Mail Archives: pgcc/1999/05/21/12:27:15
Hi!
> > It would break the x86-compatibility. The x86 only has 3 bit for the
> > registers, so you canīt have more than 8 registers without heavy
> > changes in the instruction set :-(
>
> No, it wouldn't. The ppro has much more than 8 registers through register
> renaming. Under this light having only three bits for the register might
> even improve performance due to better packed instructions.
Of course I know, that most modern x86-CPUs have hidden registers, but
IMHO
its still a difference compared with having direct access to these
registers.
If you have more variables than registers, than you have to store these
variables in memory, which is more expensive (load+alu instead of alu
only),
even if stored in fast cache. And additionally its probably quite hard
for
parallelism to have only 8 independend registers but perhaps 8 or even
more
integer execution units and be forced to avoid dependencies. The hidden
registers help a bit, but more freely accessible register would be much
better IMHO :-)
> (No kidding, I expect compressed assembly in the next years, x86 already
> comes close!)
Isnīt EPIC pretty close to this? I heard, there are 3 instructions in
one
128bit value. It would come even closer to compressed assembly IMHO :-)
> > the start of the second fab in Dresden.
>
> Oh, thats great!
AMD plans to start Dresden in Q3/99 with 0.18 alumunium and convert to
0.18
copper in Q1/2000 allready :-)
Dresden is intended to produce K7 only, so the number of produced K7
should
rise quite fast :-)
> ---==---(_)__ __ ____ __ Marc Lehmann +--
cu
Jens-Uwe
- Raw text -