http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56434
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-26 10:03:46 UTC --- (In reply to comment #4) > MALLOC_ABI_ALIGNMENT isn't properly set on most targets though. > > E.g. with glibc, the alignment of malloced memory is 2 * sizeof (void *) > (right > now, might be bigger in the future), while MALLOC_ABI_ALIGNMENT is still 8 > bits. > > It is just fine to assume malloced memory must be at aligned enough for any of > the standard C types, so if this is say on x86_64 where long double is 16-byte > aligned, then there is really no reason to bother with aligning to 16-byte > boundaries. If you have malloc implementation that doesn't align at least > that > much, just fix it, or don't use malloc attribute for that. True, but: config/i386/i386.h:#define BIGGEST_ALIGNMENT (TARGET_AVX ? 256 : 128) I don't think glibc aligns to 32 bytes currently, so the mark_reg_pointer is definitely wrong (with -mavx). Or BIGGEST_ALIGNMENT is wrong ;) MALLOC_ABI_ALIGNMENT default was set very conservatively (because we use it to derive information about pointer values). I didn't realize we already do that via the calls.c code (and even wrong ...). Feel free to implement a better default setting for MALLOC_ABI_ALIGNMENT. > Indeed. So MALLOC_ABI_ALIGNMENT should perhaps default to the largest > alignment of all the C89 types, with platform overrides as needed? But this information is not readily accessible (it was supposed to be BIGGEST_ALIGNMENT, but -mavx changed this - not sure what the psABI says and whether this makes current glibc non-conforming to the psABI).
