VM: New calling convention for generated code.

runtime/vm/stub_code_arm64.cc

 // Copyright (c) 2014, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"
 #if defined(TARGET_ARCH_ARM64)
 
 #include "vm/assembler.h"
 #include "vm/code_generator.h"
 #include "vm/compiler.h"
@@ skipping 23 matching lines @@
 //   SP : address of last argument in argument array.
 //   SP + 8*R4 - 8 : address of first argument in argument array.
 //   SP + 8*R4 : address of return value.
 //   R5 : address of the runtime function to call.
 //   R4 : number of arguments to the call.
 void StubCode::GenerateCallToRuntimeStub(Assembler* assembler) {
   const intptr_t thread_offset = NativeArguments::thread_offset();
   const intptr_t argc_tag_offset = NativeArguments::argc_tag_offset();
   const intptr_t argv_offset = NativeArguments::argv_offset();
   const intptr_t retval_offset = NativeArguments::retval_offset();
-  const intptr_t exitframe_last_param_slot_from_fp = 1;
 
   __ SetPrologueOffset();
   __ Comment("CallToRuntimeStub");
   __ EnterStubFrame();
 
   COMPILE_ASSERT((kAbiPreservedCpuRegs & (1 << R28)) != 0);
   __ LoadIsolate(R28);
 
   // Save exit frame information to enable stack walking as we are about
   // to transition to Dart VM C++ code.
@@ skipping 29 matching lines @@
 
   // There are no runtime calls to closures, so we do not need to set the tag
   // bits kClosureFunctionBit and kInstanceFunctionBit in argc_tag_.
   ASSERT(argc_tag_offset == 1 * kWordSize);
   __ mov(R1, R4);  // Set argc in NativeArguments.
 
   ASSERT(argv_offset == 2 * kWordSize);
   __ add(R2, ZR, Operand(R4, LSL, 3));
   __ add(R2, FP, Operand(R2));  // Compute argv.
   // Set argv in NativeArguments.
-  __ AddImmediate(R2, R2, exitframe_last_param_slot_from_fp * kWordSize);
+  __ AddImmediate(R2, R2, kParamEndSlotFromFp * kWordSize);
 
     ASSERT(retval_offset == 3 * kWordSize);
   __ AddImmediate(R3, R2, kWordSize);
 
   __ StoreToOffset(R0, SP, thread_offset);
   __ StoreToOffset(R1, SP, argc_tag_offset);
   __ StoreToOffset(R2, SP, argv_offset);
   __ StoreToOffset(R3, SP, retval_offset);
   __ mov(R0, SP);  // Pass the pointer to the NativeArguments.
 
@@ skipping 234 matching lines @@
 //   R4: arguments descriptor array.
 void StubCode::GenerateCallStaticFunctionStub(Assembler* assembler) {
   // Create a stub frame as we are pushing some objects on the stack before
   // calling into the runtime.
   __ EnterStubFrame();
   // Setup space on stack for return value and preserve arguments descriptor.
   __ Push(R4);
   __ PushObject(Object::null_object());
   __ CallRuntime(kPatchStaticCallRuntimeEntry, 0);
   // Get Code object result and restore arguments descriptor array.
-  __ Pop(R0);
+  __ Pop(CODE_REG);
   __ Pop(R4);
   // Remove the stub frame.
   __ LeaveStubFrame();
   // Jump to the dart function.
-  __ LoadFieldFromOffset(R0, R0, Code::entry_point_offset());
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ br(R0);
 }
 
 
 // Called from a static call only when an invalid code has been entered
 // (invalid because its function was optimized or deoptimized).
 // R4: arguments descriptor array.
 void StubCode::GenerateFixCallersTargetStub(Assembler* assembler) {
+  // Load code pointer to this stub from the thread:
+  // The one that is passed in, is not correct - it points to the code object
+  // that needs to be replaced.
+  __ ldr(CODE_REG, Address(THR, Thread::fix_callers_target_code_offset()));
   // Create a stub frame as we are pushing some objects on the stack before
   // calling into the runtime.
   __ EnterStubFrame();
   // Setup space on stack for return value and preserve arguments descriptor.
   __ Push(R4);
   __ PushObject(Object::null_object());
   __ CallRuntime(kFixCallersTargetRuntimeEntry, 0);
   // Get Code object result and restore arguments descriptor array.
-  __ Pop(R0);
+  __ Pop(CODE_REG);
   __ Pop(R4);
   // Remove the stub frame.
   __ LeaveStubFrame();
   // Jump to the dart function.
-  __ LoadFieldFromOffset(R0, R0, Code::entry_point_offset());
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ br(R0);
 }
 
 
 // Called from object allocate instruction when the allocation stub has been
 // disabled.
 void StubCode::GenerateFixAllocationStubTargetStub(Assembler* assembler) {
+  // Load code pointer to this stub from the thread:
+  // The one that is passed in, is not correct - it points to the code object
+  // that needs to be replaced.
+  __ ldr(CODE_REG, Address(THR, Thread::fix_allocation_stub_code_offset()));
   __ EnterStubFrame();
   // Setup space on stack for return value.
   __ PushObject(Object::null_object());
   __ CallRuntime(kFixAllocationStubTargetRuntimeEntry, 0);
   // Get Code object result.
-  __ Pop(R0);
+  __ Pop(CODE_REG);
   // Remove the stub frame.
   __ LeaveStubFrame();
   // Jump to the dart function.
-  __ LoadFieldFromOffset(R0, R0, Code::entry_point_offset());
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ br(R0);
 }
 
 
 // Input parameters:
 //   R2: smi-tagged argument count, may be zero.
 //   FP[kParamEndSlotFromFp + 1]: last argument.
 static void PushArgumentsArray(Assembler* assembler) {
   // Allocate array to store arguments of caller.
   __ LoadObject(R1, Object::null_object());
@@ skipping 37 matching lines @@
 // Stack after TagAndPushPP() below:
 //   +------------------+
 //   | Saved PP         | <- PP
 //   +------------------+
 //   | PC marker        | <- TOS
 //   +------------------+
 //   | Saved FP         | <- FP of stub
 //   +------------------+
 //   | return-address   |  (deoptimization point)
 //   +------------------+
+//   | Saved CODE_REG   |
+//   +------------------+
 //   | ...              | <- SP of optimized frame
 //
 // Parts of the code cannot GC, part of the code can GC.
 static void GenerateDeoptimizationSequence(Assembler* assembler,
-                                           bool preserve_result) {
+                                           DeoptStubKind kind) {
   // DeoptimizeCopyFrame expects a Dart frame, i.e. EnterDartFrame(0), but there
   // is no need to set the correct PC marker or load PP, since they get patched.
   __ EnterStubFrame();
 
   // The code in this frame may not cause GC. kDeoptimizeCopyFrameRuntimeEntry
   // and kDeoptimizeFillFrameRuntimeEntry are leaf runtime calls.
   const intptr_t saved_result_slot_from_fp =
       kFirstLocalSlotFromFp + 1 - (kNumberOfCpuRegisters - R0);
   // Result in R0 is preserved as part of pushing all registers below.
 
   // Push registers in their enumeration order: lowest register number at
   // lowest address.
   for (intptr_t i = kNumberOfCpuRegisters - 1; i >= 0; i--) {
     const Register r = static_cast<Register>(i);
-    __ str(r, Address(SP, -1 * kWordSize, Address::PreIndex));
+    if (r == CODE_REG) {
+      // Save the original value of CODE_REG pushed before invoking this stub
+      // instead of the value used to call this stub.
+      COMPILE_ASSERT(R25 > CODE_REG);
+      __ ldr(R25, Address(FP, 2 * kWordSize));
+      __ str(R25, Address(SP, -1 * kWordSize, Address::PreIndex));
+    } else {
+      __ str(r, Address(SP, -1 * kWordSize, Address::PreIndex));
+    }
   }
 
   for (intptr_t reg_idx = kNumberOfVRegisters - 1; reg_idx >= 0; reg_idx--) {
     VRegister vreg = static_cast<VRegister>(reg_idx);
     __ PushQuad(vreg);
   }
 
   __ mov(R0, SP);  // Pass address of saved registers block.
+  __ LoadImmediate(R1, kind == kLazyDeopt ? 1 : 0);
   __ ReserveAlignedFrameSpace(0);
-  __ CallRuntime(kDeoptimizeCopyFrameRuntimeEntry, 1);
+  __ CallRuntime(kDeoptimizeCopyFrameRuntimeEntry, 2);
   // Result (R0) is stack-size (FP - SP) in bytes.
 
