[PATCH] Fix PR78027

[Jeff Law]

On 12/09/2016 07:05 AM, Cesar Philippidis wrote:
> On 12/08/2016 04:37 PM, Jeff Law wrote:
>> On 12/08/2016 04:05 PM, Cesar Philippidis wrote:
>>> PR78027 was classified as a fortran bug, but the underlying problem
>>> turned out to be more generic. Basically, the IPA-ICF pass usually
>>> ignores functions with omp decl attributes as candidates for aliasing.
>>> Usually that works for OpenACC, because offloaded functions generated by
>>> OpenACC PARALLEL or KERNELS regions usually have an 'omp target
>>> entrypoint' attribute. However that is not the case in two situations:
>>>
>>> 1. ACC ROUTINES do not have an 'omp target entrypoint' attribute.
>>>
>>> 2. Nested offloaded regions inside 'omp declare target' functions do not
>>> have 'omp target entrypoint' attributes either.
>>>
>>> The solution I chose for this problem is to teach the IPA-ICF pass to
>>> ignore function with 'oacc *' decl attributes. Arguably 2) should be a
>>> parser error when an OpenACC offloaded region is embedded within an
>>> OpenMP offloaded function. The LTO linker does report an error for nvptx
>>> targets, but not the host. With that in mind, this patch is still
>>> necessary for case 1.
>>>
>>> Is this OK for trunk?
>>>
>>> Cesar
>>>
>>>
>>> trunk-pr78027.diff
>>>
>>>
>>> 2016-12-08  Cesar Philippidis  <cesar@codesourcery.com>
>>>
>>>     PR fortran/78027
>>>
>>>     gcc/
>>>     * ipa-icf.c (sem_function::parse): Don't process functions with
>>>     oacc decl attributes, as they may be OpenACC routines.
>>>
>>>     gcc/testsuite/
>>>     * c-c++-common/goacc/acc-icf.c: New test.
>>>     * gfortran.dg/goacc/pr78027.f90: New test.
>> This follows the same approach as we do for openmp.  This is fine for
>> the trunk.
>
> Is this also OK for gcc-6?
Your call -- I don't think folks are really using OpenACC much in gcc-6, 
but if you think it's worth the time, go ahead.

jeff