dummy arg in this case, making
$j = \s v -> ...
True -> $j s
(the \v alone is enough to make CPR happy) but I think it's rare
Note [Duplicating StrictArg]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The original plan had (where E is a big argument)
e.g. f E [..hole..]
==> let $j = \a -> f E a
in $j [..hole..]
But this is terrible! Here's an example:
&& E (case x of { T -> F; F -> T })
Now, && is strict so we end up simplifying the case with
an ArgOf continuation. If we let-bind it, we get
let $j = \v -> && E v
in simplExpr (case x of { T -> F; F -> T })
(ArgOf (\r -> $j r)
And after simplifying more we get
let $j = \v -> && E v
in case x of { T -> $j F; F -> $j T }
Which is a Very Bad Thing
What we do now is this
f E [..hole..]
==> let a = E
in f a [..hole..]
Now if the thing in the hole is a case expression (which is when
we'll call mkDupableCont), we'll push the function call into the
branches, which is what we want. Now RULES for f may fire, and
call-pattern specialisation. Here's an example from Trac #3116
go (n+1) (case l of
1 -> bs'
_ -> Chunk p fpc (o+1) (l-1) bs')
If we can push the call for 'go' inside the case, we get
call-pattern specialisation for 'go', which is *crucial* for
this program.
Here is the (&&) example:
&& E (case x of { T -> F; F -> T })
==> let a = E in
case x of { T -> && a F; F -> && a T }
Much better!
Notice that
* Arguments to f *after* the strict one are handled by
the ApplyTo case of mkDupableCont. Eg
f [..hole..] E
* We can only do the let-binding of E because the function
part of a StrictArg continuation is an explicit syntax
tree. In earlier versions we represented it as a function
(CoreExpr -> CoreEpxr) which we couldn't take apart.
Do *not* duplicate StrictBind and StritArg continuations. We gain
nothing by propagating them into the expressions, and we do lose a
lot.
The desire not to duplicate is the entire reason that
mkDupableCont returns a pair of continuations.
Note [Duplicating StrictBind]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Unlike StrictArg, there doesn't seem anything to gain from
duplicating a StrictBind continuation, so we don't.
Note [Single-alternative cases]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This case is just like the ArgOf case. Here's an example:
data T a = MkT !a
...(MkT (abs x))...
Then we get
case (case x of I# x' ->
case x' <# 0# of
True -> I# (negate# x')
False -> I# x') of y {
DEFAULT -> MkT y
Because the (case x) has only one alternative, we'll transform to
case x of I# x' ->
case (case x' <# 0# of
True -> I# (negate# x')
False -> I# x') of y {
DEFAULT -> MkT y
But now we do *NOT* want to make a join point etc, giving
case x of I# x' ->
let $j = \y -> MkT y
in case x' <# 0# of
True -> $j (I# (negate# x'))
False -> $j (I# x')
In this case the $j will inline again, but suppose there was a big
strict computation enclosing the orginal call to MkT. Then, it won't
"see" the MkT any more, because it's big and won't get duplicated.
And, what is worse, nothing was gained by the case-of-case transform.
So, in circumstances like these, we don't want to build join points
and push the outer case into the branches of the inner one. Instead,
don't duplicate the continuation.
When should we use this strategy? We should not use it on *every*
single-alternative case:
e.g. case (case ....) of (a,b) -> (# a,b #)
Here we must push the outer case into the inner one!
Other choices:
* Match [(DEFAULT,_,_)], but in the common case of Int,
the alternative-filling-in code turned the outer case into
case (...) of y { I# _ -> MkT y }
* Match on single alternative plus (not (isDeadBinder case_bndr))
Rationale: pushing the case inwards won't eliminate the construction.
But there's a risk of
case (...) of y { (a,b) -> let z=(a,b) in ... }
Now y looks dead, but it'll come alive again. Still, this
seems like the best option at the moment.
* Match on single alternative plus (all (isDeadBinder bndrs))
Rationale: this is essentially seq.
* Match when the rhs is *not* duplicable, and hence would lead to a
join point. This catches the disaster-case above. We can test
the *un-simplified* rhs, which is fine. It might get bigger or
smaller after simplification; if it gets smaller, this case might
fire next time round. NB also that we must test contIsDupable
case_cont *too, because case_cont might be big!
HOWEVER: I found that this version doesn't work well, because
we can get let x = case (...) of { small } in ...case x...
When x is inlined into its full context, we find that it was a bad
idea to have pushed the outer case inside the (...) case.
Note [Single-alternative-unlifted]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Here's another single-alternative where we really want to do case-of-case:
data Mk1 = Mk1 Int# | Mk2 Int#
M1.f =
\r [x_s74 y_s6X]
case
case y_s6X of tpl_s7m {
M1.Mk1 ipv_s70 -> ipv_s70;
M1.Mk2 ipv_s72 -> ipv_s72;
}
of
wild_s7c
{ __DEFAULT ->
case
case x_s74 of tpl_s7n {
M1.Mk1 ipv_s77 -> ipv_s77;
M1.Mk2 ipv_s79 -> ipv_s79;
}
of
wild1_s7b
{ __DEFAULT -> ==# [wild1_s7b wild_s7c];
};
};
So the outer case is doing *nothing at all*, other than serving as a
join-point. In this case we really want to do case-of-case and decide
whether to use a real join point or just duplicate the continuation:
let $j s7c = case x of
Mk1 ipv77 -> (==) s7c ipv77
Mk1 ipv79 -> (==) s7c ipv79
in
case y of
Mk1 ipv70 -> $j ipv70
Mk2 ipv72 -> $j ipv72
Hence: check whether the case binder's type is unlifted, because then
the outer case is *not* a seq.