[optimizing] Implement X86 intrinsic support

Implement the supported intrinsics for X86.

Enhance the graph visualizer to print <U> for unallocated locations, to
allow calling the graph dumper from within register allocation for
debugging purposes.

Change-Id: I3b0319eb70a9a4ea228f67065b4c52d13a1ae775
Signed-off-by: Mark Mendell <mark.p.mendell@intel.com>

8 files changed, 1401 insertions, 37 deletions

diff --git a/compiler/Android.mk b/compiler/Android.mk
index 6b0e6ff..0247c9d 100644
--- a/compiler/Android.mk
+++ b/compiler/Android.mk
@@ -112,6 +112,7 @@ LIBART_COMPILER_SRC_FILES := \
 	optimizing/intrinsics.cc \
 	optimizing/intrinsics_arm.cc \
 	optimizing/intrinsics_arm64.cc \
+	optimizing/intrinsics_x86.cc \
 	optimizing/intrinsics_x86_64.cc \
 	optimizing/licm.cc \
 	optimizing/locations.cc \
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 4414a65..b18cdd5 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -19,6 +19,8 @@
 #include "entrypoints/quick/quick_entrypoints.h"
 #include "entrypoints/quick/quick_entrypoints_enum.h"
 #include "gc/accounting/card_table.h"
+#include "intrinsics.h"
+#include "intrinsics_x86.h"
 #include "mirror/array-inl.h"
 #include "mirror/art_method.h"
 #include "mirror/class.h"
@@ -60,20 +62,6 @@ class InvokeRuntimeCallingConvention : public CallingConvention<Register, XmmReg
 
 #define __ reinterpret_cast<X86Assembler*>(codegen->GetAssembler())->
 
-class SlowPathCodeX86 : public SlowPathCode {
- public:
-  SlowPathCodeX86() : 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);
-};
-
 class NullCheckSlowPathX86 : public SlowPathCodeX86 {
  public:
   explicit NullCheckSlowPathX86(HNullCheck* instruction) : instruction_(instruction) {}
@@ -1140,35 +1128,30 @@ void InstructionCodeGeneratorX86::VisitReturn(HReturn* ret) {
 }
 
 void LocationsBuilderX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+  IntrinsicLocationsBuilderX86 intrinsic(GetGraph()->GetArena());
+  if (intrinsic.TryDispatch(invoke)) {
+    return;
+  }
+
   HandleInvoke(invoke);
 }
 
-void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
-  Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
-
-  // 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* codegen) {
+  if (invoke->GetLocations()->Intrinsified()) {
+    IntrinsicCodeGeneratorX86 intrinsic(codegen);
+    intrinsic.Dispatch(invoke);
+    return true;
+  }
+  return false;
+}
 
-  // temp = method;
-  codegen_->LoadCurrentMethod(temp);
-  if (!invoke->IsRecursive()) {
-    // 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->GetDexMethodIndex())));
-    // (temp + offset_of_quick_compiled_code)()
-    __ call(Address(
-        temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
-  } else {
-    __ call(codegen_->GetFrameEntryLabel());
+void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+  if (TryGenerateIntrinsicCode(invoke, codegen_)) {
+    return;
   }
 
-  DCHECK(!codegen_->IsLeafMethod());
-  codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
+  codegen_->GenerateStaticOrDirectCall(
+      invoke, invoke->GetLocations()->GetTemp(0).AsRegister<Register>());
 }
 
 void LocationsBuilderX86::VisitInvokeVirtual(HInvokeVirtual* invoke) {
@@ -2863,6 +2846,32 @@ void InstructionCodeGeneratorX86::GenerateMemoryBarrier(MemBarrierKind kind) {
 }
 
 
+void CodeGeneratorX86::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke,
+                                                  Register temp) {
+  // 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);
+  if (!invoke->IsRecursive()) {
+    // 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->GetDexMethodIndex())));
+    // (temp + offset_of_quick_compiled_code)()
+    __ call(Address(
+        temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
+  } else {
+    __ call(GetFrameEntryLabel());
+  }
+
+  DCHECK(!IsLeafMethod());
+  RecordPcInfo(invoke, invoke->GetDexPc());
+}
+
 void CodeGeneratorX86::MarkGCCard(Register temp, Register card, Register object, Register value) {
   Label is_null;
   __ testl(value, value);
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index c5763de..9b4b3db 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -228,6 +228,9 @@ class CodeGeneratorX86 : public CodeGenerator {
   // Helper method to move a 64bits value between two locations.
   void Move64(Location destination, Location source);
 
+  // Generate a call to a static or direct method.
+  void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Register temp);
+
   // Emit a write barrier.
   void MarkGCCard(Register temp, Register card, Register object, Register value);
 
@@ -261,6 +264,20 @@ class CodeGeneratorX86 : public CodeGenerator {
   DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86);
 };
 
