[Git][ghc/ghc][wip/T21470] Fix binder-swap bug
Simon Peyton Jones (@simonpj)
gitlab at gitlab.haskell.org
Tue Aug 30 09:08:26 UTC 2022
Simon Peyton Jones pushed to branch wip/T21470 at Glasgow Haskell Compiler / GHC
Commits:
a4305deb by Simon Peyton Jones at 2022-08-30T10:04:44+01:00
Fix binder-swap bug
This patch fixes #21229 / #21470 properly, by avoiding doing a
binder-swap on dictionary Ids. This is pretty subtle, and explained
in Note [Care with binder-swap on dictionaries].
This allows us to restore a feature to the specialiser that we had
to revert: see Note [Specialising polymorphic dictionaries].
I als modularised things, using a new function scrutBinderSwap_maybe
in all the places where we are (effectively) doing a binder-swap,
notably
* Simplify.Iteration.addAltUnfoldings
* SpecConstr.extendCaseBndrs
In Simplify.Iteration.addAltUnfoldings I also eliminated a guard
Many <- idMult case_bndr
because we concluded, in #22123, that it was doing no good.
- - - - -
3 changed files:
- compiler/GHC/Core/Opt/OccurAnal.hs
- compiler/GHC/Core/Opt/Simplify/Iteration.hs
- compiler/GHC/Core/Opt/SpecConstr.hs
Changes:
=====================================
compiler/GHC/Core/Opt/OccurAnal.hs
=====================================
@@ -19,7 +19,7 @@ core expression with (hopefully) improved usage information.
module GHC.Core.Opt.OccurAnal (
occurAnalysePgm,
occurAnalyseExpr,
- zapLambdaBndrs
+ zapLambdaBndrs, scrutBinderSwap_maybe
) where
import GHC.Prelude
@@ -27,11 +27,12 @@ import GHC.Prelude
import GHC.Core
import GHC.Core.FVs
import GHC.Core.Utils ( exprIsTrivial, isDefaultAlt, isExpandableApp,
- stripTicksTopE, mkTicks )
+ 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.Core.TyCo.FVs ( tyCoVarsOfMCo )
import GHC.Data.Maybe( isJust, orElse )
import GHC.Data.Graph.Directed ( SCC(..), Node(..)
@@ -2462,8 +2463,8 @@ data OccEnv
-- See Note [The binder-swap substitution]
-- If x :-> (y, co) is in the env,
- -- then please replace x by (y |> sym mco)
- -- Invariant of course: idType x = exprType (y |> sym mco)
+ -- then please replace x by (y |> mco)
+ -- Invariant of course: idType x = exprType (y |> mco)
, occ_bs_env :: !(VarEnv (OutId, MCoercion))
, occ_bs_rng :: !VarSet -- Vars free in the range of occ_bs_env
-- Domain is Global and Local Ids
@@ -2669,7 +2670,7 @@ 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, co)] to the occ_bs_env environment; this is
+ 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
@@ -2737,30 +2738,8 @@ Some tricky corners:
(BS5) We have to apply the occ_bs_env substitution uniformly,
including to (local) rules and unfoldings.
-Historical note
----------------
-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.
+(BS6) We must be very careful with dictionaries.
+ See Note [Care with binder-swap on dictionaries]
Note [Case of cast]
~~~~~~~~~~~~~~~~~~~
@@ -2770,6 +2749,54 @@ 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 scrutBinderSwap_maybe.
+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 scrutBinderSwap_maybe.
+
Note [Zap case binders in proxy bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From the original
@@ -2784,8 +2811,83 @@ 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 })
+ | Just (scrut_var, mco) <- scrutBinderSwap_maybe 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]
+
+scrutBinderSwap_maybe :: OutExpr -> Maybe (OutVar, MCoercion)
+-- If (scrutBinderSwap_maybe e = Just (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
+scrutBinderSwap_maybe (Var v) = Just (v, MRefl)
+scrutBinderSwap_maybe (Cast (Var v) co)
+ | not (isDictId v) = Just (v, MCo (mkSymCo co))
+ -- Cast: see Note [Case of cast]
+ -- isDictId: see Note [Care with binder-swap on dictionaries]
+scrutBinderSwap_maybe (Tick _ e) = scrutBinderSwap_maybe e -- Drop ticks
+scrutBinderSwap_maybe _ = Nothing
+
+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,
@@ -2844,53 +2946,8 @@ binder-swap in OccAnal:
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.
