Optimizing: Add Non Temporal Move support for x86

Add moves that don't pollute the data cache.  These can be used for
assigning large data structures.

Change-Id: I14d91ba6264f5ce2f128033d65d59b2536426643
Signed-off-by: Mark Mendell <mark.p.mendell@intel.com>
diff --git a/compiler/utils/x86/assembler_x86.cc b/compiler/utils/x86/assembler_x86.cc
index 44efc65..e01d476 100644
--- a/compiler/utils/x86/assembler_x86.cc
+++ b/compiler/utils/x86/assembler_x86.cc
@@ -145,6 +145,13 @@
   EmitLabel(lbl, dst.length_ + 5);
 }
 
+void X86Assembler::movntl(const Address& dst, Register src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x0F);
+  EmitUint8(0xC3);
+  EmitOperand(src, dst);
+}
+
 void X86Assembler::bswapl(Register dst) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x0F);
diff --git a/compiler/utils/x86/assembler_x86.h b/compiler/utils/x86/assembler_x86.h
index e2abcde..07958eb 100644
--- a/compiler/utils/x86/assembler_x86.h
+++ b/compiler/utils/x86/assembler_x86.h
@@ -231,6 +231,8 @@
   void movl(const Address& dst, const Immediate& imm);
   void movl(const Address& dst, Label* lbl);
 
+  void movntl(const Address& dst, Register src);
+
   void bswapl(Register dst);
 
   void movzxb(Register dst, ByteRegister src);
diff --git a/compiler/utils/x86/assembler_x86_test.cc b/compiler/utils/x86/assembler_x86_test.cc
index 0e8c4ae..f778608 100644
--- a/compiler/utils/x86/assembler_x86_test.cc
+++ b/compiler/utils/x86/assembler_x86_test.cc
@@ -105,6 +105,16 @@
   DriverStr(expected, "movl");
 }
 
+TEST_F(AssemblerX86Test, Movntl) {
+  GetAssembler()->movntl(x86::Address(x86::EDI, x86::EBX, x86::TIMES_4, 12), x86::EAX);
+  GetAssembler()->movntl(x86::Address(x86::EDI, 0), x86::EAX);
+  const char* expected =
+    "movntil %EAX, 0xc(%EDI,%EBX,4)\n"
+    "movntil %EAX, (%EDI)\n";
+
+  DriverStr(expected, "movntl");
+}
+
 TEST_F(AssemblerX86Test, psrlq) {
   GetAssembler()->psrlq(x86::XMM0, CreateImmediate(32));
   const char* expected = "psrlq $0x20, %xmm0\n";