[RFA] Generic stack alignment for user variables, v2

[H.J. Lu]

On Fri, Nov 5, 2010 at 8:32 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Wed, Oct 6, 2010 at 8:57 AM, Richard Henderson <rth@redhat.com> wrote:
>> On 10/01/2010 03:55 PM, Richard Henderson wrote:
>>> This patch implements the scheme described in
>>>
>>>   http://gcc.gnu.org/ml/gcc-patches/2010-08/msg01406.html
>>>
>>> Given that no target besides x86 actually implements general
>>> stack realignment, this fixes the problem of ignoring the alignment
>>> requirements of a given type when it's allocated on the stack.
>>>
>>> This only handles DECLs; it makes no attempt to align the full
>>> stack frame.  Which is ok so long as the target does not have
>>> machine modes (think SSE) that require super-alignment.
>>>
>>> An obvious future direction here is to define a
>>>
>>>   __builtin_alloca_align (size_t size, size_t align)
>>>
>>> and have all of the VLA allocations in gimplify.c and omp-low.c
>>> use that instead of bare __builtin_alloca.  In some cases that
>>> should be able to eliminate some extra padding and alignment code.
>>> In others, it will provide proper alignment for super-aligned VLAs.
>>>
>>> It has been tested on x86_64-mingw32 (which does not have the same
>>> stack realignment support that the rest of the x86 port does), and
>>> alpha-linux cross.  Native testing is proceeding on ppc64, sparc64,
>>> and ia64-linux.
>>
>> The patch has now passed bootstrap and testing on i686, x86_64, ppc64,
>> sparc64, and ia64-linux.
>>
>> The last failures that Jack Howarth mentioned elsewhere in this thread
>> were common to all i686 targets as I hadn't realized the MINIMUM_ALIGNMENT
>> hook could actually reduce the alignment of a variable.  Thus the full
>> difference between this and the most recently posted fragment is
>>
>> -+      /* Make sure we're only increasing alignment.  */
>> -+      gcc_assert (DECL_ALIGN (decl) <= align);
>> ++      /* One would think that we could assert that we're not decreasing
>> ++       alignment here, but (at least) the i386 port does exactly this
>> ++       via the MINIMUM_ALIGNMENT hook.  */
>>
>
> This patch caused:
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46325
>

This also caused:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47581


-- 
H.J.