ART: Optimizing compiler intrinsics

Add intrinsics infrastructure to the optimizing compiler.

Add almost all intrinsics supported by Quick to the x86-64 backend.
Further intrinsics require more assembler support.

Change-Id: I48de9b44c82886bb298d16e74e12a9506b8e8807

24 files changed, 1898 insertions, 75 deletions

diff --git a/compiler/Android.mk b/compiler/Android.mk
index db338f0..1f4a36c 100644
--- a/compiler/Android.mk
+++ b/compiler/Android.mk
@@ -98,6 +98,8 @@ LIBART_COMPILER_SRC_FILES := \
 	optimizing/gvn.cc \
 	optimizing/inliner.cc \
 	optimizing/instruction_simplifier.cc \
+	optimizing/intrinsics.cc \
+	optimizing/intrinsics_x86_64.cc \
 	optimizing/locations.cc \
 	optimizing/nodes.cc \
 	optimizing/optimization.cc \
diff --git a/compiler/dex/quick/dex_file_method_inliner.cc b/compiler/dex/quick/dex_file_method_inliner.cc
index 3039852..84c0d93 100644
--- a/compiler/dex/quick/dex_file_method_inliner.cc
+++ b/compiler/dex/quick/dex_file_method_inliner.cc
@@ -293,9 +293,9 @@ const DexFileMethodInliner::IntrinsicDef DexFileMethodInliner::kIntrinsicMethods
     { { kClassCache ## c, kNameCache ## n, kProtoCache ## p }, { o, kInlineIntrinsic, { d } } }
 
     INTRINSIC(JavaLangDouble, DoubleToRawLongBits, D_J, kIntrinsicDoubleCvt, 0),
-    INTRINSIC(JavaLangDouble, LongBitsToDouble, J_D, kIntrinsicDoubleCvt, 0),
+    INTRINSIC(JavaLangDouble, LongBitsToDouble, J_D, kIntrinsicDoubleCvt, kIntrinsicFlagToFloatingPoint),
     INTRINSIC(JavaLangFloat, FloatToRawIntBits, F_I, kIntrinsicFloatCvt, 0),
-    INTRINSIC(JavaLangFloat, IntBitsToFloat, I_F, kIntrinsicFloatCvt, 0),
+    INTRINSIC(JavaLangFloat, IntBitsToFloat, I_F, kIntrinsicFloatCvt, kIntrinsicFlagToFloatingPoint),
 
     INTRINSIC(JavaLangInteger, ReverseBytes, I_I, kIntrinsicReverseBytes, k32),
     INTRINSIC(JavaLangLong, ReverseBytes, J_J, kIntrinsicReverseBytes, k64),
diff --git a/compiler/optimizing/builder.cc b/compiler/optimizing/builder.cc
index f6ca6c7..9c2facb 100644
--- a/compiler/optimizing/builder.cc
+++ b/compiler/optimizing/builder.cc
@@ -604,7 +604,7 @@ bool HGraphBuilder::BuildInvoke(const Instruction& instruction,
   HInvoke* invoke = nullptr;
   if (optimized_invoke_type == kVirtual) {
     invoke = new (arena_) HInvokeVirtual(
-        arena_, number_of_arguments, return_type, dex_pc, table_index);
+        arena_, number_of_arguments, return_type, dex_pc, method_idx, table_index);
   } else if (optimized_invoke_type == kInterface) {
     invoke = new (arena_) HInvokeInterface(
         arena_, number_of_arguments, return_type, dex_pc, method_idx, table_index);
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 1862061..c4ba0fd 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -1190,7 +1190,7 @@ void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirec
       kLoadWord, temp, temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value());
   // temp = temp[index_in_cache]
   __ LoadFromOffset(
-      kLoadWord, temp, temp, CodeGenerator::GetCacheOffset(invoke->GetIndexInDexCache()));
+      kLoadWord, temp, temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex()));
   // LR = temp[offset_of_quick_compiled_code]
   __ LoadFromOffset(kLoadWord, LR, temp,
                      mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 7b19f44..6d2c3de 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -1975,7 +1975,7 @@ void InstructionCodeGeneratorARM64::VisitInvokeStaticOrDirect(HInvokeStaticOrDir
   // Make sure that ArtMethod* is passed in W0 as per the calling convention
   DCHECK(temp.Is(w0));
   size_t index_in_cache = mirror::Array::DataOffset(kHeapRefSize).SizeValue() +
-    invoke->GetIndexInDexCache() * kHeapRefSize;
+    invoke->GetDexMethodIndex() * kHeapRefSize;
 
   // TODO: Implement all kinds of calls:
   // 1) boot -> boot
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 04e36cc..1a0df44 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -1135,7 +1135,7 @@ void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirec
   // temp = temp->dex_cache_resolved_methods_;
   __ movl(temp, Address(temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value()));
   // temp = temp[index_in_cache]
-  __ movl(temp, Address(temp, CodeGenerator::GetCacheOffset(invoke->GetIndexInDexCache())));
+  __ movl(temp, Address(temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex())));
   // (temp + offset_of_quick_compiled_code)()
   __ call(Address(
       temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 5fc24f7..3d7f122 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -18,6 +18,8 @@
 
 #include "entrypoints/quick/quick_entrypoints.h"
 #include "gc/accounting/card_table.h"
+#include "intrinsics.h"
+#include "intrinsics_x86_64.h"
 #include "mirror/array-inl.h"
 #include "mirror/art_method.h"
 #include "mirror/class.h"
@@ -61,20 +63,6 @@ class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatR
 
 #define __ reinterpret_cast<X86_64Assembler*>(codegen->GetAssembler())->
 
-class SlowPathCodeX86_64 : public SlowPathCode {
- public:
-  SlowPathCodeX86_64() : entry_label_(), exit_label_() {}
-
-  Label* GetEntryLabel() { return &entry_label_; }
-  Label* GetExitLabel() { return &exit_label_; }
-
- private:
-  Label entry_label_;
-  Label exit_label_;
-
-  DISALLOW_COPY_AND_ASSIGN(SlowPathCodeX86_64);
-};
-
 class NullCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
  public:
   explicit NullCheckSlowPathX86_64(HNullCheck* instruction) : instruction_(instruction) {}
@@ -375,6 +363,31 @@ inline Condition X86_64Condition(IfCondition cond) {
   return kEqual;
 }
 
+void CodeGeneratorX86_64::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke,
+                                                     CpuRegister temp) {
+  // All registers are assumed to be correctly set up.
+
+  // TODO: Implement all kinds of calls:
+  // 1) boot -> boot
+  // 2) app -> boot
+  // 3) app -> app
+  //
+  // Currently we implement the app -> app logic, which looks up in the resolve cache.
+
+  // temp = method;
+  LoadCurrentMethod(temp);
+  // temp = temp->dex_cache_resolved_methods_;
+  __ movl(temp, Address(temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().SizeValue()));
+  // temp = temp[index_in_cache]
+  __ movl(temp, Address(temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex())));
+  // (temp + offset_of_quick_compiled_code)()
+  __ call(Address(temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
+      kX86_64WordSize).SizeValue()));
+
+  DCHECK(!IsLeafMethod());
+  RecordPcInfo(invoke, invoke->GetDexPc());
+}
+
 void CodeGeneratorX86_64::DumpCoreRegister(std::ostream& stream, int reg) const {
   stream << X86_64ManagedRegister::FromCpuRegister(Register(reg));
 }
@@ -1123,30 +1136,31 @@ Location InvokeDexCallingConventionVisitor::GetNextLocation(Primitive::Type type
 }
 
 void LocationsBuilderX86_64::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+  IntrinsicLocationsBuilderX86_64 intrinsic(GetGraph()->GetArena());
+  if (intrinsic.TryDispatch(invoke)) {
+    return;
+  }
+
   HandleInvoke(invoke);
 }
 
-void InstructionCodeGeneratorX86_64::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
-  CpuRegister temp = invoke->GetLocations()->GetTemp(0).AsRegister<CpuRegister>();
-  // TODO: Implement all kinds of calls:
-  // 1) boot -> boot
-  // 2) app -> boot
-  // 3) app -> app
-  //
-  // Currently we implement the app -> app logic, which looks up in the resolve cache.
+static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorX86_64* codegen) {
+  if (invoke->GetLocations()->Intrinsified()) {
+    IntrinsicCodeGeneratorX86_64 intrinsic(codegen);
+    intrinsic.Dispatch(invoke);
+    return true;
+  }
+  return false;
+}
 
-  // temp = method;
-  codegen_->LoadCurrentMethod(temp);
-  // temp = temp->dex_cache_resolved_methods_;
-  __ movl(temp, Address(temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().SizeValue()));
-  // temp = temp[index_in_cache]
-  __ movl(temp, Address(temp, CodeGenerator::GetCacheOffset(invoke->GetIndexInDexCache())));
-  // (temp + offset_of_quick_compiled_code)()
-  __ call(Address(temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
-      kX86_64WordSize).SizeValue()));
+void InstructionCodeGeneratorX86_64::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+  if (TryGenerateIntrinsicCode(invoke, codegen_)) {
+    return;
+  }
 
-  DCHECK(!codegen_->IsLeafMethod());
-  codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
+  codegen_->GenerateStaticOrDirectCall(
+      invoke,
+      invoke->GetLocations()->GetTemp(0).AsRegister<CpuRegister>());
 }
 
 void LocationsBuilderX86_64::HandleInvoke(HInvoke* invoke) {
@@ -1182,10 +1196,19 @@ void LocationsBuilderX86_64::HandleInvoke(HInvoke* invoke) {
 }
 
 void LocationsBuilderX86_64::VisitInvokeVirtual(HInvokeVirtual* invoke) {
+  IntrinsicLocationsBuilderX86_64 intrinsic(GetGraph()->GetArena());
+  if (intrinsic.TryDispatch(invoke)) {
+    return;
+  }
+
   HandleInvoke(invoke);
 }
 
 void InstructionCodeGeneratorX86_64::VisitInvokeVirtual(HInvokeVirtual* invoke) {
+  if (TryGenerateIntrinsicCode(invoke, codegen_)) {
+    return;
+  }
+
   CpuRegister temp = invoke->GetLocations()->GetTemp(0).AsRegister<CpuRegister>();
   size_t method_offset = mirror::Class::EmbeddedVTableOffset().SizeValue() +
           invoke->GetVTableIndex() * sizeof(mirror::Class::VTableEntry);
diff --git a/compiler/optimizing/code_generator_x86_64.h b/compiler/optimizing/code_generator_x86_64.h
index 343fba3..c501568 100644
--- a/compiler/optimizing/code_generator_x86_64.h
+++ b/compiler/optimizing/code_generator_x86_64.h
@@ -36,6 +36,8 @@ static constexpr FloatRegister kParameterFloatRegisters[] =
 static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters);
 static constexpr size_t kParameterFloatRegistersLength = arraysize(kParameterFloatRegisters);
 
+static constexpr bool kCoalescedImplicitNullCheck = false;
+
 class InvokeDexCallingConvention : public CallingConvention<Register, FloatRegister> {
  public:
   InvokeDexCallingConvention() : CallingConvention(
@@ -67,7 +69,20 @@ class InvokeDexCallingConventionVisitor {
 };
 
 class CodeGeneratorX86_64;
-class SlowPathCodeX86_64;
+
+class SlowPathCodeX86_64 : public SlowPathCode {
+ public:
+  SlowPathCodeX86_64() : entry_label_(), exit_label_() {}
+
+  Label* GetEntryLabel() { return &entry_label_; }
+  Label* GetExitLabel() { return &exit_label_; }
+
+ private:
+  Label entry_label_;
+  Label exit_label_;
+
+  DISALLOW_COPY_AND_ASSIGN(SlowPathCodeX86_64);
+};
 
 class ParallelMoveResolverX86_64 : public ParallelMoveResolver {
  public:
@@ -226,6 +241,8 @@ class CodeGeneratorX86_64 : public CodeGenerator {
     return false;
   }
 
+  void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, CpuRegister temp);
+
  private:
   // Labels for each block that will be compiled.
   GrowableArray<Label> block_labels_;
diff --git a/compiler/optimizing/inliner.cc b/compiler/optimizing/inliner.cc
index 493d93f..532167c 100644
--- a/compiler/optimizing/inliner.cc
+++ b/compiler/optimizing/inliner.cc
@@ -44,10 +44,10 @@ void HInliner::Run() {
          instr_it.Advance()) {
       HInvokeStaticOrDirect* current = instr_it.Current()->AsInvokeStaticOrDirect();
       if (current != nullptr) {
-        if (!TryInline(current, current->GetIndexInDexCache(), current->GetInvokeType())) {
+        if (!TryInline(current, current->GetDexMethodIndex(), current->GetInvokeType())) {
           if (kIsDebugBuild) {
             std::string callee_name =
-                PrettyMethod(current->GetIndexInDexCache(), *outer_compilation_unit_.GetDexFile());
+                PrettyMethod(current->GetDexMethodIndex(), *outer_compilation_unit_.GetDexFile());
             bool should_inline = callee_name.find("$inline$") != std::string::npos;
             CHECK(!should_inline) << "Could not inline " << callee_name;
           }
diff --git a/compiler/optimizing/intrinsics.cc b/compiler/optimizing/intrinsics.cc
new file mode 100644
index 0000000..fe0e7f2
--- /dev/null
+++ b/compiler/optimizing/intrinsics.cc
@@ -0,0 +1,366 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "intrinsics.h"
+
+#include "dex/quick/dex_file_method_inliner.h"
+#include "dex/quick/dex_file_to_method_inliner_map.h"
+#include "driver/compiler_driver.h"
+#include "invoke_type.h"
+#include "nodes.h"
+#include "quick/inline_method_analyser.h"
+
+namespace art {
+
+// Function that returns whether an intrinsic is static/direct or virtual.
+static inline InvokeType GetIntrinsicInvokeType(Intrinsics i) {
+  switch (i) {
+    case Intrinsics::kNone:
+      return kInterface;  // Non-sensical for intrinsic.
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) \
+    case Intrinsics::k ## Name:               \
+      return IsStatic;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+  }
+  return kInterface;
+}
+
+
+
+static Primitive::Type GetType(uint64_t data, bool is_op_size) {
+  if (is_op_size) {
+    switch (static_cast<OpSize>(data)) {
+      case kSignedByte:
+        return Primitive::Type::kPrimByte;
+      case kSignedHalf:
+        return Primitive::Type::kPrimShort;
+      case k32:
+        return Primitive::Type::kPrimInt;
+      case k64:
+        return Primitive::Type::kPrimLong;
+      default:
+        LOG(FATAL) << "Unknown/unsupported op size " << data;
+        UNREACHABLE();
+    }
+  } else {
+    if ((data & kIntrinsicFlagIsLong) != 0) {
+      return Primitive::Type::kPrimLong;
+    }
+    if ((data & kIntrinsicFlagIsObject) != 0) {
+      return Primitive::Type::kPrimNot;
+    }
+    return Primitive::Type::kPrimInt;
+  }
+}
+
+static Intrinsics GetIntrinsic(InlineMethod method) {
+  switch (method.opcode) {
+    // Floating-point conversions.
+    case kIntrinsicDoubleCvt:
+      return ((method.d.data & kIntrinsicFlagToFloatingPoint) == 0) ?
+          Intrinsics::kDoubleDoubleToRawLongBits : Intrinsics::kDoubleLongBitsToDouble;
+    case kIntrinsicFloatCvt:
+      return ((method.d.data & kIntrinsicFlagToFloatingPoint) == 0) ?
+          Intrinsics::kFloatFloatToRawIntBits : Intrinsics::kFloatIntBitsToFloat;
+
+    // Bit manipulations.
+    case kIntrinsicReverseBits:
+      switch (GetType(method.d.data, true)) {
+        case Primitive::Type::kPrimInt:
+          return Intrinsics::kIntegerReverse;
+        case Primitive::Type::kPrimLong:
+          return Intrinsics::kLongReverse;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+    case kIntrinsicReverseBytes:
+      switch (GetType(method.d.data, true)) {
+        case Primitive::Type::kPrimShort:
+          return Intrinsics::kShortReverseBytes;
+        case Primitive::Type::kPrimInt:
+          return Intrinsics::kIntegerReverseBytes;
+        case Primitive::Type::kPrimLong:
+          return Intrinsics::kLongReverseBytes;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+
+    // Abs.
+    case kIntrinsicAbsDouble:
+      return Intrinsics::kMathAbsDouble;
+    case kIntrinsicAbsFloat:
+      return Intrinsics::kMathAbsFloat;
+    case kIntrinsicAbsInt:
+      return Intrinsics::kMathAbsInt;
+    case kIntrinsicAbsLong:
+      return Intrinsics::kMathAbsLong;
+
+    // Min/max.
+    case kIntrinsicMinMaxDouble:
+      return ((method.d.data & kIntrinsicFlagMin) == 0) ?
+          Intrinsics::kMathMaxDoubleDouble : Intrinsics::kMathMinDoubleDouble;
+    case kIntrinsicMinMaxFloat:
+      return ((method.d.data & kIntrinsicFlagMin) == 0) ?
+          Intrinsics::kMathMaxFloatFloat : Intrinsics::kMathMinFloatFloat;
+    case kIntrinsicMinMaxInt:
+      return ((method.d.data & kIntrinsicFlagMin) == 0) ?
+          Intrinsics::kMathMaxIntInt : Intrinsics::kMathMinIntInt;
+    case kIntrinsicMinMaxLong:
+      return ((method.d.data & kIntrinsicFlagMin) == 0) ?
+          Intrinsics::kMathMaxLongLong : Intrinsics::kMathMinLongLong;
+
+    // Misc math.
+    case kIntrinsicSqrt:
+      return Intrinsics::kMathSqrt;
+    case kIntrinsicCeil:
+      return Intrinsics::kMathCeil;
+    case kIntrinsicFloor:
+      return Intrinsics::kMathFloor;
+    case kIntrinsicRint:
+      return Intrinsics::kMathRint;
+    case kIntrinsicRoundDouble:
+      return Intrinsics::kMathRoundDouble;
+    case kIntrinsicRoundFloat:
+      return Intrinsics::kMathRoundFloat;
+
+    // System.arraycopy.
+    case kIntrinsicSystemArrayCopyCharArray:
+      return Intrinsics::kSystemArrayCopyChar;
+
+    // Thread.currentThread.
+    case kIntrinsicCurrentThread:
+      return  Intrinsics::kThreadCurrentThread;
+
+    // Memory.peek.
+    case kIntrinsicPeek:
+      switch (GetType(method.d.data, true)) {
+        case Primitive::Type::kPrimByte:
+          return Intrinsics::kMemoryPeekByte;
+        case Primitive::Type::kPrimShort:
+          return Intrinsics::kMemoryPeekShortNative;
+        case Primitive::Type::kPrimInt:
+          return Intrinsics::kMemoryPeekIntNative;
+        case Primitive::Type::kPrimLong:
+          return Intrinsics::kMemoryPeekLongNative;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+
+    // Memory.poke.
+    case kIntrinsicPoke:
+      switch (GetType(method.d.data, true)) {
+        case Primitive::Type::kPrimByte:
+          return Intrinsics::kMemoryPokeByte;
+        case Primitive::Type::kPrimShort:
+          return Intrinsics::kMemoryPokeShortNative;
+        case Primitive::Type::kPrimInt:
+          return Intrinsics::kMemoryPokeIntNative;
+        case Primitive::Type::kPrimLong:
+          return Intrinsics::kMemoryPokeLongNative;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+
+    // String.
+    case kIntrinsicCharAt:
+      return Intrinsics::kStringCharAt;
+    case kIntrinsicCompareTo:
+      return Intrinsics::kStringCompareTo;
+    case kIntrinsicIsEmptyOrLength:
+      return ((method.d.data & kIntrinsicFlagIsEmpty) == 0) ?
+          Intrinsics::kStringLength : Intrinsics::kStringIsEmpty;
+    case kIntrinsicIndexOf:
+      return ((method.d.data & kIntrinsicFlagBase0) == 0) ?
+          Intrinsics::kStringIndexOfAfter : Intrinsics::kStringIndexOf;
+
+    case kIntrinsicCas:
+      switch (GetType(method.d.data, false)) {
+        case Primitive::Type::kPrimNot:
+          return Intrinsics::kUnsafeCASObject;
+        case Primitive::Type::kPrimInt:
+          return Intrinsics::kUnsafeCASInt;
+        case Primitive::Type::kPrimLong:
+          return Intrinsics::kUnsafeCASLong;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+    case kIntrinsicUnsafeGet: {
+      const bool is_volatile = (method.d.data & kIntrinsicFlagIsVolatile);
+      switch (GetType(method.d.data, false)) {
+        case Primitive::Type::kPrimInt:
+          return is_volatile ? Intrinsics::kUnsafeGetVolatile : Intrinsics::kUnsafeGet;
+        case Primitive::Type::kPrimLong:
+          return is_volatile ? Intrinsics::kUnsafeGetLongVolatile : Intrinsics::kUnsafeGetLong;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+    }
+    case kIntrinsicUnsafePut: {
+      enum Sync { kNoSync, kVolatile, kOrdered };
+      const Sync sync =
+          ((method.d.data & kIntrinsicFlagIsVolatile) != 0) ? kVolatile :
+          ((method.d.data & kIntrinsicFlagIsOrdered) != 0)  ? kOrdered :
+                                                              kNoSync;
+      switch (GetType(method.d.data, false)) {
+        case Primitive::Type::kPrimInt:
+          switch (sync) {
+            case kNoSync:
+              return Intrinsics::kUnsafePut;
+            case kVolatile:
+              return Intrinsics::kUnsafePutVolatile;
+            case kOrdered:
+              return Intrinsics::kUnsafePutOrdered;
+          }
+          break;
+        case Primitive::Type::kPrimLong:
+          switch (sync) {
+            case kNoSync:
+              return Intrinsics::kUnsafePutLong;
+            case kVolatile:
+              return Intrinsics::kUnsafePutLongVolatile;
+            case kOrdered:
+              return Intrinsics::kUnsafePutLongOrdered;
+          }
+          break;
+        case Primitive::Type::kPrimNot:
+          switch (sync) {
+            case kNoSync:
+              return Intrinsics::kUnsafePutObject;
+            case kVolatile:
+              return Intrinsics::kUnsafePutObjectVolatile;
+            case kOrdered:
+              return Intrinsics::kUnsafePutObjectOrdered;
+          }
+          break;
+        default:
+          LOG(FATAL) << "Unknown/unsupported op size " << method.d.data;
+          UNREACHABLE();
+      }
+      break;
+    }
+
+    // Virtual cases.
+
+    case kIntrinsicReferenceGetReferent:
+      return Intrinsics::kReferenceGetReferent;
+
+    // Quick inliner cases. Remove after refactoring. They are here so that we can use the
+    // compiler to warn on missing cases.
+
+    case kInlineOpNop:
+    case kInlineOpReturnArg:
+    case kInlineOpNonWideConst:
+    case kInlineOpIGet:
+    case kInlineOpIPut:
+      return Intrinsics::kNone;
+
+    // No default case to make the compiler warn on missing cases.
+  }
+  return Intrinsics::kNone;
+}
+
+static bool CheckInvokeType(Intrinsics intrinsic, HInvoke* invoke) {
+  // The DexFileMethodInliner should have checked whether the methods are agreeing with
+  // what we expect, i.e., static methods are called as such. Add another check here for
+  // our expectations:
+  // Whenever the intrinsic is marked as static-or-direct, report an error if we find an
+  // InvokeVirtual. The other direction is not possible: we have intrinsics for virtual
+  // functions that will perform a check inline. If the precise type is known, however,
+  // the instruction will be sharpened to an InvokeStaticOrDirect.
+  InvokeType intrinsic_type = GetIntrinsicInvokeType(intrinsic);
+  InvokeType invoke_type = invoke->IsInvokeStaticOrDirect() ?
+      invoke->AsInvokeStaticOrDirect()->GetInvokeType() :
+      invoke->IsInvokeVirtual() ? kVirtual : kSuper;
+  switch (intrinsic_type) {
+    case kStatic:
+      return (invoke_type == kStatic);
+    case kDirect:
+      return (invoke_type == kDirect);
+    case kVirtual:
+      // Call might be devirtualized.
+      return (invoke_type == kVirtual || invoke_type == kDirect);
+
+    default:
+      return false;
+  }
+}
+
+// TODO: Refactor DexFileMethodInliner and have something nicer than InlineMethod.
+void IntrinsicsRecognizer::Run() {
+  DexFileMethodInliner* inliner = driver_->GetMethodInlinerMap()->GetMethodInliner(dex_file_);
+  DCHECK(inliner != nullptr);
+
+  for (HReversePostOrderIterator it(*graph_); !it.Done(); it.Advance()) {
+    HBasicBlock* block = it.Current();
+    for (HInstructionIterator inst_it(block->GetInstructions()); !inst_it.Done();
+         inst_it.Advance()) {
+      HInstruction* inst = inst_it.Current();
+      if (inst->IsInvoke()) {
+        HInvoke* invoke = inst->AsInvoke();
+        InlineMethod method;
+        if (inliner->IsIntrinsic(invoke->GetDexMethodIndex(), &method)) {
+          Intrinsics intrinsic = GetIntrinsic(method);
+
+          if (intrinsic != Intrinsics::kNone) {
+            if (!CheckInvokeType(intrinsic, invoke)) {
+              LOG(WARNING) << "Found an intrinsic with unexpected invoke type: "
+                           << intrinsic << " for "
+                           << PrettyMethod(invoke->GetDexMethodIndex(), *dex_file_);
+            } else {
+              invoke->SetIntrinsic(intrinsic);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+std::ostream& operator<<(std::ostream& os, const Intrinsics& intrinsic) {
+  switch (intrinsic) {
+    case Intrinsics::kNone:
+      os << "No intrinsic.";
+      break;
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) \
+    case Intrinsics::k ## Name: \
+      os << # Name; \
+      break;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef STATIC_INTRINSICS_LIST
+#undef VIRTUAL_INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+  }
+  return os;
+}
+
+}  // namespace art
+
diff --git a/compiler/optimizing/intrinsics.h b/compiler/optimizing/intrinsics.h
new file mode 100644
index 0000000..29cc8ef
--- /dev/null
+++ b/compiler/optimizing/intrinsics.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_INTRINSICS_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_H_
+
+#include "nodes.h"
+#include "optimization.h"
+
+namespace art {
+
+class CompilerDriver;
+class DexFile;
+
+// Recognize intrinsics from HInvoke nodes.
+class IntrinsicsRecognizer : public HOptimization {
+ public:
+  IntrinsicsRecognizer(HGraph* graph, const DexFile* dex_file, CompilerDriver* driver)
+      : HOptimization(graph, true, "intrinsics_recognition"),
+        dex_file_(dex_file), driver_(driver) {}
+
+  void Run() OVERRIDE;
+
+ private:
+  const DexFile* dex_file_;
+  CompilerDriver* driver_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicsRecognizer);
+};
+
+class IntrinsicVisitor : public ValueObject {
+ public:
+  virtual ~IntrinsicVisitor() {}
+
+  // Dispatch logic.
+
+  void Dispatch(HInvoke* invoke) {
+    switch (invoke->GetIntrinsic()) {
+      case Intrinsics::kNone:
+        return;
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) \
+      case Intrinsics::k ## Name:             \
+        Visit ## Name(invoke);                \
+        return;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+      // Do not put a default case. That way the compiler will complain if we missed a case.
+    }
+  }
+
+  // Define visitor methods.
+
+#define OPTIMIZING_INTRINSICS(Name, IsStatic)                    \
+  virtual void Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) { \
+  }
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+ protected:
+  IntrinsicVisitor() {}
+
+ private:
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicVisitor);
+};
+
+}  // namespace art
+
+#endif  // ART_COMPILER_OPTIMIZING_INTRINSICS_H_
diff --git a/compiler/optimizing/intrinsics_list.h b/compiler/optimizing/intrinsics_list.h
new file mode 100644
index 0000000..29ca20c
--- /dev/null
+++ b/compiler/optimizing/intrinsics_list.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_INTRINSICS_LIST_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_LIST_H_
+
+// All intrinsics supported by the optimizing compiler. Format is name, then whether it is expected
+// to be a HInvokeStaticOrDirect node (compared to HInvokeVirtual).
+
+#define INTRINSICS_LIST(V) \
+  V(DoubleDoubleToRawLongBits, kStatic) \
+  V(DoubleLongBitsToDouble, kStatic) \
+  V(FloatFloatToRawIntBits, kStatic) \
+  V(FloatIntBitsToFloat, kStatic) \
+  V(IntegerReverse, kStatic) \
+  V(IntegerReverseBytes, kStatic) \
+  V(LongReverse, kStatic) \
+  V(LongReverseBytes, kStatic) \
+  V(ShortReverseBytes, kStatic) \
+  V(MathAbsDouble, kStatic) \
+  V(MathAbsFloat, kStatic) \
+  V(MathAbsLong, kStatic) \
+  V(MathAbsInt, kStatic) \
+  V(MathMinDoubleDouble, kStatic) \
+  V(MathMinFloatFloat, kStatic) \
+  V(MathMinLongLong, kStatic) \
+  V(MathMinIntInt, kStatic) \
+  V(MathMaxDoubleDouble, kStatic) \
+  V(MathMaxFloatFloat, kStatic) \
+  V(MathMaxLongLong, kStatic) \
+  V(MathMaxIntInt, kStatic) \
+  V(MathSqrt, kStatic) \
+  V(MathCeil, kStatic) \
+  V(MathFloor, kStatic) \
+  V(MathRint, kStatic) \
+  V(MathRoundDouble, kStatic) \
+  V(MathRoundFloat, kStatic) \
+  V(SystemArrayCopyChar, kStatic) \
+  V(ThreadCurrentThread, kStatic) \
+  V(MemoryPeekByte, kStatic) \
+  V(MemoryPeekIntNative, kStatic) \
+  V(MemoryPeekLongNative, kStatic) \
+  V(MemoryPeekShortNative, kStatic) \
+  V(MemoryPokeByte, kStatic) \
+  V(MemoryPokeIntNative, kStatic) \
+  V(MemoryPokeLongNative, kStatic) \
+  V(MemoryPokeShortNative, kStatic) \
+  V(StringCharAt, kDirect) \
+  V(StringCompareTo, kDirect) \
+  V(StringIsEmpty, kDirect) \
+  V(StringIndexOf, kDirect) \
+  V(StringIndexOfAfter, kDirect) \
+  V(StringLength, kDirect) \
+  V(UnsafeCASInt, kDirect) \
+  V(UnsafeCASLong, kDirect) \
+  V(UnsafeCASObject, kDirect) \
+  V(UnsafeGet, kDirect) \
+  V(UnsafeGetVolatile, kDirect) \
+  V(UnsafeGetLong, kDirect) \
+  V(UnsafeGetLongVolatile, kDirect) \
+  V(UnsafePut, kDirect) \
+  V(UnsafePutOrdered, kDirect) \
+  V(UnsafePutVolatile, kDirect) \
+  V(UnsafePutObject, kDirect) \
+  V(UnsafePutObjectOrdered, kDirect) \
+  V(UnsafePutObjectVolatile, kDirect) \
+  V(UnsafePutLong, kDirect) \
+  V(UnsafePutLongOrdered, kDirect) \
+  V(UnsafePutLongVolatile, kDirect) \
+  \
+  V(ReferenceGetReferent, kVirtual)
+
+#endif  // ART_COMPILER_OPTIMIZING_INTRINSICS_LIST_H_
+#undef ART_COMPILER_OPTIMIZING_INTRINSICS_LIST_H_   // #define is only for lint.
diff --git a/compiler/optimizing/intrinsics_x86_64.cc b/compiler/optimizing/intrinsics_x86_64.cc
new file mode 100644
index 0000000..c1f4c94
--- /dev/null
+++ b/compiler/optimizing/intrinsics_x86_64.cc
@@ -0,0 +1,984 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "intrinsics_x86_64.h"
+
+#include "code_generator_x86_64.h"
+#include "entrypoints/quick/quick_entrypoints.h"
+#include "intrinsics.h"
+#include "mirror/array-inl.h"
+#include "mirror/art_method.h"
+#include "mirror/string.h"
+#include "thread.h"
+#include "utils/x86_64/assembler_x86_64.h"
+#include "utils/x86_64/constants_x86_64.h"
+
+namespace art {
+
+namespace x86_64 {
+
+static constexpr bool kIntrinsified = true;
+
+X86_64Assembler* IntrinsicCodeGeneratorX86_64::GetAssembler() {
+  return reinterpret_cast<X86_64Assembler*>(codegen_->GetAssembler());
+}
+
+ArenaAllocator* IntrinsicCodeGeneratorX86_64::GetArena() {
+  return codegen_->GetGraph()->GetArena();
+}
+
+bool IntrinsicLocationsBuilderX86_64::TryDispatch(HInvoke* invoke) {
+  Dispatch(invoke);
+  const LocationSummary* res = invoke->GetLocations();
+  return res != nullptr && res->Intrinsified();
+}
+
+#define __ reinterpret_cast<X86_64Assembler*>(codegen->GetAssembler())->
+
+// TODO: trg as memory.
+static void MoveFromReturnRegister(Location trg,
+                                   Primitive::Type type,
+                                   CodeGeneratorX86_64* codegen) {
+  if (!trg.IsValid()) {
+    DCHECK(type == Primitive::kPrimVoid);
+    return;
+  }
+
+  switch (type) {
+    case Primitive::kPrimBoolean:
+    case Primitive::kPrimByte:
+    case Primitive::kPrimChar:
+    case Primitive::kPrimShort:
+    case Primitive::kPrimInt:
+    case Primitive::kPrimNot: {
+      CpuRegister trg_reg = trg.AsRegister<CpuRegister>();
+      if (trg_reg.AsRegister() != RAX) {
+        __ movl(trg_reg, CpuRegister(RAX));
+      }
+      break;
+    }
+    case Primitive::kPrimLong: {
+      CpuRegister trg_reg = trg.AsRegister<CpuRegister>();
+      if (trg_reg.AsRegister() != RAX) {
+        __ movq(trg_reg, CpuRegister(RAX));
+      }
+      break;
+    }
+
+    case Primitive::kPrimVoid:
+      LOG(FATAL) << "Unexpected void type for valid location " << trg;
+      UNREACHABLE();
+
+    case Primitive::kPrimDouble: {
+      XmmRegister trg_reg = trg.AsFpuRegister<XmmRegister>();
+      if (trg_reg.AsFloatRegister() != XMM0) {
+        __ movsd(trg_reg, XmmRegister(XMM0));
+      }
+      break;
+    }
+    case Primitive::kPrimFloat: {
+      XmmRegister trg_reg = trg.AsFpuRegister<XmmRegister>();
+      if (trg_reg.AsFloatRegister() != XMM0) {
+        __ movss(trg_reg, XmmRegister(XMM0));
+      }
+      break;
+    }
+  }
+}
+
+static void MoveArguments(HInvoke* invoke, ArenaAllocator* arena, CodeGeneratorX86_64* codegen) {
+  if (invoke->InputCount() == 0) {
+    return;
+  }
+
+  LocationSummary* locations = invoke->GetLocations();
+  InvokeDexCallingConventionVisitor calling_convention_visitor;
+
+  // We're moving potentially two or more locations to locations that could overlap, so we need
+  // a parallel move resolver.
+  HParallelMove parallel_move(arena);
+
+  for (size_t i = 0; i < invoke->InputCount(); i++) {
+    HInstruction* input = invoke->InputAt(i);
+    Location cc_loc = calling_convention_visitor.GetNextLocation(input->GetType());
+    Location actual_loc = locations->InAt(i);
+
+    parallel_move.AddMove(new (arena) MoveOperands(actual_loc, cc_loc, nullptr));
+  }
+
+  codegen->GetMoveResolver()->EmitNativeCode(&parallel_move);
+}
+
+// Slow-path for fallback (calling the managed code to handle the intrinsic) in an intrinsified
+// call. This will copy the arguments into the positions for a regular call.
+//
+// Note: The actual parameters are required to be in the locations given by the invoke's location
+//       summary. If an intrinsic modifies those locations before a slowpath call, they must be
+//       restored!
+class IntrinsicSlowPathX86_64 : public SlowPathCodeX86_64 {
+ public:
+  explicit IntrinsicSlowPathX86_64(HInvoke* invoke) : invoke_(invoke) { }
+
+  void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
+    CodeGeneratorX86_64* codegen = down_cast<CodeGeneratorX86_64*>(codegen_in);
+    __ Bind(GetEntryLabel());
+
+    codegen->SaveLiveRegisters(invoke_->GetLocations());
+
+    MoveArguments(invoke_, codegen->GetGraph()->GetArena(), codegen);
+
+    if (invoke_->IsInvokeStaticOrDirect()) {
+      codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(), CpuRegister(RDI));
+    } else {
+      UNIMPLEMENTED(FATAL) << "Non-direct intrinsic slow-path not yet implemented";
+      UNREACHABLE();
+    }
+
+    // Copy the result back to the expected output.
+    Location out = invoke_->GetLocations()->Out();
+    if (out.IsValid()) {
+      DCHECK(out.IsRegister());  // TODO: Replace this when we support output in memory.
+      DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
+      MoveFromReturnRegister(out, invoke_->GetType(), codegen);
+    }
+
+    codegen->RestoreLiveRegisters(invoke_->GetLocations());
+    __ jmp(GetExitLabel());
+  }
+
+ private:
+  // The instruction where this slow path is happening.
+  HInvoke* const invoke_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathX86_64);
+};
+
+#undef __
+#define __ assembler->
+
+static void CreateFPToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresFpuRegister());
+  locations->SetOut(Location::RequiresRegister());
+}
+
+static void CreateIntToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::RequiresFpuRegister());
+}
+
+static void MoveFPToInt(LocationSummary* locations, bool is64bit, X86_64Assembler* assembler) {
+  Location input = locations->InAt(0);
+  Location output = locations->Out();
+  __ movd(output.AsRegister<CpuRegister>(), input.AsFpuRegister<XmmRegister>(), is64bit);
+}
+
+static void MoveIntToFP(LocationSummary* locations, bool is64bit, X86_64Assembler* assembler) {
+  Location input = locations->InAt(0);
+  Location output = locations->Out();
+  __ movd(output.AsFpuRegister<XmmRegister>(), input.AsRegister<CpuRegister>(), is64bit);
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+  CreateFPToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+  CreateIntToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+  MoveFPToInt(invoke->GetLocations(), true, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86_64::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+  MoveIntToFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+  CreateFPToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitFloatIntBitsToFloat(HInvoke* invoke) {
+  CreateIntToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+  MoveFPToInt(invoke->GetLocations(), false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86_64::VisitFloatIntBitsToFloat(HInvoke* invoke) {
+  MoveIntToFP(invoke->GetLocations(), false, GetAssembler());
+}
+
+static void CreateIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::SameAsFirstInput());
+}
+
+static void GenReverseBytes(LocationSummary* locations,
+                            Primitive::Type size,
+                            X86_64Assembler* assembler) {
+  CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+
+  switch (size) {
+    case Primitive::kPrimShort:
+      // TODO: Can be done with an xchg of 8b registers. This is straight from Quick.
+      __ bswapl(out);
+      __ sarl(out, Immediate(16));
+      break;
+    case Primitive::kPrimInt:
+      __ bswapl(out);
+      break;
+    case Primitive::kPrimLong:
+      __ bswapq(out);
+      break;
+    default:
+      LOG(FATAL) << "Unexpected size for reverse-bytes: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitIntegerReverseBytes(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitIntegerReverseBytes(HInvoke* invoke) {
+  GenReverseBytes(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitLongReverseBytes(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitLongReverseBytes(HInvoke* invoke) {
+  GenReverseBytes(invoke->GetLocations(), Primitive::kPrimLong, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitShortReverseBytes(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitShortReverseBytes(HInvoke* invoke) {
+  GenReverseBytes(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
+}
+
+
+// TODO: Consider Quick's way of doing Double abs through integer operations, as the immediate we
+//       need is 64b.
+
+static void CreateFloatToFloatPlusTemps(ArenaAllocator* arena, HInvoke* invoke) {
+  // TODO: Enable memory operations when the assembler supports them.
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresFpuRegister());
+  // TODO: Allow x86 to work with memory. This requires assembler support, see below.
+  // locations->SetInAt(0, Location::Any());               // X86 can work on memory directly.
+  locations->SetOut(Location::SameAsFirstInput());
+  locations->AddTemp(Location::RequiresRegister());     // Immediate constant.
+  locations->AddTemp(Location::RequiresFpuRegister());  // FP version of above.
+}
+
+static void MathAbsFP(LocationSummary* locations, bool is64bit, X86_64Assembler* assembler) {
+  Location output = locations->Out();
+  CpuRegister cpu_temp = locations->GetTemp(0).AsRegister<CpuRegister>();
+
+  if (output.IsFpuRegister()) {
+    // In-register
+    XmmRegister xmm_temp = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
+
+    if (is64bit) {
+      __ movq(cpu_temp, Immediate(INT64_C(0x7FFFFFFFFFFFFFFF)));
+      __ movd(xmm_temp, cpu_temp);
+      __ andpd(output.AsFpuRegister<XmmRegister>(), xmm_temp);
+    } else {
+      __ movl(cpu_temp, Immediate(INT64_C(0x7FFFFFFF)));
+      __ movd(xmm_temp, cpu_temp);
+      __ andps(output.AsFpuRegister<XmmRegister>(), xmm_temp);
+    }
+  } else {
+    // TODO: update when assember support is available.
+    UNIMPLEMENTED(FATAL) << "Needs assembler support.";
+//  Once assembler support is available, in-memory operations look like this:
+//    if (is64bit) {
+//      DCHECK(output.IsDoubleStackSlot());
+//      // No 64b and with literal.
+//      __ movq(cpu_temp, Immediate(INT64_C(0x7FFFFFFFFFFFFFFF)));
+//      __ andq(Address(CpuRegister(RSP), output.GetStackIndex()), cpu_temp);
+//    } else {
+//      DCHECK(output.IsStackSlot());
+//      // Can use and with a literal directly.
+//      __ andl(Address(CpuRegister(RSP), output.GetStackIndex()), Immediate(INT64_C(0x7FFFFFFF)));
+//    }
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathAbsDouble(HInvoke* invoke) {
+  CreateFloatToFloatPlusTemps(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathAbsDouble(HInvoke* invoke) {
+  MathAbsFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathAbsFloat(HInvoke* invoke) {
+  CreateFloatToFloatPlusTemps(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathAbsFloat(HInvoke* invoke) {
+  MathAbsFP(invoke->GetLocations(), false, GetAssembler());
+}
+
+static void CreateIntToIntPlusTemp(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::SameAsFirstInput());
+  locations->AddTemp(Location::RequiresRegister());
+}
+
+static void GenAbsInteger(LocationSummary* locations, bool is64bit, X86_64Assembler* assembler) {
+  Location output = locations->Out();
+  CpuRegister out = output.AsRegister<CpuRegister>();
+  CpuRegister mask = locations->GetTemp(0).AsRegister<CpuRegister>();
+
+  if (is64bit) {
+    // Create mask.
+    __ movq(mask, out);
+    __ sarq(mask, Immediate(63));
+    // Add mask.
+    __ addq(out, mask);
+    __ xorq(out, mask);
+  } else {
+    // Create mask.
+    __ movl(mask, out);
+    __ sarl(mask, Immediate(31));
+    // Add mask.
+    __ addl(out, mask);
+    __ xorl(out, mask);
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathAbsInt(HInvoke* invoke) {
+  CreateIntToIntPlusTemp(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathAbsInt(HInvoke* invoke) {
+  GenAbsInteger(invoke->GetLocations(), false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathAbsLong(HInvoke* invoke) {
+  CreateIntToIntPlusTemp(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathAbsLong(HInvoke* invoke) {
+  GenAbsInteger(invoke->GetLocations(), true, GetAssembler());
+}
+
+static void GenMinMaxFP(LocationSummary* locations, bool is_min, bool is_double,
+                        X86_64Assembler* assembler) {
+  Location op1_loc = locations->InAt(0);
+  Location op2_loc = locations->InAt(1);
+  Location out_loc = locations->Out();
+  XmmRegister out = out_loc.AsFpuRegister<XmmRegister>();
+
+  // Shortcut for same input locations.
+  if (op1_loc.Equals(op2_loc)) {
+    DCHECK(out_loc.Equals(op1_loc));
+    return;
+  }
+
+  //  (out := op1)
+  //  out <=? op2
+  //  if Nan jmp Nan_label
+  //  if out is min jmp done
+  //  if op2 is min jmp op2_label
+  //  handle -0/+0
+  //  jmp done
+  // Nan_label:
+  //  out := NaN
+  // op2_label:
+  //  out := op2
+  // done:
+  //
+  // This removes one jmp, but needs to copy one input (op1) to out.
+  //
+  // TODO: This is straight from Quick (except literal pool). Make NaN an out-of-line slowpath?
+
+  XmmRegister op2 = op2_loc.AsFpuRegister<XmmRegister>();
+
+  Label nan, done, op2_label;
+  if (is_double) {
+    __ ucomisd(out, op2);
+  } else {
+    __ ucomiss(out, op2);
+  }
+
+  __ j(Condition::kParityEven, &nan);
+
+  __ j(is_min ? Condition::kAbove : Condition::kBelow, &op2_label);
+  __ j(is_min ? Condition::kBelow : Condition::kAbove, &done);
+
+  // Handle 0.0/-0.0.
+  if (is_min) {
+    if (is_double) {
+      __ orpd(out, op2);
+    } else {
+      __ orps(out, op2);
+    }
+  } else {
+    if (is_double) {
+      __ andpd(out, op2);
+    } else {
+      __ andps(out, op2);
+    }
+  }
+  __ jmp(&done);
+
+  // NaN handling.
+  __ Bind(&nan);
+  CpuRegister cpu_temp = locations->GetTemp(0).AsRegister<CpuRegister>();
+  // TODO: Literal pool. Trades 64b immediate in CPU reg for direct memory access.
+  if (is_double) {
+    __ movq(cpu_temp, Immediate(INT64_C(0x7FF8000000000000)));
+  } else {
+    __ movl(cpu_temp, Immediate(INT64_C(0x7FC00000)));
+  }
+  __ movd(out, cpu_temp, is_double);
+  __ jmp(&done);
+
+  // out := op2;
+  __ Bind(&op2_label);
+  if (is_double) {
+    __ movsd(out, op2);
+  } else {
+    __ movss(out, op2);
+  }
+
+  // Done.
+  __ Bind(&done);
+}
+
+static void CreateFPFPToFPPlusTempLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresFpuRegister());
+  locations->SetInAt(1, Location::RequiresFpuRegister());
+  // The following is sub-optimal, but all we can do for now. It would be fine to also accept
+  // the second input to be the output (we can simply swap inputs).
+  locations->SetOut(Location::SameAsFirstInput());
+  locations->AddTemp(Location::RequiresRegister());     // Immediate constant.
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMinDoubleDouble(HInvoke* invoke) {
+  CreateFPFPToFPPlusTempLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMinDoubleDouble(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), true, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMinFloatFloat(HInvoke* invoke) {
+  CreateFPFPToFPPlusTempLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMinFloatFloat(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), true, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMaxDoubleDouble(HInvoke* invoke) {
+  CreateFPFPToFPPlusTempLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMaxDoubleDouble(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), false, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMaxFloatFloat(HInvoke* invoke) {
+  CreateFPFPToFPPlusTempLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMaxFloatFloat(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), false, false, GetAssembler());
+}
+
+static void GenMinMax(LocationSummary* locations, bool is_min, bool is_long,
+                      X86_64Assembler* assembler) {
+  Location op1_loc = locations->InAt(0);
+  Location op2_loc = locations->InAt(1);
+
+  // Shortcut for same input locations.
+  if (op1_loc.Equals(op2_loc)) {
+    // Can return immediately, as op1_loc == out_loc.
+    // Note: if we ever support separate registers, e.g., output into memory, we need to check for
+    //       a copy here.
+    DCHECK(locations->Out().Equals(op1_loc));
+    return;
+  }
+
+  CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+  CpuRegister op2 = op2_loc.AsRegister<CpuRegister>();
+
+  //  (out := op1)
+  //  out <=? op2
+  //  if out is min jmp done
+  //  out := op2
+  // done:
+
+  if (is_long) {
+    __ cmpq(out, op2);
+  } else {
+    __ cmpl(out, op2);
+  }
+
+  __ cmov(is_min ? Condition::kGreater : Condition::kLess, out, op2, is_long);
+}
+
+static void CreateIntIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetInAt(1, Location::RequiresRegister());
+  locations->SetOut(Location::SameAsFirstInput());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMinIntInt(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMinIntInt(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), true, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMinLongLong(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMinLongLong(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), true, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMaxIntInt(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMaxIntInt(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), false, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathMaxLongLong(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathMaxLongLong(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), false, true, GetAssembler());
+}
+
+static void CreateFPToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresFpuRegister());
+  locations->SetOut(Location::RequiresFpuRegister());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMathSqrt(HInvoke* invoke) {
+  CreateFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMathSqrt(HInvoke* invoke) {
+  LocationSummary* locations = invoke->GetLocations();
+  XmmRegister in = locations->InAt(0).AsFpuRegister<XmmRegister>();
+  XmmRegister out = locations->Out().AsFpuRegister<XmmRegister>();
+
+  GetAssembler()->sqrtsd(out, in);
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitStringCharAt(HInvoke* invoke) {
+  // The inputs plus one temp.
+  LocationSummary* locations = new (arena_) LocationSummary(invoke,
+                                                            LocationSummary::kCallOnSlowPath,
+                                                            kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetInAt(1, Location::RequiresRegister());
+  locations->SetOut(Location::SameAsFirstInput());
+  locations->AddTemp(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitStringCharAt(HInvoke* invoke) {
+  LocationSummary* locations = invoke->GetLocations();
+
+  // Location of reference to data array
+  const int32_t value_offset = mirror::String::ValueOffset().Int32Value();
+  // Location of count
+  const int32_t count_offset = mirror::String::CountOffset().Int32Value();
+  // Starting offset within data array
+  const int32_t offset_offset = mirror::String::OffsetOffset().Int32Value();
+  // Start of char data with array_
+  const int32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value();
+
+  CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
+  CpuRegister idx = locations->InAt(1).AsRegister<CpuRegister>();
+  CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+  Location temp_loc = locations->GetTemp(0);
+  CpuRegister temp = temp_loc.AsRegister<CpuRegister>();
+
+  // Note: Nullcheck has been done before in a HNullCheck before the HInvokeVirtual. If/when we
+  //       move to (coalesced) implicit checks, we have to do a null check below.
+  DCHECK(!kCoalescedImplicitNullCheck);
+
+  // TODO: Maybe we can support range check elimination. Overall, though, I think it's not worth
+  //       the cost.
+  // TODO: For simplicity, the index parameter is requested in a register, so different from Quick
+  //       we will not optimize the code for constants (which would save a register).
+
+  SlowPathCodeX86_64* slow_path = new (GetArena()) IntrinsicSlowPathX86_64(invoke);
+  codegen_->AddSlowPath(slow_path);
+
+  X86_64Assembler* assembler = GetAssembler();
+
+  __ cmpl(idx, Address(obj, count_offset));
+  __ j(kAboveEqual, slow_path->GetEntryLabel());
+
+  // Get the actual element.
+  __ movl(temp, idx);                          // temp := idx.
+  __ addl(temp, Address(obj, offset_offset));  // temp := offset + idx.
+  __ movl(out, Address(obj, value_offset));    // obj := obj.array.
+  // out = out[2*temp].
+  __ movzxw(out, Address(out, temp, ScaleFactor::TIMES_2, data_offset));
+
+  __ Bind(slow_path->GetExitLabel());
+}
+
+static void GenPeek(LocationSummary* locations, Primitive::Type size, X86_64Assembler* assembler) {
+  CpuRegister address = locations->InAt(0).AsRegister<CpuRegister>();
+  CpuRegister out = locations->Out().AsRegister<CpuRegister>();  // == address, here for clarity.
+  // x86 allows unaligned access. We do not have to check the input or use specific instructions
+  // to avoid a SIGBUS.
+  switch (size) {
+    case Primitive::kPrimByte:
+      __ movsxb(out, Address(address, 0));
+      break;
+    case Primitive::kPrimShort:
+      __ movsxw(out, Address(address, 0));
+      break;
+    case Primitive::kPrimInt:
+      __ movl(out, Address(address, 0));
+      break;
+    case Primitive::kPrimLong:
+      __ movq(out, Address(address, 0));
+      break;
+    default:
+      LOG(FATAL) << "Type not recognized for peek: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPeekByte(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPeekByte(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimByte, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPeekIntNative(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPeekIntNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPeekLongNative(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPeekLongNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimLong, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPeekShortNative(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPeekShortNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
+}
+
+static void CreateIntIntToVoidLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetInAt(1, Location::RequiresRegister());
+}
+
+static void GenPoke(LocationSummary* locations, Primitive::Type size, X86_64Assembler* assembler) {
+  CpuRegister address = locations->InAt(0).AsRegister<CpuRegister>();
+  CpuRegister value = locations->InAt(1).AsRegister<CpuRegister>();
+  // x86 allows unaligned access. We do not have to check the input or use specific instructions
+  // to avoid a SIGBUS.
+  switch (size) {
+    case Primitive::kPrimByte:
+      __ movb(Address(address, 0), value);
+      break;
+    case Primitive::kPrimShort:
+      __ movw(Address(address, 0), value);
+      break;
+    case Primitive::kPrimInt:
+      __ movl(Address(address, 0), value);
+      break;
+    case Primitive::kPrimLong:
+      __ movq(Address(address, 0), value);
+      break;
+    default:
+      LOG(FATAL) << "Type not recognized for poke: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPokeByte(HInvoke* invoke) {
+  CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPokeByte(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimByte, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPokeIntNative(HInvoke* invoke) {
+  CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPokeIntNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPokeLongNative(HInvoke* invoke) {
+  CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPokeLongNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimLong, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitMemoryPokeShortNative(HInvoke* invoke) {
+  CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMemoryPokeShortNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitThreadCurrentThread(HInvoke* invoke) {
+  LocationSummary* locations = new (arena_) LocationSummary(invoke,
+                                                            LocationSummary::kNoCall,
+                                                            kIntrinsified);
+  locations->SetOut(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitThreadCurrentThread(HInvoke* invoke) {
+  CpuRegister out = invoke->GetLocations()->Out().AsRegister<CpuRegister>();
+  GetAssembler()->gs()->movl(out, Address::Absolute(Thread::PeerOffset<kX86_64WordSize>(), true));
+}
+
+static void GenUnsafeGet(LocationSummary* locations, bool is_long,
+                         bool is_volatile ATTRIBUTE_UNUSED, X86_64Assembler* assembler) {
+  CpuRegister base = locations->InAt(1).AsRegister<CpuRegister>();
+  CpuRegister offset = locations->InAt(2).AsRegister<CpuRegister>();
+  CpuRegister trg = locations->Out().AsRegister<CpuRegister>();
+
+  if (is_long) {
+    __ movq(trg, Address(base, offset, ScaleFactor::TIMES_1, 0));
+  } else {
+    // TODO: Distinguish object. In case we move to an actual compressed heap, retrieving an object
+    //       pointer will entail an unpack operation.
+    __ movl(trg, Address(base, offset, ScaleFactor::TIMES_1, 0));
+  }
+}
+
+static void CreateIntIntIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetInAt(1, Location::RequiresRegister());
+  locations->SetInAt(2, Location::RequiresRegister());
+  locations->SetOut(Location::SameAsFirstInput());
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitUnsafeGet(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafeGetVolatile(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafeGetLong(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitUnsafeGet(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), false, false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafeGetVolatile(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), false, true, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafeGetLong(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), true, false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), true, true, GetAssembler());
+}
+
+static void CreateIntIntIntIntToVoidPlusTempsLocations(ArenaAllocator* arena,
+                                                       Primitive::Type type,
+                                                       HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::NoLocation());
+  locations->SetInAt(1, Location::RequiresRegister());
+  locations->SetInAt(2, Location::RequiresRegister());
+  locations->SetInAt(3, Location::RequiresRegister());
+  if (type == Primitive::kPrimNot) {
+    // Need temp registers for card-marking.
+    locations->AddTemp(Location::RequiresRegister());
+    locations->AddTemp(Location::RequiresRegister());
+  }
+}
+
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePut(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutObject(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutLong(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke);
+}
+void IntrinsicLocationsBuilderX86_64::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke);
+}
+
+// We don't care for ordered: it requires an AnyStore barrier, which is already given by the x86
+// memory model.
+static void GenUnsafePut(LocationSummary* locations, Primitive::Type type, bool is_volatile,
+                         CodeGeneratorX86_64* codegen) {
+  X86_64Assembler* assembler = reinterpret_cast<X86_64Assembler*>(codegen->GetAssembler());
+  CpuRegister base = locations->InAt(1).AsRegister<CpuRegister>();
+  CpuRegister offset = locations->InAt(2).AsRegister<CpuRegister>();
+  CpuRegister value = locations->InAt(3).AsRegister<CpuRegister>();
+
+  if (type == Primitive::kPrimLong) {
+    __ movq(Address(base, offset, ScaleFactor::TIMES_1, 0), value);
+  } else {
+    __ movl(Address(base, offset, ScaleFactor::TIMES_1, 0), value);
+  }
+
+  if (is_volatile) {
+    __ mfence();
+  }
+
+  if (type == Primitive::kPrimNot) {
+    codegen->MarkGCCard(locations->GetTemp(0).AsRegister<CpuRegister>(),
+                        locations->GetTemp(1).AsRegister<CpuRegister>(),
+                        base,
+                        value);
+  }
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePut(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, true, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutObject(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, true, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutLong(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86_64::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, true, codegen_);
+}
+
+// Unimplemented intrinsics.
+
+#define UNIMPLEMENTED_INTRINSIC(Name)                                                   \
+void IntrinsicLocationsBuilderX86_64::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) { \
+}                                                                                       \
+void IntrinsicCodeGeneratorX86_64::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) {    \
+}
+
+UNIMPLEMENTED_INTRINSIC(IntegerReverse)
+UNIMPLEMENTED_INTRINSIC(LongReverse)
+UNIMPLEMENTED_INTRINSIC(MathFloor)
+UNIMPLEMENTED_INTRINSIC(MathCeil)
+UNIMPLEMENTED_INTRINSIC(MathRint)
+UNIMPLEMENTED_INTRINSIC(MathRoundDouble)
+UNIMPLEMENTED_INTRINSIC(MathRoundFloat)
+UNIMPLEMENTED_INTRINSIC(StringIsEmpty)  // Might not want to do these two anyways, inlining should
+UNIMPLEMENTED_INTRINSIC(StringLength)   // be good enough here.
+UNIMPLEMENTED_INTRINSIC(StringCompareTo)
+UNIMPLEMENTED_INTRINSIC(StringIndexOf)
+UNIMPLEMENTED_INTRINSIC(StringIndexOfAfter)
+UNIMPLEMENTED_INTRINSIC(SystemArrayCopyChar)
+UNIMPLEMENTED_INTRINSIC(UnsafeCASInt)
+UNIMPLEMENTED_INTRINSIC(UnsafeCASLong)
+UNIMPLEMENTED_INTRINSIC(UnsafeCASObject)
+UNIMPLEMENTED_INTRINSIC(ReferenceGetReferent)
+
+}  // namespace x86_64
+}  // namespace art
diff --git a/compiler/optimizing/intrinsics_x86_64.h b/compiler/optimizing/intrinsics_x86_64.h
new file mode 100644
index 0000000..c1fa99c
--- /dev/null
+++ b/compiler/optimizing/intrinsics_x86_64.h
@@ -0,0 +1,83 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_INTRINSICS_X86_64_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_X86_64_H_
+
+#include "intrinsics.h"
+
+namespace art {
+
+class ArenaAllocator;
+class HInvokeStaticOrDirect;
+class HInvokeVirtual;
+
+namespace x86_64 {
+
+class CodeGeneratorX86_64;
+class X86_64Assembler;
+
+class IntrinsicLocationsBuilderX86_64 FINAL : public IntrinsicVisitor {
+ public:
+  explicit IntrinsicLocationsBuilderX86_64(ArenaAllocator* arena) : arena_(arena) {}
+
+  // Define visitor methods.
+
+#define OPTIMIZING_INTRINSICS(Name, IsStatic)   \
+  void Visit ## Name(HInvoke* invoke) OVERRIDE;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+  // Check whether an invoke is an intrinsic, and if so, create a location summary. Returns whether
+  // a corresponding LocationSummary with the intrinsified_ flag set was generated and attached to
+  // the invoke.
+  bool TryDispatch(HInvoke* invoke);
+
+ private:
+  ArenaAllocator* arena_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicLocationsBuilderX86_64);
+};
+
+class IntrinsicCodeGeneratorX86_64 FINAL : public IntrinsicVisitor {
+ public:
+  explicit IntrinsicCodeGeneratorX86_64(CodeGeneratorX86_64* codegen) : codegen_(codegen) {}
+
+  // Define visitor methods.
+
+#define OPTIMIZING_INTRINSICS(Name, IsStatic)   \
+  void Visit ## Name(HInvoke* invoke) OVERRIDE;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+ private:
+  X86_64Assembler* GetAssembler();
+
+  ArenaAllocator* GetArena();
+
+  CodeGeneratorX86_64* codegen_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicCodeGeneratorX86_64);
+};
+
+}  // namespace x86_64
+}  // namespace art
+
+#endif  // ART_COMPILER_OPTIMIZING_INTRINSICS_X86_64_H_
diff --git a/compiler/optimizing/locations.cc b/compiler/optimizing/locations.cc
index ed5e260..9f2f9ec 100644
--- a/compiler/optimizing/locations.cc
+++ b/compiler/optimizing/locations.cc
@@ -20,7 +20,9 @@
 
 namespace art {
 
-LocationSummary::LocationSummary(HInstruction* instruction, CallKind call_kind)
+LocationSummary::LocationSummary(HInstruction* instruction,
+                                 CallKind call_kind,
+                                 bool intrinsified)
     : inputs_(instruction->GetBlock()->GetGraph()->GetArena(), instruction->InputCount()),
       temps_(instruction->GetBlock()->GetGraph()->GetArena(), 0),
       environment_(instruction->GetBlock()->GetGraph()->GetArena(),
@@ -29,7 +31,8 @@ LocationSummary::LocationSummary(HInstruction* instruction, CallKind call_kind)
       call_kind_(call_kind),
       stack_mask_(nullptr),
       register_mask_(0),
-      live_registers_() {
+      live_registers_(),
+      intrinsified_(intrinsified) {
   inputs_.SetSize(instruction->InputCount());
   for (size_t i = 0; i < instruction->InputCount(); ++i) {
     inputs_.Put(i, Location());
diff --git a/compiler/optimizing/locations.h b/compiler/optimizing/locations.h
index 7df99d4..d41b3ae 100644
--- a/compiler/optimizing/locations.h
+++ b/compiler/optimizing/locations.h
@@ -463,7 +463,9 @@ class LocationSummary : public ArenaObject<kArenaAllocMisc> {
     kCall
   };
 
-  LocationSummary(HInstruction* instruction, CallKind call_kind = kNoCall);
+  LocationSummary(HInstruction* instruction,
+                  CallKind call_kind = kNoCall,
+                  bool intrinsified = false);
 
   void SetInAt(uint32_t at, Location location) {
     DCHECK(inputs_.Get(at).IsUnallocated() || inputs_.Get(at).IsInvalid());
@@ -574,6 +576,10 @@ class LocationSummary : public ArenaObject<kArenaAllocMisc> {
     return output_overlaps_;
   }
 
+  bool Intrinsified() const {
+    return intrinsified_;
+  }
+
  private:
   GrowableArray<Location> inputs_;
   GrowableArray<Location> temps_;
@@ -593,6 +599,9 @@ class LocationSummary : public ArenaObject<kArenaAllocMisc> {
   // Registers that are in use at this position.
   RegisterSet live_registers_;
 
+  // Whether these are locations for an intrinsified call.
+  const bool intrinsified_;
+
   ART_FRIEND_TEST(RegisterAllocatorTest, ExpectedInRegisterHint);
   ART_FRIEND_TEST(RegisterAllocatorTest, SameAsFirstInputHint);
   DISALLOW_COPY_AND_ASSIGN(LocationSummary);
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index b98bc70..4838e1a 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1580,19 +1580,18 @@ class HLongConstant : public HConstant {
   DISALLOW_COPY_AND_ASSIGN(HLongConstant);
 };
 
+enum class Intrinsics {
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) k ## Name,
+#include "intrinsics_list.h"
+  kNone,
+  INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+};
+std::ostream& operator<<(std::ostream& os, const Intrinsics& intrinsic);
+
 class HInvoke : public HInstruction {
  public:
-  HInvoke(ArenaAllocator* arena,
-          uint32_t number_of_arguments,
-          Primitive::Type return_type,
-          uint32_t dex_pc)
-    : HInstruction(SideEffects::All()),
-      inputs_(arena, number_of_arguments),
-      return_type_(return_type),
-      dex_pc_(dex_pc) {
-    inputs_.SetSize(number_of_arguments);
-  }
-
   virtual size_t InputCount() const { return inputs_.Size(); }
   virtual HInstruction* InputAt(size_t i) const { return inputs_.Get(i); }
 
@@ -1612,12 +1611,38 @@ class HInvoke : public HInstruction {
 
   uint32_t GetDexPc() const { return dex_pc_; }
 
+  uint32_t GetDexMethodIndex() const { return dex_method_index_; }
+
+  Intrinsics GetIntrinsic() {
+    return intrinsic_;
+  }
+
+  void SetIntrinsic(Intrinsics intrinsic) {
+    intrinsic_ = intrinsic;
+  }
+
   DECLARE_INSTRUCTION(Invoke);
 
  protected:
+  HInvoke(ArenaAllocator* arena,
+          uint32_t number_of_arguments,
+          Primitive::Type return_type,
+          uint32_t dex_pc,
+          uint32_t dex_method_index)
+    : HInstruction(SideEffects::All()),
+      inputs_(arena, number_of_arguments),
+      return_type_(return_type),
+      dex_pc_(dex_pc),
+      dex_method_index_(dex_method_index),
+      intrinsic_(Intrinsics::kNone) {
+    inputs_.SetSize(number_of_arguments);
+  }
+
   GrowableArray<HInstruction*> inputs_;
   const Primitive::Type return_type_;
   const uint32_t dex_pc_;
+  const uint32_t dex_method_index_;
+  Intrinsics intrinsic_;
 
  private:
   DISALLOW_COPY_AND_ASSIGN(HInvoke);
@@ -1629,19 +1654,16 @@ class HInvokeStaticOrDirect : public HInvoke {
                         uint32_t number_of_arguments,
                         Primitive::Type return_type,
                         uint32_t dex_pc,
-                        uint32_t index_in_dex_cache,
+                        uint32_t dex_method_index,
                         InvokeType invoke_type)
-      : HInvoke(arena, number_of_arguments, return_type, dex_pc),
-        index_in_dex_cache_(index_in_dex_cache),
+      : HInvoke(arena, number_of_arguments, return_type, dex_pc, dex_method_index),
         invoke_type_(invoke_type) {}
 
-  uint32_t GetIndexInDexCache() const { return index_in_dex_cache_; }
   InvokeType GetInvokeType() const { return invoke_type_; }
 
   DECLARE_INSTRUCTION(InvokeStaticOrDirect);
 
  private:
-  const uint32_t index_in_dex_cache_;
   const InvokeType invoke_type_;
 
   DISALLOW_COPY_AND_ASSIGN(HInvokeStaticOrDirect);
@@ -1653,8 +1675,9 @@ class HInvokeVirtual : public HInvoke {
                  uint32_t number_of_arguments,
                  Primitive::Type return_type,
                  uint32_t dex_pc,
+                 uint32_t dex_method_index,
                  uint32_t vtable_index)
-      : HInvoke(arena, number_of_arguments, return_type, dex_pc),
+      : HInvoke(arena, number_of_arguments, return_type, dex_pc, dex_method_index),
         vtable_index_(vtable_index) {}
 
   uint32_t GetVTableIndex() const { return vtable_index_; }
@@ -1675,8 +1698,7 @@ class HInvokeInterface : public HInvoke {
                    uint32_t dex_pc,
                    uint32_t dex_method_index,
                    uint32_t imt_index)
-      : HInvoke(arena, number_of_arguments, return_type, dex_pc),
-        dex_method_index_(dex_method_index),
+      : HInvoke(arena, number_of_arguments, return_type, dex_pc, dex_method_index),
         imt_index_(imt_index) {}
 
   uint32_t GetImtIndex() const { return imt_index_; }
@@ -1685,7 +1707,6 @@ class HInvokeInterface : public HInvoke {
   DECLARE_INSTRUCTION(InvokeInterface);
 
  private:
-  const uint32_t dex_method_index_;
   const uint32_t imt_index_;
 
   DISALLOW_COPY_AND_ASSIGN(HInvokeInterface);
diff --git a/compiler/optimizing/optimizing_compiler.cc b/compiler/optimizing/optimizing_compiler.cc
index eaecbb0..6056373 100644
--- a/compiler/optimizing/optimizing_compiler.cc
+++ b/compiler/optimizing/optimizing_compiler.cc
@@ -25,6 +25,7 @@
 #include "compiler.h"
 #include "constant_folding.h"
 #include "dead_code_elimination.h"
+#include "dex/quick/dex_file_to_method_inliner_map.h"
 #include "driver/compiler_driver.h"
 #include "driver/dex_compilation_unit.h"
 #include "elf_writer_quick.h"
@@ -32,6 +33,7 @@
 #include "gvn.h"
 #include "inliner.h"
 #include "instruction_simplifier.h"
+#include "intrinsics.h"
 #include "jni/quick/jni_compiler.h"
 #include "mirror/art_method-inl.h"
 #include "nodes.h"
@@ -215,9 +217,12 @@ static void RunOptimizations(HGraph* graph,
   BoundsCheckElimination bce(graph);
   InstructionSimplifier simplify2(graph);
 
+  IntrinsicsRecognizer intrinsics(graph, dex_compilation_unit.GetDexFile(), driver);
+
   HOptimization* optimizations[] = {
     &redundant_phi,
     &dead_phi,
+    &intrinsics,
     &dce,
     &fold,
     &simplify1,
diff --git a/compiler/optimizing/register_allocator.cc b/compiler/optimizing/register_allocator.cc
index d2f4f9b..1d155f9 100644
--- a/compiler/optimizing/register_allocator.cc
+++ b/compiler/optimizing/register_allocator.cc
@@ -1217,10 +1217,17 @@ void RegisterAllocator::ConnectSiblings(LiveInterval* interval) {
         locations->SetEnvironmentAt(use->GetInputIndex(), source);
       } else {
         Location expected_location = locations->InAt(use->GetInputIndex());
-        if (expected_location.IsUnallocated()) {
-          locations->SetInAt(use->GetInputIndex(), source);
-        } else if (!expected_location.IsConstant()) {
-          AddInputMoveFor(use->GetUser(), source, expected_location);
+        // The expected (actual) location may be invalid in case the input is unused. Currently
+        // this only happens for intrinsics.
+        if (expected_location.IsValid()) {
+          if (expected_location.IsUnallocated()) {
+            locations->SetInAt(use->GetInputIndex(), source);
+          } else if (!expected_location.IsConstant()) {
+            AddInputMoveFor(use->GetUser(), source, expected_location);
+          }
+        } else {
+          DCHECK(use->GetUser()->IsInvoke());
+          DCHECK(use->GetUser()->AsInvoke()->GetIntrinsic() != Intrinsics::kNone);
         }
       }
       use = use->GetNext();
diff --git a/compiler/utils/x86_64/assembler_x86_64.cc b/compiler/utils/x86_64/assembler_x86_64.cc
index d843a72..c7414a1 100644
--- a/compiler/utils/x86_64/assembler_x86_64.cc
+++ b/compiler/utils/x86_64/assembler_x86_64.cc
@@ -184,6 +184,20 @@ void X86_64Assembler::movl(const Address& dst, const Immediate& imm) {
   EmitImmediate(imm);
 }
 
