Stream: t-compiler

Topic: severe performance regression after #65186


mw (Oct 09 2019 at 11:48, on Zulip):

https://perf.rust-lang.org/compare.html?start=59a31c8c676bdc9f50490d5798b1b4e884b7d7ae&end=d304f5c1ff0b86b07f643e733f9fe2d9cd13753e&stat=instructions:u

pnkfelix (Oct 09 2019 at 11:49, on Zulip):

eek

mw (Oct 09 2019 at 11:49, on Zulip):

deeply-nested-opt is 538126.0% slower :P

pnkfelix (Oct 09 2019 at 11:49, on Zulip):

well, at worst

mw (Oct 09 2019 at 11:49, on Zulip):

it might be a fluke, maybe we should wait for the next perf run

pnkfelix (Oct 09 2019 at 11:49, on Zulip):

On average its "only" 348058% slower

pnkfelix (Oct 09 2019 at 11:50, on Zulip):

how many runs are those averages based on?

mw (Oct 09 2019 at 11:50, on Zulip):

@simulacrum would know best. I think it's taking the fastest out of 5 for the smaller benchmarks

simulacrum (Oct 09 2019 at 11:51, on Zulip):

Heh at best 2 I think

mw (Oct 09 2019 at 11:51, on Zulip):

what? we have a sample of 2?

mw (Oct 09 2019 at 11:52, on Zulip):

that does not sound good :)

simulacrum (Oct 09 2019 at 11:53, on Zulip):

well, we take the minimum

simulacrum (Oct 09 2019 at 11:53, on Zulip):

this is in fact fairly reliable

simulacrum (Oct 09 2019 at 11:53, on Zulip):

for most "core" regressions you are actually running something like 2 * 4 (each run type) so you can usually see spurious stuff fairly well

simulacrum (Oct 09 2019 at 11:54, on Zulip):

in any case this specific regression has been there for a while I believe

simulacrum (Oct 09 2019 at 11:54, on Zulip):

i.e., we have multiple commits backing it up

simulacrum (Oct 09 2019 at 11:54, on Zulip):

I've not tried reproducing locally yet though

mw (Oct 09 2019 at 11:57, on Zulip):

@simulacrum will perf.rlo test the commits before?

simulacrum (Oct 09 2019 at 11:57, on Zulip):

hm?

simulacrum (Oct 09 2019 at 11:57, on Zulip):

before what?

simulacrum (Oct 09 2019 at 11:58, on Zulip):

you can see the queue here: https://perf.rust-lang.org/status.html

simulacrum (Oct 09 2019 at 11:58, on Zulip):

the length is part of why I suspect this is genuine

mw (Oct 09 2019 at 11:59, on Zulip):

I meant: will perf.rlo test the (as of yet untested) commit range, so we can better pinpoint the regression?

mw (Oct 09 2019 at 11:59, on Zulip):

the list on the status page probably has the answer :)

simulacrum (Oct 09 2019 at 12:00, on Zulip):

Yes, if there's bors commits in between we'll test them eventually

HeroicKatora (Oct 09 2019 at 12:02, on Zulip):

According to the graphs/numbers only baseline incremental and clean is affected that hard.

HeroicKatora (Oct 09 2019 at 12:03, on Zulip):

What's the difference about those and a clean incremental run?

simulacrum (Oct 09 2019 at 12:40, on Zulip):

whatever it is is cached by an incremental run

simulacrum (Oct 09 2019 at 12:40, on Zulip):

there's descriptions of the variants at the bottom of the compare page

mw (Oct 10 2019 at 07:38, on Zulip):

https://github.com/rust-lang/rust/pull/65249 should undo the regression. Thanks @Matthew Jasper and @nnethercote for looking into it!

Last update: Nov 22 2019 at 04:30UTC