On 07/04/10 07:14, Jim Meyering wrote:

> BTW, if someone can test this patch on a 32-bit system
> on a tree with lots of hard-linked files that have inode numbers
> of 2^32 or larger, it would be instructive to see how it compares
> to the version of du before any of these changes.  I think it will
> still save memory, but how does performance change in that case?

I looked into this and found a performance hit due to the 64-bit %
in the code I proposed.  I hadn't thought the % would be a CPU bottleneck!
Anyway, I fixed this by XORing the two halves of the 64-bit dividend before
doing a 32-bit %, and this caused the resulting code to run about as fast
as coreutils 8.5 (any difference was below the resolution of the time
stamp on my "du cu*" benchmark; I artificially modified my benchmark
by adding 2**32 to all the inode numbers).

I'll submit an updated patch shortly.



Reply via email to