This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: C Optimization Tests, 1 May 2004, tree-ssa/3.5/3.4/icc
Kevin Atkinson wrote:
You can't run 32-bit x86 binaries on your system at all?
It's more a matter of "don't" than "can't." I build the GCC compiler for
the Opteron with --disable-multilib; I'm only doing 64-bit development.
As built, the compiler won't accept any architecture setting that
doesn't generate x86_64 code.
Gentoo recently installed 32-bit compatibility libraries, so I could
compile the test suite on the Pentium 4 with -march=athlon-mp, then run
the executable on the Opteron, but that's too squirrelly for my tastes.
Aw heck, you piqued my curiosity. Here are the numbers for the
benchmarks as compiled on the P4 with -march=athlon-mp and run on the
64-bit Opteron:
AMD64/Opteron 240 (1.4GHz)
Gentoo AMD64 64-bit Linux 2.6.5
Compiled on a Pentium 4 with: -O3 -ffast-math -march=athlon-mp
-D__NO_MATH_LINES -D__NO_STRING_INLINES -mfpmath=sse
test tree-ssa mainline 3.4.1
---- -------- -------- --------
alma 22.6 51.7 69.0
evo 60.9 60.4 61.6
fft 37.6 37.3 37.4
huff 30.3 27.1 30.3
lin 31.2 30.2 30.3
mat1 29.5 29.6 29.5
mole 33.9 34.0 34.5
tree 40.6 37.3 36.7
-------- -------- --------
total 286.7 307.7 329.4
With minor exceptions, 64-bit "-march=opteron" code out-performs 32-bit
"-march=athlon-mp" code. Not surprising, really. Also, tree-ssa
obviously knows something about alma that mainline and 3.4 don't.
I'm unlikely to do the above on a regular basis, since I think this type
of pseudo-cross-compiling is unlikely to produce accurate results for a
processor I don't have.
As my original post says, I only run benchmarks on the hardware I have
at hand. You are more than welcome to run the benchmarks yourself; just
e-mail me privately, and I'll send you the beta benchmark suite.
..Scott
--
Scott Robert Ladd
Coyote Gulch Productions (http://www.coyotegulch.com)
Software Invention for High-Performance Computing