On Fri, 2007-05-04 at 23:45 +0200, Martijn van Oosterhout wrote: > On Fri, May 04, 2007 at 03:42:45PM -0400, Tom Lane wrote: > > It's hardly credible that you could do either strcmp or strcoll in 2 nsec > > on any run-of-the-mill hardware. What I think is happening is that the > > compiler is aware that these are side-effect-free functions and is > > removing the calls entirely, or at least moving them out of the loops; > > these times would be credible for loops consisting only of an increment, > > test, and branch. > > It's not the compiler, it's the C library. strcmp and strcoll are > defined as: > > extern int strcoll (__const char *__s1, __const char *__s2) > __THROW __attribute_pure__ __nonnull ((1, 2)); > > In this context "pure" is essentially what IMMUTABLE is in postgres. > > Which doesn't change the fact that strcoll is expensive. > Thanks for clearing that up. I should have done a sanity check on those numbers to begin with. By the way, I didn't see the warning Tom mentioned using -Wall on either system. Regards, Jeff Davis
