On 10/06/15 00:15, Thomas Schwinge wrote:
> Hi Tom!
>
> On Tue, 9 Jun 2015 16:12:12 +0200, Tom de Vries <Tom_deVries@mentor.com> wrote:
>> On 09/06/15 13:07, Richard Biener wrote:
>>> On Mon, 8 Jun 2015, Tom de Vries wrote:
>>>
>>>> On 17/04/15 12:08, Tom de Vries wrote:
>>>>> On 20-03-15 12:38, Tom de Vries wrote:
>>>>>> On 19-03-15 12:05, Tom de Vries wrote:
>>>>>>> On 18-03-15 18:22, Tom de Vries wrote:
>>>>>>>> this patch fixes PR65460.
>>>>>>>>
>>>>>>>> The patch marks offloaded functions as parallelized, which means the
>>>>>>>> parloops
>>>>>>>> pass no longer attempts to modify that function.
>
>>> Ok, but shouldn't it be set before calling add_new_function as
>>> add_new_function might run passes that wouldn't identify the
>>> function as parallelized?
>>
>> Hm, indeed sometimes add_new_function executes some passes itself,
>> besides queueing the function for further processing. I suppose the
>> existing settings of parallelized_function should be modified in a
>> similar way.
>
> (I took note of this small change of the trunk patch vs. the version
> present on gomp-4_0-branch, and will adapt the latter as required as part
> of the next merge.)
>
>> I'll bootstrap and reg-test attached two patches on x86_64, and commit
>> unless objections.
>
>> Mark offloaded functions as parallelized
>>
>> 2015-06-09  Tom de Vries  <tom@codesourcery.com>
>>
>> 	PR tree-optimization/65460
>> 	* omp-low.c (expand_omp_target): Set parallelized_function on
>> 	cgraph_node for child_fn.
>
> (Committed to trunk in r224303.)
>
>> --- a/gcc/omp-low.c
>> +++ b/gcc/omp-low.c
>> @@ -8959,6 +8959,8 @@ expand_omp_target (struct omp_region *region)
>>
>>         /* Inform the callgraph about the new function.  */
>>         DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties;
>> +      cgraph_node *node = cgraph_node::get_create (child_fn);
>
> Would you please committ a fix: in offloading configurations
> (ENABLE_OFFLOADING), this new node variable will clash with an existing
> definition a little later:
>
>> +      node->parallelized_function = 1;
>>         cgraph_node::add_new_function (child_fn, true);
>>
>>   #ifdef ENABLE_OFFLOADING
> |        /* Add the new function to the offload table.  */
> |        vec_safe_push (offload_funcs, child_fn);
> |  #endif
> |
> |        /* Fix the callgraph edges for child_cfun.  Those for cfun will be
> |  	 fixed in a following pass.  */
> |        push_cfun (child_cfun);
> |        cgraph_edge::rebuild_edges ();
> |
> |  #ifdef ENABLE_OFFLOADING
> |        /* Prevent IPA from removing child_fn as unreachable, since there are no
> |  	 refs from the parent function to child_fn in offload LTO mode.  */
> |        struct cgraph_node *node = cgraph_node::get (child_fn);
> |        node->mark_force_output ();
> |  #endif
>

Committed attached patch to trunk as obvious. Build and tested in 
ENABLE_OFFLOADING configuration on top of gomp-4_0-branch.

Thanks,
- Tom