Sweep interpreter caches using checkpoint during GC

Interpreter caches are mutator thread-local caches. Currently, the way
GC sweeps these caches breaks that invariant and requires
synchronization for correct behavior.

In this CL we take a different approach of sweeping the caches
during GC using checkpoint (or in a stop-the-world pause),
eliminating the need for synchronization.

Bug: 235988901
Test: art/test/testrunner/testrunner.py --host --debug --interpreter
Change-Id: I8e8bbfc2d424bdce9c1c8c662782febe18678193

1 files changed, 3 insertions, 7 deletions

diff --git a/runtime/interpreter/interpreter_cache-inl.h b/runtime/interpreter/interpreter_cache-inl.h
index 269f5fa9ab..804d382877 100644
--- a/runtime/interpreter/interpreter_cache-inl.h
+++ b/runtime/interpreter/interpreter_cache-inl.h
@@ -35,13 +35,9 @@ inline bool InterpreterCache::Get(Thread* self, const void* key, /* out */ size_
 
 inline void InterpreterCache::Set(Thread* self, const void* key, size_t value) {
   DCHECK(self->GetInterpreterCache() == this) << "Must be called from owning thread";
-
-  // For simplicity, only update the cache if weak ref accesses are enabled. If
-  // they are disabled, this means the CC GC could be processing the cache, and
-  // reading it concurrently.
-  if (!gUseReadBarrier || self->GetWeakRefAccessEnabled()) {
-    data_[IndexOf(key)] = Entry{key, value};
-  }
+  // Simple store works here as the cache is always read/written by the owning
+  // thread only (or in a stop-the-world pause).
+  data_[IndexOf(key)] = Entry{key, value};
 }
 
 }  // namespace art