https://gcc.gnu.org/bugzilla/show_bug.cgi?id=15596
--- Comment #34 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to rguent...@suse.de from comment #33) > Something like init-regs I'd not like. But the above should be > detectable by store-merging in some way - store-merging can > merge across "uninitialized" bits (by using zeroing or any > other convenient value). The problem is store merging might be too late. I have bit-field lowering right now right before PRE. that means we have lost information before we get to store-merging. I have been working on the regressions that bit-field lowering and store-merging testcases are the biggest one; I have decided to implement that inside reassoc instead as we are really reassociating BIT_INSERT_EXPRs and are able to optimize some of them into byte swaps, vector constructors, etc.
