This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
Sparc issue in GCC 2.95.2
- To: gcc-bugs at gcc dot gnu dot org
- Subject: Sparc issue in GCC 2.95.2
- From: Josh Fryman <fryman at cc dot gatech dot edu>
- Date: Mon, 13 Nov 2000 10:27:16 -0500
- Organization: CoC, GaTech
Hi,
I didn't see someplace more appropriate to send this, so here goes. Please
don't flame me, if there's someplace better put it on the GCC web pages!
This isn't a bug so much a performance issue. I have a very simple C program
that computes an infinite impulse response. When I use the Sun "cc" compiler
with no special flags, I get pretty good performance. When I use gcc with
flags to get better performance, I get performance that's
_substantially_ worse. As in, more than an order of magnitude.
I've looked at the assembly output, and the best guess I have is that the sun
"cc" compiler really knows how to schedule and unroll the source code, whereas
gcc is just doing a good guess.
I found it most annoying that when I told gcc to generate ultrasparc code, it
had _worse_ performance than without the flags!
Are there any special compiler flags beyond the obvious ones that might improve
the results from gcc? I browsed the manual for a while, and could only come up
with these two as most important: -mflat -mcpu=ultrasparc.
Would anyone like a copy of the source code? Is anyone actively investigating
issues like this?
Disclaimer: I like gcc and would prefer to use it always, this is not
criticism! I'm just trying to find out if this is an on-going area of effort
or if this is "just the way it is" for now.
Regards,
Josh Fryman