[x86] Allow (v)divsd->(v)mulsd to execute in parallel.

This tweak was already present in CrankShaft for the non-AVX case. As it
turns out, it's also relevant even with AVX. Now the same optimization
is applied in case of TurboFan as well.

R=dcarney@chromium.org

Review URL: https://codereview.chromium.org/1081033003

Cr-Commit-Position: refs/heads/master@{#27774}

src/compiler/ia32/code-generator-ia32.cc

@@ -468,6 +468,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       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));
@@ -506,6 +509,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       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));
@@ -629,6 +635,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       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: {
@@ -665,6 +674,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       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: {

src/compiler/x64/code-generator-x64.cc

@@ -726,6 +726,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       break;
     case kSSEFloat32Div:
       ASSEMBLE_SSE_BINOP(divss);
+      // 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 kSSEFloat32Abs: {
       // TODO(bmeurer): Use RIP relative 128-bit constants.
@@ -767,6 +770,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       break;
     case kSSEFloat64Div:
       ASSEMBLE_SSE_BINOP(divsd);
+      // 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 kSSEFloat64Mod: {
       __ subq(rsp, Immediate(kDoubleSize));
@@ -919,6 +925,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       break;
     case kAVXFloat32Div:
       ASSEMBLE_AVX_BINOP(vdivss);
+      // 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:
       ASSEMBLE_AVX_BINOP(vmaxss);
@@ -946,6 +955,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       break;
     case kAVXFloat64Div:
       ASSEMBLE_AVX_BINOP(vdivsd);
+      // 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:
       ASSEMBLE_AVX_BINOP(vmaxsd);

src/ia32/lithium-codegen-ia32.cc

@@ -2040,10 +2040,10 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
       } else {
         DCHECK(result.is(left));
         __ 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);
       }
+      // Don't delete this mov. It may improve performance on some CPUs,
+      // when there is a (v)mulsd depending on the result
+      __ movaps(result, result);
       break;
     case Token::MOD: {
       // Pass two doubles as arguments on the stack.

src/x64/lithium-codegen-x64.cc

@@ -2076,10 +2076,10 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
       } else {
         DCHECK(result.is(left));
         __ 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);
       }
+      // Don't delete this mov. It may improve performance on some CPUs,
+      // when there is a (v)mulsd depending on the result
+      __ movaps(result, result);
       break;
     case Token::MOD: {
       XMMRegister xmm_scratch = double_scratch0();