Fast Art interpreter
Add a Dalvik-style fast interpreter to Art.
Three primary deficiencies in the existing Art interpreter
will be addressed:
1. Structural inefficiencies (primarily the bloated
fetch/decode/execute overhead of the C++ interpreter
implementation).
2. Stack memory wastage. Each managed-language invoke
adds a full copy of the interpreter's compiler-generated
locals on the shared stack. We're at the mercy of
the compiler now in how much memory is wasted here. An
assembly based interpreter can manage memory usage more
effectively.
3. Shadow frame model, which not only spends twice the memory
to store the Dalvik virtual registers, but causes vreg stores
to happen twice.
This CL mostly deals with #1 (but does provide some stack memory
savings). Subsequent CLs will address the other issues.
Current status:
Passes all run-tests.
Phone boots interpret-only.
2.5x faster than Clang-compiled Art goto interpreter on fetch/decode/execute
microbenchmark, 5x faster than gcc-compiled goto interpreter.
1.6x faster than Clang goto on Caffeinemark overall
2.0x faster than Clang switch on Caffeinemark overall
68% of Dalvik interpreter performance on Caffeinemark (still much slower,
primarily because of poor invoke performance and lack of execute-inline)
Still nearly an order of magnitude slower than Dalvik on invokes
(but slightly better than Art Clang goto interpreter.
Importantly, saves ~200 bytes of stack memory per invoke (but still
wastes ~400 relative to Dalvik).
What's needed:
Remove the (large quantity of) bring-up hackery in place.
Integrate into the build mechanism. I'm still using the old Dalvik manual
build step to generate assembly code from the stub files.
Remove the suspend check hack. For bring-up purposes, I'm using an explicit
suspend check (like the other Art interpreters). However, we should be
doing a Dalvik style suspend check via the table base switch mechanism.
This should be done during the alternative interpreter activation.
General cleanup.
Add CFI info.
Update the new target bring-up README documentation.
Add other targets.
In later CLs:
Consolidate mterp handlers for expensive operations (such as new-instance) with
the code used by the switch interpreter. No need to duplicate the code for
heavyweight operations (but will need some refactoring to align).
Tuning - some fast paths needs to be moved down to the assembly handlers,
rather than being dealt with in the out-of-line code.
JIT profiling. Currently, the fast interpreter is used only in the fast
case - no instrumentation, no transactions and no access checks. We
will want to implement fast + JIT-profiling as the alternate fast
interpreter. All other cases can still fall back to the reference
interpreter.
Improve invoke performance. We're nearly an order of magnitude slower than
Dalvik here. Some of that is unavoidable, but I suspect we can do
better.
Add support for our other targets.
Change-Id: I43e25dc3d786fb87245705ac74a87274ad34fedc
diff --git a/runtime/stack.h b/runtime/stack.h
index a0c44cb..4fa1a4f 100644
--- a/runtime/stack.h
+++ b/runtime/stack.h
@@ -184,11 +184,12 @@
}
uint32_t GetDexPC() const {
- return dex_pc_;
+ return (dex_pc_ptr_ == nullptr) ? dex_pc_ : dex_pc_ptr_ - code_item_->insns_;
}
void SetDexPC(uint32_t dex_pc) {
dex_pc_ = dex_pc;
+ dex_pc_ptr_ = nullptr;
}
ShadowFrame* GetLink() const {
@@ -206,6 +207,20 @@
return *reinterpret_cast<const int32_t*>(vreg);
}
+ uint32_t* GetVRegAddr(size_t i) {
+ return &vregs_[i];
+ }
+
+ uint32_t* GetShadowRefAddr(size_t i) {
+ DCHECK(HasReferenceArray());
+ DCHECK_LT(i, NumberOfVRegs());
+ return &vregs_[i + NumberOfVRegs()];
+ }
+
+ void SetCodeItem(const DexFile::CodeItem* code_item) {
+ code_item_ = code_item;
+ }
+
float GetVRegFloat(size_t i) const {
DCHECK_LT(i, NumberOfVRegs());
// NOTE: Strict-aliasing?
@@ -346,6 +361,10 @@
return lock_count_data_;
}
+ static size_t LockCountDataOffset() {
+ return OFFSETOF_MEMBER(ShadowFrame, lock_count_data_);
+ }
+
static size_t LinkOffset() {
return OFFSETOF_MEMBER(ShadowFrame, link_);
}
@@ -366,6 +385,18 @@
return OFFSETOF_MEMBER(ShadowFrame, vregs_);
}
+ static size_t ResultRegisterOffset() {
+ return OFFSETOF_MEMBER(ShadowFrame, result_register_);
+ }
+
+ static size_t DexPCPtrOffset() {
+ return OFFSETOF_MEMBER(ShadowFrame, dex_pc_ptr_);
+ }
+
+ static size_t CodeItemOffset() {
+ return OFFSETOF_MEMBER(ShadowFrame, code_item_);
+ }
+
// Create ShadowFrame for interpreter using provided memory.
static ShadowFrame* CreateShadowFrameImpl(uint32_t num_vregs,
ShadowFrame* link,
@@ -375,10 +406,19 @@
return new (memory) ShadowFrame(num_vregs, link, method, dex_pc, true);
}
+ uint16_t* GetDexPCPtr() {
+ return dex_pc_ptr_;
+ }
+
+ JValue* GetResultRegister() {
+ return result_register_;
+ }
+
private:
ShadowFrame(uint32_t num_vregs, ShadowFrame* link, ArtMethod* method,
uint32_t dex_pc, bool has_reference_array)
- : number_of_vregs_(num_vregs), link_(link), method_(method), dex_pc_(dex_pc) {
+ : link_(link), method_(method), result_register_(nullptr), dex_pc_ptr_(nullptr),
+ code_item_(nullptr), number_of_vregs_(num_vregs), dex_pc_(dex_pc) {
// TODO(iam): Remove this parameter, it's an an artifact of portable removal
DCHECK(has_reference_array);
if (has_reference_array) {
@@ -399,12 +439,15 @@
const_cast<const ShadowFrame*>(this)->References());
}
- const uint32_t number_of_vregs_;
// Link to previous shadow frame or null.
ShadowFrame* link_;
ArtMethod* method_;
- uint32_t dex_pc_;
+ JValue* result_register_;
+ uint16_t* dex_pc_ptr_;
+ const DexFile::CodeItem* code_item_;
LockCountData lock_count_data_; // This may contain GC roots when lock counting is active.
+ const uint32_t number_of_vregs_;
+ uint32_t dex_pc_;
// This is a two-part array:
// - [0..number_of_vregs) holds the raw virtual registers, and each element here is always 4