8195148: Collapse G1SATBCardTableModRefBS and G1SATBCardTableLoggingModRefBS into a single G1BarrierSet (original) (raw)

Kim Barrett kim.barrett at oracle.com
Fri Mar 2 22:51:02 UTC 2018

On Mar 2, 2018, at 2:31 AM, Erik Osterlund <erik.osterlund at oracle.com> wrote:

src/hotspot/share/gc/g1/g1BarrierSet.inline.hpp 32 inline void G1BarrierSet::writereffieldpre(T* field) {

The change here doesn't seem to have anything to do with the renaming. Rather, it looks like a separate bug fix? The old code deferred the decode until after the null check, with the decoding benefitting from having already done the null check. At first glance, the new code seems like it might not perform as well. I do see why adding volatile is needed here though. I understand this might look unrelated. Here is my explanation: There has been an unfortunate implicit dependency to oop.inline.hpp. Now with some headers included in different order, it no longer compiles without adding that include. But including oop.inline.hpp causes an unfortunate include cycle that causes other problems. By loading the oop with RawAccess instead, those issues are solved. As for MOVOLATILE, I thought I might as well correct that while I am at it. Some compilers can and actually do reload the oop after the null check, at which point they may be NULL and break the algorithm. I couldn’t not put that decorator in there to solve that. So you see how that started as a necessary include dependency fix (I had to do something or it would not compile) but ended up fixing more things. Hope that is okay.

I tried this out, to try to get a better understanding of the issues, and I don't know what problems you are referring to.

I used the variant I suggested, e.g. RawAccess and conversion to T rather than collapsing to oop, with the original oopDesc-based null check and decoding. That did indeed fail to compile (not too surpisingly). But adding an #include of oop.inline.hpp seemed to just work.

While thinking about this before trying any experiments, it occurred to me that we might have a usage mistake around .inline.hpp files, but that didn't seem to arise here.

So no, I'm not (yet) okay with this part of the change.

More information about the hotspot-dev mailing list