Merge "Make the mark stack expandable for the CC collector."
diff --git a/build/Android.common_path.mk b/build/Android.common_path.mk
index 4abd191..c53479c 100644
--- a/build/Android.common_path.mk
+++ b/build/Android.common_path.mk
@@ -89,9 +89,11 @@
 HOST_CORE_DEX_FILES   := $(foreach jar,$(HOST_CORE_JARS),  $(call intermediates-dir-for,JAVA_LIBRARIES,$(jar),t,COMMON)/javalib.jar)
 TARGET_CORE_DEX_FILES := $(foreach jar,$(TARGET_CORE_JARS),$(call intermediates-dir-for,JAVA_LIBRARIES,$(jar), ,COMMON)/javalib.jar)
 
+ifeq ($(ANDROID_COMPILE_WITH_JACK),true)
 # Classpath for Jack compilation: we only need core-libart.
 HOST_JACK_CLASSPATH_DEPENDENCIES   := $(call intermediates-dir-for,JAVA_LIBRARIES,core-libart-hostdex,t,COMMON)/classes.jack
 HOST_JACK_CLASSPATH                := $(foreach dep,$(HOST_JACK_CLASSPATH_DEPENDENCIES),$(abspath $(dep)))
 TARGET_JACK_CLASSPATH_DEPENDENCIES := $(call intermediates-dir-for,JAVA_LIBRARIES,core-libart, ,COMMON)/classes.jack
 TARGET_JACK_CLASSPATH              := $(foreach dep,$(TARGET_JACK_CLASSPATH_DEPENDENCIES),$(abspath $(dep)))
+endif
 endif # ART_ANDROID_COMMON_PATH_MK
diff --git a/build/Android.gtest.mk b/build/Android.gtest.mk
index 6295e15..fdfd94c 100644
--- a/build/Android.gtest.mk
+++ b/build/Android.gtest.mk
@@ -230,6 +230,7 @@
   runtime/jni_internal_test.cc \
   runtime/proxy_test.cc \
   runtime/reflection_test.cc \
+  compiler/compiled_method_test.cc \
   compiler/dex/gvn_dead_code_elimination_test.cc \
   compiler/dex/global_value_numbering_test.cc \
   compiler/dex/local_value_numbering_test.cc \
@@ -237,6 +238,7 @@
   compiler/dex/mir_optimization_test.cc \
   compiler/dex/type_inference_test.cc \
   compiler/dwarf/dwarf_test.cc \
+  compiler/driver/compiled_method_storage_test.cc \
   compiler/driver/compiler_driver_test.cc \
   compiler/elf_writer_test.cc \
   compiler/image_test.cc \
diff --git a/compiler/Android.mk b/compiler/Android.mk
index 17f9d12..aaac126 100644
--- a/compiler/Android.mk
+++ b/compiler/Android.mk
@@ -54,6 +54,7 @@
 	dex/verification_results.cc \
 	dex/vreg_analysis.cc \
 	dex/quick_compiler_callbacks.cc \
+	driver/compiled_method_storage.cc \
 	driver/compiler_driver.cc \
 	driver/compiler_options.cc \
 	driver/dex_compilation_unit.cc \
@@ -88,6 +89,7 @@
 	optimizing/primitive_type_propagation.cc \
 	optimizing/reference_type_propagation.cc \
 	optimizing/register_allocator.cc \
+	optimizing/sharpening.cc \
 	optimizing/side_effects_analysis.cc \
 	optimizing/ssa_builder.cc \
 	optimizing/ssa_liveness_analysis.cc \
diff --git a/compiler/cfi_test.h b/compiler/cfi_test.h
index 5e345db..6fd4575 100644
--- a/compiler/cfi_test.h
+++ b/compiler/cfi_test.h
@@ -51,7 +51,7 @@
     dwarf::WriteDebugFrameCIE(is64bit, dwarf::DW_EH_PE_absptr, dwarf::Reg(8),
                               initial_opcodes, kCFIFormat, &debug_frame_data_);
     std::vector<uintptr_t> debug_frame_patches;
-    dwarf::WriteDebugFrameFDE(is64bit, 0, 0, actual_asm.size(), &actual_cfi,
+    dwarf::WriteDebugFrameFDE(is64bit, 0, 0, actual_asm.size(), ArrayRef<const uint8_t>(actual_cfi),
                               kCFIFormat, &debug_frame_data_, &debug_frame_patches);
     ReformatCfi(Objdump(false, "-W"), &lines);
     // Pretty-print assembly.
diff --git a/compiler/common_compiler_test.cc b/compiler/common_compiler_test.cc
index 58a2f96..151437b 100644
--- a/compiler/common_compiler_test.cc
+++ b/compiler/common_compiler_test.cc
@@ -54,22 +54,22 @@
                                                             method->GetDexMethodIndex()));
   }
   if (compiled_method != nullptr) {
-    const SwapVector<uint8_t>* code = compiled_method->GetQuickCode();
-    uint32_t code_size = code->size();
+    ArrayRef<const uint8_t> code = compiled_method->GetQuickCode();
+    uint32_t code_size = code.size();
     CHECK_NE(0u, code_size);
-    const SwapVector<uint8_t>* vmap_table = compiled_method->GetVmapTable();
-    uint32_t vmap_table_offset = vmap_table->empty() ? 0u
-        : sizeof(OatQuickMethodHeader) + vmap_table->size();
-    const SwapVector<uint8_t>* mapping_table = compiled_method->GetMappingTable();
-    bool mapping_table_used = mapping_table != nullptr && !mapping_table->empty();
-    size_t mapping_table_size = mapping_table_used ? mapping_table->size() : 0U;
+    ArrayRef<const uint8_t> vmap_table = compiled_method->GetVmapTable();
+    uint32_t vmap_table_offset = vmap_table.empty() ? 0u
+        : sizeof(OatQuickMethodHeader) + vmap_table.size();
+    ArrayRef<const uint8_t> mapping_table = compiled_method->GetMappingTable();
+    bool mapping_table_used = !mapping_table.empty();
+    size_t mapping_table_size = mapping_table.size();
     uint32_t mapping_table_offset = !mapping_table_used ? 0u
-        : sizeof(OatQuickMethodHeader) + vmap_table->size() + mapping_table_size;
-    const SwapVector<uint8_t>* gc_map = compiled_method->GetGcMap();
-    bool gc_map_used = gc_map != nullptr && !gc_map->empty();
-    size_t gc_map_size = gc_map_used ? gc_map->size() : 0U;
+        : sizeof(OatQuickMethodHeader) + vmap_table.size() + mapping_table_size;
+    ArrayRef<const uint8_t> gc_map = compiled_method->GetGcMap();
+    bool gc_map_used = !gc_map.empty();
+    size_t gc_map_size = gc_map.size();
     uint32_t gc_map_offset = !gc_map_used ? 0u
-        : sizeof(OatQuickMethodHeader) + vmap_table->size() + mapping_table_size + gc_map_size;
+        : sizeof(OatQuickMethodHeader) + vmap_table.size() + mapping_table_size + gc_map_size;
     OatQuickMethodHeader method_header(mapping_table_offset, vmap_table_offset, gc_map_offset,
                                        compiled_method->GetFrameSizeInBytes(),
                                        compiled_method->GetCoreSpillMask(),
@@ -77,25 +77,25 @@
 
