Fix parallel move code for swapping SIMD registers on ARM and ARM64.

runtime/vm/flow_graph_compiler_arm.cc

Index: runtime/vm/flow_graph_compiler_arm.cc
===================================================================
--- runtime/vm/flow_graph_compiler_arm.cc	(revision 42037)
+++ runtime/vm/flow_graph_compiler_arm.cc	(working copy)
@@ -1690,11 +1690,19 @@
     Exchange(source.base_reg(), source.ToStackSlotOffset(),
              destination.base_reg(), destination.ToStackSlotOffset());
   } else if (source.IsFpuRegister() && destination.IsFpuRegister()) {
-    const DRegister dst = EvenDRegisterOf(destination.fpu_reg());
-    DRegister src = EvenDRegisterOf(source.fpu_reg());
-    __ vmovd(DTMP, src);
-    __ vmovd(src, dst);
-    __ vmovd(dst, DTMP);
+    if (TargetCPUFeatures::neon_supported()) {
+      QRegister dst = destination.fpu_reg();

[zerny-google, 2014/12/01 15:23:02]

Nit: should we keep const here and below?

[Florian Schneider, 2014/12/01 15:24:40]

Done.

+      QRegister src = source.fpu_reg();
+      __ vmovq(QTMP, src);
+      __ vmovq(src, dst);
+      __ vmovq(dst, QTMP);
+    } else {
+      DRegister dst = EvenDRegisterOf(destination.fpu_reg());
+      DRegister src = EvenDRegisterOf(source.fpu_reg());
+      __ vmovd(DTMP, src);
+      __ vmovd(src, dst);
+      __ vmovd(dst, DTMP);
+    }
   } else if (source.IsFpuRegister() || destination.IsFpuRegister()) {
     ASSERT(destination.IsDoubleStackSlot() ||
            destination.IsQuadStackSlot() ||