------- Comment #12 from ebotcazou at gcc dot gnu dot org 2008-09-08 15:25 ------- > Eric, thanks a lot for your analysis. It was very helpful.
You're welcome! > IRA uses live ranges to find conflicts for spill slots during reload. Live > ranges for r376 were wrong after IR flattening. We have the following loop > structure for the case > > l8->l7->l0 > ^ > l24->l22-| > > and we remove mem<-reg coping allocno values from l8 to l0 because allocnos > corresponding to r376 are not modified in the loops. Allocno for r376 in l0 > got live ranges only from l8 and l7 but not from l22 ... Therefore IRA does > not find conflicts with another allocno (r90) in l22 and l24. Yes, my understanding was that the flattening code currently assumes that the moves are always generated at the borders. > I'll send a patch to solve the problem today. Just removing the optimization > (removing unnecessary mem<-reg) is not wise because it has significant impact > on the generated code performance. OK, thanks for the explanation. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37377
