Hi!

On 2022-03-01T17:46:20+0100, I wrote:
> On 2022-01-13T10:54:16+0100, I wrote:
>> On 2019-05-08T14:51:57+0100, Julian Brown <julian@codesourcery.com> wrote:
>>>  - The "addressable" bit is set during the kernels conversion pass for
>>>    variables that have "create" (alloc) clauses created for them in the
>>>    synthesised outer data region (instead of in the front-end, etc.,
>>>    where it can't be done accurately). Such variables actually have
>>>    their address taken during transformations made in a later pass
>>>    (omp-low, I think), but there's a phase-ordering problem that means
>>>    the flag should be set earlier.
>>
>> The actual issue is a bit different, but yes, there is a problem.
>> The related ICE has also been reported as <https://gcc.gnu.org/PR100280>
>> "ICE in lower_omp_target, at omp-low.c:12287".  (And I'm confused why we
>> didn't run into that with the OpenACC 'kernels' decomposition
>> originally.)  I've pushed to master branch
>> commit 9b32c1669aad5459dd053424f9967011348add83
>> "OpenACC 'kernels' decomposition: Mark variables used in synthesized data
>> clauses as addressable [PR100280]"

>> --- a/gcc/omp-oacc-kernels-decompose.cc
>> +++ b/gcc/omp-oacc-kernels-decompose.cc
>> @@ -793,7 +793,8 @@ make_data_region_try_statement (location_t loc, gimple *body)
>>
>>  /* If INNER_BIND_VARS holds variables, build an OpenACC data region with
>>     location LOC containing BODY and having 'create (var)' clauses for each
>> -   variable.  If INNER_CLEANUP is present, add a try-finally statement with
>> +   variable (as a side effect, such variables also get TREE_ADDRESSABLE set).
>> +   If INNER_CLEANUP is present, add a try-finally statement with
>>     this cleanup code in the finally block.  Return the new data region, or
>>     the original BODY if no data region was needed.  */
>>
>> @@ -842,6 +843,9 @@ maybe_build_inner_data_region (location_t loc, gimple *body,
>>         inner_data_clauses = new_clause;
>>
>>         prev_mapped_var = v;
>> +
>> +       /* See <https://gcc.gnu.org/PR100280>.  */
>> +       TREE_ADDRESSABLE (v) = 1;
>>       }
>>      }
>
> So, that's too simple.  ;-) [...]

> We're after gimplification, and must not just set 'TREE_ADDRESSABLE',
> because that may easily violate GIMPLE invariants, leading to ICEs later.
> There are a few open PRs, which my following changes are addressing.  To
> make "late" 'TREE_ADDRESSABLE' work, we have a precedent in OpenMP's
> 'gcc/omp-low.cc:task_shared_vars' handling, as Jakub had pointed to in
> discussion of <https://gcc.gnu.org/PR102330>.

> I'm thus proposing to generalize 'gcc/omp-low.cc:task_shared_vars' into
> 'make_addressable_vars', plus new 'OMP_CLAUSE_MAP_DECL_MAKE_ADDRESSABLE'
> that we then may use instead of the 'TREE_ADDRESSABLE (v) = 1;' quoted
> above (plus one or two additional ones to be introduced in later
> patches), and wire that up in 'gcc/omp-low.cc:scan_sharing_clauses', for
> 'OMP_CLAUSE_MAP': set 'TREE_ADDRESSABLE' and put into
> 'make_addressable_vars' for later fix-up.

Pushed to master branch commit 8935589b496f755e08cadf26d8ceddf0dd6e0968
"OMP lowering: Regimplify 'OMP_CLAUSE_MAP_DECL_MAKE_ADDRESSABLE' DECLs
[PR100280, PR104132, PR104133]", see attached.


Grüße
 Thomas


-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955