Re: [Gems-users] Regarding out of order stores in Opal

[Milo Martin <milom@xxxxxxxxxxxxx>]

One of the main advantages of out-of-order execution is to hide  
memory latency by prefetching data into the caches. Yes, sometimes  
this pollutes the cache, but as branch prediction is usually correct,  
this generally results in a performance improvement.
In many processors, TLB misses are not handled speculatively.  One  
result of this is that some wild prefetches (those without TLB  
mappings) are filtered polluting the cache (because they fail  
translation).
- Milo

On Jun 22, 2007, at 5:00 AM, salil pant wrote:

> Hi, I have noticed that stores get issued to the memory system during
> the execute stage, which means they might be issued to the memory out
> of order.
> However, stores dont really perform and act as prefetches to RUBY ( or
> any other memory model) which means upon a branch mispredict rolling
> back a store will have no correctness consequences on the memory
> system .
> However, are we not polluting the cache unnecessarily due to this ?
> If the frequency of branch mispredicts is high, I am concerned about
> the timing fidelity of Opal with respect to the caches. ( Since
> excessive stores may change the timing of subsequent loads)
> I have noticed in the SIMICS MAI that stores get issued only during
> the retire phase at which point they are guaranteed to be correct thus
> avoiding the possibility of a rollback.
> Am I correct on all of this ?
> I just wanted to confirm my analysis before moving ahead with Opal
> thanks
> smp
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
> Use Google to search the GEMS Users mailing list by adding  
> "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.
--
Milo M. K. Martin (milom@xxxxxxxxxxxxx)
http://www.cis.upenn.edu/~milom/
Assistant Professor
Computer and Information Sciences Department
University of Pennsylvania