[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: ramblings about old hardware, gzip, bz2, and pentium opts

Tomasz Wegrzanowski wrote:
>   `pentium optimisations slows down 99% of all systems'
> You claim that Celeron / Cx 6x86 / PIII / K6 / Merced etc.
> are faster with 386-optimized soft than with Pentium-optimized soft
> Do you have any stats to prove it ???
I'd say 99 might be a small over estimate, but the guy that made this claim
seems to be talking within the context of implementation dependant
optimzations.  In the past pgcc has been brought in on these discussions. 
Could someone clarify if pgcc does that kind of optimization?  If it does
the likelyhood of a performance increase on non-Intel Penitums chips is
really decreased.  

The Cyrix and AMD chips are different implementations of the 586+
instruction(for which there is a 568/686 cpu code) set meaning different
instruction timings, different order dependent instruction timings, etc. 
Though they do that same thing they work differently inside.  In order for
all these chips to recieve the same benefit from an implementation
dependent optimization, the relative timings between instructions must be
similar and preferably exactly within a scalar multiple of a true pentium
core.  Celerons/PII/PIII are more likely to reciving a boost, assuming
their core was a directly descendant.  They are all essentially PII core in
some cases with addition SSE instructions.  I can't remember how much the
PII core was overhauled coming up from the pentium, if too much was done
then the optimizations are highly likely to be not as near effective(or
effective at all) enough to warrent doing the whole distribution.  If the
gain is quite a bit, it may prove benefical to just do essental packages. 
e.g. The ones in required. 

As for the mecred the likely hood of improvement would probably be very
small.  It is not likely Intel could just add to a existing design, they
probably did a good portion of it from the ground up. An optimzated set of
packages for the 64-bit intels instruciton set would probably be more
benifical then a simple pentium optimzation in the long run as 64-bit
systems(Merced, Sledgehammer..) become more common place, anyway.

Philip Thiem /---/ ptt@umr.edu /---/ Pass on the GAS get NASM instead
    Computer Science & Mathematics Undergraduate @ UM-Rolla
    Interests: Security, Operating Systems, Numerical Computing,
               Algorithm Analysis, Discrete/Linear/Modern Algebra,

Reply to: