On Tue, Apr 23, 2013 at 12:27:43PM +0200, Richard Biener wrote: > On Mon, Apr 22, 2013 at 5:42 PM, Andi Kleen <a...@linux.intel.com> wrote: > > On Mon, Apr 22, 2013 at 01:46:58PM +0200, Richard Biener wrote: > >> On Fri, Apr 19, 2013 at 11:31 PM, Andi Kleen <a...@firstfloor.org> wrote: > >> > From: Andi Kleen <a...@linux.intel.com> > >> > > >> > For a large LTO test case The previous pointer hash change brought > >> > the collision rate for the WPA gimple type hash table from 90% to > >> > 70. This patch uses the well known murmur3 to improve it further > >> > to 64%. > >> > >> But if they are pointers then pointer_hash should be good enough... ? > > > > The original pointer hash (ptr >> 3) % hashsize and throwing away most bits > > is > > very poor. > > > > The evahash based on I sent earlier is better, but murmur3 is even better > > than > > that, at least for this case. > > I'd rather not have different pointer hashes for things where there isn't a > fundamental difference between the pointer values.
One of the reasons I did it explicitely is that the murmur3 reference code is C++ (well really only the mixed code/declarations I think) and libiberty seems to be C only. I suppose can port it to C and put it into libiberty though, and make hashtab.c always use it. My understanding is that murmur is generally stronger than evahash. -Andi -- a...@linux.intel.com -- Speaking for myself only
