On Tue, Apr 17, 2012 at 1:48 AM, Jan Hubicka <hubi...@ucw.cz> wrote: >> > Note that it would make a lot of sense to teach this heuristics predict.c >> > and properly identify chars. >> >> Indeed this would be the proper place to implement this logic. > > TO a degree - switch expansion needs more info than it can obtain from edge > profile. Having > switch > case 1,3,5,7,8,9: aaa > case 2,4,6,8,10,12: bbb > to produce well ballanced decision tree, it is not enough to know how > often the value is even and how often it is odd...
Why is that? In this case, the expanded switch case does not use BST, but testing against bit patterns. > > Thus there is a need for value histograms. None of the existing value profiler will be powerful enough for this though: the one_value profiler only tracks one value. The interval profiler can potentially be used if the switch case range is small -- otherwise the runtime memory overhead will be too large. Thanks, David >> >> > Also it is possble to get an historgrams from profile feedback into >> > switch expansion. I always wanted to do that once switch expansion code >> > is cleaned up and moved to gimple level... >> >> Indeed. At least the parts that expand switch stmts to (balanced) trees >> should be moved to the GIMPLE level, retaining only the table-jump-like >> expansions as switch stmts. > > Yep. > Honza >> >> >> >> >> >> >> The attached patch removes the heuristic. >> >> >> >> Bootstrapped and tested on powerpc-unknown-linux-gnu. OK for trunk? >> >> Ok. >> >> Thanks, >> Richard. >> >> >> Ciao! >> >> Steven >> > >> >
