On Fri, 3 Apr 2026 06:39:28 GMT, Jatin Bhateja <[email protected]> wrote:
> After last lworld+vector [merge > ](https://github.com/openjdk/valhalla/pull/1688) almost all vector API jtreg > tests were failing with default options. > > Summary of fixes:- > 1/ The C2_MISMATCHED decorator forces memory accesses to use > [TypeRawPtr::BOTTOM (alias index = > Raw)](https://github.com/openjdk/valhalla/blob/lworld/src/hotspot/share/opto/inlinetypenode.cpp#L605), > which is a different memory slice than the typed alias used by subsequent > LoadVector operations. > When VectorBoxAllocateNode is expanded, the StoreVector for the payload > writes to the Raw memory slice, but later LoadVector reads from the typed > slice preventing value forwarding. > > 2/ IR rule modification in following tests:- > test/hotspot/jtreg/compiler/vectorapi/VectorMaskCastIdentityTest.java > test/hotspot/jtreg/compiler/vectorapi/VectorMaskToLongTest.java > Modifications in above two tests is due to difference in mask boxing / > unboxing handling introduced by a prior checkin: > https://github.com/openjdk/valhalla/pull/845 > > With these two fixes all VectorAPI tests are now passing with following > configurations at various AVX levels. > - default options > - With -XX:-UseNonAtomicValueFlattening > > Best Regards, > Jatin Sure, let's get this in We need to fix .jcheck/conf in lworld+vector branch. ------------- Marked as reviewed by liach (Committer). PR Review: https://git.openjdk.org/valhalla/pull/2298#pullrequestreview-4055601638 PR Comment: https://git.openjdk.org/valhalla/pull/2298#issuecomment-4184300800
