Commit 4f38fa10 authored by Simon Peyton Jones's avatar Simon Peyton Jones Committed by David Feuer
Browse files

Add -fspec-constr-keen

I discovered that the dramatic imprvoement in perf/should_run/T9339
with the introduction of join points was really rather a fluke, and
very fragile.

The real problem (see Note [Making SpecConstr keener]) is that
SpecConstr wasn't specialising a function even though it was applied
to a freshly-allocated constructor.  The paper describes plausible
reasons for this, but I think it may well be better to be a bit more

So this patch add -fspec-constr-keen, which makes SpecConstr a bit
keener to specialise, by ignoring whether or not the argument
corresponding to a call pattern is scrutinised in the function body.
Now the gains in T9339 should be robust; and it might even be a
better default.

I'd be interested in what happens if we switched on -fspec-constr-keen
with -O2.

Reviewers: austin, bgamari

Reviewed By: bgamari

Subscribers: thomie

Differential Revision:
parent e4188b53
......@@ -432,6 +432,7 @@ data GeneralFlag
| Opt_StgCSE
| Opt_LiberateCase
| Opt_SpecConstr
| Opt_SpecConstrKeen
| Opt_DoLambdaEtaExpansion
| Opt_IgnoreAsserts
| Opt_DoEtaReduction
......@@ -3684,6 +3685,7 @@ fFlagsDeps = [
(useInstead "enable-rewrite-rules"),
flagSpec "shared-implib" Opt_SharedImplib,
flagSpec "spec-constr" Opt_SpecConstr,
flagSpec "spec-constr-keen" Opt_SpecConstrKeen,
flagSpec "specialise" Opt_Specialise,
flagSpec "specialize" Opt_Specialise,
flagSpec "specialise-aggressively" Opt_SpecialiseAggressively,
......@@ -41,7 +41,8 @@ import VarEnv
import VarSet
import Name
import BasicTypes
import DynFlags ( DynFlags(..), hasPprDebug )
import DynFlags ( DynFlags(..), GeneralFlag( Opt_SpecConstrKeen )
, gopt, hasPprDebug )
import Maybes ( orElse, catMaybes, isJust, isNothing )
import Demand
import GHC.Serialized ( deserializeWithData )
......@@ -447,7 +448,6 @@ breaks an invariant.
Note [Forcing specialisation]
With stream fusion and in other similar cases, we want to fully
specialise some (but not necessarily all!) loops regardless of their
size and the number of specialisations.
......@@ -754,6 +754,39 @@ into a work-free value again, thus
a'_shr = (a1, x_af7)
but that's more work, so until its shown to be important I'm going to
leave it for now.
Note [Making SpecConstr keener]
Consider this, in (perf/should_run/T9339)
last (filter odd [1..1000])
After optimisation, including SpecConstr, we get:
f :: Int# -> Int -> Int
f x y = case case remInt# x 2# of
__DEFAULT -> case x of
__DEFAULT -> f (+# wild_Xp 1#) (I# x)
1000000# -> ...
0# -> case x of
__DEFAULT -> f (+# wild_Xp 1#) y
1000000# -> y
Not good! We build an (I# x) box every time around the loop.
SpecConstr (as described in the paper) does not specialise f, despite
the call (f ... (I# x)) because 'y' is not scrutinied in the body.
But it is much better to specialise f for the case where the argument
is of form (I# x); then we build the box only when returning y, which
is on the cold path.
Another exmaple:
f x = ...(g x)....
Here 'x' is not scrutinised in f's body; but if we did specialise 'f'
then the call (g x) might allow 'g' to be specialised in turn.
So sc_keen controls whether or not we take account of whether argument is
scrutinised in the body. True <=> ignore that, and speicalise whenever
the function is applied to a data constructor.
data ScEnv = SCE { sc_dflags :: DynFlags,
......@@ -765,6 +798,11 @@ data ScEnv = SCE { sc_dflags :: DynFlags,
sc_recursive :: Int, -- Max # of specialisations over recursive type.
-- Stops ForceSpecConstr from diverging.
sc_keen :: Bool, -- Specialise on arguments that are known
-- constructors, even if they are not
-- scrutinised in the body. See
-- Note [Making SpecConstr keener]
sc_force :: Bool, -- Force specialisation?
-- See Note [Forcing specialisation]
......@@ -807,6 +845,7 @@ initScEnv dflags this_mod anns
sc_size = specConstrThreshold dflags,
sc_count = specConstrCount dflags,
sc_recursive = specConstrRecursive dflags,
sc_keen = gopt Opt_SpecConstrKeen dflags,
sc_force = False,
sc_subst = emptySubst,
sc_how_bound = emptyVarEnv,
......@@ -1976,11 +2015,12 @@ argToPat env in_scope val_env arg arg_occ
mkConApp dc (ty_args ++ args')) }
mb_scrut dc = case arg_occ of
ScrutOcc bs
| Just occs <- lookupUFM bs dc
-> Just (occs) -- See Note [Reboxing]
_other | sc_force env -> Just (repeat UnkOcc)
| otherwise -> Nothing
ScrutOcc bs | Just occs <- lookupUFM bs dc
-> Just (occs) -- See Note [Reboxing]
_other | sc_force env || sc_keen env
-> Just (repeat UnkOcc)
| otherwise
-> Nothing
-- Check if the argument is a variable that
-- (a) is used in an interesting way in the function body
......@@ -1989,6 +2029,9 @@ argToPat env in_scope val_env arg arg_occ
argToPat env in_scope val_env (Var v) arg_occ
| sc_force env || case arg_occ of { UnkOcc -> False; _other -> True }, -- (a)
is_value, -- (b)
-- Ignoring sc_keen here to avoid gratuitously incurring Note [Reboxing]
-- So sc_keen focused just on f (I# x), where we have freshly-allocated
-- box that we can eliminate in the caller
not (ignoreType env (varType v))
= return (True, Var v)
......@@ -522,7 +522,7 @@ list.
Turn on call-pattern specialisation; see `Call-pattern specialisation for
Haskell programs
This optimisation specializes recursive functions according to their
argument "shapes". This is best explained by example so consider: ::
......@@ -580,6 +580,16 @@ list.
body directly, allowing heavy specialisation over the recursive
.. ghc-flag:: -fspec-constr-keen
:default: off
If this flag is on, call-patten specialision will specialise a call
``(f (Just x))`` with an explicit constructor agument, even if the argument
is not scrutinised in the body of the function. This is sometimes
beneficial; e.g. the argument might be given to some other function
that can itself be specialised.
.. ghc-flag:: -fspec-constr-count=<n>
:default: 3
......@@ -462,7 +462,9 @@ test('T9339',
# 2016-08-17: 50728 Join points (#12988)
['-O2 -fspec-constr-keen'])
# For the -fspec-constr-keen see Note [Making SpecConstr keener] in SpecConstr
[stats_num_field('bytes allocated',
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment