On Sat, Oct 13, 2012 at 11:05 PM, Eric Botcazou wrote: >> I think it would be a good idea to keep things unchanged at -O1. For >> that, the patch needs a few minor modifications (remove calls to >> df_live_add_problem and make some code to update DF_LIVE_{IN,OUT} >> conditional). I can prepare an updated patch for that, if you think >> that's best. > > That was just a proposal to be put on the table. :-) Others, especially Vlad > of course, have the final say. And I somewhat sympathize with the cautious > approach here, because we were forced to downgrade from DF_LIVE to DF_LR for > resource.c at some point (but of course it's resource.c so...).
Right, that was for PR40710 (http://gcc.gnu.org/ml/gcc/2009-07/msg00241.html) because resource.c computes its own idea of liveness and it uses a similar definition as the DF_LR problem. Using incompatible views of liveness is not a good idea... For IRA it doesn't matter whether you use DF_LR or DF_LIVE, as long as you use one or the other consistently. Ciao! Steven