MIPS: Re-land Crankshaft-generated KeyedLoad stubs.

Port r13236 (cd9236c5)

BUG=
TEST=

Review URL: https://codereview.chromium.org/11801002
Patch from Akos Palfi <palfia@homejinni.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13322 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/mips/assembler-mips-inl.h

@@ -80,9 +80,36 @@ bool Operand::is_reg() const {
 }
 
 
+int Register::NumAllocatableRegisters() {
+  if (CpuFeatures::IsSupported(FPU)) {
+    return kMaxNumAllocatableRegisters;
+  } else {
+    return kMaxNumAllocatableRegisters - kGPRsPerNonFPUDouble;
+  }
+}
+
+
+int DoubleRegister::NumRegisters() {
+  if (CpuFeatures::IsSupported(FPU)) {
+    return FPURegister::kNumRegisters;
+  } else {
+    return 1;
+  }
+}
+
+
+int DoubleRegister::NumAllocatableRegisters() {
+  if (CpuFeatures::IsSupported(FPU)) {
+    return FPURegister::kMaxNumAllocatableRegisters;
+  } else {
+    return 1;
+  }
+}
+
+
 int FPURegister::ToAllocationIndex(FPURegister reg) {
   ASSERT(reg.code() % 2 == 0);
-  ASSERT(reg.code() / 2 < kNumAllocatableRegisters);
+  ASSERT(reg.code() / 2 < kMaxNumAllocatableRegisters);
   ASSERT(reg.is_valid());
   ASSERT(!reg.is(kDoubleRegZero));
   ASSERT(!reg.is(kLithiumScratchDouble));

src/mips/assembler-mips.cc

@@ -73,6 +73,33 @@ static uint64_t CpuFeaturesImpliedByCompiler() {
 }
 
 
+const char* DoubleRegister::AllocationIndexToString(int index) {
+  if (CpuFeatures::IsSupported(FPU)) {
+    ASSERT(index >= 0 && index < kMaxNumAllocatableRegisters);
+    const char* const names[] = {
+      "f0",
+      "f2",
+      "f4",
+      "f6",
+      "f8",
+      "f10",
+      "f12",
+      "f14",
+      "f16",
+      "f18",
+      "f20",
+      "f22",
+      "f24",
+      "f26"
+    };
+    return names[index];
+  } else {
+    ASSERT(index == 0);
+    return "sfpd0";
+  }
+}
+
+
 void CpuFeatures::Probe() {
   unsigned standard_features = (OS::CpuFeaturesImpliedByPlatform() |
                                 CpuFeaturesImpliedByCompiler());

src/mips/assembler-mips.h

@@ -72,20 +72,23 @@ namespace internal {
 // Core register.
 struct Register {
   static const int kNumRegisters = v8::internal::kNumRegisters;
-  static const int kNumAllocatableRegisters = 14;  // v0 through t7.
+  static const int kMaxNumAllocatableRegisters = 14;  // v0 through t7.
   static const int kSizeInBytes = 4;
+  static const int kGPRsPerNonFPUDouble = 2;
+
+  inline static int NumAllocatableRegisters();
 
   static int ToAllocationIndex(Register reg) {
     return reg.code() - 2;  // zero_reg and 'at' are skipped.
   }
 
   static Register FromAllocationIndex(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < kMaxNumAllocatableRegisters);
     return from_code(index + 2);  // zero_reg and 'at' are skipped.
   }
 
   static const char* AllocationIndexToString(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < kMaxNumAllocatableRegisters);
     const char* const names[] = {
       "v0",
       "v1",
@@ -197,38 +200,19 @@ struct FPURegister {
   //  f28: 0.0
   //  f30: scratch register.
   static const int kNumReservedRegisters = 2;
-  static const int kNumAllocatableRegisters = kNumRegisters / 2 -
+  static const int kMaxNumAllocatableRegisters = kNumRegisters / 2 -
       kNumReservedRegisters;
 
-
+  inline static int NumRegisters();
+  inline static int NumAllocatableRegisters();
   inline static int ToAllocationIndex(FPURegister reg);
+  static const char* AllocationIndexToString(int index);
 
   static FPURegister FromAllocationIndex(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < kMaxNumAllocatableRegisters);
     return from_code(index * 2);
   }
 
-  static const char* AllocationIndexToString(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
-    const char* const names[] = {
-      "f0",
-      "f2",
-      "f4",
-      "f6",
-      "f8",
-      "f10",
-      "f12",
-      "f14",
-      "f16",
-      "f18",
-      "f20",
-      "f22",
-      "f24",
-      "f26"
-    };
-    return names[index];
-  }
-
   static FPURegister from_code(int code) {
     FPURegister r = { code };
     return r;
@@ -316,6 +300,9 @@ const FPURegister f29 = { 29 };
 const FPURegister f30 = { 30 };
 const FPURegister f31 = { 31 };
 
+const Register sfpd_lo  = { kRegister_t6_Code };
+const Register sfpd_hi  = { kRegister_t7_Code };
+
 // Register aliases.
 // cp is assumed to be a callee saved register.
 // Defined using #define instead of "static const Register&" because Clang

src/mips/builtins-mips.cc

@@ -1297,6 +1297,26 @@ CODE_AGE_LIST(DEFINE_CODE_AGE_BUILTIN_GENERATOR)
 #undef DEFINE_CODE_AGE_BUILTIN_GENERATOR
 
 
+void Builtins::Generate_NotifyICMiss(MacroAssembler* masm) {
+  {
+    FrameScope scope(masm, StackFrame::INTERNAL);
+
+    // Preserve registers across notification, this is important for compiled
+    // stubs that tail call the runtime on deopts passing their parameters in
+    // registers.
+    __ MultiPush(kJSCallerSaved | kCalleeSaved);
+    // Pass the function and deoptimization type to the runtime system.
+    __ CallRuntime(Runtime::kNotifyICMiss, 0);
+    __ MultiPop(kJSCallerSaved | kCalleeSaved);
+  }
+
+  __ mov(at, ra);  // Stash the miss continuation
+  __ Addu(sp, sp, Operand(kPointerSize));  // Ignore state
+  __ pop(ra);  // Restore RA to continuation in JSFunction
+  __ Jump(at);  // Jump to miss handler
+}
+
+
 static void Generate_NotifyDeoptimizedHelper(MacroAssembler* masm,
                                              Deoptimizer::BailoutType type) {
   {

src/mips/code-stubs-mips.cc

@@ -38,6 +38,17 @@ namespace v8 {
 namespace internal {
 
 
+void KeyedLoadFastElementStub::InitializeInterfaceDescriptor(
+    Isolate* isolate,
+    CodeStubInterfaceDescriptor* descriptor) {
+  static Register registers[] = { a1, a0 };
+  descriptor->register_param_count_ = 2;
+  descriptor->register_params_ = registers;
+  descriptor->deoptimization_handler_ =
+      isolate->builtins()->KeyedLoadIC_Miss();
+}
+
+
 #define __ ACCESS_MASM(masm)
 
 static void EmitIdenticalObjectComparison(MacroAssembler* masm,
@@ -500,7 +511,7 @@ void FastCloneShallowObjectStub::Generate(MacroAssembler* masm) {
 // 52 fraction bits (20 in the first word, 32 in the second).  Zeros is a
 // scratch register.  Destroys the source register.  No GC occurs during this
 // stub so you don't have to set up the frame.
-class ConvertToDoubleStub : public CodeStub {
+class ConvertToDoubleStub : public PlatformCodeStub {
  public:
   ConvertToDoubleStub(Register result_reg_1,
                       Register result_reg_2,
@@ -3893,12 +3904,29 @@ void CodeStub::GenerateStubsAheadOfTime() {
 
 
 void CodeStub::GenerateFPStubs() {
-  CEntryStub save_doubles(1, kSaveFPRegs);
-  Handle<Code> code = save_doubles.GetCode();
-  code->set_is_pregenerated(true);
-  StoreBufferOverflowStub stub(kSaveFPRegs);
-  stub.GetCode()->set_is_pregenerated(true);
-  code->GetIsolate()->set_fp_stubs_generated(true);
+  SaveFPRegsMode mode = CpuFeatures::IsSupported(FPU)
+      ? kSaveFPRegs
+      : kDontSaveFPRegs;
+  CEntryStub save_doubles(1, mode);
+  StoreBufferOverflowStub stub(mode);
+  // These stubs might already be in the snapshot, detect that and don't
+  // regenerate, which would lead to code stub initialization state being messed
+  // up.
+  Code* save_doubles_code = NULL;
+  Code* store_buffer_overflow_code = NULL;
+  if (!save_doubles.FindCodeInCache(&save_doubles_code, ISOLATE)) {
+    if (CpuFeatures::IsSupported(FPU)) {
+      CpuFeatures::Scope scope2(FPU);
+      save_doubles_code = *save_doubles.GetCode();
+      store_buffer_overflow_code = *stub.GetCode();
+    } else {
+      save_doubles_code = *save_doubles.GetCode();
+      store_buffer_overflow_code = *stub.GetCode();
+    }
+    save_doubles_code->set_is_pregenerated(true);
+    store_buffer_overflow_code->set_is_pregenerated(true);
+  }
+  ISOLATE->set_fp_stubs_generated(true);
 }
 
 

src/mips/code-stubs-mips.h

@@ -37,7 +37,7 @@ namespace internal {
 
 // Compute a transcendental math function natively, or call the
 // TranscendentalCache runtime function.
-class TranscendentalCacheStub: public CodeStub {
+class TranscendentalCacheStub: public PlatformCodeStub {
  public:
   enum ArgumentType {
     TAGGED = 0 << TranscendentalCache::kTranscendentalTypeBits,
@@ -59,7 +59,7 @@ class TranscendentalCacheStub: public CodeStub {
 };
 
 
-class StoreBufferOverflowStub: public CodeStub {
+class StoreBufferOverflowStub: public PlatformCodeStub {
  public:
   explicit StoreBufferOverflowStub(SaveFPRegsMode save_fp)
       : save_doubles_(save_fp) { }
@@ -78,7 +78,7 @@ class StoreBufferOverflowStub: public CodeStub {
 };
 
 
-class UnaryOpStub: public CodeStub {
+class UnaryOpStub: public PlatformCodeStub {
  public:
   UnaryOpStub(Token::Value op,
               UnaryOverwriteMode mode,
@@ -220,7 +220,7 @@ enum StringAddFlags {
 };
 
 
-class StringAddStub: public CodeStub {
+class StringAddStub: public PlatformCodeStub {
  public:
   explicit StringAddStub(StringAddFlags flags) : flags_(flags) {}
 
@@ -243,7 +243,7 @@ class StringAddStub: public CodeStub {
 };
 
 
-class SubStringStub: public CodeStub {
+class SubStringStub: public PlatformCodeStub {
  public:
   SubStringStub() {}
 
@@ -255,7 +255,7 @@ class SubStringStub: public CodeStub {
 };
 
 
-class StringCompareStub: public CodeStub {
+class StringCompareStub: public PlatformCodeStub {
  public:
   StringCompareStub() { }
 
@@ -296,7 +296,7 @@ class StringCompareStub: public CodeStub {
 // This stub can convert a signed int32 to a heap number (double).  It does
 // not work for int32s that are in Smi range!  No GC occurs during this stub
 // so you don't have to set up the frame.
-class WriteInt32ToHeapNumberStub : public CodeStub {
+class WriteInt32ToHeapNumberStub : public PlatformCodeStub {
  public:
   WriteInt32ToHeapNumberStub(Register the_int,
                              Register the_heap_number,
@@ -340,7 +340,7 @@ class WriteInt32ToHeapNumberStub : public CodeStub {
 };
 
