Compile time improvement, runtime regression (ia64)

[Daniel Berlin]

On Jun 13, 2004, at 8:18 AM, Richard Guenther wrote:

> Since the last two days I experience a compile time improvement both 
> in bootstrap times (50min vs. 60min before - but also switched host 
> compiler to 3.4.0 from 3.3.4) and tramp3d-v3 (non-leafified) testcase
> (2min30 vs. 3min before).  Unfortunately tramp3d-v3 performance 
> regressed from about 7.5s/it to 8s/it.
>
> There's a hole in testing due to various bootstrapping problems on 
> ia64 in the last week, so the precise cause remains uncertain.
>
> Any ideas what causes the performance drop?
>

It's quite possible it's GVN-PRE.
It doesn't perform load motion, which SSAPRE did.
This is being worked on.
It also can find so many partially redundant values (compared to 
SSAPRE) that it can increase register pressure.
I'm working on this too.
The easy way to tell if either of these could be the problem is to turn 
off PRE and see what timings you get.