This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Overhead of SSA operands
On Tue, 2004-11-16 at 03:50, Steven Bosscher wrote:
> On Tuesday 16 November 2004 09:26, Steven Bosscher wrote:
> > Here is the code we produce for this loop, including initialization:
>
> *sigh* When I'm going to dump assembly in a mail, I should make sure I
> pick the right file. That asm was for GCC 3.3.3. Sorry about that :-/
>
We definately ought to do a better job optimizing it in 4.0 :-)
> For 4.0 we do this:
<...>
> The loop starts at .L910 now. Much better, but still a lot of overhead
> with the loop tests.
>
Is the overhead really that measurable tho? I agree there is a certain
amount of quadraticness to the algorithm in that we have a bunch of
checks for zero index and move on to the next loop. I'd be suprised if
you could measure much impact since the number of variables is pretty
small.
THat said, I suspect we could introduce a switch statement to track the
current index, and eliminate a few of the ifs. You are likely to still
get some small number of comparisons anywaytho. Im experimenting with
it in my current version which has converted the entire compiler to the
iterators.
Also, the operands are being totally reworked under the covers for 4.1,
so I'll keep this in mind :-)
Andrew
Andrew