 
-class NumberToStringStub: public CodeStub {
+class NumberToStringStub: public PlatformCodeStub {
  public:
   NumberToStringStub() { }
 
@@ -366,7 +366,7 @@ class NumberToStringStub: public CodeStub {
 };
 
 
-class RecordWriteStub: public CodeStub {
+class RecordWriteStub: public PlatformCodeStub {
  public:
   RecordWriteStub(Register object,
                   Register value,
@@ -512,7 +512,7 @@ class RecordWriteStub: public CodeStub {
     Register GetRegThatIsNotOneOf(Register r1,
                                   Register r2,
                                   Register r3) {
-      for (int i = 0; i < Register::kNumAllocatableRegisters; i++) {
+      for (int i = 0; i < Register::NumAllocatableRegisters(); i++) {
         Register candidate = Register::FromAllocationIndex(i);
         if (candidate.is(r1)) continue;
         if (candidate.is(r2)) continue;
@@ -571,7 +571,7 @@ class RecordWriteStub: public CodeStub {
 // Enter C code from generated RegExp code in a way that allows
 // the C code to fix the return address in case of a GC.
 // Currently only needed on ARM and MIPS.
-class RegExpCEntryStub: public CodeStub {
+class RegExpCEntryStub: public PlatformCodeStub {
  public:
   RegExpCEntryStub() {}
   virtual ~RegExpCEntryStub() {}
@@ -589,7 +589,7 @@ class RegExpCEntryStub: public CodeStub {
 // keep the code which called into native pinned in the memory. Currently the
 // simplest approach is to generate such stub early enough so it can never be
 // moved by GC
-class DirectCEntryStub: public CodeStub {
+class DirectCEntryStub: public PlatformCodeStub {
  public:
   DirectCEntryStub() {}
   void Generate(MacroAssembler* masm);
@@ -739,7 +739,7 @@ class FloatingPointHelper : public AllStatic {
 };
 
 
-class StringDictionaryLookupStub: public CodeStub {
+class StringDictionaryLookupStub: public PlatformCodeStub {
  public:
   enum LookupMode { POSITIVE_LOOKUP, NEGATIVE_LOOKUP };
 

src/mips/codegen-mips.h

@@ -46,6 +46,10 @@ enum TypeofState { INSIDE_TYPEOF, NOT_INSIDE_TYPEOF };
 
 class CodeGenerator: public AstVisitor {
  public:
+  CodeGenerator() {
+    InitializeAstVisitor();
+  }
+
   static bool MakeCode(CompilationInfo* info);
 
   // Printing of AST, etc. as requested by flags.
@@ -70,7 +74,7 @@ class CodeGenerator: public AstVisitor {
                               int pos,
                               bool right_here = false);
 
-  DEFINE_AST_VISITOR_SUBCLASS_METHODS();
+  DEFINE_AST_VISITOR_SUBCLASS_MEMBERS();
 
  private:
   DISALLOW_COPY_AND_ASSIGN(CodeGenerator);

src/mips/deoptimizer-mips.cc

@@ -206,7 +206,7 @@ static int LookupBailoutId(DeoptimizationInputData* data, BailoutId ast_id) {
 
 void Deoptimizer::DoComputeOsrOutputFrame() {
   DeoptimizationInputData* data = DeoptimizationInputData::cast(
-      optimized_code_->deoptimization_data());
+      compiled_code_->deoptimization_data());
   unsigned ast_id = data->OsrAstId()->value();
 
   int bailout_id = LookupBailoutId(data, BailoutId(ast_id));
@@ -240,7 +240,7 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
   unsigned input_frame_size = input_->GetFrameSize();
   ASSERT(fixed_size + height_in_bytes == input_frame_size);
 
-  unsigned stack_slot_size = optimized_code_->stack_slots() * kPointerSize;
+  unsigned stack_slot_size = compiled_code_->stack_slots() * kPointerSize;
   unsigned outgoing_height = data->ArgumentsStackHeight(bailout_id)->value();
   unsigned outgoing_size = outgoing_height * kPointerSize;
   unsigned output_frame_size = fixed_size + stack_slot_size + outgoing_size;
@@ -332,7 +332,7 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
 
     unsigned pc_offset = data->OsrPcOffset()->value();
     uint32_t pc = reinterpret_cast<uint32_t>(
-        optimized_code_->entry() + pc_offset);
+        compiled_code_->entry() + pc_offset);
     output_[0]->SetPc(pc);
   }
   Code* continuation = isolate_->builtins()->builtin(Builtins::kNotifyOSR);
@@ -445,6 +445,70 @@ void Deoptimizer::DoComputeArgumentsAdaptorFrame(TranslationIterator* iterator,
 }
 
 
+void Deoptimizer::DoCompiledStubFrame(TranslationIterator* iterator,
+                                      int frame_index) {
+  //
+  //               FROM                                  TO             <-fp
+  //    |          ....           |          |          ....           |
+  //    +-------------------------+          +-------------------------+
+  //    | JSFunction continuation |          | JSFunction continuation |
+  //    +-------------------------+          +-------------------------+<-sp
+  // |  |   saved frame (fp)      |
+  // |  +=========================+<-fp
+  // |  |   JSFunction context    |
+  // v  +-------------------------+
+  //    |   COMPILED_STUB marker  |          fp = saved frame
+  //    +-------------------------+          f8 = JSFunction context
+  //    |                         |
+  //    | ...                     |
+  //    |                         |
+  //    +-------------------------+<-sp
+  //
+  //
+  int output_frame_size = 1 * kPointerSize;
+  FrameDescription* output_frame =
+      new(output_frame_size) FrameDescription(output_frame_size, 0);
+  Code* notify_miss =
+      isolate_->builtins()->builtin(Builtins::kNotifyICMiss);
+  output_frame->SetState(Smi::FromInt(FullCodeGenerator::NO_REGISTERS));
+  output_frame->SetContinuation(
+      reinterpret_cast<intptr_t>(notify_miss->entry()));
+
+  ASSERT(compiled_code_->kind() == Code::COMPILED_STUB);
+  int major_key = compiled_code_->major_key();
+  CodeStubInterfaceDescriptor* descriptor =
+      isolate_->code_stub_interface_descriptor(major_key);
+  Handle<Code> miss_ic(descriptor->deoptimization_handler_);
+  output_frame->SetPc(reinterpret_cast<intptr_t>(miss_ic->instruction_start()));
+  unsigned input_frame_size = input_->GetFrameSize();
+  intptr_t value = input_->GetFrameSlot(input_frame_size - kPointerSize);
+  output_frame->SetFrameSlot(0, value);
+  value = input_->GetFrameSlot(input_frame_size - 2 * kPointerSize);
+  output_frame->SetRegister(fp.code(), value);
+  output_frame->SetFp(value);
+  value = input_->GetFrameSlot(input_frame_size - 3 * kPointerSize);
+  output_frame->SetRegister(cp.code(), value);
+
+  Translation::Opcode opcode =
+      static_cast<Translation::Opcode>(iterator->Next());
+  ASSERT(opcode == Translation::REGISTER);
+  USE(opcode);
+  int input_reg = iterator->Next();
+  intptr_t input_value = input_->GetRegister(input_reg);
+  output_frame->SetRegister(a1.code(), input_value);
+
+  int32_t next = iterator->Next();
+  opcode = static_cast<Translation::Opcode>(next);
+  ASSERT(opcode == Translation::REGISTER);
+  input_reg = iterator->Next();
+  input_value = input_->GetRegister(input_reg);
+  output_frame->SetRegister(a0.code(), input_value);
+
+  ASSERT(frame_index == 0);
+  output_[frame_index] = output_frame;
+}
+
+
 void Deoptimizer::DoComputeConstructStubFrame(TranslationIterator* iterator,
                                               int frame_index) {
   Builtins* builtins = isolate_->builtins();
@@ -868,7 +932,7 @@ void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
   }
   input_->SetRegister(sp.code(), reinterpret_cast<intptr_t>(frame->sp()));
   input_->SetRegister(fp.code(), reinterpret_cast<intptr_t>(frame->fp()));
-  for (int i = 0; i < DoubleRegister::kNumAllocatableRegisters; i++) {
+  for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
     input_->SetDoubleRegister(i, 0.0);
   }
 
@@ -889,7 +953,6 @@ void Deoptimizer::EntryGenerator::Generate() {
 
   Isolate* isolate = masm()->isolate();
 
-  CpuFeatures::Scope scope(FPU);
   // Unlike on ARM we don't save all the registers, just the useful ones.
   // For the rest, there are gaps on the stack, so the offsets remain the same.
   const int kNumberOfRegisters = Register::kNumRegisters;
@@ -898,15 +961,20 @@ void Deoptimizer::EntryGenerator::Generate() {
   RegList saved_regs = restored_regs | sp.bit() | ra.bit();
 
   const int kDoubleRegsSize =
-      kDoubleSize * FPURegister::kNumAllocatableRegisters;
+      kDoubleSize * FPURegister::kMaxNumAllocatableRegisters;
 
+  if (CpuFeatures::IsSupported(FPU)) {
+    CpuFeatures::Scope scope(FPU);
     // Save all FPU registers before messing with them.
     __ Subu(sp, sp, Operand(kDoubleRegsSize));
-  for (int i = 0; i < FPURegister::kNumAllocatableRegisters; ++i) {
+    for (int i = 0; i < FPURegister::kMaxNumAllocatableRegisters; ++i) {
       FPURegister fpu_reg = FPURegister::FromAllocationIndex(i);
       int offset = i * kDoubleSize;
       __ sdc1(fpu_reg, MemOperand(sp, offset));
     }
+  } else {
+    __ Subu(sp, sp, Operand(kDoubleRegsSize));
+  }
 
   // Push saved_regs (needed to populate FrameDescription::registers_).
   // Leave gaps for other registers.
@@ -977,15 +1045,18 @@ void Deoptimizer::EntryGenerator::Generate() {
     }
   }
 
+  if (CpuFeatures::IsSupported(FPU)) {
+    CpuFeatures::Scope scope(FPU);
     // Copy FPU registers to
     // double_registers_[DoubleRegister::kNumAllocatableRegisters]
     int double_regs_offset = FrameDescription::double_registers_offset();
-  for (int i = 0; i < FPURegister::kNumAllocatableRegisters; ++i) {
+    for (int i = 0; i < FPURegister::NumAllocatableRegisters(); ++i) {
       int dst_offset = i * kDoubleSize + double_regs_offset;
       int src_offset = i * kDoubleSize + kNumberOfRegisters * kPointerSize;
       __ ldc1(f0, MemOperand(sp, src_offset));
       __ sdc1(f0, MemOperand(a1, dst_offset));
     }
+  }
 
   // Remove the bailout id, eventually return address, and the saved registers
   // from the stack.
@@ -1005,11 +1076,14 @@ void Deoptimizer::EntryGenerator::Generate() {
   // frame description.
   __ Addu(a3, a1, Operand(FrameDescription::frame_content_offset()));
   Label pop_loop;
+  Label pop_loop_header;
+  __ Branch(&pop_loop_header);
   __ bind(&pop_loop);
   __ pop(t0);
   __ sw(t0, MemOperand(a3, 0));
-  __ Branch(USE_DELAY_SLOT, &pop_loop, ne, a2, Operand(sp));
-  __ addiu(a3, a3, sizeof(uint32_t));  // In delay slot.
+  __ addiu(a3, a3, sizeof(uint32_t));
+  __ bind(&pop_loop_header);
+  __ Branch(&pop_loop, ne, a2, Operand(sp));
 
   // Compute the output frame in the deoptimizer.
   __ push(a0);  // Preserve deoptimizer object across call.
@@ -1024,25 +1098,30 @@ void Deoptimizer::EntryGenerator::Generate() {
   __ pop(a0);  // Restore deoptimizer object (class Deoptimizer).
 
