[x86] Use AVX in Crankshaft when available.

R=verwaest@chromium.org

Review URL: https://codereview.chromium.org/860003002

Cr-Commit-Position: refs/heads/master@{#26159}

src/ia32/lithium-codegen-ia32.cc

@@ -1983,19 +1983,43 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
   XMMRegister result = ToDoubleRegister(instr->result());
   switch (instr->op()) {
     case Token::ADD:
-      __ addsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vaddsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ addsd(left, right);
+      }
       break;
     case Token::SUB:
-      __ subsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vsubsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ subsd(left, right);
+      }
       break;
     case Token::MUL:
-      __ mulsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vmulsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ 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);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vdivsd(result, left, right);
+      } 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);
+      }
       break;
     case Token::MOD: {
       // Pass two doubles as arguments on the stack.

src/ia32/lithium-ia32.cc

@@ -765,7 +765,8 @@ LInstruction* LChunkBuilder::DoArithmeticD(Token::Value op,
     LOperand* left = UseRegisterAtStart(instr->BetterLeftOperand());
     LOperand* right = UseRegisterAtStart(instr->BetterRightOperand());
     LArithmeticD* result = new(zone()) LArithmeticD(op, left, right);
-    return DefineSameAsFirst(result);
+    return CpuFeatures::IsSupported(AVX) ? DefineAsRegister(result)
+                                         : DefineSameAsFirst(result);
   }
 }
 

src/x64/lithium-codegen-x64.cc

@@ -2015,23 +2015,45 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
   XMMRegister left = ToDoubleRegister(instr->left());
   XMMRegister right = ToDoubleRegister(instr->right());
   XMMRegister result = ToDoubleRegister(instr->result());
-  // All operations except MOD are computed in-place.
-  DCHECK(instr->op() == Token::MOD || left.is(result));
   switch (instr->op()) {
     case Token::ADD:
-      __ addsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vaddsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ addsd(left, right);
+      }
       break;
     case Token::SUB:
-       __ subsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vsubsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ subsd(left, right);
+      }
        break;
     case Token::MUL:
-      __ mulsd(left, right);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vmulsd(result, left, right);
+      } else {
+        DCHECK(result.is(left));
+        __ 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);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope scope(masm(), AVX);
+        __ vdivsd(result, left, right);
+      } 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);
+      }
       break;
     case Token::MOD: {
       XMMRegister xmm_scratch = double_scratch0();

src/x64/lithium-x64.cc

@@ -748,7 +748,8 @@ LInstruction* LChunkBuilder::DoArithmeticD(Token::Value op,
     LOperand* left = UseRegisterAtStart(instr->BetterLeftOperand());
     LOperand* right = UseRegisterAtStart(instr->BetterRightOperand());
     LArithmeticD* result = new(zone()) LArithmeticD(op, left, right);
-    return DefineSameAsFirst(result);
+    return CpuFeatures::IsSupported(AVX) ? DefineAsRegister(result)
+                                         : DefineSameAsFirst(result);
   }
 }