+  const bool preserve_result = (kind == kLazyDeopt);
   if (preserve_result) {
     // Restore result into R1 temporarily.
     __ LoadFromOffset(R1, FP, saved_result_slot_from_fp * kWordSize);
   }
 
   // There is a Dart Frame on the stack. We must restore PP and leave frame.
+  __ RestoreCodePointer();
   __ LeaveStubFrame();
   __ sub(SP, FP, Operand(R0));
 
   // DeoptimizeFillFrame expects a Dart frame, i.e. EnterDartFrame(0), but there
   // is no need to set the correct PC marker or load PP, since they get patched.
   __ EnterStubFrame();
 
   if (preserve_result) {
     __ Push(R1);  // Preserve result as first local.
   }
   __ ReserveAlignedFrameSpace(0);
   __ mov(R0, FP);  // Pass last FP as parameter in R0.
   __ CallRuntime(kDeoptimizeFillFrameRuntimeEntry, 1);
   if (preserve_result) {
     // Restore result into R1.
     __ LoadFromOffset(R1, FP, kFirstLocalSlotFromFp * kWordSize);
   }
   // Code above cannot cause GC.
   // There is a Dart Frame on the stack. We must restore PP and leave frame.
+  __ RestoreCodePointer();
   __ LeaveStubFrame();
 
   // Frame is fully rewritten at this point and it is safe to perform a GC.
   // Materialize any objects that were deferred by FillFrame because they
   // require allocation.
   // Enter stub frame with loading PP. The caller's PP is not materialized yet.
   __ EnterStubFrame();
   if (preserve_result) {
     __ Push(R1);  // Preserve result, it will be GC-d here.
   }
   __ Push(ZR);  // Space for the result.
   __ CallRuntime(kDeoptimizeMaterializeRuntimeEntry, 0);
   // Result tells stub how many bytes to remove from the expression stack
   // of the bottom-most frame. They were used as materialization arguments.
   __ Pop(R1);
   __ SmiUntag(R1);
   if (preserve_result) {
     __ Pop(R0);  // Restore result.
   }
   __ LeaveStubFrame();
   // Remove materialization arguments.
   __ add(SP, SP, Operand(R1));
   __ ret();
 }
 
 
 void StubCode::GenerateDeoptimizeLazyStub(Assembler* assembler) {
   // Correct return address to point just after the call that is being
   // deoptimized.
-  __ AddImmediate(LR, LR, -CallPattern::kLengthInBytes);
-  GenerateDeoptimizationSequence(assembler, true);  // Preserve R0.
+  __ AddImmediate(LR, LR, -CallPattern::kDeoptCallLengthInBytes);
+  // Push zap value instead of CODE_REG for lazy deopt.
+  __ LoadImmediate(TMP, 0xf1f1f1f1);
+  __ Push(TMP);
+  GenerateDeoptimizationSequence(assembler, kLazyDeopt);
 }
 
 
 void StubCode::GenerateDeoptimizeStub(Assembler* assembler) {
-  GenerateDeoptimizationSequence(assembler, false);  // Don't preserve R0.
+  GenerateDeoptimizationSequence(assembler, kEagerDeopt);
 }
 
 
 static void GenerateDispatcherCode(Assembler* assembler,
                                    Label* call_target_function) {
   __ Comment("NoSuchMethodDispatch");
   // When lazily generated invocation dispatchers are disabled, the
   // miss-handler may return null.
   __ CompareObject(R0, Object::null_object());
   __ b(call_target_function, NE);
@@ skipping 40 matching lines @@
   __ Push(R4);
   __ CallRuntime(kMegamorphicCacheMissHandlerRuntimeEntry, 3);
   // Remove arguments.
   __ Drop(3);
   __ Pop(R0);  // Get result into R0 (target function).
 
   // Restore IC data and arguments descriptor.
   __ Pop(R4);
   __ Pop(R5);
 
+  __ RestoreCodePointer();
   __ LeaveStubFrame();
 
   if (!FLAG_lazy_dispatchers) {
     Label call_target_function;
     GenerateDispatcherCode(assembler, &call_target_function);
     __ Bind(&call_target_function);
   }
 
   // Tail-call to target function.
+  __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
   __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
   __ br(R2);
 }
 
 
 // Called for inline allocation of arrays.
 // Input parameters:
 //   LR: return address.
 //   R2: array length as Smi.
 //   R1: array element type (either NULL or an instantiated type).
@@ skipping 136 matching lines @@
   __ Pop(R2);
   __ Pop(R0);
   __ LeaveStubFrame();
   __ ret();
 }
 