   // Replace the current (input) frame with the output frames.
-  Label outer_push_loop, inner_push_loop;
+  Label outer_push_loop, inner_push_loop,
+      outer_loop_header, inner_loop_header;
   // Outer loop state: a0 = current "FrameDescription** output_",
   // a1 = one past the last FrameDescription**.
   __ lw(a1, MemOperand(a0, Deoptimizer::output_count_offset()));
   __ lw(a0, MemOperand(a0, Deoptimizer::output_offset()));  // a0 is output_.
   __ sll(a1, a1, kPointerSizeLog2);  // Count to offset.
   __ addu(a1, a0, a1);  // a1 = one past the last FrameDescription**.
+  __ jmp(&outer_loop_header);
   __ bind(&outer_push_loop);
   // Inner loop state: a2 = current FrameDescription*, a3 = loop index.
   __ lw(a2, MemOperand(a0, 0));  // output_[ix]
   __ lw(a3, MemOperand(a2, FrameDescription::frame_size_offset()));
+  __ jmp(&inner_loop_header);
   __ bind(&inner_push_loop);
   __ Subu(a3, a3, Operand(sizeof(uint32_t)));
   __ Addu(t2, a2, Operand(a3));
   __ lw(t3, MemOperand(t2, FrameDescription::frame_content_offset()));
   __ push(t3);
+  __ bind(&inner_loop_header);
   __ Branch(&inner_push_loop, ne, a3, Operand(zero_reg));
 
   __ Addu(a0, a0, Operand(kPointerSize));
+  __ bind(&outer_loop_header);
   __ Branch(&outer_push_loop, lt, a0, Operand(a1));
 
 

src/mips/lithium-codegen-mips.cc

@@ -77,6 +77,7 @@ bool LCodeGen::GenerateCode() {
   return GeneratePrologue() &&
       GenerateBody() &&
       GenerateDeferredCode() &&
+      GenerateDeoptJumpTable() &&
       GenerateSafepointTable();
 }
 
@@ -116,6 +117,7 @@ void LCodeGen::Comment(const char* format, ...) {
 bool LCodeGen::GeneratePrologue() {
   ASSERT(is_generating());
 
+  if (info()->IsOptimizing()) {
     ProfileEntryHookStub::MaybeCallEntryHook(masm_);
 
 #ifdef DEBUG
@@ -143,15 +145,20 @@ bool LCodeGen::GeneratePrologue() {
       __ sw(a2, MemOperand(sp, receiver_offset));
       __ bind(&ok);
     }
+  }
 
   info()->set_prologue_offset(masm_->pc_offset());
+  if (NeedsEagerFrame()) {
     // The following three instructions must remain together and unmodified for
     // code aging to work properly.
     __ Push(ra, fp, cp, a1);
     // Add unused load of ip to ensure prologue sequence is identical for
     // full-codegen and lithium-codegen.
     __ LoadRoot(at, Heap::kUndefinedValueRootIndex);
-  __ Addu(fp, sp, Operand(2 * kPointerSize));  // Adj. FP to point to saved FP.
+    // Adj. FP to point to saved FP.
+    __ Addu(fp, sp, Operand(2 * kPointerSize));
+    frame_is_built_ = true;
+  }
 
   // Reserve space for the stack slots needed by the code.
   int slots = GetStackSlotCount();
@@ -170,7 +177,7 @@ bool LCodeGen::GeneratePrologue() {
   }
 
   // Possibly allocate a local context.
-  int heap_slots = scope()->num_heap_slots() - Context::MIN_CONTEXT_SLOTS;
+  int heap_slots = info()->num_heap_slots() - Context::MIN_CONTEXT_SLOTS;
   if (heap_slots > 0) {
     Comment(";;; Allocate local context");
     // Argument to NewContext is the function, which is in a1.
@@ -206,7 +213,7 @@ bool LCodeGen::GeneratePrologue() {
   }
 
   // Trace the call.
-  if (FLAG_trace) {
+  if (FLAG_trace && info()->IsOptimizing()) {
     __ CallRuntime(Runtime::kTraceEnter, 0);
   }
   EnsureSpaceForLazyDeopt();
@@ -264,10 +271,31 @@ bool LCodeGen::GenerateDeferredCode() {
     for (int i = 0; !is_aborted() && i < deferred_.length(); i++) {
       LDeferredCode* code = deferred_[i];
       __ bind(code->entry());
+      if (NeedsDeferredFrame()) {
+        Comment(";;; Deferred build frame",
+                code->instruction_index(),
+                code->instr()->Mnemonic());
+        ASSERT(!frame_is_built_);
+        ASSERT(info()->IsStub());
+        frame_is_built_ = true;
+        __ MultiPush(cp.bit() | fp.bit() | ra.bit());
+        __ li(scratch0(), Operand(Smi::FromInt(StackFrame::STUB)));
+        __ push(scratch0());
+        __ Addu(fp, sp, Operand(2 * kPointerSize));
+      }
       Comment(";;; Deferred code @%d: %s.",
               code->instruction_index(),
               code->instr()->Mnemonic());
       code->Generate();
+      if (NeedsDeferredFrame()) {
+        Comment(";;; Deferred destroy frame",
+                code->instruction_index(),
+                code->instr()->Mnemonic());
+        ASSERT(frame_is_built_);
+        __ pop(at);
+        __ MultiPop(cp.bit() | fp.bit() | ra.bit());
+        frame_is_built_ = false;
+      }
       __ jmp(code->exit());
     }
   }
@@ -279,10 +307,72 @@ bool LCodeGen::GenerateDeferredCode() {
 
 
 bool LCodeGen::GenerateDeoptJumpTable() {
-  // TODO(plind): not clear that this will have advantage for MIPS.
-  // Skipping it for now. Raised issue #100 for this.
-  Abort("Unimplemented: GenerateDeoptJumpTable");
-  return false;
+  // Check that the jump table is accessible from everywhere in the function
+  // code, i.e. that offsets to the table can be encoded in the 16bit signed
+  // immediate of a branch instruction.
+  // To simplify we consider the code size from the first instruction to the
+  // end of the jump table.
+  if (!is_int16((masm()->pc_offset() / Assembler::kInstrSize) +
+      deopt_jump_table_.length() * 12)) {
+    Abort("Generated code is too large");
+  }
+
+  Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm_);
+  __ RecordComment("[ Deoptimization jump table");
+  Label table_start;
+  __ bind(&table_start);
+  Label needs_frame_not_call;
+  Label needs_frame_is_call;
+  for (int i = 0; i < deopt_jump_table_.length(); i++) {
+    __ bind(&deopt_jump_table_[i].label);
+    Address entry = deopt_jump_table_[i].address;
+    __ li(t9, Operand(ExternalReference::ForDeoptEntry(entry)));
+    if (deopt_jump_table_[i].needs_frame) {
+      if (deopt_jump_table_[i].is_lazy_deopt) {
+        if (needs_frame_is_call.is_bound()) {
+          __ Branch(&needs_frame_is_call);
+        } else {
+          __ bind(&needs_frame_is_call);
+          __ MultiPush(cp.bit() | fp.bit() | ra.bit());
+          // This variant of deopt can only be used with stubs. Since we don't
+          // have a function pointer to install in the stack frame that we're
+          // building, install a special marker there instead.
+          ASSERT(info()->IsStub());
+          __ li(scratch0(), Operand(Smi::FromInt(StackFrame::STUB)));
+          __ push(scratch0());
+          __ Addu(fp, sp, Operand(2 * kPointerSize));
+          __ Call(t9);
+        }
+      } else {
+        if (needs_frame_not_call.is_bound()) {
+          __ Branch(&needs_frame_not_call);
+        } else {
+          __ bind(&needs_frame_not_call);
+          __ MultiPush(cp.bit() | fp.bit() | ra.bit());
+          // This variant of deopt can only be used with stubs. Since we don't
+          // have a function pointer to install in the stack frame that we're
+          // building, install a special marker there instead.
+          ASSERT(info()->IsStub());
+          __ li(scratch0(), Operand(Smi::FromInt(StackFrame::STUB)));
+          __ push(scratch0());
+          __ Addu(fp, sp, Operand(2 * kPointerSize));
+          __ Jump(t9);
+        }
+      }
+    } else {
+      if (deopt_jump_table_[i].is_lazy_deopt) {
+        __ Call(t9);
+      } else {
+        __ Jump(t9);
+      }
+    }
+  }
+  __ RecordComment("]");
+
+  // The deoptimization jump table is the last part of the instruction
+  // sequence. Mark the generated code as done unless we bailed out.
+  if (!is_aborted()) status_ = DONE;
+  return !is_aborted();
 }
 
 
@@ -482,7 +572,9 @@ void LCodeGen::WriteTranslation(LEnvironment* environment,
                    translation,
                    arguments_index,
                    arguments_count);
-  int closure_id = *info()->closure() != *environment->closure()
+  bool has_closure_id = !info()->closure().is_null() &&
+      *info()->closure() != *environment->closure();
+  int closure_id = has_closure_id
       ? DefineDeoptimizationLiteral(environment->closure())
       : Translation::kSelfLiteralId;
 
@@ -503,6 +595,9 @@ void LCodeGen::WriteTranslation(LEnvironment* environment,
       ASSERT(height == 0);
       translation->BeginSetterStubFrame(closure_id);
       break;
+    case STUB:
+      translation->BeginCompiledStubFrame();
+      break;
     case ARGUMENTS_ADAPTOR:
       translation->BeginArgumentsAdaptorFrame(closure_id, translation_size);
       break;
@@ -689,7 +784,11 @@ void LCodeGen::DeoptimizeIf(Condition cc,
   RegisterEnvironmentForDeoptimization(environment, Safepoint::kNoLazyDeopt);
   ASSERT(environment->HasBeenRegistered());
   int id = environment->deoptimization_index();
-  Address entry = Deoptimizer::GetDeoptimizationEntry(id, Deoptimizer::EAGER);
+
+  Deoptimizer::BailoutType bailout_type = info()->IsStub()
+      ? Deoptimizer::LAZY
+      : Deoptimizer::EAGER;
+  Address entry = Deoptimizer::GetDeoptimizationEntry(id, bailout_type);
   if (entry == NULL) {
     Abort("bailout was not prepared");
     return;
@@ -712,9 +811,22 @@ void LCodeGen::DeoptimizeIf(Condition cc,
     __ bind(&skip);
   }
 
-  // TODO(plind): The Arm port is a little different here, due to their
-  // DeOpt jump table, which is not used for Mips yet.
+  bool needs_lazy_deopt = info()->IsStub();
+  ASSERT(info()->IsStub() || frame_is_built_);
+  if (cc == al && !needs_lazy_deopt) {
     __ Jump(entry, RelocInfo::RUNTIME_ENTRY, cc, src1, src2);
+  } else {
+    // We often have several deopts to the same entry, reuse the last
+    // jump entry if this is the case.
+    if (deopt_jump_table_.is_empty() ||
+        (deopt_jump_table_.last().address != entry) ||
+        (deopt_jump_table_.last().is_lazy_deopt != needs_lazy_deopt) ||
+        (deopt_jump_table_.last().needs_frame != !frame_is_built_)) {
+      JumpTableEntry table_entry(entry, !frame_is_built_, needs_lazy_deopt);
+      deopt_jump_table_.Add(table_entry, zone());
+    }
+    __ Branch(&deopt_jump_table_.last().label, cc, src1, src2);
+  }
 }
 