+class SlowPathCodeX86 : public SlowPathCode {
+ public:
+  SlowPathCodeX86() : 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);
+};
+
 }  // namespace x86
 }  // namespace art
 
diff --git a/compiler/optimizing/graph_visualizer.cc b/compiler/optimizing/graph_visualizer.cc
index cabfa48..49c0d38 100644
--- a/compiler/optimizing/graph_visualizer.cc
+++ b/compiler/optimizing/graph_visualizer.cc
@@ -149,6 +149,8 @@ class HGraphVisualizerPrinter : public HGraphVisitor {
       codegen_.DumpCoreRegister(output_, location.low());
       output_ << " and ";
       codegen_.DumpCoreRegister(output_, location.high());
+    } else if (location.IsUnallocated()) {
+      output_ << "<U>";
     } else {
       DCHECK(location.IsDoubleStackSlot());
       output_ << "2x" << location.GetStackIndex() << "(sp)";
diff --git a/compiler/optimizing/intrinsics_x86.cc b/compiler/optimizing/intrinsics_x86.cc
new file mode 100644
index 0000000..bcf947f
--- /dev/null
+++ b/compiler/optimizing/intrinsics_x86.cc
@@ -0,0 +1,1180 @@
+/*
+ * 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.h"
+
+#include "code_generator_x86.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/assembler_x86.h"
+#include "utils/x86/constants_x86.h"
+
+namespace art {
+
+namespace x86 {
+
+static constexpr int kDoubleNaNHigh = 0x7FF80000;
+static constexpr int kDoubleNaNLow = 0x00000000;
+static constexpr int kFloatNaN = 0x7FC00000;
+
+X86Assembler* IntrinsicCodeGeneratorX86::GetAssembler() {
+  return reinterpret_cast<X86Assembler*>(codegen_->GetAssembler());
+}
+
+ArenaAllocator* IntrinsicCodeGeneratorX86::GetAllocator() {
+  return codegen_->GetGraph()->GetArena();
+}
+
+bool IntrinsicLocationsBuilderX86::TryDispatch(HInvoke* invoke) {
+  Dispatch(invoke);
+  LocationSummary* res = invoke->GetLocations();
+  return res != nullptr && res->Intrinsified();
+}
+
+#define __ reinterpret_cast<X86Assembler*>(codegen->GetAssembler())->
+
+// TODO: target as memory.
+static void MoveFromReturnRegister(Location target,
+                                   Primitive::Type type,
+                                   CodeGeneratorX86* codegen) {
+  if (!target.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: {
+      Register target_reg = target.AsRegister<Register>();
+      if (target_reg != EAX) {
+        __ movl(target_reg, EAX);
+      }
+      break;
+    }
+    case Primitive::kPrimLong: {
+      Register target_reg_lo = target.AsRegisterPairLow<Register>();
+      Register target_reg_hi = target.AsRegisterPairHigh<Register>();
+      if (target_reg_lo != EAX) {
+        __ movl(target_reg_lo, EAX);
+      }
+      if (target_reg_hi != EDX) {
+        __ movl(target_reg_hi, EDX);
+      }
+      break;
+    }
+
+    case Primitive::kPrimVoid:
+      LOG(FATAL) << "Unexpected void type for valid location " << target;
+      UNREACHABLE();
+
+    case Primitive::kPrimDouble: {
+      XmmRegister target_reg = target.AsFpuRegister<XmmRegister>();
+      if (target_reg != XMM0) {
+        __ movsd(target_reg, XMM0);
+      }
+      break;
+    }
+    case Primitive::kPrimFloat: {
+      XmmRegister target_reg = target.AsFpuRegister<XmmRegister>();
+      if (target_reg != XMM0) {
+        __ movss(target_reg, XMM0);
+      }
+      break;
+    }
+  }
+}
+
+static void MoveArguments(HInvoke* invoke, ArenaAllocator* arena, CodeGeneratorX86* 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(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 : public SlowPathCodeX86 {
+ public:
+  explicit IntrinsicSlowPathX86(HInvoke* invoke, Register temp)
+    : invoke_(invoke) {
+      // The temporary register has to be EAX for x86 invokes.
+      DCHECK_EQ(temp, EAX);
+    }
+
+  void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
+    CodeGeneratorX86* codegen = down_cast<CodeGeneratorX86*>(codegen_in);
+    __ Bind(GetEntryLabel());
+
+    SaveLiveRegisters(codegen, invoke_->GetLocations());
+
+    MoveArguments(invoke_, codegen->GetGraph()->GetArena(), codegen);
+
+    if (invoke_->IsInvokeStaticOrDirect()) {
+      codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(), EAX);
+    } 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);
+    }
+
+    RestoreLiveRegisters(codegen, invoke_->GetLocations());
+    __ jmp(GetExitLabel());
+  }
+
+ private:
+  // The instruction where this slow path is happening.
+  HInvoke* const invoke_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathX86);
+};
+
+#undef __
+#define __ assembler->
+
+static void CreateFPToIntLocations(ArenaAllocator* arena, HInvoke* invoke, bool is64bit) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresFpuRegister());
+  locations->SetOut(Location::RequiresRegister());
+  if (is64bit) {
+    locations->AddTemp(Location::RequiresFpuRegister());
+  }
+}
+
+static void CreateIntToFPLocations(ArenaAllocator* arena, HInvoke* invoke, bool is64bit) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::RequiresFpuRegister());
+  if (is64bit) {
+    locations->AddTemp(Location::RequiresFpuRegister());
+    locations->AddTemp(Location::RequiresFpuRegister());
+  }
+}
+
+static void MoveFPToInt(LocationSummary* locations, bool is64bit, X86Assembler* assembler) {
+  Location input = locations->InAt(0);
+  Location output = locations->Out();
+  if (is64bit) {
+    // Need to use the temporary.
+    XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+    __ movsd(temp, input.AsFpuRegister<XmmRegister>());
+    __ movd(output.AsRegisterPairLow<Register>(), temp);
+    __ psrlq(temp, Immediate(32));
+    __ movd(output.AsRegisterPairHigh<Register>(), temp);
+  } else {
+    __ movd(output.AsRegister<Register>(), input.AsFpuRegister<XmmRegister>());
+  }
+}
+
+static void MoveIntToFP(LocationSummary* locations, bool is64bit, X86Assembler* assembler) {
+  Location input = locations->InAt(0);
+  Location output = locations->Out();
+  if (is64bit) {
+    // Need to use the temporary.
+    XmmRegister temp1 = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+    XmmRegister temp2 = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
+    __ movd(temp1, input.AsRegisterPairLow<Register>());
+    __ movd(temp2, input.AsRegisterPairHigh<Register>());
+    __ punpckldq(temp1, temp2);
+    __ movsd(output.AsFpuRegister<XmmRegister>(), temp1);
+  } else {
+    __ movd(output.AsFpuRegister<XmmRegister>(), input.AsRegister<Register>());
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+  CreateFPToIntLocations(arena_, invoke, true);
+}
+void IntrinsicLocationsBuilderX86::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+  CreateIntToFPLocations(arena_, invoke, true);
+}
+
+void IntrinsicCodeGeneratorX86::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+  MoveFPToInt(invoke->GetLocations(), true, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+  MoveIntToFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+  CreateFPToIntLocations(arena_, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitFloatIntBitsToFloat(HInvoke* invoke) {
+  CreateIntToFPLocations(arena_, invoke, false);
+}
+
+void IntrinsicCodeGeneratorX86::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+  MoveFPToInt(invoke->GetLocations(), false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::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 CreateLongToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::RequiresRegister());
+}
+
+static void CreateLongToLongLocations(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
+}
+
+static void GenReverseBytes(LocationSummary* locations,
+                            Primitive::Type size,
+                            X86Assembler* assembler) {
+  Register out = locations->Out().AsRegister<Register>();
+
+  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;
+    default:
+      LOG(FATAL) << "Unexpected size for reverse-bytes: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitIntegerReverseBytes(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitIntegerReverseBytes(HInvoke* invoke) {
+  GenReverseBytes(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitShortReverseBytes(HInvoke* invoke) {
+  CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::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 CreateFloatToFloat(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());
+}
+
+static void MathAbsFP(LocationSummary* locations, bool is64bit, X86Assembler* assembler) {
+  Location output = locations->Out();
+
+  if (output.IsFpuRegister()) {
+    // Create the right constant on an aligned stack.
+    if (is64bit) {
+      __ subl(ESP, Immediate(8));
+      __ pushl(Immediate(0x7FFFFFFF));
+      __ pushl(Immediate(0xFFFFFFFF));
+      __ andpd(output.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
+    } else {
+      __ subl(ESP, Immediate(12));
+      __ pushl(Immediate(0x7FFFFFFF));
+      __ andps(output.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
+    }
+    __ addl(ESP, Immediate(16));
+  } 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());
+//      __ andl(Address(Register(RSP), output.GetHighStackIndex(kX86WordSize)),
+//              Immediate(0x7FFFFFFF));
+//    } else {
+//      DCHECK(output.IsStackSlot());
+//      // Can use and with a literal directly.
+//      __ andl(Address(Register(RSP), output.GetStackIndex()), Immediate(0x7FFFFFFF));
+//    }
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathAbsDouble(HInvoke* invoke) {
+  CreateFloatToFloat(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathAbsDouble(HInvoke* invoke) {
+  MathAbsFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathAbsFloat(HInvoke* invoke) {
+  CreateFloatToFloat(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathAbsFloat(HInvoke* invoke) {
+  MathAbsFP(invoke->GetLocations(), false, GetAssembler());
+}
+
+static void CreateAbsIntLocation(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RegisterLocation(EAX));
+  locations->SetOut(Location::SameAsFirstInput());
+  locations->AddTemp(Location::RegisterLocation(EDX));
+}
+
+static void GenAbsInteger(LocationSummary* locations, X86Assembler* assembler) {
+  Location output = locations->Out();
+  Register out = output.AsRegister<Register>();
+  DCHECK_EQ(out, EAX);
+  Register temp = locations->GetTemp(0).AsRegister<Register>();
+  DCHECK_EQ(temp, EDX);
+
+  // Sign extend EAX into EDX.
+  __ cdq();
+
+  // XOR EAX with sign.
+  __ xorl(EAX, EDX);
+
+  // Subtract out sign to correct.
+  __ subl(EAX, EDX);
+
+  // The result is in EAX.
+}
+
+static void CreateAbsLongLocation(ArenaAllocator* arena, HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
+  locations->AddTemp(Location::RequiresRegister());
+}
+
+static void GenAbsLong(LocationSummary* locations, X86Assembler* assembler) {
