Re: [PATCH, PR46032] Handle BUILT_IN_GOMP_PARALLEL in ipa-pta

[Christophe Lyon <christophe dot lyon at linaro dot org>]

On 30 November 2015 at 18:55, Tom de Vries <Tom_deVries@mentor.com> wrote:
> On 30/11/15 17:48, Jakub Jelinek wrote:
>>
>> On Mon, Nov 30, 2015 at 05:36:25PM +0100, Tom de Vries wrote:
>>>
>>> +int
>>> +main (void)
>>> +{
>>> +  unsigned results[nEvents];
>>> +  unsigned pData[nEvents];
>>> +  unsigned coeff = 2;
>>> +
>>> +  init (&results[0], &pData[0]);
>>> +
>>> +#pragma omp parallel for
>>> +  for (int idx = 0; idx < (int)nEvents; idx++)
>>> +    results[idx] = coeff * pData[idx];
>>
>>
>> Could you please add another testcase, where you have say pData
>> and some other pointer that init sets to alias with pData, and verify
>> that such loop (would need to be say normal loop inside #pragma omp single
>> or master) is not vectorized?
>
>
> I've:
> - added a simpler (not vectorizer-based) version of the testcase as
>   pr46032-2.c, and
> - copied pr46032-2.c to pr46032-3.c and modified it such that two
>   pointers are aliasing
>
> Committed to trunk.
>

Hi,

I've committed the attached patch as obvious: it adds
dg-require-effective-target fopenmp to these new tests
so that they are skipped e.g. on arm bare-metal targets
(using newlib).

Note that pr46032.c has some failures:
FAIL:  gcc.dg/pr46032.c scan-tree-dump-times vect "note: vectorized 1 loop" 1
on arm-none-linux-gnueabi, on arm-none-linux-gnueabihf with -mfpu=vfp*,
and on armeb-none-linux-gnueabihf

I haven't looked at the details yet; see
http://people.linaro.org/~christophe.lyon/cross-validation/gcc/trunk/231076/report-build-info.html
for more information.

Thanks,

Christophe.

2015-12-01  Christophe Lyon  <christophe.lyon@linaro.org>

    * gcc.dg/pr46032.c: Add dg-require-effective-target fopenmp.
    * gcc.dg/pr46032-2.c: Likewise.
    * gcc.dg/pr46032-3.c: Likewise.


> Thanks,
> - Tom
>

Attachment: pr46032.patch.txt
Description: Text document