Initial SIMDBC interpreter.

runtime/vm/simulator_dbc.cc

+// Copyright (c) 2016, 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 <setjmp.h>  // NOLINT
+#include <stdlib.h>
+
+#include "vm/globals.h"
+#if defined(TARGET_ARCH_DBC)
+
+#if !defined(USING_SIMULATOR)
+#error "DBC is a simulated architecture"
+#endif
+
+#include "vm/simulator.h"

[zra, 2016/04/14 18:27:49]

You can probably safely add "vm/simulator_dbc.h" here to silence a linter error
if there is one.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

No other file does this.

+
+#include "vm/assembler.h"
+#include "vm/compiler.h"
+#include "vm/constants_dbc.h"
+#include "vm/cpu.h"
+#include "vm/dart_entry.h"
+#include "vm/debugger.h"
+#include "vm/disassembler.h"
+#include "vm/lockers.h"
+#include "vm/native_arguments.h"
+#include "vm/native_entry.h"
+#include "vm/object.h"
+#include "vm/object_store.h"
+#include "vm/os_thread.h"
+#include "vm/stack_frame.h"
+
+namespace dart {
+
+DEFINE_FLAG(uint64_t, trace_sim_after, ULLONG_MAX,
+            "Trace simulator execution after instruction count reached.");
+DEFINE_FLAG(uint64_t, stop_sim_at, ULLONG_MAX,
+            "Instruction address or instruction count to stop simulator at.");
+
+// SimulatorSetjmpBuffer are linked together, and the last created one
+// is referenced by the Simulator. When an exception is thrown, the exception
+// runtime looks at where to jump and finds the corresponding
+// SimulatorSetjmpBuffer based on the stack pointer of the exception handler.
+// The runtime then does a Longjmp on that buffer to return to the simulator.
+class SimulatorSetjmpBuffer {
+ public:
+  void Longjmp() {
+    // "This" is now the last setjmp buffer.
+    simulator_->set_last_setjmp_buffer(this);
+    longjmp(buffer_, 1);
+  }
+
+  explicit SimulatorSetjmpBuffer(Simulator* sim) {
+    simulator_ = sim;
+    link_ = sim->last_setjmp_buffer();
+    sim->set_last_setjmp_buffer(this);
+    top_ = sim->top_;
+    base_ = sim->base_;
+  }
+
+  ~SimulatorSetjmpBuffer() {
+    ASSERT(simulator_->last_setjmp_buffer() == this);
+    simulator_->set_last_setjmp_buffer(link_);
+  }
+
+  SimulatorSetjmpBuffer* link() { return link_; }

[zra, 2016/04/14 18:27:49]

const?

[Vyacheslav Egorov (Google), 2016/04/18 19:19:56]

Done.

+
+  uintptr_t* top() const { return top_; }
+  uword base() const { return reinterpret_cast<uword>(base_); }
+
+  jmp_buf buffer_;

[zra, 2016/04/14 18:27:50]

This is private on other architectures. There, the class has a public method
Setjmp that forwards to setjmp(). Why not do the same thing here?

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

You can't have a method that forwards to setjmp, because "The stack context will
be invalidated if the function which called setjmp() returns.".

The code like this:

bool SimulatorSetjmpBuffer::Setjmp() { return setjmp(&buffer_); } 

if (buffer->Setjmp()) {
  // ...
} else {
  // ...
}

does NOT work. On other architectures Setjmp is present but is actually never
used. Instead we access jmp_buf directly because Simulator is a friend of this
class. 

[zra, 2016/04/18 16:50:14]

Please remove or add TODO to remove in the other simulators.

[Vyacheslav Egorov (Google), 2016/04/18 19:19:56]

Done.

+
+ private:
+  uintptr_t* top_;
+  uintptr_t* base_;
+  Simulator* simulator_;
+  SimulatorSetjmpBuffer* link_;
+
+  friend class Simulator;

[zra, 2016/04/14 18:27:50]

DISALLOW_ALLOCATION()? DISALLOW_COPY_AND_ASSIGN()?

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+};
+
+
+DART_FORCE_INLINE static RawObject** SavedCallerFP(RawObject** FP) {
+  return reinterpret_cast<RawObject**>(FP[kSavedCallerFpSlotFromFp]);
+}
+
+
+DART_FORCE_INLINE static RawCode* FrameCode(RawObject** FP) {
+  return static_cast<RawCode*>(FP[kPcMarkerSlotFromFp]);
+}
+
+
+DART_FORCE_INLINE static void SetFrameCode(RawObject** FP, RawCode* code) {
+  FP[kPcMarkerSlotFromFp] = code;
+}
+
+
+DART_FORCE_INLINE static RawObject** FrameArguments(RawObject** FP,
+                                                    intptr_t argc) {
+  return FP - (kDartFrameFixedSize + argc);
+}
+
+
+class SimulatorHelpers {
+ public:
+  DART_FORCE_INLINE static RawSmi* GetClassIdAsSmi(RawObject* obj) {
+    return Smi::New(obj->IsHeapObject() ? obj->GetClassId() : kSmiCid);
+  }
+
+  DART_FORCE_INLINE static intptr_t GetClassId(RawObject* obj) {
+    return obj->IsHeapObject() ? obj->GetClassId() : kSmiCid;
+  }
+
+  DART_FORCE_INLINE static void IncrementUsageCounter(RawICData* icdata) {
+    reinterpret_cast<RawFunction*>(icdata->ptr()->owner_)
+        ->ptr()
+        ->usage_counter_++;
+  }
+
+  DART_FORCE_INLINE static bool IsStrictEqualWithNumberCheck(RawObject* lhs,
+                                                             RawObject* rhs) {
+    if (lhs == rhs) {
+      return true;
+    }
+
+    if (lhs->IsHeapObject() && rhs->IsHeapObject()) {
+      const intptr_t lhs_cid = lhs->GetClassId();
+      const intptr_t rhs_cid = rhs->GetClassId();
+      if (lhs_cid == rhs_cid) {
+        switch (lhs_cid) {
+          case kDoubleCid:
+            return (bit_cast<uint64_t, double>(
+                        static_cast<RawDouble*>(lhs)->ptr()->value_) ==
+                    bit_cast<uint64_t, double>(
+                        static_cast<RawDouble*>(rhs)->ptr()->value_));
+
+          case kMintCid:
+            return (static_cast<RawMint*>(lhs)->ptr()->value_ ==
+                    static_cast<RawMint*>(rhs)->ptr()->value_);
+
+          case kBigintCid:
+            return (DLRT_BigintCompare(static_cast<RawBigint*>(lhs),
+                                       static_cast<RawBigint*>(rhs)) == 0);
+        }
+      }
+    }
+
+    return false;
+  }
+
+  template <typename T>
+  DART_FORCE_INLINE static T* Untag(T* tagged) {
+    return tagged->ptr();
+  }
+
+  static bool ObjectArraySetIndexed(Thread* thread,
+                                    RawObject** B,
+                                    RawObject** result) {
+    if (thread->isolate()->type_checks()) {
+      return false;
+    }
+
+    RawObject** args = FrameArguments(B, 3);
+    RawSmi* index = static_cast<RawSmi*>(args[1]);
+    RawArray* array = static_cast<RawArray*>(args[0]);
+    if (!index->IsHeapObject() && reinterpret_cast<intptr_t>(index) >= 0 &&

[zra, 2016/04/14 18:27:50]

parens around comparisons.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+        reinterpret_cast<intptr_t>(index) <
+            reinterpret_cast<intptr_t>(array->ptr()->length_)) {
+      array->StorePointer(array->ptr()->data() + Smi::Value(index), args[2]);
+      return true;
+    }
+    return false;
+  }
+
+  static bool ObjectArrayGetIndexed(Thread* thread,
+                                    RawObject** B,
+                                    RawObject** result) {
+    RawObject** args = FrameArguments(B, 2);
+    RawSmi* index = static_cast<RawSmi*>(args[1]);
+    RawArray* array = static_cast<RawArray*>(args[0]);
+    if (!index->IsHeapObject() && reinterpret_cast<intptr_t>(index) >= 0 &&

