Relax some CASes for the CC collector.
That is, removing some unnecessary memory fences.
We can use the relaxed CAS for the mark bitmap and reference field/GC
root updates because only the atomicity of the updated word matters
there.
We can use the release CAS for the read barrier bits in the lock word
because it needs to make sure the reference field updates are visible
when the object changes black from gray (the field update stores won't
be reordered after the CAS.)
The CC collector's Ritz EAAC GC time decreases from 34.7s to
29.1s (-16%) on N5.
Bug: 12687968
Change-Id: If082d5911a25fac695df66263a8f55ce8149b199
diff --git a/runtime/read_barrier-inl.h b/runtime/read_barrier-inl.h
index 85ac4aa..4998a6a 100644
--- a/runtime/read_barrier-inl.h
+++ b/runtime/read_barrier-inl.h
@@ -63,7 +63,7 @@
ref = reinterpret_cast<MirrorType*>(Mark(old_ref));
// Update the field atomically. This may fail if mutator updates before us, but it's ok.
if (ref != old_ref) {
- obj->CasFieldStrongSequentiallyConsistentObjectWithoutWriteBarrier<false, false>(
+ obj->CasFieldStrongRelaxedObjectWithoutWriteBarrier<false, false>(
offset, old_ref, ref);
}
}
@@ -101,7 +101,7 @@
// Update the field atomically. This may fail if mutator updates before us, but it's ok.
if (ref != old_ref) {
Atomic<mirror::Object*>* atomic_root = reinterpret_cast<Atomic<mirror::Object*>*>(root);
- atomic_root->CompareExchangeStrongSequentiallyConsistent(old_ref, ref);
+ atomic_root->CompareExchangeStrongRelaxed(old_ref, ref);
}
}
AssertToSpaceInvariant(gc_root_source, ref);
@@ -140,7 +140,7 @@
if (new_ref.AsMirrorPtr() != old_ref.AsMirrorPtr()) {
auto* atomic_root =
reinterpret_cast<Atomic<mirror::CompressedReference<MirrorType>>*>(root);
- atomic_root->CompareExchangeStrongSequentiallyConsistent(old_ref, new_ref);
+ atomic_root->CompareExchangeStrongRelaxed(old_ref, new_ref);
}
}
AssertToSpaceInvariant(gc_root_source, ref);