https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92925
--- Comment #6 from Eric Botcazou <ebotcazou at gcc dot gnu.org> --- > So I wonder why we bother to feed 'bitpos' to > set_mem_attributes_minus_bitpos when we expect the MEM_ATTRS to be > created from solely 'T' (do we?) and then later apply 'bitpos' via > adjust_address or friends. Because we derive information from T itself for MEM_EXPR and MEM_OFFSET, which means that we need to preventively subtract BITPOS as I showed; if we don't do the latter, the outcome of adjust_address would be incorrect. If we don't want to pass BITPOS, then we need to get rid of all the cases setting APPLY_BITPOS in set_mem_attributes_minus_bitpos, i.e. basically derive nothing from T. That being said, I agree that the mechanism is fragile.
