Hmmm...a little known fact is that the current binutils and glibc 2.2.4 provide a feature called combreloc which allows glibc to sort the relocations for programs with lots of relocations like qt/kde and speed them up dramatically. That is the reason Debian abandoned objprelink since the same speedup is obtained by rebuilding glibc and all your programs against a binutils with -z combreloc enabled. Fortunately this has been easy for us with Debian sid. I believe RedHat 7.2 has their binutils with -z combreloc enabled and since they rebuild several times before shipping qt/kde should show the speed up. I'm not sure what the situation is with other linux distros. I did some benchmarks on Debian ppc sid using a glibc with hp-timing enabled and saw that the tutorials spent 90% less time in the dynamic loader than without combreloc. Of course when prelink is mature it will reduce this further. I did some benchmarks with prelink and it reduces the same tutorials down to 1% of the original time in the dynamic loader. This is because prelink actually eliminates the relocations. The down side of prelink will be that you have to prelink binaries again anytime one of their support libs have been replaced. This is pretty much what happens on SGI's Irix when you install software and requickrestart is called at the end. Finally as another example of combreloc we just had OpenOffice 641C rebuilt with the latest devtools on ppclinux and found that the coarse load time decreased from 11 sec down to 7 sec. Not bad for just needing a recompile.
Slashdot Mirror
Hmmm...a little known fact is that the current binutils and glibc 2.2.4 provide a feature called combreloc which allows glibc to sort the relocations for programs with lots of relocations like qt/kde and speed them up dramatically. That is the reason Debian abandoned objprelink since the same speedup is obtained by rebuilding glibc and all your programs against a binutils with -z combreloc enabled. Fortunately this has been easy for us with Debian sid. I believe RedHat 7.2 has their binutils with -z combreloc enabled and since they rebuild several times before shipping qt/kde should show the speed up. I'm not sure what the situation is with other linux distros. I did some benchmarks on Debian ppc sid using a glibc with hp-timing enabled and saw that the tutorials spent 90% less time in the dynamic loader than without combreloc. Of course when prelink is mature it will reduce this further. I did some benchmarks with prelink and it reduces the same tutorials down to 1% of the original time in the dynamic loader. This is because prelink actually eliminates the relocations. The down side of prelink will be that you have to prelink binaries again anytime one of their support libs have been replaced. This is pretty much what happens on SGI's Irix when you install software and requickrestart is called at the end. Finally as another example of combreloc we just had OpenOffice 641C rebuilt with the latest devtools on ppclinux and found that the coarse load time decreased from 11 sec down to 7 sec. Not bad for just needing a recompile.