[zra, 2016/04/14 18:27:49]

ditto

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+        reinterpret_cast<intptr_t>(index) <
+            reinterpret_cast<intptr_t>(array->ptr()->length_)) {
+      *result = array->ptr()->data()[Smi::Value(index)];
+      return true;
+    }
+    return false;
+  }
+
+  static bool GrowableArraySetIndexed(Thread* thread,
+                                      RawObject** B,
+                                      RawObject** result) {
+    if (thread->isolate()->type_checks()) {
+      return false;
+    }
+
+    RawObject** args = FrameArguments(B, 3);
+    RawSmi* index = static_cast<RawSmi*>(args[1]);
+    RawGrowableObjectArray* array =
+        static_cast<RawGrowableObjectArray*>(args[0]);
+    if (!index->IsHeapObject() && reinterpret_cast<intptr_t>(index) >= 0 &&

[zra, 2016/04/14 18:27:49]

ditto. A small cleanup might be to pull this check out into a function. It's a
bit messy and repeated 3 or 4 times.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+        reinterpret_cast<intptr_t>(index) <
+            reinterpret_cast<intptr_t>(array->ptr()->length_)) {
+      RawArray* data = array->ptr()->data_;
+      data->StorePointer(data->ptr()->data() + Smi::Value(index), args[2]);
+      return true;
+    }
+    return false;
+  }
+
+  static bool GrowableArrayGetIndexed(Thread* thread,
+                                      RawObject** B,
+                                      RawObject** result) {
+    RawObject** args = FrameArguments(B, 2);
+    RawSmi* index = static_cast<RawSmi*>(args[1]);
+    RawGrowableObjectArray* array =
+        static_cast<RawGrowableObjectArray*>(args[0]);
+    if (!index->IsHeapObject() && reinterpret_cast<intptr_t>(index) >= 0 &&

[zra, 2016/04/14 18:27:49]

ditto

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+        reinterpret_cast<intptr_t>(index) <
+            reinterpret_cast<intptr_t>(array->ptr()->length_)) {
+      *result = array->ptr()->data_->ptr()->data()[Smi::Value(index)];
+      return true;
+    }
+    return false;
+  }
+};
+
+
+DART_FORCE_INLINE static uint32_t* SavedCallerPC(RawObject** FP) {
+  return reinterpret_cast<uint32_t*>(FP[kSavedCallerPcSlotFromFp]);
+}
+
+
+DART_FORCE_INLINE static RawFunction* FrameFunction(RawObject** FP) {
+  RawFunction* function = static_cast<RawFunction*>(FP[kFunctionSlotFromFp]);
+  ASSERT(SimulatorHelpers::GetClassId(function) == kFunctionCid);
+  return function;
+}
+
+
+IntrinsicHandler Simulator::intrinsics_[Simulator::kIntrinsicCount];
+
+
+// Synchronization primitives support.
+void Simulator::InitOnce() {
+  for (intptr_t i = 0; i < kIntrinsicCount; i++) {
+    intrinsics_[i] = 0;
+  }
+
+  intrinsics_[kObjectArraySetIndexedIntrinsic] =
+      SimulatorHelpers::ObjectArraySetIndexed;
+  intrinsics_[kObjectArrayGetIndexedIntrinsic] =
+      SimulatorHelpers::ObjectArrayGetIndexed;
+  intrinsics_[kGrowableArraySetIndexedIntrinsic] =
+      SimulatorHelpers::GrowableArraySetIndexed;
+  intrinsics_[kGrowableArrayGetIndexedIntrinsic] =
+      SimulatorHelpers::GrowableArrayGetIndexed;
+}
+
+
+Simulator::Simulator() {
+  // Setup simulator support first. Some of this information is needed to
+  // setup the architecture state.
+  // We allocate the stack here, the size is computed as the sum of
+  // the size specified by the user and the buffer space needed for
+  // handling stack overflow exceptions. To be safe in potential
+  // stack underflows we also add some underflow buffer space.
+  stack_ = new uintptr_t[(OSThread::GetSpecifiedStackSize() +
+                          OSThread::kStackSizeBuffer +
+                          kSimulatorStackUnderflowSize) /
+                         sizeof(uintptr_t)];
+  base_ = top_ = stack_;
+  last_setjmp_buffer_ = NULL;
+  top_exit_frame_info_ = 0;
+}
+
+
+Simulator::~Simulator() {
+  delete[] stack_;
+  Isolate* isolate = Isolate::Current();
+  if (isolate != NULL) {
+    isolate->set_simulator(NULL);
+  }
+}
+
+
+// Get the active Simulator for the current isolate.
+Simulator* Simulator::Current() {
+  Simulator* simulator = Isolate::Current()->simulator();
+  if (simulator == NULL) {
+    simulator = new Simulator();
+    Isolate::Current()->set_simulator(simulator);
+  }
+  return simulator;
+}
+
+
+// Returns the top of the stack area to enable checking for stack pointer
+// validity.
+uword Simulator::StackTop() const {
+  // To be safe in potential stack underflows we leave some buffer above and
+  // set the stack top.
+  return StackBase() +
+         (OSThread::GetSpecifiedStackSize() + OSThread::kStackSizeBuffer);
+}
+
+
+// Calls into the Dart runtime are based on this interface.
+typedef void (*SimulatorRuntimeCall)(NativeArguments arguments);
+
+// Calls to leaf Dart runtime functions are based on this interface.
+typedef int32_t (*SimulatorLeafRuntimeCall)(int32_t r0,
+                                            int32_t r1,
+                                            int32_t r2,
+                                            int32_t r3);
+
+// Calls to leaf float Dart runtime functions are based on this interface.
+typedef double (*SimulatorLeafFloatRuntimeCall)(double d0, double d1);
+
+// Calls to native Dart functions are based on this interface.
+typedef void (*SimulatorBootstrapNativeCall)(NativeArguments* arguments);
+typedef void (*SimulatorNativeCall)(NativeArguments* arguments, uword target);
+
+
+void Simulator::Exit(Thread* thread,
+                     RawObject** base,
+                     RawObject** frame,
+                     uint32_t* pc) {
+  frame[0] = Function::null();
+  frame[1] = Code::null();
+  frame[2] = reinterpret_cast<RawObject*>(pc);
+  frame[3] = reinterpret_cast<RawObject*>(base);
+  base_ = top_ = reinterpret_cast<uintptr_t*>(frame + kDartFrameFixedSize);
+  thread->set_top_exit_frame_info(reinterpret_cast<uword>(top_));
+}
+
+
+#if defined(__has_builtin)
+#if __has_builtin(__builtin_smul_overflow)
+#define HAS_MUL_OVERFLOW
+#endif
+#if __has_builtin(__builtin_sadd_overflow)
+#define HAS_ADD_OVERFLOW
+#endif
+#if __has_builtin(__builtin_ssub_overflow)
+#define HAS_SUB_OVERFLOW
+#endif
+#endif
+
+
+DART_FORCE_INLINE static int32_t SignedAddWithOverflow(int32_t lhs,
+                                                       int32_t rhs,
+                                                       intptr_t* out) {
+  int32_t res = 1;
+#if defined(HAS_ADD_OVERFLOW)
+  res = __builtin_sadd_overflow(lhs, rhs, out);
+#elif defined(__i386__)
+  asm volatile(

[zra, 2016/04/14 18:27:49]

How do you plan to ensure that the inline assembly here and below has test
coverage?

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

I welcome suggestions. Ultimately it's not that different from all other
handwritten assembly code. We can move this function to a header and add a test.

