GCC vs Sun CC on SPARC

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi,

I recently ran into this article, which benchmarks GCC 2.95.3 vs GCC 3.3.2 vs 
Sun ONE Studio 7 CC on an old UltraSPARC 5 machine:

   http://www.osnews.com/story.php?news_id=5830

Summary: in 32-bit mode, GCC 3.3.2 code runs slightly slower than GCC 2.95.3 
code and slower than Sun CC code (time delta: ~15%).  In 64-bit mode, there 
is again a ~15% gap between GCC 3.3.2 and Sun CC.


As I have access to a more modern SPARC technology, I decided to do the gzip 
benchmark with GCC 2.95.3 vs GCC 3.3.2 vs GCC 3.4pre vs Sun ONE Studio 8 CC 
on a Sun Fire V250.  The complete results are attached.

Summary:
- the Sun compiler produces very large (stripped) binaries with -fast, whose 
size is only matched by GCC with -funroll-all-loops (this latter option 
doesn't help wrt to -O2 and -O3, except for GCC 3.3.2).
- in 32-bit mode, GCC 2.95.3 code runs slower than Sun CC code (time delta: 
13.5%), while GCC 3.3.2 code is on par with GCC 2.95.3 code (but only with  
-funroll-all-loops).  GCC 3.4pre code is always faster than the 3 others 
(time delta: -2.5% against Sun CC).
- in 64-bit mode, GCC 3.3.2 code runs slower than Sun CC code (time delta: 
20%) and GCC 3.4pre code runs slower than Sun CC code too, although by 
roughly half the previous margin (time delta: 11%).


The results for GCC 3.4pre are pretty encouraging (amazing ?), especially in 
32-bit mode, but even in 64-bit mode given the larger initial gap.  I'm 
going to try to figure out where the difference comes from.  Any ideas?

-- 
Eric Botcazou

Attachment: sparc_gzip.txt
Description: Text document