openj9
openj9 copied to clipboard
Support MODE_BROADCAST for masks
Depends on: https://github.com/eclipse-omr/omr/pull/7785
Some corrections.
Rebased
I think I forgot to convert the child of msplats into TR::Int8.
Ready for review.
Vectorization LGTM.
For
doScalarizationthough, it's not obvious to me. I see that it's just doing splats (either vector or mask), since long bits to mask has been ruled out incheckScalarizationmode. It looks like it will represent each lane as a 32-bit integer with value 0 (for false) or -1 (for true). Is that the existing representation for masks when scalarizing? FromILOpcodeFromVectorAPIOpcode(), it looks like we getBadILOpwhen trying to scalarize operations involving a mask, and if that's always true, then maybe there is no existing representation and we can just choose one (so the one implied by the code here would be OK)
Yes, you are right. In fact, I see a problem even with vsplats. I disabled for now and added a TODO comment.
Jenkins test sanity all jdk21
Build failure seems to be due to #22144.
I'll try again once #22160 is merged
Jenkins test sanity all jdk21
Failures since https://github.com/eclipse-omr/omr/pull/7785 was merged but not promoted yet.
OK, we can still test in the meantime. Just need to remember to make sure not to merge before it promotes
Jenkins test sanity all jdk21 depends eclipse-omr/omr#7785
The failures are cmdLineTester_pltest_[aix_]0 with an unexpected page size from j9vmem_find_valid_page_size(), which is #22143.
Now we just need to wait for eclipse-omr/omr#7785 to promote before merging. Unfortunately the same issue #22143 is blocking OMR promotion at the moment.
Merging now that the dependency has promoted