ARM64 Baker's read barrier fast path implementation.

Introduce an ARM64 fast path implementation in Optimizing
for Baker's read barriers (for both heap reference loads and
GC root loads).  The marking phase of the read barrier is
performed by a slow path, invoking the runtime entry point
artReadBarrierMark.

Other read barrier algorithms continue to use the original
slow path based implementation, which has been renamed as
GenerateReadBarrierSlow/GenerateReadBarrierForRootSlow.

Bug: 12687968
Bug: 26601270
Change-Id: I60da15249b58a8ee1a065ed9be2c4e438ee17150
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 2cb2741..399718d 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -584,6 +584,56 @@
   }
 }
 
+// Slow path marking an object during a read barrier.
+class ReadBarrierMarkSlowPathARM64 : public SlowPathCodeARM64 {
+ public:
+  ReadBarrierMarkSlowPathARM64(HInstruction* instruction, Location out, Location obj)
+      : instruction_(instruction), out_(out), obj_(obj) {
+    DCHECK(kEmitCompilerReadBarrier);
+  }
+
+  const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM64"; }
+
+  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+    LocationSummary* locations = instruction_->GetLocations();
+    Primitive::Type type = Primitive::kPrimNot;
+    DCHECK(locations->CanCall());
+    DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg()));
+    DCHECK(instruction_->IsInstanceFieldGet() ||
+           instruction_->IsStaticFieldGet() ||
+           instruction_->IsArrayGet() ||
+           instruction_->IsLoadClass() ||
+           instruction_->IsLoadString() ||
+           instruction_->IsInstanceOf() ||
+           instruction_->IsCheckCast())
+        << "Unexpected instruction in read barrier marking slow path: "
+        << instruction_->DebugName();
+
+    __ Bind(GetEntryLabel());
+    SaveLiveRegisters(codegen, locations);
+
+    InvokeRuntimeCallingConvention calling_convention;
+    CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
+    arm64_codegen->MoveLocation(LocationFrom(calling_convention.GetRegisterAt(0)), obj_, type);
+    arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark),
+                                 instruction_,
+                                 instruction_->GetDexPc(),
+                                 this);
+    CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>();
+    arm64_codegen->MoveLocation(out_, calling_convention.GetReturnLocation(type), type);
+
+    RestoreLiveRegisters(codegen, locations);
+    __ B(GetExitLabel());
+  }
+
+ private:
+  HInstruction* const instruction_;
+  const Location out_;
+  const Location obj_;
+
+  DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM64);
+};
+
 // Slow path generating a read barrier for a heap reference.
 class ReadBarrierForHeapReferenceSlowPathARM64 : public SlowPathCodeARM64 {
  public:
@@ -605,7 +655,7 @@
     // to be instrumented, e.g.:
     //
     //   __ Ldr(out, HeapOperand(out, class_offset);
-    //   codegen_->GenerateReadBarrier(instruction, out_loc, out_loc, out_loc, offset);
+    //   codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset);
     //
     // In that case, we have lost the information about the original
     // object, and the emitted read barrier cannot work properly.
@@ -621,7 +671,9 @@
     DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg()));
     DCHECK(!instruction_->IsInvoke() ||
            (instruction_->IsInvokeStaticOrDirect() &&
-            instruction_->GetLocations()->Intrinsified()));
+            instruction_->GetLocations()->Intrinsified()))
+        << "Unexpected instruction in read barrier for heap reference slow path: "
+        << instruction_->DebugName();
     // The read barrier instrumentation does not support the
     // HArm64IntermediateAddress instruction yet.
     DCHECK(!(instruction_->IsArrayGet() &&
@@ -769,14 +821,18 @@
 class ReadBarrierForRootSlowPathARM64 : public SlowPathCodeARM64 {
  public:
   ReadBarrierForRootSlowPathARM64(HInstruction* instruction, Location out, Location root)
-      : instruction_(instruction), out_(out), root_(root) {}
+      : instruction_(instruction), out_(out), root_(root) {
+    DCHECK(kEmitCompilerReadBarrier);
+  }
 
   void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
     LocationSummary* locations = instruction_->GetLocations();
     Primitive::Type type = Primitive::kPrimNot;
     DCHECK(locations->CanCall());
     DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg()));
-    DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString());
+    DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString())
+        << "Unexpected instruction in read barrier for GC root slow path: "
+        << instruction_->DebugName();
 
     __ Bind(GetEntryLabel());
     SaveLiveRegisters(codegen, locations);
@@ -1338,7 +1394,8 @@
 
 void CodeGeneratorARM64::LoadAcquire(HInstruction* instruction,
                                      CPURegister dst,
-                                     const MemOperand& src) {
+                                     const MemOperand& src,
+                                     bool needs_null_check) {
   MacroAssembler* masm = GetVIXLAssembler();
   BlockPoolsScope block_pools(masm);
   UseScratchRegisterScope temps(masm);
@@ -1354,20 +1411,28 @@
   switch (type) {
     case Primitive::kPrimBoolean:
       __ Ldarb(Register(dst), base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       break;
     case Primitive::kPrimByte:
       __ Ldarb(Register(dst), base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       __ Sbfx(Register(dst), Register(dst), 0, Primitive::ComponentSize(type) * kBitsPerByte);
       break;
     case Primitive::kPrimChar:
       __ Ldarh(Register(dst), base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       break;
     case Primitive::kPrimShort:
       __ Ldarh(Register(dst), base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       __ Sbfx(Register(dst), Register(dst), 0, Primitive::ComponentSize(type) * kBitsPerByte);
       break;
     case Primitive::kPrimInt:
@@ -1375,7 +1440,9 @@
     case Primitive::kPrimLong:
       DCHECK_EQ(dst.Is64Bits(), Primitive::Is64BitType(type));
       __ Ldar(Register(dst), base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       break;
     case Primitive::kPrimFloat:
     case Primitive::kPrimDouble: {
@@ -1384,7 +1451,9 @@
 
       Register temp = dst.Is64Bits() ? temps.AcquireX() : temps.AcquireW();
       __ Ldar(temp, base);
-      MaybeRecordImplicitNullCheck(instruction);
+      if (needs_null_check) {
+        MaybeRecordImplicitNullCheck(instruction);
+      }
       __ Fmov(FPRegister(dst), temp);
       break;
     }
@@ -1505,7 +1574,7 @@
   __ Bind(slow_path->GetExitLabel());
 }
 
-void InstructionCodeGeneratorARM64::GenerateMemoryBarrier(MemBarrierKind kind) {
+void CodeGeneratorARM64::GenerateMemoryBarrier(MemBarrierKind kind) {
   BarrierType type = BarrierAll;
 
   switch (kind) {
@@ -1641,33 +1710,62 @@
 void InstructionCodeGeneratorARM64::HandleFieldGet(HInstruction* instruction,
                                                    const FieldInfo& field_info) {
   DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
+  LocationSummary* locations = instruction->GetLocations();
+  Location base_loc = locations->InAt(0);
+  Location out = locations->Out();
+  uint32_t offset = field_info.GetFieldOffset().Uint32Value();
   Primitive::Type field_type = field_info.GetFieldType();
   BlockPoolsScope block_pools(GetVIXLAssembler());
 
