On Fri, 13 Oct 2023 14:03:34 GMT, Martin Doerr <mdo...@openjdk.org> wrote:
>>> Okay, so it sounds like Java double on AIX is still 8-byte aligned, but the
>>> native double is 4-byte aligned?
>>
>> Correct within structures. I believe they are always 8-byte aligned except
>> in structures which are packed more densely.
>>
>>> In that case, I'd say that passing `JAVA_DOUBLE` as an argument resulting
>>> in an exception, is expected behavior.
>>
>> Hmm... Should we disallow to pass 8-byte aligned double values? Fixing the
>> tests is a good thing. But, I have a different concern: I guess many
>> developers may use them without testing on AIX. This would cause problems
>> which we could avoid by supporting 8-byte aligned doubles in addition.
>>
>> What about adding
>>
>> layouts.put("jdouble", ValueLayout.JAVA_DOUBLE.withByteAlignment(4));
>> layouts.put("aligneddouble", ValueLayout.JAVA_DOUBLE); // allowed for usage
>> outside of structures
>>
>> ?
>
>> Just to clarify, is the ABI equal to what is described in this table:
>> https://www.ibm.com/docs/en/xl-c-and-cpp-aix/16.1?topic=data-using-alignment-modes
>> ?
>>
>> This shows `double` having 8-byte alignment, but being 4-byte aligned when
>> not the first member of an aggregate.
>
> Interesting! This matches my understanding. I can see that doubles are 4-byte
> aligned within structures. The 8-byte alignment of the first member is
> already established by the allocation alignment.
>
> I have only read the ABI documentation for linux. IBM will have to cover AIX
> specific details if there are more questions.
Discussed this a bit with Maurizio. There are a few things to consider:
- `cannonicalLayouts` allows for just a single mapping from type name to
layout. I think `double` should map to the 8-byte aligned layout. (the same
alignment you get from `_Alignof(double)` in C).
- We have to relax the checking done by the linker on AIX. Maybe change
`AbstractLinker::checkLayoutRecursive` call some `checkStructMemberLayout`
method, that by default just calls `checkLayoutRecursive`, but which the AIX
linker can override to implement it's special rules for doubles.
- We need a small spec update to allow for struct field layouts that are not
canonical layouts. I'll take a stab at that, and then get back to you.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/16179#discussion_r1358344403
