fixed for arm32 js frame

issue:[Add JSFunction EntryFrame & LeaveFrame for arm32](https://gitee.com/openharmony/ark_js_runtime/issues/I4WRJP?from=project-issue)
Signed-off-by: getingke <getingke@huawei.com>

Change-Id: If65fd5064fa3c626f8f0ae3bae550465414af5e7

ecmascript/compiler/circuit_builder.cpp

@@ -338,7 +338,7 @@ GateRef CircuitBuilder::NewRuntimeCallGate(GateRef glue, GateRef target,
         inputs.push_back(arg);
     }
     OpCode opcode(OpCode::RUNTIME_CALL);
-    StubDescriptor *descriptor = GET_STUBDESCRIPTOR(CallRuntimeTrampoline);
+    StubDescriptor *descriptor = GET_STUBDESCRIPTOR(RuntimeCallTrampoline);
     MachineType machineType = GetCallMachineTypeFromVariableType(descriptor->GetReturnType());
     GateType type = VariableType2GateType(descriptor->GetReturnType());
     // 2 : 2 means extra two input gates (target glue)

ecmascript/compiler/compiler_macros.h

@@ -19,7 +19,7 @@
 #include "ecmascript/base/config.h"
 #include "ecmascript/ecma_macros.h"
 
-#define ECMASCRIPT_ENABLE_COMPILER_LOG 0
+#define ECMASCRIPT_ENABLE_COMPILER_LOG 1
 
 #define COMPILER_LOG(level) ECMASCRIPT_ENABLE_COMPILER_LOG && LOG_ECMA(level)
 

ecmascript/compiler/fast_stub_define.h

@@ -135,6 +135,7 @@ namespace panda::ecmascript::kungfu {
     V(Mul2Dyn, 3)                            \
     V(Div2Dyn, 3)                            \
     V(Mod2Dyn, 3)                            \
+<<<<<<< HEAD
     V(GetLexicalEnv, 1)                      \
     V(LoadValueFromConstantPool, 3)          \
     V(CreateEmptyObject, 1)                  \
@@ -160,10 +161,14 @@ namespace panda::ecmascript::kungfu {
     V(SuperCall, 5)                          \
     I(DoubleToInt, 1)                        \
     I(CallRuntimeTrampoline, 4)
+=======
+    I(RuntimeCallTrampoline, 4)
+>>>>>>> 90e2164... fixed for debug compile arm32
 
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 
 #define FAST_STUB_LIST(V)        \
+<<<<<<< HEAD
     V(FastAdd, 3)                       \
     V(FastSub, 3)                       \
     V(FastMul, 3)                       \
@@ -182,6 +187,9 @@ namespace panda::ecmascript::kungfu {
     V(TryLoadICByValue, 5)              \
     V(TryStoreICByName, 5)              \
     V(TryStoreICByValue, 6)
+=======
+    V(FastMulGCTest, 3)
+>>>>>>> 90e2164... fixed for debug compile arm32
 
 #if ECMASCRIPT_COMPILE_INTERPRETER_ASM
 #define FAST_RUNTIME_STUB_LIST(V)       \
@@ -194,7 +202,13 @@ namespace panda::ecmascript::kungfu {
 
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define TEST_FUNC_LIST(V)           \
+<<<<<<< HEAD
     V(FastMulGCTest, 3)             \
+=======
+    V(PhiGateTest, 1)               \
+    V(LoopTest, 1)                  \
+    V(LoopTest1, 1)                 \
+>>>>>>> 90e2164... fixed for debug compile arm32
     V(TestAbsoluteAddressRelocation, 2)
 
