http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51628
--- Comment #16 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-02 08:30:16 UTC --- (In reply to comment #14) > > What Ada does looks just like a workaround for what should be done properly > > in > > the expander. So no, IMHO we shouldn't be changing all other FEs and the > > middle-end (when it wants to generate them e.g. for memcpy) to do what Ada > > does, but we should change the expander. > > Of course I have the exactly opposite viewpoint, since I think that we should > keep the invariants we have: integral modes are naturally aligned and > TYPE_ALIGN >= MODE_ALIGN (TYPE_MODE). Breaking them for a few pathological > cases that the C compiler has shun for years doesn't seem the best course of > action. Well. I'm fine with forcing BLKmode for unaligned accesses, but not sure if that is less invasive. You'd basically have to change modes whenever more precise knowledge about alignment appears ...