   MemOperand field = HeapOperand(InputRegisterAt(instruction, 0), field_info.GetFieldOffset());
   bool use_acquire_release = codegen_->GetInstructionSetFeatures().PreferAcquireRelease();
 
-  if (field_info.IsVolatile()) {
-    if (use_acquire_release) {
-      // NB: LoadAcquire will record the pc info if needed.
-      codegen_->LoadAcquire(instruction, OutputCPURegister(instruction), field);
+  if (field_type == Primitive::kPrimNot && kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
+    // Object FieldGet with Baker's read barrier case.
+    MacroAssembler* masm = GetVIXLAssembler();
+    UseScratchRegisterScope temps(masm);
+    // /* HeapReference<Object> */ out = *(base + offset)
+    Register base = RegisterFrom(base_loc, Primitive::kPrimNot);
+    Register temp = temps.AcquireW();
+    // Note that potential implicit null checks are handled in this
+    // CodeGeneratorARM64::GenerateFieldLoadWithBakerReadBarrier call.
+    codegen_->GenerateFieldLoadWithBakerReadBarrier(
+        instruction,
+        out,
+        base,
+        offset,
+        temp,
+        /* needs_null_check */ true,
+        field_info.IsVolatile() && use_acquire_release);
+    if (field_info.IsVolatile() && !use_acquire_release) {
+      // For IRIW sequential consistency kLoadAny is not sufficient.
+      codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
+    }
+  } else {
+    // General case.
+    if (field_info.IsVolatile()) {
+      if (use_acquire_release) {
+        // Note that a potential implicit null check is handled in this
+        // CodeGeneratorARM64::LoadAcquire call.
+        // NB: LoadAcquire will record the pc info if needed.
+        codegen_->LoadAcquire(
+            instruction, OutputCPURegister(instruction), field, /* needs_null_check */ true);
+      } else {
+        codegen_->Load(field_type, OutputCPURegister(instruction), field);
+        codegen_->MaybeRecordImplicitNullCheck(instruction);
+        // For IRIW sequential consistency kLoadAny is not sufficient.
+        codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
+      }
     } else {
       codegen_->Load(field_type, OutputCPURegister(instruction), field);
       codegen_->MaybeRecordImplicitNullCheck(instruction);
-      // For IRIW sequential consistency kLoadAny is not sufficient.
-      GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
     }
-  } else {
-    codegen_->Load(field_type, OutputCPURegister(instruction), field);
-    codegen_->MaybeRecordImplicitNullCheck(instruction);
-  }
-
-  if (field_type == Primitive::kPrimNot) {
-    LocationSummary* locations = instruction->GetLocations();
-    Location base = locations->InAt(0);
-    Location out = locations->Out();
-    uint32_t offset = field_info.GetFieldOffset().Uint32Value();
-    codegen_->MaybeGenerateReadBarrier(instruction, out, out, base, offset);
+    if (field_type == Primitive::kPrimNot) {
+      // If read barriers are enabled, emit read barriers other than
+      // Baker's using a slow path (and also unpoison the loaded
+      // reference, if heap poisoning is enabled).
+      codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset);
+    }
   }
 }
 
@@ -1713,10 +1811,10 @@
         codegen_->StoreRelease(field_type, source, HeapOperand(obj, offset));
         codegen_->MaybeRecordImplicitNullCheck(instruction);
       } else {
-        GenerateMemoryBarrier(MemBarrierKind::kAnyStore);
+        codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore);
         codegen_->Store(field_type, source, HeapOperand(obj, offset));
         codegen_->MaybeRecordImplicitNullCheck(instruction);
-        GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
+        codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
       }
     } else {
       codegen_->Store(field_type, source, HeapOperand(obj, offset));
@@ -2021,50 +2119,62 @@
   LocationSummary* locations = instruction->GetLocations();
   Location index = locations->InAt(1);
   uint32_t offset = mirror::Array::DataOffset(Primitive::ComponentSize(type)).Uint32Value();
-  MemOperand source = HeapOperand(obj);
-  CPURegister dest = OutputCPURegister(instruction);
+  Location out = locations->Out();
 
   MacroAssembler* masm = GetVIXLAssembler();
   UseScratchRegisterScope temps(masm);
   // Block pools between `Load` and `MaybeRecordImplicitNullCheck`.
   BlockPoolsScope block_pools(masm);
 
-  if (index.IsConstant()) {
-    offset += Int64ConstantFrom(index) << Primitive::ComponentSizeShift(type);
-    source = HeapOperand(obj, offset);
+  if (type == Primitive::kPrimNot && kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
+    // Object ArrayGet with Baker's read barrier case.
+    Register temp = temps.AcquireW();
+    // The read barrier instrumentation does not support the
+    // HArm64IntermediateAddress instruction yet.
+    DCHECK(!instruction->GetArray()->IsArm64IntermediateAddress());
+    // Note that a potential implicit null check is handled in the
+    // CodeGeneratorARM64::GenerateArrayLoadWithBakerReadBarrier call.
+    codegen_->GenerateArrayLoadWithBakerReadBarrier(
+        instruction, out, obj.W(), offset, index, temp, /* needs_null_check */ true);
   } else {
-    Register temp = temps.AcquireSameSizeAs(obj);
-    if (instruction->GetArray()->IsArm64IntermediateAddress()) {
-      // The read barrier instrumentation does not support the
-      // HArm64IntermediateAddress instruction yet.
-      DCHECK(!kEmitCompilerReadBarrier);
-      // We do not need to compute the intermediate address from the array: the
-      // input instruction has done it already. See the comment in
-      // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`.
-      if (kIsDebugBuild) {
-        HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress();
-        DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == offset);
-      }
-      temp = obj;
-    } else {
-      __ Add(temp, obj, offset);
-    }
-    source = HeapOperand(temp, XRegisterFrom(index), LSL, Primitive::ComponentSizeShift(type));
-  }
-
-  codegen_->Load(type, dest, source);
-  codegen_->MaybeRecordImplicitNullCheck(instruction);
-
-  if (type == Primitive::kPrimNot) {
-    static_assert(
-        sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
-        "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
-    Location obj_loc = locations->InAt(0);
-    Location out = locations->Out();
+    // General case.
+    MemOperand source = HeapOperand(obj);
     if (index.IsConstant()) {
-      codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset);
+      offset += Int64ConstantFrom(index) << Primitive::ComponentSizeShift(type);
+      source = HeapOperand(obj, offset);
     } else {
-      codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset, index);
+      Register temp = temps.AcquireSameSizeAs(obj);
+      if (instruction->GetArray()->IsArm64IntermediateAddress()) {
+        // The read barrier instrumentation does not support the
+        // HArm64IntermediateAddress instruction yet.
+        DCHECK(!kEmitCompilerReadBarrier);
+        // We do not need to compute the intermediate address from the array: the
+        // input instruction has done it already. See the comment in
+        // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`.
+        if (kIsDebugBuild) {
+          HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress();
+          DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), offset);
+        }
+        temp = obj;
+      } else {
+        __ Add(temp, obj, offset);
+      }
+      source = HeapOperand(temp, XRegisterFrom(index), LSL, Primitive::ComponentSizeShift(type));
+    }
+
+    codegen_->Load(type, OutputCPURegister(instruction), source);
+    codegen_->MaybeRecordImplicitNullCheck(instruction);
+
+    if (type == Primitive::kPrimNot) {
+      static_assert(
+          sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
+          "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
+      Location obj_loc = locations->InAt(0);
+      if (index.IsConstant()) {
+        codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, obj_loc, offset);
+      } else {
+        codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, obj_loc, offset, index);
+      }
     }
   }
 }
@@ -2194,12 +2304,12 @@
           //   __ Mov(temp2, temp);
           //   // /* HeapReference<Class> */ temp = temp->component_type_
           //   __ Ldr(temp, HeapOperand(temp, component_offset));
-          //   codegen_->GenerateReadBarrier(
+          //   codegen_->GenerateReadBarrierSlow(
           //       instruction, temp_loc, temp_loc, temp2_loc, component_offset);
           //
           //   // /* HeapReference<Class> */ temp2 = value->klass_
           //   __ Ldr(temp2, HeapOperand(Register(value), class_offset));
-          //   codegen_->GenerateReadBarrier(
+          //   codegen_->GenerateReadBarrierSlow(
           //       instruction, temp2_loc, temp2_loc, value_loc, class_offset, temp_loc);
           //
           //   __ Cmp(temp, temp2);
@@ -2921,6 +3031,14 @@
   HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull());
 }
 
+static bool TypeCheckNeedsATemporary(TypeCheckKind type_check_kind) {
+  return kEmitCompilerReadBarrier &&
+      (kUseBakerReadBarrier ||
+       type_check_kind == TypeCheckKind::kAbstractClassCheck ||
+       type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
+       type_check_kind == TypeCheckKind::kArrayObjectCheck);
+}
+
 void LocationsBuilderARM64::VisitInstanceOf(HInstanceOf* instruction) {
   LocationSummary::CallKind call_kind = LocationSummary::kNoCall;
   TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
@@ -2947,21 +3065,22 @@
   locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
   // When read barriers are enabled, we need a temporary register for
   // some cases.
-  if (kEmitCompilerReadBarrier &&
-      (type_check_kind == TypeCheckKind::kAbstractClassCheck ||
-       type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
-       type_check_kind == TypeCheckKind::kArrayObjectCheck)) {
+  if (TypeCheckNeedsATemporary(type_check_kind)) {
     locations->AddTemp(Location::RequiresRegister());
   }
 }
 
 void InstructionCodeGeneratorARM64::VisitInstanceOf(HInstanceOf* instruction) {
+  TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
   LocationSummary* locations = instruction->GetLocations();
   Location obj_loc = locations->InAt(0);
   Register obj = InputRegisterAt(instruction, 0);
   Register cls = InputRegisterAt(instruction, 1);
   Location out_loc = locations->Out();
   Register out = OutputRegister(instruction);
+  Location temp_loc = TypeCheckNeedsATemporary(type_check_kind) ?
+      locations->GetTemp(0) :
+      Location::NoLocation();
   uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
   uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
   uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
@@ -2977,10 +3096,9 @@
   }
 
   // /* HeapReference<Class> */ out = obj->klass_
-  __ Ldr(out, HeapOperand(obj.W(), class_offset));
-  codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, obj_loc, class_offset);
+  GenerateReferenceLoadTwoRegisters(instruction, out_loc, obj_loc, class_offset, temp_loc);
 
-  switch (instruction->GetTypeCheckKind()) {
+  switch (type_check_kind) {
     case TypeCheckKind::kExactCheck: {
       __ Cmp(out, cls);
       __ Cset(out, eq);
@@ -2995,17 +3113,8 @@
       // object to avoid doing a comparison we know will fail.
       vixl::Label loop, success;
       __ Bind(&loop);
-      Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `out` into `temp` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp = WRegisterFrom(temp_loc);
-        __ Mov(temp, out);
-      }
       // /* HeapReference<Class> */ out = out->super_class_
-      __ Ldr(out, HeapOperand(out, super_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset);
+      GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, temp_loc);
       // If `out` is null, we use it for the result, and jump to `done`.
       __ Cbz(out, &done);
       __ Cmp(out, cls);
@@ -3023,17 +3132,8 @@
       __ Bind(&loop);
       __ Cmp(out, cls);
       __ B(eq, &success);
-      Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `out` into `temp` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp = WRegisterFrom(temp_loc);
-        __ Mov(temp, out);
-      }
       // /* HeapReference<Class> */ out = out->super_class_
-      __ Ldr(out, HeapOperand(out, super_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset);
+      GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, temp_loc);
       __ Cbnz(out, &loop);
       // If `out` is null, we use it for the result, and jump to `done`.
       __ B(&done);
@@ -3051,17 +3151,8 @@
       __ Cmp(out, cls);
       __ B(eq, &exact_check);
       // Otherwise, we need to check that the object's class is a non-primitive array.
-      Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `out` into `temp` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp = WRegisterFrom(temp_loc);
-        __ Mov(temp, out);
-      }
       // /* HeapReference<Class> */ out = out->component_type_
-      __ Ldr(out, HeapOperand(out, component_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, component_offset);
+      GenerateReferenceLoadOneRegister(instruction, out_loc, component_offset, temp_loc);
       // If `out` is null, we use it for the result, and jump to `done`.
       __ Cbz(out, &done);
       __ Ldrh(out, HeapOperand(out, primitive_offset));
@@ -3100,6 +3191,13 @@
       // HInstanceOf instruction (following the runtime calling
       // convention), which might be cluttered by the potential first
       // read barrier emission at the beginning of this method.
+      //
+      // TODO: Introduce a new runtime entry point taking the object
+      // to test (instead of its class) as argument, and let it deal
+      // with the read barrier issues. This will let us refactor this
+      // case of the `switch` code as it was previously (with a direct
+      // call to the runtime not using a type checking slow path).
+      // This should also be beneficial for the other cases above.
       DCHECK(locations->OnlyCallsOnSlowPath());
       slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM64(instruction,
                                                                       /* is_fatal */ false);
@@ -3152,30 +3250,29 @@
   locations->SetInAt(1, Location::RequiresRegister());
   // Note that TypeCheckSlowPathARM64 uses this "temp" register too.
   locations->AddTemp(Location::RequiresRegister());
-  locations->AddTemp(Location::RequiresRegister());
   // When read barriers are enabled, we need an additional temporary
   // register for some cases.
-  if (kEmitCompilerReadBarrier &&
-      (type_check_kind == TypeCheckKind::kAbstractClassCheck ||
-       type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
-       type_check_kind == TypeCheckKind::kArrayObjectCheck)) {
-     locations->AddTemp(Location::RequiresRegister());
+  if (TypeCheckNeedsATemporary(type_check_kind)) {
+    locations->AddTemp(Location::RequiresRegister());
   }
 }
 
 void InstructionCodeGeneratorARM64::VisitCheckCast(HCheckCast* instruction) {
+  TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
   LocationSummary* locations = instruction->GetLocations();
   Location obj_loc = locations->InAt(0);
   Register obj = InputRegisterAt(instruction, 0);
   Register cls = InputRegisterAt(instruction, 1);
   Location temp_loc = locations->GetTemp(0);
+  Location temp2_loc = TypeCheckNeedsATemporary(type_check_kind) ?
+      locations->GetTemp(1) :
+      Location::NoLocation();
   Register temp = WRegisterFrom(temp_loc);
   uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
   uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
   uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
   uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value();
 