+
+void X86_64Assembler::cmov(Condition c, CpuRegister dst, CpuRegister src) {
+  cmov(c, dst, src, true);
+}
+
+void X86_64Assembler::cmov(Condition c, CpuRegister dst, CpuRegister src, bool is64bit) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, is64bit, dst.NeedsRex(), false, src.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0x40 + c);
+  EmitRegisterOperand(dst.LowBits(), src.LowBits());
+}
+
+
 void X86_64Assembler::movzxb(CpuRegister dst, CpuRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitOptionalByteRegNormalizingRex32(dst, src);
@@ -369,19 +383,26 @@ void X86_64Assembler::movsxd(CpuRegister dst, const Address& src) {
 
 
 void X86_64Assembler::movd(XmmRegister dst, CpuRegister src) {
+  movd(dst, src, true);
+}
+
+void X86_64Assembler::movd(CpuRegister dst, XmmRegister src) {
+  movd(dst, src, true);
+}
+
+void X86_64Assembler::movd(XmmRegister dst, CpuRegister src, bool is64bit) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
-  EmitRex64(dst, src);
+  EmitOptionalRex(false, is64bit, dst.NeedsRex(), false, src.NeedsRex());
   EmitUint8(0x0F);
   EmitUint8(0x6E);
   EmitOperand(dst.LowBits(), Operand(src));
 }
 
-
-void X86_64Assembler::movd(CpuRegister dst, XmmRegister src) {
+void X86_64Assembler::movd(CpuRegister dst, XmmRegister src, bool is64bit) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
-  EmitRex64(src, dst);
+  EmitOptionalRex(false, is64bit, src.NeedsRex(), false, dst.NeedsRex());
   EmitUint8(0x0F);
   EmitUint8(0x7E);
   EmitOperand(src.LowBits(), Operand(dst));
@@ -826,6 +847,39 @@ void X86_64Assembler::andpd(XmmRegister dst, const Address& src) {
   EmitOperand(dst.LowBits(), src);
 }
 
+void X86_64Assembler::andpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::andps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::orpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::orps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
 
 void X86_64Assembler::fldl(const Address& src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
@@ -1757,6 +1811,20 @@ void X86_64Assembler::setcc(Condition condition, CpuRegister dst) {
   EmitUint8(0xC0 + dst.LowBits());
 }
 
+void X86_64Assembler::bswapl(CpuRegister dst) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, false, false, false, dst.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0xC8 + dst.LowBits());
+}
+
+void X86_64Assembler::bswapq(CpuRegister dst) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, true, false, false, dst.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0xC8 + dst.LowBits());
+}
+
 
 void X86_64Assembler::LoadDoubleConstant(XmmRegister dst, double value) {
   // TODO: Need to have a code constants table.
diff --git a/compiler/utils/x86_64/assembler_x86_64.h b/compiler/utils/x86_64/assembler_x86_64.h
index ac8bc9a..5c8d608 100644
--- a/compiler/utils/x86_64/assembler_x86_64.h
+++ b/compiler/utils/x86_64/assembler_x86_64.h
@@ -276,6 +276,9 @@ class X86_64Assembler FINAL : public Assembler {
   void movl(const Address& dst, CpuRegister src);
   void movl(const Address& dst, const Immediate& imm);
 
+  void cmov(Condition c, CpuRegister dst, CpuRegister src);  // This is the 64b version.
+  void cmov(Condition c, CpuRegister dst, CpuRegister src, bool is64bit);
+
   void movzxb(CpuRegister dst, CpuRegister src);
   void movzxb(CpuRegister dst, const Address& src);
   void movsxb(CpuRegister dst, CpuRegister src);
@@ -303,8 +306,10 @@ class X86_64Assembler FINAL : public Assembler {
   void movsxd(CpuRegister dst, CpuRegister src);
   void movsxd(CpuRegister dst, const Address& src);
 
-  void movd(XmmRegister dst, CpuRegister src);
-  void movd(CpuRegister dst, XmmRegister src);
+  void movd(XmmRegister dst, CpuRegister src);  // Note: this is the r64 version, formally movq.
+  void movd(CpuRegister dst, XmmRegister src);  // Note: this is the r64 version, formally movq.
+  void movd(XmmRegister dst, CpuRegister src, bool is64bit);
+  void movd(CpuRegister dst, XmmRegister src, bool is64bit);
 
   void addss(XmmRegister dst, XmmRegister src);
   void addss(XmmRegister dst, const Address& src);
@@ -360,6 +365,11 @@ class X86_64Assembler FINAL : public Assembler {
   void xorps(XmmRegister dst, XmmRegister src);
 
   void andpd(XmmRegister dst, const Address& src);
+  void andpd(XmmRegister dst, XmmRegister src);
+  void andps(XmmRegister dst, XmmRegister src);
+
+  void orpd(XmmRegister dst, XmmRegister src);
+  void orps(XmmRegister dst, XmmRegister src);
 
   void flds(const Address& src);
   void fstps(const Address& dst);
@@ -504,6 +514,9 @@ class X86_64Assembler FINAL : public Assembler {
 
   void setcc(Condition condition, CpuRegister dst);
 
+  void bswapl(CpuRegister dst);
+  void bswapq(CpuRegister dst);
+
   //
   // Macros for High-level operations.
   //
diff --git a/compiler/utils/x86_64/assembler_x86_64_test.cc b/compiler/utils/x86_64/assembler_x86_64_test.cc
index e93f45c..6df4144 100644
--- a/compiler/utils/x86_64/assembler_x86_64_test.cc
+++ b/compiler/utils/x86_64/assembler_x86_64_test.cc
@@ -689,6 +689,22 @@ TEST_F(AssemblerX86_64Test, Xorpd) {
   DriverStr(RepeatFF(&x86_64::X86_64Assembler::xorpd, "xorpd %{reg2}, %{reg1}"), "xorpd");
 }
 
+TEST_F(AssemblerX86_64Test, Andps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::andps, "andps %{reg2}, %{reg1}"), "andps");
+}
+
+TEST_F(AssemblerX86_64Test, Andpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::andpd, "andpd %{reg2}, %{reg1}"), "andpd");
+}
+
+TEST_F(AssemblerX86_64Test, Orps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::orps, "orps %{reg2}, %{reg1}"), "orps");
+}
+
+TEST_F(AssemblerX86_64Test, Orpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::orpd, "orpd %{reg2}, %{reg1}"), "orpd");
+}
+
 // X87
 
 std::string x87_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED,
@@ -757,6 +773,14 @@ TEST_F(AssemblerX86_64Test, RetAndLeave) {
 // MISC //
 //////////
 
+TEST_F(AssemblerX86_64Test, Bswapl) {
+  DriverStr(Repeatr(&x86_64::X86_64Assembler::bswapl, "bswap %{reg}"), "bswapl");
+}
+
+TEST_F(AssemblerX86_64Test, Bswapq) {
+  DriverStr(RepeatR(&x86_64::X86_64Assembler::bswapq, "bswap %{reg}"), "bswapq");
+}
+
 std::string setcc_test_fn(AssemblerX86_64Test::Base* assembler_test,
                           x86_64::X86_64Assembler* assembler) {
   // From Condition
diff --git a/runtime/quick/inline_method_analyser.h b/runtime/quick/inline_method_analyser.h
index 72b696b..3463025 100644
--- a/runtime/quick/inline_method_analyser.h
+++ b/runtime/quick/inline_method_analyser.h
@@ -103,6 +103,9 @@ enum IntrinsicFlags {
   kIntrinsicFlagIsObject   = 4,
   // kIntrinsicUnsafePut
   kIntrinsicFlagIsOrdered  = 8,
+
+  // kIntrinsicDoubleCvt, kIntrinsicFloatCvt.
+  kIntrinsicFlagToFloatingPoint = kIntrinsicFlagMin,
 };
 
 struct InlineIGetIPutData {
diff --git a/test/082-inline-execute/src/Main.java b/test/082-inline-execute/src/Main.java
index 56972ff..862fe06 100644
--- a/test/082-inline-execute/src/Main.java
+++ b/test/082-inline-execute/src/Main.java
@@ -119,6 +119,9 @@ public class Main {
     }
   }
 
+  // Break up the charAt tests. The optimizing compiler doesn't optimize methods with try-catch yet,
+  // so we need to separate out the tests that are expected to throw exception
+
   public static void test_String_charAt() {
     String testStr = "Now is the time";
 
@@ -127,6 +130,12 @@ public class Main {
     Assert.assertEquals(' ', testStr.charAt(10));
     Assert.assertEquals('e', testStr.charAt(testStr.length()-1));
 
+    test_String_charAtExc();
+    test_String_charAtExc2();
+  }
+
+  private static void test_String_charAtExc() {
+    String testStr = "Now is the time";
     try {
       testStr.charAt(-1);
       Assert.fail();
@@ -146,6 +155,19 @@ public class Main {
     }
   }
 
+  private static void test_String_charAtExc2() {
+    try {
+      test_String_charAtExc3();
+      Assert.fail();
+    } catch (StringIndexOutOfBoundsException expected) {
+    }
+  }
+
+  private static void test_String_charAtExc3() {
+    String testStr = "Now is the time";
+    Assert.assertEquals('N', testStr.charAt(-1));
+  }
+
   static int start;
   private static int[] negIndex = { -100000 };
   public static void test_String_indexOf() {