[openacc] implicit non-scalars data mapping in kernels

This patch fixes a bug where the implicit data mapping for "non-scalar"
variables in acc kernels regions uses a copy clause instead of a
present_or_copy when the variable is a reference type. This situation
can arise in fortran, when a dummy array is used inside a kernels
region. I noticed that there was already code in place to handle pointer
types for parallel data mappings, so in this patch I moved that type
extraction earlier so that both kernels and parallel regions can use it.

Is this patch ok for trunk?

Thanks,
Cesar

On Wed, Jan 13, 2016 at 02:29:21PM -0800, Cesar Philippidis wrote:
> --- a/gcc/gimplify.c
> +++ b/gcc/gimplify.c
> @@ -5994,6 +5994,11 @@ oacc_default_clause (struct gimplify_omp_ctx *ctx, tree decl, unsigned flags)
>  {
>    const char *rkind;
>    bool on_device = false;
> +  tree type = TREE_TYPE (decl);
> +
> +  if (TREE_CODE (type) == REFERENCE_TYPE
> +      || POINTER_TYPE_P (type))

I think this should be just POINTER_TYPE_P--this macro checks for
REFERENCE_TYPE as well.

	Marek

On Wed, Jan 13, 2016 at 11:35:08PM +0100, Marek Polacek wrote:
> On Wed, Jan 13, 2016 at 02:29:21PM -0800, Cesar Philippidis wrote:
> > --- a/gcc/gimplify.c
> > +++ b/gcc/gimplify.c
> > @@ -5994,6 +5994,11 @@ oacc_default_clause (struct gimplify_omp_ctx *ctx, tree decl, unsigned flags)
> >  {
> >    const char *rkind;
> >    bool on_device = false;
> > +  tree type = TREE_TYPE (decl);
> > +
> > +  if (TREE_CODE (type) == REFERENCE_TYPE
> > +      || POINTER_TYPE_P (type))
> 
> I think this should be just POINTER_TYPE_P--this macro checks for
> REFERENCE_TYPE as well.

Unless the spec says that for all pointers you should consider what it
points to, this really looks weird to me.

Say for OpenMP, I usually use
lang_hooks.decls.omp_privatize_by_reference (decl)
hook for similar purposes where the FEs have the possibility to say
if you are interested in the pointer/reference itself or what it points to,
which is just always false for C, for C++ REFERENCE_TYPE vars and
invisiref stuff, and for Fortran various.
So I'd expect to see instead
  tree type = TREE_TYPE (decl);
  if (lang_hooks.decls.omp_privatize_by_reference (decl))
    type = TREE_TYPE (type);
or so.  Just looking at TREE_TYPE of arbitrary POINTER_TYPE_P has also the
disadvantage that the pointer can be void *, or pointer to incomplete type,
...

	Jakub

On 01/13/16 17:44, Jakub Jelinek wrote:
> On Wed, Jan 13, 2016 at 11:35:08PM +0100, Marek Polacek wrote:
>> On Wed, Jan 13, 2016 at 02:29:21PM -0800, Cesar Philippidis wrote:
>>> --- a/gcc/gimplify.c
>>> +++ b/gcc/gimplify.c
>>> @@ -5994,6 +5994,11 @@ oacc_default_clause (struct gimplify_omp_ctx *ctx, tree decl, unsigned flags)
>>>   {
>>>     const char *rkind;
>>>     bool on_device = false;
>>> +  tree type = TREE_TYPE (decl);
>>> +
>>> +  if (TREE_CODE (type) == REFERENCE_TYPE
>>> +      || POINTER_TYPE_P (type))
>>
>> I think this should be just POINTER_TYPE_P--this macro checks for
>> REFERENCE_TYPE as well.
>
> Unless the spec says that for all pointers you should consider what it
> points to, this really looks weird to me.

I think the spec is silent.  I remember looking at this before and thinking 
looking through pointers was strange.  IMHO it should behave in a manner as 
similar as possible to OpenMP (which I presume is precise about whether one 
should look through arbitrary pointer types).

Is this a case of the Fortran FE OpenACC bits not annotating some artificial 
object correctly?

nathan