+      "add %2, %1\n"
+      "jo 1f;\n"
+      "xor %0, %0\n"
+      "mov %1, 0(%3)\n"
+      "1: "
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc");
+#elif defined(__arm__)
+  asm volatile(
+      "adds %1, %1, %2;\n"
+      "bvs 1f;\n"
+      "mov %0, $0;\n"
+      "str %1, [%3, #0]\n"
+      "1:"
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc", "r12");
+#else
+#error "Unsupported platform"
+#endif
+  return res;
+}
+
+
+DART_FORCE_INLINE static int32_t SignedSubWithOverflow(int32_t lhs,
+                                                       int32_t rhs,
+                                                       intptr_t* out) {
+  int32_t res = 1;
+#if defined(HAS_SUB_OVERFLOW)
+  res = __builtin_ssub_overflow(lhs, rhs, out);
+#elif defined(__i386__)
+  asm volatile(
+      "sub %2, %1\n"
+      "jo 1f;\n"
+      "xor %0, %0\n"
+      "mov %1, 0(%3)\n"
+      "1: "
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc");
+#elif defined(__arm__)
+  asm volatile(
+      "subs %1, %1, %2;\n"
+      "bvs 1f;\n"
+      "mov %0, $0;\n"
+      "str %1, [%3, #0]\n"
+      "1:"
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc", "r12");
+#else
+#error "Unsupported platform"
+#endif
+  return res;
+}
+
+
+DART_FORCE_INLINE static int32_t SignedMulWithOverflow(int32_t lhs,
+                                                       int32_t rhs,
+                                                       intptr_t* out) {
+  int32_t res = 1;
+#if defined(HAS_MUL_OVERFLOW)
+  res = __builtin_smul_overflow(lhs, rhs, out);
+#elif defined(__i386__)
+  asm volatile(
+      "imul %2, %1\n"
+      "jo 1f;\n"
+      "xor %0, %0\n"
+      "mov %1, 0(%3)\n"
+      "1: "
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc");
+#elif defined(__arm__)
+  asm volatile(
+      "smull %1, ip, %1, %2;\n"
+      "cmp ip, %1, ASR #31;\n"
+      "bne 1f;\n"
+      "mov %0, $0;\n"
+      "str %1, [%3, #0]\n"
+      "1:"
+      : "+r"(res), "+r"(lhs)
+      : "r"(rhs), "r"(out)
+      : "cc", "r12");
+#else
+#error "Unsupported platform"
+#endif
+  return res;
+}
+
+
+#define LIKELY(cond) __builtin_expect((cond), 1)

[zra, 2016/04/14 18:27:49]

How big is the performance gain from this hint?

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Was around 5%-8% when measured originally. Might need re-measurement.

+
+
+DART_FORCE_INLINE static bool AreBothSmis(intptr_t a, intptr_t b) {
+  return ((a | b) & kHeapObjectTag) == 0;
+}
+
+
+#define SMI_MUL(lhs, rhs, pres) SignedMulWithOverflow((lhs), (rhs) >> 1, pres)
+#define SMI_COND(cond, lhs, rhs, pres) \
+  ((*(pres) = ((lhs cond rhs) ? true_value : false_value)), false)
+#define SMI_EQ(lhs, rhs, pres) SMI_COND(==, lhs, rhs, pres)
+#define SMI_LT(lhs, rhs, pres) SMI_COND(<, lhs, rhs, pres)
+#define SMI_GT(lhs, rhs, pres) SMI_COND(>, lhs, rhs, pres)
+#define SMI_BITOR(lhs, rhs, pres) ((*(pres) = (lhs | rhs)), false)
+#define SMI_BITAND(lhs, rhs, pres) ((*(pres) = (lhs & rhs)), false)
+
+
+void Simulator::CallRuntime(Thread* thread,
+                            RawObject** base,
+                            RawObject** exit_frame,
+                            uint32_t* pc,
+                            intptr_t argc_tag,
+                            RawObject** args,
+                            RawObject** result,
+                            uword target) {
+  Exit(thread, base, exit_frame, pc);
+  NativeArguments native_args(thread, argc_tag, args, result);
+  reinterpret_cast<RuntimeFunction>(target)(native_args);
+}
+
+
+DART_FORCE_INLINE void Simulator::Invoke(Thread* thread,
+                                         RawObject** call_base,
+                                         RawObject** call_top,
+                                         RawObjectPool** pp,
+                                         uint32_t** pc,
+                                         RawObject*** B,
+                                         RawObject*** SP) {
+  RawObject** callee_fp = call_top + kDartFrameFixedSize;
+
+  RawFunction* function = FrameFunction(callee_fp);
+  RawCode* code = function->ptr()->code_;
+  callee_fp[kPcMarkerSlotFromFp] = code;
+  callee_fp[kSavedCallerPcSlotFromFp] = reinterpret_cast<RawObject*>(*pc);
+  callee_fp[kSavedCallerFpSlotFromFp] = reinterpret_cast<RawObject*>(*B);
+  *pp = code->ptr()->object_pool_->ptr();
+  *pc = reinterpret_cast<uint32_t*>(code->ptr()->entry_point_);
+  *B = callee_fp;
+  *SP = *B - 1;
+}
+
+
+void Simulator::InlineCacheMiss(int checked_args,
+                                Thread* thread,
+                                RawICData* icdata,
+                                RawObject** call_base,
+                                RawObject** top,
+                                uint32_t* pc,
+                                RawObject** B,
+                                RawObject** SP) {
+  RawObject** args = call_base;
+  for (intptr_t i = 0; i < checked_args; i++) {
+    top[1 + i] = args[i];
+  }
+  top[1 + checked_args] = icdata;
+  RuntimeFunction handler = NULL;
+  switch (checked_args) {
+    case 1:
+      handler = DRT_InlineCacheMissHandlerOneArg;
+      break;
+    case 2:
+      handler = DRT_InlineCacheMissHandlerTwoArgs;
+      break;
+    case 3:
+      handler = DRT_InlineCacheMissHandlerThreeArgs;
+      break;
+    default:
+      UNREACHABLE();
+      break;
+  }
+  CallRuntime(thread, B, top + checked_args + 1 + 1, pc, checked_args + 1,

[zra, 2016/04/14 18:27:49]

It might be helpful to add some comments explaining these calculations.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+              top + 1, top, reinterpret_cast<uword>(handler));
+}
+
+
+DART_FORCE_INLINE void Simulator::InstanceCall1(Thread* thread,
+                                                RawICData* icdata,
+                                                RawObject** call_base,
+                                                RawObject** top,
+                                                Array* argdesc,
+                                                RawObjectPool** pp,
+                                                uint32_t** pc,
+                                                RawObject*** B,
+                                                RawObject*** SP) {
+  ASSERT(icdata->GetClassId() == kICDataCid);
+  SimulatorHelpers::IncrementUsageCounter(icdata);
+
+  const intptr_t kCheckedArgs = 1;
+  RawObject** args = call_base;
+  RawArray* cache = icdata->ptr()->ic_data_->ptr();
+
+  RawSmi* receiver_cid = SimulatorHelpers::GetClassIdAsSmi(args[0]);
+
+  bool found = false;
+  for (intptr_t i = 0, length = Smi::Value(cache->length_);

[zra, 2016/04/14 18:27:50]

This might be easier to read if length is initialized before the for loop.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+       i < (length - (kCheckedArgs + 2)); i += (kCheckedArgs + 2)) {
+    if (cache->data()[i + 0] == receiver_cid) {
+      top[0] = cache->data()[i + kCheckedArgs];
+      found = true;
+      break;
+    }
+  }
+
+  if (!found) {
+    InlineCacheMiss(kCheckedArgs, thread, icdata, call_base, top, *pc, *B, *SP);
+  }
+
+  *argdesc = icdata->ptr()->args_descriptor_;
+  Invoke(thread, call_base, top, pp, pc, B, SP);
+}
+
+
+DART_FORCE_INLINE void Simulator::InstanceCall2(Thread* thread,
+                                                RawICData* icdata,
+                                                RawObject** call_base,
+                                                RawObject** top,
+                                                Array* argdesc,
+                                                RawObjectPool** pp,
+                                                uint32_t** pc,
+                                                RawObject*** B,
+                                                RawObject*** SP) {
+  ASSERT(icdata->GetClassId() == kICDataCid);
+  SimulatorHelpers::IncrementUsageCounter(icdata);
+
+  const intptr_t kCheckedArgs = 2;
+  RawObject** args = call_base;
+  RawArray* cache = icdata->ptr()->ic_data_->ptr();
+
+  RawSmi* receiver_cid = SimulatorHelpers::GetClassIdAsSmi(args[0]);
+  RawSmi* arg0_cid = SimulatorHelpers::GetClassIdAsSmi(args[1]);
+
+  bool found = false;
+  for (intptr_t i = 0, length = Smi::Value(cache->length_);

[zra, 2016/04/14 18:27:49]

ditto

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+       i < (length - (kCheckedArgs + 2)); i += (kCheckedArgs + 2)) {
+    if (cache->data()[i + 0] == receiver_cid &&

[zra, 2016/04/14 18:27:50]

parens around equality tests

[Vyacheslav Egorov (Google), 2016/04/18 15:56:42]

Done.

+        cache->data()[i + 1] == arg0_cid) {
+      top[0] = cache->data()[i + kCheckedArgs];
+      found = true;
+      break;
+    }
+  }
+
+  if (!found) {
+    InlineCacheMiss(kCheckedArgs, thread, icdata, call_base, top, *pc, *B, *SP);
+  }
+
+  *argdesc = icdata->ptr()->args_descriptor_;
+  Invoke(thread, call_base, top, pp, pc, B, SP);
+}
+
+
+DART_FORCE_INLINE void Simulator::InstanceCall3(Thread* thread,
+                                                RawICData* icdata,
+                                                RawObject** call_base,
+                                                RawObject** top,
+                                                Array* argdesc,
+                                                RawObjectPool** pp,
+                                                uint32_t** pc,
+                                                RawObject*** B,
+                                                RawObject*** SP) {
+  ASSERT(icdata->GetClassId() == kICDataCid);
+  SimulatorHelpers::IncrementUsageCounter(icdata);
+
+  const intptr_t kCheckedArgs = 3;
+  RawObject** args = call_base;
+  RawArray* cache = icdata->ptr()->ic_data_->ptr();
+
+  RawSmi* receiver_cid = SimulatorHelpers::GetClassIdAsSmi(args[0]);
+  RawSmi* arg0_cid = SimulatorHelpers::GetClassIdAsSmi(args[1]);
+  RawSmi* arg1_cid = SimulatorHelpers::GetClassIdAsSmi(args[2]);
+
+  bool found = false;
+  for (intptr_t i = 0, length = Smi::Value(cache->length_);

[zra, 2016/04/14 18:27:49]

ditto.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Done.

+       i < (length - (kCheckedArgs + 2)); i += (kCheckedArgs + 2)) {
+    if (cache->data()[i + 0] == receiver_cid &&

[zra, 2016/04/14 18:27:49]

parens around equality tests and maybe one per line.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:42]

Done.

+        cache->data()[i + 1] == arg0_cid && cache->data()[i + 2] == arg1_cid) {
+      top[0] = cache->data()[i + kCheckedArgs];
+      found = true;
+      break;
+    }
+  }
+
+  if (!found) {
+    InlineCacheMiss(kCheckedArgs, thread, icdata, call_base, top, *pc, *B, *SP);
+  }
+
+  *argdesc = icdata->ptr()->args_descriptor_;
+  Invoke(thread, call_base, top, pp, pc, B, SP);
+}
+
+
+static __attribute__((noinline)) bool InvokeRuntime(

[zra, 2016/04/14 18:27:50]

Does inlining these break setjmp/longjmp? If so, it seems like we have problems
elsewhere as well.

It looks like we've been putting function attributes in platform/globals.h. e.g.
DART_NOINLINE. MSVC should respect __declspec(noinline).

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

Not inlinling these helps to recover a bit of performance on ARM.

+    Thread* thread,
+    Simulator* sim,
+    RuntimeFunction drt,
+    const NativeArguments& args) {
+  SimulatorSetjmpBuffer buffer(sim);
+  if (!setjmp(buffer.buffer_)) {
+    thread->set_vm_tag(reinterpret_cast<uword>(drt));
+    drt(args);
+    thread->set_vm_tag(VMTag::kDartTagId);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+
+static __attribute__((noinline)) bool InvokeNative(
+    Thread* thread,
+    Simulator* sim,
+    SimulatorBootstrapNativeCall f,
+    NativeArguments* args) {
+  SimulatorSetjmpBuffer buffer(sim);
+  if (!setjmp(buffer.buffer_)) {
+    thread->set_vm_tag(reinterpret_cast<uword>(f));
+    f(args);
+    thread->set_vm_tag(VMTag::kDartTagId);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+
+static __attribute__((noinline)) bool InvokeNativeWrapper(
+    Thread* thread,
+    Simulator* sim,
+    Dart_NativeFunction f,
+    NativeArguments* args) {
+  SimulatorSetjmpBuffer buffer(sim);
+  if (!setjmp(buffer.buffer_)) {
+    thread->set_vm_tag(reinterpret_cast<uword>(f));
+    NativeEntry::NativeCallWrapper(reinterpret_cast<Dart_NativeArguments>(args),
+                                   f);
+    thread->set_vm_tag(VMTag::kDartTagId);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+
+RawObject* Simulator::Call(const Code& code,

[zra, 2016/04/14 18:27:49]

Slava, I can't read this code. I'm don't want to stop this from going in, but
this needs to be rewritten so that a normal human can read it.

[Vyacheslav Egorov (Google), 2016/04/18 15:56:43]

I added more comments. Please take a look again. 

I would certainly like this function to be readable by other people - but you
should approach it with an idea that its a function written in assembly - not
C++.

[zra, 2016/04/18 16:52:34]

Much easier now. Thanks!