 
@@ -1289,6 +1401,7 @@ void LCodeGen::DoConstantI(LConstantI* instr) {
 void LCodeGen::DoConstantD(LConstantD* instr) {
   ASSERT(instr->result()->IsDoubleRegister());
   DoubleRegister result = ToDoubleRegister(instr->result());
+  CpuFeatures::Scope scope(FPU);
   double v = instr->value();
   __ Move(result, v);
 }
@@ -1486,6 +1599,7 @@ void LCodeGen::DoMathMinMax(LMathMinMax* instr) {
     __ bind(&done);
   } else {
     ASSERT(instr->hydrogen()->representation().IsDouble());
+    CpuFeatures::Scope scope(FPU);
     FPURegister left_reg = ToDoubleRegister(left);
     FPURegister right_reg = ToDoubleRegister(right);
     FPURegister result_reg = ToDoubleRegister(instr->result());
@@ -1526,6 +1640,7 @@ void LCodeGen::DoMathMinMax(LMathMinMax* instr) {
 
 
 void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister left = ToDoubleRegister(instr->left());
   DoubleRegister right = ToDoubleRegister(instr->right());
   DoubleRegister result = ToDoubleRegister(instr->result());
@@ -1635,6 +1750,7 @@ void LCodeGen::DoBranch(LBranch* instr) {
     Register reg = ToRegister(instr->value());
     EmitBranch(true_block, false_block, ne, reg, Operand(zero_reg));
   } else if (r.IsDouble()) {
+    CpuFeatures::Scope scope(FPU);
     DoubleRegister reg = ToDoubleRegister(instr->value());
     // Test the double value. Zero and NaN are false.
     EmitBranchF(true_block, false_block, ne, reg, kDoubleRegZero);
@@ -1712,6 +1828,7 @@ void LCodeGen::DoBranch(LBranch* instr) {
       }
 
       if (expected.Contains(ToBooleanStub::HEAP_NUMBER)) {
+        CpuFeatures::Scope scope(FPU);
         // heap number -> false iff +0, -0, or NaN.
         DoubleRegister dbl_scratch = double_scratch0();
         Label not_heap_number;
@@ -1791,6 +1908,7 @@ void LCodeGen::DoCmpIDAndBranch(LCmpIDAndBranch* instr) {
     EmitGoto(next_block);
   } else {
     if (instr->is_double()) {
+      CpuFeatures::Scope scope(FPU);
       // Compare left and right as doubles and load the
       // resulting flags into the normal status register.
       FPURegister left_reg = ToDoubleRegister(left);
@@ -2338,16 +2456,21 @@ void LCodeGen::DoCmpT(LCmpT* instr) {
 
 
 void LCodeGen::DoReturn(LReturn* instr) {
-  if (FLAG_trace) {
+  if (FLAG_trace && info()->IsOptimizing()) {
     // Push the return value on the stack as the parameter.
     // Runtime::TraceExit returns its parameter in v0.
     __ push(v0);
     __ CallRuntime(Runtime::kTraceExit, 1);
   }
+  if (NeedsEagerFrame()) {
     int32_t sp_delta = (GetParameterCount() + 1) * kPointerSize;
     __ mov(sp, fp);
     __ Pop(ra, fp);
     __ Addu(sp, sp, Operand(sp_delta));
+  }
+  if (info()->IsStub()) {
+    __ lw(cp, MemOperand(fp, StandardFrameConstants::kContextOffset));
+  }
   __ Jump(ra);
 }
 
@@ -2706,13 +2829,62 @@ void LCodeGen::DoLoadKeyedExternalArray(LLoadKeyed* instr) {
       __ sll(scratch0(), key, shift_size);
       __ Addu(scratch0(), scratch0(), external_pointer);
     }
-
+    if (CpuFeatures::IsSupported(FPU)) {
+      CpuFeatures::Scope scope(FPU);
       if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
         __ lwc1(result, MemOperand(scratch0(), additional_offset));
         __ cvt_d_s(result, result);
       } else  {  // i.e. elements_kind == EXTERNAL_DOUBLE_ELEMENTS
         __ ldc1(result, MemOperand(scratch0(), additional_offset));
       }
+    } else {
+      if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
+        Register value = external_pointer;
+        __ lw(value, MemOperand(scratch0(), additional_offset));
+        __ And(sfpd_lo, value, Operand(kBinary32MantissaMask));
+
+        __ srl(scratch0(), value, kBinary32MantissaBits);
+        __ And(scratch0(), scratch0(),
+                Operand(kBinary32ExponentMask >> kBinary32MantissaBits));
+
+        Label exponent_rebiased;
+        __ Xor(at, scratch0(), Operand(0x00));
+        __ Branch(&exponent_rebiased, eq, at, Operand(zero_reg));
+
+        __ Xor(at, scratch0(), Operand(0xff));
+        Label skip;
+        __ Branch(&skip, ne, at, Operand(zero_reg));
+        __ li(scratch0(), Operand(0x7ff));
+        __ bind(&skip);
+        __ Branch(&exponent_rebiased, eq, at, Operand(zero_reg));
+
+        // Rebias exponent.
+        __ Addu(scratch0(),
+                scratch0(),
+                Operand(-kBinary32ExponentBias + HeapNumber::kExponentBias));
+
+        __ bind(&exponent_rebiased);
+        __ And(sfpd_hi, value, Operand(kBinary32SignMask));
+        __ sll(at, scratch0(), HeapNumber::kMantissaBitsInTopWord);
+        __ Or(sfpd_hi, sfpd_hi, at);
+
+        // Shift mantissa.
+        static const int kMantissaShiftForHiWord =
+            kBinary32MantissaBits - HeapNumber::kMantissaBitsInTopWord;
+
+        static const int kMantissaShiftForLoWord =
+            kBitsPerInt - kMantissaShiftForHiWord;
+
+        __ srl(at, sfpd_lo, kMantissaShiftForHiWord);
+        __ Or(sfpd_hi, sfpd_hi, at);
+        __ sll(sfpd_lo, sfpd_lo, kMantissaShiftForLoWord);
+
+      } else {
+        __ lw(sfpd_lo, MemOperand(scratch0(), additional_offset));
+        __ lw(sfpd_hi, MemOperand(scratch0(),
+                                   additional_offset + kPointerSize));
+      }
+    }
   } else {
     Register result = ToRegister(instr->result());
     MemOperand mem_operand = PrepareKeyedOperand(
@@ -2780,25 +2952,28 @@ void LCodeGen::DoLoadKeyedFixedDoubleArray(LLoadKeyed* instr) {
     key = ToRegister(instr->key());
   }
 
-  if (key_is_constant) {
-    __ Addu(elements, elements,
-        Operand(((constant_key + instr->additional_index()) <<
-                 element_size_shift) +
-                FixedDoubleArray::kHeaderSize - kHeapObjectTag));
-  } else {
+  int base_offset = (FixedDoubleArray::kHeaderSize - kHeapObjectTag) +
+      ((constant_key + instr->additional_index()) << element_size_shift);
+  if (!key_is_constant) {
     __ sll(scratch, key, shift_size);
-    __ Addu(elements, elements, Operand(scratch));
-    __ Addu(elements, elements,
-            Operand((FixedDoubleArray::kHeaderSize - kHeapObjectTag) +
-                    (instr->additional_index() << element_size_shift)));
+    __ Addu(elements, elements, scratch);
   }
-
+  if (CpuFeatures::IsSupported(FPU)) {
+    CpuFeatures::Scope scope(FPU);
+    __ Addu(elements, elements, Operand(base_offset));
+    __ ldc1(result, MemOperand(elements));
     if (instr->hydrogen()->RequiresHoleCheck()) {
       __ lw(scratch, MemOperand(elements, sizeof(kHoleNanLower32)));
       DeoptimizeIf(eq, instr->environment(), scratch, Operand(kHoleNanUpper32));
     }
-
-  __ ldc1(result, MemOperand(elements));
+  } else {
+      __ lw(sfpd_hi, MemOperand(elements, base_offset + kPointerSize));
+      __ lw(sfpd_lo, MemOperand(elements, base_offset));
+    if (instr->hydrogen()->RequiresHoleCheck()) {
+      ASSERT(kPointerSize == sizeof(kHoleNanLower32));
+      DeoptimizeIf(eq, instr->environment(), sfpd_hi, Operand(kHoleNanUpper32));
+    }
+  }
 }
 
 
@@ -3241,6 +3416,7 @@ void LCodeGen::EmitIntegerMathAbs(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   // Class for deferred case.
   class DeferredMathAbsTaggedHeapNumber: public LDeferredCode {
    public:
@@ -3277,6 +3453,7 @@ void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   Register result = ToRegister(instr->result());
   Register scratch1 = scratch0();
@@ -3305,6 +3482,7 @@ void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   Register result = ToRegister(instr->result());
   DoubleRegister double_scratch1 = ToDoubleRegister(instr->temp());
@@ -3381,6 +3559,7 @@ void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());
   __ sqrt_d(result, input);
@@ -3388,6 +3567,7 @@ void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());
   DoubleRegister temp = ToDoubleRegister(instr->temp());
@@ -3412,6 +3592,7 @@ void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoPower(LPower* instr) {
+  CpuFeatures::Scope scope(FPU);
   Representation exponent_type = instr->hydrogen()->right()->representation();
   // Having marked this as a call, we can use any registers.
   // Just make sure that the input/output registers are the expected ones.
@@ -3442,6 +3623,7 @@ void LCodeGen::DoPower(LPower* instr) {
 
 
 void LCodeGen::DoRandom(LRandom* instr) {
+  CpuFeatures::Scope scope(FPU);
   class DeferredDoRandom: public LDeferredCode {
    public:
     DeferredDoRandom(LCodeGen* codegen, LRandom* instr)
@@ -3518,6 +3700,7 @@ void LCodeGen::DoDeferredRandom(LRandom* instr) {
 
 
 void LCodeGen::DoMathExp(LMathExp* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());
   DoubleRegister double_scratch1 = ToDoubleRegister(instr->double_temp());
@@ -3813,6 +3996,7 @@ void LCodeGen::DoBoundsCheck(LBoundsCheck* instr) {
 
 
 void LCodeGen::DoStoreKeyedExternalArray(LStoreKeyed* instr) {
+  CpuFeatures::Scope scope(FPU);
   Register external_pointer = ToRegister(instr->elements());
   Register key = no_reg;
   ElementsKind elements_kind = instr->elements_kind();
@@ -3886,6 +4070,7 @@ void LCodeGen::DoStoreKeyedExternalArray(LStoreKeyed* instr) {
 
 
 void LCodeGen::DoStoreKeyedFixedDoubleArray(LStoreKeyed* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister value = ToDoubleRegister(instr->value());
   Register elements = ToRegister(instr->elements());
   Register key = no_reg;
@@ -4169,6 +4354,7 @@ void LCodeGen::DoStringLength(LStringLength* instr) {
 
 
 void LCodeGen::DoInteger32ToDouble(LInteger32ToDouble* instr) {
+  CpuFeatures::Scope scope(FPU);
   LOperand* input = instr->value();
   ASSERT(input->IsRegister() || input->IsStackSlot());
   LOperand* output = instr->result();
@@ -4186,6 +4372,7 @@ void LCodeGen::DoInteger32ToDouble(LInteger32ToDouble* instr) {
 
 
 void LCodeGen::DoUint32ToDouble(LUint32ToDouble* instr) {
+  CpuFeatures::Scope scope(FPU);
   LOperand* input = instr->value();
   LOperand* output = instr->result();
 
