Interpreter tweaks
Increase the amount of inlining and hot function hints in the interpreter
to encourage the hot Execute function to be faster.
Performance is 3x slower than Dalvik+JIT on FibonacciFast and similar
microbenchmarks.
Change-Id: I2b1a0c7545f86036b9b1b5ccac881d06292356d8
diff --git a/src/interpreter/interpreter.cc b/src/interpreter/interpreter.cc
index 33bdf9f..e315710 100644
--- a/src/interpreter/interpreter.cc
+++ b/src/interpreter/interpreter.cc
@@ -385,8 +385,7 @@
static void DoInvoke(Thread* self, MethodHelper& mh, ShadowFrame& shadow_frame,
const Instruction* inst, InvokeType type, bool is_range,
- JValue* result)
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ JValue* result) {
uint32_t vregC = (is_range) ? inst->VRegC_3rc() : inst->VRegC_35c();
Object* receiver;
if (type == kStatic) {
@@ -474,7 +473,11 @@
static void DoFieldGet(Thread* self, ShadowFrame& shadow_frame,
const Instruction* inst, FindFieldType find_type,
Primitive::Type field_type)
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) ALWAYS_INLINE;
+
+static inline void DoFieldGet(Thread* self, ShadowFrame& shadow_frame,
+ const Instruction* inst, FindFieldType find_type,
+ Primitive::Type field_type) {
bool is_static = (find_type == StaticObjectRead) || (find_type == StaticPrimitiveRead);
uint32_t field_idx = is_static ? inst->VRegB_21c() : inst->VRegC_22c();
Field* f = FindFieldFromCode(field_idx, shadow_frame.GetMethod(), self,
@@ -524,7 +527,11 @@
static void DoFieldPut(Thread* self, ShadowFrame& shadow_frame,
const Instruction* inst, FindFieldType find_type,
Primitive::Type field_type)
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) ALWAYS_INLINE;
+
+static inline void DoFieldPut(Thread* self, ShadowFrame& shadow_frame,
+ const Instruction* inst, FindFieldType find_type,
+ Primitive::Type field_type) {
bool is_static = (find_type == StaticObjectWrite) || (find_type == StaticPrimitiveWrite);
uint32_t field_idx = is_static ? inst->VRegB_21c() : inst->VRegC_22c();
Field* f = FindFieldFromCode(field_idx, shadow_frame.GetMethod(), self,
@@ -572,7 +579,7 @@
}
}
-static String* ResolveString(Thread* self, MethodHelper& mh, uint32_t string_idx) {
+static inline String* ResolveString(Thread* self, MethodHelper& mh, uint32_t string_idx) {
Class* java_lang_string_class = String::GetJavaLangString();
if (UNLIKELY(!java_lang_string_class->IsInitialized())) {
ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
@@ -585,7 +592,7 @@
return mh.ResolveString(string_idx);
}
-static void DoIntDivide(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
+static inline void DoIntDivide(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
int32_t dividend, int32_t divisor) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
if (UNLIKELY(divisor == 0)) {
ThrowArithmeticExceptionDivideByZero(self);
@@ -596,7 +603,7 @@
}
}
-static void DoIntRemainder(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
+static inline void DoIntRemainder(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
int32_t dividend, int32_t divisor) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
if (UNLIKELY(divisor == 0)) {
ThrowArithmeticExceptionDivideByZero(self);
@@ -607,7 +614,7 @@
}
}
-static void DoLongDivide(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
+static inline void DoLongDivide(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
int64_t dividend, int64_t divisor) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
if (UNLIKELY(divisor == 0)) {
ThrowArithmeticExceptionDivideByZero(self);
@@ -618,7 +625,7 @@
}
}
-static void DoLongRemainder(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
+static inline void DoLongRemainder(Thread* self, ShadowFrame& shadow_frame, size_t result_reg,
int64_t dividend, int64_t divisor) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
if (UNLIKELY(divisor == 0)) {
ThrowArithmeticExceptionDivideByZero(self);
@@ -629,12 +636,20 @@
}
}
-static const Instruction* FindNextInstructionFollowingException(Thread* self,
- ShadowFrame& shadow_frame,
- uint32_t dex_pc,
- const uint16_t* const insns,
- SirtRef<Object>& this_object_ref,
- instrumentation::Instrumentation* instrumentation) {
+static inline const Instruction* FindNextInstructionFollowingException(Thread* self,
+ ShadowFrame& shadow_frame,
+ uint32_t dex_pc,
+ const uint16_t* insns,
+ SirtRef<Object>& this_object_ref,
+ instrumentation::Instrumentation* instrumentation)
+ ALWAYS_INLINE;
+
+static inline const Instruction* FindNextInstructionFollowingException(Thread* self,
+ ShadowFrame& shadow_frame,
+ uint32_t dex_pc,
+ const uint16_t* insns,
+ SirtRef<Object>& this_object_ref,
+ instrumentation::Instrumentation* instrumentation) {
self->VerifyStack();
ThrowLocation throw_location;
mirror::Throwable* exception = self->GetException(&throw_location);
@@ -670,9 +685,20 @@
inst = inst-> next_function (); \
}
+static void UnexpectedOpcode(const Instruction* inst, MethodHelper& mh)
+ __attribute__ ((cold, noreturn, noinline));
+
+static void UnexpectedOpcode(const Instruction* inst, MethodHelper& mh) {
+ LOG(FATAL) << "Unexpected instruction: " << inst->DumpString(&mh.GetDexFile());
+ exit(0); // Unreachable, keep GCC happy.