-  TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
   bool is_type_check_slow_path_fatal =
       (type_check_kind == TypeCheckKind::kExactCheck ||
        type_check_kind == TypeCheckKind::kAbstractClassCheck ||
@@ -3194,8 +3291,7 @@
   }
 
   // /* HeapReference<Class> */ temp = obj->klass_
-  __ Ldr(temp, HeapOperand(obj, class_offset));
-  codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+  GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, temp2_loc);
 
   switch (type_check_kind) {
     case TypeCheckKind::kExactCheck:
@@ -3212,18 +3308,8 @@
       // object to avoid doing a comparison we know will fail.
       vixl::Label loop, compare_classes;
       __ Bind(&loop);
-      Location temp2_loc =
-          kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `temp` into `temp2` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp2 = WRegisterFrom(temp2_loc);
-        __ Mov(temp2, temp);
-      }
       // /* HeapReference<Class> */ temp = temp->super_class_
-      __ Ldr(temp, HeapOperand(temp, super_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset);
+      GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, temp2_loc);
 
       // If the class reference currently in `temp` is not null, jump
       // to the `compare_classes` label to compare it with the checked
@@ -3235,8 +3321,7 @@
       // going into the slow path, as it has been overwritten in the
       // meantime.
       // /* HeapReference<Class> */ temp = obj->klass_
-      __ Ldr(temp, HeapOperand(obj, class_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+      GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, temp2_loc);
       __ B(type_check_slow_path->GetEntryLabel());
 
       __ Bind(&compare_classes);
@@ -3252,18 +3337,8 @@
       __ Cmp(temp, cls);
       __ B(eq, &done);
 
-      Location temp2_loc =
-          kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `temp` into `temp2` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp2 = WRegisterFrom(temp2_loc);
-        __ Mov(temp2, temp);
-      }
       // /* HeapReference<Class> */ temp = temp->super_class_
-      __ Ldr(temp, HeapOperand(temp, super_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset);
+      GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, temp2_loc);
 
       // If the class reference currently in `temp` is not null, jump
       // back at the beginning of the loop.
@@ -3274,8 +3349,7 @@
       // going into the slow path, as it has been overwritten in the
       // meantime.
       // /* HeapReference<Class> */ temp = obj->klass_
-      __ Ldr(temp, HeapOperand(obj, class_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+      GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, temp2_loc);
       __ B(type_check_slow_path->GetEntryLabel());
       break;
     }
@@ -3287,19 +3361,8 @@
       __ B(eq, &done);
 
       // Otherwise, we need to check that the object's class is a non-primitive array.
-      Location temp2_loc =
-          kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
-      if (kEmitCompilerReadBarrier) {
-        // Save the value of `temp` into `temp2` before overwriting it
-        // in the following move operation, as we will need it for the
-        // read barrier below.
-        Register temp2 = WRegisterFrom(temp2_loc);
-        __ Mov(temp2, temp);
-      }
       // /* HeapReference<Class> */ temp = temp->component_type_
-      __ Ldr(temp, HeapOperand(temp, component_offset));
-      codegen_->MaybeGenerateReadBarrier(
-          instruction, temp_loc, temp_loc, temp2_loc, component_offset);
+      GenerateReferenceLoadOneRegister(instruction, temp_loc, component_offset, temp2_loc);
 
       // If the component type is not null (i.e. the object is indeed
       // an array), jump to label `check_non_primitive_component_type`
@@ -3312,8 +3375,7 @@
       // going into the slow path, as it has been overwritten in the
       // meantime.
       // /* HeapReference<Class> */ temp = obj->klass_
-      __ Ldr(temp, HeapOperand(obj, class_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+      GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, temp2_loc);
       __ B(type_check_slow_path->GetEntryLabel());
 
       __ Bind(&check_non_primitive_component_type);
@@ -3322,8 +3384,7 @@
       __ Cbz(temp, &done);
       // Same comment as above regarding `temp` and the slow path.
       // /* HeapReference<Class> */ temp = obj->klass_
-      __ Ldr(temp, HeapOperand(obj, class_offset));
-      codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+      GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, temp2_loc);
       __ B(type_check_slow_path->GetEntryLabel());
       break;
     }
@@ -3340,6 +3401,13 @@
       // instruction (following the runtime calling convention), which
       // might be cluttered by the potential first read barrier
       // emission at the beginning of this method.
+      //
+      // TODO: Introduce a new runtime entry point taking the object
+      // to test (instead of its class) as argument, and let it deal
+      // with the read barrier issues. This will let us refactor this
+      // case of the `switch` code as it was previously (with a direct
+      // call to the runtime not using a type checking slow path).
+      // This should also be beneficial for the other cases above.
       __ B(type_check_slow_path->GetEntryLabel());
       break;
   }
@@ -3465,7 +3533,7 @@
 HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM64::GetSupportedInvokeStaticOrDirectDispatch(
       const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
       MethodReference target_method ATTRIBUTE_UNUSED) {
-  // On arm64 we support all dispatch types.
+  // On ARM64 we support all dispatch types.
   return desired_dispatch_info;
 }
 
@@ -3742,32 +3810,17 @@
   if (cls->IsReferrersClass()) {
     DCHECK(!cls->CanCallRuntime());
     DCHECK(!cls->MustGenerateClinitCheck());
-    uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value();
-    if (kEmitCompilerReadBarrier) {
-      // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_)
-      __ Add(out.X(), current_method.X(), declaring_class_offset);
-      // /* mirror::Class* */ out = out->Read()
-      codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc);
-    } else {
-      // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
-      __ Ldr(out, MemOperand(current_method, declaring_class_offset));
-    }
+    // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
+    GenerateGcRootFieldLoad(
+        cls, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value());
   } else {
     MemberOffset resolved_types_offset = ArtMethod::DexCacheResolvedTypesOffset(kArm64PointerSize);
     // /* GcRoot<mirror::Class>[] */ out =
     //        current_method.ptr_sized_fields_->dex_cache_resolved_types_
     __ Ldr(out.X(), MemOperand(current_method, resolved_types_offset.Int32Value()));
-
-    size_t cache_offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex());
-    if (kEmitCompilerReadBarrier) {
-      // /* GcRoot<mirror::Class>* */ out = &out[type_index]
-      __ Add(out.X(), out.X(), cache_offset);
-      // /* mirror::Class* */ out = out->Read()
-      codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc);
-    } else {
-      // /* GcRoot<mirror::Class> */ out = out[type_index]
-      __ Ldr(out, MemOperand(out.X(), cache_offset));
-    }
+    // /* GcRoot<mirror::Class> */ out = out[type_index]
+    GenerateGcRootFieldLoad(
+        cls, out_loc, out.X(), CodeGenerator::GetCacheOffset(cls->GetTypeIndex()));
 
     if (!cls->IsInDexCache() || cls->MustGenerateClinitCheck()) {
       DCHECK(cls->CanCallRuntime());
@@ -3830,30 +3883,14 @@
   Register out = OutputRegister(load);
   Register current_method = InputRegisterAt(load, 0);
 
