[x64] Emit vmovmskpd when AVX is enabled.

BUG=v8:4406
LOG=N

Review URL: https://codereview.chromium.org/1411023002

Cr-Commit-Position: refs/heads/master@{#31350}

src/x64/assembler-x64.cc

@@ -3558,15 +3558,6 @@ void Assembler::vmovq(Register dst, XMMRegister src) {
 }
 
 
-void Assembler::vmovapd(XMMRegister dst, XMMRegister src) {
-  DCHECK(IsEnabled(AVX));
-  EnsureSpace ensure_space(this);
-  emit_vex_prefix(dst, xmm0, src, kL128, k66, k0F, kWIG);
-  emit(0x28);
-  emit_sse_operand(dst, src);
-}
-
-
 void Assembler::vsd(byte op, XMMRegister dst, XMMRegister src1,
                     XMMRegister src2, SIMDPrefix pp, LeadingOpcode m, VexW w) {
   DCHECK(IsEnabled(AVX));

src/x64/assembler-x64.h

@@ -1281,7 +1281,6 @@ class Assembler : public AssemblerBase {
   void vmovq(XMMRegister dst, const Operand& src);
   void vmovq(Register dst, XMMRegister src);
 
-  void vmovapd(XMMRegister dst, XMMRegister src);
   void vmovsd(XMMRegister dst, const Operand& src) {
     vsd(0x10, dst, xmm0, src);
   }
@@ -1596,6 +1595,12 @@ class Assembler : public AssemblerBase {
   void rorxl(Register dst, Register src, byte imm8);
   void rorxl(Register dst, const Operand& src, byte imm8);
 
+  void vmovapd(XMMRegister dst, XMMRegister src) { vpd(0x28, dst, xmm0, src); }
+  void vmovmskpd(Register dst, XMMRegister src) {
+    XMMRegister idst = {dst.code()};
+    vpd(0x50, idst, xmm0, src);
+  }
+
 #define PACKED_OP_LIST(V) \
   V(and, 0x54)            \
   V(xor, 0x57)

src/x64/disasm-x64.cc

@@ -1201,6 +1201,10 @@ int DisassemblerX64::AVXInstruction(byte* data) {
         AppendToBuffer("vucomisd %s,", NameOfXMMRegister(regop));
         current += PrintRightXMMOperand(current);
         break;
+      case 0x50:
+        AppendToBuffer("vmovmskpd %s,", NameOfCPURegister(regop));
+        current += PrintRightXMMOperand(current);
+        break;
       case 0x54:
         AppendToBuffer("vandpd %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));

src/x64/lithium-codegen-x64.cc

@@ -2410,7 +2410,7 @@ void LCodeGen::DoCompareMinusZeroAndBranch(LCompareMinusZeroAndBranch* instr) {
     __ Xorpd(xmm_scratch, xmm_scratch);
     __ ucomisd(xmm_scratch, value);
     EmitFalseBranch(instr, not_equal);
-    __ movmskpd(kScratchRegister, value);
+    __ Movmskpd(kScratchRegister, value);
     __ testl(kScratchRegister, Immediate(1));
     EmitBranch(instr, not_zero);
   } else {
@@ -3623,8 +3623,8 @@ void LCodeGen::DoMathFloor(LMathFloor* instr) {
       // Check for negative zero.
       Label positive_sign;
       __ j(above, &positive_sign, Label::kNear);
-      __ movmskpd(output_reg, input_reg);
-      __ testq(output_reg, Immediate(1));
+      __ Movmskpd(output_reg, input_reg);
+      __ testl(output_reg, Immediate(1));
       DeoptimizeIf(not_zero, instr, Deoptimizer::kMinusZero);
       __ Set(output_reg, 0);
       __ jmp(&done);
@@ -4926,8 +4926,8 @@ void LCodeGen::EmitNumberUntagD(LNumberUntagD* instr, Register input_reg,
       __ Xorpd(xmm_scratch, xmm_scratch);
       __ ucomisd(xmm_scratch, result_reg);
       __ j(not_equal, &done, Label::kNear);
-      __ movmskpd(kScratchRegister, result_reg);
-      __ testq(kScratchRegister, Immediate(1));
+      __ Movmskpd(kScratchRegister, result_reg);
+      __ testl(kScratchRegister, Immediate(1));
       DeoptimizeIf(not_zero, instr, Deoptimizer::kMinusZero);
     }
     __ jmp(&done, Label::kNear);
@@ -5001,7 +5001,7 @@ void LCodeGen::DoDeferredTaggedToI(LTaggedToI* instr, Label* done) {
     if (instr->hydrogen()->GetMinusZeroMode() == FAIL_ON_MINUS_ZERO) {
       __ testl(input_reg, input_reg);
       __ j(not_zero, done);
-      __ movmskpd(input_reg, xmm0);
+      __ Movmskpd(input_reg, xmm0);
       __ andl(input_reg, Immediate(1));
       DeoptimizeIf(not_zero, instr, Deoptimizer::kMinusZero);
     }