+}
+
static JValue Execute(Thread* self, MethodHelper& mh, const DexFile::CodeItem* code_item,
ShadowFrame& shadow_frame, JValue result_register)
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) __attribute__ ((hot));
+
+static JValue Execute(Thread* self, MethodHelper& mh, const DexFile::CodeItem* code_item,
+ ShadowFrame& shadow_frame, JValue result_register) {
if (UNLIKELY(!shadow_frame.HasReferenceArray())) {
LOG(FATAL) << "Invalid shadow frame for interpreter use";
return JValue();
@@ -692,7 +718,7 @@
shadow_frame.GetMethod(), 0);
}
while (true) {
- if (self->TestAllFlags()) {
+ if (UNLIKELY(self->TestAllFlags())) {
CheckSuspend(self);
}
const uint32_t dex_pc = inst->GetDexPc(insns);
@@ -1180,12 +1206,10 @@
float val1 = shadow_frame.GetVRegFloat(inst->VRegB_23x());
float val2 = shadow_frame.GetVRegFloat(inst->VRegC_23x());
int32_t result;
- // TODO: we should not test float equality like this. Reorder comparisons
- // or use a different comparison mechanism.
- if (val1 == val2) {
- result = 0;
- } else if (val1 > val2) {
+ if (val1 > val2) {
result = 1;
+ } else if (val1 == val2) {
+ result = 0;
} else {
result = -1;
}
@@ -1197,12 +1221,10 @@
float val1 = shadow_frame.GetVRegFloat(inst->VRegB_23x());
float val2 = shadow_frame.GetVRegFloat(inst->VRegC_23x());
int32_t result;
- // TODO: we should not test float equality like this. Reorder comparisons
- // or use a different comparison mechanism.
- if (val1 == val2) {
- result = 0;
- } else if (val1 < val2) {
+ if (val1 < val2) {
result = -1;
+ } else if (val1 == val2) {
+ result = 0;
} else {
result = 1;
}
@@ -1214,12 +1236,10 @@
double val1 = shadow_frame.GetVRegDouble(inst->VRegB_23x());
double val2 = shadow_frame.GetVRegDouble(inst->VRegC_23x());
int32_t result;
- // TODO: we should not test double equality like this. Reorder comparisons
- // or use a different comparison mechanism.
- if (val1 == val2) {
- result = 0;
- } else if (val1 > val2) {
+ if (val1 > val2) {
result = 1;
+ } else if (val1 == val2) {
+ result = 0;
} else {
result = -1;
}
@@ -1232,12 +1252,10 @@
double val1 = shadow_frame.GetVRegDouble(inst->VRegB_23x());
double val2 = shadow_frame.GetVRegDouble(inst->VRegC_23x());
int32_t result;
- // TODO: we should not test double equality like this. Reorder comparisons
- // or use a different comparison mechanism.
- if (val1 == val2) {
- result = 0;
- } else if (val1 < val2) {
+ if (val1 < val2) {
result = -1;
+ } else if (val1 == val2) {
+ result = 0;
} else {
result = 1;
}
@@ -2433,9 +2451,12 @@
(inst->VRegC_22b() & 0x1f));
inst = inst->Next_2xx();
break;
- default:
- LOG(FATAL) << "Unexpected instruction: " << inst->DumpString(&mh.GetDexFile());
- break;
+ case Instruction::UNUSED_3E ... Instruction::UNUSED_43:
+ case Instruction::UNUSED_E3 ... Instruction::UNUSED_FF:
+ case Instruction::UNUSED_73:
+ case Instruction::UNUSED_79:
+ case Instruction::UNUSED_7A:
+ UnexpectedOpcode(inst, mh);
}
}
}
diff --git a/src/mirror/class-inl.h b/src/mirror/class-inl.h
index 62740be..6819fb2 100644
--- a/src/mirror/class-inl.h
+++ b/src/mirror/class-inl.h
@@ -189,7 +189,7 @@
// src's super should be java_lang_Object, since it is an array.
Class* java_lang_Object = src->GetSuperClass();
DCHECK(java_lang_Object != NULL) << PrettyClass(src);
- DCHECK(java_lang_Object->GetSuperClass() == NULL) << PrettyClass(src);
+ DCHECK(java_lang_Object->GetSuperClass() == NULL) << PrettyClass(src);
return this == java_lang_Object;
}
return IsArrayAssignableFromArray(src);
diff --git a/src/mirror/class.h b/src/mirror/class.h
index dfbe815..0661b42 100644
--- a/src/mirror/class.h
+++ b/src/mirror/class.h
@@ -429,8 +429,7 @@
// downcast would be necessary. Similarly for interfaces, a class that implements (or an interface
// that extends) another can be assigned to its parent, but not vice-versa. All Classes may assign
// to themselves. Classes for primitive types may not assign to each other.