 
 // Called when invoking Dart code from C++ (VM code).
 // Input parameters:
 //   LR : points to return address.
-//   R0 : entrypoint of the Dart function to call.
+//   R0 : code object of the Dart function to call.
 //   R1 : arguments descriptor array.
 //   R2 : arguments array.
 //   R3 : current thread.
 void StubCode::GenerateInvokeDartCodeStub(Assembler* assembler) {
   __ Comment("InvokeDartCodeStub");
 
   // Copy the C stack pointer (R31) into the stack pointer we'll actually use
   // to access the stack, and put the C stack pointer at the stack limit.
   __ SetupDartSP(Isolate::GetSpecifiedStackSize());
   __ EnterFrame(0);
 
+  // Push code object to PC marker slot.
+  __ ldr(TMP, Address(R3, Thread::invoke_dart_code_stub_offset()));
+  __ Push(TMP);
+
   // Save the callee-saved registers.
   for (int i = kAbiFirstPreservedCpuReg; i <= kAbiLastPreservedCpuReg; i++) {
     const Register r = static_cast<Register>(i);
     // We use str instead of the Push macro because we will be pushing the PP
     // register when it is not holding a pool-pointer since we are coming from
     // C++ code.
     __ str(r, Address(SP, -1 * kWordSize, Address::PreIndex));
   }
 
   // Save the bottom 64-bits of callee-saved V registers.
   for (int i = kAbiFirstPreservedFpuReg; i <= kAbiLastPreservedFpuReg; i++) {
     const VRegister r = static_cast<VRegister>(i);
     __ PushDouble(r);
   }
 
-  // We now load the pool pointer(PP) as we are about to invoke dart code and we
-  // could potentially invoke some intrinsic functions which need the PP to be
-  // set up.
-  __ LoadPoolPointer();
-
   // Set up THR, which caches the current thread in Dart code.
   if (THR != R3) {
     __ mov(THR, R3);
   }
   // Load Isolate pointer into temporary register R5.
   __ LoadIsolate(R5);
 
   // Save the current VMTag on the stack.
   __ LoadFromOffset(R4, R5, Isolate::vm_tag_offset());
   __ Push(R4);
 
   // Mark that the isolate is executing Dart code.
   __ LoadImmediate(R6, VMTag::kDartTagId);
   __ StoreToOffset(R6, R5, Isolate::vm_tag_offset());
 
   // Save top resource and top exit frame info. Use R6 as a temporary register.
   // StackFrameIterator reads the top exit frame info saved in this frame.
   __ LoadFromOffset(R6, THR, Thread::top_resource_offset());
   __ StoreToOffset(ZR, THR, Thread::top_resource_offset());
   __ Push(R6);
   __ LoadFromOffset(R6, THR, Thread::top_exit_frame_info_offset());
   __ StoreToOffset(ZR, THR, Thread::top_exit_frame_info_offset());
   // kExitLinkSlotFromEntryFp must be kept in sync with the code below.
-  ASSERT(kExitLinkSlotFromEntryFp == -21);
+  ASSERT(kExitLinkSlotFromEntryFp == -22);
   __ Push(R6);
 
   // Load arguments descriptor array into R4, which is passed to Dart code.
   __ LoadFromOffset(R4, R1, VMHandles::kOffsetOfRawPtrInHandle);
 
   // Load number of arguments into S5.
   __ LoadFieldFromOffset(R5, R4, ArgumentsDescriptor::count_offset());
   __ SmiUntag(R5);
 
   // Compute address of 'arguments array' data area into R2.
   __ LoadFromOffset(R2, R2, VMHandles::kOffsetOfRawPtrInHandle);
   __ AddImmediate(R2, R2, Array::data_offset() - kHeapObjectTag);
 
   // Set up arguments for the Dart call.
   Label push_arguments;
   Label done_push_arguments;
   __ cmp(R5, Operand(0));
   __ b(&done_push_arguments, EQ);  // check if there are arguments.
   __ LoadImmediate(R1, 0);
   __ Bind(&push_arguments);
   __ ldr(R3, Address(R2));
   __ Push(R3);
   __ add(R1, R1, Operand(1));
   __ add(R2, R2, Operand(kWordSize));
   __ cmp(R1, Operand(R5));
   __ b(&push_arguments, LT);
   __ Bind(&done_push_arguments);
 
