Date: Thu, 4 Jun 1998 13:59:21 -0500 (EST) From: "John S. Dyson" <dyson@FreeBSD.ORG> To: eivind@yes.no (Eivind Eklund) Cc: mike@smith.net.au, dmm125@bellatlantic.net, dyson@FreeBSD.ORG, hackers@FreeBSD.ORG Subject: Re: Recompiling sources with "-O2 -m486 -pipe" Message-ID: <199806041859.NAA02065@dyson.iquest.net> In-Reply-To: <19980604134839.09175@follo.net> from Eivind Eklund at "Jun 4, 98 01:48:40 pm"
next in thread | previous in thread | raw e-mail | index | archive | help
Eivind Eklund said:
>
> The _only_ thing I've found that affect PPro timing much is
> alignement, and alignment of inner loops can make a 50% speed
> difference. It caused me a lot of hassle until I found out about it
> ("Why the **** does changing that instruction, which is 20
> instructions before the main loop, speed up things by 30%?")
>
PPro is also sensitive to dirty cache lines. It is best
to check a data item before storing to it (for large transfers,
like bzero.) Careful scheduling can make some difference
also, but the effort is most often not worth the returns.
My local versions of swtch.s are carefully scheduled, and
they do make a difference, but not much.
The PPro can reorder instructions very aggressively, and
the order of long:1:1 isn't necessary, since the processor
can reorder those very effectively. The P5 is less
forgiving. I have spent many hours trying to make a P6
go faster, and it is really hard, and not generally
worthwhile.
--
John | Never try to teach a pig to sing,
dyson@freebsd.org | it just makes you look stupid,
jdyson@nc.com | and it irritates the pig.
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199806041859.NAA02065>