diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 7b6c97c..5bdaac2 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -633,23 +633,58 @@
   }
 }
 
-// Abstract base class for read barrier slow paths marking a reference
-// `ref`.
+// Slow path marking an object reference `ref` during a read
+// barrier. The field `obj.field` in the object `obj` holding this
+// reference does not get updated by this slow path after marking (see
+// ReadBarrierMarkAndUpdateFieldSlowPathARM64 below for that).
 //
-// Argument `entrypoint` must be a register location holding the read
-// barrier marking runtime entry point to be invoked.
-class ReadBarrierMarkSlowPathBaseARM64 : public SlowPathCodeARM64 {
- protected:
-  ReadBarrierMarkSlowPathBaseARM64(HInstruction* instruction, Location ref, Location entrypoint)
-      : SlowPathCodeARM64(instruction), ref_(ref), entrypoint_(entrypoint) {
+// This means that after the execution of this slow path, `ref` will
+// always be up-to-date, but `obj.field` may not; i.e., after the
+// flip, `ref` will be a to-space reference, but `obj.field` will
+// probably still be a from-space reference (unless it gets updated by
+// another thread, or if another thread installed another object
+// reference (different from `ref`) in `obj.field`).
+//
+// If `entrypoint` is a valid location it is assumed to already be
+// holding the entrypoint. The case where the entrypoint is passed in
+// is when the decision to mark is based on whether the GC is marking.
+class ReadBarrierMarkSlowPathARM64 : public SlowPathCodeARM64 {
+ public:
+  ReadBarrierMarkSlowPathARM64(HInstruction* instruction,
+                               Location ref,
+                               Location entrypoint = Location::NoLocation())
+      : SlowPathCodeARM64(instruction),
+        ref_(ref),
+        entrypoint_(entrypoint) {
     DCHECK(kEmitCompilerReadBarrier);
   }
 
-  const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathBaseARM64"; }
+  const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM64"; }
 
-  // Generate assembly code calling the read barrier marking runtime
-  // entry point (ReadBarrierMarkRegX).
-  void GenerateReadBarrierMarkRuntimeCall(CodeGenerator* codegen) {
+  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+    LocationSummary* locations = instruction_->GetLocations();
+    DCHECK(locations->CanCall());
+    DCHECK(ref_.IsRegister()) << ref_;
+    DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_.reg())) << ref_.reg();
+    DCHECK(instruction_->IsInstanceFieldGet() ||
+           instruction_->IsStaticFieldGet() ||
+           instruction_->IsArrayGet() ||
+           instruction_->IsArraySet() ||
+           instruction_->IsLoadClass() ||
+           instruction_->IsLoadString() ||
+           instruction_->IsInstanceOf() ||
+           instruction_->IsCheckCast() ||
+           (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()) ||
+           (instruction_->IsInvokeStaticOrDirect() && instruction_->GetLocations()->Intrinsified()))
+        << "Unexpected instruction in read barrier marking slow path: "
+        << instruction_->DebugName();
+    // The read barrier instrumentation of object ArrayGet
+    // instructions does not support the HIntermediateAddress
+    // instruction.
+    DCHECK(!(instruction_->IsArrayGet() &&
+             instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress()));
+
+    __ Bind(GetEntryLabel());
     // No need to save live registers; it's taken care of by the
     // entrypoint. Also, there is no need to update the stack mask,
     // as this runtime call will not trigger a garbage collection.
@@ -685,253 +720,52 @@
       // This runtime call does not require a stack map.
       arm64_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this);
     }
+    __ B(GetExitLabel());
   }
 
+ private:
   // The location (register) of the marked object reference.
   const Location ref_;
 
   // The location of the entrypoint if it is already loaded.
   const Location entrypoint_;
 