+                           const Array& arguments_descriptor,
+                           const Array& arguments,
+                           Thread* thread) {
+  static const void* dispatch[] = {
+#define TARGET(name, fmt, fmta, fmtb, fmtc) &&bc##name,
+      BYTECODES_LIST(TARGET)
+#undef TARGET
+  };
+
+  const uword vm_tag = thread->vm_tag();
+  thread->set_vm_tag(VMTag::kDartTagId);
+
+  StackResource* top_resource = thread->top_resource();
+  thread->set_top_resource(NULL);
+
+  if (top_ == NULL) {
+    top_ = base_ = stack_;
+  }
+
+  top_[0] = thread->top_exit_frame_info();
+
+  // Unpack arguments.
+  {
+    const intptr_t argc = arguments.Length();
+    RawObject** call_base = reinterpret_cast<RawObject**>(top_ + 1);
+    RawObject** call_top = call_base + argc;
+    for (intptr_t i = 0; i < argc; i++) {
+      call_base[i] = arguments.At(i);
+    }
+
+    call_top[0] = code.function();
+    call_top[1] = code.raw();
+    call_top[2] =
+        reinterpret_cast<RawObject*>((argc << 2) | 2);  // Return address.
+    call_top[3] = reinterpret_cast<RawObject*>(top_);   // Caller FP
+  }
+
+  RawObject** B = reinterpret_cast<RawObject**>(top_ + arguments.Length() +
+                                                kDartFrameFixedSize + 1);
+  RawObject** SP = B - 1;
+  uint32_t* pc = reinterpret_cast<uint32_t*>(code.raw()->ptr()->entry_point_);
+  RawObjectPool* pp = code.object_pool();
+
+  pp = pp->ptr();
+  RawBool* true_value = Bool::True().raw();
+  RawBool* false_value = Bool::False().raw();
+  RawObject* null_value = Object::null();
+  RawObject* empty_context = thread->isolate()->object_store()->empty_context();
+
+  Array& argdesc = Array::Handle();
+#if defined(DEBUG)
+  Function& function_h = Function::Handle();
+#endif
+  argdesc = arguments_descriptor.raw();
+
+#define DECLARE_A_B_C \
+  uint16_t rB, rC;    \
+  USE(rB);            \
+  USE(rC)
+#define DECODE_A_B_C(type, shiftB, shiftC, mask) \
+  rB = ((__op >> shiftB) & mask) /* | rBx */;    \
+  rC = ((__op >> shiftC) & mask) /* | rCx */;
+
+#define DECLARE_0
+#define DECODE_0(type, shiftB, shiftC, mask)
+
+#define DECLARE_A
+#define DECODE_A(type, shiftB, shiftC, mask)
+
+#define DECLARE_A_D \
+  uint32_t rD;      \
+  USE(rD)
+#define DECODE_A_D(type, shiftB, shiftC, mask) rD = (__op >> shiftB);
+
+#define DECLARE___D \
+  uint32_t rD;      \
+  USE(rD)
+#define DECODE___D(type, shiftB, shiftC, mask) rD = (__op >> shiftB);
+
+#define DECLARE_A_X \
+  int32_t rD;       \
+  USE(rD)
+#define DECODE_A_X(type, shiftB, shiftC, mask) \
+  rD = (static_cast<int32_t>(__op) >> shiftB);
+
+#define BYTECODE_HEADER(Name, Declare, Decode) \
+  Declare;                                     \
+  bc##Name : Decode(uint8_t, 16, 24, 0xFF);    \
+  goto bc##Name##Entry;                        \
+  bc##Name##Entry:
+
+#define BYTECODE(Name, type) \
+  BYTECODE_HEADER(Name, DECLARE_##type, DECODE_##type)
+
+#define DISPATCH_OP(val)         \
+  do {                           \
+    __op = (val);                \
+    rA = ((__op >> 8) & 0xFF);   \
+    goto* dispatch[__op & 0xFF]; \
+  } while (0)
+
+#define DISPATCH() DISPATCH_OP(*pc++)
+
+#define SMI_FASTPATH_TOS(ResultT, F)                           \
+  {                                                            \
+    const intptr_t lhs = reinterpret_cast<intptr_t>(SP[-1]);   \
+    const intptr_t rhs = reinterpret_cast<intptr_t>(SP[-0]);   \
+    ResultT* slot = reinterpret_cast<ResultT*>(SP - 1);        \
+    if (LIKELY(AreBothSmis(lhs, rhs) && !F(lhs, rhs, slot))) { \
+      pc += 1;                                                 \
+      --SP;                                                    \
+    }                                                          \
+    DISPATCH();                                                \
+  }
+
+#define FALLTHROUGH
+
+#define HANDLE_EXCEPTION(name, ON_EXCEPTION)                                   \
+  do {                                                                         \
+    B = reinterpret_cast<RawObject**>(base_);                                  \
+    pc = reinterpret_cast<uint32_t*>(pc_);                                     \
+    if ((reinterpret_cast<uword>(pc) & 2) != 0) {                              \
+      base_ = top_ = reinterpret_cast<uintptr_t*>(base_[0]);                   \
+      thread->set_top_exit_frame_info(reinterpret_cast<uword>(top_));          \
+      thread->set_top_resource(top_resource);                                  \
+      thread->set_vm_tag(vm_tag);                                              \
+      return special_[0];                                                      \
+    }                                                                          \
+    pp = FrameCode(B)->ptr()->object_pool_->ptr();                             \
+    ON_EXCEPTION;                                                              \
+  } while (0)
+
+#define INVOKE_RUNTIME(func, args, onexception)   \
+  if (!InvokeRuntime(thread, this, func, args)) { \
+    HANDLE_EXCEPTION(InvokeRuntime, onexception); \
+  }
+
+#define INVOKE_NATIVE(func, args, onexception)     \
+  if (!InvokeNative(thread, this, func, &args)) {  \
+    HANDLE_EXCEPTION(InvokeRuntime, onexception);  \
+  }
+
+#define INVOKE_NATIVE_WRAPPER(func, args, onexception)   \
+  if (!InvokeNativeWrapper(thread, this, func, &args)) { \
+    HANDLE_EXCEPTION(InvokeRuntime, onexception);        \
+  }
+
+  uint16_t rA;
+  uint32_t __op;
+  DISPATCH();
+  while (true) {
+    {
+      {
+      ClosureNoSuchMethod:
+#if defined(DEBUG)
+        function_h ^= FrameFunction(B);
+        ASSERT(function_h.IsClosureFunction());
+#endif
+        pc = SavedCallerPC(B);
+
+        const bool has_dart_callee = (reinterpret_cast<uword>(pc) & 2) == 0;
+        const intptr_t argc = has_dart_callee
+                                  ? Bytecode::DecodeArgc(pc[-1])
+                                  : (reinterpret_cast<uword>(pc) >> 2);
+
+        SP = FrameArguments(B, 0);
+        RawObject** args = SP - argc;
+        B = SavedCallerFP(B);
+        if (has_dart_callee) {
+          pp = FrameCode(B)->ptr()->object_pool_->ptr();
+        }
+
+        *++SP = null_value;
+        *++SP = args[0];
+        *++SP = argdesc.raw();
+        *++SP = null_value;  // will contain array
+        {
+          SP[1] = Smi::New(argc);  // length
+          SP[2] = null_value;      // type
+          Exit(thread, B, SP + 3, pc);
+          NativeArguments native_args(thread, 2, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_AllocateArray, native_args,
+                         goto ClosureNoSuchMethodDispatch);

[zra, 2016/04/14 18:27:49]

No.

[zra, 2016/04/15 14:16:18]

I realized I meant to remove this but thought I'd elaborate a bit. It was hard
to find this. If you don't see this it looks like the while loop dives right
into InvokeClosureNoSuchMethod. Passing goto statements to macros is difficult
to notice and decipher. It would be good at least to have more comments
scattered around this code, and probably also a guide somewhere in the simulator
that describes the overall strategy.

