Word-by-word card table scanning

[petermz]

This is a port of a Hotspot fix. The idea is, currently card table is scanned byte by byte. But the majority of cards are clean, so we could move by 8-byte steps as long as none of those 8 cards are marked, and fall back to single-byte steps otherwise.

Testing shows this speeds up things a bit, especially when heap is large. The charts below show length of the blackenDirtyCardRoots stage for HyperAlloc running with OnlyIncrementally collection policy and different heap sizes.

[oubidar-Abderrahim]

Hi, Thank you for contributing to GraalVM, Is there a specific member of the GraalVM team you prefer to review your PR?

[petermz]

Hi Abderrahim --
No, I don't think I have any preference

[oubidar-Abderrahim]

Hi @wirthi, could you please assign someone from the team to review this PR? Once the PR gets approved I can integrate it internally, Thank you.

[christianhaeubl]

Thanks for working on the performance of the card table scanning. The change looks good, I will integrate this change after I did some additional testing and benchmarking.

[christianhaeubl]

I did some benchmarking and your PR had a negative impact on a small number of benchmarks. I opened a PR that potentially replaces your PR: #5963. The implementation there, is closer to the one in HotSpot and doesn't seem to have any negative performance impacts.

If possible, please re-run your benchmarks with the PR from above and verify that the speedups are similar to the ones that you saw with your implementation.

[petermz]

I've measured speedups with your PR, they are about the same. Thanks!

[christianhaeubl]

Thanks for checking, then I am closing this PR in favor of #5963.