refCount alignment issue about cachelib

[Svelar]

No really, that is caused by mistake of 'compiler optimization'. I still need more time to verify that.

But you could close this issue anytime, thanks!

[jaesoo-fb]

Hi @Svelar

Thanks for the report. What is the exact processor architecture? Just in case, is it 32 bit architecture?

As explained here, the refcount is third members and should start aligned on 4B since the access hook and mm hook are both aligned on 8B as long as the alloc sizes are aligned on 4B. In fact, generateAllocSizes(...) should give all alloc sizes aligned on 8B.

So, this might be just a test issue where we are passing some arbitrary alloc sizes (e.g., 50B) for testing. I think we can enforce the 4B alignment of alloc sizes and amend the tests accordingly for now.

For disclaimer, we are not testing on ARM for now, so the build and the operation are not guaranteed. We might need to consider the ARM support in the future for various reasons, but we always welcome any contribution in the meantime.

[Svelar]

Hi @jaesoo-fb , thanks for your reply.

I tested on kunpeng 920 (aarch64) server.

Yes, you are right. Nothing is affected but tests. Actually, I modified a little bit codes to enforce allocSize aligned on 8B and worked fine. I will raise a PR once it passed all tests.

I tested on ARM with centOS and openEuler, and there are not many codes should be added. Pin me if I could offer any help!

[jaesoo-fb]

Thanks for the update. We really appreciate if you can submit it as PR.

Just in case, is this related to another issue #276 as well?