- private:
-  DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathBaseARM64);
-};
-
-// Slow path marking an object reference `ref` during a read
-// barrier. The field `obj.field` in the object `obj` holding this
-// reference does not get updated by this slow path after marking.
-//
-// This means that after the execution of this slow path, `ref` will
-// always be up-to-date, but `obj.field` may not; i.e., after the
-// flip, `ref` will be a to-space reference, but `obj.field` will
-// probably still be a from-space reference (unless it gets updated by
-// another thread, or if another thread installed another object
-// reference (different from `ref`) in `obj.field`).
-//
-// If `entrypoint` is a valid location it is assumed to already be
-// holding the entrypoint. The case where the entrypoint is passed in
-// is when the decision to mark is based on whether the GC is marking.
-class ReadBarrierMarkSlowPathARM64 : public ReadBarrierMarkSlowPathBaseARM64 {
- public:
-  ReadBarrierMarkSlowPathARM64(HInstruction* instruction,
-                               Location ref,
-                               Location entrypoint = Location::NoLocation())
-      : ReadBarrierMarkSlowPathBaseARM64(instruction, ref, entrypoint) {
-    DCHECK(kEmitCompilerReadBarrier);
-  }
-
-  const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM64"; }
-
-  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
-    LocationSummary* locations = instruction_->GetLocations();
-    DCHECK(locations->CanCall());
-    DCHECK(ref_.IsRegister()) << ref_;
-    DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_.reg())) << ref_.reg();
-    DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString())
-        << "Unexpected instruction in read barrier marking slow path: "
-        << instruction_->DebugName();
-
-    __ Bind(GetEntryLabel());
-    GenerateReadBarrierMarkRuntimeCall(codegen);
-    __ B(GetExitLabel());
-  }
-
- private:
   DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM64);
 };
 
