------- Comment #2 from mikpe at it dot uu dot se 2008-03-16 16:49 -------
(In reply to comment #1)
> This happens with 4.1, 4.2 and trunk on old ABI. Apparently it doesn't
> happen with EABI.
>
I see the problem too, on Linux/ARM/OABI with gcc-4.1.2.
However, the problem is in the test case itself. Specifically the attempt to
align _b:
> if (((unsigned long)_b)&sizeof(_b)) _b++;
_b is void*, so this is "if (_b & 4) _b += 1;" at the machine level, which is
totally bogus. Om my machine alloca() returned 0xbea1d97c (a multiple of 4 but
not of 8), which this code changed to 0xbea1d97d. The following uses of _b as a
pointer caused numerous alignment exceptions (logged by the kernel).
On the other machines I tested this on (sparc, powerpc, x86), alloca() returned
an address that was a multiple of 8, so the bogus alignment code didn't
trigger.
If you want to align _b to a multiple of 2*sizeof(void*), do something like:
_b = (void*)(((unsigned long)_b + 2*sizeof(void*)-1) & ~(2*sizeof(void*)-1));
Not a gcc bug, IMO.
--
mikpe at it dot uu dot se changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |mikpe at it dot uu dot se
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34652
