Re the patch you just reported in: <http://lists.gnu.org/archive/html/bug-gnulib/2010-07/msg00005.html>
I assume this is to support the du performance improvement that was proposed in <http://debbugs.gnu.org/db/65/6524.html>. I see that you incorporated a further improvement in the gnulib patch, namely, support for NULL keys. The gnulib change seems fine, but I noticed some problems with the coreutils part of that earlier proposal. Among other things, it makes du dump core on a (large) test case that I have locally. I don't know why (perhaps there were further fixes to the gnulib part that I didn't get right?). I found out about the core dump only because I had independently prepared a patch that I think is better: it uses a bit less memory and is quite a bit simpler. I'll try to merge my patch with your gnulib change and send it off to bug-coreutils in the next day or two. My patch, by the way, doesn't hash pointer keys: it just uses size_t values and casts them to void * (which is what the hash package wants). This trick works on all architectures that I know about, but it isn't guaranteed by C or POSIX and the casts are a bit offputting, so it'd be nice if the hash package supported hashing size_t keys directly. That's lower priority, though.
