MIPS: Re-land Crankshaft-generated KeyedLoad stubs.

src/mips/lithium-codegen-mips.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 59 matching lines @@
   CodeStub::GenerateFPStubs();
 
   // Open a frame scope to indicate that there is a frame on the stack.  The
   // NONE indicates that the scope shouldn't actually generate code to set up
   // the frame (that is done in GeneratePrologue).
   FrameScope frame_scope(masm_, StackFrame::NONE);
 
   return GeneratePrologue() &&
       GenerateBody() &&
       GenerateDeferredCode() &&
+      GenerateDeoptJumpTable() &&
       GenerateSafepointTable();
 }
 
 
 void LCodeGen::FinishCode(Handle<Code> code) {
   ASSERT(is_done());
   code->set_stack_slots(GetStackSlotCount());
   code->set_safepoint_table_offset(safepoints_.GetCodeOffset());
   PopulateDeoptimizationData(code);
 }
@@ skipping 19 matching lines @@
   size_t length = builder.position();
   Vector<char> copy = Vector<char>::New(length + 1);
   memcpy(copy.start(), builder.Finalize(), copy.length());
   masm()->RecordComment(copy.start());
 }
 
 
 bool LCodeGen::GeneratePrologue() {
   ASSERT(is_generating());
 
-  ProfileEntryHookStub::MaybeCallEntryHook(masm_);
+  if (info()->IsOptimizing()) {
+    ProfileEntryHookStub::MaybeCallEntryHook(masm_);
 
 #ifdef DEBUG
-  if (strlen(FLAG_stop_at) > 0 &&
-      info_->function()->name()->IsEqualTo(CStrVector(FLAG_stop_at))) {
-    __ stop("stop_at");
-  }
+    if (strlen(FLAG_stop_at) > 0 &&
+        info_->function()->name()->IsEqualTo(CStrVector(FLAG_stop_at))) {
+      __ stop("stop_at");
+    }
 #endif
 
-  // a1: Callee's JS function.
-  // cp: Callee's context.
-  // fp: Caller's frame pointer.
-  // lr: Caller's pc.
+    // a1: Callee's JS function.
+    // cp: Callee's context.
+    // fp: Caller's frame pointer.
+    // lr: Caller's pc.
 
-  // Strict mode functions and builtins need to replace the receiver
-  // with undefined when called as functions (without an explicit
-  // receiver object). r5 is zero for method calls and non-zero for
-  // function calls.
-  if (!info_->is_classic_mode() || info_->is_native()) {
-    Label ok;
-    __ Branch(&ok, eq, t1, Operand(zero_reg));
+    // Strict mode functions and builtins need to replace the receiver
+    // with undefined when called as functions (without an explicit
+    // receiver object). r5 is zero for method calls and non-zero for
+    // function calls.
+    if (!info_->is_classic_mode() || info_->is_native()) {
+      Label ok;
+      __ Branch(&ok, eq, t1, Operand(zero_reg));
 
-    int receiver_offset = scope()->num_parameters() * kPointerSize;
-    __ LoadRoot(a2, Heap::kUndefinedValueRootIndex);
-    __ sw(a2, MemOperand(sp, receiver_offset));
-    __ bind(&ok);
+      int receiver_offset = scope()->num_parameters() * kPointerSize;
+      __ LoadRoot(a2, Heap::kUndefinedValueRootIndex);
+      __ sw(a2, MemOperand(sp, receiver_offset));
+      __ bind(&ok);
+    }
   }
 
   info()->set_prologue_offset(masm_->pc_offset());
-  // 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.
+  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);
+    // 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();
   if (slots > 0) {
     if (FLAG_debug_code) {
       __ li(a0, Operand(slots));
       __ li(a2, Operand(kSlotsZapValue));
       Label loop;
       __ bind(&loop);
       __ push(a2);
       __ Subu(a0, a0, 1);
       __ Branch(&loop, ne, a0, Operand(zero_reg));
     } else {
       __ Subu(sp, sp, Operand(slots * kPointerSize));
     }
   }
 
   // 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.
     __ push(a1);
     if (heap_slots <= FastNewContextStub::kMaximumSlots) {
       FastNewContextStub stub(heap_slots);
       __ CallStub(&stub);
     } else {
       __ CallRuntime(Runtime::kNewFunctionContext, 1);
     }
@@ skipping 15 matching lines @@
         __ sw(a0, target);
         // Update the write barrier. This clobbers a3 and a0.
         __ RecordWriteContextSlot(
             cp, target.offset(), a0, a3, kRAHasBeenSaved, kSaveFPRegs);
       }
     }
     Comment(";;; End allocate local context");
   }
 
   // Trace the call.
-  if (FLAG_trace) {
+  if (FLAG_trace && info()->IsOptimizing()) {
     __ CallRuntime(Runtime::kTraceEnter, 0);
   }
   EnsureSpaceForLazyDeopt();
   return !is_aborted();
 }
 
 
 bool LCodeGen::GenerateBody() {
   ASSERT(is_generating());
   bool emit_instructions = true;
@@ skipping 37 matching lines @@
   return !is_aborted();
 }
 
 
 bool LCodeGen::GenerateDeferredCode() {
   ASSERT(is_generating());
   if (deferred_.length() > 0) {
     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());
     }
   }
   // Deferred code 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();
 }
 
 
 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();
 }
 
 
 bool LCodeGen::GenerateSafepointTable() {
   ASSERT(is_done());
   safepoints_.Emit(masm(), GetStackSlotCount());
   return !is_aborted();
 }
 
 
@@ skipping 179 matching lines @@
   // arguments index points to the first element of a sequence of tagged
   // values on the stack that represent the arguments. This needs to be
   // kept in sync with the LArgumentsElements implementation.
   *arguments_index = -environment->parameter_count();
   *arguments_count = environment->parameter_count();
 
   WriteTranslation(environment->outer(),
                    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;
 
   switch (environment->frame_type()) {
     case JS_FUNCTION:
       translation->BeginJSFrame(environment->ast_id(), closure_id, height);
       break;
     case JS_CONSTRUCT:
       translation->BeginConstructStubFrame(closure_id, translation_size);
       break;
     case JS_GETTER:
       ASSERT(translation_size == 1);
       ASSERT(height == 0);
       translation->BeginGetterStubFrame(closure_id);
       break;
     case JS_SETTER:
       ASSERT(translation_size == 2);
       ASSERT(height == 0);
       translation->BeginSetterStubFrame(closure_id);
       break;
+    case STUB:
+      translation->BeginCompiledStubFrame();
+      break;
     case ARGUMENTS_ADAPTOR:
       translation->BeginArgumentsAdaptorFrame(closure_id, translation_size);
       break;
   }
 
   // Inlined frames which push their arguments cause the index to be
   // bumped and a new stack area to be used for materialization.
   if (environment->entry() != NULL &&
       environment->entry()->arguments_pushed()) {
     *arguments_index = *arguments_index < 0
@@ skipping 166 matching lines @@
 }
 
 
 void LCodeGen::DeoptimizeIf(Condition cc,
                             LEnvironment* environment,
                             Register src1,
                             const Operand& src2) {
   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;
   }
 
   ASSERT(FLAG_deopt_every_n_times < 2);  // Other values not supported on MIPS.
 
   if (FLAG_deopt_every_n_times == 1 &&
       info_->shared_info()->opt_count() == id) {
     __ Jump(entry, RelocInfo::RUNTIME_ENTRY);
     return;
   }
 
   if (FLAG_trap_on_deopt) {
     Label skip;
     if (cc != al) {
       __ Branch(&skip, NegateCondition(cc), src1, src2);
     }
     __ stop("trap_on_deopt");
     __ 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.
-  __ Jump(entry, RelocInfo::RUNTIME_ENTRY, cc, src1, src2);
+  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);
+  }
 }
 
 
 void LCodeGen::PopulateDeoptimizationData(Handle<Code> code) {
   int length = deoptimizations_.length();
   if (length == 0) return;
   Handle<DeoptimizationInputData> data =
       factory()->NewDeoptimizationInputData(length, TENURED);
 
   Handle<ByteArray> translations = translations_.CreateByteArray();
@@ skipping 546 matching lines @@
 
 void LCodeGen::DoConstantI(LConstantI* instr) {
   ASSERT(instr->result()->IsRegister());
   __ li(ToRegister(instr->result()), Operand(instr->value()));
 }
 
 
 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);
 }
 
 
 void LCodeGen::DoConstantT(LConstantT* instr) {
   Handle<Object> value = instr->value();
   if (value->IsSmi()) {
     __ li(ToRegister(instr->result()), Operand(value));
   } else {
@@ skipping 177 matching lines @@
       __ Branch(&return_right, NegateCondition(condition), left_reg, right_op);
       __ mov(result_reg, left_reg);
       __ Branch(&done);
     }
     __ Branch(&done, condition, left_reg, right_op);
     __ bind(&return_right);
     __ Addu(result_reg, zero_reg, right_op);
     __ 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());
     Label check_nan_left, check_zero, return_left, return_right, done;
     __ BranchF(&check_zero, &check_nan_left, eq, left_reg, right_reg);
     __ BranchF(&return_left, NULL, condition, left_reg, right_reg);
     __ Branch(&return_right);
 
     __ bind(&check_zero);
     // left == right != 0.
@@ skipping 20 matching lines @@
     __ bind(&return_left);
     if (!left_reg.is(result_reg)) {
       __ mov_d(result_reg, left_reg);
     }
     __ bind(&done);
   }
 }
 
 
 void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister left = ToDoubleRegister(instr->left());
   DoubleRegister right = ToDoubleRegister(instr->right());
   DoubleRegister result = ToDoubleRegister(instr->result());
   switch (instr->op()) {
     case Token::ADD:
       __ add_d(result, left, right);
       break;
     case Token::SUB:
       __ sub_d(result, left, right);
       break;
@@ skipping 89 matching lines @@
 
 void LCodeGen::DoBranch(LBranch* instr) {
   int true_block = chunk_->LookupDestination(instr->true_block_id());
   int false_block = chunk_->LookupDestination(instr->false_block_id());
 
   Representation r = instr->hydrogen()->value()->representation();
   if (r.IsInteger32()) {
     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);
   } else {
     ASSERT(r.IsTagged());
     Register reg = ToRegister(instr->value());
     HType type = instr->hydrogen()->value()->type();
     if (type.IsBoolean()) {
       __ LoadRoot(at, Heap::kTrueValueRootIndex);
       EmitBranch(true_block, false_block, eq, reg, Operand(at));
@@ skipping 57 matching lines @@
         Label not_string;
         __ lbu(at, FieldMemOperand(map, Map::kInstanceTypeOffset));
         __ Branch(&not_string, ge , at, Operand(FIRST_NONSTRING_TYPE));
         __ lw(at, FieldMemOperand(reg, String::kLengthOffset));
         __ Branch(true_label, ne, at, Operand(zero_reg));
         __ Branch(false_label);
         __ bind(&not_string);
       }
 
       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;
         __ LoadRoot(at, Heap::kHeapNumberMapRootIndex);
         __ Branch(&not_heap_number, ne, map, Operand(at));
         __ ldc1(dbl_scratch, FieldMemOperand(reg, HeapNumber::kValueOffset));
         __ BranchF(true_label, false_label, ne, dbl_scratch, kDoubleRegZero);
         // Falls through if dbl_scratch == 0.
         __ Branch(false_label);
         __ bind(&not_heap_number);
@@ skipping 59 matching lines @@
   if (left->IsConstantOperand() && right->IsConstantOperand()) {
     // We can statically evaluate the comparison.
     double left_val = ToDouble(LConstantOperand::cast(left));
     double right_val = ToDouble(LConstantOperand::cast(right));
     int next_block =
       EvalComparison(instr->op(), left_val, right_val) ? true_block
                                                        : false_block;
     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);
       FPURegister right_reg = ToDoubleRegister(right);
 
       // If a NaN is involved, i.e. the result is unordered,
       // jump to false block label.
       __ BranchF(NULL, chunk_->GetAssemblyLabel(false_block), eq,
                  left_reg, right_reg);
 
@@ skipping 527 matching lines @@
   Label done;
   __ Branch(USE_DELAY_SLOT, &done, condition, v0, Operand(zero_reg));
   __ LoadRoot(ToRegister(instr->result()), Heap::kTrueValueRootIndex);
   __ LoadRoot(ToRegister(instr->result()), Heap::kFalseValueRootIndex);
   ASSERT_EQ(3, masm()->InstructionsGeneratedSince(&done));
   __ bind(&done);
 }
 
 
 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);
   }
-  int32_t sp_delta = (GetParameterCount() + 1) * kPointerSize;
-  __ mov(sp, fp);
-  __ Pop(ra, fp);
-  __ Addu(sp, sp, Operand(sp_delta));
+  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);
 }
 
 
 void LCodeGen::DoLoadGlobalCell(LLoadGlobalCell* instr) {
   Register result = ToRegister(instr->result());
   __ li(at, Operand(Handle<Object>(instr->hydrogen()->cell())));
   __ lw(result, FieldMemOperand(at, JSGlobalPropertyCell::kValueOffset));
   if (instr->hydrogen()->RequiresHoleCheck()) {
     __ LoadRoot(at, Heap::kTheHoleValueRootIndex);
@@ skipping 338 matching lines @@
 
   if (elements_kind == EXTERNAL_FLOAT_ELEMENTS ||
       elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
     FPURegister result = ToDoubleRegister(instr->result());
     if (key_is_constant) {
       __ Addu(scratch0(), external_pointer, constant_key << element_size_shift);
     } else {
       __ 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));
 
-    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));
+        __ 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(
         key, external_pointer, key_is_constant, constant_key,
         element_size_shift, shift_size,
         instr->additional_index(), additional_offset);
     switch (elements_kind) {
       case EXTERNAL_BYTE_ELEMENTS:
         __ lb(result, mem_operand);
@@ skipping 48 matching lines @@
   int constant_key = 0;
   if (key_is_constant) {
     constant_key = ToInteger32(LConstantOperand::cast(instr->key()));
     if (constant_key & 0xF0000000) {
       Abort("array index constant value too big.");
     }
   } else {
     key = ToRegister(instr->key());
   }
 
-  if (key_is_constant) {
-    __ Addu(elements, elements,
-        Operand(((constant_key + instr->additional_index()) <<
-                 element_size_shift) +
-                FixedDoubleArray::kHeaderSize - kHeapObjectTag));
+  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, 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));
+    }
   } else {
-    __ sll(scratch, key, shift_size);
-    __ Addu(elements, elements, Operand(scratch));
-    __ Addu(elements, elements,
-            Operand((FixedDoubleArray::kHeaderSize - kHeapObjectTag) +
-                    (instr->additional_index() << element_size_shift)));
+      __ 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));
+    }
   }
-
-  if (instr->hydrogen()->RequiresHoleCheck()) {
-    __ lw(scratch, MemOperand(elements, sizeof(kHoleNanLower32)));
-    DeoptimizeIf(eq, instr->environment(), scratch, Operand(kHoleNanUpper32));
-  }
-
-  __ ldc1(result, MemOperand(elements));
 }
 
 
 void LCodeGen::DoLoadKeyedFixedArray(LLoadKeyed* instr) {
   Register elements = ToRegister(instr->elements());
   Register result = ToRegister(instr->result());
   Register scratch = scratch0();
   Register store_base = scratch;
   int offset = 0;
 
@@ skipping 422 matching lines @@
   __ mov(result, input);
   ASSERT_EQ(2, masm()->InstructionsGeneratedSince(&done));
   __ subu(result, zero_reg, input);
   // Overflow if result is still negative, i.e. 0x80000000.
   DeoptimizeIf(lt, instr->environment(), result, Operand(zero_reg));
   __ bind(&done);
 }
 
 
 void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   // Class for deferred case.
   class DeferredMathAbsTaggedHeapNumber: public LDeferredCode {
    public:
     DeferredMathAbsTaggedHeapNumber(LCodeGen* codegen,
                                     LUnaryMathOperation* instr)
         : LDeferredCode(codegen), instr_(instr) { }
     virtual void Generate() {
       codegen()->DoDeferredMathAbsTaggedHeapNumber(instr_);
     }
     virtual LInstruction* instr() { return instr_; }
@@ skipping 16 matching lines @@
     // Smi check.
     __ JumpIfNotSmi(input, deferred->entry());
     // If smi, handle it directly.
     EmitIntegerMathAbs(instr);
     __ bind(deferred->exit());
   }
 }
 
 
 void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   Register result = ToRegister(instr->result());
   Register scratch1 = scratch0();
   Register except_flag = ToRegister(instr->temp());
 
   __ EmitFPUTruncate(kRoundToMinusInf,
                      result,
                      input,
                      scratch1,
                      double_scratch0(),
                      except_flag);
 
   // Deopt if the operation did not succeed.
   DeoptimizeIf(ne, instr->environment(), except_flag, Operand(zero_reg));
 
   if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
     // Test for -0.
     Label done;
     __ Branch(&done, ne, result, Operand(zero_reg));
     __ mfc1(scratch1, input.high());
     __ And(scratch1, scratch1, Operand(HeapNumber::kSignMask));
     DeoptimizeIf(ne, instr->environment(), scratch1, Operand(zero_reg));
     __ bind(&done);
   }
 }
 
 
 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());
   Register scratch = scratch0();
   Label done, check_sign_on_zero;
 
   // Extract exponent bits.
   __ mfc1(result, input.high());
   __ Ext(scratch,
          result,
@@ skipping 56 matching lines @@
     __ bind(&check_sign_on_zero);
     __ mfc1(scratch, input.high());
     __ And(scratch, scratch, Operand(HeapNumber::kSignMask));
     DeoptimizeIf(ne, instr->environment(), scratch, Operand(zero_reg));
   }
   __ bind(&done);
 }
 
 
 void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());
   __ sqrt_d(result, input);
 }
 
 
 void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());
   DoubleRegister temp = ToDoubleRegister(instr->temp());
 
   ASSERT(!input.is(result));
 
   // Note that according to ECMA-262 15.8.2.13:
   // Math.pow(-Infinity, 0.5) == Infinity
   // Math.sqrt(-Infinity) == NaN
   Label done;
   __ Move(temp, -V8_INFINITY);
   __ BranchF(USE_DELAY_SLOT, &done, NULL, eq, temp, input);
   // Set up Infinity in the delay slot.
   // result is overwritten if the branch is not taken.
   __ neg_d(result, temp);
 
   // Add +0 to convert -0 to +0.
   __ add_d(result, input, kDoubleRegZero);
   __ sqrt_d(result, result);
   __ bind(&done);
 }
 
 
 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.
   ASSERT(!instr->right()->IsDoubleRegister() ||
          ToDoubleRegister(instr->right()).is(f4));
   ASSERT(!instr->right()->IsRegister() ||
          ToRegister(instr->right()).is(a2));
   ASSERT(ToDoubleRegister(instr->left()).is(f2));
   ASSERT(ToDoubleRegister(instr->result()).is(f0));
 
@@ skipping 10 matching lines @@
     __ CallStub(&stub);
   } else {
     ASSERT(exponent_type.IsDouble());
     MathPowStub stub(MathPowStub::DOUBLE);
     __ CallStub(&stub);
   }
 }
 
 
 void LCodeGen::DoRandom(LRandom* instr) {
+  CpuFeatures::Scope scope(FPU);
   class DeferredDoRandom: public LDeferredCode {
    public:
     DeferredDoRandom(LCodeGen* codegen, LRandom* instr)
         : LDeferredCode(codegen), instr_(instr) { }
     virtual void Generate() { codegen()->DoDeferredRandom(instr_); }
     virtual LInstruction* instr() { return instr_; }
    private:
     LRandom* instr_;
   };
 
@@ skipping 56 matching lines @@
 }
 
 void LCodeGen::DoDeferredRandom(LRandom* instr) {
   __ PrepareCallCFunction(1, scratch0());
   __ CallCFunction(ExternalReference::random_uint32_function(isolate()), 1);
   // Return value is in v0.
 }
 
 
 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());
   DoubleRegister double_scratch2 = double_scratch0();
   Register temp1 = ToRegister(instr->temp1());
   Register temp2 = ToRegister(instr->temp2());
 
   MathExpGenerator::EmitMathExp(
       masm(), input, result, double_scratch1, double_scratch2,
       temp1, temp2, scratch0());
@@ skipping 275 matching lines @@
   } else {
     DeoptimizeIf(hs,
                  instr->environment(),
                  ToRegister(instr->index()),
                  Operand(ToRegister(instr->length())));
   }
 }
 
 
 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();
   bool key_is_constant = instr->key()->IsConstantOperand();
   int constant_key = 0;
   if (key_is_constant) {
     constant_key = ToInteger32(LConstantOperand::cast(instr->key()));
     if (constant_key & 0xF0000000) {
       Abort("array index constant value too big.");
     }
@@ skipping 53 matching lines @@
       case DICTIONARY_ELEMENTS:
       case NON_STRICT_ARGUMENTS_ELEMENTS:
         UNREACHABLE();
         break;
     }
   }
 }
 
 
 void LCodeGen::DoStoreKeyedFixedDoubleArray(LStoreKeyed* instr) {
+  CpuFeatures::Scope scope(FPU);
   DoubleRegister value = ToDoubleRegister(instr->value());
   Register elements = ToRegister(instr->elements());
   Register key = no_reg;
   Register scratch = scratch0();
   bool key_is_constant = instr->key()->IsConstantOperand();
   int constant_key = 0;
   Label not_nan;
 
   // Calculate the effective address of the slot in the array to store the
   // double value.
@@ skipping 263 matching lines @@
 
 
 void LCodeGen::DoStringLength(LStringLength* instr) {
   Register string = ToRegister(instr->string());
   Register result = ToRegister(instr->result());
   __ lw(result, FieldMemOperand(string, String::kLengthOffset));
 }
 
 
 void LCodeGen::DoInteger32ToDouble(LInteger32ToDouble* instr) {
+  CpuFeatures::Scope scope(FPU);
   LOperand* input = instr->value();
   ASSERT(input->IsRegister() || input->IsStackSlot());
   LOperand* output = instr->result();
   ASSERT(output->IsDoubleRegister());
   FPURegister single_scratch = double_scratch0().low();
   if (input->IsStackSlot()) {
     Register scratch = scratch0();
     __ lw(scratch, ToMemOperand(input));
     __ mtc1(scratch, single_scratch);
   } else {
     __ mtc1(ToRegister(input), single_scratch);
   }
   __ cvt_d_w(ToDoubleRegister(output), single_scratch);
 }
 
 
 void LCodeGen::DoUint32ToDouble(LUint32ToDouble* instr) {
+  CpuFeatures::Scope scope(FPU);
   LOperand* input = instr->value();
   LOperand* output = instr->result();
 
   FPURegister dbl_scratch = double_scratch0();
   __ mtc1(ToRegister(input), dbl_scratch);
   __ Cvt_d_uw(ToDoubleRegister(output), dbl_scratch, f22);
 }
 
 
 void LCodeGen::DoNumberTagI(LNumberTagI* instr) {
@@ skipping 41 matching lines @@
   ASSERT(input->IsRegister() && input->Equals(instr->result()));
   Register reg = ToRegister(input);
 
   DeferredNumberTagU* deferred = new(zone()) DeferredNumberTagU(this, instr);
   __ Branch(deferred->entry(), hi, reg, Operand(Smi::kMaxValue));
   __ SmiTag(reg, reg);
   __ bind(deferred->exit());
 }
 
 
+// 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);
 
   Label done;
   if (signedness == SIGNED_INT32) {
     // There was overflow, so bits 30 and 31 of the original integer
     // disagree. Try to allocate a heap number in new space and store
     // the value in there. If that fails, call the runtime system.
     if (dst.is(src)) {
       __ SmiUntag(src, dst);
       __ Xor(src, src, Operand(0x80000000));
     }
-    __ mtc1(src, dbl_scratch);
-    __ cvt_d_w(dbl_scratch, dbl_scratch);
+    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 {
-    __ mtc1(src, dbl_scratch);
-    __ Cvt_d_uw(dbl_scratch, dbl_scratch, f22);
+    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);
   }
 
   // Slow case: Call the runtime system to do the number allocation.
   __ bind(&slow);
 
   // TODO(3095996): Put a valid pointer value in the stack slot where the result
   // register is stored, as this register is in the pointer map, but contains an
   // integer value.
   __ StoreToSafepointRegisterSlot(zero_reg, dst);
   CallRuntimeFromDeferred(Runtime::kAllocateHeapNumber, 0, instr);
   __ Move(dst, v0);
   __ Subu(dst, dst, kHeapObjectTag);
 
   // Done. Put the value in dbl_scratch into the value of the allocated heap
   // number.
   __ bind(&done);
-  __ sdc1(dbl_scratch, MemOperand(dst, HeapNumber::kValueOffset));
+  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);
 }
 
 
 void LCodeGen::DoNumberTagD(LNumberTagD* instr) {
   class DeferredNumberTagD: public LDeferredCode {
    public:
     DeferredNumberTagD(LCodeGen* codegen, LNumberTagD* instr)
         : LDeferredCode(codegen), instr_(instr) { }
@@ skipping 12 matching lines @@
   DeferredNumberTagD* deferred = new(zone()) DeferredNumberTagD(this, instr);
   if (FLAG_inline_new) {
     __ LoadRoot(scratch, Heap::kHeapNumberMapRootIndex);
     // We want the untagged address first for performance
     __ AllocateHeapNumber(reg, temp1, temp2, scratch, deferred->entry(),
                           DONT_TAG_RESULT);
   } else {
     __ Branch(deferred->entry());
   }
   __ bind(deferred->exit());
-  __ sdc1(input_reg, MemOperand(reg, HeapNumber::kValueOffset));
+  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);
 }
 
 
 void LCodeGen::DoDeferredNumberTagD(LNumberTagD* instr) {
   // TODO(3095996): Get rid of this. For now, we need to make the
   // result register contain a valid pointer because it is already
   // contained in the register pointer map.
   Register reg = ToRegister(instr->result());
@@ skipping 27 matching lines @@
   }
 }
 
 
 void LCodeGen::EmitNumberUntagD(Register input_reg,
                                 DoubleRegister result_reg,
                                 bool deoptimize_on_undefined,
                                 bool deoptimize_on_minus_zero,
                                 LEnvironment* env) {
   Register scratch = scratch0();
+  CpuFeatures::Scope scope(FPU);
 
   Label load_smi, heap_number, done;
 
   // Smi check.
   __ UntagAndJumpIfSmi(scratch, input_reg, &load_smi);
 
   // Heap number map check.
   __ lw(scratch, FieldMemOperand(input_reg, HeapObject::kMapOffset));
   __ LoadRoot(at, Heap::kHeapNumberMapRootIndex);
   if (deoptimize_on_undefined) {
@@ skipping 44 matching lines @@
   Label done;
 
   // The input is a tagged HeapObject.
   // Heap number map check.
   __ lw(scratch1, FieldMemOperand(input_reg, HeapObject::kMapOffset));
   __ LoadRoot(at, Heap::kHeapNumberMapRootIndex);
   // This 'at' value and scratch1 map value are used for tests in both clauses
   // 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) &&
            !scratch3.is(scratch1) &&
            !scratch3.is(scratch2));
     // Performs a truncating conversion of a floating point number as used by
     // the JS bitwise operations.
     Label heap_number;
     __ Branch(&heap_number, eq, scratch1, Operand(at));  // HeapNumber map?
     // Check for undefined. Undefined is converted to zero for truncating
@@ skipping 216 matching lines @@
     __ CompareMapAndBranch(
         reg, scratch, map, &success, eq, &success, REQUIRE_EXACT_MAP);
   }
   Handle<Map> map = map_set->last();
   DoCheckMapCommon(reg, scratch, map, REQUIRE_EXACT_MAP, instr->environment());
   __ bind(&success);
 }
 
 
 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());
   __ ClampDoubleToUint8(result_reg, value_reg, temp_reg);
 }
 
 
 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);
 }
 
 
 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());
   DoubleRegister temp_reg = ToDoubleRegister(instr->temp());
   Label is_smi, done, heap_number;
 
   // Both smi and heap number cases are handled.
   __ UntagAndJumpIfSmi(scratch, input_reg, &is_smi);
 
   // Check for heap number
@@ skipping 601 matching lines @@
             Operand(Smi::FromInt(StackFrame::ARGUMENTS_ADAPTOR)));
   __ lw(temp1, MemOperand(temp1, StandardFrameConstants::kCallerFPOffset));
 
   // Check the marker in the calling frame.
   __ bind(&check_frame_marker);
   __ lw(temp1, MemOperand(temp1, StandardFrameConstants::kMarkerOffset));
 }
 
 
 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();
   int patch_size = Deoptimizer::patch_size();
   if (current_pc < last_lazy_deopt_pc_ + patch_size) {
     int padding_size = last_lazy_deopt_pc_ + patch_size - current_pc;
     ASSERT_EQ(0, padding_size % Assembler::kInstrSize);
     while (padding_size > 0) {
       __ nop();
       padding_size -= Assembler::kInstrSize;
@@ skipping 204 matching lines @@
   __ Subu(scratch, result, scratch);
   __ lw(result, FieldMemOperand(scratch,
                                 FixedArray::kHeaderSize - kPointerSize));
   __ bind(&done);
 }
 
 
 #undef __
 
 } }  // namespace v8::internal