------- Additional Comments From greenrd at greenrd dot org 2005-08-30 01:17 ------- (In reply to comment #7) > This seems like something glibc's memcmp should be doing also, could > you report a bug to glibc about this comparison?
Actually I was wrong - it's not glibc's memcmp that's being used here, it's the gcc inlined memcmp (see http://sources.redhat.com/bugzilla/show_bug.cgi?id=1262 ), which uses the "CISC-style" instructions REP MOVSB. This performs badly in _every_ case on athlon-xp compared to a "RISC-style" loop. (Indeed this is documented - it's surprising that gcc has been using the wrong substitution for so long!) So I will submit a more conservative bug about changing the inlined memcmp into a loop, against gcc. And will test it on pentium 4 as well. The block-compare is harder to argue for at the gcc level because it is quite complicated for the case of memcmp on big-endian (see the glibc implementation) and therefore probably isn't a good case for inlining, at least not on x86. However, the block-compare makes perfect sense for the simpler case of java.lang.String.equals, as my results show. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23495
