[Git][ghc/ghc][wip/T24471] 2 commits: ghc-experimental: Add dummy dependencies to work around #23942
Ben Gamari (@bgamari)
gitlab at gitlab.haskell.org
Wed Mar 6 00:47:03 UTC 2024
Ben Gamari pushed to branch wip/T24471 at Glasgow Haskell Compiler / GHC
Commits:
3e892fe5 by Ben Gamari at 2024-03-05T19:45:07-05:00
ghc-experimental: Add dummy dependencies to work around #23942
This is a temporary measure to improve CI reliability until a proper
solution is developed.
Works around #23942.
- - - - -
8d1384b1 by Simon Peyton Jones at 2024-03-05T19:45:53-05:00
Three compile perf improvements with deep nesting
These were changes are all triggered by #24471.
1. Make GHC.Core.Opt.SetLevels.lvlMFE behave better when there are
many free variables. See Note [Large free-variable sets].
2. Make GHC.Core.Opt.Arity.floatIn a bit lazier in its Cost argument.
This benefits the common case where the ArityType turns out to
be nullary. See Note [Care with nested expressions]
3. Make GHC.CoreToStg.Prep.cpeArg behave for deeply-nested
expressions. See Note [Eta expansion of arguments in CorePrep]
wrinkle (EA2).
Compile times go down by up to 4.5%, and much more in artificial
cases. (Geo mean of compiler/perf changes is -0.4%.)
Metric Decrease:
CoOpt_Read
T10421
T12425
- - - - -
8 changed files:
- compiler/GHC/Core/Opt/Arity.hs
- compiler/GHC/Core/Opt/SetLevels.hs
- compiler/GHC/CoreToStg/Prep.hs
- libraries/ghc-experimental/src/Data/Sum/Experimental.hs
- libraries/ghc-experimental/src/Data/Tuple/Experimental.hs
- + testsuite/tests/perf/compiler/T24471.hs
- + testsuite/tests/perf/compiler/T24471a.hs
- testsuite/tests/perf/compiler/all.T
Changes:
=====================================
compiler/GHC/Core/Opt/Arity.hs
=====================================
@@ -1270,8 +1270,14 @@ arityLam id (AT oss div)
floatIn :: Cost -> ArityType -> ArityType
-- We have something like (let x = E in b),
-- where b has the given arity type.
-floatIn IsCheap at = at
-floatIn IsExpensive at = addWork at
+-- NB: be as lazy as possible in the Cost-of-E argument;
+-- we can often get away without ever looking at it
+-- See Note [Care with nested expressions]
+floatIn ch at@(AT lams div)
+ = case lams of
+ [] -> at
+ (IsExpensive,_):_ -> at
+ (_,os):lams -> AT ((ch,os):lams) div
addWork :: ArityType -> ArityType
-- Add work to the outermost level of the arity type
@@ -1354,6 +1360,25 @@ That gives \1.T (see Note [Combining case branches: andWithTail],
first bullet). So 'go2' gets an arityType of \(C?)(C1).T, which is
what we want.
+Note [Care with nested expressions]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Consider
+ arityType (Just <big-expressions>)
+We will take
+ arityType Just = AT [(IsCheap,os)] topDiv
+and then do
+ arityApp (AT [(IsCheap os)] topDiv) (exprCost <big-expression>)
+The result will be AT [] topDiv. It doesn't matter what <big-expresison>
+is! The same is true of
+ arityType (let x = <rhs> in <body>)
+where the cost of <rhs> doesn't matter unless <body> has a useful
+arityType.
+
+TL;DR in `floatIn`, do not to look at the Cost argument until you have to.
+
+I found this when looking at #24471, although I don't think it was really
+the main culprit.
+
Note [Combining case branches: andWithTail]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When combining the ArityTypes for two case branches (with andArityType)
@@ -1576,7 +1601,7 @@ arityType env (Case scrut bndr _ alts)
= alts_type
| otherwise -- In the remaining cases we may not push
- = addWork alts_type -- evaluation of the scrutinee in
+ = addWork alts_type -- evaluation of the scrutinee in
where
env' = delInScope env bndr
arity_type_alt (Alt _con bndrs rhs) = arityType (delInScopeList env' bndrs) rhs
=====================================
compiler/GHC/Core/Opt/SetLevels.hs
=====================================
@@ -643,7 +643,8 @@ lvlMFE env strict_ctxt e@(_, AnnCase {})
= lvlExpr env e -- See Note [Case MFEs]
lvlMFE env strict_ctxt ann_expr
- | floatTopLvlOnly env && not (isTopLvl dest_lvl)
+ | not float_me
+ || floatTopLvlOnly env && not (isTopLvl dest_lvl)
-- Only floating to the top level is allowed.
|| hasFreeJoin env fvs -- If there is a free join, don't float
-- See Note [Free join points]
@@ -652,8 +653,9 @@ lvlMFE env strict_ctxt ann_expr
-- how it will be represented at runtime.