+
+          RawArray* array = static_cast<RawArray*>(SP[0]);
+          ASSERT(array->GetClassId() == kArrayCid);
+          for (intptr_t i = 0; i < argc; i++) {
+            array->ptr()->data()[i] = args[i];
+          }
+        }
+        {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments native_args(thread, 3, SP - 2, SP - 3);
+          INVOKE_RUNTIME(DRT_InvokeClosureNoSuchMethod, native_args,
+                         goto ClosureNoSuchMethodDispatch);
+          UNREACHABLE();
+        }
+      ClosureNoSuchMethodDispatch:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Entry, A_B_C);
+        const uint16_t num_locals = rB;
+        const uint16_t context_reg = rC;
+        const intptr_t pos_count = Smi::Value(*reinterpret_cast<RawSmi**>(
+            reinterpret_cast<uword>(argdesc.raw()->ptr()) +
+            Array::element_offset(ArgumentsDescriptor::kPositionalCountIndex)));
+
+        if (pos_count != rA) {
+          goto ClosureNoSuchMethod;
+        }
+
+        {
+          RawObject** L = B;
+          for (intptr_t i = 0; i < num_locals; i++) {
+            L[i] = null_value;
+          }
+          L[context_reg] = empty_context;
+          SP = B + num_locals - 1;
+        }
+
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(DebugStep, A);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args, FALLTHROUGH);
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(DebugBreak, A);
+        {
+          const uint32_t original_bc =
+              static_cast<uint32_t>(reinterpret_cast<uintptr_t>(
+                  thread->isolate()->debugger()->GetPatchedStubAddress(
+                      reinterpret_cast<uword>(pc))));
+
+          SP[1] = null_value;
+          Exit(thread, B, SP + 2, pc);
+          NativeArguments args(thread, 0, NULL, SP + 1);
+          INVOKE_RUNTIME(DRT_BreakpointRuntimeHandler, args, FALLTHROUGH) else {
+            DISPATCH_OP(original_bc);
+          }
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(CheckStack, A);
+        {
+          if (reinterpret_cast<uword>(SP) >= thread->isolate()->stack_limit()) {
+            Exit(thread, B, SP + 1, pc);
+            NativeArguments args(thread, 0, NULL, NULL);
+            INVOKE_RUNTIME(DRT_StackOverflow, args, FALLTHROUGH);
+          }
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(EntryOpt, A_B_C);
+        const uint16_t num_fixed_params = rA;
+        const uint16_t num_opt_pos_params = rB;
+        const uint16_t num_opt_named_params = rC;
+
+        const intptr_t min_num_pos_args = num_fixed_params;
+        const intptr_t max_num_pos_args = num_fixed_params + num_opt_pos_params;
+        const intptr_t arg_count = Smi::Value(*reinterpret_cast<RawSmi**>(
+            reinterpret_cast<uword>(argdesc.raw()->ptr()) +
+            Array::element_offset(ArgumentsDescriptor::kCountIndex)));
+        const intptr_t pos_count = Smi::Value(*reinterpret_cast<RawSmi**>(
+            reinterpret_cast<uword>(argdesc.raw()->ptr()) +
+            Array::element_offset(ArgumentsDescriptor::kPositionalCountIndex)));
+        const intptr_t named_count = (arg_count - pos_count);
+        if (!(min_num_pos_args <= pos_count && pos_count <= max_num_pos_args)) {
+          goto ClosureNoSuchMethod;
+        }
+
+        // Copy all passed position arguments.Entry
+        RawObject** first_arg = FrameArguments(B, arg_count);
+        memmove(B, first_arg, pos_count * kWordSize);
+        if (num_opt_named_params != 0) {
+          // RawObject** opt_copy = first_arg + pos_count;
+
+          intptr_t i = named_count - 1;
+          intptr_t j = num_opt_named_params - 1;
+          while (j >= 0 && i >= 0) {
+            const uint32_t load_name = pc[2 * j];
+            const uint32_t load_value = pc[2 * j + 1];
+            ASSERT((load_name & 0xFF) == Bytecode::kLoadConstant);
+            ASSERT((load_value & 0xFF) == Bytecode::kLoadConstant);
+            ASSERT((load_value & 0xFF00) == (load_name & 0xFF00));
+            const uint8_t reg = (load_name >> 8) & 0xFF;
+            RawString* name = reinterpret_cast<RawString*>(
+                pp->data()[load_name >> 16].raw_obj_);
+            if (name == argdesc.At(ArgumentsDescriptor::name_index(i))) {
+              const intptr_t copy_idx = Smi::Value(static_cast<RawSmi*>(
+                  argdesc.At(ArgumentsDescriptor::position_index(i))));
+              B[reg] = first_arg[copy_idx];
+              i--;
+            } else {
+              RawObject* value = pp->data()[load_value >> 16].raw_obj_;
+              B[reg] = value;
+            }
+            j--;
+          }
+
+          while (j >= 0) {
+            const uint32_t load_name = pc[2 * j];
+            const uint32_t load_value = pc[2 * j + 1];
+            ASSERT((load_name & 0xFF) == Bytecode::kLoadConstant);
+            ASSERT((load_value & 0xFF) == Bytecode::kLoadConstant);
+            ASSERT((load_value & 0xFF00) == (load_name & 0xFF00));
+            const uint8_t reg = (load_name >> 8) & 0xFF;
+            RawObject* value = pp->data()[load_value >> 16].raw_obj_;
+            B[reg] = value;
+            j--;
+          }
+
+          if (i != -1) {
+            goto ClosureNoSuchMethod;
+          }
+
+          pc += num_opt_named_params * 2;
+          SP = B + num_fixed_params + num_opt_named_params - 1;
+        } else {
+          ASSERT(num_opt_pos_params != 0);
+          if (named_count != 0) {
+            goto ClosureNoSuchMethod;
+          }
+
+          RawObject** opt_base = B + rA;
+          for (intptr_t i = pos_count - rA; i < num_opt_pos_params; i++) {
+            const uint32_t load_value = pc[i];
+            ASSERT((load_value & 0xFF) == Bytecode::kLoadConstant);
+            const uint8_t reg = (load_value >> 8) & 0xFF;
+            RawObject* value = pp->data()[load_value >> 16].raw_obj_;
+            ASSERT((i + rA) == reg);
+            opt_base[i] = value;
+          }
+
+          pc += num_opt_pos_params;
+          SP = B + max_num_pos_args - 1;
+        }
+
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InstantiateType, A_D);
+        RawObject* type = pp->data()[rD].raw_obj_;
+        SP[1] = type;
+        SP[2] = SP[0];
+        SP[0] = null_value;
+        Exit(thread, B, SP + 3, pc);
+        {
+          NativeArguments args(thread, 2, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_InstantiateType, args, FALLTHROUGH);
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InstantiateTypeArgumentsTOS, A_D);
+        RawTypeArguments* type_arguments =
+            static_cast<RawTypeArguments*>(pp->data()[rD].raw_obj_);
+
+        RawObject* instantiator = SP[0];
+        // If the instantiator is null and if the type argument vector
+        // instantiated from null becomes a vector of dynamic, then use null as
+        // the type arguments.
+        if (rA == 0 || null_value != instantiator) {
+          RawArray* instantiations = type_arguments->ptr()->instantiations_;
+          for (intptr_t i = 0;
+               instantiations->ptr()->data()[i] != NULL;  // kNoInstantiator
+               i += 2) {
+            if (instantiations->ptr()->data()[i] == instantiator) {
+              SP[0] = instantiations->ptr()->data()[i + 1];
+              goto InstantiateTypeArgumentsTOSDone;
+            }
+          }
+
+          SP[1] = type_arguments;
+          SP[2] = instantiator;
+
+          Exit(thread, B, SP + 3, pc);
+          NativeArguments args(thread, 2, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_InstantiateTypeArguments, args, FALLTHROUGH);
+        }
+
+      InstantiateTypeArgumentsTOSDone:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Throw, A);
+        {
+          SP[1] = 0;
+          Exit(thread, B, SP + 2, pc);
+          if (rA == 0) {
+            NativeArguments args(thread, 1, SP, SP + 1);
+            INVOKE_RUNTIME(DRT_Throw, args, FALLTHROUGH);
+          } else {
+            NativeArguments args(thread, 2, SP - 1, SP + 1);
+            INVOKE_RUNTIME(DRT_ReThrow, args, FALLTHROUGH);
+          }
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Frame, A_D);
+        const uint16_t num_locals = rD;
+        for (intptr_t i = 1; i <= num_locals; i++) {
+          SP[i] = null_value;
+        }
+        SP += num_locals;
+
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(SetFrame, A);
+        SP = B + rA - 1;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Drop1, 0);
+        SP--;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Drop, 0);
+        SP -= rA;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(DropR, 0);
+        RawObject* result = SP[0];
+        SP -= rA;
+        SP[0] = result;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(LoadConstant, A_D);
+        B[rA] = pp->data()[rD].raw_obj_;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(PushConstant, __D);