 #define CALL_STUB_LIST(V)           \

ecmascript/compiler/llvm_ir_builder.cpp

@@ -43,7 +43,7 @@ LLVMIRBuilder::LLVMIRBuilder(const std::vector<std::vector<GateRef>> *schedule,
     if (circuit_->GetFrameType() == FrameType::INTERPRETER_FRAME) {
         LLVMSetFunctionCallConv(function_, LLVMGHCCallConv);
     } else if (circuit_->GetFrameType() == FrameType::OPTIMIZED_FRAME) {
-        LLVMSetFunctionCallConv(function_, LLVMWebKitJSCallConv);
+        // LLVMSetFunctionCallConv(function_, LLVMWebKitJSCallConv);
     }
     LLVMSetGC(function_, "statepoint-example");
     if (compCfg_->Is32Bit()) {
@@ -300,9 +300,9 @@ void LLVMIRBuilder::GenPrologue(LLVMModuleRef &module, LLVMBuilderRef &builder)
         -8    type
         -16   current sp before call other function
     */
-    if (compCfg_->Is32Bit()) {
-        return;
-    }
+    // if (compCfg_->Is32Bit()) {
+    //     return;
+    // }
     auto frameType = circuit_->GetFrameType();
     if (frameType != FrameType::OPTIMIZED_FRAME) {
         return;
@@ -458,12 +458,12 @@ void LLVMIRBuilder::VisitRuntimeCall(GateRef gate, const std::vector<GateRef> &i
     ASSERT(stubModule_ != nullptr);
     LLVMValueRef callee;
     LLVMValueRef rtoffset;
-    LLVMTypeRef rtfuncType = stubModule_->GetStubFunctionType(FAST_STUB_ID(CallRuntimeTrampoline));
+    LLVMTypeRef rtfuncType = stubModule_->GetStubFunctionType(FAST_STUB_ID(RuntimeCallTrampoline));
     LLVMTypeRef rtfuncTypePtr = LLVMPointerType(rtfuncType, 0);
     LLVMValueRef glue = gateToLLVMMaps_[inList[2]];  // 2 : 2 means skip two input gates (target glue)
     LLVMTypeRef glue_type = LLVMTypeOf(glue);
     rtoffset = LLVMConstInt(glue_type, compCfg_->GetGlueOffset(JSThread::GlueID::RUNTIME_FUNCTIONS) +
-                                (FAST_STUB_ID(CallRuntimeTrampoline) - FAST_STUB_MAXCOUNT) * slotSize_, 0);
+                                (FAST_STUB_ID(RuntimeCallTrampoline) - FAST_STUB_MAXCOUNT) * slotSize_, 0);
     LLVMValueRef rtbaseoffset = LLVMBuildAdd(builder_, glue, rtoffset, "");
     LLVMValueRef rtbaseAddr = LLVMBuildIntToPtr(builder_, rtbaseoffset, LLVMPointerType(glue_type, 0), "");
     LLVMValueRef llvmAddr = LLVMBuildLoad(builder_, rtbaseAddr, "");
@@ -484,7 +484,7 @@ void LLVMIRBuilder::VisitRuntimeCall(GateRef gate, const std::vector<GateRef> &i
         return;
     }
     LLVMValueRef runtimeCall = LLVMBuildCall(builder_, callee, params, inList.size() + 1, "");
-    LLVMSetInstructionCallConv(runtimeCall, LLVMWebKitJSCallConv);
+    // LLVMSetInstructionCallConv(runtimeCall, LLVMWebKitJSCallConv);
     gateToLLVMMaps_[gate] = runtimeCall;
 }
 
@@ -584,9 +584,19 @@ void LLVMIRBuilder::VisitCall(GateRef gate, const std::vector<GateRef> &inList)
     callee = LLVMBuildIntToPtr(builder_, llvmAddr, rtfuncTypePtr, "cast");
     // 16 : params limit
     LLVMValueRef params[16];
+    int extraParameterCnt = 0;
     for (size_t paraIdx = paraStartIndex; paraIdx < inList.size(); ++paraIdx) {
         GateRef gateTmp = inList[paraIdx];
         params[paraIdx - paraStartIndex] = gateToLLVMMaps_[gateTmp];
+        if (compCfg_->Is32Bit() && (paraIdx - paraStartIndex == 0)) {
+            // for is32 push multiple paramer in order to use r0-r4
+            if (calleeDescriptor->GetStubKind() != StubDescriptor::CallStubKind::RUNTIME_STUB) {
+                for (int i = 0; i < reservedSlotForArm32; i++) {
+                    params[paraIdx - paraStartIndex] = gateToLLVMMaps_[gateTmp];
+                }
+                extraParameterCnt += reservedSlotForArm32;
+            }
+        }
     }
     if (callee == nullptr) {
         COMPILER_LOG(ERROR) << "callee nullptr";
@@ -596,7 +606,7 @@ void LLVMIRBuilder::VisitCall(GateRef gate, const std::vector<GateRef> &inList)
         SaveCurrentSP();
     }
 
-    gateToLLVMMaps_[gate] = LLVMBuildCall(builder_, callee, params, inList.size() - paraStartIndex, "");
+    gateToLLVMMaps_[gate] = LLVMBuildCall(builder_, callee, params, inList.size() - paraStartIndex + extraParameterCnt, "");
     return;
 }
 
