[PTLsim-devel] Is a perfect branch predictor possible?

Fri Nov 2 02:51:29 EDT 2007

On Wednesday 31 October 2007 19:19, Krist Sun wrote:
> Hi Matt,
> I am working on a PTLsim LSQ project and the simulation result is very
> frustrating. I believe the branch predictor mistakenly flushed the pipeline
> too often. So I am looking for a way to tackle this issue. Is it possible
> to use a perfect branch predictor in PTLsim?
>

Since PTLsim is an execution simulator, the outcome of a branch is not known 
until it's actually executed in the ALU stage, just like a real processor. 
There's no easy way to model perfect branch prediction in this design.

If you really need this feature, you could probably write all branch outcomes 
passing through the commit stage into a file, then run the simulation again, 
but replace the fetch stage's predictions with the branch directions recorded 
in that file. It won't be perfect, and it may get out of alignment if you 
modify the microarchitecture after doing this. Also remember that things like 
load/store aliasing can cause uops to be refetched; the recorded branch 
predictions have no way to account for this.

- Matt

-------------------------------------------------------
 Matt T. Yourst                    yourst at peptidal.com
 Peptidal Research Inc., Co-Founder and Lead Architect
-------------------------------------------------------