Implement X87 stack tracking and x87 multiplication

src/ia32/lithium-codegen-ia32.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 335 matching lines @@
               current_instruction_,
               instr->hydrogen_value()->id(),
               instr->Mnemonic());
     }
 
     if (!CpuFeatures::IsSupported(SSE2)) FlushX87StackIfNecessary(instr);
 
     instr->CompileToNative(this);
 
     if (!CpuFeatures::IsSupported(SSE2)) {
-      ASSERT(!instr->HasDoubleRegisterResult() || x87_stack_depth_ == 1);
       if (FLAG_debug_code && FLAG_enable_slow_asserts) {
         __ VerifyX87StackDepth(x87_stack_depth_);
       }
     }
   }
   EnsureSpaceForLazyDeopt();
   return !is_aborted();
 }
 
 
@@ skipping 127 matching lines @@
   safepoints_.Emit(masm(), GetStackSlotCount());
   return !is_aborted();
 }
 
 
 Register LCodeGen::ToRegister(int index) const {
   return Register::FromAllocationIndex(index);
 }
 
 
+X87Register LCodeGen::ToX87Register(int index) const {
+  return X87Register::FromAllocationIndex(index);
+}
+
+
 XMMRegister LCodeGen::ToDoubleRegister(int index) const {
   return XMMRegister::FromAllocationIndex(index);
 }
 
 
-bool LCodeGen::IsX87TopOfStack(LOperand* op) const {
-  return op->IsDoubleRegister();
+void LCodeGen::X87LoadForUsage(X87Register reg) {
+  ASSERT(X87StackContains(reg));
+  X87Fxch(reg);
+  x87_stack_depth_--;
 }
 
 
-void LCodeGen::ReadX87Operand(Operand dst) {
-  ASSERT(x87_stack_depth_ == 1);
+void LCodeGen::X87Fxch(X87Register reg, int other_slot) {
+  ASSERT(X87StackContains(reg) && x87_stack_depth_ > other_slot);
+  int i  = X87ArrayIndex(reg);
+  int st = x87_st2idx(i);
+  if (st != other_slot) {
+    int other_i = x87_st2idx(other_slot);
+    X87Register other   = x87_stack_[other_i];
+    x87_stack_[other_i] = reg;
+    x87_stack_[i]       = other;
+    if (st == 0) {
+      __ fxch(other_slot);
+    } else if (other_slot == 0) {
+      __ fxch(st);
+    } else {
+      __ fxch(st);
+      __ fxch(other_slot);
+      __ fxch(st);
+    }
+  }
+}
+
+
+int LCodeGen::x87_st2idx(int pos) {
+  return x87_stack_depth_ - pos - 1;
+}
+
+
+int LCodeGen::X87ArrayIndex(X87Register reg) {
+  for (int i = 0; i < x87_stack_depth_; i++) {
+    if (x87_stack_[i].is(reg)) return i;
+  }
+  UNREACHABLE();
+  return -1;
+}
+
+
+bool LCodeGen::X87StackContains(X87Register reg) {
+  for (int i = 0; i < x87_stack_depth_; i++) {
+    if (x87_stack_[i].is(reg)) return true;
+  }
+  return false;
+}
+
+
+void LCodeGen::X87Free(X87Register reg) {
+  ASSERT(X87StackContains(reg));
+  int i  = X87ArrayIndex(reg);
+  int st = x87_st2idx(i);
+  if (st > 0) {
+    // keep track of how fstp(i) changes the order of elements
+    int tos_i = x87_st2idx(0);
+    x87_stack_[i] = x87_stack_[tos_i];
+  }
+  x87_stack_depth_--;
+  __ fstp(st);
+}
+
+
+void LCodeGen::X87Mov(X87Register dst, Operand src, X87OperandType opts) {
+  if (X87StackContains(dst)) {
+    X87Fxch(dst);
+    __ fstp(0);
+  } else {
+    ASSERT(x87_stack_depth_ < X87Register::kNumAllocatableRegisters);
+    x87_stack_[x87_stack_depth_] = dst;
+    x87_stack_depth_++;
+  }
+  X87Fld(src, opts);
+}
+
+
+void LCodeGen::X87Fld(Operand src, X87OperandType opts) {
+  if (opts == kX87DoubleOperand) {
+    __ fld_d(src);
+  } else if (opts == kX87FloatOperand) {
+    __ fld_s(src);
+  } else if (opts == kX87IntOperand) {
+    __ fild_s(src);
+  } else {
+    UNREACHABLE();
+  }
+}
+
+
+void LCodeGen::X87Mov(Operand dst, X87Register src) {
+  X87Fxch(src);
   __ fst_d(dst);
 }
 
 
-void LCodeGen::PushX87DoubleOperand(Operand src) {
-  ASSERT(x87_stack_depth_ == 0);
-  x87_stack_depth_++;
-  __ fld_d(src);
+void LCodeGen::X87PrepareToWrite(X87Register reg) {
+  if (X87StackContains(reg)) {
+    X87Free(reg);
+  }
+  // Mark this register as the next register to write to
+  x87_stack_[x87_stack_depth_] = reg;
 }
 
 
-void LCodeGen::PushX87FloatOperand(Operand src) {
-  ASSERT(x87_stack_depth_ == 0);
+void LCodeGen::X87CommitWrite(X87Register reg) {
+  // Assert the reg is prepared to write, but not on the virtual stack yet
+  ASSERT(!X87StackContains(reg) && x87_stack_[x87_stack_depth_].is(reg) &&
+      x87_stack_depth_ < X87Register::kNumAllocatableRegisters);
   x87_stack_depth_++;
-  __ fld_s(src);
 }
 
 
-void LCodeGen::PopX87() {
-  ASSERT(x87_stack_depth_ == 1);
-  x87_stack_depth_--;
-  __ fstp(0);
+void LCodeGen::X87PrepareBinaryOp(
+    X87Register left, X87Register right, X87Register result) {
+  // You need to use DefineSameAsFirst for x87 instructions
+  ASSERT(result.is(left));
+  X87Fxch(right, 1);
+  X87Fxch(left);
 }
 
 
-void LCodeGen::CurrentInstructionReturnsX87Result() {
-  ASSERT(x87_stack_depth_ <= 1);
-  if (x87_stack_depth_ == 0) {
-    x87_stack_depth_ = 1;
+void LCodeGen::FlushX87StackIfNecessary(LInstruction* instr) {
+  if (x87_stack_depth_ > 0 && instr->ClobbersDoubleRegisters()) {
+    bool double_inputs = instr->HasDoubleRegisterInput();
+
+    // Flush stack from tos down, since FreeX87() will mess with tos
+    for (int i = x87_stack_depth_-1; i >= 0; i--) {
+      X87Register reg = x87_stack_[i];
+      // Skip registers which contain the inputs for the next instruction
+      // when flushing the stack
+      if (double_inputs && instr->IsDoubleInput(reg, this)) {
+        continue;
+      }
+      X87Free(reg);
+      if (i < x87_stack_depth_-1) i++;
+    }
+  }
+  if (instr->IsReturn()) {
+    while (x87_stack_depth_ > 0) {
+      __ fstp(0);
+      x87_stack_depth_--;
+    }
   }
 }
 
 
-void LCodeGen::FlushX87StackIfNecessary(LInstruction* instr) {
-  if (x87_stack_depth_ > 0) {
-    if ((instr->ClobbersDoubleRegisters() ||
-         instr->HasDoubleRegisterResult()) &&
-        !instr->HasDoubleRegisterInput()) {
-      PopX87();
-    }
-  }
+void LCodeGen::EmitFlushX87ForDeopt() {
+  for (int i = 0; i < x87_stack_depth_; i++) __ fstp(0);
 }
 
 
 Register LCodeGen::ToRegister(LOperand* op) const {
   ASSERT(op->IsRegister());
   return ToRegister(op->index());
 }
 
 
+X87Register LCodeGen::ToX87Register(LOperand* op) const {
+  ASSERT(op->IsDoubleRegister());
+  return ToX87Register(op->index());
+}
+
+
 XMMRegister LCodeGen::ToDoubleRegister(LOperand* op) const {
   ASSERT(op->IsDoubleRegister());
   return ToDoubleRegister(op->index());
 }
 
 
 int LCodeGen::ToInteger32(LConstantOperand* op) const {
   HConstant* constant = chunk_->LookupConstant(op);
   return constant->Integer32Value();
 }
@@ skipping 252 matching lines @@
     deoptimizations_.Add(environment, zone());
   }
 }
 
 
 void LCodeGen::DeoptimizeIf(Condition cc,
                             LEnvironment* environment,
                             Deoptimizer::BailoutType bailout_type) {
   RegisterEnvironmentForDeoptimization(environment, Safepoint::kNoLazyDeopt);
   ASSERT(environment->HasBeenRegistered());
-  // It's an error to deoptimize with the x87 fp stack in use.
-  ASSERT(x87_stack_depth_ == 0);
   int id = environment->deoptimization_index();
   ASSERT(info()->IsOptimizing() || info()->IsStub());
   Address entry =
       Deoptimizer::GetDeoptimizationEntry(isolate(), id, bailout_type);
   if (entry == NULL) {
     Abort("bailout was not prepared");
     return;
   }
 
   if (FLAG_deopt_every_n_times != 0 && !info()->IsStub()) {
@@ skipping 17 matching lines @@
     __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
 
     __ bind(&no_deopt);
     __ mov(FieldOperand(ebx, SharedFunctionInfo::kStressDeoptCounterOffset),
            eax);
     __ pop(ebx);
     __ pop(eax);
     __ popfd();
   }
 
+  // Before Instructions which can deopt, we normally flush the x87 stack. But
+  // we can have inputs or outputs of the current instruction on the stack,
+  // thus we need to flush them here from the physical stack to leave it in a
+  // consistent state.
+  if (x87_stack_depth_ > 0) {
+    Label done;
+    if (cc != no_condition) __ j(NegateCondition(cc), &done, Label::kNear);
+    EmitFlushX87ForDeopt();
+    __ bind(&done);
+  }
+
   if (FLAG_trap_on_deopt && info()->IsOptimizing()) {
     Label done;
-    if (cc != no_condition) {
-      __ j(NegateCondition(cc), &done, Label::kNear);
-    }
+    if (cc != no_condition) __ j(NegateCondition(cc), &done, Label::kNear);
     __ int3();
     __ bind(&done);
   }
 
   ASSERT(info()->IsStub() || frame_is_built_);
   if (cc == no_condition && frame_is_built_) {
     if (bailout_type == Deoptimizer::LAZY) {
       __ call(entry, RelocInfo::RUNTIME_ENTRY);
     } else {
       __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
@@ skipping 822 matching lines @@
 }
 
 
 void LCodeGen::DoConstantD(LConstantD* instr) {
   double v = instr->value();
   uint64_t int_val = BitCast<uint64_t, double>(v);
   int32_t lower = static_cast<int32_t>(int_val);
   int32_t upper = static_cast<int32_t>(int_val >> (kBitsPerInt));
 
   if (!CpuFeatures::IsSafeForSnapshot(SSE2)) {
+    __ push(Immediate(upper));
     __ push(Immediate(lower));
-    __ push(Immediate(upper));
-    PushX87DoubleOperand(Operand(esp, 0));
+    X87Mov(ToX87Register(instr->result()), Operand(esp, 0));
     __ add(Operand(esp), Immediate(kDoubleSize));
-    CurrentInstructionReturnsX87Result();
   } else {
     CpuFeatureScope scope1(masm(), SSE2);
     ASSERT(instr->result()->IsDoubleRegister());
     XMMRegister res = ToDoubleRegister(instr->result());
     if (int_val == 0) {
       __ xorps(res, res);
     } else {
       Register temp = ToRegister(instr->temp());
       if (CpuFeatures::IsSupported(SSE4_1)) {
         CpuFeatureScope scope2(masm(), SSE4_1);
@@ skipping 244 matching lines @@
     __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
     __ bind(&return_right);
     __ movsd(left_reg, right_reg);
 
     __ bind(&return_left);
   }
 }
 
 
 void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
-  CpuFeatureScope scope(masm(), SSE2);
-  XMMRegister left = ToDoubleRegister(instr->left());
-  XMMRegister right = ToDoubleRegister(instr->right());
-  XMMRegister result = ToDoubleRegister(instr->result());
-  // Modulo uses a fixed result register.
-  ASSERT(instr->op() == Token::MOD || left.is(result));
-  switch (instr->op()) {
-    case Token::ADD:
-      __ addsd(left, right);
-      break;
-    case Token::SUB:
-      __ subsd(left, right);
-      break;
-    case Token::MUL:
-      __ mulsd(left, right);
-      break;
-    case Token::DIV:
-      __ divsd(left, right);
-      // Don't delete this mov. It may improve performance on some CPUs,
-      // when there is a mulsd depending on the result
-      __ movaps(left, left);
-      break;
-    case Token::MOD: {
-      // Pass two doubles as arguments on the stack.
-      __ PrepareCallCFunction(4, eax);
-      __ movdbl(Operand(esp, 0 * kDoubleSize), left);
-      __ movdbl(Operand(esp, 1 * kDoubleSize), right);
-      __ CallCFunction(
-          ExternalReference::double_fp_operation(Token::MOD, isolate()),
-          4);
+  if (CpuFeatures::IsSafeForSnapshot(SSE2)) {
+    CpuFeatureScope scope(masm(), SSE2);
+    XMMRegister left = ToDoubleRegister(instr->left());
+    XMMRegister right = ToDoubleRegister(instr->right());
+    XMMRegister result = ToDoubleRegister(instr->result());
+    // Modulo uses a fixed result register.
+    ASSERT(instr->op() == Token::MOD || left.is(result));
+    switch (instr->op()) {
+      case Token::ADD:
+        __ addsd(left, right);
+        break;
+      case Token::SUB:
+        __ subsd(left, right);
+        break;
+      case Token::MUL:
+        __ mulsd(left, right);
+        break;
+      case Token::DIV:
+        __ divsd(left, right);
+        // Don't delete this mov. It may improve performance on some CPUs,
+        // when there is a mulsd depending on the result
+        __ movaps(left, left);
+        break;
+      case Token::MOD: {
+        // Pass two doubles as arguments on the stack.
+        __ PrepareCallCFunction(4, eax);
+        __ movdbl(Operand(esp, 0 * kDoubleSize), left);
+        __ movdbl(Operand(esp, 1 * kDoubleSize), right);
+        __ CallCFunction(
+            ExternalReference::double_fp_operation(Token::MOD, isolate()),
+            4);
 
-      // Return value is in st(0) on ia32.
-      // Store it into the (fixed) result register.
-      __ sub(Operand(esp), Immediate(kDoubleSize));
-      __ fstp_d(Operand(esp, 0));
-      __ movdbl(result, Operand(esp, 0));
-      __ add(Operand(esp), Immediate(kDoubleSize));
-      break;
+        // Return value is in st(0) on ia32.
+        // Store it into the (fixed) result register.
+        __ sub(Operand(esp), Immediate(kDoubleSize));
+        __ fstp_d(Operand(esp, 0));
+        __ movdbl(result, Operand(esp, 0));
+        __ add(Operand(esp), Immediate(kDoubleSize));
+        break;
+      }
+      default:
+        UNREACHABLE();
+        break;
     }
-    default:
-      UNREACHABLE();
-      break;
+  } else {
+    X87Register left = ToX87Register(instr->left());
+    X87Register right = ToX87Register(instr->right());
+    X87Register result = ToX87Register(instr->result());
+    X87PrepareBinaryOp(left, right, result);
+    switch (instr->op()) {
+      case Token::MUL:
+        __ fmul_i(1);
+        break;
+      default:
+        UNREACHABLE();
+        break;
+    }
   }
 }
 
 
-void LCodeGen::DoNegateNoSSE2D(LNegateNoSSE2D* instr) {
-  __ push(Immediate(-1));
-  __ fild_s(Operand(esp, 0));
-  __ add(esp, Immediate(kPointerSize));
-  __ fmulp();
-  CurrentInstructionReturnsX87Result();
-}
-
-
-
 void LCodeGen::DoArithmeticT(LArithmeticT* instr) {
   ASSERT(ToRegister(instr->context()).is(esi));
   ASSERT(ToRegister(instr->left()).is(edx));
   ASSERT(ToRegister(instr->right()).is(eax));
   ASSERT(ToRegister(instr->result()).is(eax));
 
   BinaryOpStub stub(instr->op(), NO_OVERWRITE);
   CallCode(stub.GetCode(isolate()), RelocInfo::CODE_TARGET, instr);
   __ nop();  // Signals no inlined code.
 }
@@ skipping 897 matching lines @@
   HObjectAccess access = instr->hydrogen()->access();
   int offset = access.offset();
   Register object = ToRegister(instr->object());
   if (FLAG_track_double_fields &&
       instr->hydrogen()->representation().IsDouble()) {
     if (CpuFeatures::IsSupported(SSE2)) {
       CpuFeatureScope scope(masm(), SSE2);
       XMMRegister result = ToDoubleRegister(instr->result());
       __ movdbl(result, FieldOperand(object, offset));
     } else {
-      PushX87DoubleOperand(FieldOperand(object, offset));
-      CurrentInstructionReturnsX87Result();
+      X87Mov(ToX87Register(instr->result()), FieldOperand(object, offset));
     }
     return;
   }
 
   Register result = ToRegister(instr->result());
   if (access.IsInobject()) {
     __ mov(result, FieldOperand(object, offset));
   } else {
     __ mov(result, FieldOperand(object, JSObject::kPropertiesOffset));
     __ mov(result, FieldOperand(result, offset));
@@ skipping 224 matching lines @@
       elements_kind,
       0,
       instr->additional_index()));
   if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
     if (CpuFeatures::IsSupported(SSE2)) {
       CpuFeatureScope scope(masm(), SSE2);
       XMMRegister result(ToDoubleRegister(instr->result()));
       __ movss(result, operand);
       __ cvtss2sd(result, result);
     } else {
-      PushX87FloatOperand(operand);
-      CurrentInstructionReturnsX87Result();
+      X87Mov(ToX87Register(instr->result()), operand, kX87FloatOperand);
     }
   } else if (elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
     if (CpuFeatures::IsSupported(SSE2)) {
       CpuFeatureScope scope(masm(), SSE2);
       __ movdbl(ToDoubleRegister(instr->result()), operand);
     } else {
-      PushX87DoubleOperand(operand);
-      CurrentInstructionReturnsX87Result();
+      X87Mov(ToX87Register(instr->result()), operand);
     }
   } else {
     Register result(ToRegister(instr->result()));
     switch (elements_kind) {
       case EXTERNAL_BYTE_ELEMENTS:
         __ movsx_b(result, operand);
         break;
       case EXTERNAL_PIXEL_ELEMENTS:
       case EXTERNAL_UNSIGNED_BYTE_ELEMENTS:
         __ movzx_b(result, operand);
@@ skipping 50 matching lines @@
       instr->key(),
       instr->hydrogen()->key()->representation(),
       FAST_DOUBLE_ELEMENTS,
       FixedDoubleArray::kHeaderSize - kHeapObjectTag,
       instr->additional_index());
   if (CpuFeatures::IsSupported(SSE2)) {
     CpuFeatureScope scope(masm(), SSE2);
     XMMRegister result = ToDoubleRegister(instr->result());
     __ movdbl(result, double_load_operand);
   } else {
-    PushX87DoubleOperand(double_load_operand);
-    CurrentInstructionReturnsX87Result();
+    X87Mov(ToX87Register(instr->result()), double_load_operand);
   }
 }
 
 
 void LCodeGen::DoLoadKeyedFixedArray(LLoadKeyed* instr) {
   Register result = ToRegister(instr->result());
 
   // Load the result.
   __ mov(result,
          BuildFastArrayOperand(instr->elements(),
@@ skipping 973 matching lines @@
     }
   } else if (FLAG_track_double_fields && representation.IsDouble()) {
     ASSERT(transition.is_null());
     ASSERT(access.IsInobject());
     ASSERT(!instr->hydrogen()->NeedsWriteBarrier());
     if (CpuFeatures::IsSupported(SSE2)) {
       CpuFeatureScope scope(masm(), SSE2);
       XMMRegister value = ToDoubleRegister(instr->value());
       __ movdbl(FieldOperand(object, offset), value);
     } else {
-      __ fstp_d(FieldOperand(object, offset));
+      X87Register value = ToX87Register(instr->value());
+      X87Mov(FieldOperand(object, offset), value);
     }
     return;
   }
 
   if (!transition.is_null()) {
     if (!instr->hydrogen()->NeedsWriteBarrierForMap()) {
       __ mov(FieldOperand(object, HeapObject::kMapOffset), transition);
     } else {
       Register temp = ToRegister(instr->temp());
       Register temp_map = ToRegister(instr->temp_map());
@@ skipping 105 matching lines @@
       __ movss(operand, xmm0);
     } else {
       __ fld(0);
       __ fstp_s(operand);
     }
   } else if (elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
     if (CpuFeatures::IsSafeForSnapshot(SSE2)) {
       CpuFeatureScope scope(masm(), SSE2);
       __ movdbl(operand, ToDoubleRegister(instr->value()));
     } else {
-      __ fst_d(operand);
+      X87Mov(operand, ToX87Register(instr->value()));
     }
   } else {
     Register value = ToRegister(instr->value());
     switch (elements_kind) {
       case EXTERNAL_PIXEL_ELEMENTS:
       case EXTERNAL_UNSIGNED_BYTE_ELEMENTS:
       case EXTERNAL_BYTE_ELEMENTS:
         __ mov_b(operand, value);
         break;
       case EXTERNAL_SHORT_ELEMENTS:
@@ skipping 61 matching lines @@
       Operand double_store_operand2 = BuildFastArrayOperand(
           instr->elements(),
           instr->key(),
           instr->hydrogen()->key()->representation(),
           FAST_DOUBLE_ELEMENTS,
           FixedDoubleArray::kHeaderSize - kHeapObjectTag + kPointerSize,
           instr->additional_index());
       __ mov(double_store_operand2, Immediate(upper));
     } else {
       Label no_special_nan_handling;
-      ASSERT(x87_stack_depth_ > 0);
+      X87Register value = ToX87Register(instr->value());
+      X87Fxch(value);
 
       if (instr->NeedsCanonicalization()) {
         __ fld(0);
         __ fld(0);
         __ FCmp();
 
         __ j(parity_odd, &no_special_nan_handling);
         __ sub(esp, Immediate(kDoubleSize));
         __ fst_d(MemOperand(esp, 0));
         __ cmp(MemOperand(esp, sizeof(kHoleNanLower32)),
@@ skipping 449 matching lines @@
 
   Register reg = ToRegister(instr->result());
 
   bool convert_hole = false;
   HValue* change_input = instr->hydrogen()->value();
   if (change_input->IsLoadKeyed()) {
     HLoadKeyed* load = HLoadKeyed::cast(change_input);
     convert_hole = load->UsesMustHandleHole();
   }
 
+  bool use_sse2 = CpuFeatures::IsSupported(SSE2);
+  if (!use_sse2) {
+    // Put the value to the top of stack
+    X87Register src = ToX87Register(instr->value());
+    X87LoadForUsage(src);
+  }
+
   Label no_special_nan_handling;
   Label done;
   if (convert_hole) {
-    bool use_sse2 = CpuFeatures::IsSupported(SSE2);
     if (use_sse2) {
       CpuFeatureScope scope(masm(), SSE2);
       XMMRegister input_reg = ToDoubleRegister(instr->value());
       __ ucomisd(input_reg, input_reg);
     } else {
       __ fld(0);
-      __ fld(0);
       __ FCmp();
     }
 
     __ j(parity_odd, &no_special_nan_handling);
     __ sub(esp, Immediate(kDoubleSize));
     if (use_sse2) {
       CpuFeatureScope scope(masm(), SSE2);
       XMMRegister input_reg = ToDoubleRegister(instr->value());
       __ movdbl(MemOperand(esp, 0), input_reg);
     } else {
@@ skipping 33 matching lines @@
     __ jmp(deferred->entry());
   }
   __ bind(deferred->exit());
   if (CpuFeatures::IsSupported(SSE2)) {
     CpuFeatureScope scope(masm(), SSE2);
     XMMRegister input_reg = ToDoubleRegister(instr->value());
     __ movdbl(FieldOperand(reg, HeapNumber::kValueOffset), input_reg);
   } else {
     __ fst_d(FieldOperand(reg, HeapNumber::kValueOffset));
   }
+  if (!use_sse2) {
+    // clean up the stack
+    __ fstp(0);
+  }
   __ bind(&done);
 }
 
 
 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());
   __ Set(reg, Immediate(0));
@@ skipping 29 matching lines @@
     DeoptimizeIf(not_zero, instr->environment());
   } else {
     __ AssertSmi(result);
   }
   __ SmiUntag(result);
 }
 
 
 void LCodeGen::EmitNumberUntagDNoSSE2(Register input_reg,
                                       Register temp_reg,
+                                      X87Register res_reg,
                                       bool allow_undefined_as_nan,
                                       bool deoptimize_on_minus_zero,
                                       LEnvironment* env,
                                       NumberUntagDMode mode) {
   Label load_smi, done;
 
+  X87PrepareToWrite(res_reg);
   STATIC_ASSERT(NUMBER_CANDIDATE_IS_ANY_TAGGED_CONVERT_HOLE >
                 NUMBER_CANDIDATE_IS_ANY_TAGGED);
   if (mode >= NUMBER_CANDIDATE_IS_ANY_TAGGED) {
     // Smi check.
     __ JumpIfSmi(input_reg, &load_smi, Label::kNear);
 
     // Heap number map check.
     __ cmp(FieldOperand(input_reg, HeapObject::kMapOffset),
            factory()->heap_number_map());
     if (!allow_undefined_as_nan) {
@@ skipping 40 matching lines @@
     ASSERT(mode == NUMBER_CANDIDATE_IS_SMI);
   }
 
   __ bind(&load_smi);
   __ SmiUntag(input_reg);  // Untag smi before converting to float.
   __ push(input_reg);
   __ fild_s(Operand(esp, 0));
   __ pop(input_reg);
   __ SmiTag(input_reg);  // Retag smi.
   __ bind(&done);
+  X87CommitWrite(res_reg);
 }
 
 
 void LCodeGen::EmitNumberUntagD(Register input_reg,
                                 Register temp_reg,
                                 XMMRegister result_reg,
                                 bool allow_undefined_as_nan,
                                 bool deoptimize_on_minus_zero,
                                 LEnvironment* env,
                                 NumberUntagDMode mode) {
@@ skipping 361 matching lines @@
     EmitNumberUntagD(input_reg,
                      temp_reg,
                      result_reg,
                      instr->hydrogen()->allow_undefined_as_nan(),
                      deoptimize_on_minus_zero,
                      instr->environment(),
                      mode);
   } else {
     EmitNumberUntagDNoSSE2(input_reg,
                            temp_reg,
+                           ToX87Register(instr->result()),
                            instr->hydrogen()->allow_undefined_as_nan(),
                            deoptimize_on_minus_zero,
                            instr->environment(),
                            mode);
-    CurrentInstructionReturnsX87Result();
   }
 }
 
 
 void LCodeGen::DoDoubleToI(LDoubleToI* instr) {
   LOperand* input = instr->value();
   ASSERT(input->IsDoubleRegister());
   LOperand* result = instr->result();
   ASSERT(result->IsRegister());
   CpuFeatureScope scope(masm(), SSE2);
@@ skipping 1001 matching lines @@
                               FixedArray::kHeaderSize - kPointerSize));
   __ bind(&done);
 }
 
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_IA32