- bool IsAssignableFrom(const Class* src) const
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ inline bool IsAssignableFrom(const Class* src) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
DCHECK(src != NULL);
if (this == src) {
// Can always assign to things of the same type.
diff --git a/src/oat/runtime/support_dexcache.cc b/src/oat/runtime/support_dexcache.cc
index 6811d20..3e8ebc6 100644
--- a/src/oat/runtime/support_dexcache.cc
+++ b/src/oat/runtime/support_dexcache.cc
@@ -15,7 +15,9 @@
*/
#include "callee_save_frame.h"
+#include "gc/card_table-inl.h"
#include "class_linker-inl.h"
+#include "dex_file-inl.h"
#include "mirror/abstract_method-inl.h"
#include "mirror/object_array-inl.h"
#include "mirror/object-inl.h"
diff --git a/src/runtime_support.cc b/src/runtime_support.cc
index b096431..9242c87 100644
--- a/src/runtime_support.cc
+++ b/src/runtime_support.cc
@@ -333,39 +333,6 @@
}
}
-mirror::Class* ResolveVerifyAndClinit(uint32_t type_idx, const mirror::AbstractMethod* referrer,
- Thread* self, bool can_run_clinit, bool verify_access) {
- ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
- mirror::Class* klass = class_linker->ResolveType(type_idx, referrer);
- if (UNLIKELY(klass == NULL)) {
- CHECK(self->IsExceptionPending());
- return NULL; // Failure - Indicate to caller to deliver exception
- }
- // Perform access check if necessary.
- mirror::Class* referring_class = referrer->GetDeclaringClass();
- if (verify_access && UNLIKELY(!referring_class->CanAccess(klass))) {
- ThrowIllegalAccessErrorClass(referring_class, klass);
- return NULL; // Failure - Indicate to caller to deliver exception
- }
- // If we're just implementing const-class, we shouldn't call <clinit>.
- if (!can_run_clinit) {
- return klass;
- }
- // If we are the <clinit> of this class, just return our storage.
- //
- // Do not set the DexCache InitializedStaticStorage, since that implies <clinit> has finished
- // running.
- if (klass == referring_class && MethodHelper(referrer).IsClassInitializer()) {
- return klass;
- }
- if (!class_linker->EnsureInitialized(klass, true, true)) {
- CHECK(self->IsExceptionPending());
- return NULL; // Failure - Indicate to caller to deliver exception
- }
- referrer->GetDexCacheInitializedStaticStorage()->Set(type_idx, klass);
- return klass;
-}
-
void ThrowStackOverflowError(Thread* self) {
CHECK(!self->IsHandlingStackOverflow()) << "Recursive stack overflow.";
diff --git a/src/runtime_support.h b/src/runtime_support.h
index 1c39214..5fc8da5 100644
--- a/src/runtime_support.h
+++ b/src/runtime_support.h
@@ -248,10 +248,41 @@
Thread* self, bool access_check, InvokeType type)
SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
-extern mirror::Class* ResolveVerifyAndClinit(uint32_t type_idx,
- const mirror::AbstractMethod* referrer, Thread* self,
- bool can_run_clinit, bool verify_access)
- SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
+static inline mirror::Class* ResolveVerifyAndClinit(uint32_t type_idx,
+ const mirror::AbstractMethod* referrer,
+ Thread* self, bool can_run_clinit,
+ bool verify_access)
+ SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
+ ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
+ mirror::Class* klass = class_linker->ResolveType(type_idx, referrer);
+ if (UNLIKELY(klass == NULL)) {
+ CHECK(self->IsExceptionPending());
+ return NULL; // Failure - Indicate to caller to deliver exception
+ }
+ // Perform access check if necessary.
+ mirror::Class* referring_class = referrer->GetDeclaringClass();
+ if (verify_access && UNLIKELY(!referring_class->CanAccess(klass))) {
+ ThrowIllegalAccessErrorClass(referring_class, klass);
+ return NULL; // Failure - Indicate to caller to deliver exception
+ }
+ // If we're just implementing const-class, we shouldn't call <clinit>.
+ if (!can_run_clinit) {
+ return klass;
+ }
+ // If we are the <clinit> of this class, just return our storage.
+ //
+ // Do not set the DexCache InitializedStaticStorage, since that implies <clinit> has finished
+ // running.
+ if (klass == referring_class && MethodHelper(referrer).IsClassInitializer()) {
+ return klass;
+ }
+ if (!class_linker->EnsureInitialized(klass, true, true)) {
+ CHECK(self->IsExceptionPending());
+ return NULL; // Failure - Indicate to caller to deliver exception
+ }
+ referrer->GetDexCacheInitializedStaticStorage()->Set(type_idx, klass);
+ return klass;
+}
extern void ThrowStackOverflowError(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
