On 09/06/15 00:05, Tom de Vries wrote:
> On 08/06/15 17:55, Thomas Schwinge wrote:
>> Hi Tom!
>>
>> On Mon, 8 Jun 2015 12:43:01 +0200, Tom de Vries
>> <Tom_deVries@mentor.com> wrote:
>>> There are two problems in try_transform_to_exit_first_loop_alt:
>>> 1. In case the latch is not a singleton bb, the function should return
>>>      false rather than true.
>>> 2. The check for singleton bb should ignore debug-insns.
>>>
>>> Attached patch fixes these problems.
>>
>>> Fix try_transform_to_exit_first_loop_alt
>>
>>>     PR tree-optimization/66442
>>>     * gimple-iterator.h (gimple_seq_nondebug_singleton_p): Add function.
>>>     * tree-parloops.c (try_transform_to_exit_first_loop_alt): Return
>>> false
>>>     if the loop latch is not a singleton.  Use
>>>     gimple_seq_nondebug_singleton_p instead of gimple_seq_singleton_p.
>>
>> Per my testing, the backport of this patch that you committed to
>> gomp-4_0-branch, r224219, introduces a number of regressions in your
>> OpenACC kernels test cases, specifically the »scan-tree-dump-times
>> parloops_oacc_kernels "(?n)pragma omp target
>> oacc_parallel.*num_gangs\\(32\\)" 1« tests.  Would you please have a
>> look?
>>
>>
>
> Hi Thomas,
>
> I seem to have committed (to both trunk and gomp-4_0-branch) an older
> version of the patch, which contained an incorrect version of
> gimple_seq_nondebug_singleton_p.
>
> I'll correct the mistake tomorrow morning.

Committed attached patch to trunk and propagated to gomp-4_0-branch.

Committed as obvious, since it changes gimple_seq_nondebug_singleton_p 
into the tested and approved version.

Thanks,
- Tom