On Mon, Dec 12, 2011 at 12:40 PM, Eric Botcazou <ebotca...@adacore.com> wrote:
>> Well, I can live with this change (though I cannot approve anything).
>> On the other hand, the real underlying problem is that expander cannot
>> handle unaligned MEM_REFs where strict alignment is required. SRA is
>> of course much more prone to create such situations than anything else
>> but I wonder whether they can creep up elsewhere too. It also takes
>> us in the opposite direction than the one initially intended with
>> MEM_REFs, doesn't it?
>
> Certainly, but we need to fix the regression in a relatively safe manner.
>
>> That said, I looked into the expander briefly in summer but given my
>> level of experience in that area I did not nearly have enough time. I
>> still plan to look into this issue in expander but for the same
>> reasons I cannot guarantee any quick success. So I acknowledge this is
>> the only working approach to a long-standing difficult bug... and most
>> probably the most appropriate for the 4.6 branch.
>
> Thanks. This is still the same very old issue: misalignment cannot be handled
> indirectly (because we don't really have misaligned pointers) so MEM_REFs can
> be used safely only when everything is properly aligned.
We do have misaligned accesses - TYPE_ALIGN of TREE_TYPE of
the MEM_REF reflects that. Similar to how would do
typedef int myint __attribute__((aligned(1)));
int foo (myint *p)
{
return *p;
}
which is a testcase that is "miscompiled" since forever on
STRICT_ALIGNMENT targets (well, maybe apart from now for
those who implement movmisalign).
The fix is to fix the above testcase (which is a good idea anyway)
and then to make sure to transition misaligned information to
TREE_TYPE of the MEM_REF we create.
Richard.