-  uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value();
-  if (kEmitCompilerReadBarrier) {
-    // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_)
-    __ Add(out.X(), current_method.X(), declaring_class_offset);
-    // /* mirror::Class* */ out = out->Read()
-    codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc);
-  } else {
-    // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
-    __ Ldr(out, MemOperand(current_method, declaring_class_offset));
-  }
-
+  // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
+  GenerateGcRootFieldLoad(
+      load, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value());
   // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_
   __ Ldr(out.X(), HeapOperand(out, mirror::Class::DexCacheStringsOffset().Uint32Value()));
-
-  size_t cache_offset = CodeGenerator::GetCacheOffset(load->GetStringIndex());
-  if (kEmitCompilerReadBarrier) {
-    // /* GcRoot<mirror::String>* */ out = &out[string_index]
-    __ Add(out.X(), out.X(), cache_offset);
-    // /* mirror::String* */ out = out->Read()
-    codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc);
-  } else {
-    // /* GcRoot<mirror::String> */ out = out[string_index]
-    __ Ldr(out, MemOperand(out.X(), cache_offset));
-  }
+  // /* GcRoot<mirror::String> */ out = out[string_index]
+  GenerateGcRootFieldLoad(
+      load, out_loc, out.X(), CodeGenerator::GetCacheOffset(load->GetStringIndex()));
 
   if (!load->IsInDexCache()) {
     SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM64(load);
@@ -4222,7 +4259,7 @@
 }
 
 void InstructionCodeGeneratorARM64::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) {
-  GenerateMemoryBarrier(memory_barrier->GetBarrierKind());
+  codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind());
 }
 
 void LocationsBuilderARM64::VisitReturn(HReturn* instruction) {
@@ -4607,14 +4644,288 @@
   }
 }
 
-void CodeGeneratorARM64::GenerateReadBarrier(HInstruction* instruction,
-                                             Location out,
-                                             Location ref,
-                                             Location obj,
-                                             uint32_t offset,
-                                             Location index) {
+void InstructionCodeGeneratorARM64::GenerateReferenceLoadOneRegister(HInstruction* instruction,
+                                                                     Location out,
+                                                                     uint32_t offset,
+                                                                     Location temp) {
+  Primitive::Type type = Primitive::kPrimNot;
+  Register out_reg = RegisterFrom(out, type);
+  Register temp_reg = RegisterFrom(temp, type);
+  if (kEmitCompilerReadBarrier) {
+    if (kUseBakerReadBarrier) {
+      // Load with fast path based Baker's read barrier.
+      // /* HeapReference<Object> */ out = *(out + offset)
+      codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction,
+                                                      out,
+                                                      out_reg,
+                                                      offset,
+                                                      temp_reg,
+                                                      /* needs_null_check */ false,
+                                                      /* use_load_acquire */ false);
+    } else {
+      // Load with slow path based read barrier.
+      // Save the value of `out` into `temp` before overwriting it
+      // in the following move operation, as we will need it for the
+      // read barrier below.
+      __ Mov(temp_reg, out_reg);
+      // /* HeapReference<Object> */ out = *(out + offset)
+      __ Ldr(out_reg, HeapOperand(out_reg, offset));
+      codegen_->GenerateReadBarrierSlow(instruction, out, out, temp, offset);
+    }
+  } else {
+    // Plain load with no read barrier.
+    // /* HeapReference<Object> */ out = *(out + offset)
+    __ Ldr(out_reg, HeapOperand(out_reg, offset));
+    GetAssembler()->MaybeUnpoisonHeapReference(out_reg);
+  }
+}
+
+void InstructionCodeGeneratorARM64::GenerateReferenceLoadTwoRegisters(HInstruction* instruction,
+                                                                      Location out,
+                                                                      Location obj,
+                                                                      uint32_t offset,
+                                                                      Location temp) {
+  Primitive::Type type = Primitive::kPrimNot;
+  Register out_reg = RegisterFrom(out, type);
+  Register obj_reg = RegisterFrom(obj, type);
+  if (kEmitCompilerReadBarrier) {
+    if (kUseBakerReadBarrier) {
+      // Load with fast path based Baker's read barrier.
+      Register temp_reg = RegisterFrom(temp, type);
+      // /* HeapReference<Object> */ out = *(obj + offset)
+      codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction,
+                                                      out,
+                                                      obj_reg,
+                                                      offset,
+                                                      temp_reg,
+                                                      /* needs_null_check */ false,
+                                                      /* use_load_acquire */ false);
+    } else {
+      // Load with slow path based read barrier.
+      // /* HeapReference<Object> */ out = *(obj + offset)
+      __ Ldr(out_reg, HeapOperand(obj_reg, offset));
+      codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset);
+    }
+  } else {
+    // Plain load with no read barrier.
+    // /* HeapReference<Object> */ out = *(obj + offset)
+    __ Ldr(out_reg, HeapOperand(obj_reg, offset));
+    GetAssembler()->MaybeUnpoisonHeapReference(out_reg);
+  }
+}
+
+void InstructionCodeGeneratorARM64::GenerateGcRootFieldLoad(HInstruction* instruction,
+                                                            Location root,
+                                                            vixl::Register obj,
+                                                            uint32_t offset) {
+  Register root_reg = RegisterFrom(root, Primitive::kPrimNot);
+  if (kEmitCompilerReadBarrier) {
+    if (kUseBakerReadBarrier) {
+      // Fast path implementation of art::ReadBarrier::BarrierForRoot when
+      // Baker's read barrier are used:
+      //
+      //   root = obj.field;
+      //   if (Thread::Current()->GetIsGcMarking()) {
+      //     root = ReadBarrier::Mark(root)
+      //   }
+
+      // /* GcRoot<mirror::Object> */ root = *(obj + offset)
+      __ Ldr(root_reg, MemOperand(obj, offset));
+      static_assert(
+          sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>),
+          "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> "
+          "have different sizes.");
+      static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t),
+                    "art::mirror::CompressedReference<mirror::Object> and int32_t "
+                    "have different sizes.");
+
+      // Slow path used to mark the GC root `root`.
+      SlowPathCodeARM64* slow_path =
+          new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, root, root);
+      codegen_->AddSlowPath(slow_path);
+
+      MacroAssembler* masm = GetVIXLAssembler();
+      UseScratchRegisterScope temps(masm);
+      Register temp = temps.AcquireW();
+      // temp = Thread::Current()->GetIsGcMarking()
+      __ Ldr(temp, MemOperand(tr, Thread::IsGcMarkingOffset<kArm64WordSize>().Int32Value()));
+      __ Cbnz(temp, slow_path->GetEntryLabel());
+      __ Bind(slow_path->GetExitLabel());
+    } else {
+      // GC root loaded through a slow path for read barriers other
+      // than Baker's.
+      // /* GcRoot<mirror::Object>* */ root = obj + offset
+      __ Add(root_reg.X(), obj.X(), offset);
+      // /* mirror::Object* */ root = root->Read()
+      codegen_->GenerateReadBarrierForRootSlow(instruction, root, root);
+    }
+  } else {
+    // Plain GC root load with no read barrier.
+    // /* GcRoot<mirror::Object> */ root = *(obj + offset)
+    __ Ldr(root_reg, MemOperand(obj, offset));
+    // Note that GC roots are not affected by heap poisoning, thus we
+    // do not have to unpoison `root_reg` here.
+  }
+}
+
+void CodeGeneratorARM64::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
+                                                               Location ref,
+                                                               vixl::Register obj,
+                                                               uint32_t offset,
+                                                               Register temp,
+                                                               bool needs_null_check,
+                                                               bool use_load_acquire) {
+  DCHECK(kEmitCompilerReadBarrier);
+  DCHECK(kUseBakerReadBarrier);
+
+  // /* HeapReference<Object> */ ref = *(obj + offset)
+  Location no_index = Location::NoLocation();
+  GenerateReferenceLoadWithBakerReadBarrier(
+      instruction, ref, obj, offset, no_index, temp, needs_null_check, use_load_acquire);
+}
+
+void CodeGeneratorARM64::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction,
+                                                               Location ref,
+                                                               vixl::Register obj,
+                                                               uint32_t data_offset,
+                                                               Location index,
+                                                               Register temp,
+                                                               bool needs_null_check) {
+  DCHECK(kEmitCompilerReadBarrier);
+  DCHECK(kUseBakerReadBarrier);
+
+  // Array cells are never volatile variables, therefore array loads
+  // never use Load-Acquire instructions on ARM64.
+  const bool use_load_acquire = false;
+
+  // /* HeapReference<Object> */ ref =
+  //     *(obj + data_offset + index * sizeof(HeapReference<Object>))
+  GenerateReferenceLoadWithBakerReadBarrier(
+      instruction, ref, obj, data_offset, index, temp, needs_null_check, use_load_acquire);
+}
+
+void CodeGeneratorARM64::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction,
+                                                                   Location ref,
+                                                                   vixl::Register obj,
+                                                                   uint32_t offset,
+                                                                   Location index,
+                                                                   Register temp,
+                                                                   bool needs_null_check,
+                                                                   bool use_load_acquire) {
+  DCHECK(kEmitCompilerReadBarrier);
+  DCHECK(kUseBakerReadBarrier);
+  // If `index` is a valid location, then we are emitting an array
+  // load, so we shouldn't be using a Load Acquire instruction.
+  // In other words: `index.IsValid()` => `!use_load_acquire`.
+  DCHECK(!index.IsValid() || !use_load_acquire);
+
+  MacroAssembler* masm = GetVIXLAssembler();
+  UseScratchRegisterScope temps(masm);
+
+  // In slow path based read barriers, the read barrier call is
+  // inserted after the original load. However, in fast path based
+  // Baker's read barriers, we need to perform the load of
+  // mirror::Object::monitor_ *before* the original reference load.
+  // This load-load ordering is required by the read barrier.
+  // The fast path/slow path (for Baker's algorithm) should look like:
+  //
+  //   uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState();
+  //   lfence;  // Load fence or artificial data dependency to prevent load-load reordering
+  //   HeapReference<Object> ref = *src;  // Original reference load.
+  //   bool is_gray = (rb_state == ReadBarrier::gray_ptr_);
+  //   if (is_gray) {
+  //     ref = ReadBarrier::Mark(ref);  // Performed by runtime entrypoint slow path.
+  //   }
+  //
+  // Note: the original implementation in ReadBarrier::Barrier is
+  // slightly more complex as it performs additional checks that we do
+  // not do here for performance reasons.
+
+  Primitive::Type type = Primitive::kPrimNot;
+  Register ref_reg = RegisterFrom(ref, type);
+  DCHECK(obj.IsW());
+  uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value();
+
+  // /* int32_t */ monitor = obj->monitor_
+  __ Ldr(temp, HeapOperand(obj, monitor_offset));
+  if (needs_null_check) {
+    MaybeRecordImplicitNullCheck(instruction);
+  }
+  // /* LockWord */ lock_word = LockWord(monitor)
+  static_assert(sizeof(LockWord) == sizeof(int32_t),
+                "art::LockWord and int32_t have different sizes.");
+  // /* uint32_t */ rb_state = lock_word.ReadBarrierState()
+  __ Lsr(temp, temp, LockWord::kReadBarrierStateShift);
+  __ And(temp, temp, Operand(LockWord::kReadBarrierStateMask));
+  static_assert(
+      LockWord::kReadBarrierStateMask == ReadBarrier::rb_ptr_mask_,
+      "art::LockWord::kReadBarrierStateMask is not equal to art::ReadBarrier::rb_ptr_mask_.");
+
+  // Introduce a dependency on the high bits of rb_state, which shall
+  // be all zeroes, to prevent load-load reordering, and without using
+  // a memory barrier (which would be more expensive).
+  // temp2 = rb_state & ~LockWord::kReadBarrierStateMask = 0
+  Register temp2 = temps.AcquireW();
+  __ Bic(temp2, temp, Operand(LockWord::kReadBarrierStateMask));
+  // obj is unchanged by this operation, but its value now depends on
+  // temp2, which depends on temp.
+  __ Add(obj, obj, Operand(temp2));
+  temps.Release(temp2);
+
+  // The actual reference load.
+  if (index.IsValid()) {
+    static_assert(
+        sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
+        "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
+    temp2 = temps.AcquireW();
+    // /* HeapReference<Object> */ ref =
+    //     *(obj + offset + index * sizeof(HeapReference<Object>))
+    MemOperand source = HeapOperand(obj);
+    if (index.IsConstant()) {
+      uint32_t computed_offset =
+          offset + (Int64ConstantFrom(index) << Primitive::ComponentSizeShift(type));
+      source = HeapOperand(obj, computed_offset);
+    } else {
+      __ Add(temp2, obj, offset);
+      source = HeapOperand(temp2, XRegisterFrom(index), LSL, Primitive::ComponentSizeShift(type));
+    }
+    Load(type, ref_reg, source);
+    temps.Release(temp2);
+  } else {
+    // /* HeapReference<Object> */ ref = *(obj + offset)
+    MemOperand field = HeapOperand(obj, offset);
+    if (use_load_acquire) {
+      LoadAcquire(instruction, ref_reg, field, /* needs_null_check */ false);
+    } else {
+      Load(type, ref_reg, field);
+    }
+  }
+
+  // Object* ref = ref_addr->AsMirrorPtr()
+  GetAssembler()->MaybeUnpoisonHeapReference(ref_reg);
+
+  // Slow path used to mark the object `ref` when it is gray.
+  SlowPathCodeARM64* slow_path =
+      new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, ref, ref);
+  AddSlowPath(slow_path);
+
+  // if (rb_state == ReadBarrier::gray_ptr_)
+  //   ref = ReadBarrier::Mark(ref);
+  __ Cmp(temp, ReadBarrier::gray_ptr_);
+  __ B(eq, slow_path->GetEntryLabel());
+  __ Bind(slow_path->GetExitLabel());
+}
+
+void CodeGeneratorARM64::GenerateReadBarrierSlow(HInstruction* instruction,
+                                                 Location out,
+                                                 Location ref,
+                                                 Location obj,
+                                                 uint32_t offset,
+                                                 Location index) {
   DCHECK(kEmitCompilerReadBarrier);
 
