Re: New patch: [AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe.

[Christophe Lyon <christophe.lyon@linaro.org>]

On 25 January 2015 at 22:33, Christophe Lyon <christophe.lyon@linaro.org> wrote:
> On 17 December 2014 at 18:02, Tejas Belagod <tejas.belagod@arm.com> wrote:
>> On 17/12/14 16:46, Marcus Shawcroft wrote:
>>>
>>> On 17 December 2014 at 15:15, Tejas Belagod <tejas.belagod@arm.com> wrote:
>>>
>>>>> It isn;t clear to me how far through the various BE patches we need to
>>>>> get before 59810 is actually resolved?
>>>>
>>>>
>>>>
>>>> David's 2 patches
>>>>
>>>> https://gcc.gnu.org/ml/gcc-patches/2014-11/msg01431.html
>>>> https://gcc.gnu.org/ml/gcc-patches/2014-10/msg01099.html
>>>>
>>>> and Alan's 2 patches:
>>>>
>>>> https://gcc.gnu.org/ml/gcc-patches/2014-11/msg02797.html
>>>
>>>
>>> OK, thanks, my understanding is that all of the above are now blocked
>>> waiting some resolution on this patch to rtlanal.c:
>>
>>
>> I believe so.
>>
>> Thanks,
>> Tejas.
>>
>
> Hi,
> If I'm not mistaken, this has been committed as r219959, and is
> causing regressions on aarch64_be for several AdvSimd intrinsic tests:
> vldX_lane, vtrn, vuzp, vzip, as well as vldN_1 and vstN_1
>
> See:
> http://abe.tcwglab.linaro.org/logs/validations/cross-validation/gcc/trunk/219959/report-build-info.html
>
> These tests started passing at the previous commit (r219958) with the
> other half of this patch.
>
> I haven't looked at the details yet.
>

I've been looking at these failures for some time. I currently have a
patch which mostly reverts the one discussed in this thread, and makes
the AdvSIMD tests pass.

However, it creates regressions in some of the vectorizer tests. When
r219959 was committed, I observed that these tests started to pass:
gcc.dg/torture/pr52028.c
gcc.dg/torture/pr53366-1.c
gcc.dg/vect/pr37539.c
gcc.dg/vect/pr40074.c
gcc.dg/vect/pr51074.c
gcc.dg/vect/pr59354.c
gcc.dg/vect/pr64252.c
gcc.dg/vect/slp-12b.c
gcc.dg/vect/slp-19b.c
gcc.dg/vect/slp-perm-8.c
gcc.dg/vect/slp-perm-9.c
gcc.dg/vect/vect-107.c
gcc.dg/vect/vect-over-widen-1-big-array.c
gcc.dg/vect/vect-over-widen-1.c
gcc.dg/vect/vect-over-widen-2-big-array.c
gcc.dg/vect/vect-over-widen-2.c
gcc.dg/vect/vect-over-widen-3-big-array.c
gcc.dg/vect/vect-over-widen-3.c
gcc.dg/vect/vect-over-widen-4-big-array.c
gcc.dg/vect/vect-over-widen-4.c
gcc.dg/vect/vect-strided-store-a-u8-i2.c
gcc.dg/vect/vect-strided-store-u16-i4.c
gcc.dg/vect/vect-strided-store-u32-i2.c
gcc.dg/vect/vect-strided-u16-i3.c

I am looking at the pr59354.c testcase, which seems to be one the
simplest showing a regression. The vectorizer does a pretty good at
generating obfuscated code :-) and I find it quite difficult to debug,
given the lack of debug environment.

In fact, I still don't understand what r219959 actually fixed. I
believe that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59810 is
covered by the AdvSIMD tests, which r219959 actually breaks.

In summary, I have the feeling that r219959 should be somewhat
reverted such that the AdvSIMD tests pass on aarch64_be, but that this
would expose a bug related to vectorization.

Any advice appreciated.

Thanks,

Christophe.

> Christophe.
>
>>>
>>>> https://gcc.gnu.org/ml/gcc-patches/2014-12/msg01087.html
>>>
>>>
>>> /Marcus
>>>
>>
>>