-- See Note [Representation polymorphism invariants] in GHC.Core
|| notWorthFloating expr abs_vars
- || not float_me
- = -- Don't float it out
+ -- Test notWorhtFloating last;
+ -- See Note [Large free-variable sets]
+ = -- Don't float it out
lvlExpr env ann_expr
| float_is_new_lam || exprIsTopLevelBindable expr expr_ty
@@ -822,6 +824,28 @@ early loses opportunities for RULES which (needless to say) are
important in some nofib programs (gcd is an example). [SPJ note:
I think this is obsolete; the flag seems always on.]
+Note [Large free-variable sets]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+In #24471 we had something like
+ x1 = I# 1
+ ...
+ x1000 = I# 1000
+ foo = f x1 (f x2 (f x3 ....))
+So every sub-expression in `foo` has lots and lots of free variables. But
+none of these sub-expressions float anywhere; the entire float-out pass is a
+no-op.
+
+In lvlMFE, we want to find out quickly if the MFE is not-floatable; that is
+the common case. In #24471 it turned out that we were testing `abs_vars` (a
+relatively complicated calculation that takes at least O(n-free-vars) time to
+compute) for every sub-expression.
+
+Better instead to test `float_me` early. That still involves looking at
+dest_lvl, which means looking at every free variable, but the constant factor
+is a lot better.
+
+ToDo: find a way to fix the bad asymptotic complexity.
+
Note [Floating join point bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Mostly we only float a join point if it can /stay/ a join point. But
=====================================
compiler/GHC/CoreToStg/Prep.hs
=====================================
@@ -1469,8 +1469,7 @@ cpeArg env dmd arg
= do { (floats1, arg1) <- cpeRhsE env arg -- arg1 can be a lambda
; let arg_ty = exprType arg1
is_unlifted = isUnliftedType arg_ty
- dec = wantFloatLocal NonRecursive dmd is_unlifted
- floats1 arg1
+ dec = wantFloatLocal NonRecursive dmd is_unlifted floats1 arg1
; (floats2, arg2) <- executeFloatDecision dec floats1 arg1
-- Else case: arg1 might have lambdas, and we can't
-- put them inside a wrapBinds
@@ -1482,23 +1481,29 @@ cpeArg env dmd arg
then return (floats2, arg2)
else do { v <- newVar arg_ty
-- See Note [Eta expansion of arguments in CorePrep]
- ; let arg3 = cpeEtaExpandArg env arg2
+ ; let arity = cpeArgArity env dec arg2
+ arg3 = cpeEtaExpand arity arg2
arg_float = mkNonRecFloat env dmd is_unlifted v arg3
; return (snocFloat floats2 arg_float, varToCoreExpr v) }
}
-cpeEtaExpandArg :: CorePrepEnv -> CoreArg -> CoreArg
+cpeArgArity :: CorePrepEnv -> FloatDecision -> CoreArg -> Arity
-- ^ See Note [Eta expansion of arguments in CorePrep]
-cpeEtaExpandArg env arg = cpeEtaExpand arity arg
- where
- arity | Just ao <- cp_arityOpts (cpe_config env) -- Just <=> -O1 or -O2
- , not (has_join_in_tail_context arg)
+-- Returning 0 means "no eta-expansion"; see cpeEtaExpand
+cpeArgArity env float_decision arg
+ | FloatNone <- float_decision
+ = 0 -- Crucial short-cut
+ -- See wrinkle (EA2) in Note [Eta expansion of arguments in CorePrep]
+
+ | Just ao <- cp_arityOpts (cpe_config env) -- Just <=> -O1 or -O2
+ , not (has_join_in_tail_context arg)
-- See Wrinkle (EA1) of Note [Eta expansion of arguments in CorePrep]
- = case exprEtaExpandArity ao arg of
- Nothing -> 0
- Just at -> arityTypeArity at
- | otherwise
- = exprArity arg -- this is cheap enough for -O0
+ = case exprEtaExpandArity ao arg of
+ Nothing -> 0
+ Just at -> arityTypeArity at
+
+ | otherwise
+ = exprArity arg -- this is cheap enough for -O0
has_join_in_tail_context :: CoreExpr -> Bool
-- ^ Identify the cases where we'd generate invalid `CpeApp`s as described in
@@ -1510,34 +1515,10 @@ has_join_in_tail_context (Tick _ e) = has_join_in_tail_context e
has_join_in_tail_context (Case _ _ _ alts) = any has_join_in_tail_context (rhssOfAlts alts)
has_join_in_tail_context _ = False
-{-
-Note [Eta expansion of arguments with join heads]
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-See Note [Eta expansion for join points] in GHC.Core.Opt.Arity
-Eta expanding the join point would introduce crap that we can't
-generate code for
-
-------------------------------------------------------------------------------
--- Building the saturated syntax
--- ---------------------------------------------------------------------------
-
-Note [Eta expansion of hasNoBinding things in CorePrep]
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-maybeSaturate deals with eta expanding to saturate things that can't deal with
-unsaturated applications (identified by 'hasNoBinding', currently
-foreign calls, unboxed tuple/sum constructors, and representation-polymorphic
-primitives such as 'coerce' and 'unsafeCoerce#').
