[Git][ghc/ghc][wip/T22084] Make SpecConstr bale out less often
Simon Peyton Jones (@simonpj)
gitlab at gitlab.haskell.org
Mon Oct 10 15:46:24 UTC 2022
Simon Peyton Jones pushed to branch wip/T22084 at Glasgow Haskell Compiler / GHC
Commits:
b749bf93 by Simon Peyton Jones at 2022-10-10T16:45:49+01:00
Make SpecConstr bale out less often
When doing performance debugging on #22084 / !8901, I found that the
algorithm in SpecConstr.decreaseSpecCount was so aggressive that if
there were /more/ specialisations available for an outer function,
that could more or less kill off specialisation for an /inner/
function. (An example was in nofib/spectral/fibheaps.)
This patch makes it a bit more aggressive, by dividing by 2, rather
than by the number of outer specialisations.
This makes the program bigger, temporarily:
T19695(normal) ghc/alloc +11.3% BAD
because we get more specialisation. But lots of other programs
compile a bit faster and the geometric mean in perf/compiler
is 0.0%.
Metric Increase:
T19695
- - - - -
1 changed file:
- compiler/GHC/Core/Opt/SpecConstr.hs
Changes:
=====================================
compiler/GHC/Core/Opt/SpecConstr.hs
=====================================
@@ -881,7 +881,7 @@ data SpecConstrOpts = SpecConstrOpts
, sc_count :: !(Maybe Int)
-- ^ Max # of specialisations for any one function. Nothing => no limit.
- -- See Note [Avoiding exponential blowup].
+ -- See Note [Avoiding exponential blowup] and decreaseSpecCount
, sc_recursive :: !Int
-- ^ Max # of specialisations over recursive type. Stops
@@ -1098,16 +1098,20 @@ extendCaseBndrs env scrut case_bndr con alt_bndrs
decreaseSpecCount :: ScEnv -> Int -> ScEnv
-- See Note [Avoiding exponential blowup]
-decreaseSpecCount env n_specs
+decreaseSpecCount env _n_specs
= env { sc_force = False -- See Note [Forcing specialisation]
- , sc_opts = (sc_opts env)
- { sc_count = case sc_count $ sc_opts env of
- Nothing -> Nothing
- Just n -> Just $! (n `div` (n_specs + 1))
+ , sc_opts = opts { sc_count = case sc_count opts of
+ Nothing -> Nothing
+ Just n -> Just $! dec n
}
}
- -- The "+1" takes account of the original function;
- -- See Note [Avoiding exponential blowup]
+ where
+ opts = sc_opts env
+ dec n = n `div` 2 -- See Note [Avoiding exponential blowup]
+
+ -- Or: n `div` (n_specs + 1)
+ -- See the historical note part of Note [Avoiding exponential blowup]
+ -- The "+1" takes account of the original function;
---------------------------------------------------
-- See Note [Forcing specialisation]
@@ -1183,9 +1187,20 @@ we can specialise $j2, and similarly $j3. Even if we make just *one*
specialisation of each, because we also have the original we'll get 2^n
copies of $j3, which is not good.
-So when recursively specialising we divide the sc_count by the number of
-copies we are making at this level, including the original.
-
+So when recursively specialising we divide the sc_count (the maximum
+number of specialisations, in the ScEnv) by two. You might think that
+gives us n*(n/2)*(n/4)... copies of the innnermost thing, which is
+still exponential the depth. But we use integer division, rounding
+down, so if the starting sc_count is 3, we'll get 3 -> 1 -> 0, and
+stop. In fact, simply subtracting 1 would be good enough, for the same
+reason.
+
+Historical note: in the past we divided by (n_specs+1), where n_specs
+is the number of specialisations at this level; but that gets us down
+to zero jolly quickly, which I found led to some regressions. (An
+example is nofib/spectral/fibheaps, the getMin' function inside the
+outer function $sfibToList, which has several interesting call
+patterns.)