@@ -816,7 +826,7 @@ void LLVMIRBuilder::VisitConstant(GateRef gate, std::bitset<64> value) // 64: bi
             llvmValue = LLVMBuildIntToPtr(builder_, llvmValue, type, "");
         } else if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
             LLVMValueRef tmp1Value =
-                    LLVMBuildLShr(builder_, llvmValue, LLVMConstInt(LLVMInt32Type(), 32, 0), ""); // 32: offset
+                    LLVMBuildLShr(builder_, llvmValue, LLVMConstInt(LLVMInt64Type(), 32, 0), ""); // 32: offset
             LLVMValueRef tmp2Value = LLVMBuildIntCast(builder_, llvmValue, LLVMInt32Type(), ""); // low
             LLVMValueRef emptyValue = LLVMGetUndef(type);
             tmp1Value = LLVMBuildIntToPtr(builder_, tmp1Value, LLVMPointerType(LLVMInt8Type(), 1), "");
@@ -880,7 +890,11 @@ void LLVMIRBuilder::VisitParameter(GateRef gate)
 {
     int argth = circuit_->LoadGatePtrConst(gate)->GetBitField();
     COMPILER_LOG(DEBUG) << " Parameter value" << argth;
+    if (compCfg_->Is32Bit() && argth > 0) {
+        argth += reservedSlotForArm32;
+    }
     LLVMValueRef value = LLVMGetParam(function_, argth);
+
     ASSERT(LLVMTypeOf(value) == ConvertLLVMTypeFromGate(gate));
     gateToLLVMMaps_[gate] = value;
     COMPILER_LOG(DEBUG) << "VisitParameter set gate:" << gate << "  value:" << value;
@@ -1544,7 +1558,7 @@ void LLVMIRBuilder::VisitZExtInt(GateRef gate, GateRef e1)
     COMPILER_LOG(DEBUG) << "operand 0: " << LLVMValueToString(e1Value);
     ASSERT(GetBitWidthFromMachineType(circuit_->LoadGatePtrConst(e1)->GetMachineType()) <=
         GetBitWidthFromMachineType(circuit_->LoadGatePtrConst(gate)->GetMachineType()));
-    LLVMValueRef result = LLVMBuildZExt(builder_, e1Value, ConvertLLVMTypeFromGate(gate), "");
+    LLVMValueRef result = LLVMBuildZExt(builder_, e1Value, ConvertLLVMTypeFromGate(gate), "gtk");
     gateToLLVMMaps_[gate] = result;
     COMPILER_LOG(DEBUG) << "result: " << LLVMValueToString(result);
 }
@@ -1614,7 +1628,7 @@ void LLVMIRBuilder::VisitChangeInt64ToTagged(GateRef gate, GateRef e1)
     LLVMValueRef result;
     if (compCfg_->Is32Bit()) {
         LLVMValueRef tmp1Value =
-            LLVMBuildLShr(builder_, e1Value, LLVMConstInt(LLVMInt32Type(), 32, 0), ""); // 32: offset
+            LLVMBuildLShr(builder_, e1Value, LLVMConstInt(LLVMInt64Type(), 32, 0), ""); // 32: offset
         LLVMValueRef tmp2Value = LLVMBuildIntCast(builder_, e1Value, LLVMInt32Type(), ""); // low
         LLVMTypeRef vectorType = LLVMVectorType(LLVMPointerType(LLVMInt8Type(), 1), 2);  // 2: packed vector type
         LLVMValueRef emptyValue = LLVMGetUndef(vectorType);
@@ -1711,11 +1725,21 @@ LLVMTypeRef LLVMStubModule::GetLLVMFunctionTypeStubDescriptor(StubDescriptor *st
     LLVMTypeRef returnType = ConvertLLVMTypeFromVariableType(stubDescriptor->GetReturnType());
     std::vector<LLVMTypeRef> paramTys;
     auto paramCount = stubDescriptor->GetParametersCount();
+    int extraParameterCnt = 0;
     for (int i = 0; i < paramCount; i++) {
         auto paramsType = stubDescriptor->GetParametersType();
+        if (compCfg_.Is32Bit() && (i == 0)) {
+            if (stubDescriptor->GetStubKind() != StubDescriptor::CallStubKind::RUNTIME_STUB) {
+                // for is32 push extra multiple paramer in order to use r0-r4
+                for (int t = 0; t < reservedSlotForArm32; t++) {
+                    paramTys.push_back(ConvertLLVMTypeFromVariableType(paramsType[i]));
+                }
+                extraParameterCnt += reservedSlotForArm32;
+            }
+        }
         paramTys.push_back(ConvertLLVMTypeFromVariableType(paramsType[i]));
     }
-    auto functype = LLVMFunctionType(returnType, paramTys.data(), paramCount, stubDescriptor->GetVariableArgs());
+    auto functype = LLVMFunctionType(returnType, paramTys.data(), paramCount + extraParameterCnt, stubDescriptor->GetVariableArgs());
     return functype;
 }
 

ecmascript/compiler/llvm_ir_builder.h

@@ -155,6 +155,7 @@ private:
 #endif
     LLVMModuleRef module_;
     CompilationConfig compCfg_;
+    static constexpr int reservedSlotForArm32 = 3; // r1-r3
 };
 
