On Mon, 7 Oct 2024 12:54:00 GMT, Vladimir Kozelkov <d...@openjdk.org> wrote:
> > I think the new rules should explicitly ban zero-element sequence layouts.
>
> Even in the rules with the latest edits, I don't see what exactly should
> prohibit such sequences.
>
> Also, if they are banned, won't that break jextract and some user code?
> Additionally, I think many people might use zero-length sequences for layouts
> that are architecture-dependent.
Let me clarify. The javadoc rules we're talking about try to define a set of
suported layout across _all linkers_. As such, the rules have to try and find
some lowest common denominator between the various platforms.
While GCC does seem to support empty struct fields and empty sequence fields in
structs, other compilers (e.g. MSVC) do not. As such, the set of minimally
supported layouts should not allow such fields.
Here's a refined set of rules to see if a group layout is in this minimally
supported set:
{@code L} is a group layout {@code G} and all the following conditions hold:
1. the alignment constraint of {@code G} is set to its <a
href="MemoryLayout.html#layout-align">natural alignment</a>;</li>
2. the size of {@code G} is a multiple of its alignment constraint;
3. each member layout in {@code G.memberLayouts()} is either a padding layout
or a layout supported by {@code NL}
4. {@code G} contains at least a non-padding member layout
5. each padding member layout in {@code G} is naturaly aligned
6. each non-padding member layout in {@code G} follows an optional padding
member layout, whose size is the minimum size required to align E;
7. {@code G} contains an optional trailing padding member layout, whose size is
the minimum size that satisfies (2)
(and, we need an additional restriction, to say that zero-element sequence
layouts are not part of the minimally supported set).
-------------
PR Comment: https://git.openjdk.org/jdk/pull/21041#issuecomment-2397064613
