Vendor uses -funroll-all-loops that generates slower code which is
documented in gcc(1) manpage. If user cares about small optimization
(unlike asm) then he'd probably benefit more from plain `-O3' on gcc46.

How-To-Repeat: # user time from decoding an album's mp3 directory
$ ministat -n *
x base gcc, default cflags
+ base gcc, default cflags + --with-optimization=0
* gcc46 + my cflags
% gcc46 + my cflags + --with-optimization=0
# gcc46 + my cflags + generic_fpu
@ gcc46 + my cflags + generic_fpu + --with-optimization=0
    N           Min           Max        Median           Avg        Stddev
x  10          9.39          9.62           9.5         9.493   0.078464429
+  10           9.3          9.54          9.43         9.436   0.083426614
*  10           8.7          8.92          8.78         8.781   0.071250731
%  10           8.3          8.54           8.5         8.459   0.076514341
#  10         12.97         13.23         13.12        13.104   0.094068533
@  10         12.73         12.93         12.83        12.832   0.059029183

and numeric data generated by

#! /bin/sh

times=10
dir='/path/to/album'

while [ $((i+=1)) -le $times ]; do
    find "$dir" -iname \*mp3 -exec time -p mpg123 -tq {} + 2>&1
done | while read desc time; do
    case $desc in
	user) echo $time
    esac
done