+  Location input = locations->InAt(0);
+  Register input_lo = input.AsRegisterPairLow<Register>();
+  Register input_hi = input.AsRegisterPairHigh<Register>();
+  Location output = locations->Out();
+  Register output_lo = output.AsRegisterPairLow<Register>();
+  Register output_hi = output.AsRegisterPairHigh<Register>();
+  Register temp = locations->GetTemp(0).AsRegister<Register>();
+
+  // Compute the sign into the temporary.
+  __ movl(temp, input_hi);
+  __ sarl(temp, Immediate(31));
+
+  // Store the sign into the output.
+  __ movl(output_lo, temp);
+  __ movl(output_hi, temp);
+
+  // XOR the input to the output.
+  __ xorl(output_lo, input_lo);
+  __ xorl(output_hi, input_hi);
+
+  // Subtract the sign.
+  __ subl(output_lo, temp);
+  __ sbbl(output_hi, temp);
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathAbsInt(HInvoke* invoke) {
+  CreateAbsIntLocation(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathAbsInt(HInvoke* invoke) {
+  GenAbsInteger(invoke->GetLocations(), GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathAbsLong(HInvoke* invoke) {
+  CreateAbsLongLocation(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathAbsLong(HInvoke* invoke) {
+  GenAbsLong(invoke->GetLocations(), GetAssembler());
+}
+
+static void GenMinMaxFP(LocationSummary* locations, bool is_min, bool is_double,
+                        X86Assembler* 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);
+  if (is_double) {
+    __ pushl(Immediate(kDoubleNaNHigh));
+    __ pushl(Immediate(kDoubleNaNLow));
+    __ movsd(out, Address(ESP, 0));
+    __ addl(ESP, Immediate(8));
+  } else {
+    __ pushl(Immediate(kFloatNaN));
+    __ movss(out, Address(ESP, 0));
+    __ addl(ESP, Immediate(4));
+  }
+  __ jmp(&done);
+
+  // out := op2;
+  __ Bind(&op2_label);
+  if (is_double) {
+    __ movsd(out, op2);
+  } else {
+    __ movss(out, op2);
+  }
+
+  // Done.
+  __ Bind(&done);
+}
+
+static void CreateFPFPToFPLocations(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());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMinDoubleDouble(HInvoke* invoke) {
+  CreateFPFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMinDoubleDouble(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), true, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMinFloatFloat(HInvoke* invoke) {
+  CreateFPFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMinFloatFloat(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), true, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMaxDoubleDouble(HInvoke* invoke) {
+  CreateFPFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMaxDoubleDouble(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), false, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMaxFloatFloat(HInvoke* invoke) {
+  CreateFPFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMaxFloatFloat(HInvoke* invoke) {
+  GenMinMaxFP(invoke->GetLocations(), false, false, GetAssembler());
+}
+
+static void GenMinMax(LocationSummary* locations, bool is_min, bool is_long,
+                      X86Assembler* 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;
+  }
+
+  if (is_long) {
+    // Need to perform a subtract to get the sign right.
+    // op1 is already in the same location as the output.
+    Location output = locations->Out();
+    Register output_lo = output.AsRegisterPairLow<Register>();
+    Register output_hi = output.AsRegisterPairHigh<Register>();
+
+    Register op2_lo = op2_loc.AsRegisterPairLow<Register>();
+    Register op2_hi = op2_loc.AsRegisterPairHigh<Register>();
+
+    // Spare register to compute the subtraction to set condition code.
+    Register temp = locations->GetTemp(0).AsRegister<Register>();
+
+    // Subtract off op2_low.
+    __ movl(temp, output_lo);
+    __ subl(temp, op2_lo);
+
+    // Now use the same tempo and the borrow to finish the subtraction of op2_hi.
+    __ movl(temp, output_hi);
+    __ sbbl(temp, op2_hi);
+
+    // Now the condition code is correct.
+    Condition cond = is_min ? Condition::kGreaterEqual : Condition::kLess;
+    __ cmovl(cond, output_lo, op2_lo);
+    __ cmovl(cond, output_hi, op2_hi);
+  } else {
+    Register out = locations->Out().AsRegister<Register>();
+    Register op2 = op2_loc.AsRegister<Register>();
+
+    //  (out := op1)
+    //  out <=? op2
+    //  if out is min jmp done
+    //  out := op2
+    // done:
+
+    __ cmpl(out, op2);
+    Condition cond = is_min ? Condition::kGreater : Condition::kLess;
+    __ cmovl(cond, out, op2);
+  }
+}
+
+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());
+}
+
+static void CreateLongLongToLongLocations(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());
+  // Register to use to perform a long subtract to set cc.
+  locations->AddTemp(Location::RequiresRegister());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMinIntInt(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMinIntInt(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), true, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMinLongLong(HInvoke* invoke) {
+  CreateLongLongToLongLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMinLongLong(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), true, true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMaxIntInt(HInvoke* invoke) {
+  CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMathMaxIntInt(HInvoke* invoke) {
+  GenMinMax(invoke->GetLocations(), false, false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMathMaxLongLong(HInvoke* invoke) {
+  CreateLongLongToLongLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::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::VisitMathSqrt(HInvoke* invoke) {
+  CreateFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::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::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());
+  // Needs to be EAX for the invoke.
+  locations->AddTemp(Location::RegisterLocation(EAX));
+}
+
+void IntrinsicCodeGeneratorX86::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();
+
+  Register obj = locations->InAt(0).AsRegister<Register>();
+  Register idx = locations->InAt(1).AsRegister<Register>();
+  Register out = locations->Out().AsRegister<Register>();
+  Location temp_loc = locations->GetTemp(0);
+  Register temp = temp_loc.AsRegister<Register>();
+
+  // 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* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke, temp);
+  codegen_->AddSlowPath(slow_path);
+
+  X86Assembler* assembler = GetAssembler();
+
+  __ cmpl(idx, Address(obj, count_offset));
+  codegen_->MaybeRecordImplicitNullCheck(invoke);
+  __ 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, X86Assembler* assembler) {
+  Register address = locations->InAt(0).AsRegisterPairLow<Register>();
+  Location out_loc = locations->Out();
+  // 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_loc.AsRegister<Register>(), Address(address, 0));
+      break;
+    case Primitive::kPrimShort:
+      __ movsxw(out_loc.AsRegister<Register>(), Address(address, 0));
+      break;
+    case Primitive::kPrimInt:
+      __ movl(out_loc.AsRegister<Register>(), Address(address, 0));
+      break;
+    case Primitive::kPrimLong:
+      __ movl(out_loc.AsRegisterPairLow<Register>(), Address(address, 0));
+      __ movl(out_loc.AsRegisterPairHigh<Register>(), Address(address, 4));
+      break;
+    default:
+      LOG(FATAL) << "Type not recognized for peek: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPeekByte(HInvoke* invoke) {
+  CreateLongToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPeekByte(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimByte, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPeekIntNative(HInvoke* invoke) {
+  CreateLongToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPeekIntNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPeekLongNative(HInvoke* invoke) {
+  CreateLongToLongLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPeekLongNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimLong, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPeekShortNative(HInvoke* invoke) {
+  CreateLongToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPeekShortNative(HInvoke* invoke) {
+  GenPeek(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
+}
+
+static void CreateLongIntToVoidLocations(ArenaAllocator* arena, Primitive::Type size,
+                                         HInvoke* invoke) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::RequiresRegister());
+  HInstruction *value = invoke->InputAt(1);
+  if (size == Primitive::kPrimByte) {
+    locations->SetInAt(1, Location::ByteRegisterOrConstant(EDX, value));
+  } else {
+    locations->SetInAt(1, Location::RegisterOrConstant(value));
+  }
+}
+
+static void GenPoke(LocationSummary* locations, Primitive::Type size, X86Assembler* assembler) {
+  Register address = locations->InAt(0).AsRegisterPairLow<Register>();
+  Location value_loc = locations->InAt(1);
+  // 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:
+      if (value_loc.IsConstant()) {
+        __ movb(Address(address, 0),
+                Immediate(value_loc.GetConstant()->AsIntConstant()->GetValue()));
+      } else {
+        __ movb(Address(address, 0), value_loc.AsRegister<ByteRegister>());
+      }
+      break;
+    case Primitive::kPrimShort:
+      if (value_loc.IsConstant()) {
+        __ movw(Address(address, 0),
+                Immediate(value_loc.GetConstant()->AsIntConstant()->GetValue()));
+      } else {
+        __ movw(Address(address, 0), value_loc.AsRegister<Register>());
+      }
+      break;
+    case Primitive::kPrimInt:
+      if (value_loc.IsConstant()) {
+        __ movl(Address(address, 0),
+                Immediate(value_loc.GetConstant()->AsIntConstant()->GetValue()));
+      } else {
+        __ movl(Address(address, 0), value_loc.AsRegister<Register>());
+      }
+      break;
+    case Primitive::kPrimLong:
+      if (value_loc.IsConstant()) {
+        int64_t value = value_loc.GetConstant()->AsLongConstant()->GetValue();
+        __ movl(Address(address, 0), Immediate(Low32Bits(value)));
+        __ movl(Address(address, 4), Immediate(High32Bits(value)));
+      } else {
+        __ movl(Address(address, 0), value_loc.AsRegisterPairLow<Register>());
+        __ movl(Address(address, 4), value_loc.AsRegisterPairHigh<Register>());
+      }
+      break;
+    default:
+      LOG(FATAL) << "Type not recognized for poke: " << size;
+      UNREACHABLE();
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPokeByte(HInvoke* invoke) {
