In message <199704120849.KAA00471@Kyra.FutureGroove.de> you wrote:

> > In message <199704101949.VAA00692@Kyra.FutureGroove.de> you wrote:
> > 
> > > The most overhead currently is trap and interupt handling I guess. 
> > 
> > .. therefore avoiding this huge overhead.
> > 
> > If I got it right, his fpe module would replace the instruction 
> > at call time with a jump to an fp function - so the next time 
> > the same bit of code is executed, the exception/decoding etc. doesn't
> > take place. 
> 
> This would result in a 'self modifying code'. Since the invention of the 
> instruction cache in the processors this (my meaning) is no longer efficient 
> enough because you have to switch off the cache for this (I think someone else 
> said this already some days before in this thread).

Yes, this is slow, of course - but it would only happen *once* for every
instruction. But then, it would be even better to do it at load time, 
before the code is executed at all.

> A better solution is to write it as softfloat library for the gcc. (can not a 
> existing softfloat library from a other machine be compiled/ported to our 
> system ?)

Yes, but *if* an fpu appears for the SA (well, uhmmm, perhaps not), you'll
have to recompile everything.

> Next idea: If a DSP would be present in the system, would this improve the FP 
> performance if the floatingpoint requests from a (special) softfloat library
> would be calculated by the DSP ?
> I'm not shure whether the encoding/transmitting/decoding overhead for this 
> mechanism still keeps the speed improvement of the DSP. (Native DSP soft is 
> regions faster).
> (I ask this not without a reason ;)

My knowledge of DSPs is somewhat limited, but AFAIK, DSPs are good at handling
SIMD type tasks, ie do the same processing to lots of data. And, don't 
forget this, there are integer DSPs and FP DSPs as well. Wether a DSP 
will be usefull for FP calculations will depend on this - and I think 
it isn't very likely that there will be an FP DSP for the ARM before 
there'll be a "normal" FPA.

Of course an integer DSP could be usefull as well (I'm not that familiar 
with the algorithms used), but the same can be said for any faster 
processor - ie a DSP wouldn't "change everything".

Kind regards,


-- 
Thomas Boroske