|
From: Josef W. <Jos...@gm...> - 2013-01-04 20:21:36
|
Am 28.12.2012 11:52, schrieb Julian Seward: > > Hi Josef, > > IIRC you posted a message some time back about some experiments you > did, which allow handling of strange (non-power-of-2) sized LLs in a > more generic way, and would allow the removal of existing kludges that > change the associativity vs #lines. Yes. I had to search a bit myself. It's attached. The patch simply uses the modulo operation for mapping addresses to cache sets for the LL. > What's the status on this? I was undecided on what to do. On average, it slows down cachegrind by 5% on amd64, with the worst case being 17% for ffbench (see below). If the perf loss is small, it would be good > to get this landed, so that we don't have future breakage on strange > sized caches. I agree, but just unconditionally always using modulo for LL (as the patch does now) seems not to be the best solution. Alternatives below. Josef > > Thanks, > > J > |