[patch][Fortran] Actually permit OpenMP's 'target simd'

[Thomas König]

Hi Bernhard,

> Am 14.10.2019 um 11:58 schrieb Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>:
> 
> On 14 October 2019 10:50:55 CEST, Jakub Jelinek <jakub@redhat.com> wrote:
>>> On Tue, Oct 08, 2019 at 02:04:17PM +0200, Tobias Burnus wrote:
>>>    testsuite/
>>>    * gfortran.dg/gomp/target-simd.f90: New.
>> 
>> John Reporte in bugzilla the testcase fails on hppa with an error that
>> looks like heap corruption, and indeed, running the test under valgrind
>> fails on x86_64-linux too, the allocatable arrays are 1..100, but the
>> loops
>> were reading and writing 0..100.  Fixed thusly, tested on x86_64-linux,
>> committed to trunk.
> 
> Which suggests that we ought to diagnose this, obviously.

We already do for arrays with static bounds.

Allocatable - not straightforward. This would need value propagation from the ALLOCATE statement, which in turn would suggest splitting up the FE represenation into basic blocks, which of course can be higher level than what the middle end can do after scalarization.

Anyway, I have just submitted PR 92087 for this. Feel free to comment or, even better, participate 😉