src/x64/macro-assembler-x64.cc

@@ -2619,6 +2619,16 @@ void MacroAssembler::Movq(Register dst, XMMRegister src) {
 }
 
 
+void MacroAssembler::Movmskpd(Register dst, XMMRegister src) {
+  if (CpuFeatures::IsSupported(AVX)) {
+    CpuFeatureScope scope(this, AVX);
+    vmovmskpd(dst, src);
+  } else {
+    movmskpd(dst, src);
+  }
+}
+
+
 void MacroAssembler::Xorpd(XMMRegister dst, XMMRegister src) {
   if (CpuFeatures::IsSupported(AVX)) {
     CpuFeatureScope scope(this, AVX);
@@ -3420,7 +3430,7 @@ void MacroAssembler::DoubleToI(Register result_reg, XMMRegister input_reg,
     // only have to test if we got -0 as an input.
     testl(result_reg, result_reg);
     j(not_zero, &done, Label::kNear);
-    movmskpd(result_reg, input_reg);
+    Movmskpd(result_reg, input_reg);
     // Bit 0 contains the sign of the double in input_reg.
     // If input was positive, we are ok and return 0, otherwise
     // jump to minus_zero.

src/x64/macro-assembler-x64.h

@@ -915,6 +915,8 @@ class MacroAssembler: public Assembler {
   void Movq(XMMRegister dst, Register src);
   void Movq(Register dst, XMMRegister src);
 
+  void Movmskpd(Register dst, XMMRegister src);
+
   void Xorpd(XMMRegister dst, XMMRegister src);
 
   // Control Flow

test/cctest/test-assembler-x64.cc

@@ -1400,6 +1400,17 @@ TEST(AssemblerX64AVX_sd) {
     __ cmpq(rcx, rdx);
     __ j(not_equal, &exit);
 
+    // Test vmovmskpd
+    __ movl(rax, Immediate(12));
+    __ movq(rdx, V8_INT64_C(0x426D1A94A2000000));  // 1.0e12
+    __ vmovq(xmm6, rdx);
+    __ movq(rdx, V8_INT64_C(0xC26D1A94A2000000));  // -1.0e12
+    __ vmovq(xmm7, rdx);
+    __ shufps(xmm6, xmm7, 0x44);
+    __ vmovmskpd(rdx, xmm6);
+    __ cmpl(rdx, Immediate(2));
+    __ j(not_equal, &exit);
+
     __ movl(rdx, Immediate(6));
     __ vcvtlsi2sd(xmm6, xmm6, rdx);
     __ movl(Operand(rsp, 0), Immediate(5));

test/cctest/test-disasm-x64.cc

@@ -518,7 +518,6 @@ TEST(DisasmX64) {
       __ vmovq(xmm9, Operand(rbx, rcx, times_4, 10000));
       __ vmovq(r9, xmm6);
 
-      __ vmovapd(xmm7, xmm0);
       __ vmovsd(xmm6, xmm2);
       __ vmovsd(xmm9, Operand(rbx, rcx, times_4, 10000));
       __ vmovsd(Operand(rbx, rcx, times_4, 10000), xmm0);
@@ -549,6 +548,9 @@ TEST(DisasmX64) {
       __ vcvttsd2siq(rdi, xmm9);
       __ vcvttsd2siq(r8, Operand(r9, rbx, times_4, 10000));
 
+      __ vmovapd(xmm7, xmm0);
+      __ vmovmskpd(r9, xmm4);
+
       __ vandps(xmm0, xmm9, xmm2);
       __ vandps(xmm9, xmm1, Operand(rbx, rcx, times_4, 10000));
       __ vxorps(xmm0, xmm1, xmm9);