 
@@ -275,6 +276,7 @@ private:
     std::set<OpCode::Op> opCodeHandleIgnore;
     int slotSize_;
     LLVMTypeRef slotType_;
+    static constexpr int reservedSlotForArm32 = 3; // r1-r3
 };
 }  // namespace panda::ecmascript::kungfu
 #endif  // PANDA_RUNTIME_ECMASCRIPT_COMPILER_LLVM_IR_BUILDER_H

ecmascript/compiler/stub_descriptor.cpp

@@ -1045,9 +1045,10 @@ CALL_STUB_INIT_DESCRIPTOR(IterNext)
     descriptor->SetStubKind(StubDescriptor::CallStubKind::RUNTIME_STUB);
 }
 
-CALL_STUB_INIT_DESCRIPTOR(CallRuntimeTrampoline)
+CALL_STUB_INIT_DESCRIPTOR(RuntimeCallTrampoline)
 {
     /* 4 : 4 input parameters */
+<<<<<<< HEAD
     StubDescriptor CallRuntimeTrampoline("CallRuntimeTrampoline", 0, 4,
         ArgumentsOrder::DEFAULT_ORDER, VariableType::JS_ANY());
     *descriptor = CallRuntimeTrampoline;
@@ -1056,6 +1057,16 @@ CALL_STUB_INIT_DESCRIPTOR(CallRuntimeTrampoline)
         VariableType::INT64(),
         VariableType::INT64(),
         VariableType::INT64(),
+=======
+    StubDescriptor runtimeCallTrampoline("RuntimeCallTrampoline", 0, 4,
+        ArgumentsOrder::DEFAULT_ORDER, StubMachineType::TAGGED);
+    *descriptor = runtimeCallTrampoline;
+    std::array<StubMachineType, 4> params = { /* 4 : 4 input parameters */
+        StubMachineType::NATIVE_POINTER,
+        StubMachineType::UINT64,
+        StubMachineType::UINT64,
+        StubMachineType::UINT64,
+>>>>>>> 90e2164... fixed for debug compile arm32
     };
     descriptor->SetVariableArgs(true);
     descriptor->SetParameters(params.data());

ecmascript/frames.h

@@ -220,7 +220,7 @@
 
 namespace panda::ecmascript {
 class JSThread;
-enum class FrameType: uint64_t {
+enum class FrameType: uint64_t  {
     OPTIMIZED_FRAME = 0,
     OPTIMIZED_ENTRY_FRAME = 1,
     INTERPRETER_FRAME = 2,
@@ -245,8 +245,8 @@ public:
 #ifdef PANDA_TARGET_ARM32
     static constexpr int SP_DWARF_REG_NUM = 13;
     static constexpr int FP_DWARF_REG_NUM = 11;
-    static constexpr int CALLSITE_SP_TO_FP_DELTA = 0;
-    static constexpr int INTERPER_FRAME_FP_TO_FP_DELTA = 0;
+    static constexpr int CALLSITE_SP_TO_FP_DELTA = -2;
+    static constexpr int INTERPER_FRAME_FP_TO_FP_DELTA = -2;
 #else
     static constexpr int SP_DWARF_REG_NUM = 0;
     static constexpr int FP_DWARF_REG_NUM = 0;
@@ -443,7 +443,9 @@ struct OptimizedLeaveFrame {
     FrameType type;
     uintptr_t callsiteFp; // thread sp set here
     uintptr_t returnAddr;
+#ifndef PANDA_TARGET_32
     uint64_t argRuntimeId;
+#endif
     uint64_t argPatchId;
     uint64_t argc;
     // argv[] is dynamic
@@ -454,7 +456,11 @@ struct OptimizedLeaveFrame {
     }
     uintptr_t GetCallSiteSp()
     {
+#ifndef PANDA_TARGET_32
         return ToUintPtr(this) + MEMBER_OFFSET(OptimizedLeaveFrame, argRuntimeId);
+#else
+        return ToUintPtr(this) + MEMBER_OFFSET(OptimizedLeaveFrame, argPatchId);
+#endif
     }
 };
 