************************************************************************
* *
@@ -1794,16 +1809,19 @@ specialise env bind_calls (RI { ri_fn = fn, ri_lam_bndrs = arg_bndrs
, not (null arg_bndrs) -- Only specialise functions
, Just all_calls <- lookupVarEnv bind_calls fn -- Some calls to it
= -- pprTrace "specialise entry {" (ppr fn <+> ppr all_calls) $
- do { (boring_call, new_pats) <- callsToNewPats env fn spec_info arg_occs all_calls
+ do { (boring_call, pats_discarded, new_pats)
+ <- callsToNewPats env fn spec_info arg_occs all_calls
; let n_pats = length new_pats
--- ; if (not (null new_pats) || isJust mb_unspec) then
--- pprTrace "specialise" (vcat [ ppr fn <+> text "with" <+> int n_pats <+> text "good patterns"
--- , text "mb_unspec" <+> ppr (isJust mb_unspec)
--- , text "arg_occs" <+> ppr arg_occs
--- , text "good pats" <+> ppr new_pats]) $
--- return ()
--- else return ()
+-- ; when (not (null new_pats) || isJust mb_unspec) $
+-- pprTraceM "specialise" (vcat [ ppr fn <+> text "with" <+> int n_pats <+> text "good patterns"
+-- , text "boring_call:" <+> ppr boring_call
+-- , text "pats_discarded:" <+> ppr pats_discarded
+-- , text "old spec_count" <+> ppr spec_count
+-- , text "spec count limit" <+> ppr (sc_count (sc_opts env))
+-- , text "mb_unspec" <+> ppr (isJust mb_unspec)
+-- , text "arg_occs" <+> ppr arg_occs
+-- , text "new_pats" <+> ppr new_pats])
; let spec_env = decreaseSpecCount env n_pats
; (spec_usgs, new_specs) <- mapAndUnzipM (spec_one spec_env fn arg_bndrs body)
@@ -1812,7 +1830,7 @@ specialise env bind_calls (RI { ri_fn = fn, ri_lam_bndrs = arg_bndrs
; let spec_usg = combineUsages spec_usgs
- unspec_rhs_needed = boring_call || isExportedId fn
+ unspec_rhs_needed = pats_discarded || boring_call || isExportedId fn
-- If there were any boring calls among the seeds (= all_calls), then those
-- calls will call the un-specialised function. So we should use the seeds
@@ -1823,15 +1841,14 @@ specialise env bind_calls (RI { ri_fn = fn, ri_lam_bndrs = arg_bndrs
-> (spec_usg `combineUsage` rhs_usg, Nothing)
_ -> (spec_usg, mb_unspec)
--- ; pprTrace "specialise return }"
--- (vcat [ ppr fn
--- , text "boring_call:" <+> ppr boring_call
--- , text "new calls:" <+> ppr (scu_calls new_usg)]) $
--- return ()
+-- ; pprTraceM "specialise return }" $
+-- vcat [ ppr fn
+-- , text "unspec_rhs_needed:" <+> ppr unspec_rhs_needed
+-- , text "new calls:" <+> ppr (scu_calls new_usg)]
- ; return (new_usg, SI { si_specs = new_specs ++ specs
- , si_n_specs = spec_count + n_pats
- , si_mb_unspec = mb_unspec' }) }
+ ; return (new_usg, SI { si_specs = new_specs ++ specs
+ , si_n_specs = spec_count + n_pats
+ , si_mb_unspec = mb_unspec' }) }
| otherwise -- No calls, inactive, or not a function
-- Behave as if there was a single, boring call
@@ -1874,7 +1891,9 @@ spec_one :: ScEnv
spec_one env fn arg_bndrs body (call_pat, rule_number)
| CP { cp_qvars = qvars, cp_args = pats, cp_strict_args = cbv_args } <- call_pat
- = do { spec_uniq <- getUniqueM
+ = do { -- pprTraceM "spec_one {" (ppr fn <+> ppr pats)
+
+ ; spec_uniq <- getUniqueM
; let env1 = extendScSubstList (extendScInScope env qvars)
(arg_bndrs `zip` pats)
(body_env, extra_bndrs) = extendBndrs env1 (dropList pats arg_bndrs)
@@ -1900,9 +1919,6 @@ spec_one env fn arg_bndrs body (call_pat, rule_number)
-- ; pprTraceM "body_subst_for" $ ppr (spec_occ) $$ ppr (sc_subst body_env)
; (spec_usg, spec_body) <- scExpr body_env body
--- ; pprTrace "done spec_one }" (ppr fn $$ ppr (scu_calls spec_usg)) $
--- return ()
-
-- And build the results
; (qvars', pats') <- generaliseDictPats qvars pats
; let spec_body_ty = exprType spec_body
@@ -1946,21 +1962,22 @@ spec_one env fn arg_bndrs body (call_pat, rule_number)
fn_name qvars' pats' rule_rhs
-- See Note [Transfer activation]
- -- ; pprTraceM "spec_one {" (vcat [ text "function:" <+> ppr fn <+> braces (ppr (idUnique fn))
- -- , text "sc_count:" <+> ppr (sc_count env)
- -- , text "pats:" <+> ppr pats
- -- , text "call_pat:" <+> ppr call_pat
- -- , text "-->" <+> ppr spec_name
- -- , text "bndrs" <+> ppr arg_bndrs
- -- , text "extra_bndrs" <+> ppr extra_bndrs
- -- , text "cbv_args" <+> ppr cbv_args
- -- , text "spec_lam_args" <+> ppr spec_lam_args
- -- , text "spec_call_args" <+> ppr spec_call_args
- -- , text "rule_rhs" <+> ppr rule_rhs
- -- , text "adds_void_worker_arg" <+> ppr add_void_arg
