Re: gcc loop unrolling problem ??

[Robert Dewar <dewar at gnat dot com>]

Robert Dewar wrote:

> Jan Hubicka wrote:
>
> > The bound on 8 unrollings is pretty artifical.  We do have other bound
> > specifying the size of loop to reach.  Perhaps you can try benchmark on
> > the code and tell us how it works when the loop is unrolled 9 times.
>
>
> Sure, but it would be surprising on modern processor architectures to
> find a case where unrolling more than 8 times helps. I suppose if
> symbolic interpretation handles straightline code and not loops,
> you could get an advantage :-)

Note incidentally that trying to measure this sort of thing is not
easy. That's because the kind of small programs you write to test
things don't usually suffer from i-cache pressure problems :-)