-
-Historical Note: Note that eta expansion in CorePrep used to be very fragile
-due to the "prediction" of CAFfyness that we used to make during tidying.
-We previously saturated primop
-applications here as well but due to this fragility (see #16846) we now deal
-with this another way, as described in Note [Primop wrappers] in GHC.Builtin.PrimOps.
--}
-
maybeSaturate :: Id -> CpeApp -> Int -> [CoreTickish] -> UniqSM CpeRhs
maybeSaturate fn expr n_args unsat_ticks
| hasNoBinding fn -- There's no binding
+ -- See Note [Eta expansion of hasNoBinding things in CorePrep]
= return $ wrapLamBody (\body -> foldr mkTick body unsat_ticks) sat_expr
| mark_arity > 0 -- A call-by-value function. See Note [CBV Function Ids]
@@ -1567,24 +1548,14 @@ maybeSaturate fn expr n_args unsat_ticks
fn_arity = idArity fn
excess_arity = (max fn_arity mark_arity) - n_args
sat_expr = cpeEtaExpand excess_arity expr
- applied_marks = n_args >= (length . dropWhile (not . isMarkedCbv) . reverse . expectJust "maybeSaturate" $ (idCbvMarks_maybe fn))
+ applied_marks = n_args >= (length . dropWhile (not . isMarkedCbv) .
+ reverse . expectJust "maybeSaturate" $ (idCbvMarks_maybe fn))
-- For join points we never eta-expand (See Note [Do not eta-expand join points])
- -- so we assert all arguments that need to be passed cbv are visible so that the backend can evalaute them if required..
-{-
-************************************************************************
-* *
- Simple GHC.Core operations
-* *
-************************************************************************
--}
+ -- so we assert all arguments that need to be passed cbv are visible so that the
+ -- backend can evalaute them if required..
-{-
--- -----------------------------------------------------------------------------
--- Eta reduction
--- -----------------------------------------------------------------------------
-
-Note [Eta expansion]
-~~~~~~~~~~~~~~~~~~~~~
+{- Note [Eta expansion]
+~~~~~~~~~~~~~~~~~~~~~~~
Eta expand to match the arity claimed by the binder Remember,
CorePrep must not change arity
@@ -1603,6 +1574,19 @@ NB2: we have to be careful that the result of etaExpand doesn't
an SCC note - we're now careful in etaExpand to make sure the
SCC is pushed inside any new lambdas that are generated.
+Note [Eta expansion of hasNoBinding things in CorePrep]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+maybeSaturate deals with eta expanding to saturate things that can't deal
+with unsaturated applications (identified by 'hasNoBinding', currently
+foreign calls, unboxed tuple/sum constructors, and representation-polymorphic
+primitives such as 'coerce' and 'unsafeCoerce#').
+
+Historical Note: Note that eta expansion in CorePrep used to be very fragile
+due to the "prediction" of CAFfyness that we used to make during tidying. We
+previously saturated primop applications here as well but due to this
+fragility (see #16846) we now deal with this another way, as described in
+Note [Primop wrappers] in GHC.Builtin.PrimOps.
+
Note [Eta expansion and the CorePrep invariants]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It turns out to be much much easier to do eta expansion
@@ -1685,6 +1669,22 @@ There is a nasty Wrinkle:
This scenario occurs rarely; hence it's OK to generate sub-optimal code.
The alternative would be to fix Note [Eta expansion for join points], but
that's quite challenging due to unfoldings of (recursive) join points.
+
+(EA2) In cpeArgArity, if float_decision = FloatNone) the `arg` will look like
+ let <binds> in rhs
+ where <binds> is non-empty and can't be floated out of a lazy context (see
+ `wantFloatLocal`). So we can't eta-expand it anyway, so we can return 0
+ forthwith. Without this short-cut we will call exprEtaExpandArity on the
+ `arg`, and <binds> might be enormous. exprEtaExpandArity be very expensive
+ on this: it uses arityType, and may look at <binds>.
