Of course, it helps if you tell gcc to generate a 64 bit executable

... when you're using 64 bit operations:

tercel-2:$ time bin32/tbray4 datasets/hundred-o10k.ap 
user    0m5.644s
tercel-2:$ time bin32/tbray6 datasets/hundred-o10k.ap 
user    0m4.048s
tercel-2:$ time bin/tbray4 datasets/hundred-o10k.ap 
user    0m5.725s
tercel-2:$ time bin/tbray6 datasets/hundred-o10k.ap 
user    0m3.839s

I'd assumed it would default as it does on linux, but no. Which also explains why I was confused about sizeof(size_t) being less than sizeof(long long) here.

It's still around 8 clock cycles per byte - the 5% gain is not enough to alter yesterday's estimates, which is a little encouraging actually, as I wasn't sure whether or not 32 bit Sparc instruction generators, such as gnu lightning (used in rubinius) or whatever erlang's hipe uses would cause a significant slow down.


Pete