     header_code_and_maps_chunks_.push_back(std::vector<uint8_t>());
     std::vector<uint8_t>* chunk = &header_code_and_maps_chunks_.back();
-    size_t size = sizeof(method_header) + code_size + vmap_table->size() + mapping_table_size +
+    size_t size = sizeof(method_header) + code_size + vmap_table.size() + mapping_table_size +
         gc_map_size;
     size_t code_offset = compiled_method->AlignCode(size - code_size);
     size_t padding = code_offset - (size - code_size);
     chunk->reserve(padding + size);
     chunk->resize(sizeof(method_header));
     memcpy(&(*chunk)[0], &method_header, sizeof(method_header));
-    chunk->insert(chunk->begin(), vmap_table->begin(), vmap_table->end());
+    chunk->insert(chunk->begin(), vmap_table.begin(), vmap_table.end());
     if (mapping_table_used) {
-      chunk->insert(chunk->begin(), mapping_table->begin(), mapping_table->end());
+      chunk->insert(chunk->begin(), mapping_table.begin(), mapping_table.end());
     }
     if (gc_map_used) {
-      chunk->insert(chunk->begin(), gc_map->begin(), gc_map->end());
+      chunk->insert(chunk->begin(), gc_map.begin(), gc_map.end());
     }
     chunk->insert(chunk->begin(), padding, 0);
-    chunk->insert(chunk->end(), code->begin(), code->end());
+    chunk->insert(chunk->end(), code.begin(), code.end());
     CHECK_EQ(padding + size, chunk->size());
     const void* code_ptr = &(*chunk)[code_offset];
-    MakeExecutable(code_ptr, code->size());
+    MakeExecutable(code_ptr, code.size());
     const void* method_code = CompiledMethod::CodePointer(code_ptr,
                                                           compiled_method->GetInstructionSet());
     LOG(INFO) << "MakeExecutable " << PrettyMethod(method) << " code=" << method_code;
diff --git a/compiler/compiled_method.cc b/compiler/compiled_method.cc
index 74ef35e..9551d22 100644
--- a/compiler/compiled_method.cc
+++ b/compiler/compiled_method.cc
@@ -15,27 +15,22 @@
  */
 