-// Slow path loading `obj`'s lock word, loading a reference from
-// object `*(obj + offset + (index << scale_factor))` into `ref`, and
-// marking `ref` if `obj` is gray according to the lock word (Baker
-// read barrier). The field `obj.field` in the object `obj` holding
-// this reference does not get updated by this slow path after marking
-// (see LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64
-// below for that).
-//
-// This means that after the execution of this slow path, `ref` will
-// always be up-to-date, but `obj.field` may not; i.e., after the
-// flip, `ref` will be a to-space reference, but `obj.field` will
-// probably still be a from-space reference (unless it gets updated by
-// another thread, or if another thread installed another object
-// reference (different from `ref`) in `obj.field`).
-//
-// Argument `entrypoint` must be a register location holding the read
-// barrier marking runtime entry point to be invoked.
-class LoadReferenceWithBakerReadBarrierSlowPathARM64 : public ReadBarrierMarkSlowPathBaseARM64 {
- public:
-  LoadReferenceWithBakerReadBarrierSlowPathARM64(HInstruction* instruction,
-                                                 Location ref,
-                                                 Register obj,
-                                                 uint32_t offset,
-                                                 Location index,
-                                                 size_t scale_factor,
-                                                 bool needs_null_check,
-                                                 bool use_load_acquire,
-                                                 Register temp,
-                                                 Location entrypoint)
-      : ReadBarrierMarkSlowPathBaseARM64(instruction, ref, entrypoint),
-        obj_(obj),
-        offset_(offset),
-        index_(index),
-        scale_factor_(scale_factor),
-        needs_null_check_(needs_null_check),
-        use_load_acquire_(use_load_acquire),
-        temp_(temp) {
-    DCHECK(kEmitCompilerReadBarrier);
-    DCHECK(kUseBakerReadBarrier);
-  }
-
-  const char* GetDescription() const OVERRIDE {
-    return "LoadReferenceWithBakerReadBarrierSlowPathARM64";
-  }
-
-  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
-    LocationSummary* locations = instruction_->GetLocations();
-    DCHECK(locations->CanCall());
-    DCHECK(ref_.IsRegister()) << ref_;
-    DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_.reg())) << ref_.reg();
-    DCHECK(obj_.IsW());
-    DCHECK_NE(ref_.reg(), LocationFrom(temp_).reg());
-    DCHECK(instruction_->IsInstanceFieldGet() ||
-           instruction_->IsStaticFieldGet() ||
-           instruction_->IsArrayGet() ||
-           instruction_->IsArraySet() ||
-           instruction_->IsInstanceOf() ||
-           instruction_->IsCheckCast() ||
-           (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()) ||
-           (instruction_->IsInvokeStaticOrDirect() && instruction_->GetLocations()->Intrinsified()))
-        << "Unexpected instruction in read barrier marking slow path: "
-        << instruction_->DebugName();
-    // The read barrier instrumentation of object ArrayGet
-    // instructions does not support the HIntermediateAddress
-    // instruction.
-    DCHECK(!(instruction_->IsArrayGet() &&
-             instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress()));
-
-    __ Bind(GetEntryLabel());
-
-    // When using MaybeGenerateReadBarrierSlow, 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<mirror::Object> ref = *src;  // Original reference load.
-    //   bool is_gray = (rb_state == ReadBarrier::GrayState());
-    //   if (is_gray) {
-    //     ref = entrypoint(ref);  // ref = ReadBarrier::Mark(ref);  // Runtime entry point call.
-    //   }
-    //
-    // 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.
-
-    // /* int32_t */ monitor = obj->monitor_
-    uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value();
-    __ Ldr(temp_, HeapOperand(obj_, monitor_offset));
-    if (needs_null_check_) {
-      codegen->MaybeRecordImplicitNullCheck(instruction_);
-    }
-    // /* LockWord */ lock_word = LockWord(monitor)
-    static_assert(sizeof(LockWord) == sizeof(int32_t),
-                  "art::LockWord and int32_t have different sizes.");
-
-    // Introduce a dependency on the lock_word including rb_state,
-    // to prevent load-load reordering, and without using
-    // a memory barrier (which would be more expensive).
-    // `obj` is unchanged by this operation, but its value now depends
-    // on `temp`.
-    __ Add(obj_.X(), obj_.X(), Operand(temp_.X(), LSR, 32));
-
-    // The actual reference load.
-    // A possible implicit null check has already been handled above.
-    CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
-    arm64_codegen->GenerateRawReferenceLoad(instruction_,
-                                            ref_,
-                                            obj_,
-                                            offset_,
-                                            index_,
-                                            scale_factor_,
-                                            /* needs_null_check */ false,
-                                            use_load_acquire_);
-
-    // Mark the object `ref` when `obj` is gray.
-    //
-    //   if (rb_state == ReadBarrier::GrayState())
-    //     ref = ReadBarrier::Mark(ref);
-    //
-    // Given the numeric representation, it's enough to check the low bit of the rb_state.
-    static_assert(ReadBarrier::WhiteState() == 0, "Expecting white to have value 0");
-    static_assert(ReadBarrier::GrayState() == 1, "Expecting gray to have value 1");
-    __ Tbz(temp_, LockWord::kReadBarrierStateShift, GetExitLabel());
-    GenerateReadBarrierMarkRuntimeCall(codegen);
-
-    __ B(GetExitLabel());
-  }
-
- private:
-  // The register containing the object holding the marked object reference field.
-  Register obj_;
-  // The offset, index and scale factor to access the reference in `obj_`.
-  uint32_t offset_;
-  Location index_;
-  size_t scale_factor_;
-  // Is a null check required?
-  bool needs_null_check_;
-  // Should this reference load use Load-Acquire semantics?
-  bool use_load_acquire_;
-  // A temporary register used to hold the lock word of `obj_`.
-  Register temp_;
-
-  DISALLOW_COPY_AND_ASSIGN(LoadReferenceWithBakerReadBarrierSlowPathARM64);
-};
-
-// Slow path loading `obj`'s lock word, loading a reference from
-// object `*(obj + offset + (index << scale_factor))` into `ref`, and
-// marking `ref` if `obj` is gray according to the lock word (Baker
-// read barrier). If needed, this slow path also atomically updates
-// the field `obj.field` in the object `obj` holding this reference
-// after marking (contrary to
-// LoadReferenceWithBakerReadBarrierSlowPathARM64 above, which never
-// tries to update `obj.field`).
+// Slow path marking an object reference `ref` during a read barrier,
+// and if needed, atomically updating the field `obj.field` in the
+// object `obj` holding this reference after marking (contrary to
+// ReadBarrierMarkSlowPathARM64 above, which never tries to update
+// `obj.field`).
 //
 // This means that after the execution of this slow path, both `ref`
 // and `obj.field` will be up-to-date; i.e., after the flip, both will
 // hold the same to-space reference (unless another thread installed
 // another object reference (different from `ref`) in `obj.field`).
 //
-// Argument `entrypoint` must be a register location holding the read
-// barrier marking runtime entry point to be invoked.
-class LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64
-    : public ReadBarrierMarkSlowPathBaseARM64 {
+// If `entrypoint` is a valid location it is assumed to already be
+// holding the entrypoint. The case where the entrypoint is passed in
+// is when the decision to mark is based on whether the GC is marking.
+class ReadBarrierMarkAndUpdateFieldSlowPathARM64 : public SlowPathCodeARM64 {
  public:
-  LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64(HInstruction* instruction,
-                                                               Location ref,
-                                                               Register obj,
-                                                               uint32_t offset,
-                                                               Location index,
-                                                               size_t scale_factor,
-                                                               bool needs_null_check,
-                                                               bool use_load_acquire,
-                                                               Register temp,
-                                                               Location entrypoint)
-      : ReadBarrierMarkSlowPathBaseARM64(instruction, ref, entrypoint),
+  ReadBarrierMarkAndUpdateFieldSlowPathARM64(HInstruction* instruction,
+                                             Location ref,
+                                             Register obj,
+                                             Location field_offset,
+                                             Register temp,
+                                             Location entrypoint = Location::NoLocation())
+      : SlowPathCodeARM64(instruction),
+        ref_(ref),
         obj_(obj),
-        offset_(offset),
-        index_(index),
-        scale_factor_(scale_factor),
-        needs_null_check_(needs_null_check),
-        use_load_acquire_(use_load_acquire),
-        temp_(temp) {
+        field_offset_(field_offset),
+        temp_(temp),
+        entrypoint_(entrypoint) {
     DCHECK(kEmitCompilerReadBarrier);
-    DCHECK(kUseBakerReadBarrier);
   }
 
