review (S) for 6793828
Thomas.Rodriguez at Sun.COM
Wed Feb 4 18:25:46 PST 2009
On Feb 4, 2009, at 5:30 PM, John Rose wrote:
> I agree that it's surprising that safepoints don't modify raw
> memory. A pure safepoint cannot change managed pointers (as far as
> the compiler can see) so I would think that it would modify *only*
> raw memory. I don't remember how we got to this design...
I suspect that all our other uses of raw memory are pretty tightly
control dependent so they don't trip on this. I know that Steve had a
lot of trouble getting the G1 writer barrier to work correctly because
of issues like this though it's somewhat amazing we didn't actually
hit it before.
Steve also has some work with conditional card marks that important
for the scalability of oop related concurrent access. There were some
problems there having to do with improving our ability to elide
repeated card marks to reduce any overhead of conditional card marks
and getting the memory deps right was part of the issue. I suspect if
we want to do conditional card marks effectively we'll need to fix
SafePointNodes to produce a new raw memory.
> -- John
> On Feb 4, 2009, at 5:02 PM, Tom Rodriguez wrote:
More information about the hotspot-compiler-dev