+
+ On the other hand, if float_decision = FloatAll, there will be no
+ let-bindings around 'arg'; they will have floated out. So
+ exprEtaExpandArity is cheap.
+
+ This can make a huge difference on deeply nested expressions like
+ f (f (f (f (f ...))))
+ #24471 is a good example, where Prep took 25% of compile time!
-}
cpeEtaExpand :: Arity -> CpeRhs -> CpeRhs
@@ -1899,7 +1899,7 @@ instance Outputable FloatInfo where
-- See Note [Floating in CorePrep]
-- and Note [BindInfo and FloatInfo]
data FloatingBind
- = Float !CoreBind !BindInfo !FloatInfo
+ = Float !CoreBind !BindInfo !FloatInfo -- Never a join-point binding
| UnsafeEqualityCase !CoreExpr !CoreBndr !AltCon ![CoreBndr]
| FloatTick CoreTickish
@@ -2126,19 +2126,16 @@ data FloatDecision
| FloatAll
executeFloatDecision :: FloatDecision -> Floats -> CpeRhs -> UniqSM (Floats, CpeRhs)
-executeFloatDecision dec floats rhs = do
- let (float,stay) = case dec of
- _ | isEmptyFloats floats -> (emptyFloats,emptyFloats)
- FloatNone -> (emptyFloats, floats)
- FloatAll -> (floats, emptyFloats)
- -- Wrap `stay` around `rhs`.
- -- NB: `rhs` might have lambdas, and we can't
- -- put them inside a wrapBinds, which expects a `CpeBody`.
- if isEmptyFloats stay -- Fast path where we don't need to call `rhsToBody`
- then return (float, rhs)
- else do
- (floats', body) <- rhsToBody rhs
- return (float, wrapBinds stay $ wrapBinds floats' body)
+executeFloatDecision dec floats rhs
+ = case dec of
+ FloatAll -> return (floats, rhs)
+ FloatNone
+ | isEmptyFloats floats -> return (emptyFloats, rhs)
+ | otherwise -> do { (floats', body) <- rhsToBody rhs
+ ; return (emptyFloats, wrapBinds floats $
+ wrapBinds floats' body) }
+ -- FloatNone case: `rhs` might have lambdas, and we can't
+ -- put them inside a wrapBinds, which expects a `CpeBody`.
wantFloatTop :: Floats -> FloatDecision
wantFloatTop fs
=====================================
libraries/ghc-experimental/src/Data/Sum/Experimental.hs
=====================================
@@ -80,4 +80,5 @@ module Data.Sum.Experimental (
Sum63#,
) where
+import GHC.Num.BigNat () -- for build ordering
import GHC.Types
=====================================
libraries/ghc-experimental/src/Data/Tuple/Experimental.hs
=====================================
@@ -161,3 +161,4 @@ module Data.Tuple.Experimental (
import GHC.Tuple
import GHC.Types
import GHC.Classes
+import GHC.Num.BigNat () -- for build ordering
=====================================
testsuite/tests/perf/compiler/T24471.hs
=====================================
@@ -0,0 +1,8 @@
+{-# LANGUAGE TemplateHaskell #-}
+module T24471 where
+
+import T24471a
+
+{-# OPAQUE foo #-}
+foo :: (List_ Int a -> a) -> a
+foo alg = $$(between [|| alg ||] 0 1000)
=====================================
testsuite/tests/perf/compiler/T24471a.hs
=====================================
@@ -0,0 +1,8 @@
+{-# LANGUAGE TemplateHaskell #-}
+module T24471a where
+
+data List_ a f = Nil_ | Cons_ a f deriving Functor
+
+between alg a b
+ | a == b = [|| $$alg Nil_ ||]
+ | otherwise = [|| $$alg (Cons_ a $$(between alg (a + 1) b)) ||]
=====================================
testsuite/tests/perf/compiler/all.T
=====================================
@@ -712,3 +712,7 @@ test ('LookupFusion',
[collect_stats('bytes allocated',2), when(wordsize(32), skip)],
compile_and_run,
['-O2 -package base'])
+
+test('T24471',
+ [ collect_compiler_stats('all', 5) ],
+ multimod_compile, ['T24471', '-v0 -O'])
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/5c227bca89edb27b4328b9ed8b28c393d4cc312a...8d1384b1156be380b3d64eefbba4de756cd3f366
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/5c227bca89edb27b4328b9ed8b28c393d4cc312a...8d1384b1156be380b3d64eefbba4de756cd3f366
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/20240305/2265cff0/attachment-0001.html>
More information about the ghc-commits
mailing list