   const char* GetDescription() const OVERRIDE {
-    return "LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64";
+    return "ReadBarrierMarkAndUpdateFieldSlowPathARM64";
   }
 
   void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
@@ -940,82 +774,70 @@
     DCHECK(locations->CanCall());
     DCHECK(ref_.IsRegister()) << ref_;
     DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_.reg())) << ref_.reg();
-    DCHECK(obj_.IsW());
-    DCHECK_NE(ref_.reg(), LocationFrom(temp_).reg());
-
-    // This slow path is only used by the UnsafeCASObject intrinsic at the moment.
+    // This slow path is only used by the UnsafeCASObject intrinsic.
     DCHECK((instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()))
         << "Unexpected instruction in read barrier marking and field updating slow path: "
         << instruction_->DebugName();
     DCHECK(instruction_->GetLocations()->Intrinsified());
     DCHECK_EQ(instruction_->AsInvoke()->GetIntrinsic(), Intrinsics::kUnsafeCASObject);
-    DCHECK_EQ(offset_, 0u);
-    DCHECK_EQ(scale_factor_, 0u);
-    DCHECK_EQ(use_load_acquire_, false);
-    // The location of the offset of the marked reference field within `obj_`.
-    Location field_offset = index_;
-    DCHECK(field_offset.IsRegister()) << field_offset;
+    DCHECK(field_offset_.IsRegister()) << field_offset_;
 
     __ Bind(GetEntryLabel());
 
-    // /* int32_t */ monitor = obj->monitor_
-    uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value();
-    __ Ldr(temp_, HeapOperand(obj_, monitor_offset));
-    if (needs_null_check_) {
-      codegen->MaybeRecordImplicitNullCheck(instruction_);
-    }
-    // /* LockWord */ lock_word = LockWord(monitor)
-    static_assert(sizeof(LockWord) == sizeof(int32_t),
-                  "art::LockWord and int32_t have different sizes.");
-
-    // Introduce a dependency on the lock_word including rb_state,
-    // to prevent load-load reordering, and without using
-    // a memory barrier (which would be more expensive).
-    // `obj` is unchanged by this operation, but its value now depends
-    // on `temp`.
-    __ Add(obj_.X(), obj_.X(), Operand(temp_.X(), LSR, 32));
-
-    // The actual reference load.
-    // A possible implicit null check has already been handled above.
-    CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
-    arm64_codegen->GenerateRawReferenceLoad(instruction_,
-                                            ref_,
-                                            obj_,
-                                            offset_,
-                                            index_,
-                                            scale_factor_,
-                                            /* needs_null_check */ false,
-                                            use_load_acquire_);
-
-    // Mark the object `ref` when `obj` is gray.
-    //
-    //   if (rb_state == ReadBarrier::GrayState())
-    //     ref = ReadBarrier::Mark(ref);
-    //
-    // Given the numeric representation, it's enough to check the low bit of the rb_state.
-    static_assert(ReadBarrier::WhiteState() == 0, "Expecting white to have value 0");
-    static_assert(ReadBarrier::GrayState() == 1, "Expecting gray to have value 1");
-    __ Tbz(temp_, LockWord::kReadBarrierStateShift, GetExitLabel());
-
-    // Save the old value of the reference before marking it.
+    // Save the old reference.
     // Note that we cannot use IP to save the old reference, as IP is
     // used internally by the ReadBarrierMarkRegX entry point, and we
     // need the old reference after the call to that entry point.
     DCHECK_NE(LocationFrom(temp_).reg(), IP0);
     __ Mov(temp_.W(), ref_reg);
 