+  // Insert a slow path based read barrier *after* the reference load.
+  //
   // If heap poisoning is enabled, the unpoisoning of the loaded
   // reference will be carried out by the runtime within the slow
   // path.
@@ -4628,57 +4939,41 @@
       ReadBarrierForHeapReferenceSlowPathARM64(instruction, out, ref, obj, offset, index);
   AddSlowPath(slow_path);
 
-  // TODO: When read barrier has a fast path, add it here.
-  /* Currently the read barrier call is inserted after the original load.
-   * However, if we have a fast path, we need to perform the load of obj.LockWord *before* the
-   * original load. This load-load ordering is required by the read barrier.
-   * The fast path/slow path (for Baker's algorithm) should look like:
-   *
-   * bool isGray = obj.LockWord & kReadBarrierMask;
-   * lfence;  // load fence or artificial data dependence to prevent load-load reordering
-   * ref = obj.field;    // this is the original load
-   * if (isGray) {
-   *   ref = Mark(ref);  // ideally the slow path just does Mark(ref)
-   * }
-   */
-
   __ B(slow_path->GetEntryLabel());
   __ Bind(slow_path->GetExitLabel());
 }
 
-void CodeGeneratorARM64::MaybeGenerateReadBarrier(HInstruction* instruction,
-                                                  Location out,
-                                                  Location ref,
-                                                  Location obj,
-                                                  uint32_t offset,
-                                                  Location index) {
+void CodeGeneratorARM64::MaybeGenerateReadBarrierSlow(HInstruction* instruction,
+                                                      Location out,
+                                                      Location ref,
+                                                      Location obj,
+                                                      uint32_t offset,
+                                                      Location index) {
   if (kEmitCompilerReadBarrier) {
+    // Baker's read barriers shall be handled by the fast path
+    // (CodeGeneratorARM64::GenerateReferenceLoadWithBakerReadBarrier).
+    DCHECK(!kUseBakerReadBarrier);
     // If heap poisoning is enabled, unpoisoning will be taken care of
     // by the runtime within the slow path.
-    GenerateReadBarrier(instruction, out, ref, obj, offset, index);
+    GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index);
   } else if (kPoisonHeapReferences) {
     GetAssembler()->UnpoisonHeapReference(WRegisterFrom(out));
   }
 }
 
-void CodeGeneratorARM64::GenerateReadBarrierForRoot(HInstruction* instruction,
-                                                    Location out,
-                                                    Location root) {
+void CodeGeneratorARM64::GenerateReadBarrierForRootSlow(HInstruction* instruction,
+                                                        Location out,
+                                                        Location root) {
   DCHECK(kEmitCompilerReadBarrier);
 
+  // Insert a slow path based read barrier *after* the GC root load.
+  //
   // Note that GC roots are not affected by heap poisoning, so we do
   // not need to do anything special for this here.
   SlowPathCodeARM64* slow_path =
       new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM64(instruction, out, root);
   AddSlowPath(slow_path);
 
-  // TODO: Implement a fast path for ReadBarrierForRoot, performing
-  // the following operation (for Baker's algorithm):
-  //
-  //   if (thread.tls32_.is_gc_marking) {
-  //     root = Mark(root);
-  //   }
-
   __ B(slow_path->GetEntryLabel());
   __ Bind(slow_path->GetExitLabel());
 }
diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index 8eb9fcc..1346bc8 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -208,14 +208,47 @@
 
  private:
   void GenerateClassInitializationCheck(SlowPathCodeARM64* slow_path, vixl::Register class_reg);
-  void GenerateMemoryBarrier(MemBarrierKind kind);
   void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
   void HandleBinaryOp(HBinaryOperation* instr);
+
   void HandleFieldSet(HInstruction* instruction,
                       const FieldInfo& field_info,
                       bool value_can_be_null);
   void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
   void HandleCondition(HCondition* instruction);
+
+  // Generate a heap reference load using one register `out`:
+  //
+  //   out <- *(out + offset)
+  //
+  // while honoring heap poisoning and/or read barriers (if any).
+  // Register `temp` is used when generating a read barrier.
+  void GenerateReferenceLoadOneRegister(HInstruction* instruction,
+                                        Location out,
+                                        uint32_t offset,
+                                        Location temp);
+  // Generate a heap reference load using two different registers
+  // `out` and `obj`:
+  //
+  //   out <- *(obj + offset)
+  //
+  // while honoring heap poisoning and/or read barriers (if any).
+  // Register `temp` is used when generating a Baker's read barrier.
+  void GenerateReferenceLoadTwoRegisters(HInstruction* instruction,
+                                         Location out,
+                                         Location obj,
+                                         uint32_t offset,
+                                         Location temp);
+  // Generate a GC root reference load:
+  //
+  //   root <- *(obj + offset)
+  //
+  // while honoring read barriers (if any).
+  void GenerateGcRootFieldLoad(HInstruction* instruction,
+                               Location root,
+                               vixl::Register obj,
+                               uint32_t offset);
+
   void HandleShift(HBinaryOperation* instr);
   void GenerateImplicitNullCheck(HNullCheck* instruction);
   void GenerateExplicitNullCheck(HNullCheck* instruction);
@@ -337,6 +370,8 @@
   // Emit a write barrier.
   void MarkGCCard(vixl::Register object, vixl::Register value, bool value_can_be_null);
 
+  void GenerateMemoryBarrier(MemBarrierKind kind);
+
   // Register allocation.
 
   void SetupBlockedRegisters() const OVERRIDE;
@@ -386,9 +421,12 @@
   void AddLocationAsTemp(Location location, LocationSummary* locations) OVERRIDE;
 
   void Load(Primitive::Type type, vixl::CPURegister dst, const vixl::MemOperand& src);
-  void Store(Primitive::Type type, vixl::CPURegister rt, const vixl::MemOperand& dst);
-  void LoadAcquire(HInstruction* instruction, vixl::CPURegister dst, const vixl::MemOperand& src);
-  void StoreRelease(Primitive::Type type, vixl::CPURegister rt, const vixl::MemOperand& dst);
+  void Store(Primitive::Type type, vixl::CPURegister src, const vixl::MemOperand& dst);
+  void LoadAcquire(HInstruction* instruction,
+                   vixl::CPURegister dst,
+                   const vixl::MemOperand& src,
+                   bool needs_null_check);
+  void StoreRelease(Primitive::Type type, vixl::CPURegister src, const vixl::MemOperand& dst);
 