ecmascript/interpreter/frame_handler.cpp

@@ -334,18 +334,11 @@ void FrameIterator::Iterate(const RootVisitor &v0, const RootRangeVisitor &v1) c
             OptimizedLeaveFrame *frame = OptimizedLeaveFrame::GetFrameFromSp(current);
             OptimizedLeaveFrameHandler(reinterpret_cast<uintptr_t *>(current)).Iterate(v0,
                 v1, derivedPointers, isVerifying);
-            //  arm32 only support stub, optimized entry frame don't exist, when interpret call stub
-            // don't customed prologue handle. when stub call runtime, generate optimized Leave Frame.
-            //  arm64 and x86_64 support stub and aot, when aot/stub call runtime, generate Optimized
+            //  arm32, arm64 and x86_64 support stub and aot, when aot/stub call runtime, generate Optimized
             // Leave Frame.
-#ifdef PANDA_TARGET_ARM32
-            current = reinterpret_cast<JSTaggedType *>(frame->callsiteFp);
-            ASSERT(FrameHandler(current).IsInterpretedFrame());
-#else
             current = reinterpret_cast<JSTaggedType *>(frame->callsiteFp);
             ASSERT(FrameHandler(current).GetFrameType() == FrameType::OPTIMIZED_ENTRY_FRAME ||
             FrameHandler(current).GetFrameType() == FrameType::OPTIMIZED_FRAME);
-#endif
         }
     }
 }

ecmascript/interpreter/interpreter-inl.h

@@ -16,6 +16,7 @@
 #ifndef ECMASCRIPT_INTERPRETER_INTERPRETER_INL_H
 #define ECMASCRIPT_INTERPRETER_INTERPRETER_INL_H
 
+#include "ecmascript/trampoline/trampoline.h"
 #include "ecmascript/class_linker/program_object-inl.h"
 #if defined(ECMASCRIPT_SUPPORT_CPUPROFILER)
 #include "ecmascript/dfx/cpu_profiler/cpu_profiler.h"
@@ -1296,7 +1297,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
     }
     HANDLE_OPCODE(HANDLE_TYPEOFDYN_PREF) {
         LOG_INST() << "intrinsics::typeofdyn";
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
         auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(FastTypeOf));
         typedef JSTaggedType (*PFFastTypeOf)(uintptr_t, JSTaggedType);
         auto fastTypeOfPtr = reinterpret_cast<PFFastTypeOf>(stubAddr);
@@ -1472,10 +1473,17 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
         JSTaggedValue left = GET_VREG_VALUE(v0);
         JSTaggedValue right = acc;
 #ifdef ECMASCRIPT_ENABLE_STUB_AOT
-        auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(FastMul));
-        typedef JSTaggedType (*PFFastMul)(JSTaggedType, JSTaggedType);
-        auto fastMulPtr = reinterpret_cast<PFFastMul>(stubAddr);
-        JSTaggedValue value = JSTaggedValue(fastMulPtr(left.GetRawData(), right.GetRawData()));
+        auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(FastMulGCTest));
+        // typedef JSTaggedType (*PFFastMul)(JSTaggedType, JSTaggedType);
+        // auto fastMulPtr = reinterpret_cast<PFFastMul>(stubAddr);
+        // JSTaggedValue value = JSTaggedValue(fastMulPtr(left.GetRawData(), right.GetRawData()));
+        JSTaggedType argV[2] = {0x0a, 0x0a};
+        argV[0] = left.GetRawData();
+        argV[1] = right.GetRawData();
+        JSTaggedValue res1(JSFunctionEntry(thread->GetGlueAddr(),
+            reinterpret_cast<uintptr_t>(thread->GetCurrentSPFrame()), 2, 2, argV, stubAddr));
+        (void)res1;
+        JSTaggedValue value = FastRuntimeStub::FastMul(left, right);
 #else
         JSTaggedValue value = FastRuntimeStub::FastMul(left, right);
 #endif
@@ -1516,7 +1524,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
         JSTaggedValue left = GET_VREG_VALUE(vs);
         JSTaggedValue right = GET_ACC();
 
