RFR: 8317545: AIX PPC64: Implementation of Foreign Function & Memory API [v3]
Martin Doerr
mdoerr at openjdk.org
Mon Oct 16 21:41:21 UTC 2023
On Mon, 16 Oct 2023 19:24:51 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:
> Ok. I'm assuming there are currently some failing tests due to the mismatch in ABI? (assuming the test libraries are compiled for the packed ABI)
We should be careful with the names:
- "#pragma align (power)" is the default which is currently used to compile the test libraries and which uses 4-byte alignment for double in structs after first member. It's compatible with libraries which were built with default settings.
- "#pragma align (packed)" is a special mode which packs even more, but I don't think we need to support that.
- "#pragma align (natural)" is the one which makes the layout compatible with other platforms and which is recommended for performance.
Almost all of the jdk/java/foreign are currently failing because of https://bugs.openjdk.org/browse/JDK-8317799. Suchi has a preliminary fix and with that most tests have passed with the 4-byte aligned double members. Almost all tests should work after both subtasks are resolved. (There are still a couple of things to investigate which are not related to Up-/Downcalls.)
My standalone tests are working fine with this PR. I've tried both, changing the layout to 4-byte double alignment and "#pragma align (natural)".
-------------
PR Comment: https://git.openjdk.org/jdk/pull/16179#issuecomment-1765319281
More information about the hotspot-compiler-dev
mailing list