+  CreateLongIntToVoidLocations(arena_, Primitive::kPrimByte, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPokeByte(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimByte, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPokeIntNative(HInvoke* invoke) {
+  CreateLongIntToVoidLocations(arena_, Primitive::kPrimInt, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPokeIntNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimInt, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPokeLongNative(HInvoke* invoke) {
+  CreateLongIntToVoidLocations(arena_, Primitive::kPrimLong, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPokeLongNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimLong, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitMemoryPokeShortNative(HInvoke* invoke) {
+  CreateLongIntToVoidLocations(arena_, Primitive::kPrimShort, invoke);
+}
+
+void IntrinsicCodeGeneratorX86::VisitMemoryPokeShortNative(HInvoke* invoke) {
+  GenPoke(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderX86::VisitThreadCurrentThread(HInvoke* invoke) {
+  LocationSummary* locations = new (arena_) LocationSummary(invoke,
+                                                            LocationSummary::kNoCall,
+                                                            kIntrinsified);
+  locations->SetOut(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorX86::VisitThreadCurrentThread(HInvoke* invoke) {
+  Register out = invoke->GetLocations()->Out().AsRegister<Register>();
+  GetAssembler()->fs()->movl(out, Address::Absolute(Thread::PeerOffset<kX86WordSize>()));
+}
+
+static void GenUnsafeGet(LocationSummary* locations, Primitive::Type type,
+                         bool is_volatile, X86Assembler* assembler) {
+  Register base = locations->InAt(1).AsRegister<Register>();
+  Register offset = locations->InAt(2).AsRegisterPairLow<Register>();
+  Location output = locations->Out();
+
+  switch (type) {
+    case Primitive::kPrimInt:
+    case Primitive::kPrimNot:
+      __ movl(output.AsRegister<Register>(), Address(base, offset, ScaleFactor::TIMES_1, 0));
+      break;
+
+    case Primitive::kPrimLong: {
+        Register output_lo = output.AsRegisterPairLow<Register>();
+        Register output_hi = output.AsRegisterPairHigh<Register>();
+        if (is_volatile) {
+          // Need to use a XMM to read atomically.
+          XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+          __ movsd(temp, Address(base, offset, ScaleFactor::TIMES_1, 0));
+          __ movd(output_lo, temp);
+          __ psrlq(temp, Immediate(32));
+          __ movd(output_hi, temp);
+        } else {
+          __ movl(output_lo, Address(base, offset, ScaleFactor::TIMES_1, 0));
+          __ movl(output_hi, Address(base, offset, ScaleFactor::TIMES_1, 4));
+        }
+      }
+      break;
+
+    default:
+      LOG(FATAL) << "Unsupported op size " << type;
+      UNREACHABLE();
+  }
+}
+
+static void CreateIntIntIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke,
+                                          bool is_long, bool is_volatile) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::NoLocation());        // Unused receiver.
+  locations->SetInAt(1, Location::RequiresRegister());
+  locations->SetInAt(2, Location::RequiresRegister());
+  if (is_long) {
+    if (is_volatile) {
+      // Need to use XMM to read volatile.
+      locations->AddTemp(Location::RequiresFpuRegister());
+      locations->SetOut(Location::RequiresRegister());
+    } else {
+      locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
+    }
+  } else {
+    locations->SetOut(Location::RequiresRegister());
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitUnsafeGet(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, false, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafeGetVolatile(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, false, true);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafeGetLong(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, false, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, true, true);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafeGetObject(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, false, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafeGetObjectVolatile(HInvoke* invoke) {
+  CreateIntIntIntToIntLocations(arena_, invoke, false, true);
+}
+
+
+void IntrinsicCodeGeneratorX86::VisitUnsafeGet(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimInt, false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafeGetVolatile(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimInt, true, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafeGetLong(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimLong, false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimLong, true, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafeGetObject(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimNot, false, GetAssembler());
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafeGetObjectVolatile(HInvoke* invoke) {
+  GenUnsafeGet(invoke->GetLocations(), Primitive::kPrimNot, true, GetAssembler());
+}
+
+
+static void CreateIntIntIntIntToVoidPlusTempsLocations(ArenaAllocator* arena,
+                                                       Primitive::Type type,
+                                                       HInvoke* invoke,
