Yes you are right. I'm getting a very noticeable improvement. Maybe I 
should try to recompile the VM with unsigned int instead of int in every 
longAt and longAtPut macro.

    Yoel

Richard A. O'Keefe wrote:

>Yoel Jacobsen <yoel at emet.co.il> wrote:
>	By the way, I made these tests because I was troubled by how slow squeak 
>	is on Solaris (~25M bytecode/sec on Ultra 10 440MHz, ~45M on this Blade 
>	1000)... I'm quite sure it will get no better unless squeak will be 
>	ported to 64bit.
>	
>I don't know if it applies to the latest UltraSPARCs, but in the earlier
>models, 32-bit *signed* ints are slower than 32-bit *unsigned* ints.
>When you load a 32-bit signed int, it's sign extended, which takes an
>extra cycle _unless_ the CPU was already expecting that.  Unsigned ints
>are zero extended, which doesn't take that extra cycle.  So a mix of
>signed and unsigned loads is slower than a straight set of signed loads
>or a straight set of unsigned loads.  It was quite easy to measure the
>difference.
>
>
>
>