diff options
author | Simon Peyton Jones <simonpj@microsoft.com> | 2021-02-11 09:47:42 +0000 |
---|---|---|
committer | Simon Peyton Jones <simonpj@microsoft.com> | 2021-02-12 17:59:40 +0000 |
commit | cedcdd27742acfee17196cef155bac17ca7b6a83 (patch) | |
tree | 9ea409c268da18dab2184350eeb24a8754cff063 /MAKEHELP.md | |
parent | 40983d2331fe34c0af6925db7588d5ac6a19ae36 (diff) | |
download | haskell-wip/T19347.tar.gz |
Fix over-eager inlining in SimpleOptwip/T19347
In GHC.Core.SimpleOpt, I found that its inlining could duplicate
an arbitary redex inside a lambda! Consider (\xyz. x+y). The
occurrence-analysis treats the lamdda as a group, and says that
both x and y occur once, even though the occur under the lambda-z.
See Note [Occurrence analysis for lambda binders] in OccurAnal.
When the lambda is under-applied in a call, the Simplifier is
careful to zap the occ-info on x,y, because they appear under the \z.
(See the call to zapLamBndrs in simplExprF1.) But SimpleOpt
missed this test, resulting in #19347.
So this patch
* commons up the binder-zapping in GHC.Core.Utils.zapLamBndrs.
* Calls this new function from GHC.Core.Opt.Simplify
* Adds a call to zapLamBndrs to GHC.Core.SimpleOpt.simple_app
This change makes test T12990 regress somewhat, but it was always
very delicate, so I'm going to put up with that.
In this voyage I also discovered a small, rather unrelated infelicity
in the Simplifier:
* In GHC.Core.Opt.Simplify.simplNonRecX we should apply isStrictId
to the OutId not the InId. See Note [Dark corner with levity polymorphism]
It may never "bite", because SimpleOpt should have inlined all
the levity-polymorphic compulsory inlnings already, but somehow
it bit me at one point and it's generally a more solid thing
to do.
Fixing the main bug increases runtime allocation in test
perf/should_run/T12990, for (acceptable) reasons explained in a
comement on
Metric Increase:
T12990
Diffstat (limited to 'MAKEHELP.md')
0 files changed, 0 insertions, 0 deletions