-    GenerateReadBarrierMarkRuntimeCall(codegen);
+    // No need to save live registers; it's taken care of by the
+    // entrypoint. Also, there is no need to update the stack mask,
+    // as this runtime call will not trigger a garbage collection.
+    CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
+    DCHECK_NE(ref_.reg(), LR);
+    DCHECK_NE(ref_.reg(), WSP);
+    DCHECK_NE(ref_.reg(), WZR);
+    // IP0 is used internally by the ReadBarrierMarkRegX entry point
+    // as a temporary, it cannot be the entry point's input/output.
+    DCHECK_NE(ref_.reg(), IP0);
+    DCHECK(0 <= ref_.reg() && ref_.reg() < kNumberOfWRegisters) << ref_.reg();
+    // "Compact" slow path, saving two moves.
+    //
+    // Instead of using the standard runtime calling convention (input
+    // and output in W0):
+    //
+    //   W0 <- ref
+    //   W0 <- ReadBarrierMark(W0)
+    //   ref <- W0
+    //
+    // we just use rX (the register containing `ref`) as input and output
+    // of a dedicated entrypoint:
+    //
+    //   rX <- ReadBarrierMarkRegX(rX)
+    //
+    if (entrypoint_.IsValid()) {
+      arm64_codegen->ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction_, this);
+      __ Blr(XRegisterFrom(entrypoint_));
+    } else {
+      // Entrypoint is not already loaded, load from the thread.
+      int32_t entry_point_offset =
+          CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArm64PointerSize>(ref_.reg());
+      // This runtime call does not require a stack map.
+      arm64_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this);
+    }
 
     // If the new reference is different from the old reference,
-    // update the field in the holder (`*(obj_ + field_offset)`).
+    // update the field in the holder (`*(obj_ + field_offset_)`).
     //
     // Note that this field could also hold a different object, if
     // another thread had concurrently changed it. In that case, the
     // LDXR/CMP/BNE sequence of instructions in the compare-and-set
     // (CAS) operation below would abort the CAS, leaving the field
     // as-is.
+    vixl::aarch64::Label done;
     __ Cmp(temp_.W(), ref_reg);
-    __ B(eq, GetExitLabel());
+    __ B(eq, &done);
 
     // Update the the holder's field atomically.  This may fail if
     // mutator updates before us, but it's OK.  This is achieved
@@ -1028,7 +850,7 @@
 
     // Convenience aliases.
     Register base = obj_.W();
-    Register offset = XRegisterFrom(field_offset);
+    Register offset = XRegisterFrom(field_offset_);
     Register expected = temp_.W();
     Register value = ref_reg;
     Register tmp_ptr = temps.AcquireX();    // Pointer to actual memory.
@@ -1072,26 +894,24 @@
       }
     }
 
+    __ Bind(&done);
     __ B(GetExitLabel());
   }
 
  private:
+  // The location (register) of the marked object reference.
+  const Location ref_;
   // The register containing the object holding the marked object reference field.
   const Register obj_;
-  // The offset, index and scale factor to access the reference in `obj_`.
-  uint32_t offset_;
-  Location index_;
-  size_t scale_factor_;
-  // Is a null check required?
-  bool needs_null_check_;
-  // Should this reference load use Load-Acquire semantics?
-  bool use_load_acquire_;
-  // A temporary register used to hold the lock word of `obj_`; and
-  // also to hold the original reference value, when the reference is
-  // marked.
+  // The location of the offset of the marked reference field within `obj_`.
+  Location field_offset_;
+
   const Register temp_;
 
-  DISALLOW_COPY_AND_ASSIGN(LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64);
+  // The location of the entrypoint if it is already loaded.
+  const Location entrypoint_;
+
+  DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkAndUpdateFieldSlowPathARM64);
 };
 
 // Slow path generating a read barrier for a heap reference.
@@ -5953,11 +5773,13 @@
   // `instruction->IsArrayGet()` => `!use_load_acquire`.
   DCHECK(!instruction->IsArrayGet() || !use_load_acquire);
 