+                                                       bool is_volatile) {
+  LocationSummary* locations = new (arena) LocationSummary(invoke,
+                                                           LocationSummary::kNoCall,
+                                                           kIntrinsified);
+  locations->SetInAt(0, Location::NoLocation());        // Unused receiver.
+  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());
+    // Ensure the value is in a byte register.
+    locations->AddTemp(Location::RegisterLocation(ECX));
+  } else if (type == Primitive::kPrimLong && is_volatile) {
+    locations->AddTemp(Location::RequiresFpuRegister());
+    locations->AddTemp(Location::RequiresFpuRegister());
+  }
+}
+
+void IntrinsicLocationsBuilderX86::VisitUnsafePut(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimInt, invoke, true);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutObject(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimNot, invoke, true);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutLong(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke, false);
+}
+void IntrinsicLocationsBuilderX86::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+  CreateIntIntIntIntToVoidPlusTempsLocations(arena_, Primitive::kPrimLong, invoke, true);
+}
+
+// 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* codegen) {
+  X86Assembler* assembler = reinterpret_cast<X86Assembler*>(codegen->GetAssembler());
+  Register base = locations->InAt(1).AsRegister<Register>();
+  Register offset = locations->InAt(2).AsRegisterPairLow<Register>();
+  Location value_loc = locations->InAt(3);
+
+  if (type == Primitive::kPrimLong) {
+    Register value_lo = value_loc.AsRegisterPairLow<Register>();
+    Register value_hi = value_loc.AsRegisterPairHigh<Register>();
+    if (is_volatile) {
+      XmmRegister temp1 = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+      XmmRegister temp2 = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
+      __ movd(temp1, value_lo);
+      __ movd(temp2, value_hi);
+      __ punpckldq(temp1, temp2);
+      __ movsd(Address(base, offset, ScaleFactor::TIMES_1, 0), temp1);
+    } else {
+      __ movl(Address(base, offset, ScaleFactor::TIMES_1, 0), value_lo);
+      __ movl(Address(base, offset, ScaleFactor::TIMES_1, 4), value_hi);
+    }
+  } else {
+    __ movl(Address(base, offset, ScaleFactor::TIMES_1, 0), value_loc.AsRegister<Register>());
+  }
+
+  if (is_volatile) {
+    __ mfence();
+  }
+
+  if (type == Primitive::kPrimNot) {
+    codegen->MarkGCCard(locations->GetTemp(0).AsRegister<Register>(),
+                        locations->GetTemp(1).AsRegister<Register>(),
+                        base,
+                        value_loc.AsRegister<Register>());
+  }
+}
+
+void IntrinsicCodeGeneratorX86::VisitUnsafePut(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, true, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutObject(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, true, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutLong(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, codegen_);
+}
+void IntrinsicCodeGeneratorX86::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+  GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, true, codegen_);
+}
+
+// Unimplemented intrinsics.
+
+#define UNIMPLEMENTED_INTRINSIC(Name)                                                   \
+void IntrinsicLocationsBuilderX86::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) { \
+}                                                                                       \
+void IntrinsicCodeGeneratorX86::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) {    \
+}
+
+UNIMPLEMENTED_INTRINSIC(IntegerReverse)
+UNIMPLEMENTED_INTRINSIC(LongReverse)
+UNIMPLEMENTED_INTRINSIC(LongReverseBytes)
+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
+}  // namespace art
diff --git a/compiler/optimizing/intrinsics_x86.h b/compiler/optimizing/intrinsics_x86.h
new file mode 100644
index 0000000..e1e8260
--- /dev/null
+++ b/compiler/optimizing/intrinsics_x86.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_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_X86_H_
+
+#include "intrinsics.h"
+
+namespace art {
+
+class ArenaAllocator;
+class HInvokeStaticOrDirect;
+class HInvokeVirtual;
+
+namespace x86 {
+
+class CodeGeneratorX86;
+class X86Assembler;
+
+class IntrinsicLocationsBuilderX86 FINAL : public IntrinsicVisitor {
+ public:
+  explicit IntrinsicLocationsBuilderX86(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);
+};
+
+class IntrinsicCodeGeneratorX86 FINAL : public IntrinsicVisitor {
+ public:
+  explicit IntrinsicCodeGeneratorX86(CodeGeneratorX86* 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:
+  X86Assembler* GetAssembler();
+
+  ArenaAllocator* GetAllocator();
+
+  CodeGeneratorX86* codegen_;
+
+  DISALLOW_COPY_AND_ASSIGN(IntrinsicCodeGeneratorX86);
+};
+
+}  // namespace x86
+}  // namespace art
+
+#endif  // ART_COMPILER_OPTIMIZING_INTRINSICS_X86_H_
diff --git a/compiler/utils/x86/assembler_x86.cc b/compiler/utils/x86/assembler_x86.cc
index 90170ce..5773459 100644
--- a/compiler/utils/x86/assembler_x86.cc
+++ b/compiler/utils/x86/assembler_x86.cc
@@ -146,6 +146,12 @@ void X86Assembler::movl(const Address& dst, Label* lbl) {
   EmitLabel(lbl, dst.length_ + 5);
 }
 
