Do not use atomic increment in allocation as fence.

A sequentially consistent fetch_and_add implemented with ARM v8 acquire
release operations is not a fence.  Don't use it as one.

The result may also be somewhat faster, since a sequentially consistent
increment requires more fencing than needed for the increment.

Bug: 16377103
Change-Id: I5b1add098d3488aa755f140612e54521b80aa749

2 files changed, 6 insertions, 2 deletions

diff --git a/runtime/atomic.h b/runtime/atomic.h
index 0faa3c69c6..d4a7f37bc6 100644
--- a/runtime/atomic.h
+++ b/runtime/atomic.h
@@ -278,6 +278,10 @@ class PACKED(sizeof(T)) Atomic : public std::atomic<T> {
     return this->fetch_add(value, std::memory_order_seq_cst);  // Return old_value.
   }
 
+  T FetchAndAddRelaxed(const T value) {
+    return this->fetch_add(value, std::memory_order_relaxed);  // Return old_value.
+  }
+
   T FetchAndSubSequentiallyConsistent(const T value) {
     return this->fetch_sub(value, std::memory_order_seq_cst);  // Return old value.
   }
diff --git a/runtime/gc/heap-inl.h b/runtime/gc/heap-inl.h
index d1ab587aea..f437830e11 100644
--- a/runtime/gc/heap-inl.h
+++ b/runtime/gc/heap-inl.h
@@ -145,9 +145,9 @@ inline mirror::Object* Heap::AllocObjectWithAllocator(Thread* self,
       WriteBarrierField(obj, mirror::Object::ClassOffset(), klass);
     }
     pre_fence_visitor(obj, usable_size);
+    QuasiAtomic::ThreadFenceForConstructor();
     new_num_bytes_allocated = static_cast<size_t>(
-        num_bytes_allocated_.FetchAndAddSequentiallyConsistent(bytes_tl_bulk_allocated))
-        + bytes_tl_bulk_allocated;
+        num_bytes_allocated_.FetchAndAddRelaxed(bytes_tl_bulk_allocated)) + bytes_tl_bulk_allocated;
   }
   if (kIsDebugBuild && Runtime::Current()->IsStarted()) {
     CHECK_LE(obj->SizeOf(), usable_size);