[Git][ghc/ghc][wip/T22404] Make the occurrence analyser smarter about join points
Simon Peyton Jones (@simonpj)
gitlab at gitlab.haskell.org
Sun Jul 30 13:14:34 UTC 2023
Simon Peyton Jones pushed to branch wip/T22404 at Glasgow Haskell Compiler / GHC
Commits:
d0369802 by Simon Peyton Jones at 2023-07-30T09:24:48+01:00
Make the occurrence analyser smarter about join points
This MR addresses #22404. There is a big Note
Note [Occurrence analysis for join points]
that explains it all. Significant changes
* New field occ_join_points in OccEnv
* The NonRec case of occAnalBind splits into two cases:
one for existing join points (which does the special magic for
Note [Occurrence analysis for join points], and one for other
bindings.
* mkOneOcc adds in info from occ_join_points.
* All "bring into scope" activity is centralised in the
new function `addInScope`.
* I made a local data type LocalOcc for use inside the occurrence analyser
It is like OccInfo, but lacks IAmDead and IAmALoopBreaker, which in turn
makes computationns over it simpler and more efficient.
* I found quite a bit of allocation in GHC.Core.Rules.getRules
so I optimised it a bit.
More minor changes
* I found I was using (Maybe Arity) a lot, so I defined a new data
type JoinPointHood and used it everwhere. This touches a lot of
non-occ-anal files, but it makes everything more perspicuous.
* Renamed data constructor WithUsageDetails to WUD, and
WithTailUsageDetails to WTUD
This also fixes #21128, on the way.
--------- Compiler perf -----------
I spent quite a time on performance tuning, so even though it
does more than before, the occurrence analyser runs slightly faster
on average. Here are the compile-time allocation changes over 0.5%
CoOpt_Read(normal) ghc/alloc 766,025,520 754,561,992 -1.5%
CoOpt_Singletons(normal) ghc/alloc 759,436,840 762,925,512 +0.5%
LargeRecord(normal) ghc/alloc 1,814,482,440 1,799,530,456 -0.8%
PmSeriesT(normal) ghc/alloc 68,159,272 67,519,720 -0.9%
T10858(normal) ghc/alloc 120,805,224 118,746,968 -1.7%
T11374(normal) ghc/alloc 164,901,104 164,070,624 -0.5%
T11545(normal) ghc/alloc 79,851,808 78,964,704 -1.1%
T12150(optasm) ghc/alloc 73,903,664 71,237,544 -3.6% GOOD
T12227(normal) ghc/alloc 333,663,200 331,625,864 -0.6%
T12234(optasm) ghc/alloc 52,583,224 52,340,344 -0.5%
T12425(optasm) ghc/alloc 81,943,216 81,566,720 -0.5%
T13056(optasm) ghc/alloc 294,517,928 289,642,512 -1.7%
T13253-spj(normal) ghc/alloc 118,271,264 59,859,040 -49.4% GOOD
T15164(normal) ghc/alloc 1,102,630,352 1,091,841,296 -1.0%
T15304(normal) ghc/alloc 1,196,084,000 1,166,733,000 -2.5%
T15630(normal) ghc/alloc 148,729,632 147,261,064 -1.0%
T15703(normal) ghc/alloc 379,366,664 377,600,008 -0.5%
T16875(normal) ghc/alloc 32,907,120 32,670,976 -0.7%
T17516(normal) ghc/alloc 1,658,001,888 1,627,863,848 -1.8%
T17836(normal) ghc/alloc 395,329,400 393,080,248 -0.6%
T18140(normal) ghc/alloc 71,968,824 73,243,040 +1.8%
T18223(normal) ghc/alloc 456,852,568 453,059,088 -0.8%
T18282(normal) ghc/alloc 129,105,576 131,397,064 +1.8%
T18304(normal) ghc/alloc 71,311,712 70,722,720 -0.8%
T18698a(normal) ghc/alloc 208,795,112 210,102,904 +0.6%
T18698b(normal) ghc/alloc 230,320,736 232,697,976 +1.0% BAD
