https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67288
--- Comment #16 from Segher Boessenkool <segher at gcc dot gnu.org> --- Many things in combine assume that they can move an earlier insn to later, if none of the registers it uses is set in the intermediate insns (etc.) This isn't correct if you make combine work on EBBs. Keeping the reg_stat info based on dominance relations doesn't work the way things are set up now. I have long wished for a dataflow problem that would compute known 0/1/ext bits, or value ranges, etc. Ideally this can replace reg_stat completely. This would be both more powerful and simpler and much less problematic code :-)
