Re: [tree-ssa] New alias grouping heuristic

[Vladimir Makarov <vmakarov at redhat dot com>]

Diego Novillo wrote:

> Having said that, a 2-4% difference in SPEC is really nothing to be
> excited about.  SPEC is pretty volatile.  You typically get that
> difference using the same compiler with the exact same flags.  It's
> differences greater than 5% that are worth looking into.
>
>
>  
I am absolutely disagree with this.  The first, 2-4% for SPEC is really 
exciting (it is about 1 year work on improving compiler according to 
Proebsting's law).

The second, I found that SPEC results are pretty stable if you run it 
right on a free machine.  I have stable results in range 2-4 for score 
about 800 for P4.   Although the volatility could be more for single 
tests.  But the geometry meaning is pretty the same.

I saw bigger volatility on some machines (e.g. Pentium-M) but if you run 
peak right after base, the % difference is not so volatile again 
(because the peak and base results change in  the same proportion).

Vlad