> Is this piece of code necessary for generic Squashes?
>
I'd agree that this does not seem correct and may be something exposed with
the deeper pipeline and recently changed fault-checking mechanisms.
The code should be OK with squashing all the younger instructions after
either a misspeculated branch or an event that causes the pipeline to be
flushed/squashed. It sounds like that code is assuming that only the first
case is the squash possibility.
Maybe the code should do something like this:
if (pred_hist.front().seqNum != squashed_sn) {
DPRINTF(InOrderBPred, "Front sn %i != Squash sn %i.\n",
pred_hist.front().seqNum, squashed_sn);
assert(pred_hist.front().seqNum > squashed_sn);
pred_hist.erase(hist_it);
return;
}
Since the rest of the code is updating the BTB, it seems that it should
erase the prediction history at that point. I'm actually thinking that a
pred_hist.clear() might be more appropriate since if you have a fault in
the last stage of the pipeline, it should be safe to clear all prediction
history from the pipe.
In any case, those would be the potential fixes that I see offhand. Can you
try them out?
--
- Korey
_______________________________________________
gem5-users mailing list
gem5-users@gem5.org
http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
