https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122227
--- Comment #9 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Zack Buhman from comment #7)
> This behavior is undesirable.
Which one? -O0 or -O2? Both are valid based on what how C treats alignment of
types.
> GCC does not know and can not prove the (un)alignment of `data`.
Yes that is why at -O2 get the loads/stores via byte.
For -O0 as I said the alignment is based on the type which is alignment of 4
bytes.
If you want unaligned stores there then you need to use an type which has an
aligned of 1.
Like doing:
typedef unsigned int unaligned_uint __attribute__((aligned(1)));
void test_write_a(void)
{
*(volatile unaligned_uint*)data = 0;
}
If you want to always do an aligned stores then mark the data as an aligned
like:
extern volatile unsigned char data[8] __attribute__((aligned(4)));
The behavior of GCC here might seen inconsistent but it is not because you are
invoking undefined behavior anyways due to alignment requirements of types and
such.
