On 25/04/12 15:31, Richard Guenther wrote:
> On Wed, Apr 25, 2012 at 4:27 PM, Greta Yorsh <greta.yo...@arm.com> wrote:
>> Richard Guenther wrote:
>>> On Wed, Apr 25, 2012 at 3:34 PM, Greta Yorsh <greta.yo...@arm.com>
>>> wrote:
>>>> Richard Guenther wrote:
>>>>> On Wed, Apr 25, 2012 at 1:51 PM, Greta Yorsh <greta.yo...@arm.com>
>>>>> wrote:
>>>>>> The test gcc.dg/vect/slp-perm-8.c fails on arm-none-eabi with neon
>>>>> enabled:
>>>>>> FAIL: gcc.dg/vect/slp-perm-8.c scan-tree-dump-times vect
>>> "vectorized
>>>>> 1
>>>>>> loops" 2
>>>>>>
>>>>>> The test expects 2 loops to be vectorized, while gcc successfully
>>>>> vectorizes
>>>>>> 3 loops in this test using neon on arm. This patch adjusts the
>>>>> expected
>>>>>> output. Fixed test passes on qemu for arm and powerpc.
>>>>>>
>>>>>> OK for trunk?
>>>>>
>>>>> I think the proper fix is to instead of
>>>>>
>>>>> for (i = 0; i < N; i++)
>>>>> {
>>>>> input[i] = i;
>>>>> output[i] = 0;
>>>>> if (input[i] > 256)
>>>>> abort ();
>>>>> }
>>>>>
>>>>> use
>>>>>
>>>>> for (i = 0; i < N; i++)
>>>>> {
>>>>> input[i] = i;
>>>>> output[i] = 0;
>>>>> __asm__ volatile ("");
>>>>> }
>>>>>
>>>>> to prevent vectorization of initialization loops.
>>>>
>>>> Actually, it looks like both arm and powerpc vectorize this
>>> initialization loop (line 31), because the control flow is hoisted
>>> outside the loop by previous optimizations. In addition, arm with neon
>>> vectorizes the second loop (line 39), but powerpc does not:
>>>>
>>>> 39: not vectorized: relevant stmt not supported: D.2163_8 = i_40 * 9;
>>>>
>>>> If this is the expected behaviour for powerpc, then the patch I
>>> proposed is still needed to fix the test failure on arm. Also, there
>>> would be no need to disable vectorization of the initialization loop,
>>> right?
>>>
>>> Ah, I thought that was what changed. Btw, the if () abort () tries to
>>> disable
>>> vectorization but does not succeed in doing so.
>>>
>>> Richard.
>>
>> Here is an updated patch. It prevents vectorization of the initialization
>> loop, as Richard suggested, and updates the expected number of vectorized
>> loops accordingly. This patch assumes that the second loop in main (line 39)
>> should only be vectorized on arm with neon. The test passes for arm and
>> powerpc.
>>
>> OK for trunk?
>
> If arm cannot handle 9 * i then the approrpiate condition would be
> vect_int_mult, not arm_neon_ok.
>
The issue is that arm has (well, should be marked has having)
vect_char_mult. The difference in count of vectorized loops is based on
that.
R.
> Ok with that change.
>
> Richard.
>
>> Thank you,
>> Greta
>>
>> gcc/testsuite/ChangeLog
>>
>> 2012-04-25 Greta Yorsh <greta.yo...@arm.com>
>>
>> * gcc.dg/vect/slp-perm-8.c (main): Prevent
>> vectorization of initialization loop.
>> (dg-final): Adjust the expected number of
>> vectorized loops.
>>
>>
>>
>>
>