   // Generate code to invoke a runtime entry point.
   void InvokeRuntime(QuickEntrypointEnum entrypoint,
@@ -423,7 +461,27 @@
 
   void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
 
-  // Generate a read barrier for a heap reference within `instruction`.
+  // Fast path implementation of ReadBarrier::Barrier for a heap
+  // reference field load when Baker's read barriers are used.
+  void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
+                                             Location ref,
+                                             vixl::Register obj,
+                                             uint32_t offset,
+                                             vixl::Register temp,
+                                             bool needs_null_check,
+                                             bool use_load_acquire);
+  // Fast path implementation of ReadBarrier::Barrier for a heap
+  // reference array load when Baker's read barriers are used.
+  void GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction,
+                                             Location ref,
+                                             vixl::Register obj,
+                                             uint32_t data_offset,
+                                             Location index,
+                                             vixl::Register temp,
+                                             bool needs_null_check);
+
+  // Generate a read barrier for a heap reference within `instruction`
+  // using a slow path.
   //
   // A read barrier for an object reference read from the heap is
   // implemented as a call to the artReadBarrierSlow runtime entry
@@ -440,23 +498,25 @@
   // When `index` is provided (i.e. for array accesses), the offset
   // value passed to artReadBarrierSlow is adjusted to take `index`
   // into account.
-  void GenerateReadBarrier(HInstruction* instruction,
-                           Location out,
-                           Location ref,
-                           Location obj,
-                           uint32_t offset,
-                           Location index = Location::NoLocation());
+  void GenerateReadBarrierSlow(HInstruction* instruction,
+                               Location out,
+                               Location ref,
+                               Location obj,
+                               uint32_t offset,
+                               Location index = Location::NoLocation());
 
-  // If read barriers are enabled, generate a read barrier for a heap reference.
-  // If heap poisoning is enabled, also unpoison the reference in `out`.
-  void MaybeGenerateReadBarrier(HInstruction* instruction,
-                                Location out,
-                                Location ref,
-                                Location obj,
-                                uint32_t offset,
-                                Location index = Location::NoLocation());
+  // If read barriers are enabled, generate a read barrier for a heap
+  // reference using a slow path. If heap poisoning is enabled, also
+  // unpoison the reference in `out`.
+  void MaybeGenerateReadBarrierSlow(HInstruction* instruction,
+                                    Location out,
+                                    Location ref,
+                                    Location obj,
+                                    uint32_t offset,
+                                    Location index = Location::NoLocation());
 
-  // Generate a read barrier for a GC root within `instruction`.
+  // Generate a read barrier for a GC root within `instruction` using
+  // a slow path.
   //
   // A read barrier for an object reference GC root is implemented as
   // a call to the artReadBarrierForRootSlow runtime entry point,
@@ -466,9 +526,20 @@
   //
   // The `out` location contains the value returned by
   // artReadBarrierForRootSlow.
-  void GenerateReadBarrierForRoot(HInstruction* instruction, Location out, Location root);
+  void GenerateReadBarrierForRootSlow(HInstruction* instruction, Location out, Location root);
 
  private:
+  // Factored implementation of GenerateFieldLoadWithBakerReadBarrier
+  // and GenerateArrayLoadWithBakerReadBarrier.
+  void GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction,
+                                                 Location ref,
+                                                 vixl::Register obj,
+                                                 uint32_t offset,
+                                                 Location index,
+                                                 vixl::Register temp,
+                                                 bool needs_null_check,
+                                                 bool use_load_acquire);
+
   using Uint64ToLiteralMap = ArenaSafeMap<uint64_t, vixl::Literal<uint64_t>*>;
   using MethodToLiteralMap = ArenaSafeMap<MethodReference,
                                           vixl::Literal<uint64_t>*,
diff --git a/compiler/optimizing/intrinsics_arm64.cc b/compiler/optimizing/intrinsics_arm64.cc
index 81cab86..e857f6f 100644
--- a/compiler/optimizing/intrinsics_arm64.cc
+++ b/compiler/optimizing/intrinsics_arm64.cc
@@ -752,21 +752,33 @@
   Register trg = RegisterFrom(trg_loc, type);
   bool use_acquire_release = codegen->GetInstructionSetFeatures().PreferAcquireRelease();
 
-  MemOperand mem_op(base.X(), offset);
-  if (is_volatile) {
-    if (use_acquire_release) {
-      codegen->LoadAcquire(invoke, trg, mem_op);
-    } else {
-      codegen->Load(type, trg, mem_op);
+  if (type == Primitive::kPrimNot && kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
+    // UnsafeGetObject/UnsafeGetObjectVolatile with Baker's read barrier case.
+    UseScratchRegisterScope temps(masm);
+    Register temp = temps.AcquireW();
+    codegen->GenerateArrayLoadWithBakerReadBarrier(
+        invoke, trg_loc, base, 0U, offset_loc, temp, /* needs_null_check */ false);
+    if (is_volatile && !use_acquire_release) {
       __ Dmb(InnerShareable, BarrierReads);
     }
   } else {
-    codegen->Load(type, trg, mem_op);
-  }
+    // Other cases.
+    MemOperand mem_op(base.X(), offset);
+    if (is_volatile) {
+      if (use_acquire_release) {
+        codegen->LoadAcquire(invoke, trg, mem_op, /* needs_null_check */ true);
+      } else {
+        codegen->Load(type, trg, mem_op);
+        __ Dmb(InnerShareable, BarrierReads);
+      }
+    } else {
+      codegen->Load(type, trg, mem_op);
+    }
 
-  if (type == Primitive::kPrimNot) {
-    DCHECK(trg.IsW());
-    codegen->MaybeGenerateReadBarrier(invoke, trg_loc, trg_loc, base_loc, 0U, offset_loc);
+    if (type == Primitive::kPrimNot) {
+      DCHECK(trg.IsW());
+      codegen->MaybeGenerateReadBarrierSlow(invoke, trg_loc, trg_loc, base_loc, 0U, offset_loc);
+    }
   }
 }
 
@@ -1026,10 +1038,15 @@
   vixl::Label loop_head, exit_loop;
   if (use_acquire_release) {
     __ Bind(&loop_head);
-    __ Ldaxr(tmp_value, MemOperand(tmp_ptr));
-    // TODO: Do we need a read barrier here when `type == Primitive::kPrimNot`?
+    // TODO: When `type == Primitive::kPrimNot`, add a read barrier for
+    // the reference stored in the object before attempting the CAS,
+    // similar to the one in the art::Unsafe_compareAndSwapObject JNI
+    // implementation.
+    //
     // Note that this code is not (yet) used when read barriers are
     // enabled (see IntrinsicLocationsBuilderARM64::VisitUnsafeCASObject).
+    DCHECK(!(type == Primitive::kPrimNot && kEmitCompilerReadBarrier));
+    __ Ldaxr(tmp_value, MemOperand(tmp_ptr));
     __ Cmp(tmp_value, expected);
     __ B(&exit_loop, ne);
     __ Stlxr(tmp_32, value, MemOperand(tmp_ptr));