--}
-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 })
- | Just (scrut_var, mco) <- get_scrut_var (stripTicksTopE (const True) 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
- get_scrut_var :: OutExpr -> Maybe (OutVar, MCoercion)
- get_scrut_var (Var v) = Just (v, MRefl)
- get_scrut_var (Cast (Var v) co) = Just (v, MCo co) -- See Note [Case of cast]
- get_scrut_var _ = Nothing
-
- case_bndr' = zapIdOccInfo case_bndr
- -- See Note [Zap case binders in proxy bindings]
-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) -> (add_cast fun mco, fun_id) }
-
- where
- add_cast fun MRefl = fun
- add_cast fun (MCo co) = Cast fun (mkSymCo co)
- -- We must switch that 'co' to 'sym co';
- -- see the comment with occ_bs_env
- -- No need to test for isReflCo, because 'co' came from
- -- a (Cast e co) and hence is unlikely to be Refl
-
-{-
************************************************************************
* *
\subsection[OccurAnal-types]{OccEnv}
=====================================
compiler/GHC/Core/Opt/Simplify/Iteration.hs
=====================================
@@ -22,7 +22,7 @@ import GHC.Core.Opt.Simplify.Monad
import GHC.Core.Type hiding ( substTy, substTyVar, extendTvSubst, extendCvSubst )
import GHC.Core.Opt.Simplify.Env
import GHC.Core.Opt.Simplify.Utils
-import GHC.Core.Opt.OccurAnal ( occurAnalyseExpr, zapLambdaBndrs )
+import GHC.Core.Opt.OccurAnal ( occurAnalyseExpr, zapLambdaBndrs, scrutBinderSwap_maybe )
import GHC.Core.Make ( FloatBind, mkImpossibleExpr, castBottomExpr )
import qualified GHC.Core.Make
import GHC.Core.Coercion hiding ( substCo, substCoVar )
@@ -3240,19 +3240,21 @@ zapIdOccInfoAndSetEvald str v =
-- see Note [Case alternative occ info]
addAltUnfoldings :: SimplEnv -> Maybe OutExpr -> OutId -> OutExpr -> SimplM SimplEnv
-addAltUnfoldings env scrut case_bndr con_app
+addAltUnfoldings env mb_scrut case_bndr con_app
= do { let con_app_unf = mk_simple_unf con_app
env1 = addBinderUnfolding env case_bndr con_app_unf
-- See Note [Add unfolding for scrutinee]
- env2 | Many <- idMult case_bndr = case scrut of
- Just (Var v) -> addBinderUnfolding env1 v con_app_unf
- Just (Cast (Var v) co) -> addBinderUnfolding env1 v $
- mk_simple_unf (Cast con_app (mkSymCo co))
- _ -> env1
+ env2 | Just scrut <- mb_scrut
+ , Just (v,mco) <- scrutBinderSwap_maybe scrut
+ = addBinderUnfolding env1 v $
+ if isReflMCo mco -- isReflMCo: avoid calling mk_simple_unf
+ then con_app_unf -- twice in the common case
+ else mk_simple_unf (mkCastMCo con_app mco)
+
| otherwise = env1
- ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr scrut, ppr con_app])
+ ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr mb_scrut, ppr con_app])
; return env2 }
where
-- Force the opts, so that the whole SimplEnv isn't retained
@@ -3315,9 +3317,6 @@ it's also good for case-elimination -- 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 GHC.Core.Opt.SpecConstr;
-see Note [Add scrutinee to ValueEnv too] in GHC.Core.Opt.SpecConstr
-
HOWEVER, given
case x of y { Just a -> r1; Nothing -> r2 }
we do not want to add the unfolding x -> y to 'x', which might seem cool,
@@ -3328,8 +3327,11 @@ piece of information.
So instead we add the unfolding x -> Just a, and x -> Nothing in the
respective RHSs.
-Since this transformation is tantamount to a binder swap, the same caveat as in
-Note [Suppressing binder-swaps on linear case] in OccurAnal apply.
+Since this transformation is tantamount to a binder swap, we use
+GHC.Core.Opt.OccurAnal.scrutBinderSwap_maybe to do the check.
+
+Exactly the same issue arises in GHC.Core.Opt.SpecConstr;
+see Note [Add scrutinee to ValueEnv too] in GHC.Core.Opt.SpecConstr
************************************************************************
=====================================
compiler/GHC/Core/Opt/SpecConstr.hs
=====================================
@@ -31,6 +31,7 @@ import GHC.Core.Unfold
import GHC.Core.FVs ( exprsFreeVarsList, exprFreeVars )
import GHC.Core.Opt.Monad
import GHC.Core.Opt.WorkWrap.Utils
+import GHC.Core.Opt.OccurAnal( scrutBinderSwap_maybe )
import GHC.Core.DataCon
import GHC.Core.Class( classTyVars )
import GHC.Core.Coercion hiding( substCo )
@@ -1057,8 +1058,8 @@ extendCaseBndrs env scrut case_bndr con alt_bndrs
= (env2, alt_bndrs')
where
live_case_bndr = not (isDeadBinder case_bndr)
- env1 | Var v <- stripTicksTopE (const True) scrut
- = extendValEnv env v cval
+ env1 | Just (v, mco) <- scrutBinderSwap_maybe scrut
+ , isReflMCo mco = extendValEnv env v cval
| otherwise = env -- See Note [Add scrutinee to ValueEnv too]
env2 | live_case_bndr = extendValEnv env1 case_bndr cval
| otherwise = env1
@@ -1148,6 +1149,10 @@ though the simplifier has systematically replaced uses of 'x' with 'y'
and 'b' with 'c' in the code. The use of 'b' in the ValueEnv came
from outside the case. See #4908 for the live example.
+It's very like the binder-swap story, so we use scrutBinderSwap_maybe
+to identify suitable scrutinees -- but only if there is no cast
+(isReflMCo) because that's all that the ValueEnv allows.
+
Note [Avoiding exponential blowup]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The sc_count field of the ScEnv says how many times we are prepared to
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/a4305deb90e81b713433f2f3701162056ba1c3a5
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/a4305deb90e81b713433f2f3701162056ba1c3a5
You're receiving this email because of your account on gitlab.haskell.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20220830/77d72a30/attachment-0001.html>
More information about the ghc-commits
mailing list