- -- , text "body" <+> ppr body
- -- , text "spec_rhs" <+> ppr spec_rhs
- -- , text "how_bound" <+> ppr (sc_how_bound env) ])
+-- ; pprTraceM "spec_one end }" $
+-- vcat [ text "function:" <+> ppr fn <+> braces (ppr (idUnique fn))
+-- , text "pats:" <+> ppr pats
+-- , text "call_pat:" <+> ppr call_pat
+-- , text "-->" <+> ppr spec_name
+-- , text "bndrs" <+> ppr arg_bndrs
+-- , text "extra_bndrs" <+> ppr extra_bndrs
+-- , text "cbv_args" <+> ppr cbv_args
+-- , text "spec_lam_args" <+> ppr spec_lam_args
+-- , text "spec_call_args" <+> ppr spec_call_args
+-- , text "rule_rhs" <+> ppr rule_rhs
+-- , text "adds_void_worker_arg" <+> ppr add_void_arg
+---- , text "body" <+> ppr body
+---- , text "spec_rhs" <+> ppr spec_rhs
+---- , text "how_bound" <+> ppr (sc_how_bound env) ]
+-- ]
; return (spec_usg, OS { os_pat = call_pat, os_rule = rule
, os_id = spec_id
, os_rhs = spec_rhs }) }
@@ -2330,7 +2347,9 @@ instance Outputable CallPat where
callsToNewPats :: ScEnv -> Id
-> SpecInfo
-> [ArgOcc] -> [Call]
- -> UniqSM (Bool, [CallPat])
+ -> UniqSM ( Bool -- At least one boring call
+ , Bool -- Patterns were discarded
+ , [CallPat] ) -- Patterns to specialise
-- Result has no duplicate patterns,
-- nor ones mentioned in si_specs (hence "new" patterns)
-- Bool indicates that there was at least one boring pattern
@@ -2362,12 +2381,11 @@ callsToNewPats env fn spec_info@(SI { si_specs = done_specs }) bndr_occs calls
-- Discard specialisations if there are too many of them
(pats_were_discarded, trimmed_pats) = trim_pats env fn spec_info small_pats
--- ; pprTrace "callsToPats" (vcat [ text "calls to" <+> ppr fn <> colon <+> ppr calls
--- , text "done_specs:" <+> ppr (map os_pat done_specs)
--- , text "good_pats:" <+> ppr good_pats ]) $
--- return ()
+-- ; pprTraceM "callsToPats" (vcat [ text "calls to" <+> ppr fn <> colon <+> ppr calls
+-- , text "done_specs:" <+> ppr (map os_pat done_specs)
+-- , text "trimmed_pats:" <+> ppr trimmed_pats ])
- ; return (have_boring_call || pats_were_discarded, trimmed_pats) }
+ ; return (have_boring_call, pats_were_discarded, trimmed_pats) }
-- If any of the calls does not give rise to a specialisation, either
-- because it is boring, or because there are too many specialisations,
-- return a flag to say so, so that we know to keep the original function.
@@ -2476,29 +2494,29 @@ callToPats env bndr_occs call@(Call fn args con_env)
sanitise id = updateIdTypeAndMult expandTypeSynonyms id
-- See Note [Free type variables of the qvar types]
- -- Bad coercion variables: see Note [SpecConstr and casts]
- bad_covars :: CoVarSet
+
+ -- Check for bad coercion variables: see Note [SpecConstr and casts]
+ ; let bad_covars :: CoVarSet
bad_covars = mapUnionVarSet get_bad_covars pats
get_bad_covars :: CoreArg -> CoVarSet
get_bad_covars (Type ty) = filterVarSet bad_covar (tyCoVarsOfType ty)
get_bad_covars _ = emptyVarSet
bad_covar v = isId v && not (is_in_scope v)
- ; -- pprTrace "callToPats" (ppr args $$ ppr bndr_occs) $
- warnPprTrace (not (isEmptyVarSet bad_covars))
+ ; warnPprTrace (not (isEmptyVarSet bad_covars))
"SpecConstr: bad covars"
(ppr bad_covars $$ ppr call) $
+
if interesting && isEmptyVarSet bad_covars
- then do
- -- pprTraceM "callToPatsOut" (
- -- text "fn:" <+> ppr fn $$
- -- text "args:" <+> ppr args $$
- -- text "in_scope:" <+> ppr in_scope $$
- -- -- text "in_scope:" <+> ppr in_scope $$
- -- text "pat_fvs:" <+> ppr pat_fvs
- -- )
- -- ppr (CP { cp_qvars = qvars', cp_args = pats })) >>
- return (Just (CP { cp_qvars = qvars', cp_args = pats, cp_strict_args = concat cbv_ids }))
+ then do { let cp_res = CP { cp_qvars = qvars', cp_args = pats
+ , cp_strict_args = concat cbv_ids }
+-- ; pprTraceM "callToPatsOut" $
+-- vcat [ text "fn:" <+> ppr fn
+-- , text "args:" <+> ppr args
+-- , text "bndr_occs:" <+> ppr bndr_occs
+-- , text "pat_fvs:" <+> ppr pat_fvs
+-- , text "cp_res:" <+> ppr cp_res ]
+ ; return (Just cp_res) }
else return Nothing }
-- argToPat takes an actual argument, and returns an abstracted
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/b749bf933f5a4ee5c7c26c3f29eba0873fc9a371
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/b749bf933f5a4ee5c7c26c3f29eba0873fc9a371
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/20221010/14b16725/attachment-0001.html>
More information about the ghc-commits
mailing list