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
         auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(FastMod));
         typedef JSTaggedType (*PFFastMod)(uintptr_t, JSTaggedType, JSTaggedType);
         auto fastModPtr = reinterpret_cast<PFFastMod>(stubAddr);
@@ -1542,7 +1550,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
                    << " v" << v0;
         JSTaggedValue left = GET_VREG_VALUE(v0);
         JSTaggedValue right = acc;
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
         auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(FastEqual));
         typedef JSTaggedType (*PFFastEqual)(JSTaggedType, JSTaggedType);
         auto fastEqualPtr = reinterpret_cast<PFFastEqual>(stubAddr);
@@ -2490,7 +2498,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
             JSTaggedValue value = GET_ACC();
             // fast path
             SAVE_ACC();
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(SetPropertyByNameWithOwn));
             typedef JSTaggedType (*PFSetPropertyByName)(uintptr_t, JSTaggedType, JSTaggedType, JSTaggedType);
             auto setPropertyByNamePtr = reinterpret_cast<PFSetPropertyByName>(stubAddr);
@@ -2842,7 +2850,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
         JSTaggedValue receiver = GET_VREG_VALUE(v0);
         // fast path
         if (LIKELY(receiver.IsHeapObject())) {
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(GetPropertyByIndex));
             typedef JSTaggedType (*PFGetPropertyByIndex)(uintptr_t, JSTaggedType, uint32_t);
             auto getPropertyByIndex = reinterpret_cast<PFGetPropertyByIndex>(stubAddr);
@@ -2875,7 +2883,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
             SAVE_ACC();
             JSTaggedValue value = GET_ACC();
             // fast path
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(SetPropertyByIndex));
             typedef JSTaggedType (*PFSetPropertyByIndex)(uintptr_t, JSTaggedType, uint32_t, JSTaggedType);
             auto setPropertyByIndex = reinterpret_cast<PFSetPropertyByIndex>(stubAddr);
@@ -2938,7 +2946,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
 #endif
         // fast path
         if (LIKELY(receiver.IsHeapObject())) {
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(GetPropertyByValue));
             typedef JSTaggedType (*PFGetPropertyByValue)(uintptr_t, JSTaggedType, JSTaggedType);
             auto getPropertyByValuePtr = reinterpret_cast<PFGetPropertyByValue>(stubAddr);
@@ -3240,7 +3248,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
             JSTaggedValue value = GET_ACC();
             // fast path
             SAVE_ACC();
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(SetPropertyByNameWithOwn));
             typedef JSTaggedType (*PFSetPropertyByName)(uintptr_t, JSTaggedType, JSTaggedType, JSTaggedType);
             auto setPropertyByNamePtr = reinterpret_cast<PFSetPropertyByName>(stubAddr);
@@ -3338,7 +3346,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
 
         if (LIKELY(receiver.IsHeapObject())) {
             // fast path
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(GetPropertyByName));
             typedef JSTaggedType (*PFGetPropertyByName)(uintptr_t, JSTaggedType, JSTaggedType);
             auto getPropertyByNamePtr = reinterpret_cast<PFGetPropertyByName>(stubAddr);
@@ -3377,7 +3385,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
 
             if (LIKELY(firstValue.IsHeapObject())) {
                 JSTaggedValue secondValue = profileTypeArray->Get(slotId + 1);
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
                 auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(TryStoreICByName));
                 typedef JSTaggedType (*PFTryStoreICByName)(uintptr_t,
                     JSTaggedType, JSTaggedType, JSTaggedType, JSTaggedType);
@@ -3413,7 +3421,7 @@ NO_UB_SANITIZE void EcmaInterpreter::RunInternal(JSThread *thread, ConstantPool
             value = GET_ACC();
             // fast path
             SAVE_ACC();
-#ifdef ECMASCRIPT_ENABLE_STUB_AOT
+#ifdef ECMASCRIPT_ENABLE_STUB_AOT1
             auto stubAddr = thread->GetFastStubEntry(FAST_STUB_ID(SetPropertyByName));
             typedef JSTaggedType (*PFSetPropertyByName)(uintptr_t, JSTaggedType, JSTaggedType, JSTaggedType);
             auto setPropertyByNamePtr = reinterpret_cast<PFSetPropertyByName>(stubAddr);

ecmascript/runtime_trampolines.h

@@ -21,7 +21,7 @@
 #include "ecmascript/js_thread.h"
 
 namespace panda::ecmascript {
-extern "C" JSTaggedType CallRuntimeTrampoline(uintptr_t glue, uint64_t runtime_id, uint64_t patch_id,
+extern "C" JSTaggedType RuntimeCallTrampoline(uintptr_t glue, uint64_t runtime_id, uint64_t patch_id,
                                               uint64_t argc, ...);
 class RuntimeTrampolines {
 public:

ecmascript/trampoline/call_trampoline_amd64.S

@@ -26,6 +26,7 @@
     pushq %rdi
 .endm
 
+ /* JSFunctionEntry Callee Register restore Macro */
 .macro POP_GENERAL_REGS
     /* caller restore */
     popq %rdi
@@ -98,7 +99,7 @@ JSFunctionEntry:
 