+        *++SP = pp->data()[rD].raw_obj_;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Push, A_X);
+        *++SP = B[rD];
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Move, A_X);
+        B[rA] = B[rD];
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StoreLocal, A_X);
+        B[rD] = *SP;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(PopLocal, A_X);
+        B[rD] = *SP--;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(MoveSpecial, A_D);
+        B[rA] = special_[rD];
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(BooleanNegateTOS, 0);
+        SP[0] = (SP[0] == true_value) ? false_value : true_value;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StaticCall, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args, goto StaticCallDispatch);
+        }
+
+        {
+          const uint16_t argc = rA;
+          RawObject** call_base = SP - argc;
+          RawObject** call_top = SP;  // *SP contains function
+          argdesc ^= pp->data()[rD].raw_obj_;
+          Invoke(thread, call_base, call_top, &pp, &pc, &B, &SP);
+        }
+
+      StaticCallDispatch:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InstanceCall, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args,
+                         goto InstanceCallDispatch);
+        }
+
+        {
+          const uint16_t argc = rA;
+          const uint16_t kidx = rD;
+
+          RawObject** call_base = SP - argc + 1;
+          RawObject** call_top = SP + 1;
+          InstanceCall1(thread,
+                        static_cast<RawICData*>(pp->data()[kidx].raw_obj_),
+                        call_base, call_top, &argdesc, &pp, &pc, &B, &SP);
+        }
+
+      InstanceCallDispatch:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InstanceCall2, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args,
+                         goto InstanceCall2Dispatch);
+        }
+
+        {
+          const uint16_t argc = rA;
+          const uint16_t kidx = rD;
+
+          RawObject** call_base = SP - argc + 1;
+          RawObject** call_top = SP + 1;
+          InstanceCall2(thread,
+                        static_cast<RawICData*>(pp->data()[kidx].raw_obj_),
+                        call_base, call_top, &argdesc, &pp, &pc, &B, &SP);
+        }
+
+      InstanceCall2Dispatch:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InstanceCall3, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args,
+                         goto InstanceCall3Dispatch);
+        }
+
+        {
+          const uint16_t argc = rA;
+          const uint16_t kidx = rD;
+
+          RawObject** call_base = SP - argc + 1;
+          RawObject** call_top = SP + 1;
+          InstanceCall3(thread,
+                        static_cast<RawICData*>(pp->data()[kidx].raw_obj_),
+                        call_base, call_top, &argdesc, &pp, &pc, &B, &SP);
+        }
+
+      InstanceCall3Dispatch:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(NativeBootstrapCall, 0);
+        RawFunction* function = FrameFunction(B);
+        RawObject** incoming_args =
+            function->ptr()->num_optional_parameters_ == 0
+                ? FrameArguments(B, function->ptr()->num_fixed_parameters_)
+                : B;
+
+        SimulatorBootstrapNativeCall native_target =
+            reinterpret_cast<SimulatorBootstrapNativeCall>(SP[-1]);
+        intptr_t argc_tag = reinterpret_cast<intptr_t>(SP[-0]);
+        SP[-0] = 0;  // Note: argc_tag is not smi-tagged.
+        SP[-1] = null_value;
+        Exit(thread, B, SP + 1, pc);
+        NativeArguments args(thread, argc_tag, incoming_args, SP - 1);
+        INVOKE_NATIVE(native_target, args, FALLTHROUGH) else {
+          SP -= 1;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(NativeCall, 0);
+        RawFunction* function = FrameFunction(B);
+        RawObject** incoming_args =
+            function->ptr()->num_optional_parameters_ == 0
+                ? FrameArguments(B, function->ptr()->num_fixed_parameters_)
+                : B;
+
+        Dart_NativeFunction native_target =
+            reinterpret_cast<Dart_NativeFunction>(SP[-1]);
+        intptr_t argc_tag = reinterpret_cast<intptr_t>(SP[-0]);
+        SP[-0] = 0;  // argc_tag is not smi tagged!
+        SP[-1] = null_value;
+        Exit(thread, B, SP + 1, pc);
+        NativeArguments args(thread, argc_tag, incoming_args, SP - 1);
+        INVOKE_NATIVE_WRAPPER(native_target, args, FALLTHROUGH) else {
+          SP -= 1;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(AddTOS, A_B_C);
+        SMI_FASTPATH_TOS(intptr_t, SignedAddWithOverflow);
+      }
+      {
+        BYTECODE(SubTOS, A_B_C);
+        SMI_FASTPATH_TOS(intptr_t, SignedSubWithOverflow);
+      }
+      {
+        BYTECODE(MulTOS, A_B_C);
+        SMI_FASTPATH_TOS(intptr_t, SMI_MUL);
+      }
+      {
+        BYTECODE(BitOrTOS, A_B_C);
+        SMI_FASTPATH_TOS(intptr_t, SMI_BITOR);
+      }
+      {
+        BYTECODE(BitAndTOS, A_B_C);
+        SMI_FASTPATH_TOS(intptr_t, SMI_BITAND);
+      }
+      {
+        BYTECODE(EqualTOS, A_B_C);
+        SMI_FASTPATH_TOS(RawObject*, SMI_EQ);
+      }
+      {
+        BYTECODE(LessThanTOS, A_B_C);
+        SMI_FASTPATH_TOS(RawObject*, SMI_LT);
+      }
+      {
+        BYTECODE(GreaterThanTOS, A_B_C);
+        SMI_FASTPATH_TOS(RawObject*, SMI_GT);
+      }
+
+      {
+        BYTECODE(Compile, 0);
+        B[0] = FrameFunction(B);
+        B[1] = 0;
+        Exit(thread, B, B + 2, pc);
+        NativeArguments args(thread, 1, B, B + 1);
+        INVOKE_RUNTIME(DRT_CompileFunction, args, FALLTHROUGH) else {
+          // Function should be compiled now.
+          RawCode* code = FrameFunction(B)->ptr()->code_;
+          SetFrameCode(B, code);
+          pp = code->ptr()->object_pool_->ptr();
+          pc = reinterpret_cast<uint32_t*>(code->ptr()->entry_point_);
+        }
+        DISPATCH();
+      }
+
+      {
+        RawObject* result;
+
+        BYTECODE(Intrinsic, A);
+        result = null_value;
+        if (!intrinsics_[rA](thread, B, &result)) {
+          DISPATCH();
+        }
+        goto ReturnImpl;
+
+        BYTECODE(Return, A);
+        result = B[rA];
+        goto ReturnImpl;
+
+        BYTECODE(ReturnTOS, 0);
+        result = *SP;
+
+      ReturnImpl:
+        pc = SavedCallerPC(B);
+        if ((reinterpret_cast<uword>(pc) & 2) != 0) {
+          const intptr_t argc = reinterpret_cast<uword>(pc) >> 2;
+          base_ = top_ =
+              reinterpret_cast<uintptr_t*>(FrameArguments(B, argc + 1)[0]);
+          thread->set_top_exit_frame_info(reinterpret_cast<uword>(top_));
+          thread->set_top_resource(top_resource);
+          thread->set_vm_tag(vm_tag);
+          return result;
+        }
+        const uint8_t argc = Bytecode::DecodeArgc(pc[-1]);
+        SP = FrameArguments(B, argc);
+        B = SavedCallerFP(B);
+        *SP = result;
+        RawCode* code = FrameCode(B);
+        pp = code->ptr()->object_pool_->ptr();
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StoreStaticTOS, A_D);
+        RawField* field = reinterpret_cast<RawField*>(pp->data()[rD].raw_obj_);
+        RawInstance* value = static_cast<RawInstance*>(*SP--);
+        field->StorePointer(&field->ptr()->value_.static_value_, value);
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(PushStatic, A_D);
+        RawField* field = reinterpret_cast<RawField*>(pp->data()[rD].raw_obj_);
+        *SP = field->ptr()->value_.static_value_;  // Note: field on the stack.
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StoreField, A_B_C);
+        const uint16_t offset_in_words = rB;
+        const uint16_t value_reg = rC;
+
+        RawInstance* instance = reinterpret_cast<RawInstance*>(B[rA]);
+        RawObject* value = reinterpret_cast<RawObject*>(B[value_reg]);
+
+        instance->StorePointer(
+            reinterpret_cast<RawObject**>(instance->ptr()) + offset_in_words,
+            value);
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StoreFieldTOS, A_D);
+        const uint16_t offset_in_words = rD;
+        RawInstance* instance = reinterpret_cast<RawInstance*>(SP[-1]);
+        RawObject* value = reinterpret_cast<RawObject*>(SP[0]);
+        SP -= 2;
+        instance->StorePointer(
+            reinterpret_cast<RawObject**>(instance->ptr()) + offset_in_words,
+            value);
+
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(LoadField, A_B_C);
+        const uint16_t instance_reg = rB;
+        const uint16_t offset_in_words = rC;
+        RawInstance* instance = reinterpret_cast<RawInstance*>(B[instance_reg]);
+        B[rA] = reinterpret_cast<RawObject**>(instance->ptr())[offset_in_words];
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(LoadFieldTOS, A_D);
+        const uint16_t offset_in_words = rD;
+        RawInstance* instance = static_cast<RawInstance*>(SP[0]);
+        SP[0] = reinterpret_cast<RawObject**>(instance->ptr())[offset_in_words];