+void X86Assembler::bswapl(Register dst) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x0F);
+  EmitUint8(0xC8 + dst);
+}
+
 void X86Assembler::movzxb(Register dst, ByteRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x0F);
@@ -725,6 +731,32 @@ void X86Assembler::xorpd(XmmRegister dst, XmmRegister src) {
 }
 
 
+void X86Assembler::andps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::andpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::orpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
 void X86Assembler::xorps(XmmRegister dst, const Address& src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x0F);
@@ -733,6 +765,14 @@ void X86Assembler::xorps(XmmRegister dst, const Address& src) {
 }
 
 
+void X86Assembler::orps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
 void X86Assembler::xorps(XmmRegister dst, XmmRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x0F);
@@ -741,6 +781,14 @@ void X86Assembler::xorps(XmmRegister dst, XmmRegister src) {
 }
 
 
+void X86Assembler::andps(XmmRegister dst, const Address& src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitOperand(dst, src);
+}
+
+
 void X86Assembler::andpd(XmmRegister dst, const Address& src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
@@ -1090,6 +1138,13 @@ void X86Assembler::subl(Register reg, const Address& address) {
 }
 
 
+void X86Assembler::subl(const Address& address, Register reg) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x29);
+  EmitOperand(reg, address);
+}
+
+
 void X86Assembler::cdq() {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x99);
@@ -1175,6 +1230,13 @@ void X86Assembler::sbbl(Register dst, const Address& address) {
 }
 
 
+void X86Assembler::sbbl(const Address& address, Register src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x19);
+  EmitOperand(src, address);
+}
+
+
 void X86Assembler::incl(Register reg) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x40 + reg);
diff --git a/compiler/utils/x86/assembler_x86.h b/compiler/utils/x86/assembler_x86.h
index 4d20db0..6ccf2e3 100644
--- a/compiler/utils/x86/assembler_x86.h
+++ b/compiler/utils/x86/assembler_x86.h
@@ -231,6 +231,8 @@ class X86Assembler FINAL : public Assembler {
   void movl(const Address& dst, const Immediate& imm);
   void movl(const Address& dst, Label* lbl);
 
+  void bswapl(Register dst);
+
   void movzxb(Register dst, ByteRegister src);
   void movzxb(Register dst, const Address& src);
   void movsxb(Register dst, ByteRegister src);
@@ -318,7 +320,13 @@ class X86Assembler FINAL : public Assembler {
   void xorps(XmmRegister dst, const Address& src);
   void xorps(XmmRegister dst, XmmRegister src);
 
+  void andpd(XmmRegister dst, XmmRegister src);
   void andpd(XmmRegister dst, const Address& src);
+  void andps(XmmRegister dst, XmmRegister src);
+  void andps(XmmRegister dst, const Address& src);
+
+  void orpd(XmmRegister dst, XmmRegister src);
+  void orps(XmmRegister dst, XmmRegister src);
 
   void flds(const Address& src);
   void fstps(const Address& dst);
@@ -389,6 +397,7 @@ class X86Assembler FINAL : public Assembler {
   void subl(Register dst, Register src);
   void subl(Register reg, const Immediate& imm);
   void subl(Register reg, const Address& address);
+  void subl(const Address& address, Register src);
 
   void cdq();
 
@@ -407,6 +416,7 @@ class X86Assembler FINAL : public Assembler {
   void sbbl(Register dst, Register src);
   void sbbl(Register reg, const Immediate& imm);
   void sbbl(Register reg, const Address& address);
+  void sbbl(const Address& address, Register src);
 
   void incl(Register reg);
   void incl(const Address& address);