Thread-local mark stacks for the CC collector.
Thread-local mark stacks are assigned to mutators where they push
references in read barriers to reduce the (CAS) synchronization cost
in a global mark stack/queue.
We step through three mark stack modes (thread-local, shared,
GC-exclusive) and use per-thread flags to disable/enable system weak
accesses (only for the CC collector) instead of the existing global
one to safely perform the marking phase. The reasons are 1)
thread-local mark stacks for mutators need to be revoked using a
checkpoint to avoid races (incorrectly leaving a reference on mark
stacks) when terminating marking, and 2) we can’t use a checkpoint
while system weak accesses are disabled (or a deadlock would
happen). More details are described in the code comments.
Performance improvements in Ritzperf EAAC: a ~2.8% improvement
(13290->12918) in run time and a ~23% improvement (51.6s->39.8s) in
the total GC time on N5.
Bug: 12687968
Change-Id: I5d234d7e48bf115cd773d38bdb62ad24ce9116c7
diff --git a/runtime/gc/reference_processor.h b/runtime/gc/reference_processor.h
index a44319b..284d13c 100644
--- a/runtime/gc/reference_processor.h
+++ b/runtime/gc/reference_processor.h
@@ -54,6 +54,7 @@
// Only allow setting this with mutators suspended so that we can avoid using a lock in the
// GetReferent fast path as an optimization.
void EnableSlowPath() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
+ void BroadcastForSlowPath(Thread* self);
// Decode the referent, may block if references are being processed.
mirror::Object* GetReferent(Thread* self, mirror::Reference* reference)
SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) LOCKS_EXCLUDED(Locks::reference_processor_lock_);