RFR: 8242038: G1: Lazily initialize RSHashTables

[Kim Barrett]

> On Apr 2, 2020, at 12:07 PM, Claes Redestad <claes.redestad at oracle.com> wrote:
> 
> Hi,
> 
> make RSHashTable initialization lazy by replacing the initial table for
> each region with a shared empty sentinel. Using an empty_table might be
> a bit tricky/"clever", but it ensures logic and lookups behave properly
> before we add a card to a regions RS.
> 
> Initialization to a unique table then piggy-backs on the test that
> checks whether we need to expand the RSHashTable, so GC overhead should
> be neutral (or even an improvement in cases where most PRTs stay empty).
> 
> Webrev: http://cr.openjdk.java.net/~redestad/8242038/open.00/
> Bug:    https://bugs.openjdk.java.net/browse/JDK-8242038
> 
> Testing: tier1-3, verified a small improvement to startup
> 
> Thanks!
> 
> /Claes

Not a complete review yet, but this caught my eye:

src/hotspot/share/gc/g1/sparsePRT.cpp
  92 static RSHashTable empty_table(1);

Static initialization of non-trivial objects has been a not-infrequent
source of problems.  In this case we're going to do C-heap allocation
during initialization and C-heap deallocation during process shutdown.
I'm not sure whether that's kosher or not.

Morover, the RSHashTable constructor calls RSHashTable::clear(). That
calls SparsePRTEntry::size(), whose value is computed from
G1RSetSparseRegionEntries, which is a product flag which doesn't have
its final value at static initialization time. Oops!