On 1/29/2012 4:45 PM, John David Anglin wrote:
> Here is take take 2.  It's against linux-stable v3.2.2.  I also have a 
> very similar
> version against v3.1.10.
>
Here is take 3.  Since take 2, I have tried mainly to explore 
performance issues.

The big difference is a new implementation of flush_cache_mm where I try to
avoid the brutal flush of the whole cache.  This drops the full GCC 
build time at
-j4 from about six hours nine minutes to three hours ten minutes on 
rp3440 (i.e.,
almost by a factor two).  Build and check time is still a bit slower 
than HP-UX.
There's probably more to tweak here.

Other changes made little difference (e.g., clear_page_asm).  I also 
explored
the affect of prefetch operations in clear_page_asm and copy_page_asm, but
the difference if any was in the noise.

Dave

-- 
John David Anglin    dave.anglin@bell.net