Mail Archives: djgpp/1996/12/16/18:21:15
DJ Delorie wrote:
>
> > > Without being too facetious, the problem is that the Intel architecture is
> > > so poorly suited for such calculations that the optimal is still none too
> > > good. But yes, DJGPP is pretty good.
> >
> > Ummm...not meaning to start any wars... but your joking right? :)
> > The FPU in the pentium is pretty damn good... 3 cycles for a mul, 1
> > for an add, while being able to run in parallel with the integer
> > unit...seems pretty cool to me! :)
>
> There are other considerations besides opcode speed. For example, the
> fpu only has eight registers; other types have up to 32, and you must
> access the registers in a stack-like way, instead of through standard
> numbered register. These may also affect the overall efficiency of
> any FPU code.
Well, the stack method is pretty crumby (I used to use the 040, and it
was nice... :) but it gains some forgivness with the fxch being free,
generally being the equivalent of being able to load into any register
(IMHO BTW). I can't see this really ever changing though - Motorola
created a completely new CPU to get past the limit of the 680x0 range,
and I don't think Intel will be doing that... ;)
Leathal.
- Raw text -