Subject: Re: Performance tweak to Prefetch Abort path
To: , <email@example.com>
From: Chris Gilbert <firstname.lastname@example.org>
Date: 03/25/2002 20:45:01
doing a benchmark (the time make configure of gmake) gives:
16.5u 37.3s 1:13.04 73.7% 0+0k 26+1560io 233pf+0w
against -current from last night:
16.9u 39.1s 1:14.21 75.5% 0+0k 26+1558io 231pf+0w
So yet another gain 8)
----- Original Message -----
From: "Jason R Thorpe" <email@example.com>
Sent: Monday, March 25, 2002 1:59 AM
Subject: Performance tweak to Prefetch Abort path
> I've just made another tweak to the ARM VM code, and this time
> it should definitely improve performance.
> In the Prefetch Abort path, the old code would make sure things looked
> kosher, and then actually perform a data access on the faulting PC to
> bring in the page. While this had the desired effect, it meant that
> we took a nother fault, for no good reason at all.
> What I did was change the Prefetch Abort handler to do the UVM dance
> directly, thus eliminating the second fault.
> If you have any problems, please file a bug report. That said, it works
> fine on my StrongARM and XScale systems.
> -- Jason R. Thorpe <firstname.lastname@example.org>