Class Hierarchy Analysis (CHA)
The class linker now tracks whether a method has a single implementation
and if so, the JIT compiler will try to devirtualize a virtual call for
the method into a direct call. If the single-implementation assumption
is violated due to additional class linking, compiled code that makes the
assumption is invalidated. Deoptimization is triggered for compiled code
live on stack. Instead of patching return pc's on stack, a CHA guard is
added which checks a hidden should_deoptimize flag for deoptimization.
This approach limits the number of deoptimization points.
This CL does not devirtualize abstract/interface method invocation.
Slides on CHA:
https://docs.google.com/a/google.com/presentation/d/1Ax6cabP1vM44aLOaJU3B26n5fTE9w5YU-1CRevIDsBc/edit?usp=sharing
Change-Id: I18bf716a601b6413b46312e925a6ad9e4008efa4
Test: ART_TEST_JIT=true m test-art-host/target-run-test test-art-host-gtest
diff --git a/runtime/art_method.h b/runtime/art_method.h
index 00fab65..2ae8147 100644
--- a/runtime/art_method.h
+++ b/runtime/art_method.h
@@ -85,9 +85,29 @@
template <ReadBarrierOption kReadBarrierOption = kWithReadBarrier>
ALWAYS_INLINE uint32_t GetAccessFlags();
+ // This version should only be called when it's certain there is no
+ // concurrency so there is no need to guarantee atomicity. For example,
+ // before the method is linked.
void SetAccessFlags(uint32_t new_access_flags) {
- // Not called within a transaction.
- access_flags_ = new_access_flags;
+ access_flags_.store(new_access_flags, std::memory_order_relaxed);
+ }
+
+ // This setter guarantees atomicity.
+ void AddAccessFlags(uint32_t flag) {
+ uint32_t old_access_flags = access_flags_.load(std::memory_order_relaxed);
+ uint32_t new_access_flags;
+ do {
+ new_access_flags = old_access_flags | flag;
+ } while (!access_flags_.compare_exchange_weak(old_access_flags, new_access_flags));
+ }
+
+ // This setter guarantees atomicity.
+ void ClearAccessFlags(uint32_t flag) {
+ uint32_t old_access_flags = access_flags_.load(std::memory_order_relaxed);
+ uint32_t new_access_flags;
+ do {
+ new_access_flags = old_access_flags & ~flag;
+ } while (!access_flags_.compare_exchange_weak(old_access_flags, new_access_flags));
}
// Approximate what kind of method call would be used for this method.
@@ -142,39 +162,7 @@
return (GetAccessFlags() & kAccIntrinsic) != 0;
}
- void SetIntrinsic(uint32_t intrinsic) {
- DCHECK(IsUint<8>(intrinsic));
- uint32_t new_value = (GetAccessFlags() & kAccFlagsNotUsedByIntrinsic) |
- kAccIntrinsic |
- (intrinsic << POPCOUNT(kAccFlagsNotUsedByIntrinsic));
- if (kIsDebugBuild) {
- uint32_t java_flags = (GetAccessFlags() & kAccJavaFlagsMask);
- bool is_constructor = IsConstructor();
- bool is_synchronized = IsSynchronized();
- bool skip_access_checks = SkipAccessChecks();
- bool is_fast_native = IsFastNative();
- bool is_copied = IsCopied();
- bool is_miranda = IsMiranda();
- bool is_default = IsDefault();
- bool is_default_conflict = IsDefaultConflicting();
- bool is_compilable = IsCompilable();
- bool must_count_locks = MustCountLocks();
- SetAccessFlags(new_value);
- DCHECK_EQ(java_flags, (GetAccessFlags() & kAccJavaFlagsMask));
- DCHECK_EQ(is_constructor, IsConstructor());
- DCHECK_EQ(is_synchronized, IsSynchronized());
- DCHECK_EQ(skip_access_checks, SkipAccessChecks());
- DCHECK_EQ(is_fast_native, IsFastNative());
- DCHECK_EQ(is_copied, IsCopied());
- DCHECK_EQ(is_miranda, IsMiranda());
- DCHECK_EQ(is_default, IsDefault());
- DCHECK_EQ(is_default_conflict, IsDefaultConflicting());
- DCHECK_EQ(is_compilable, IsCompilable());
- DCHECK_EQ(must_count_locks, MustCountLocks());
- } else {
- SetAccessFlags(new_value);
- }
- }
+ ALWAYS_INLINE void SetIntrinsic(uint32_t intrinsic) REQUIRES_SHARED(Locks::mutator_lock_);
uint32_t GetIntrinsic() {
DCHECK(IsIntrinsic());
@@ -259,7 +247,7 @@
void SetSkipAccessChecks() {
DCHECK(!SkipAccessChecks());
- SetAccessFlags(GetAccessFlags() | kAccSkipAccessChecks);
+ AddAccessFlags(kAccSkipAccessChecks);
}
// Should this method be run in the interpreter and count locks (e.g., failed structured-
@@ -461,6 +449,26 @@
return DataOffset(kRuntimePointerSize);
}
+ ALWAYS_INLINE bool HasSingleImplementation() REQUIRES_SHARED(Locks::mutator_lock_);
+
+ ALWAYS_INLINE void SetHasSingleImplementation(bool single_impl) {
+ DCHECK(!IsIntrinsic()) << "conflict with intrinsic bits";
+ if (single_impl) {
+ AddAccessFlags(kAccSingleImplementation);
+ } else {
+ ClearAccessFlags(kAccSingleImplementation);
+ }
+ }
+
+ ArtMethod* GetSingleImplementation()
+ REQUIRES_SHARED(Locks::mutator_lock_);
+
+ ALWAYS_INLINE void SetSingleImplementation(ArtMethod* method, PointerSize pointer_size) {
+ DCHECK(!IsNative());
+ DCHECK(IsAbstract()); // Non-abstract method's single implementation is just itself.
+ SetDataPtrSize(method, pointer_size);
+ }
+
void* GetEntryPointFromJni() {
DCHECK(IsNative());
return GetEntryPointFromJniPtrSize(kRuntimePointerSize);
@@ -655,7 +663,10 @@
GcRoot<mirror::Class> declaring_class_;
// Access flags; low 16 bits are defined by spec.
- uint32_t access_flags_;
+ // Getting and setting this flag needs to be atomic when concurrency is
+ // possible, e.g. after this method's class is linked. Such as when setting
+ // verifier flags and single-implementation flag.
+ std::atomic<std::uint32_t> access_flags_;
/* Dex file fields. The defining dex file is available via declaring_class_->dex_cache_ */