+  // We now load the pool pointer(PP) with a GC safe value as we are about to
+  // invoke dart code. We don't need a real object pool here.
+  // Smi zero does not work because ARM64 assumes PP to be untagged.
+  __ LoadObject(PP, Object::null_object());
+
   // Call the Dart code entrypoint.
+  __ ldr(CODE_REG, Address(R0, VMHandles::kOffsetOfRawPtrInHandle));
+  __ ldr(R0, FieldAddress(CODE_REG, Code::entry_point_offset()));
   __ blr(R0);  // R4 is the arguments descriptor array.
   __ Comment("InvokeDartCodeStub return");
 
-  // Restore constant pool pointer after return.
-  __ LoadPoolPointer();
-
   // Get rid of arguments pushed on the stack.
   __ AddImmediate(SP, FP, kExitLinkSlotFromEntryFp * kWordSize);
 
   __ LoadIsolate(R28);
 
   // Restore the saved top exit frame info and top resource back into the
   // Isolate structure. Uses R6 as a temporary register for this.
   __ Pop(R6);
   __ StoreToOffset(R6, THR, Thread::top_exit_frame_info_offset());
   __ Pop(R6);
@@ skipping 11 matching lines @@
 
   // Restore C++ ABI callee-saved registers.
   for (int i = kAbiLastPreservedCpuReg; i >= kAbiFirstPreservedCpuReg; i--) {
     Register r = static_cast<Register>(i);
     // We use ldr instead of the Pop macro because we will be popping the PP
     // register when it is not holding a pool-pointer since we are returning to
     // C++ code. We also skip the dart stack pointer SP, since we are still
     // using it as the stack pointer.
     __ ldr(r, Address(SP, 1 * kWordSize, Address::PostIndex));
   }
-  __ set_constant_pool_allowed(false);
 
   // Restore the frame pointer and C stack pointer and return.
   __ LeaveFrame();
   __ mov(CSP, SP);
   __ ret();
 }
 
 
 // Called for inline allocation of contexts.
 // Input:
@@ skipping 165 matching lines @@
   // Restore callee-saved registers, tear down frame.
   __ LeaveCallRuntimeFrame();
   __ ret();
 }
 
 
 // Called for inline allocation of objects.
 // Input parameters:
 //   LR : return address.
 //   SP + 0 : type arguments object (only if class is parameterized).
-void StubCode::GenerateAllocationStubForClass(
-    Assembler* assembler, const Class& cls,
-    uword* entry_patch_offset, uword* patch_code_pc_offset) {
-  *entry_patch_offset = assembler->CodeSize();
+void StubCode::GenerateAllocationStubForClass(Assembler* assembler,
+                                              const Class& cls) {
   // The generated code is different if the class is parameterized.
   const bool is_cls_parameterized = cls.NumTypeArguments() > 0;
   ASSERT(!is_cls_parameterized ||
          (cls.type_arguments_field_offset() != Class::kNoTypeArguments));
   // kInlineInstanceSize is a constant used as a threshold for determining
   // when the object initialization should be done as a loop or as
   // straight line code.
   const int kInlineInstanceSize = 12;
   const intptr_t instance_size = cls.instance_size();
   ASSERT(instance_size > 0);
@@ skipping 99 matching lines @@
     // Push null type arguments.
     __ PushObject(Object::null_object());
   }
   __ CallRuntime(kAllocateObjectRuntimeEntry, 2);  // Allocate object.
   __ Drop(2);  // Pop arguments.
   __ Pop(R0);  // Pop result (newly allocated object).
   // R0: new object
   // Restore the frame pointer.
   __ LeaveStubFrame();
   __ ret();
-  *patch_code_pc_offset = assembler->CodeSize();
-  __ BranchPatchable(*StubCode::FixAllocationStubTarget_entry());
 }
 
 
 // Called for invoking "dynamic noSuchMethod(Invocation invocation)" function
 // from the entry code of a dart function after an error in passed argument
 // name or number is detected.
 // Input parameters:
 //  LR : return address.
 //  SP : address of last argument.
 //  R4: arguments descriptor array.
@@ skipping 296 matching lines @@
   // Pass IC data object.
   __ Push(R5);
   __ CallRuntime(handle_ic_miss, num_args + 1);
   // Remove the call arguments pushed earlier, including the IC data object.
   __ Drop(num_args + 1);
   // Pop returned function object into R0.
   // Restore arguments descriptor array and IC data array.
   __ Pop(R0);  // Pop returned function object into R0.
   __ Pop(R5);  // Restore IC Data.
   __ Pop(R4);  // Restore arguments descriptor array.
+  if (range_collection_mode == kCollectRanges) {
+    __ RestoreCodePointer();
+  }
   __ LeaveStubFrame();
   Label call_target_function;
   if (!FLAG_lazy_dispatchers) {
     GenerateDispatcherCode(assembler, &call_target_function);
   } else {
     __ b(&call_target_function);
   }
 
