RFR: 8368722: Vector API intrinsics enabled wrongly on platforms without support for misaligned vector access [v3]
Dingli Zhang
dzhang at openjdk.org
Tue Oct 21 13:35:46 UTC 2025
On Fri, 10 Oct 2025 09:23:31 GMT, Hamlin Li <mli at openjdk.org> wrote:
> > Mind you that the following tests explicitly are testing `-XX:-AlignVector` and won't work on hardware platforms without misaligned vector. But it doesn't seem to me reasonable to add a requirement like `@requires vm.opt.EnableVectorSupport == true` for them as they are superword tests. These tests were once enabled for RISC-V by: https://bugs.openjdk.org/browse/JDK-8352529. Maybe we should simply keep them disabled for this platform?
>
> Simply distable them on riscv seems not a good idea. Maybe add `@requires (os.arch != "riscv64" | !vm.opt.AlignVector))`?
These tests all report IR-related errors because the first commit did not handle `AlignVector`, only modifying `EnableVectorSupport`.
This prevented the corresponding nodes from being printed. After implementing AlignVector handling, cases that fail to satisfy constraints are now skipped and no longer generate errors.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/27506#issuecomment-3426678036
More information about the hotspot-dev
mailing list