 
 
-// uint64_t CallRuntimeTrampoline(uintptr_t glue, uint64_t runtime_id, uint64_t patch_id, uint64_t argc, ...);
+// uint64_t RuntimeCallTrampoline(uintptr_t glue, uint64_t runtime_id, uint64_t patch_id, uint64_t argc, ...);
 // webkit_jscc calling convention call runtime_id's runtion function(c-abi)
 // JSTaggedType (*)(uintptr_t argGlue, uint64_t argc, JSTaggedType argv[])
 // Input:
@@ -133,9 +134,9 @@ JSFunctionEntry:
 //  sp - 16: rbp <---------current rbp & current sp
 //  current sp - 8:  type
 
-.global CallRuntimeTrampoline
-.type CallRuntimeTrampoline, %function
-CallRuntimeTrampoline:
+.global RuntimeCallTrampoline
+.type RuntimeCallTrampoline, %function
+RuntimeCallTrampoline:
     pushq   %rbp
     movq    %rsp, %rbp     // set frame pointer
     movq    %rbp, ASM_GLUE_CURRENT_FRAME_OFFSET(%rax)   // save to thread->currentFrame_

ecmascript/trampoline/call_trampoline_arm32.S

@@ -13,22 +13,30 @@
  * limitations under the License.
  */
 
-// uint64_t JSFunctionCallTrampoline(uintptr_t glue, uintptr_t prevFp, uint32_t expectedNumArgs,
+#include "ecma_asm_defines.h"
+
+// uint64_t JSFunctionEntry(uintptr_t glue, uintptr_t prevFp, uint32_t expectedNumArgs,
 //                                uint32_t actualNumArgs, const JSTaggedType argV[], uintptr_t codeAddr);
 // Input:
 // %r0 - glue
 // %r1 - prevFp
 // %r2 - expectedNumArgs
 // %r3 - actualNumArgs
-// %r5  - argV
-// %r12  - codeAddr
-.global JSFunctionCallTrampoline
-.type JSFunctionCallTrampoline, %function
-JSFunctionCallTrampoline:
-    push    {r4, r5, r6, r7, r8, r9, r10, r11, lr}
-    add     r11, sp, #28
-    sub     sp, sp, #4
-    mov     r10, r2
+
+.global JSFunctionEntry
+.type JSFunctionEntry, %function
+JSFunctionEntry:
+    push    {r4, r5, r6, r7, r8, r9, r10, lr}   // callee save
+    push    {r11}                               // rbp
+    mov     r11, sp                             // set frame pointer
+    // construct frame
+    mov     r6, #0
+    mov     r8, #JS_ENTRY_FRAME_TYPE
+    push    {r6, r8}                            // frame type
+    push    {r1}                                // prev managed frame
+    push    {r0}
+
+    mov     r10, r2                             // save expected 
     cmp     r2, r3
     bls     .LCopyArguments
     mov     r6, #0
@@ -46,64 +54,91 @@ JSFunctionCallTrampoline:
     rscs    r5, r4, #0
     blo     .LCopyExtraAument
 .LCopyArguments:
-    ldr     r12, [r11, #12]
+    ldr     lr, [r11, #40]   // load js function code address
     cmp     r10, r3
-    add     lr, r1, #8
-    add     r2, r1, #16
     movls   r3, r10
-    add     r3, r3, #3          // add 3 for calltarget newtarget this LCopyArguments
-    ldr     r4, [r11, #8]
-    mov     r7, #0
+    cmp     r3, #0
+    beq     .LInvokeJSFunction
+    ldr     r4, [r11, #36]   // load argv address
     add     r5, r4, r3, lsl #3
-.LCopyLoop:                                @ =>This Inner Loop Header: Depth=1
-    ldr     r6, [r5, #4]
-    ldr     r8, [r5]
+.LCopyLoop:
     sub     r5, r5, #8
     subs    r3, r3, #1
-    mov     r9, r6
+    ldr     r8, [r5, #4]
+    ldr     r6, [r5]
     push    {r6, r8}
 
     bne     .LCopyLoop
-
-    mov     r2, r12
-    bl      InvokeCompiledJSFunction
-    mov     r2, #24
+.LInvokeJSFunction:
+    mov     r10, #OPTIMIZE_FRAME_TYPE  // set frame type
+    blx     lr
+    mov     r2, #0                    // need fix for calltarget thisTarget NewTarget
     add     r2, r2, r10, lsl #3
     add     sp, sp, r2
-    sub     sp, r11, #28
-    pop     {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+
+    pop     {r4}
+    pop     {r1}
+    str     r1, [r4, #ASM_GLUE_CURRENT_FRAME_OFFSET]
+    add     sp, sp, #8
+    pop     {r11}
+    pop     {r4, r5, r6, r7, r8, r9, r10, lr}
     bx      lr
 
