Hi Sandra,
good work! I tested with arm-none-eabi. Now both variants, with and without -fstrict-volatile-bitfields produce the expected results for EABI-compliant structures, and still create reasonable code with non-compliant packed structures. > I think there is still a case where this version of the patch could > still end up reading or writing past the end of an object. For example, > consider a packed struct that is 3 bytes long, with a bitfield that has > int type. I am not sure how to detect this situation at this point in > the code, but leaving just that case un-fixed for now seems better than > leaving the current mess in place. Perhaps it could be addressed by a > follow-up patch. Also, assuming the change in part 4 of the new series > to make -fstrict-volatile-bitfields not be the default on any target any > more is required as a condition of due to overriding the new memory > model, programmers at least will be less likely to be taken by surprise > when it does something weird with packed structures. > > Is this version OK to commit, or at least getting closer? I'd like to > wind this project up soon, somehow or another. > The patch fixes both aspects of PR56341 (unaligned access fault and write incorrect value). Regardless of -fstrict-volatile-bitfields setting. Fixes also PR48748 + PR56997, except for the access outside the structure. Apparently these accesses do not happen if the struct is an external object, they only happen if the struct is a comdat object (and the compiler possibly knows the extra bytes exist and are not used by anything?). Nevertheless this patch is a significant improvement, and those left-overs should be fixed after this is committed. Regards Bernd.
