On ppc it is hard to tell what is faster. For example loading a 64-bit register may take up to 4 instructions. Small immediate values (up to 25-bit I believe) don't take anymore due to sign extension. Also if your data structures in memory are 64-bit now instead of 32-bit you pay a cache and memory bandwidth penalty.

On intel you actually often win with 64-bit. As long as you do not change the layout of your structures in memory to 64-bit then you do not pay the cache and memory bandwidth penalties, but you do get 16 GPRs (general purpose registers). In 32-bit code you have 8 GPRs (handwavingly since the 286 4 are very general purpose, the other 4 have some specialness to them). I actually saw up to 30% increase in performance recompiling some of my code as x86_64.

[ Reply to This | # ]