   __ Bind(&found);
   __ Comment("Update caller's counter");
   // R6: pointer to an IC data check group.
   const intptr_t target_offset = ICData::TargetIndexFor(num_args) * kWordSize;
   const intptr_t count_offset = ICData::CountIndexFor(num_args) * kWordSize;
   __ LoadFromOffset(R0, R6, target_offset);
 
   if (FLAG_optimization_counter_threshold >= 0) {
     // Update counter.
     __ LoadFromOffset(R1, R6, count_offset);
     __ adds(R1, R1, Operand(Smi::RawValue(1)));
     __ LoadImmediate(R2, Smi::RawValue(Smi::kMaxValue));
     __ csel(R1, R2, R1, VS);  // Overflow.
     __ StoreToOffset(R1, R6, count_offset);
   }
 
   __ Comment("Call target");
   __ Bind(&call_target_function);
   // R0: target function.
-  __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
   if (range_collection_mode == kCollectRanges) {
+    __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
     __ ldr(R1, Address(SP, 0 * kWordSize));
     if (num_args == 2) {
       __ ldr(R3, Address(SP, 1 * kWordSize));
     }
     __ EnterStubFrame();
     __ Push(R5);
     if (num_args == 2) {
       __ Push(R3);
     }
     __ Push(R1);
+    __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
     __ blr(R2);
 
     Label done;
     __ ldr(R5, Address(FP, kFirstLocalSlotFromFp * kWordSize));
     __ UpdateRangeFeedback(R0, 2, R5, R1, R4, &done);
     __ Bind(&done);
     __ LeaveStubFrame();
     __ ret();
   } else {
+    __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
+    __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
     __ br(R2);
   }
 
   if (FLAG_support_debugger && !optimized) {
     __ Bind(&stepping);
     __ EnterStubFrame();
     __ Push(R5);  // Preserve IC data.
     __ CallRuntime(kSingleStepHandlerRuntimeEntry, 0);
     __ Pop(R5);
+    __ RestoreCodePointer();
     __ LeaveStubFrame();
     __ b(&done_stepping);
   }
 }
 
 
 // Use inline cache data array to invoke the target or continue in inline
 // cache miss handler. Stub for 1-argument check (receiver class).
 //  LR: return address.
 //  R5: inline cache data object.
@@ skipping 124 matching lines @@
     __ LoadImmediate(R2, Smi::RawValue(Smi::kMaxValue));
     __ csel(R1, R2, R1, VS);  // Overflow.
     __ StoreToOffset(R1, R6, count_offset);
   }
 
   // Load arguments descriptor into R4.
   __ LoadFieldFromOffset(R4, R5, ICData::arguments_descriptor_offset());
 
   // Get function and call it, if possible.
   __ LoadFromOffset(R0, R6, target_offset);
+  __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
   __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
   __ br(R2);
 
   if (FLAG_support_debugger) {
     __ Bind(&stepping);
     __ EnterStubFrame();
     __ Push(R5);  // Preserve IC data.
     __ CallRuntime(kSingleStepHandlerRuntimeEntry, 0);
     __ Pop(R5);
+    __ RestoreCodePointer();
     __ LeaveStubFrame();
     __ b(&done_stepping);
   }
 }
 
 
 void StubCode::GenerateOneArgUnoptimizedStaticCallStub(Assembler* assembler) {
   GenerateUsageCounterIncrement(assembler, R6);
   GenerateNArgsCheckInlineCacheStub(
       assembler, 1, kStaticCallMissHandlerOneArgRuntimeEntry, Token::kILLEGAL,
@@ skipping 18 matching lines @@
   __ EnterStubFrame();
   __ Push(R5);  // Save IC Data.
   __ Push(R4);  // Save arg. desc.
   __ Push(R0);  // Pass function.
   __ CallRuntime(kCompileFunctionRuntimeEntry, 1);
   __ Pop(R0);  // Restore argument.
   __ Pop(R4);  // Restore arg desc.
   __ Pop(R5);  // Restore IC Data.
   __ LeaveStubFrame();
 