T19695(normal) ghc/alloc 1,483,648,128 1,504,702,976 +1.4%
T20049(normal) ghc/alloc 85,612,024 85,114,376 -0.6%
T21839c(normal) ghc/alloc 415,080,992 410,906,216 -1.0% GOOD
T4801(normal) ghc/alloc 247,590,920 250,726,272 +1.3%
T6048(optasm) ghc/alloc 95,699,416 95,080,680 -0.6%
T783(normal) ghc/alloc 335,323,384 332,988,120 -0.7%
T9233(normal) ghc/alloc 709,641,224 685,947,008 -3.3% GOOD
T9630(normal) ghc/alloc 965,635,712 948,356,120 -1.8%
T9675(optasm) ghc/alloc 444,604,152 428,987,216 -3.5% GOOD
T9961(normal) ghc/alloc 303,064,592 308,798,800 +1.9% BAD
WWRec(normal) ghc/alloc 503,728,832 498,102,272 -1.1%
geo. mean -1.0%
minimum -49.4%
maximum +1.9%
In fact these figures seem to vary between platforms; generally worse
on i386 for some reason. The Windows numbers vary by 1% espec in
benchmarks where the total allocation is low. But the geom mean stays
solidly negative, which is good. The "increase/decrease" list below
covers all platforms.
The big win on T13253-spj comes because it has a big nest of join
points, each occurring twice in the next one. The new occ-anal takes
only one iteration of the simplifier to do the inlining; the old one
took four. Moreover, we get much smaller code with the new one:
New: Result size of Tidy Core
= {terms: 429, types: 84, coercions: 0, joins: 14/14}
Old: Result size of Tidy Core
= {terms: 2,437, types: 304, coercions: 0, joins: 10/10}
--------- Runtime perf -----------
No significant changes in nofib results, except a 1% reduction in
compiler allocation.
Metric Decrease:
CoOpt_Read
T13253-spj
T9233
T9630
T9675
T12150
T21839c
LargeRecord
MultiComponentModulesRecomp
T10421
T13701
T10421
T13701
T12425
Metric Increase:
T18140
T9961
T18282
T18698a
T18698b
T19695
- - - - -
30 changed files:
- compiler/GHC/Core/Lint.hs
- compiler/GHC/Core/Opt/CSE.hs
- compiler/GHC/Core/Opt/DmdAnal.hs
- compiler/GHC/Core/Opt/Exitify.hs
- compiler/GHC/Core/Opt/FloatIn.hs
- compiler/GHC/Core/Opt/FloatOut.hs
- compiler/GHC/Core/Opt/OccurAnal.hs
- compiler/GHC/Core/Opt/SetLevels.hs
- compiler/GHC/Core/Opt/Simplify/Env.hs
- compiler/GHC/Core/Opt/Simplify/Iteration.hs
- compiler/GHC/Core/Opt/SpecConstr.hs
- compiler/GHC/Core/Opt/WorkWrap.hs
- compiler/GHC/Core/Ppr.hs
- compiler/GHC/Core/Rules.hs
- compiler/GHC/Core/SimpleOpt.hs
- compiler/GHC/Core/Tidy.hs
- compiler/GHC/Core/Unfold.hs
- compiler/GHC/Core/Unfold/Make.hs
- compiler/GHC/CoreToIface.hs
- compiler/GHC/CoreToStg/Prep.hs
- compiler/GHC/Iface/Syntax.hs
- compiler/GHC/IfaceToCore.hs
- compiler/GHC/Stg/Lint.hs
- compiler/GHC/Types/Basic.hs
- compiler/GHC/Types/Id.hs
- compiler/GHC/Types/Var.hs
- compiler/GHC/Utils/Binary.hs
- compiler/GHC/Utils/Outputable.hs
- + testsuite/tests/simplCore/should_compile/T22404.hs
- + testsuite/tests/simplCore/should_compile/T22404.stderr
The diff was not included because it is too large.
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/d03698023891b9d474915ad1cccdef8c8ba78e78
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/d03698023891b9d474915ad1cccdef8c8ba78e78
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/20230730/250baf6e/attachment.html>
More information about the ghc-commits
mailing list