From: Godmar Back (gback@cs.utah.edu)
Date: Sun Dec 06 1998 - 13:53:55 EST
>
> On Dec 5, 1998, Godmar Back <gback@cs.utah.edu> wrote:
>
> > emptyloop: 23.759843 17.515912 16.802838
>
> I don't understand this difference, specially such a huge difference.
> Are you sure the load remained constant while you run the tests?
> Remember that the program relies on real time, not virtual time. I
> think this invalidates all the remaining tests, but we might get a
> general idea of the overall impact with the println test:
>
> > println: 70995.777874 79507.162940 93135.850864
>
I ran the tests on a dedicated test machine to minimize such problems.
Pat is currently running tests for his thesis, and he discovers similar
phenomena: adding seemingly unrelated code drastically changes the
times of other benchmarks.
I already mentioned one possible reason: the non-alignment of branch targets.
In the example above, the insertion of the softcall could have had the
side effect of aligning the back branch target of the empty loop.
How would I ensure that branch targets are always word aligned?
- Godmar
This archive was generated by hypermail 2b29 : Sat Sep 23 2000 - 19:57:06 EDT