@@ -4247,13 +4434,51 @@ void LCodeGen::DoNumberTagU(LNumberTagU* instr) {
 }
 
 
+// Convert unsigned integer with specified number of leading zeroes in binary
+// representation to IEEE 754 double.
+// Integer to convert is passed in register hiword.
+// Resulting double is returned in registers hiword:loword.
+// This functions does not work correctly for 0.
+static void GenerateUInt2Double(MacroAssembler* masm,
+                                Register hiword,
+                                Register loword,
+                                Register scratch,
+                                int leading_zeroes) {
+  const int meaningful_bits = kBitsPerInt - leading_zeroes - 1;
+  const int biased_exponent = HeapNumber::kExponentBias + meaningful_bits;
+
+  const int mantissa_shift_for_hi_word =
+      meaningful_bits - HeapNumber::kMantissaBitsInTopWord;
+  const int mantissa_shift_for_lo_word =
+      kBitsPerInt - mantissa_shift_for_hi_word;
+  masm->li(scratch, Operand(biased_exponent << HeapNumber::kExponentShift));
+  if (mantissa_shift_for_hi_word > 0) {
+    masm->sll(loword, hiword, mantissa_shift_for_lo_word);
+    masm->srl(hiword, hiword, mantissa_shift_for_hi_word);
+    masm->Or(hiword, scratch, hiword);
+  } else {
+    masm->mov(loword, zero_reg);
+    masm->sll(hiword, hiword, mantissa_shift_for_hi_word);
+    masm->Or(hiword, scratch, hiword);
+  }
+
+  // If least significant bit of biased exponent was not 1 it was corrupted
+  // by most significant bit of mantissa so we should fix that.
+  if (!(biased_exponent & 1)) {
+    masm->li(scratch, 1 << HeapNumber::kExponentShift);
+    masm->nor(scratch, scratch, scratch);
+    masm->and_(hiword, hiword, scratch);
+  }
+}
+
+
 void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
                                     LOperand* value,
                                     IntegerSignedness signedness) {
   Label slow;
   Register src = ToRegister(value);
   Register dst = ToRegister(instr->result());
-  FPURegister dbl_scratch = double_scratch0();
+  DoubleRegister dbl_scratch = double_scratch0();
 
   // Preserve the value of all registers.
   PushSafepointRegistersScope scope(this, Safepoint::kWithRegisters);
@@ -4267,16 +4492,40 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
       __ SmiUntag(src, dst);
       __ Xor(src, src, Operand(0x80000000));
     }
+    if (CpuFeatures::IsSupported(FPU)) {
+      CpuFeatures::Scope scope(FPU);
       __ mtc1(src, dbl_scratch);
       __ cvt_d_w(dbl_scratch, dbl_scratch);
     } else {
+      FloatingPointHelper::Destination dest =
+          FloatingPointHelper::kCoreRegisters;
+      FloatingPointHelper::ConvertIntToDouble(masm(), src, dest, f0,
+                                              sfpd_lo, sfpd_hi,
+                                              scratch0(), f2);
+    }
+  } else {
+    if (CpuFeatures::IsSupported(FPU)) {
+      CpuFeatures::Scope scope(FPU);
       __ mtc1(src, dbl_scratch);
       __ Cvt_d_uw(dbl_scratch, dbl_scratch, f22);
+    } else {
+      Label no_leading_zero, done;
+      __ And(at, src, Operand(0x80000000));
+      __ Branch(&no_leading_zero, ne, at, Operand(zero_reg));
+
+      // Integer has one leading zeros.
+      GenerateUInt2Double(masm(), sfpd_hi, sfpd_lo, t0, 1);
+      __ Branch(&done);
+
+      __ bind(&no_leading_zero);
+      GenerateUInt2Double(masm(), sfpd_hi, sfpd_lo, t0, 0);
+      __ Branch(&done);
+    }
   }
 
   if (FLAG_inline_new) {
-    __ LoadRoot(t2, Heap::kHeapNumberMapRootIndex);
-    __ AllocateHeapNumber(t1, a3, t0, t2, &slow, DONT_TAG_RESULT);
+    __ LoadRoot(scratch0(), Heap::kHeapNumberMapRootIndex);
+    __ AllocateHeapNumber(t1, a3, t0, scratch0(), &slow, DONT_TAG_RESULT);
     __ Move(dst, t1);
     __ Branch(&done);
   }
@@ -4295,7 +4544,13 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
   // Done. Put the value in dbl_scratch into the value of the allocated heap
   // number.
   __ bind(&done);
+  if (CpuFeatures::IsSupported(FPU)) {
+    CpuFeatures::Scope scope(FPU);
     __ sdc1(dbl_scratch, MemOperand(dst, HeapNumber::kValueOffset));
+  } else {
+    __ sw(sfpd_lo, MemOperand(dst, HeapNumber::kMantissaOffset));
+    __ sw(sfpd_hi, MemOperand(dst, HeapNumber::kExponentOffset));
+  }
   __ Addu(dst, dst, kHeapObjectTag);
   __ StoreToSafepointRegisterSlot(dst, dst);
 }
@@ -4328,7 +4583,13 @@ void LCodeGen::DoNumberTagD(LNumberTagD* instr) {
     __ Branch(deferred->entry());
   }
   __ bind(deferred->exit());
+  if (CpuFeatures::IsSupported(FPU)) {
+    CpuFeatures::Scope scope(FPU);
     __ sdc1(input_reg, MemOperand(reg, HeapNumber::kValueOffset));
+  } else {
+    __ sw(sfpd_lo, MemOperand(reg, HeapNumber::kValueOffset));
+    __ sw(sfpd_hi, MemOperand(reg, HeapNumber::kValueOffset + kPointerSize));
+  }
   // Now that we have finished with the object's real address tag it
   __ Addu(reg, reg, kHeapObjectTag);
 }
@@ -4376,6 +4637,7 @@ void LCodeGen::EmitNumberUntagD(Register input_reg,
                                 bool deoptimize_on_minus_zero,
                                 LEnvironment* env) {
   Register scratch = scratch0();
+  CpuFeatures::Scope scope(FPU);
 
   Label load_smi, heap_number, done;
 
@@ -4440,6 +4702,7 @@ void LCodeGen::DoDeferredTaggedToI(LTaggedToI* instr) {
   // of the if.
 
   if (instr->truncating()) {
+    CpuFeatures::Scope scope(FPU);
     Register scratch3 = ToRegister(instr->temp2());
     FPURegister single_scratch = double_scratch.low();
     ASSERT(!scratch3.is(input_reg) &&
@@ -4676,6 +4939,7 @@ void LCodeGen::DoCheckMaps(LCheckMaps* instr) {
 
 
 void LCodeGen::DoClampDToUint8(LClampDToUint8* instr) {
+  CpuFeatures::Scope vfp_scope(FPU);
   DoubleRegister value_reg = ToDoubleRegister(instr->unclamped());
   Register result_reg = ToRegister(instr->result());
   DoubleRegister temp_reg = ToDoubleRegister(instr->temp());
@@ -4684,6 +4948,7 @@ void LCodeGen::DoClampDToUint8(LClampDToUint8* instr) {
 
 
 void LCodeGen::DoClampIToUint8(LClampIToUint8* instr) {
+  CpuFeatures::Scope vfp_scope(FPU);
   Register unclamped_reg = ToRegister(instr->unclamped());
   Register result_reg = ToRegister(instr->result());
   __ ClampUint8(result_reg, unclamped_reg);
@@ -4691,6 +4956,7 @@ void LCodeGen::DoClampIToUint8(LClampIToUint8* instr) {
 
 
 void LCodeGen::DoClampTToUint8(LClampTToUint8* instr) {
+  CpuFeatures::Scope vfp_scope(FPU);
   Register scratch = scratch0();
   Register input_reg = ToRegister(instr->unclamped());
   Register result_reg = ToRegister(instr->result());
@@ -5314,6 +5580,7 @@ void LCodeGen::EmitIsConstructCall(Register temp1, Register temp2) {
 
 
 void LCodeGen::EnsureSpaceForLazyDeopt() {
+  if (info()->IsStub()) return;
   // Ensure that we have enough space after the previous lazy-bailout
   // instruction for patching the code here.
   int current_pc = masm()->pc_offset();

src/mips/lithium-codegen-mips.h

@@ -61,6 +61,7 @@ class LCodeGen BASE_EMBEDDED {
         deferred_(8, info->zone()),
         osr_pc_offset_(-1),
         last_lazy_deopt_pc_(0),
+        frame_is_built_(false),
         safepoints_(info->zone()),
         resolver_(this),
         expected_safepoint_kind_(Safepoint::kSimple) {
@@ -76,6 +77,15 @@ class LCodeGen BASE_EMBEDDED {
   Heap* heap() const { return isolate()->heap(); }
   Zone* zone() const { return zone_; }
 
+  bool NeedsEagerFrame() const {
+    return GetStackSlotCount() > 0 ||
+        info()->is_non_deferred_calling() ||
+        !info()->IsStub();
+  }
+  bool NeedsDeferredFrame() const {
+    return !NeedsEagerFrame() && info()->is_deferred_calling();
+  }
+
   // Support for converting LOperands to assembler types.
   // LOperand must be a register.
   Register ToRegister(LOperand* op) const;
@@ -189,7 +199,7 @@ class LCodeGen BASE_EMBEDDED {
                        Register temporary2);
 
   int GetStackSlotCount() const { return chunk()->spill_slot_count(); }
-  int GetParameterCount() const { return scope()->num_parameters(); }
+  int GetParameterCount() const { return info()->num_parameters(); }
 
   void Abort(const char* reason);
   void Comment(const char* format, ...);
@@ -368,11 +378,15 @@ class LCodeGen BASE_EMBEDDED {
                     int* offset);
 
   struct JumpTableEntry {
-    explicit inline JumpTableEntry(Address entry)
+    inline JumpTableEntry(Address entry, bool frame, bool is_lazy)
         : label(),
-          address(entry) { }
+          address(entry),
+          needs_frame(frame),
+          is_lazy_deopt(is_lazy) { }
     Label label;
     Address address;
+    bool needs_frame;
+    bool is_lazy_deopt;
   };
 
   void EnsureSpaceForLazyDeopt();
@@ -401,6 +415,7 @@ class LCodeGen BASE_EMBEDDED {
   ZoneList<LDeferredCode*> deferred_;
   int osr_pc_offset_;
   int last_lazy_deopt_pc_;
+  bool frame_is_built_;
 
   // Builder that keeps track of safepoints in the code. The table
   // itself is emitted at the end of the generated code.
@@ -416,6 +431,7 @@ class LCodeGen BASE_EMBEDDED {
     PushSafepointRegistersScope(LCodeGen* codegen,
                                 Safepoint::Kind kind)
         : codegen_(codegen) {
+      ASSERT(codegen_->info()->is_calling());
       ASSERT(codegen_->expected_safepoint_kind_ == Safepoint::kSimple);
       codegen_->expected_safepoint_kind_ = kind;
 

src/mips/lithium-gap-resolver-mips.cc

@@ -172,8 +172,10 @@ void LGapResolver::BreakCycle(int index) {
   } else if (source->IsStackSlot()) {
     __ lw(kLithiumScratchReg, cgen_->ToMemOperand(source));
   } else if (source->IsDoubleRegister()) {
+    CpuFeatures::Scope scope(FPU);
     __ mov_d(kLithiumScratchDouble, cgen_->ToDoubleRegister(source));
   } else if (source->IsDoubleStackSlot()) {
+    CpuFeatures::Scope scope(FPU);
     __ ldc1(kLithiumScratchDouble, cgen_->ToMemOperand(source));
   } else {
     UNREACHABLE();
@@ -193,9 +195,11 @@ void LGapResolver::RestoreValue() {
   } else if (saved_destination_->IsStackSlot()) {
     __ sw(kLithiumScratchReg, cgen_->ToMemOperand(saved_destination_));
   } else if (saved_destination_->IsDoubleRegister()) {
+    CpuFeatures::Scope scope(FPU);
     __ mov_d(cgen_->ToDoubleRegister(saved_destination_),
             kLithiumScratchDouble);
   } else if (saved_destination_->IsDoubleStackSlot()) {
+    CpuFeatures::Scope scope(FPU);
     __ sdc1(kLithiumScratchDouble,
             cgen_->ToMemOperand(saved_destination_));
   } else {
@@ -232,6 +236,7 @@ void LGapResolver::EmitMove(int index) {
       MemOperand destination_operand = cgen_->ToMemOperand(destination);
       if (in_cycle_) {
         if (!destination_operand.OffsetIsInt16Encodable()) {
+          CpuFeatures::Scope scope(FPU);
           // 'at' is overwritten while saving the value to the destination.
           // Therefore we can't use 'at'.  It is OK if the read from the source
           // destroys 'at', since that happens before the value is read.
@@ -271,6 +276,7 @@ void LGapResolver::EmitMove(int index) {
     }
 
   } else if (source->IsDoubleRegister()) {
+    CpuFeatures::Scope scope(FPU);
     DoubleRegister source_register = cgen_->ToDoubleRegister(source);
     if (destination->IsDoubleRegister()) {
       __ mov_d(cgen_->ToDoubleRegister(destination), source_register);
@@ -281,6 +287,7 @@ void LGapResolver::EmitMove(int index) {
     }
 
   } else if (source->IsDoubleStackSlot()) {
+    CpuFeatures::Scope scope(FPU);
     MemOperand source_operand = cgen_->ToMemOperand(source);
     if (destination->IsDoubleRegister()) {
       __ ldc1(cgen_->ToDoubleRegister(destination), source_operand);

src/mips/lithium-mips.cc

@@ -42,10 +42,10 @@ LITHIUM_CONCRETE_INSTRUCTION_LIST(DEFINE_COMPILE)
 #undef DEFINE_COMPILE
 
 LOsrEntry::LOsrEntry() {
-  for (int i = 0; i < Register::kNumAllocatableRegisters; ++i) {
+  for (int i = 0; i < Register::NumAllocatableRegisters(); ++i) {
     register_spills_[i] = NULL;
   }
-  for (int i = 0; i < DoubleRegister::kNumAllocatableRegisters; ++i) {
+  for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); ++i) {
     double_register_spills_[i] = NULL;
   }
 }
@@ -616,6 +616,7 @@ LInstruction* LChunkBuilder::AssignEnvironment(LInstruction* instr) {
 LInstruction* LChunkBuilder::MarkAsCall(LInstruction* instr,
                                         HInstruction* hinstr,
                                         CanDeoptimize can_deoptimize) {
+  info()->MarkAsNonDeferredCalling();
 #ifdef DEBUG
   instr->VerifyCall();
 #endif
@@ -1583,6 +1584,7 @@ LInstruction* LChunkBuilder::DoChange(HChange* instr) {
   Representation to = instr->to();
   if (from.IsTagged()) {
     if (to.IsDouble()) {
+      info()->MarkAsDeferredCalling();
       LOperand* value = UseRegister(instr->value());
       LNumberUntagD* res = new(zone()) LNumberUntagD(value);
       return AssignEnvironment(DefineAsRegister(res));
@@ -1607,6 +1609,7 @@ LInstruction* LChunkBuilder::DoChange(HChange* instr) {
     }
   } else if (from.IsDouble()) {
     if (to.IsTagged()) {
+      info()->MarkAsDeferredCalling();
       LOperand* value = UseRegister(instr->value());
       LOperand* temp1 = TempRegister();
       LOperand* temp2 = TempRegister();
@@ -1626,6 +1629,7 @@ LInstruction* LChunkBuilder::DoChange(HChange* instr) {
       return AssignEnvironment(DefineAsRegister(res));
     }
   } else if (from.IsInteger32()) {
+    info()->MarkAsDeferredCalling();
     if (to.IsTagged()) {
       HValue* val = instr->value();
       LOperand* value = UseRegisterAtStart(val);
@@ -1864,7 +1868,16 @@ LInstruction* LChunkBuilder::DoLoadKeyed(HLoadKeyed* instr) {
         (instr->representation().IsDouble() &&
          ((elements_kind == EXTERNAL_FLOAT_ELEMENTS) ||
           (elements_kind == EXTERNAL_DOUBLE_ELEMENTS))));
-    LOperand* external_pointer = UseRegister(instr->elements());
+    // float->double conversion on non-VFP2 requires an extra scratch
+    // register. For convenience, just mark the elements register as "UseTemp"
+    // so that it can be used as a temp during the float->double conversion
+    // after it's no longer needed after the float load.
+    bool needs_temp =
+        !CpuFeatures::IsSupported(FPU) &&
+        (elements_kind == EXTERNAL_FLOAT_ELEMENTS);
+    LOperand* external_pointer = needs_temp
+        ? UseTempRegister(instr->elements())
+        : UseRegister(instr->elements());
     result = new(zone()) LLoadKeyed(external_pointer, key);
   }
 
@@ -2083,8 +2096,17 @@ LInstruction* LChunkBuilder::DoOsrEntry(HOsrEntry* instr) {
 
 
 LInstruction* LChunkBuilder::DoParameter(HParameter* instr) {
+  LParameter* result = new(zone()) LParameter;
+  if (info()->IsOptimizing()) {
     int spill_index = chunk()->GetParameterStackSlot(instr->index());
-  return DefineAsSpilled(new(zone()) LParameter, spill_index);
+    return DefineAsSpilled(result, spill_index);
+  } else {
+    ASSERT(info()->IsStub());
+    CodeStubInterfaceDescriptor* descriptor =
+        info()->code_stub()->GetInterfaceDescriptor(info()->isolate());
+    Register reg = descriptor->register_params_[instr->index()];
+    return DefineFixed(result, reg);
+  }
 }
 
 

src/mips/lithium-mips.h

@@ -251,6 +251,11 @@ class LInstruction: public ZoneObject {
 
   void MarkAsCall() { is_call_ = true; }
 
+  // Interface to the register allocator and iterators.
+  bool ClobbersTemps() const { return is_call_; }
+  bool ClobbersRegisters() const { return is_call_; }
+  bool ClobbersDoubleRegisters() const { return is_call_; }
+
   // Interface to the register allocator and iterators.
   bool IsMarkedAsCall() const { return is_call_; }
 
@@ -2304,8 +2309,9 @@ class LOsrEntry: public LTemplateInstruction<0, 0, 0> {
   // slot, i.e., that must also be restored to the spill slot on OSR entry.
   // NULL if the register has no assigned spill slot.  Indexed by allocation
   // index.
-  LOperand* register_spills_[Register::kNumAllocatableRegisters];
-  LOperand* double_register_spills_[DoubleRegister::kNumAllocatableRegisters];
+  LOperand* register_spills_[Register::kMaxNumAllocatableRegisters];
+  LOperand* double_register_spills_[
+      DoubleRegister::kMaxNumAllocatableRegisters];
 };
 
 

src/mips/macro-assembler-mips.cc

@@ -118,8 +118,8 @@ void MacroAssembler::PopSafepointRegisters() {
 
 void MacroAssembler::PushSafepointRegistersAndDoubles() {
   PushSafepointRegisters();
-  Subu(sp, sp, Operand(FPURegister::kNumAllocatableRegisters * kDoubleSize));
-  for (int i = 0; i < FPURegister::kNumAllocatableRegisters; i+=2) {
+  Subu(sp, sp, Operand(FPURegister::NumAllocatableRegisters() * kDoubleSize));
+  for (int i = 0; i < FPURegister::NumAllocatableRegisters(); i+=2) {
     FPURegister reg = FPURegister::FromAllocationIndex(i);
     sdc1(reg, MemOperand(sp, i * kDoubleSize));
   }
@@ -127,11 +127,11 @@ void MacroAssembler::PushSafepointRegistersAndDoubles() {
 
 
 void MacroAssembler::PopSafepointRegistersAndDoubles() {
-  for (int i = 0; i < FPURegister::kNumAllocatableRegisters; i+=2) {
+  for (int i = 0; i < FPURegister::NumAllocatableRegisters(); i+=2) {
     FPURegister reg = FPURegister::FromAllocationIndex(i);
     ldc1(reg, MemOperand(sp, i * kDoubleSize));
   }
-  Addu(sp, sp, Operand(FPURegister::kNumAllocatableRegisters * kDoubleSize));
+  Addu(sp, sp, Operand(FPURegister::NumAllocatableRegisters() * kDoubleSize));
   PopSafepointRegisters();
 }
 
@@ -167,7 +167,7 @@ MemOperand MacroAssembler::SafepointRegisterSlot(Register reg) {
 MemOperand MacroAssembler::SafepointRegistersAndDoublesSlot(Register reg) {
   UNIMPLEMENTED_MIPS();
   // General purpose registers are pushed last on the stack.
-  int doubles_size = FPURegister::kNumAllocatableRegisters * kDoubleSize;
+  int doubles_size = FPURegister::NumAllocatableRegisters() * kDoubleSize;
   int register_offset = SafepointRegisterStackIndex(reg.code()) * kPointerSize;
   return MemOperand(sp, doubles_size + register_offset);
 }
@@ -4250,7 +4250,10 @@ void MacroAssembler::CallRuntimeSaveDoubles(Runtime::FunctionId id) {
   const Runtime::Function* function = Runtime::FunctionForId(id);
   PrepareCEntryArgs(function->nargs);
   PrepareCEntryFunction(ExternalReference(function, isolate()));
-  CEntryStub stub(1, kSaveFPRegs);
+  SaveFPRegsMode mode = CpuFeatures::IsSupported(FPU)
+      ? kSaveFPRegs
+      : kDontSaveFPRegs;
+  CEntryStub stub(1, mode);
   CallStub(&stub);
 }
 

src/mips/macro-assembler-mips.h

@@ -1514,9 +1514,9 @@ class MacroAssembler: public Assembler {
   // This handle will be patched with the code object on installation.
   Handle<Object> code_object_;
 
-  // Needs access to SafepointRegisterStackIndex for optimized frame
+  // Needs access to SafepointRegisterStackIndex for compiled frame
   // traversal.
-  friend class OptimizedFrame;
+  friend class StandardFrame;
 };
 
 

src/mips/stub-cache-mips.cc

@@ -1054,46 +1054,6 @@ static void StoreIntAsFloat(MacroAssembler* masm,
 }
 
 
-// Convert unsigned integer with specified number of leading zeroes in binary
-// representation to IEEE 754 double.
-// Integer to convert is passed in register hiword.
-// Resulting double is returned in registers hiword:loword.
-// This functions does not work correctly for 0.
-static void GenerateUInt2Double(MacroAssembler* masm,
-                                Register hiword,
-                                Register loword,
-                                Register scratch,
-                                int leading_zeroes) {
-  const int meaningful_bits = kBitsPerInt - leading_zeroes - 1;
-  const int biased_exponent = HeapNumber::kExponentBias + meaningful_bits;
-
-  const int mantissa_shift_for_hi_word =
-      meaningful_bits - HeapNumber::kMantissaBitsInTopWord;
-
-  const int mantissa_shift_for_lo_word =
-      kBitsPerInt - mantissa_shift_for_hi_word;
-
-  __ li(scratch, biased_exponent << HeapNumber::kExponentShift);
-  if (mantissa_shift_for_hi_word > 0) {
-    __ sll(loword, hiword, mantissa_shift_for_lo_word);
-    __ srl(hiword, hiword, mantissa_shift_for_hi_word);
-    __ or_(hiword, scratch, hiword);
-  } else {
-    __ mov(loword, zero_reg);
-    __ sll(hiword, hiword, mantissa_shift_for_hi_word);
-    __ or_(hiword, scratch, hiword);
-  }
-
-  // If least significant bit of biased exponent was not 1 it was corrupted
-  // by most significant bit of mantissa so we should fix that.
-  if (!(biased_exponent & 1)) {
-    __ li(scratch, 1 << HeapNumber::kExponentShift);
-    __ nor(scratch, scratch, scratch);
-    __ and_(hiword, hiword, scratch);
-  }
-}
-
-
 #undef __
 #define __ ACCESS_MASM(masm())
 
@@ -3316,9 +3276,17 @@ Handle<Code> KeyedLoadStubCompiler::CompileLoadElement(
   //  -- a1    : receiver
   // -----------------------------------
   ElementsKind elements_kind = receiver_map->elements_kind();
-  Handle<Code> stub = KeyedLoadElementStub(elements_kind).GetCode();
-
+  if (receiver_map->has_fast_elements() ||
+      receiver_map->has_external_array_elements()) {
+    Handle<Code> stub = KeyedLoadFastElementStub(
+        receiver_map->instance_type() == JS_ARRAY_TYPE,
+        elements_kind).GetCode();
+    __ DispatchMap(a1, a2, receiver_map, stub, DO_SMI_CHECK);
+  } else {
+    Handle<Code> stub =
+        KeyedLoadDictionaryElementStub().GetCode();
     __ DispatchMap(a1, a2, receiver_map, stub, DO_SMI_CHECK);
+  }
 
   Handle<Code> ic = isolate()->builtins()->KeyedLoadIC_Miss();
   __ Jump(ic, RelocInfo::CODE_TARGET);
@@ -3730,355 +3698,6 @@ static void GenerateSmiKeyCheck(MacroAssembler* masm,
 }
 
 
-void KeyedLoadStubCompiler::GenerateLoadExternalArray(
-    MacroAssembler* masm,
-    ElementsKind elements_kind) {
-  // ---------- S t a t e --------------
-  //  -- ra     : return address
-  //  -- a0     : key
-  //  -- a1     : receiver
-  // -----------------------------------
-  Label miss_force_generic, slow, failed_allocation;
-
-  Register key = a0;
-  Register receiver = a1;
-
-  // This stub is meant to be tail-jumped to, the receiver must already
-  // have been verified by the caller to not be a smi.
-
-  // Check that the key is a smi or a heap number convertible to a smi.
-  GenerateSmiKeyCheck(masm, key, t0, t1, f2, f4, &miss_force_generic);
-
-  __ lw(a3, FieldMemOperand(receiver, JSObject::kElementsOffset));
-  // a3: elements array
-
-  // Check that the index is in range.
-  __ lw(t1, FieldMemOperand(a3, ExternalArray::kLengthOffset));
-  __ sra(t2, key, kSmiTagSize);
-  // Unsigned comparison catches both negative and too-large values.
-  __ Branch(&miss_force_generic, Ugreater_equal, key, Operand(t1));
-
-  __ lw(a3, FieldMemOperand(a3, ExternalArray::kExternalPointerOffset));
-  // a3: base pointer of external storage
-
-  // We are not untagging smi key and instead work with it
-  // as if it was premultiplied by 2.
-  STATIC_ASSERT((kSmiTag == 0) && (kSmiTagSize == 1));
-
-  Register value = a2;
-  switch (elements_kind) {
-    case EXTERNAL_BYTE_ELEMENTS:
-      __ srl(t2, key, 1);
-      __ addu(t3, a3, t2);
-      __ lb(value, MemOperand(t3, 0));
-      break;
-    case EXTERNAL_PIXEL_ELEMENTS:
-    case EXTERNAL_UNSIGNED_BYTE_ELEMENTS:
-      __ srl(t2, key, 1);
-      __ addu(t3, a3, t2);
-      __ lbu(value, MemOperand(t3, 0));
-      break;
-    case EXTERNAL_SHORT_ELEMENTS:
-      __ addu(t3, a3, key);
-      __ lh(value, MemOperand(t3, 0));
-      break;
-    case EXTERNAL_UNSIGNED_SHORT_ELEMENTS:
-      __ addu(t3, a3, key);
-      __ lhu(value, MemOperand(t3, 0));
-      break;
-    case EXTERNAL_INT_ELEMENTS:
-    case EXTERNAL_UNSIGNED_INT_ELEMENTS:
-      __ sll(t2, key, 1);
-      __ addu(t3, a3, t2);
-      __ lw(value, MemOperand(t3, 0));
-      break;
-    case EXTERNAL_FLOAT_ELEMENTS:
-      __ sll(t3, t2, 2);
-      __ addu(t3, a3, t3);
-      if (CpuFeatures::IsSupported(FPU)) {
-        CpuFeatures::Scope scope(FPU);
-        __ lwc1(f0, MemOperand(t3, 0));
-      } else {
-        __ lw(value, MemOperand(t3, 0));
-      }
-      break;
-    case EXTERNAL_DOUBLE_ELEMENTS:
-      __ sll(t2, key, 2);
-      __ addu(t3, a3, t2);
-      if (CpuFeatures::IsSupported(FPU)) {
-        CpuFeatures::Scope scope(FPU);
-        __ ldc1(f0, MemOperand(t3, 0));
-      } else {
-        // t3: pointer to the beginning of the double we want to load.
-        __ lw(a2, MemOperand(t3, 0));
-        __ lw(a3, MemOperand(t3, Register::kSizeInBytes));
-      }
-      break;
-    case FAST_ELEMENTS:
-    case FAST_SMI_ELEMENTS:
-    case FAST_DOUBLE_ELEMENTS:
-    case FAST_HOLEY_ELEMENTS:
-    case FAST_HOLEY_SMI_ELEMENTS:
-    case FAST_HOLEY_DOUBLE_ELEMENTS:
-    case DICTIONARY_ELEMENTS:
-    case NON_STRICT_ARGUMENTS_ELEMENTS:
-      UNREACHABLE();
-      break;
-  }
-
-  // For integer array types:
-  // a2: value
-  // For float array type:
-  // f0: value (if FPU is supported)
-  // a2: value (if FPU is not supported)
-  // For double array type:
-  // f0: value (if FPU is supported)
-  // a2/a3: value (if FPU is not supported)
-
-  if (elements_kind == EXTERNAL_INT_ELEMENTS) {
-    // For the Int and UnsignedInt array types, we need to see whether
-    // the value can be represented in a Smi. If not, we need to convert
-    // it to a HeapNumber.
-    Label box_int;
-    __ Subu(t3, value, Operand(0xC0000000));  // Non-smi value gives neg result.
-    __ Branch(&box_int, lt, t3, Operand(zero_reg));
-    // Tag integer as smi and return it.
-    __ sll(v0, value, kSmiTagSize);
-    __ Ret();
-
-    __ bind(&box_int);
-
-    if (CpuFeatures::IsSupported(FPU)) {
-      CpuFeatures::Scope scope(FPU);
-      // Allocate a HeapNumber for the result and perform int-to-double
-      // conversion.
-      // The arm version uses a temporary here to save r0, but we don't need to
-      // (a0 is not modified).
-      __ LoadRoot(t1, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, a3, t0, t1, &slow, DONT_TAG_RESULT);
-      __ mtc1(value, f0);
-      __ cvt_d_w(f0, f0);
-      __ sdc1(f0, MemOperand(v0, HeapNumber::kValueOffset));
-      __ Addu(v0, v0, kHeapObjectTag);
-      __ Ret();
-    } else {
-      // Allocate a HeapNumber for the result and perform int-to-double
-      // conversion.
-      // The arm version uses a temporary here to save r0, but we don't need to
-      // (a0 is not modified).
-      __ LoadRoot(t1, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, a3, t0, t1, &slow, TAG_RESULT);
-      Register dst_mantissa = t2;
-      Register dst_exponent = t3;
-      FloatingPointHelper::Destination dest =
-          FloatingPointHelper::kCoreRegisters;
-      FloatingPointHelper::ConvertIntToDouble(masm,
-                                              value,
-                                              dest,
-                                              f0,
-                                              dst_mantissa,
-                                              dst_exponent,
-                                              t1,
-                                              f2);
-      __ sw(dst_mantissa, FieldMemOperand(v0, HeapNumber::kMantissaOffset));
-      __ sw(dst_exponent, FieldMemOperand(v0, HeapNumber::kExponentOffset));
-      __ Ret();
-    }
-  } else if (elements_kind == EXTERNAL_UNSIGNED_INT_ELEMENTS) {
-    // The test is different for unsigned int values. Since we need
-    // the value to be in the range of a positive smi, we can't
-    // handle either of the top two bits being set in the value.
-    if (CpuFeatures::IsSupported(FPU)) {
-      CpuFeatures::Scope scope(FPU);
-      Label pl_box_int;
-      __ And(t2, value, Operand(0xC0000000));
-      __ Branch(&pl_box_int, ne, t2, Operand(zero_reg));
-
-      // It can fit in an Smi.
-      // Tag integer as smi and return it.
-      __ sll(v0, value, kSmiTagSize);
-      __ Ret();
-
-      __ bind(&pl_box_int);
-      // Allocate a HeapNumber for the result and perform int-to-double
-      // conversion. Don't use a0 and a1 as AllocateHeapNumber clobbers all
-      // registers - also when jumping due to exhausted young space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, t2, t3, t6, &slow, DONT_TAG_RESULT);
-
-      // This is replaced by a macro:
-      // __ mtc1(value, f0);     // LS 32-bits.
-      // __ mtc1(zero_reg, f1);  // MS 32-bits are all zero.
-      // __ cvt_d_l(f0, f0); // Use 64 bit conv to get correct unsigned 32-bit.
-
-      __ Cvt_d_uw(f0, value, f22);
-
-      __ sdc1(f0, MemOperand(v0, HeapNumber::kValueOffset));
-
-      __ Addu(v0, v0, kHeapObjectTag);
-      __ Ret();
-    } else {
-      // Check whether unsigned integer fits into smi.
-      Label box_int_0, box_int_1, done;
-      __ And(t2, value, Operand(0x80000000));
-      __ Branch(&box_int_0, ne, t2, Operand(zero_reg));
-      __ And(t2, value, Operand(0x40000000));
-      __ Branch(&box_int_1, ne, t2, Operand(zero_reg));
-
-      // Tag integer as smi and return it.
-      __ sll(v0, value, kSmiTagSize);
-      __ Ret();
-
-      Register hiword = value;  // a2.
-      Register loword = a3;
-
-      __ bind(&box_int_0);
-      // Integer does not have leading zeros.
-      GenerateUInt2Double(masm, hiword, loword, t0, 0);
-      __ Branch(&done);
-
-      __ bind(&box_int_1);
-      // Integer has one leading zero.
-      GenerateUInt2Double(masm, hiword, loword, t0, 1);
-
-
-      __ bind(&done);
-      // Integer was converted to double in registers hiword:loword.
-      // Wrap it into a HeapNumber. Don't use a0 and a1 as AllocateHeapNumber
-      // clobbers all registers - also when jumping due to exhausted young
-      // space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(t2, t3, t5, t6, &slow, TAG_RESULT);
-
-      __ sw(hiword, FieldMemOperand(t2, HeapNumber::kExponentOffset));
-      __ sw(loword, FieldMemOperand(t2, HeapNumber::kMantissaOffset));
-
-      __ mov(v0, t2);
-      __ Ret();
-    }
-  } else if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
-    // For the floating-point array type, we need to always allocate a
-    // HeapNumber.
-    if (CpuFeatures::IsSupported(FPU)) {
-      CpuFeatures::Scope scope(FPU);
-      // Allocate a HeapNumber for the result. Don't use a0 and a1 as
-      // AllocateHeapNumber clobbers all registers - also when jumping due to
-      // exhausted young space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, t3, t5, t6, &slow, DONT_TAG_RESULT);
-      // The float (single) value is already in fpu reg f0 (if we use float).
-      __ cvt_d_s(f0, f0);
-      __ sdc1(f0, MemOperand(v0, HeapNumber::kValueOffset));
-
-      __ Addu(v0, v0, kHeapObjectTag);
-      __ Ret();
-    } else {
-      // Allocate a HeapNumber for the result. Don't use a0 and a1 as
-      // AllocateHeapNumber clobbers all registers - also when jumping due to
-      // exhausted young space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, t3, t5, t6, &slow, TAG_RESULT);
-      // FPU is not available, do manual single to double conversion.
-
-      // a2: floating point value (binary32).
-      // v0: heap number for result
-
-      // Extract mantissa to t4.
-      __ And(t4, value, Operand(kBinary32MantissaMask));
-
-      // Extract exponent to t5.
-      __ srl(t5, value, kBinary32MantissaBits);
-      __ And(t5, t5, Operand(kBinary32ExponentMask >> kBinary32MantissaBits));
-
-      Label exponent_rebiased;
-      __ Branch(&exponent_rebiased, eq, t5, Operand(zero_reg));
-
-      __ li(t0, 0x7ff);
-      __ Xor(t1, t5, Operand(0xFF));
-      __ Movz(t5, t0, t1);  // Set t5 to 0x7ff only if t5 is equal to 0xff.
-      __ Branch(&exponent_rebiased, eq, t1, Operand(zero_reg));
-
-      // Rebias exponent.
-      __ Addu(t5,
-              t5,
-              Operand(-kBinary32ExponentBias + HeapNumber::kExponentBias));
-
-      __ bind(&exponent_rebiased);
-      __ And(a2, value, Operand(kBinary32SignMask));
-      value = no_reg;
-      __ sll(t0, t5, HeapNumber::kMantissaBitsInTopWord);
-      __ or_(a2, a2, t0);
-
-      // Shift mantissa.
-      static const int kMantissaShiftForHiWord =
-          kBinary32MantissaBits - HeapNumber::kMantissaBitsInTopWord;
-
-      static const int kMantissaShiftForLoWord =
-          kBitsPerInt - kMantissaShiftForHiWord;
-
-      __ srl(t0, t4, kMantissaShiftForHiWord);
-      __ or_(a2, a2, t0);
-      __ sll(a0, t4, kMantissaShiftForLoWord);
-
-      __ sw(a2, FieldMemOperand(v0, HeapNumber::kExponentOffset));
-      __ sw(a0, FieldMemOperand(v0, HeapNumber::kMantissaOffset));
-      __ Ret();
-    }
-
-  } else if (elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
-    if (CpuFeatures::IsSupported(FPU)) {
-      CpuFeatures::Scope scope(FPU);
-      // Allocate a HeapNumber for the result. Don't use a0 and a1 as
-      // AllocateHeapNumber clobbers all registers - also when jumping due to
-      // exhausted young space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, t3, t5, t6, &slow, DONT_TAG_RESULT);
-      // The double value is already in f0
-      __ sdc1(f0, MemOperand(v0, HeapNumber::kValueOffset));
-
-      __ Addu(v0, v0, kHeapObjectTag);
-      __ Ret();
-    } else {
-      // Allocate a HeapNumber for the result. Don't use a0 and a1 as
-      // AllocateHeapNumber clobbers all registers - also when jumping due to
-      // exhausted young space.
-      __ LoadRoot(t6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(v0, t3, t5, t6, &slow, TAG_RESULT);
-
-      __ sw(a2, FieldMemOperand(v0, HeapNumber::kMantissaOffset));
-      __ sw(a3, FieldMemOperand(v0, HeapNumber::kExponentOffset));
-      __ Ret();
-    }
-
-  } else {
-    // Tag integer as smi and return it.
-    __ sll(v0, value, kSmiTagSize);
-    __ Ret();
-  }
-
-  // Slow case, key and receiver still in a0 and a1.
-  __ bind(&slow);
-  __ IncrementCounter(
-      masm->isolate()->counters()->keyed_load_external_array_slow(),
-      1, a2, a3);
-
-  // ---------- S t a t e --------------
-  //  -- ra     : return address
-  //  -- a0     : key
-  //  -- a1     : receiver
-  // -----------------------------------
-
-  __ Push(a1, a0);
-
-  __ TailCallRuntime(Runtime::kKeyedGetProperty, 2, 1);
-
-  __ bind(&miss_force_generic);
-  Handle<Code> stub =
-      masm->isolate()->builtins()->KeyedLoadIC_MissForceGeneric();
-  __ Jump(stub, RelocInfo::CODE_TARGET);
-}
-
-
 void KeyedStoreStubCompiler::GenerateStoreExternalArray(
     MacroAssembler* masm,
     ElementsKind elements_kind) {
@@ -4478,115 +4097,6 @@ void KeyedStoreStubCompiler::GenerateStoreExternalArray(
 }
 
 
-void KeyedLoadStubCompiler::GenerateLoadFastElement(MacroAssembler* masm) {
-  // ----------- S t a t e -------------
-  //  -- ra    : return address
-  //  -- a0    : key
-  //  -- a1    : receiver
-  // -----------------------------------
-  Label miss_force_generic;
-
-  // This stub is meant to be tail-jumped to, the receiver must already
-  // have been verified by the caller to not be a smi.
-
-  // Check that the key is a smi or a heap number convertible to a smi.
-  GenerateSmiKeyCheck(masm, a0, t0, t1, f2, f4, &miss_force_generic);
-
-  // Get the elements array.
-  __ lw(a2, FieldMemOperand(a1, JSObject::kElementsOffset));
-  __ AssertFastElements(a2);
-
-  // Check that the key is within bounds.
-  __ lw(a3, FieldMemOperand(a2, FixedArray::kLengthOffset));
-  __ Branch(USE_DELAY_SLOT, &miss_force_generic, hs, a0, Operand(a3));
-
-  // Load the result and make sure it's not the hole.
-  __ Addu(a3, a2, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
-  STATIC_ASSERT(kSmiTag == 0 && kSmiTagSize < kPointerSizeLog2);
-  __ sll(t0, a0, kPointerSizeLog2 - kSmiTagSize);
-  __ Addu(t0, t0, a3);
-  __ lw(t0, MemOperand(t0));
-  __ LoadRoot(t1, Heap::kTheHoleValueRootIndex);
-  __ Branch(&miss_force_generic, eq, t0, Operand(t1));
-  __ Ret(USE_DELAY_SLOT);
-  __ mov(v0, t0);
-
-  __ bind(&miss_force_generic);
-  Handle<Code> stub =
-      masm->isolate()->builtins()->KeyedLoadIC_MissForceGeneric();
-  __ Jump(stub, RelocInfo::CODE_TARGET);
-}
-
-
-void KeyedLoadStubCompiler::GenerateLoadFastDoubleElement(
-    MacroAssembler* masm) {
-  // ----------- S t a t e -------------
-  //  -- ra    : return address
-  //  -- a0    : key
-  //  -- a1    : receiver
-  // -----------------------------------
-  Label miss_force_generic, slow_allocate_heapnumber;
-
-  Register key_reg = a0;
-  Register receiver_reg = a1;
-  Register elements_reg = a2;
-  Register heap_number_reg = a2;
-  Register indexed_double_offset = a3;
-  Register scratch = t0;
-  Register scratch2 = t1;
-  Register scratch3 = t2;
-  Register heap_number_map = t3;
-
-  // This stub is meant to be tail-jumped to, the receiver must already
-  // have been verified by the caller to not be a smi.
-
-  // Check that the key is a smi or a heap number convertible to a smi.
-  GenerateSmiKeyCheck(masm, key_reg, t0, t1, f2, f4, &miss_force_generic);
-
-  // Get the elements array.
-  __ lw(elements_reg,
-        FieldMemOperand(receiver_reg, JSObject::kElementsOffset));
-
-  // Check that the key is within bounds.
-  __ lw(scratch, FieldMemOperand(elements_reg, FixedArray::kLengthOffset));
-  __ Branch(&miss_force_generic, hs, key_reg, Operand(scratch));
-
-  // Load the upper word of the double in the fixed array and test for NaN.
-  __ sll(scratch2, key_reg, kDoubleSizeLog2 - kSmiTagSize);
-  __ Addu(indexed_double_offset, elements_reg, Operand(scratch2));
-  uint32_t upper_32_offset = FixedArray::kHeaderSize + sizeof(kHoleNanLower32);
-  __ lw(scratch, FieldMemOperand(indexed_double_offset, upper_32_offset));
-  __ Branch(&miss_force_generic, eq, scratch, Operand(kHoleNanUpper32));
-
-  // Non-NaN. Allocate a new heap number and copy the double value into it.
-  __ LoadRoot(heap_number_map, Heap::kHeapNumberMapRootIndex);
-  __ AllocateHeapNumber(heap_number_reg, scratch2, scratch3,
-                        heap_number_map, &slow_allocate_heapnumber, TAG_RESULT);
-
-  // Don't need to reload the upper 32 bits of the double, it's already in
-  // scratch.
-  __ sw(scratch, FieldMemOperand(heap_number_reg,
-                                 HeapNumber::kExponentOffset));
-  __ lw(scratch, FieldMemOperand(indexed_double_offset,
-                                 FixedArray::kHeaderSize));
-  __ sw(scratch, FieldMemOperand(heap_number_reg,
-                                 HeapNumber::kMantissaOffset));
-
-  __ mov(v0, heap_number_reg);
-  __ Ret();
-
-  __ bind(&slow_allocate_heapnumber);
-  Handle<Code> slow_ic =
-      masm->isolate()->builtins()->KeyedLoadIC_Slow();
-  __ Jump(slow_ic, RelocInfo::CODE_TARGET);
-
-  __ bind(&miss_force_generic);
-  Handle<Code> miss_ic =
-      masm->isolate()->builtins()->KeyedLoadIC_MissForceGeneric();
-  __ Jump(miss_ic, RelocInfo::CODE_TARGET);
-}
-
-
 void KeyedStoreStubCompiler::GenerateStoreFastElement(
     MacroAssembler* masm,
     bool is_js_array,