+  __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
   __ LoadFieldFromOffset(R2, R0, Function::entry_point_offset());
   __ br(R2);
 }
 
 
 // R5: Contains an ICData.
 void StubCode::GenerateICCallBreakpointStub(Assembler* assembler) {
   __ EnterStubFrame();
   __ Push(R5);
   __ PushObject(Object::null_object());  // Space for result.
   __ CallRuntime(kBreakpointRuntimeHandlerRuntimeEntry, 0);
-  __ Pop(R0);
+  __ Pop(CODE_REG);
   __ Pop(R5);
   __ LeaveStubFrame();
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ br(R0);
 }
 
 
 void StubCode::GenerateRuntimeCallBreakpointStub(Assembler* assembler) {
   __ EnterStubFrame();
   __ PushObject(Object::null_object());  // Space for result.
   __ CallRuntime(kBreakpointRuntimeHandlerRuntimeEntry, 0);
-  __ Pop(R0);
+  __ Pop(CODE_REG);
   __ LeaveStubFrame();
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ br(R0);
 }
 
 // Called only from unoptimized code. All relevant registers have been saved.
 void StubCode::GenerateDebugStepCheckStub(
     Assembler* assembler) {
   // Check single stepping.
   Label stepping, done_stepping;
   __ LoadIsolate(R1);
   __ LoadFromOffset(
@@ skipping 157 matching lines @@
 // R6: function to be re-optimized.
 // R4: argument descriptor (preserved).
 void StubCode::GenerateOptimizeFunctionStub(Assembler* assembler) {
   __ EnterStubFrame();
   __ Push(R4);
   // Setup space on stack for the return value.
   __ PushObject(Object::null_object());
   __ Push(R6);
   __ CallRuntime(kOptimizeInvokedFunctionRuntimeEntry, 1);
   __ Pop(R0);  // Discard argument.
-  __ Pop(R0);  // Get Code object
+  __ Pop(CODE_REG);  // Get Code object
   __ Pop(R4);  // Restore argument descriptor.
-  __ LoadFieldFromOffset(R0, R0, Code::entry_point_offset());
+  __ LoadFieldFromOffset(R0, CODE_REG, Code::entry_point_offset());
   __ LeaveStubFrame();
   __ br(R0);
   __ brk(0);
 }
 
 
 // Does identical check (object references are equal or not equal) with special
 // checks for boxed numbers.
 // Left and right are pushed on stack.
 // Return Zero condition flag set if equal.
@@ skipping 72 matching lines @@
   const Register right = R0;
   __ LoadFromOffset(left, SP, 1 * kWordSize);
   __ LoadFromOffset(right, SP, 0 * kWordSize);
   GenerateIdenticalWithNumberCheckStub(assembler, left, right);
   __ ret();
 
   if (FLAG_support_debugger) {
     __ Bind(&stepping);
     __ EnterStubFrame();
     __ CallRuntime(kSingleStepHandlerRuntimeEntry, 0);
+    __ RestoreCodePointer();
     __ LeaveStubFrame();
     __ b(&done_stepping);
   }
 }
 
 
 // Called from optimized code only.
 // LR: return address.
 // SP + 4: left operand.
 // SP + 0: right operand.
@@ skipping 38 matching lines @@
   __ CompareRegisters(R4, R0);
   __ b(&update, NE);
 
   __ Bind(&call_target_function);
   // Call the target found in the cache.  For a class id match, this is a
   // proper target for the given name and arguments descriptor.  If the
   // illegal class id was found, the target is a cache miss handler that can
   // be invoked as a normal Dart function.
   __ add(TMP, R2, Operand(R3, LSL, 3));
   __ LoadFieldFromOffset(R0, TMP, base + kWordSize);
+  __ LoadFieldFromOffset(CODE_REG, R0, Function::code_offset());
   __ LoadFieldFromOffset(R1, R0, Function::entry_point_offset());
 }
 
 
 // Called from megamorphic calls.
 //  R0: receiver.
 //  R1: lookup cache.
 // Result:
 //  R1: entry point.
 void StubCode::GenerateMegamorphicLookupStub(Assembler* assembler) {
   EmitMegamorphicLookup(assembler, R0, R1, R1);
   __ ret();
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_ARM64