Hi Max, thanks for the patch. I'll add it to my queue for the development tree and make sure it gets in. I'll also check that assertion condition you mention as well and see if there is a workaround for that.
-Korey On Thu, Mar 18, 2010 at 12:52 PM, Maximilien Breughe < [email protected]> wrote: > Hi all, > > I've discovered a bug in the in-order branch prediction system. > I noticed that changing the size of the branch predictor didn't affect the > number of taken/not taken branches at all! > The reason for this is that all BTB accesses are misses. > > I figured out that the BTB actually never gets updated! > I dug into the code and saw that there are 2 squash-functions > BPredUnit::squash (in cpu/inorder/resources/bpred_unit.cc) with different > arguments. > One of them gets called by BranchPredictor::squash (in > cpu/inorder/resources/branc_predictor.cc), but in my opinion it should be > the 4-arguments one. This one will update the BTB as it should. > To clarify this I made a diff and attached it to this e-mail. > > Please note that I added a condition around the BPredUnit::squash - > function with 4 arguments. If I didn't do this an assertion would fail > during the execution. > Although my patch is working, there might be a better way to solve this > problem. Again I looked more into it to figure out what's going on. > I noticed that sometimes an instruction with the same sequence number calls > the squash-function twice! This means that the squash function tries more > than once to remove the associated predHist[tid] object. Since it's already > removed the assertion fails.. > > I've attached a small extract from the trace produced by the hello world > program. This trace was generated without applying bpredPatch. > If we analyze this we will see that instruction with sn 218 calls squash 2 > times! Once in the ID stage and once in the EX stage. > Actually a "return" instruction always causes a squash in the ID stage. Is > this squash necessary? From inorder/resources/fetch_seq_unit.cc we can see > that this also happens for other branch-instructions. > > Hopefully this was a little bit clear and helped you guys! > > > Max > > _______________________________________________ > m5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users > -- - Korey
_______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
