Perf

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

Perf

Simon Peyton Jones

Joachim

Harbormaster seems to be saying that GHC’s perf got a lot worse

https://phabricator.haskell.org/harbormaster/

after Buildable 9015, which omitted the oneshot info.  The typechecker seems to allocate a lot more.

That is not good.  Revert and investigate?

Simon

 


_______________________________________________
ghc-devs mailing list
[hidden email]
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
Reply | Threaded
Open this post in threaded view
|

Re: Perf

Joachim Breitner-2
Hi,

Am Donnerstag, den 31.03.2016, 14:56 +0000 schrieb Simon Peyton Jones:

> Joachim
>
> Harbormaster seems to be saying that GHC’s perf got a lot worse
>
> https://phabricator.haskell.org/harbormaster/
>
> after Buildable 9015, which omitted the oneshot info.  The
> typechecker seems to allocate a lot more.
>
> That is not good.  Revert and investigate?
yes, I noticed as well, and perf.haskell.org has now also confirmed
this. Just pushed a reversal. But I wonder what went wrong when I ran
./validate --slow
locally.

Can we please have harbormaster builds of DRs back? I felt just so much
more secure about pushing when the commits were checked by a more
reliable party than me.

Greetings,
Joachim


--
Joachim “nomeata” Breitner
  [hidden email]https://www.joachim-breitner.de/
  XMPP: [hidden email] • OpenPGP-Key: 0xF0FBF51F
  Debian Developer: [hidden email]


_______________________________________________
ghc-devs mailing list
[hidden email]
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

signature.asc (836 bytes) Download Attachment
Reply | Threaded
Open this post in threaded view
|

cryptarithm1 performance discontinuity (Was: Perf)

Joachim Breitner-2
Hi,

Am Donnerstag, den 31.03.2016, 18:33 +0200 schrieb Joachim Breitner:
> yes, I noticed as well, and perf.haskell.org has now also confirmed
> this. Just pushed a reversal. But I wonder what went wrong when I ran
> ./validate --slow
> locally.

interesting, my patch, as bad as it was for compiler performance,
improved the runtime of cryptarithm1 by the amount it regressed due to 

commit d1179c4bff6d05cc9e86eee3e2d2cee707983c90
Author: Ben Gamari <[hidden email]>
Date:   Wed Mar 30 10:09:36 2016 +0200

    ghc-prim: Delay inlining of {gt,ge,lt,le}Int to phase 1

as you can nicely see in this graph:
https://perf.haskell.org/ghc/#graph/nofib/time/cryptarithm1;hl=d1179c4bff6d05cc9e86eee3e2d2cee707983c90
(sorry, the relevant bits will move out of the window as more commits
come in. This is something I need to improve in gipeda).

There seems to be a performance non-continuity triggered by very small
changes. *shrug*, for now. But if anyone feels like investigating, I’d
be eager to hear you find.

Greetings,
Joachim

--
Joachim “nomeata” Breitner
  [hidden email]https://www.joachim-breitner.de/
  XMPP: [hidden email] • OpenPGP-Key: 0xF0FBF51F
  Debian Developer: [hidden email]


_______________________________________________
ghc-devs mailing list
[hidden email]
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

signature.asc (836 bytes) Download Attachment