By: Paul A. Clayton (paaronclayton.delete@this.gmail.com), May 6, 2013 9:35 pm
Room: Moderated Discussions
rwessel (robertwessel.delete@this.yahoo.com) on May 6, 2013 9:04 pm wrote:
[snip]
> Page 2 specifies a 4 entry RSB, and page 3 a 16 entry RSB.
These are two different RSBs. The first is associated with the (fast/early) BTB predictor and "can recover from certain types of corruption caused by branch mispredictions". The second is associated with the gshare predictor and "there is no RSB renaming". (I seem to recall that at least one mainstream x86 used two RSBs one with fancy indirection and the other being a simple stack, though both were [I think] the same size. Silvermont's reduced speculation would presumably make the tradeoffs in sizing different than for mainstream x86.)
[snip]
> Page 2 specifies a 4 entry RSB, and page 3 a 16 entry RSB.
These are two different RSBs. The first is associated with the (fast/early) BTB predictor and "can recover from certain types of corruption caused by branch mispredictions". The second is associated with the gshare predictor and "there is no RSB renaming". (I seem to recall that at least one mainstream x86 used two RSBs one with fancy indirection and the other being a simple stack, though both were [I think] the same size. Silvermont's reduced speculation would presumably make the tradeoffs in sizing different than for mainstream x86.)