Inline two argument shuffle operations on IA32 and X64.

runtime/vm/intermediate_language_x64.cc

Index: runtime/vm/intermediate_language_x64.cc
diff --git a/runtime/vm/intermediate_language_x64.cc b/runtime/vm/intermediate_language_x64.cc
index 3ab4d0f7e1d0f850d9de3dfd19bd0df3fb934152..34d4805f5e98d22c2063c4d43b187bbbc9737cac 100644
--- a/runtime/vm/intermediate_language_x64.cc
+++ b/runtime/vm/intermediate_language_x64.cc
@@ -3412,6 +3412,45 @@ void Float32x4ToUint32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
+LocationSummary* Float32x4TwoArgShuffleInstr::MakeLocationSummary() const {
+  const intptr_t kNumInputs = 2;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  summary->set_in(0, Location::RequiresFpuRegister());
+  summary->set_in(1, Location::RequiresFpuRegister());
+  summary->set_out(Location::SameAsFirstInput());
+  return summary;
+}
+
+
+void Float32x4TwoArgShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  XmmRegister left = locs()->in(0).fpu_reg();
+  XmmRegister right = locs()->in(1).fpu_reg();
+
+  ASSERT(locs()->out().fpu_reg() == left);
+
+  switch (op_kind()) {
+    case MethodRecognizer::kFloat32x4WithZWInXY:
+      __ movhlps(left, right);
+    break;
+    case MethodRecognizer::kFloat32x4InterleaveXY:
+      __ unpcklps(left, right);
+    break;
+    case MethodRecognizer::kFloat32x4InterleaveZW:
+      __ unpckhps(left, right);
+    break;
+    case MethodRecognizer::kFloat32x4InterleaveXYPairs:
+      __ unpcklpd(left, right);
+    break;
+    case MethodRecognizer::kFloat32x4InterleaveZWPairs:
+      __ unpckhpd(left, right);
+    break;
+    default: UNREACHABLE();
+  }
+}
+
+
 LocationSummary* Uint32x4BoolConstructorInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 4;
   const intptr_t kNumTemps = 1;