> On Wed, Mar 6, 2013 at 5:02 PM, Jan Hubicka wrote: > >> > >> An interesting question is, how can you identify bitmaps that could > >> benefit from viewing it as a tree (at least for some time). I have no > >> ideas here. Something with detailed memory stats, of course, but then > >> how to derive some measure for a trade-off between list or tree view. > >> Thoughts? > > > > Well, I guess we can simply accumulate the counter on linked list walks > > (when > > the one element cache is missed) and divide it by number of iterations. > > Where > > this number thends to grow and not be counstant bounded, we have nonlinear > > behaviour, right? > > Well, yes and no. This is not fine-grained enough to see if there are > specific usages of bitmaps that can be better represented as a > linked-list or as a tree. But yes, it's a place to start, and it's > what I've started doing last night for a set of files (large PR test > cases, cc1-i files, etc.).
Well, I meant with the --enable-gather-detailed-stats code. it should give you data about individual bitmaps used thorough the compiler. Honza > > Ciao! > Steven
