[jdk20] RFR: 8299715: IR test: VectorGatherScatterTest.java fails with SVE randomly
Xiaohong Gong
xgong at openjdk.org
Tue Jan 10 02:20:05 UTC 2023
On Tue, 10 Jan 2023 00:11:09 GMT, Paul Sandoz <psandoz at openjdk.org> wrote:
> Presuming that randomized mask data is not important this looks good. Otherwise we can check for all bits set and and if so reset a randomly selected bit?
Thanks for looking at this PR, Paul!
Yes, I don't think it's so important making the mask data randomized. The main purpose of this test is testing the register misuse issue in AArch64 backend match rules we fixed before (https://bugs.openjdk.org/browse/JDK-8288397). The masked gather_load/scatter_store result is not checked in the tests.
To keep the randomness of the mask data, we can either use a loop inside the mask which uses all the generated mask elements instead of the first vector size part, or just make the index of the mask array input randomly such as:
int index = RD.nextInt(100);
VectorMask<Long> mask = VectorMask.fromArray(L_SPECIES, m, index);
But as a comparison, the current fixing version is fine to me as a neat fixing.
Thanks,
Xiaohong
-------------
PR: https://git.openjdk.org/jdk20/pull/90
More information about the hotspot-compiler-dev
mailing list