-  // Query `art::Thread::Current()->GetIsGcMarking()` to decide
-  // whether we need to enter the slow path to mark the reference.
-  // Then, in the slow path, check the gray bit in the lock word of
-  // the reference's holder (`obj`) to decide whether to mark `ref` or
-  // not.
+  // After loading the reference from `obj.field` into `ref`, query
+  // `art::Thread::Current()->GetIsGcMarking()` to decide whether we
+  // need to enter the slow path to mark the reference. This
+  // optimistic strategy (we expect the GC to not be marking most of
+  // the time) does not check `obj`'s lock word (to see if it is a
+  // gray object or not), so may sometimes mark an already marked
+  // object.
   //
   // Note that we do not actually check the value of `GetIsGcMarking()`;
   // instead, we load into `temp2` the read barrier mark entry point
@@ -5965,17 +5787,10 @@
   // that `GetIsGcMarking()` is false, and vice versa.
   //
   //   temp2 = Thread::Current()->pReadBarrierMarkReg ## root.reg()
+  //   HeapReference<mirror::Object> ref = *src;  // Original reference load.
   //   if (temp2 != nullptr) {  // <=> Thread::Current()->GetIsGcMarking()
   //     // Slow path.
-  //     uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState();
-  //     lfence;  // Load fence or artificial data dependency to prevent load-load reordering
-  //     HeapReference<mirror::Object> ref = *src;  // Original reference load.
-  //     bool is_gray = (rb_state == ReadBarrier::GrayState());
-  //     if (is_gray) {
-  //       ref = temp2(ref);  // ref = ReadBarrier::Mark(ref);  // Runtime entry point call.
-  //     }
-  //   } else {
-  //     HeapReference<mirror::Object> ref = *src;  // Original reference load.
+  //     ref = temp2(ref);  // ref = ReadBarrier::Mark(ref);  // Runtime entry point call.
   //   }
 
   // Slow path marking the object `ref` when the GC is marking. The
@@ -5984,38 +5799,17 @@
   Location temp2_loc = LocationFrom(temp2);
   SlowPathCodeARM64* slow_path;
   if (always_update_field) {
-    // LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64
-    // only supports address of the form `obj + field_offset`, where
-    // `obj` is a register and `field_offset` is a register. Thus
-    // `offset` and `scale_factor` above are expected to be null in
-    // this code path.
+    // ReadBarrierMarkAndUpdateFieldSlowPathARM64 only supports
+    // address of the form `obj + field_offset`, where `obj` is a
+    // register and `field_offset` is a register. Thus `offset` and
+    // `scale_factor` above are expected to be null in this code path.
     DCHECK_EQ(offset, 0u);
     DCHECK_EQ(scale_factor, 0u);  /* "times 1" */
-    Location field_offset = index;
-    slow_path =
-        new (GetGraph()->GetArena()) LoadReferenceWithBakerReadBarrierAndUpdateFieldSlowPathARM64(
-            instruction,
-            ref,
-            obj,
-            offset,
-            /* index */ field_offset,
-            scale_factor,
-            needs_null_check,
-            use_load_acquire,
-            temp,
-            /* entrypoint */ temp2_loc);
+    slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkAndUpdateFieldSlowPathARM64(
+        instruction, ref, obj, /* field_offset */ index, temp, /* entrypoint */ temp2_loc);
   } else {
-    slow_path = new (GetGraph()->GetArena()) LoadReferenceWithBakerReadBarrierSlowPathARM64(
-        instruction,
-        ref,
-        obj,
-        offset,
-        index,
-        scale_factor,
-        needs_null_check,
-        use_load_acquire,
-        temp,
-        /* entrypoint */ temp2_loc);
+    slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(
+        instruction, ref, /* entrypoint */ temp2_loc);
   }
   AddSlowPath(slow_path);
 
@@ -6025,12 +5819,12 @@
   // Loading the entrypoint does not require a load acquire since it is only changed when
   // threads are suspended or running a checkpoint.
   __ Ldr(temp2, MemOperand(tr, entry_point_offset));
+  // The reference load.
+  GenerateRawReferenceLoad(
+      instruction, ref, obj, offset, index, scale_factor, needs_null_check, use_load_acquire);
   // The entrypoint is null when the GC is not marking, this prevents one load compared to
   // checking GetIsGcMarking.
   __ Cbnz(temp2, slow_path->GetEntryLabel());
-  // Fast path: just load the reference.
-  GenerateRawReferenceLoad(
-      instruction, ref, obj, offset, index, scale_factor, needs_null_check, use_load_acquire);
   __ Bind(slow_path->GetExitLabel());
 }