[foreign-memaccess+abi] RFR: 8284300: Explicitly reject layouts that don't have the correct alignment during linking

[Jorn Vernee]

On Mon, 19 Sep 2022 17:39:23 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

> I realized that rejecting layouts that don't have natural alignment also prevents us from supporting packed structs in the future. I tested out the packed struct support and it works on Windows pretty much out of the box. It doesn't work on SysV, some changes to CallArranger are needed. So, there are some options:
> 1. reject layouts that don't have natural alignment on all platforms (Windows won't be able to use packed structs)
> 2. reject layouts that don't have natural alignment on platforms besides Windows (yay Windows users)
> 3. align the status quos: Windows and AArch64 re-compute the alignment. It's not used on Windows, so it can be dropped there. On AArch64 we could "trust" the alignment that is attached to the layout given by the user, same as SysV, but it would fail for packed structs. With the intention of fixing the failure in the future.
> 
> I'm leaning most towards 1 at the moment, and that's what this patch implements. We can change that approach when we add reliable support for packed structs across platforms.

I've addressed the review comments. I've reverted the changes to the aarch64 CallArranger, since they really don't do anything at the moment.

I also switched to using `AbstractLayout::hasNaturalAlignment` called from AbstractLinker. For aggregates we also have to recursively check the member layouts.

The old alignment computing code was no longer used, so I removed.

I also had a look at the layouts clang gives us for the different platforms when using arrays, and I don't see any 16 byte alignments for arrays larger than 16 bytes or incomplete arrays, which the old alignment computation code used, so I think we don't need to take that into account.

-------------

PR: https://git.openjdk.org/panama-foreign/pull/728