[kaffe] Re: INTRP, JIT1 & JIT3 on m68k platform

[Tony Wyatt]

Hi Gwenole,

On 06-May-03, you wrote:

> Do you conjecture that JIT3 is slower than JIT1 on platforms with a decent
> amount of registers to allocate to translated blocks? I am just wondering
> since I am still hesitating for the AMD64 JIT.
> 
You've got me there. I'll have to go try it on a machine with no registers
to speak of (Intel) and see what the results are. The above results are of
course on an m68k with all eight integer registers and five address
registers available. Also it's a very rudimetary cache flushing process
(flush the lot, every time) which could be improved, and degrades the JIT
performance.

I haven't looked at the translated code to see if it would benefit from
different register allocations. That's a bit over my head. I might add some
peephole optimisation, though.

tony