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(InitStaticTOS, A);
+        RawField* field = static_cast<RawField*>(*SP--);
+        RawObject* value = field->ptr()->value_.static_value_;
+        if ((value == Object::sentinel().raw()) ||
+            (value == Object::transition_sentinel().raw())) {
+          /* SP[1] field */
+          SP[2] = 0;
+          Exit(thread, B, SP + 3, pc);
+          NativeArguments args(thread, 1, SP + 1, SP + 2);
+          INVOKE_RUNTIME(DRT_InitStaticField, args, FALLTHROUGH);
+        }
+        DISPATCH();
+      }
+
+      // TODO(vegorov) allocation bytecodes can benefit from the new-space
+      // allocation fast-path that does not transition into the runtime system.
+      {
+        BYTECODE(AllocateContext, A_D);
+        const uint16_t num_context_variables = rD;
+        {
+          *++SP = 0;
+          SP[1] = Smi::New(num_context_variables);
+          Exit(thread, B, SP + 2, pc);
+          NativeArguments args(thread, 1, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_AllocateContext, args, FALLTHROUGH);
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(CloneContext, A);
+        {
+          SP[1] = SP[0];
+          Exit(thread, B, SP + 2, pc);
+          NativeArguments args(thread, 1, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_CloneContext, args, FALLTHROUGH);
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Allocate, A_D);
+        SP[1] = 0;  // result
+        SP[2] = pp->data()[rD].raw_obj_;
+        SP[3] = null_value;
+        Exit(thread, B, SP + 4, pc);
+        NativeArguments args(thread, 2, SP + 2, SP + 1);
+        INVOKE_RUNTIME(DRT_AllocateObject, args, FALLTHROUGH);
+        SP++;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(AllocateT, 0);
+        SP[1] = SP[-0];  // class
+        SP[2] = SP[-1];  // type arguments
+        Exit(thread, B, SP + 3, pc);
+        NativeArguments args(thread, 2, SP + 1, SP - 1);
+        INVOKE_RUNTIME(DRT_AllocateObject, args, FALLTHROUGH);
+        SP -= 1;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(CreateArrayTOS, 0);
+        SP[1] = SP[-0];  // length
+        SP[2] = SP[-1];  // type
+        Exit(thread, B, SP + 3, pc);
+        NativeArguments args(thread, 2, SP + 1, SP - 1);
+        INVOKE_RUNTIME(DRT_AllocateArray, args, FALLTHROUGH);
+        SP -= 1;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(AssertAssignable, A_D);  // instance, type args, type, name
+        RawObject** args = SP - 3;
+        if (args[0] != null_value) {
+          RawSubtypeTestCache* cache =
+              static_cast<RawSubtypeTestCache*>(pp->data()[rD].raw_obj_);
+          if (cache != null_value) {
+            RawInstance* instance = static_cast<RawInstance*>(args[0]);
+            RawTypeArguments* instantiator_type_arguments =
+                static_cast<RawTypeArguments*>(args[1]);
+
+            const intptr_t cid = SimulatorHelpers::GetClassId(instance);
+
+            RawTypeArguments* instance_type_arguments =
+                static_cast<RawTypeArguments*>(null_value);
+            RawObject* instance_cid_or_function;
+            if (cid == kClosureCid) {
+              RawClosure* closure = static_cast<RawClosure*>(instance);
+              instance_type_arguments = closure->ptr()->type_arguments_;
+              instance_cid_or_function = closure->ptr()->function_;
+            } else {
+              instance_cid_or_function = Smi::New(cid);
+
+              RawClass* instance_class =
+                  thread->isolate()->class_table()->At(cid);
+              if (instance_class->ptr()->num_type_arguments_ < 0) {
+                goto AssertAssignableCallRuntime;
+              } else if (instance_class->ptr()->num_type_arguments_ > 0) {
+                instance_type_arguments = reinterpret_cast<RawTypeArguments**>(
+                    instance
+                        ->ptr())[instance_class->ptr()
+                                     ->type_arguments_field_offset_in_words_];
+              }
+            }
+
+            for (RawObject** entries = cache->ptr()->cache_->ptr()->data();
+                 entries[0] != null_value;
+                 entries += SubtypeTestCache::kTestEntryLength) {
+              if (entries[SubtypeTestCache::kInstanceClassIdOrFunction] ==
+                      instance_cid_or_function &&
+                  entries[SubtypeTestCache::kInstanceTypeArguments] ==
+                      instance_type_arguments &&
+                  entries[SubtypeTestCache::kInstantiatorTypeArguments] ==
+                      instantiator_type_arguments) {
+                if (true_value == entries[SubtypeTestCache::kTestResult]) {
+                  goto AssertAssignableOk;
+                } else {
+                  break;
+                }
+              }
+            }
+          }
+
+        AssertAssignableCallRuntime:
+          SP[1] = args[0];  // instance
+          SP[2] = args[2];  // type
+          SP[3] = args[1];  // type args
+          SP[4] = args[3];  // name
+          SP[5] = cache;
+          Exit(thread, B, SP + 6, pc);
+          NativeArguments args(thread, 5, SP + 1, SP - 3);
+          INVOKE_RUNTIME(DRT_TypeCheck, args, FALLTHROUGH);
+        }
+
+      AssertAssignableOk:
+        SP -= 3;
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(AssertBoolean, A);
+        RawObject* value = SP[0];
+        if (rA) {
+          if (value == true_value || value == false_value) {
+            goto AssertBooleanOk;
+          }
+        } else if (value != null_value) {
+          goto AssertBooleanOk;
+        }
+
+        {
+          SP[1] = SP[0];  // instance
+          Exit(thread, B, SP + 2, pc);
+          NativeArguments args(thread, 1, SP + 1, SP);
+          INVOKE_RUNTIME(DRT_NonBoolTypeError, args, FALLTHROUGH);
+        }
+
+      AssertBooleanOk:
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(IfEqStrictTOS, A_D);
+        SP -= 2;
+        if (SP[1] != SP[2]) {
+          pc++;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(IfNeStrictTOS, A_D);
+        SP -= 2;
+        if (SP[1] == SP[2]) {
+          pc++;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(IfEqStrictNumTOS, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args, FALLTHROUGH);
+        }
+
+        SP -= 2;
+        if (!SimulatorHelpers::IsStrictEqualWithNumberCheck(SP[1], SP[2])) {
+          pc++;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(IfNeStrictNumTOS, A_D);
+        if (thread->isolate()->single_step()) {
+          Exit(thread, B, SP + 1, pc);
+          NativeArguments args(thread, 0, NULL, NULL);
+          INVOKE_RUNTIME(DRT_SingleStepHandler, args, FALLTHROUGH);
+        }
+
+        SP -= 2;
+        if (SimulatorHelpers::IsStrictEqualWithNumberCheck(SP[1], SP[2])) {
+          pc++;
+        }
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Jump, 0);
+        const int32_t tgt = static_cast<int32_t>(__op) >> 8;
+        pc += (tgt - 1);
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(StoreIndexedTOS, 0);
+        SP -= 3;
+        RawArray* array = static_cast<RawArray*>(SP[1]);
+        RawSmi* index = static_cast<RawSmi*>(SP[2]);
+        RawObject* value = SP[3];
+        ASSERT(array->GetClassId() == kArrayCid);
+        ASSERT(!index->IsHeapObject());
+        array->StorePointer(array->ptr()->data() + Smi::Value(index), value);
+        DISPATCH();
+      }
+
+      {
+        BYTECODE(Trap, 0);
+        UNIMPLEMENTED();
+        DISPATCH();
+      }
+    }
+  }
+
+  UNREACHABLE();
+  return 0;
+}
+
+void Simulator::Longjmp(uword pc,
+                        uword sp,
+                        uword fp,
+                        RawObject* raw_exception,
+                        RawObject* raw_stacktrace,
+                        Thread* thread) {
+  // Walk over all setjmp buffers (simulated --> C++ transitions)
+  // and try to find the setjmp associated with the simulated stack pointer.
+  SimulatorSetjmpBuffer* buf = last_setjmp_buffer();
+  while ((buf->link() != NULL) && (buf->link()->base() > fp)) {
+    buf = buf->link();
+  }
+  ASSERT(buf != NULL);
+  ASSERT(last_setjmp_buffer() == buf);
+
+  // The C++ caller has not cleaned up the stack memory of C++ frames.
+  // Prepare for unwinding frames by destroying all the stack resources
+  // in the previous C++ frames.
+  StackResource::Unwind(thread);
+
+  // Set the tag.
+  thread->set_vm_tag(VMTag::kDartTagId);
+  // Clear top exit frame.
+  thread->set_top_exit_frame_info(0);
+
+  ASSERT(raw_exception != Object::null());
+  top_ = reinterpret_cast<uintptr_t*>(sp);
+  base_ = reinterpret_cast<uintptr_t*>(fp);
+  pc_ = pc;
+  special_[kExceptionSpecialIndex] = raw_exception;
+  special_[kStacktraceSpecialIndex] = raw_stacktrace;
+  buf->Longjmp();
+  UNREACHABLE();
+}
+
+}  // namespace dart
+
+
+#endif  // defined TARGET_ARCH_DBC