https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90569
Florian Weimer <fw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |fw at gcc dot gnu.org
--- Comment #8 from Florian Weimer <fw at gcc dot gnu.org> ---
(In reply to Jonathan Wakely from comment #1)
> (I'm starting to think that __float128 support should have been disabled on
> targets where it requires greater alignment than malloc guarantees, instead
> of making GCC's max_align_t lie).
The problem is that popular mallocs do not care about ABI and return unaligned
pointers for allocations smaller than the max_align_t alignment. As a result,
__float128 support with 16-byte alignment would end up with a potentially
rather short list of supported targets (including targets which have a
16-byte-aligned long double already).
We spent a lot energy fixing this in glibc without breaking backwards
compatibility with existing Emacs binaries:
<https://sourceware.org/bugzilla/show_bug.cgi?id=6527>
Basically, we had to include a heap rewriter that converts the heap from the
old to the new format. But other mallocs wouldn't even have to do that, and
they still don't want to fix their default alignment choices unfortunately.
I also found this for the first time:
<http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2293.htm>
So if the standard moves to the weak guarantee, GCC probably shouldn't make
assumptions in the other direction.
