Re: [patch, fortran] PR 84487

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

Re: [patch, fortran] PR 84487

Thomas Koenig-6
Hi,

Jakub, do you have any comments?  After Seve's OK I plan to commit in a
couple of days unless I read anything to the contrary.

Regards

        Thomas

>> this patch makes sure that the __def_init variables, which have been
>> generated for normal allocatable arrays for quite some time, do not fill
>> up huge amounts of space in the object files with zeros. This is done by
>> not marking them read-only, which means that they are put into the BSS.
>>
>> Setting DECL_ARTIFICIAL on the __def_init variable makes sure it
>> is handled as predetermined shared in gfc_omp_predetermined_sharing .
>>
>> This is not an optimum solution. As the xfail shows, we are now missing
>> out on an optimization (as seen by the xfail that is now needed), and
>> having large all-zero variables seems wrong. However, this patch solves
>> the most urgent problem in this respect.
>>
>> This is an 8/9/10 regression, so I would like to commit this to
>> all of these branches (waiting before gcc 9 reopens, of course).
>>
>> I wold then close the PR and open an enchancement PR for the xfail
>> and the design improvement.
>>
>> Test case... I'm not sure what to test for.
>>
>> Regression-tested. OK for all affected branches?
>>
>
> Your approach seems reasonable to me.  You may want to
> ping Jakub or Tobias as this affects openmp.
>