This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: gcc4.0 recent >15% runtime performance regression


On Mon, 03 Jan 2005 10:06:30 +0100, Uros Bizjak <uros@kss-loka.si> wrote:
> It would be nice to have a small testcase, or perhaps access to source
> code. I googled for mono-ray, but didn't found anything usable.
I said mono-ray as opposed to ray packets (4 rays bundled together
which nicely suit SSE), that is the classical aproach of shooting one
ray at a time: in that path i don't use intrinsics and totally rely on
the compiler.

In that regard gcc4.0 used to produce the best code; as it's a bit
intricate i posted here first to see if that rang a bell regarding a
recent change... i mean i'm not sure it's because of increased use of
the fpu instead of pure SSE. All that i know it's that it's slower.

> I suggest that you file a bugreport
> (http://gcc.gnu.org/bugzilla/enter_bug.cgi), where all discussion
> regarding this bug should take place.
Well, i wanted to avoid to have to make a testcase for something i'm
not totally sure but if that's what it takes... Thanks anyway.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]