Michael S (already5chosen@yahoo.com) on 8/18/09 wrote:
---------------------------
>_Arthur (arthur_@sympatico.ca) on 8/18/09 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 8/18/09 wrote:
>>---------------------------
>>>Well that's great - you found one specific combination of an algorithm and a data
>>>set where branch prediction **might not work**.
>>>
>>>Forgive me if I'm unimpressed.
>>>
>>>Also, the statistical distribution of T/NT isn't as important as the dynamic ordering.
>>>
>>>David
>>
>>There are plenty of algorithms where you're working in a tight loop on essentialy
>>random data, and prediction will NOT WORK.
>>So, you end up doing the work twice, paying the full misprediction price.
>>
>>These are not rare, nor special cases.
>>
>>Some processors, like the Sun Niagara, pay no misprediction penalty, because they
>>don't predict.
>>Of course there are also downsides to in-order execution...
>
>Strictly speaking, branch prediction and in/out-of-order execution are two orthogonal
>dimensions of processor architecture.
>

Don't even have to be strict, there's an easy counter-example: POWER6