-// uint64_t InvokeCompiledJSFunction
+// uint64_t RuntimeCallTrampoline(uintptr_t glue, uint64_t runtime_id, uint64_t patch_id, uint64_t argc, ...);
+// webkit_jscc calling convention call runtime_id's runtion function(c-abi)
+// JSTaggedType (*)(uintptr_t argGlue, uint64_t argc, JSTaggedType argv[])
 // Input:
 // %r0 - glue
-// %r1 - prev fp
-// %r2 - code address
+// %r2, %r3 - runtime_id 
+// stack layout:
+// sp + N*8 argvN
+// ........
+// sp + 24: argv1
+// sp + 16: argv0
+// sp + 8: argc
+// sp:  patch_id
+// construct Leave Frame:
+//   +--------------------------+
+//   |      argv[]              |
+//   +--------------------------+ ---
+//   |       argc               |   ^
+//   |- - - - - - -- - - - - - -|   |
+//   |       patchID            | Fixed
+//   |- - - - - - - -- - --  - -| OptLeaveFrame
+//   |       returnAddr         |   |
+//   |- - - - - - - - -- - - - -|   |
+//   |       callsiteFp         |   |
+//   |-------------------- - - -|   |
+//   |     frameType            |   v
+//   +--------------------------+ ---
+//   |  callee save registers   |
+//   +--------------------------+
 
-.global InvokeCompiledJSFunction
-.type InvokeCompiledJSFunction, %function
-InvokeCompiledJSFunction:
-    // construct the frame  
-    push   {r11, lr}
-    mov     lr, #1
-    push    {lr, r1}            // js entry frame type, prev managerd fp
-    add     r11, sp, #8
-    mov     lr, r2
-    blx     lr
-    add     sp, sp, #8
-    pop     {r11, lr}
-    bx      lr
+.global RuntimeCallTrampoline
+.type RuntimeCallTrampoline, %function
+RuntimeCallTrampoline:
+    push    {r11, lr}                                   // save register for fp, rip
+    mov     r11, sp                                     // set frame pointer = callsiteFp
+    str     r11, [r0, #ASM_GLUE_CURRENT_FRAME_OFFSET]   // save to thread->currentFrame_
+    add     sp, sp, #-8                                 // increase frame size
+    push    {r4, r5}                                    // callee save
+
+    // construct leave frame
+    mov     r4, #0
+    mov     r5, #LEAVE_FRAME_TYPE
+    str     r4, [r11, #-4]
+    str     r5, [r11, #-8]                              // save frame type
 
 
-// uint64_t CallRuntimeTrampoline(uintptr_t glue, uint32_t runtime_id, uint32_t patch_id, uint64_t argc, ...);
-// webkit_jscc calling convention
-.global CallRuntimeTrampoline
-.type CallRuntimeTrampoline, %function
-CallRuntimeTrampoline:   // todo : need to rewrite
-    push    {r4, r5, r6, r7, r8, r9, r10, r11, lr}
-    add     r11, sp, #28
-    sub     sp, sp, #4
-    bx lr
+    // load runtime trampoline address
+    add     r2, r0, r2, lsl #2
+    ldr     r4, [r2, #ASM_GLUE_RUNTIME_FUNCTIONS_OFFSET]
+    ldr     r1, [r11, #16]  // argc
+    add     r2, r11,  #24   // argv[]
+    bx      r4
+
+    // descontruct leave frame and callee save register
+    pop     {r4, r5}
     add     sp, sp, #8
     pop     {r11, lr}
     bx      lr

js_runtime_config.gni

@@ -13,9 +13,9 @@
 
 ark_root = "//ark/runtime_core"
 js_root = "//ark/js_runtime"
-compile_llvm_online = false
+compile_llvm_online = true
 run_with_asan = false
-enable_stub_aot = false
+enable_stub_aot = true
 asan_lib_path = "/usr/lib/llvm-10/lib/clang/10.0.0/lib/linux"
 
 # For OpenHarmony build, always link with the static lib: