X87: enable the X87 turbofan support.

src/compiler/x87/code-generator-x87.cc

 // Copyright 2013 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/compiler/code-generator.h"
 
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
 #include "src/compiler/node-matchers.h"
-#include "src/ia32/assembler-ia32.h"
-#include "src/ia32/macro-assembler-ia32.h"
 #include "src/scopes.h"
+#include "src/x87/assembler-x87.h"
+#include "src/x87/macro-assembler-x87.h"
 
 namespace v8 {
 namespace internal {
 namespace compiler {
 
 #define __ masm()->
 
 
-#define kScratchDoubleReg xmm0
-
-
-// Adds IA-32 specific methods for decoding operands.
-class IA32OperandConverter : public InstructionOperandConverter {
+// Adds X87 specific methods for decoding operands.
+class X87OperandConverter : public InstructionOperandConverter {
  public:
-  IA32OperandConverter(CodeGenerator* gen, Instruction* instr)
+  X87OperandConverter(CodeGenerator* gen, Instruction* instr)
       : InstructionOperandConverter(gen, instr) {}
 
   Operand InputOperand(size_t index, int extra = 0) {
     return ToOperand(instr_->InputAt(index), extra);
   }
 
   Immediate InputImmediate(size_t index) {
     return ToImmediate(instr_->InputAt(index));
   }
 
   Operand OutputOperand() { return ToOperand(instr_->Output()); }
 
   Operand ToOperand(InstructionOperand* op, int extra = 0) {
     if (op->IsRegister()) {
       DCHECK(extra == 0);
       return Operand(ToRegister(op));
     } else if (op->IsDoubleRegister()) {
       DCHECK(extra == 0);
-      return Operand(ToDoubleRegister(op));
+      UNIMPLEMENTED();
     }
     DCHECK(op->IsStackSlot() || op->IsDoubleStackSlot());
     // The linkage computes where all spill slots are located.
     FrameOffset offset = linkage()->GetFrameOffset(
         AllocatedOperand::cast(op)->index(), frame(), extra);
     return Operand(offset.from_stack_pointer() ? esp : ebp, offset.offset());
   }
 
   Operand HighOperand(InstructionOperand* op) {
     DCHECK(op->IsDoubleStackSlot());
@@ skipping 123 matching lines @@
 
   void Generate() final { __ xor_(result_, result_); }
 
  private:
   Register const result_;
 };
 
 
 class OutOfLineLoadFloat final : public OutOfLineCode {
  public:
-  OutOfLineLoadFloat(CodeGenerator* gen, XMMRegister result)
+  OutOfLineLoadFloat(CodeGenerator* gen, X87Register result)
       : OutOfLineCode(gen), result_(result) {}
 
-  void Generate() final { __ pcmpeqd(result_, result_); }
+  void Generate() final {
+    DCHECK(result_.code() == 0);
+    USE(result_);
+    __ fstp(0);
+    __ push(Immediate(0xffffffff));
+    __ push(Immediate(0x7fffffff));
+    __ fld_d(MemOperand(esp, 0));
+    __ lea(esp, Operand(esp, kDoubleSize));
+  }
 
  private:
-  XMMRegister const result_;
+  X87Register const result_;
 };
 
 
 class OutOfLineTruncateDoubleToI final : public OutOfLineCode {
  public:
   OutOfLineTruncateDoubleToI(CodeGenerator* gen, Register result,
-                             XMMRegister input)
+                             X87Register input)
       : OutOfLineCode(gen), result_(result), input_(input) {}
 
   void Generate() final {
-    __ sub(esp, Immediate(kDoubleSize));
-    __ movsd(MemOperand(esp, 0), input_);
-    __ SlowTruncateToI(result_, esp, 0);
-    __ add(esp, Immediate(kDoubleSize));
+    UNIMPLEMENTED();
+    USE(result_);
+    USE(input_);
   }
 
  private:
   Register const result_;
-  XMMRegister const input_;
+  X87Register const input_;
 };
 
 }  // namespace
 
 
 #define ASSEMBLE_CHECKED_LOAD_FLOAT(asm_instr)                          \
   do {                                                                  \
     auto result = i.OutputDoubleRegister();                             \
     auto offset = i.InputRegister(0);                                   \
+    DCHECK(result.code() == 0);                                         \
     if (instr->InputAt(1)->IsRegister()) {                              \
       __ cmp(offset, i.InputRegister(1));                               \
     } else {                                                            \
       __ cmp(offset, i.InputImmediate(1));                              \
     }                                                                   \
     OutOfLineCode* ool = new (zone()) OutOfLineLoadFloat(this, result); \
     __ j(above_equal, ool->entry());                                    \
-    __ asm_instr(result, i.MemoryOperand(2));                           \
+    __ fstp(0);                                                         \
+    __ asm_instr(i.MemoryOperand(2));                                   \
     __ bind(ool->exit());                                               \
   } while (false)
 
 
 #define ASSEMBLE_CHECKED_LOAD_INTEGER(asm_instr)                          \
   do {                                                                    \
     auto result = i.OutputRegister();                                     \
     auto offset = i.InputRegister(0);                                     \
     if (instr->InputAt(1)->IsRegister()) {                                \
       __ cmp(offset, i.InputRegister(1));                                 \
     } else {                                                              \
       __ cmp(offset, i.InputImmediate(1));                                \
     }                                                                     \
     OutOfLineCode* ool = new (zone()) OutOfLineLoadInteger(this, result); \
     __ j(above_equal, ool->entry());                                      \
     __ asm_instr(result, i.MemoryOperand(2));                             \
     __ bind(ool->exit());                                                 \
   } while (false)
 
 
-#define ASSEMBLE_CHECKED_STORE_FLOAT(asm_instr)                 \
-  do {                                                          \
-    auto offset = i.InputRegister(0);                           \
-    if (instr->InputAt(1)->IsRegister()) {                      \
-      __ cmp(offset, i.InputRegister(1));                       \
-    } else {                                                    \
-      __ cmp(offset, i.InputImmediate(1));                      \
-    }                                                           \
-    Label done;                                                 \
-    __ j(above_equal, &done, Label::kNear);                     \
-    __ asm_instr(i.MemoryOperand(3), i.InputDoubleRegister(2)); \
-    __ bind(&done);                                             \
+#define ASSEMBLE_CHECKED_STORE_FLOAT(asm_instr)   \
+  do {                                            \
+    auto offset = i.InputRegister(0);             \
+    if (instr->InputAt(1)->IsRegister()) {        \
+      __ cmp(offset, i.InputRegister(1));         \
+    } else {                                      \
+      __ cmp(offset, i.InputImmediate(1));        \
+    }                                             \
+    Label done;                                   \
+    DCHECK(i.InputDoubleRegister(2).code() == 0); \
+    __ j(above_equal, &done, Label::kNear);       \
+    __ asm_instr(i.MemoryOperand(3));             \
+    __ bind(&done);                               \
   } while (false)
 
 
 #define ASSEMBLE_CHECKED_STORE_INTEGER(asm_instr)            \
   do {                                                       \
     auto offset = i.InputRegister(0);                        \
     if (instr->InputAt(1)->IsRegister()) {                   \
       __ cmp(offset, i.InputRegister(1));                    \
     } else {                                                 \
       __ cmp(offset, i.InputImmediate(1));                   \
@@ skipping 21 matching lines @@
                                    kPointerSize)
             : 0;
     __ pop(Operand(esp, bytes_to_pop));
     __ add(esp, Immediate(bytes_to_pop));
   }
 }
 
 
 // Assembles an instruction after register allocation, producing machine code.
 void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
-  IA32OperandConverter i(this, instr);
+  X87OperandConverter i(this, instr);
 
   switch (ArchOpcodeField::decode(instr->opcode())) {
     case kArchCallCodeObject: {
       EnsureSpaceForLazyDeopt();
       if (HasImmediateInput(instr, 0)) {
         Handle<Code> code = Handle<Code>::cast(i.InputHeapObject(0));
         __ call(code, RelocInfo::CODE_TARGET);
       } else {
         Register reg = i.InputRegister(0);
         __ call(Operand(reg, Code::kHeaderSize - kHeapObjectTag));
       }
       RecordCallPosition(instr);
+      bool double_result =
+          instr->HasOutput() && instr->Output()->IsDoubleRegister();
+      if (double_result) {
+        __ lea(esp, Operand(esp, -kDoubleSize));
+        __ fstp_d(Operand(esp, 0));
+      }
+      __ fninit();
+      if (double_result) {
+        __ fld_d(Operand(esp, 0));
+        __ lea(esp, Operand(esp, kDoubleSize));
+      } else {
+        __ fld1();
+      }
       break;
     }
     case kArchTailCallCodeObject: {
       AssembleDeconstructActivationRecord();
       if (HasImmediateInput(instr, 0)) {
         Handle<Code> code = Handle<Code>::cast(i.InputHeapObject(0));
         __ jmp(code, RelocInfo::CODE_TARGET);
       } else {
         Register reg = i.InputRegister(0);
         __ jmp(Operand(reg, Code::kHeaderSize - kHeapObjectTag));
       }
       break;
     }
     case kArchCallJSFunction: {
       EnsureSpaceForLazyDeopt();
       Register func = i.InputRegister(0);
       if (FLAG_debug_code) {
         // Check the function's context matches the context argument.
         __ cmp(esi, FieldOperand(func, JSFunction::kContextOffset));
         __ Assert(equal, kWrongFunctionContext);
       }
       __ call(FieldOperand(func, JSFunction::kCodeEntryOffset));
       RecordCallPosition(instr);
+      bool double_result =
+          instr->HasOutput() && instr->Output()->IsDoubleRegister();
+      if (double_result) {
+        __ lea(esp, Operand(esp, -kDoubleSize));
+        __ fstp_d(Operand(esp, 0));
+      }
+      __ fninit();
+      if (double_result) {
+        __ fld_d(Operand(esp, 0));
+        __ lea(esp, Operand(esp, kDoubleSize));
+      } else {
+        __ fld1();
+      }
       break;
     }
     case kArchTailCallJSFunction: {
       Register func = i.InputRegister(0);
       if (FLAG_debug_code) {
         // Check the function's context matches the context argument.
         __ cmp(esi, FieldOperand(func, JSFunction::kContextOffset));
         __ Assert(equal, kWrongFunctionContext);
       }
       AssembleDeconstructActivationRecord();
@@ skipping 14 matching lines @@
       break;
     case kArchDeoptimize: {
       int deopt_state_id =
           BuildTranslation(instr, -1, 0, OutputFrameStateCombine::Ignore());
       AssembleDeoptimizerCall(deopt_state_id, Deoptimizer::EAGER);
       break;
     }
     case kArchRet:
       AssembleReturn();
       break;
+    case kArchFramePointer:
+      __ mov(i.OutputRegister(), ebp);
+      break;
     case kArchStackPointer:
       __ mov(i.OutputRegister(), esp);
       break;
-    case kArchFramePointer:
-      __ mov(i.OutputRegister(), ebp);
-      break;
     case kArchTruncateDoubleToI: {
-      auto result = i.OutputRegister();
       auto input = i.InputDoubleRegister(0);
-      auto ool = new (zone()) OutOfLineTruncateDoubleToI(this, result, input);
-      __ cvttsd2si(result, Operand(input));
-      __ cmp(result, 1);
-      __ j(overflow, ool->entry());
-      __ bind(ool->exit());
-      break;
-    }
-    case kIA32Add:
+      USE(input);
+      DCHECK(input.code() == 0);
+      auto result_reg = i.OutputRegister();
+      __ TruncateX87TOSToI(result_reg);
+      break;
+    }
+    case kX87Add:
       if (HasImmediateInput(instr, 1)) {
         __ add(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ add(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32And:
+    case kX87And:
       if (HasImmediateInput(instr, 1)) {
         __ and_(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ and_(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Cmp:
+    case kX87Cmp:
       if (HasImmediateInput(instr, 1)) {
         __ cmp(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ cmp(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Test:
+    case kX87Test:
       if (HasImmediateInput(instr, 1)) {
         __ test(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ test(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Imul:
+    case kX87Imul:
       if (HasImmediateInput(instr, 1)) {
         __ imul(i.OutputRegister(), i.InputOperand(0), i.InputInt32(1));
       } else {
         __ imul(i.OutputRegister(), i.InputOperand(1));
       }
       break;
-    case kIA32ImulHigh:
+    case kX87ImulHigh:
       __ imul(i.InputRegister(1));
       break;
-    case kIA32UmulHigh:
+    case kX87UmulHigh:
       __ mul(i.InputRegister(1));
       break;
-    case kIA32Idiv:
+    case kX87Idiv:
       __ cdq();
       __ idiv(i.InputOperand(1));
       break;
-    case kIA32Udiv:
+    case kX87Udiv:
       __ Move(edx, Immediate(0));
       __ div(i.InputOperand(1));
       break;
-    case kIA32Not:
+    case kX87Not:
       __ not_(i.OutputOperand());
       break;
-    case kIA32Neg:
+    case kX87Neg:
       __ neg(i.OutputOperand());
       break;
-    case kIA32Or:
+    case kX87Or:
       if (HasImmediateInput(instr, 1)) {
         __ or_(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ or_(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Xor:
+    case kX87Xor:
       if (HasImmediateInput(instr, 1)) {
         __ xor_(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ xor_(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Sub:
+    case kX87Sub:
       if (HasImmediateInput(instr, 1)) {
         __ sub(i.InputOperand(0), i.InputImmediate(1));
       } else {
         __ sub(i.InputRegister(0), i.InputOperand(1));
       }
       break;
-    case kIA32Shl:
+    case kX87Shl:
       if (HasImmediateInput(instr, 1)) {
         __ shl(i.OutputOperand(), i.InputInt5(1));
       } else {
         __ shl_cl(i.OutputOperand());
       }
       break;
-    case kIA32Shr:
+    case kX87Shr:
       if (HasImmediateInput(instr, 1)) {
         __ shr(i.OutputOperand(), i.InputInt5(1));
       } else {
         __ shr_cl(i.OutputOperand());
       }
       break;
-    case kIA32Sar:
+    case kX87Sar:
       if (HasImmediateInput(instr, 1)) {
         __ sar(i.OutputOperand(), i.InputInt5(1));
       } else {
         __ sar_cl(i.OutputOperand());
       }
       break;
-    case kIA32Ror:
+    case kX87Ror:
       if (HasImmediateInput(instr, 1)) {
         __ ror(i.OutputOperand(), i.InputInt5(1));
       } else {
         __ ror_cl(i.OutputOperand());
       }
       break;
-    case kIA32Lzcnt:
+    case kX87Lzcnt:
       __ Lzcnt(i.OutputRegister(), i.InputOperand(0));
       break;
-    case kSSEFloat32Cmp:
-      __ ucomiss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Add:
-      __ addss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Sub:
-      __ subss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Mul:
-      __ mulss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Div:
-      __ divss(i.InputDoubleRegister(0), i.InputOperand(1));
-      // Don't delete this mov. It may improve performance on some CPUs,
-      // when there is a (v)mulss depending on the result.
-      __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
-      break;
-    case kSSEFloat32Max:
-      __ maxss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Min:
-      __ minss(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat32Sqrt:
-      __ sqrtss(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat32Abs: {
-      // TODO(bmeurer): Use 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psrlq(kScratchDoubleReg, 33);
-      __ andps(i.OutputDoubleRegister(), kScratchDoubleReg);
-      break;
-    }
-    case kSSEFloat32Neg: {
-      // TODO(bmeurer): Use 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psllq(kScratchDoubleReg, 31);
-      __ xorps(i.OutputDoubleRegister(), kScratchDoubleReg);
-      break;
-    }
-    case kSSEFloat64Cmp:
-      __ ucomisd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Add:
-      __ addsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Sub:
-      __ subsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Mul:
-      __ mulsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Div:
-      __ divsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      // Don't delete this mov. It may improve performance on some CPUs,
-      // when there is a (v)mulsd depending on the result.
-      __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
-      break;
-    case kSSEFloat64Max:
-      __ maxsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Min:
-      __ minsd(i.InputDoubleRegister(0), i.InputOperand(1));
-      break;
-    case kSSEFloat64Mod: {
-      // TODO(dcarney): alignment is wrong.
+    case kX87LoadFloat64Constant: {
+      InstructionOperand* source = instr->InputAt(0);
+      InstructionOperand* destination = instr->Output();
+      DCHECK(source->IsConstant());
+      X87OperandConverter g(this, NULL);
+      Constant src_constant = g.ToConstant(source);
+
+      DCHECK_EQ(Constant::kFloat64, src_constant.type());
+      uint64_t src = bit_cast<uint64_t>(src_constant.ToFloat64());
+      uint32_t lower = static_cast<uint32_t>(src);
+      uint32_t upper = static_cast<uint32_t>(src >> 32);
+      if (destination->IsDoubleRegister()) {
+        __ sub(esp, Immediate(kDoubleSize));
+        __ mov(MemOperand(esp, 0), Immediate(lower));
+        __ mov(MemOperand(esp, kInt32Size), Immediate(upper));
+        __ fstp(0);
+        __ fld_d(MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
+      } else {
+        UNREACHABLE();
+      }
+      break;
+    }
+    case kX87Float32Cmp: {
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ FCmp();
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      break;
+    }
+    case kX87Float32Add: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, 0));
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ faddp();
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float32Sub: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ fsubp();
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float32Mul: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ fmulp();
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float32Div: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ fdivp();
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float32Max: {
+      Label check_nan_left, check_zero, return_left, return_right;
+      Condition condition = below;
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ fld(1);
+      __ fld(1);
+      __ FCmp();
+      __ j(parity_even, &check_nan_left, Label::kNear);  // At least one NaN.
+      __ j(equal, &check_zero, Label::kNear);            // left == right.
+      __ j(condition, &return_left, Label::kNear);
+      __ jmp(&return_right, Label::kNear);
+
+      __ bind(&check_zero);
+      __ fld(0);
+      __ fldz();
+      __ FCmp();
+      __ j(not_equal, &return_left, Label::kNear);  // left == right != 0.
+
+      __ fadd(1);
+      __ jmp(&return_left, Label::kNear);
+
+      __ bind(&check_nan_left);
+      __ fld(0);
+      __ fld(0);
+      __ FCmp();                                      // NaN check.
+      __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
+
+      __ bind(&return_right);
+      __ fxch();
+
+      __ bind(&return_left);
+      __ fstp(0);
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      break;
+    }
+    case kX87Float32Min: {
+      Label check_nan_left, check_zero, return_left, return_right;
+      Condition condition = above;
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, kFloatSize));
+      __ fld_s(MemOperand(esp, 0));
+      __ fld(1);
+      __ fld(1);
+      __ FCmp();
+      __ j(parity_even, &check_nan_left, Label::kNear);  // At least one NaN.
+      __ j(equal, &check_zero, Label::kNear);            // left == right.
+      __ j(condition, &return_left, Label::kNear);
+      __ jmp(&return_right, Label::kNear);
+
+      __ bind(&check_zero);
+      __ fld(0);
+      __ fldz();
+      __ FCmp();
+      __ j(not_equal, &return_left, Label::kNear);  // left == right != 0.
+      // At this point, both left and right are either 0 or -0.
+      // Push st0 and st1 to stack, then pop them to temp registers and OR them,
+      // load it to left.
+      __ push(eax);
+      __ fld(1);
+      __ fld(1);
+      __ sub(esp, Immediate(2 * kPointerSize));
+      __ fstp_s(MemOperand(esp, 0));
+      __ fstp_s(MemOperand(esp, kPointerSize));
+      __ pop(eax);
+      __ xor_(MemOperand(esp, 0), eax);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, 0));
+      __ pop(eax);  // restore esp
+      __ pop(eax);  // restore esp
+      __ jmp(&return_left, Label::kNear);
+
+      __ bind(&check_nan_left);
+      __ fld(0);
+      __ fld(0);
+      __ FCmp();                                      // NaN check.
+      __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
+
+      __ bind(&return_right);
+      __ fxch();
+
+      __ bind(&return_left);
+      __ fstp(0);
+      __ lea(esp, Operand(esp, 2 * kFloatSize));
+      break;
+    }
+    case kX87Float32Sqrt: {
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, 0));
+      __ fsqrt();
+      __ lea(esp, Operand(esp, kFloatSize));
+      break;
+    }
+    case kX87Float32Abs: {
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, 0));
+      __ fabs();
+      __ lea(esp, Operand(esp, kFloatSize));
+      break;
+    }
+    case kX87Float64Add: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, 0));
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ faddp();
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float64Sub: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fsub_d(MemOperand(esp, 0));
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float64Mul: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fmul_d(MemOperand(esp, 0));
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float64Div: {
+      __ X87SetFPUCW(0x027F);
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fdiv_d(MemOperand(esp, 0));
+      // Clear stack.
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      // Restore the default value of control word.
+      __ X87SetFPUCW(0x037F);
+      break;
+    }
+    case kX87Float64Mod: {
+      FrameScope frame_scope(&masm_, StackFrame::MANUAL);
+      __ mov(eax, esp);
+      __ PrepareCallCFunction(4, eax);
+      __ fstp(0);
+      __ fld_d(MemOperand(eax, 0));
+      __ fstp_d(Operand(esp, 1 * kDoubleSize));
+      __ fld_d(MemOperand(eax, kDoubleSize));
+      __ fstp_d(Operand(esp, 0));
+      __ CallCFunction(ExternalReference::mod_two_doubles_operation(isolate()),
+                       4);
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      break;
+    }
+    case kX87Float64Max: {
+      Label check_nan_left, check_zero, return_left, return_right;
+      Condition condition = below;
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fld_d(MemOperand(esp, 0));
+      __ fld(1);
+      __ fld(1);
+      __ FCmp();
+      __ j(parity_even, &check_nan_left, Label::kNear);  // At least one NaN.
+      __ j(equal, &check_zero, Label::kNear);            // left == right.
+      __ j(condition, &return_left, Label::kNear);
+      __ jmp(&return_right, Label::kNear);
+
+      __ bind(&check_zero);
+      __ fld(0);
+      __ fldz();
+      __ FCmp();
+      __ j(not_equal, &return_left, Label::kNear);  // left == right != 0.
+
+      __ fadd(1);
+      __ jmp(&return_left, Label::kNear);
+
+      __ bind(&check_nan_left);
+      __ fld(0);
+      __ fld(0);
+      __ FCmp();                                      // NaN check.
+      __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
+
+      __ bind(&return_right);
+      __ fxch();
+
+      __ bind(&return_left);
+      __ fstp(0);
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      break;
+    }
+    case kX87Float64Min: {
+      Label check_nan_left, check_zero, return_left, return_right;
+      Condition condition = above;
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fld_d(MemOperand(esp, 0));
+      __ fld(1);
+      __ fld(1);
+      __ FCmp();
+      __ j(parity_even, &check_nan_left, Label::kNear);  // At least one NaN.
+      __ j(equal, &check_zero, Label::kNear);            // left == right.
+      __ j(condition, &return_left, Label::kNear);
+      __ jmp(&return_right, Label::kNear);
+
+      __ bind(&check_zero);
+      __ fld(0);
+      __ fldz();
+      __ FCmp();
+      __ j(not_equal, &return_left, Label::kNear);  // left == right != 0.
+      // At this point, both left and right are either 0 or -0.
+      // Push st0 and st1 to stack, then pop them to temp registers and OR them,
+      // load it to left.
+      __ push(eax);
+      __ fld(1);
+      __ fld(1);
+      __ sub(esp, Immediate(2 * kPointerSize));
+      __ fstp_s(MemOperand(esp, 0));
+      __ fstp_s(MemOperand(esp, kPointerSize));
+      __ pop(eax);
+      __ xor_(MemOperand(esp, 0), eax);
+      __ fstp(0);
+      __ fld_s(MemOperand(esp, 0));
+      __ pop(eax);  // restore esp
+      __ pop(eax);  // restore esp
+      __ jmp(&return_left, Label::kNear);
+
+      __ bind(&check_nan_left);
+      __ fld(0);
+      __ fld(0);
+      __ FCmp();                                      // NaN check.
+      __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
+
+      __ bind(&return_right);
+      __ fxch();
+
+      __ bind(&return_left);
+      __ fstp(0);
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      break;
+    }
+    case kX87Float64Abs: {
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, 0));
+      __ fabs();
+      __ lea(esp, Operand(esp, kDoubleSize));
+      break;
+    }
+    case kX87Int32ToFloat64: {
+      InstructionOperand* input = instr->InputAt(0);
+      DCHECK(input->IsRegister() || input->IsStackSlot());
+      __ fstp(0);
+      if (input->IsRegister()) {
+        Register input_reg = i.InputRegister(0);
+        __ push(input_reg);
+        __ fild_s(Operand(esp, 0));
+        __ pop(input_reg);
+      } else {
+        __ fild_s(i.InputOperand(0));
+      }
+      break;
+    }
+    case kX87Float32ToFloat64: {
+      InstructionOperand* input = instr->InputAt(0);
+      if (input->IsDoubleRegister()) {
+        __ sub(esp, Immediate(kDoubleSize));
+        __ fstp_d(MemOperand(esp, 0));
+        __ fld_d(MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
+      } else {
+        DCHECK(input->IsDoubleStackSlot());
+        __ fstp(0);
+        __ fld_s(i.InputOperand(0));
+      }
+      break;
+    }
+    case kX87Uint32ToFloat64: {
+      __ fstp(0);
+      __ LoadUint32NoSSE2(i.InputRegister(0));
+      break;
+    }
+    case kX87Float64ToInt32: {
+      if (!instr->InputAt(0)->IsDoubleRegister()) {
+        __ fld_d(i.InputOperand(0));
+      }
+      __ TruncateX87TOSToI(i.OutputRegister(0));
+      if (!instr->InputAt(0)->IsDoubleRegister()) {
+        __ fstp(0);
+      }
+      break;
+    }
+    case kX87Float64ToFloat32: {
+      InstructionOperand* input = instr->InputAt(0);
+      if (input->IsDoubleRegister()) {
+        __ sub(esp, Immediate(kDoubleSize));
+        __ fstp_s(MemOperand(esp, 0));
+        __ fld_s(MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
+      } else {
+        DCHECK(input->IsDoubleStackSlot());
+        __ fstp(0);
+        __ fld_d(i.InputOperand(0));
+        __ sub(esp, Immediate(kDoubleSize));
+        __ fstp_s(MemOperand(esp, 0));
+        __ fld_s(MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
+      }
+      break;
+    }
+    case kX87Float64ToUint32: {
+      __ push_imm32(-2147483648);
+      if (!instr->InputAt(0)->IsDoubleRegister()) {
+        __ fld_d(i.InputOperand(0));
+      }
+      __ fild_s(Operand(esp, 0));
+      __ fadd(1);
+      __ fstp(0);
+      __ TruncateX87TOSToI(i.OutputRegister(0));
+      __ add(esp, Immediate(kInt32Size));
+      __ add(i.OutputRegister(), Immediate(0x80000000));
+      if (!instr->InputAt(0)->IsDoubleRegister()) {
+        __ fstp(0);
+      }
+      break;
+    }
+    case kX87Float64ExtractHighWord32: {
+      if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ sub(esp, Immediate(kDoubleSize));
+        __ fst_d(MemOperand(esp, 0));
+        __ mov(i.OutputRegister(), MemOperand(esp, kDoubleSize / 2));
+        __ add(esp, Immediate(kDoubleSize));
+      } else {
+        InstructionOperand* input = instr->InputAt(0);
+        USE(input);
+        DCHECK(input->IsDoubleStackSlot());
+        __ mov(i.OutputRegister(), i.InputOperand(0, kDoubleSize / 2));
+      }
+      break;
+    }
+    case kX87Float64ExtractLowWord32: {
+      if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ sub(esp, Immediate(kDoubleSize));
+        __ fst_d(MemOperand(esp, 0));
+        __ mov(i.OutputRegister(), MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
+      } else {
+        InstructionOperand* input = instr->InputAt(0);
+        USE(input);
+        DCHECK(input->IsDoubleStackSlot());
+        __ mov(i.OutputRegister(), i.InputOperand(0));
+      }
+      break;
+    }
+    case kX87Float64InsertHighWord32: {
       __ sub(esp, Immediate(kDoubleSize));
-      // Move values to st(0) and st(1).
-      __ movsd(Operand(esp, 0), i.InputDoubleRegister(1));
-      __ fld_d(Operand(esp, 0));
-      __ movsd(Operand(esp, 0), i.InputDoubleRegister(0));
-      __ fld_d(Operand(esp, 0));
-      // Loop while fprem isn't done.
-      Label mod_loop;
-      __ bind(&mod_loop);
-      // This instructions traps on all kinds inputs, but we are assuming the
-      // floating point control word is set to ignore them all.
-      __ fprem();
-      // The following 2 instruction implicitly use eax.
-      __ fnstsw_ax();
-      __ sahf();
-      __ j(parity_even, &mod_loop);
-      // Move output to stack and clean up.
-      __ fstp(1);
-      __ fstp_d(Operand(esp, 0));
-      __ movsd(i.OutputDoubleRegister(), Operand(esp, 0));
+      __ fstp_d(MemOperand(esp, 0));
+      __ mov(MemOperand(esp, kDoubleSize / 2), i.InputRegister(1));
+      __ fld_d(MemOperand(esp, 0));
       __ add(esp, Immediate(kDoubleSize));
       break;
     }
-    case kSSEFloat64Abs: {
-      // TODO(bmeurer): Use 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psrlq(kScratchDoubleReg, 1);
-      __ andpd(i.OutputDoubleRegister(), kScratchDoubleReg);
-      break;
-    }
-    case kSSEFloat64Neg: {
-      // TODO(bmeurer): Use 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psllq(kScratchDoubleReg, 63);
-      __ xorpd(i.OutputDoubleRegister(), kScratchDoubleReg);
-      break;
-    }
-    case kSSEFloat64Sqrt:
-      __ sqrtsd(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat64Round: {
-      CpuFeatureScope sse_scope(masm(), SSE4_1);
-      RoundingMode const mode =
+    case kX87Float64InsertLowWord32: {
+      __ sub(esp, Immediate(kDoubleSize));
+      __ fstp_d(MemOperand(esp, 0));
+      __ mov(MemOperand(esp, 0), i.InputRegister(1));
+      __ fld_d(MemOperand(esp, 0));
+      __ add(esp, Immediate(kDoubleSize));
+      break;
+    }
+    case kX87Float64Sqrt: {
+      __ fstp(0);
+      __ fld_d(MemOperand(esp, 0));
+      __ fsqrt();
+      __ lea(esp, Operand(esp, kDoubleSize));
+      break;
+    }
+    case kX87Float64Round: {
+      RoundingMode mode =
           static_cast<RoundingMode>(MiscField::decode(instr->opcode()));
-      __ roundsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0), mode);
-      break;
-    }
-    case kSSEFloat32ToFloat64:
-      __ cvtss2sd(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat64ToFloat32:
-      __ cvtsd2ss(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat64ToInt32:
-      __ cvttsd2si(i.OutputRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat64ToUint32: {
-      __ Move(kScratchDoubleReg, -2147483648.0);
-      __ addsd(kScratchDoubleReg, i.InputOperand(0));
-      __ cvttsd2si(i.OutputRegister(), kScratchDoubleReg);
-      __ add(i.OutputRegister(), Immediate(0x80000000));
-      break;
-    }
-    case kSSEInt32ToFloat64:
-      __ cvtsi2sd(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEUint32ToFloat64:
-      __ LoadUint32(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kSSEFloat64ExtractLowWord32:
-      if (instr->InputAt(0)->IsDoubleStackSlot()) {
-        __ mov(i.OutputRegister(), i.InputOperand(0));
-      } else {
-        __ movd(i.OutputRegister(), i.InputDoubleRegister(0));
-      }
-      break;
-    case kSSEFloat64ExtractHighWord32:
-      if (instr->InputAt(0)->IsDoubleStackSlot()) {
-        __ mov(i.OutputRegister(), i.InputOperand(0, kDoubleSize / 2));
-      } else {
-        __ Pextrd(i.OutputRegister(), i.InputDoubleRegister(0), 1);
-      }
-      break;
-    case kSSEFloat64InsertLowWord32:
-      __ Pinsrd(i.OutputDoubleRegister(), i.InputOperand(1), 0);
-      break;
-    case kSSEFloat64InsertHighWord32:
-      __ Pinsrd(i.OutputDoubleRegister(), i.InputOperand(1), 1);
-      break;
-    case kSSEFloat64LoadLowWord32:
-      __ movd(i.OutputDoubleRegister(), i.InputOperand(0));
-      break;
-    case kAVXFloat32Add: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vaddss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat32Sub: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vsubss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat32Mul: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vmulss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat32Div: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vdivss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      // Don't delete this mov. It may improve performance on some CPUs,
-      // when there is a (v)mulss depending on the result.
-      __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
-      break;
-    }
-    case kAVXFloat32Max: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vmaxss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat32Min: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vminss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat64Add: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vaddsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat64Sub: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vsubsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat64Mul: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vmulsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat64Div: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vdivsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      // Don't delete this mov. It may improve performance on some CPUs,
-      // when there is a (v)mulsd depending on the result.
-      __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
-      break;
-    }
-    case kAVXFloat64Max: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vmaxsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat64Min: {
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vminsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-                i.InputOperand(1));
-      break;
-    }
-    case kAVXFloat32Abs: {
-      // TODO(bmeurer): Use RIP relative 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psrlq(kScratchDoubleReg, 33);
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vandps(i.OutputDoubleRegister(), kScratchDoubleReg, i.InputOperand(0));
-      break;
-    }
-    case kAVXFloat32Neg: {
-      // TODO(bmeurer): Use RIP relative 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psllq(kScratchDoubleReg, 31);
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vxorps(i.OutputDoubleRegister(), kScratchDoubleReg, i.InputOperand(0));
-      break;
-    }
-    case kAVXFloat64Abs: {
-      // TODO(bmeurer): Use RIP relative 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psrlq(kScratchDoubleReg, 1);
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vandpd(i.OutputDoubleRegister(), kScratchDoubleReg, i.InputOperand(0));
-      break;
-    }
-    case kAVXFloat64Neg: {
-      // TODO(bmeurer): Use RIP relative 128-bit constants.
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psllq(kScratchDoubleReg, 63);
-      CpuFeatureScope avx_scope(masm(), AVX);
-      __ vxorpd(i.OutputDoubleRegister(), kScratchDoubleReg, i.InputOperand(0));
-      break;
-    }
-    case kIA32Movsxbl:
+      if (mode == MiscField::encode(kRoundDown)) {
+        __ X87SetRC(0x0400);
+      } else {
+        __ X87SetRC(0x0c00);
+      }
+
+      if (!instr->InputAt(0)->IsDoubleRegister()) {
+        InstructionOperand* input = instr->InputAt(0);
+        USE(input);
+        DCHECK(input->IsDoubleStackSlot());
+        __ fstp(0);
+        __ fld_d(i.InputOperand(0));
+      }
+      __ frndint();
+      __ X87SetRC(0x0000);
+      break;
+    }
+    case kX87Float64Cmp: {
+      __ fld_d(MemOperand(esp, kDoubleSize));
+      __ fld_d(MemOperand(esp, 0));
+      __ FCmp();
+      __ lea(esp, Operand(esp, 2 * kDoubleSize));
+      break;
+    }
+    case kX87Movsxbl:
       __ movsx_b(i.OutputRegister(), i.MemoryOperand());
       break;
-    case kIA32Movzxbl:
+    case kX87Movzxbl:
       __ movzx_b(i.OutputRegister(), i.MemoryOperand());
       break;
-    case kIA32Movb: {
+    case kX87Movb: {
       size_t index = 0;
       Operand operand = i.MemoryOperand(&index);
       if (HasImmediateInput(instr, index)) {
         __ mov_b(operand, i.InputInt8(index));
       } else {
         __ mov_b(operand, i.InputRegister(index));
       }
       break;
     }
-    case kIA32Movsxwl:
+    case kX87Movsxwl:
       __ movsx_w(i.OutputRegister(), i.MemoryOperand());
       break;
-    case kIA32Movzxwl:
+    case kX87Movzxwl:
       __ movzx_w(i.OutputRegister(), i.MemoryOperand());
       break;
-    case kIA32Movw: {
+    case kX87Movw: {
       size_t index = 0;
       Operand operand = i.MemoryOperand(&index);
       if (HasImmediateInput(instr, index)) {
         __ mov_w(operand, i.InputInt16(index));
       } else {
         __ mov_w(operand, i.InputRegister(index));
       }
       break;
     }
-    case kIA32Movl:
+    case kX87Movl:
       if (instr->HasOutput()) {
         __ mov(i.OutputRegister(), i.MemoryOperand());
       } else {
         size_t index = 0;
         Operand operand = i.MemoryOperand(&index);
         if (HasImmediateInput(instr, index)) {
           __ mov(operand, i.InputImmediate(index));
         } else {
           __ mov(operand, i.InputRegister(index));
         }
       }
       break;
-    case kIA32Movsd:
+    case kX87Movsd: {
       if (instr->HasOutput()) {
-        __ movsd(i.OutputDoubleRegister(), i.MemoryOperand());
+        X87Register output = i.OutputDoubleRegister();
+        USE(output);
+        DCHECK(output.code() == 0);
+        __ fstp(0);
+        __ fld_d(i.MemoryOperand());
       } else {
         size_t index = 0;
         Operand operand = i.MemoryOperand(&index);
-        __ movsd(operand, i.InputDoubleRegister(index));
+        __ fst_d(operand);
       }
       break;
-    case kIA32Movss:
+    }
+    case kX87Movss: {
       if (instr->HasOutput()) {
-        __ movss(i.OutputDoubleRegister(), i.MemoryOperand());
+        X87Register output = i.OutputDoubleRegister();
+        USE(output);
+        DCHECK(output.code() == 0);
+        __ fstp(0);
+        __ fld_s(i.MemoryOperand());
       } else {
         size_t index = 0;
         Operand operand = i.MemoryOperand(&index);
-        __ movss(operand, i.InputDoubleRegister(index));
+        __ fst_s(operand);
       }
       break;
-    case kIA32Lea: {
+    }
+    case kX87Lea: {
       AddressingMode mode = AddressingModeField::decode(instr->opcode());
       // Shorten "leal" to "addl", "subl" or "shll" if the register allocation
       // and addressing mode just happens to work out. The "addl"/"subl" forms
       // in these cases are faster based on measurements.
       if (mode == kMode_MI) {
         __ Move(i.OutputRegister(), Immediate(i.InputInt32(0)));
       } else if (i.InputRegister(0).is(i.OutputRegister())) {
         if (mode == kMode_MRI) {
           int32_t constant_summand = i.InputInt32(1);
           if (constant_summand > 0) {
@@ skipping 14 matching lines @@
         } else if (mode == kMode_M8) {
           __ shl(i.OutputRegister(), 3);
         } else {
           __ lea(i.OutputRegister(), i.MemoryOperand());
         }
       } else {
         __ lea(i.OutputRegister(), i.MemoryOperand());
       }
       break;
     }
-    case kIA32Push:
+    case kX87Push:
       if (HasImmediateInput(instr, 0)) {
         __ push(i.InputImmediate(0));
       } else {
         __ push(i.InputOperand(0));
       }
       break;
-    case kIA32StoreWriteBarrier: {
+    case kX87PushFloat32:
+      __ lea(esp, Operand(esp, -kFloatSize));
+      if (instr->InputAt(0)->IsDoubleStackSlot()) {
+        __ fld_s(i.InputOperand(0));
+        __ fstp_s(MemOperand(esp, 0));
+      } else if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ fst_s(MemOperand(esp, 0));
+      } else {
+        UNREACHABLE();
+      }
+      break;
+    case kX87PushFloat64:
+      __ lea(esp, Operand(esp, -kDoubleSize));
+      if (instr->InputAt(0)->IsDoubleStackSlot()) {
+        __ fld_d(i.InputOperand(0));
+        __ fstp_d(MemOperand(esp, 0));
+      } else if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ fst_d(MemOperand(esp, 0));
+      } else {
+        UNREACHABLE();
+      }
+      break;
+    case kX87StoreWriteBarrier: {
       Register object = i.InputRegister(0);
       Register value = i.InputRegister(2);
       SaveFPRegsMode mode =
           frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
       if (HasImmediateInput(instr, 1)) {
         int index = i.InputInt32(1);
         Register scratch = i.TempRegister(1);
         __ mov(Operand(object, index), value);
         __ RecordWriteContextSlot(object, index, value, scratch, mode);
       } else {
@@ skipping 13 matching lines @@
     case kCheckedLoadInt16:
       ASSEMBLE_CHECKED_LOAD_INTEGER(movsx_w);
       break;
     case kCheckedLoadUint16:
       ASSEMBLE_CHECKED_LOAD_INTEGER(movzx_w);
       break;
     case kCheckedLoadWord32:
       ASSEMBLE_CHECKED_LOAD_INTEGER(mov);
       break;
     case kCheckedLoadFloat32:
-      ASSEMBLE_CHECKED_LOAD_FLOAT(movss);
+      ASSEMBLE_CHECKED_LOAD_FLOAT(fld_s);
       break;
     case kCheckedLoadFloat64:
-      ASSEMBLE_CHECKED_LOAD_FLOAT(movsd);
+      ASSEMBLE_CHECKED_LOAD_FLOAT(fld_d);
       break;
     case kCheckedStoreWord8:
       ASSEMBLE_CHECKED_STORE_INTEGER(mov_b);
       break;
     case kCheckedStoreWord16:
       ASSEMBLE_CHECKED_STORE_INTEGER(mov_w);
       break;
     case kCheckedStoreWord32:
       ASSEMBLE_CHECKED_STORE_INTEGER(mov);
       break;
     case kCheckedStoreFloat32:
-      ASSEMBLE_CHECKED_STORE_FLOAT(movss);
+      ASSEMBLE_CHECKED_STORE_FLOAT(fst_s);
       break;
     case kCheckedStoreFloat64:
-      ASSEMBLE_CHECKED_STORE_FLOAT(movsd);
+      ASSEMBLE_CHECKED_STORE_FLOAT(fst_d);
       break;
-    case kIA32StackCheck: {
+    case kX87StackCheck: {
       ExternalReference const stack_limit =
           ExternalReference::address_of_stack_limit(isolate());
       __ cmp(esp, Operand::StaticVariable(stack_limit));
       break;
     }
   }
 }  // NOLINT(readability/fn_size)
 
 
 // Assembles a branch after an instruction.
 void CodeGenerator::AssembleArchBranch(Instruction* instr, BranchInfo* branch) {
-  IA32OperandConverter i(this, instr);
+  X87OperandConverter i(this, instr);
   Label::Distance flabel_distance =
       branch->fallthru ? Label::kNear : Label::kFar;
   Label* tlabel = branch->true_label;
   Label* flabel = branch->false_label;
   switch (branch->condition) {
     case kUnorderedEqual:
       __ j(parity_even, flabel, flabel_distance);
     // Fall through.
     case kEqual:
       __ j(equal, tlabel);
@@ skipping 41 matching lines @@
 
 
 void CodeGenerator::AssembleArchJump(RpoNumber target) {
   if (!IsNextInAssemblyOrder(target)) __ jmp(GetLabel(target));
 }
 
 
 // Assembles boolean materializations after an instruction.
 void CodeGenerator::AssembleArchBoolean(Instruction* instr,
                                         FlagsCondition condition) {
-  IA32OperandConverter i(this, instr);
+  X87OperandConverter i(this, instr);
   Label done;
 
   // Materialize a full 32-bit 1 or 0 value. The result register is always the
   // last output of the instruction.
   Label check;
   DCHECK_NE(0u, instr->OutputCount());
   Register reg = i.OutputRegister(instr->OutputCount() - 1);
   Condition cc = no_condition;
   switch (condition) {
     case kUnorderedEqual:
@@ skipping 55 matching lines @@
     __ Move(reg, Immediate(0));
     __ jmp(&done, Label::kNear);
     __ bind(&set);
     __ mov(reg, Immediate(1));
   }
   __ bind(&done);
 }
 
 
 void CodeGenerator::AssembleArchLookupSwitch(Instruction* instr) {
-  IA32OperandConverter i(this, instr);
+  X87OperandConverter i(this, instr);
   Register input = i.InputRegister(0);
   for (size_t index = 2; index < instr->InputCount(); index += 2) {
     __ cmp(input, Immediate(i.InputInt32(index + 0)));
     __ j(equal, GetLabel(i.InputRpo(index + 1)));
   }
   AssembleArchJump(i.InputRpo(1));
 }
 
 
 void CodeGenerator::AssembleArchTableSwitch(Instruction* instr) {
-  IA32OperandConverter i(this, instr);
+  X87OperandConverter i(this, instr);
   Register input = i.InputRegister(0);
   size_t const case_count = instr->InputCount() - 2;
   Label** cases = zone()->NewArray<Label*>(case_count);
   for (size_t index = 0; index < case_count; ++index) {
     cases[index] = GetLabel(i.InputRpo(index + 2));
   }
   Label* const table = AddJumpTable(cases, case_count);
   __ cmp(input, Immediate(case_count));
   __ j(above_equal, GetLabel(i.InputRpo(1)));
   __ jmp(Operand::JumpTable(input, times_4, table));
 }
 
 
 void CodeGenerator::AssembleDeoptimizerCall(
     int deoptimization_id, Deoptimizer::BailoutType bailout_type) {
   Address deopt_entry = Deoptimizer::GetDeoptimizationEntry(
       isolate(), deoptimization_id, bailout_type);
   __ call(deopt_entry, RelocInfo::RUNTIME_ENTRY);
 }
 
 
-// The calling convention for JSFunctions on IA32 passes arguments on the
+// The calling convention for JSFunctions on X87 passes arguments on the
 // stack and the JSFunction and context in EDI and ESI, respectively, thus
 // the steps of the call look as follows:
 
 // --{ before the call instruction }--------------------------------------------
 //                                                         |  caller frame |
 //                                                         ^ esp           ^ ebp
 
 // --{ push arguments and setup ESI, EDI }--------------------------------------
 //                                       | args + receiver |  caller frame |
 //                                       ^ esp                             ^ ebp
@@ skipping 34 matching lines @@
 // --{ pop ebp }-----------------------------------------------------------
 // |                               | RET | args + receiver |  caller frame |
 //                                 ^ esp                                   ^ ebp
 
 // --{ ret #A+1 }-----------------------------------------------------------
 // |                                                       |  caller frame |
 //                                                         ^ esp           ^ ebp
 
 
 // Runtime function calls are accomplished by doing a stub call to the
-// CEntryStub (a real code object). On IA32 passes arguments on the
+// CEntryStub (a real code object). On X87 passes arguments on the
 // stack, the number of arguments in EAX, the address of the runtime function
 // in EBX, and the context in ESI.
 
 // --{ before the call instruction }--------------------------------------------
 //                                                         |  caller frame |
 //                                                         ^ esp           ^ ebp
 
 // --{ push arguments and setup EAX, EBX, and ESI }-----------------------------
 //                                       | args + receiver |  caller frame |
 //                                       ^ esp                             ^ ebp
@@ skipping 102 matching lines @@
     // TODO(titzer): cannot address target function == local #-1
     __ mov(edi, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset));
     DCHECK(stack_slots >= frame()->GetOsrStackSlotCount());
     stack_slots -= frame()->GetOsrStackSlotCount();
   }
 
   if (stack_slots > 0) {
     // Allocate the stack slots used by this frame.
     __ sub(esp, Immediate(stack_slots * kPointerSize));
   }
+
+  // Initailize FPU state.
+  __ fninit();
+  __ fld1();
 }
 
 
 void CodeGenerator::AssembleReturn() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
   int stack_slots = frame()->GetSpillSlotCount();
   if (descriptor->kind() == CallDescriptor::kCallAddress) {
     const RegList saves = descriptor->CalleeSavedRegisters();
     if (frame()->GetRegisterSaveAreaSize() > 0) {
       // Remove this frame's spill slots first.
@@ skipping 23 matching lines @@
                         : 0;
     __ Ret(pop_count * kPointerSize, ebx);
   } else {
     __ ret(0);
   }
 }
 
 
 void CodeGenerator::AssembleMove(InstructionOperand* source,
                                  InstructionOperand* destination) {
-  IA32OperandConverter g(this, NULL);
+  X87OperandConverter g(this, NULL);
   // Dispatch on the source and destination operand kinds.  Not all
   // combinations are possible.
   if (source->IsRegister()) {
     DCHECK(destination->IsRegister() || destination->IsStackSlot());
     Register src = g.ToRegister(source);
     Operand dst = g.ToOperand(destination);
     __ mov(dst, src);
   } else if (source->IsStackSlot()) {
     DCHECK(destination->IsRegister() || destination->IsStackSlot());
     Operand src = g.ToOperand(source);
@@ skipping 37 matching lines @@
     } else if (destination->IsRegister()) {
       Register dst = g.ToRegister(destination);
       __ Move(dst, g.ToImmediate(source));
     } else if (destination->IsStackSlot()) {
       Operand dst = g.ToOperand(destination);
       __ Move(dst, g.ToImmediate(source));
     } else if (src_constant.type() == Constant::kFloat32) {
       // TODO(turbofan): Can we do better here?
       uint32_t src = bit_cast<uint32_t>(src_constant.ToFloat32());
       if (destination->IsDoubleRegister()) {
-        XMMRegister dst = g.ToDoubleRegister(destination);
-        __ Move(dst, src);
+        __ sub(esp, Immediate(kInt32Size));
+        __ mov(MemOperand(esp, 0), Immediate(src));
+        // always only push one value into the x87 stack.
+        __ fstp(0);
+        __ fld_s(MemOperand(esp, 0));
+        __ add(esp, Immediate(kInt32Size));
       } else {
         DCHECK(destination->IsDoubleStackSlot());
         Operand dst = g.ToOperand(destination);
         __ Move(dst, Immediate(src));
       }
     } else {
       DCHECK_EQ(Constant::kFloat64, src_constant.type());
       uint64_t src = bit_cast<uint64_t>(src_constant.ToFloat64());
       uint32_t lower = static_cast<uint32_t>(src);
       uint32_t upper = static_cast<uint32_t>(src >> 32);
       if (destination->IsDoubleRegister()) {
-        XMMRegister dst = g.ToDoubleRegister(destination);
-        __ Move(dst, src);
+        __ sub(esp, Immediate(kDoubleSize));
+        __ mov(MemOperand(esp, 0), Immediate(lower));
+        __ mov(MemOperand(esp, kInt32Size), Immediate(upper));
+        // always only push one value into the x87 stack.
+        __ fstp(0);
+        __ fld_d(MemOperand(esp, 0));
+        __ add(esp, Immediate(kDoubleSize));
       } else {
         DCHECK(destination->IsDoubleStackSlot());
         Operand dst0 = g.ToOperand(destination);
         Operand dst1 = g.HighOperand(destination);
         __ Move(dst0, Immediate(lower));
         __ Move(dst1, Immediate(upper));
       }
     }
   } else if (source->IsDoubleRegister()) {
-    XMMRegister src = g.ToDoubleRegister(source);
-    if (destination->IsDoubleRegister()) {
-      XMMRegister dst = g.ToDoubleRegister(destination);
-      __ movaps(dst, src);
-    } else {
-      DCHECK(destination->IsDoubleStackSlot());
-      Operand dst = g.ToOperand(destination);
-      __ movsd(dst, src);
+    DCHECK(destination->IsDoubleStackSlot());
+    Operand dst = g.ToOperand(destination);
+    auto allocated = AllocatedOperand::cast(*source);
+    switch (allocated.machine_type()) {
+      case kRepFloat32:
+        __ fst_s(dst);
+        break;
+      case kRepFloat64:
+        __ fst_d(dst);
+        break;
+      default:
+        UNREACHABLE();
     }
   } else if (source->IsDoubleStackSlot()) {
     DCHECK(destination->IsDoubleRegister() || destination->IsDoubleStackSlot());
     Operand src = g.ToOperand(source);
+    auto allocated = AllocatedOperand::cast(*source);
     if (destination->IsDoubleRegister()) {
-      XMMRegister dst = g.ToDoubleRegister(destination);
-      __ movsd(dst, src);
+      // always only push one value into the x87 stack.
+      __ fstp(0);
+      switch (allocated.machine_type()) {
+        case kRepFloat32:
+          __ fld_s(src);
+          break;
+        case kRepFloat64:
+          __ fld_d(src);
+          break;
+        default:
+          UNREACHABLE();
+      }
     } else {
       Operand dst = g.ToOperand(destination);
-      __ movsd(kScratchDoubleReg, src);
-      __ movsd(dst, kScratchDoubleReg);
+      switch (allocated.machine_type()) {
+        case kRepFloat32:
+          __ fld_s(src);
+          __ fstp_s(dst);
+          break;
+        case kRepFloat64:
+          __ fld_d(src);
+          __ fstp_d(dst);
+          break;
+        default:
+          UNREACHABLE();
+      }
     }
   } else {
     UNREACHABLE();
   }
 }
 
 
 void CodeGenerator::AssembleSwap(InstructionOperand* source,
                                  InstructionOperand* destination) {
-  IA32OperandConverter g(this, NULL);
+  X87OperandConverter g(this, NULL);
   // Dispatch on the source and destination operand kinds.  Not all
   // combinations are possible.
   if (source->IsRegister() && destination->IsRegister()) {
     // Register-register.
     Register src = g.ToRegister(source);
     Register dst = g.ToRegister(destination);
     __ xchg(dst, src);
   } else if (source->IsRegister() && destination->IsStackSlot()) {
     // Register-memory.
     __ xchg(g.ToRegister(source), g.ToOperand(destination));
   } else if (source->IsStackSlot() && destination->IsStackSlot()) {
     // Memory-memory.
     Operand src = g.ToOperand(source);
     Operand dst = g.ToOperand(destination);
     __ push(dst);
     __ push(src);
     __ pop(dst);
     __ pop(src);
   } else if (source->IsDoubleRegister() && destination->IsDoubleRegister()) {
-    // XMM register-register swap.
-    XMMRegister src = g.ToDoubleRegister(source);
-    XMMRegister dst = g.ToDoubleRegister(destination);
-    __ movaps(kScratchDoubleReg, src);
-    __ movaps(src, dst);
-    __ movaps(dst, kScratchDoubleReg);
+    UNREACHABLE();
   } else if (source->IsDoubleRegister() && destination->IsDoubleStackSlot()) {
-    // XMM register-memory swap.
-    XMMRegister reg = g.ToDoubleRegister(source);
-    Operand other = g.ToOperand(destination);
-    __ movsd(kScratchDoubleReg, other);
-    __ movsd(other, reg);
-    __ movaps(reg, kScratchDoubleReg);
+    auto allocated = AllocatedOperand::cast(*source);
+    switch (allocated.machine_type()) {
+      case kRepFloat32:
+        __ fld_s(g.ToOperand(destination));
+        __ fxch();
+        __ fstp_s(g.ToOperand(destination));
+        break;
+      case kRepFloat64:
+        __ fld_d(g.ToOperand(destination));
+        __ fxch();
+        __ fstp_d(g.ToOperand(destination));
+        break;
+      default:
+        UNREACHABLE();
+    }
   } else if (source->IsDoubleStackSlot() && destination->IsDoubleStackSlot()) {
-    // Double-width memory-to-memory.
-    Operand src0 = g.ToOperand(source);
-    Operand src1 = g.HighOperand(source);
-    Operand dst0 = g.ToOperand(destination);
-    Operand dst1 = g.HighOperand(destination);
-    __ movsd(kScratchDoubleReg, dst0);  // Save destination in scratch register.
-    __ push(src0);  // Then use stack to copy source to destination.
-    __ pop(dst0);
-    __ push(src1);
-    __ pop(dst1);
-    __ movsd(src0, kScratchDoubleReg);
+    auto allocated = AllocatedOperand::cast(*source);
+    switch (allocated.machine_type()) {
+      case kRepFloat32:
+        __ fld_s(g.ToOperand(source));
+        __ fld_s(g.ToOperand(destination));
+        __ fstp_s(g.ToOperand(source));
+        __ fstp_s(g.ToOperand(destination));
+        break;
+      case kRepFloat64:
+        __ fld_d(g.ToOperand(source));
+        __ fld_d(g.ToOperand(destination));
+        __ fstp_d(g.ToOperand(source));
+        __ fstp_d(g.ToOperand(destination));
+        break;
+      default:
+        UNREACHABLE();
+    }
   } else {
     // No other combinations are possible.
     UNREACHABLE();
   }
 }
 
 
 void CodeGenerator::AssembleJumpTable(Label** targets, size_t target_count) {
   for (size_t index = 0; index < target_count; ++index) {
     __ dd(targets[index]);
@@ skipping 16 matching lines @@
     }
   }
   MarkLazyDeoptSite();
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8