 #include "compiled_method.h"
+
+#include "driver/compiled_method_storage.h"
 #include "driver/compiler_driver.h"
+#include "utils/swap_space.h"
 
 namespace art {
 
 CompiledCode::CompiledCode(CompilerDriver* compiler_driver, InstructionSet instruction_set,
-                           const ArrayRef<const uint8_t>& quick_code, bool owns_code_array)
-    : compiler_driver_(compiler_driver), instruction_set_(instruction_set),
-      owns_code_array_(owns_code_array), quick_code_(nullptr) {
-  if (owns_code_array_) {
-    // If we are supposed to own the code, don't deduplicate it.
-    quick_code_ = new SwapVector<uint8_t>(quick_code.begin(), quick_code.end(),
-                                          compiler_driver_->GetSwapSpaceAllocator());
-  } else {
-    quick_code_ = compiler_driver_->DeduplicateCode(quick_code);
-  }
+                           const ArrayRef<const uint8_t>& quick_code)
+    : compiler_driver_(compiler_driver),
+      instruction_set_(instruction_set),
+      quick_code_(compiler_driver_->GetCompiledMethodStorage()->DeduplicateCode(quick_code)) {
 }
 
 CompiledCode::~CompiledCode() {
-  if (owns_code_array_) {
-    delete quick_code_;
-  }
+  compiler_driver_->GetCompiledMethodStorage()->ReleaseCode(quick_code_);
 }
 
 bool CompiledCode::operator==(const CompiledCode& rhs) const {
@@ -104,59 +99,28 @@
   }
 }
 
-const std::vector<uint32_t>& CompiledCode::GetOatdataOffsetsToCompliledCodeOffset() const {
-  CHECK_NE(0U, oatdata_offsets_to_compiled_code_offset_.size());
-  return oatdata_offsets_to_compiled_code_offset_;
-}
-
-void CompiledCode::AddOatdataOffsetToCompliledCodeOffset(uint32_t offset) {
-  oatdata_offsets_to_compiled_code_offset_.push_back(offset);
-}
-
 CompiledMethod::CompiledMethod(CompilerDriver* driver,
                                InstructionSet instruction_set,
                                const ArrayRef<const uint8_t>& quick_code,
                                const size_t frame_size_in_bytes,
                                const uint32_t core_spill_mask,
                                const uint32_t fp_spill_mask,
-                               DefaultSrcMap* src_mapping_table,
+                               const ArrayRef<const SrcMapElem>& src_mapping_table,
                                const ArrayRef<const uint8_t>& mapping_table,
                                const ArrayRef<const uint8_t>& vmap_table,
                                const ArrayRef<const uint8_t>& native_gc_map,
                                const ArrayRef<const uint8_t>& cfi_info,
                                const ArrayRef<const LinkerPatch>& patches)
-    : CompiledCode(driver, instruction_set, quick_code, !driver->DedupeEnabled()),
-      owns_arrays_(!driver->DedupeEnabled()),
+    : CompiledCode(driver, instruction_set, quick_code),
       frame_size_in_bytes_(frame_size_in_bytes), core_spill_mask_(core_spill_mask),
       fp_spill_mask_(fp_spill_mask),
-      patches_(patches.begin(), patches.end(), driver->GetSwapSpaceAllocator()) {
-  if (owns_arrays_) {
-    if (src_mapping_table == nullptr) {
-      src_mapping_table_ = new SwapSrcMap(driver->GetSwapSpaceAllocator());
-    } else {
-      src_mapping_table_ = new SwapSrcMap(src_mapping_table->begin(), src_mapping_table->end(),
-                                          driver->GetSwapSpaceAllocator());
-    }
-    mapping_table_ = mapping_table.empty() ?
-        nullptr : new SwapVector<uint8_t>(mapping_table.begin(), mapping_table.end(),
-                                          driver->GetSwapSpaceAllocator());
-    vmap_table_ = new SwapVector<uint8_t>(vmap_table.begin(), vmap_table.end(),
-                                          driver->GetSwapSpaceAllocator());
-    gc_map_ = native_gc_map.empty() ? nullptr :
-        new SwapVector<uint8_t>(native_gc_map.begin(), native_gc_map.end(),
-                                driver->GetSwapSpaceAllocator());
-    cfi_info_ = cfi_info.empty() ? nullptr :
-        new SwapVector<uint8_t>(cfi_info.begin(), cfi_info.end(), driver->GetSwapSpaceAllocator());
-  } else {
-    src_mapping_table_ = src_mapping_table == nullptr ?
-        driver->DeduplicateSrcMappingTable(ArrayRef<SrcMapElem>()) :
-        driver->DeduplicateSrcMappingTable(ArrayRef<SrcMapElem>(*src_mapping_table));
-    mapping_table_ = mapping_table.empty() ?
-        nullptr : driver->DeduplicateMappingTable(mapping_table);
-    vmap_table_ = driver->DeduplicateVMapTable(vmap_table);
-    gc_map_ = native_gc_map.empty() ? nullptr : driver->DeduplicateGCMap(native_gc_map);
-    cfi_info_ = cfi_info.empty() ? nullptr : driver->DeduplicateCFIInfo(cfi_info);
-  }
+      src_mapping_table_(
+          driver->GetCompiledMethodStorage()->DeduplicateSrcMappingTable(src_mapping_table)),
+      mapping_table_(driver->GetCompiledMethodStorage()->DeduplicateMappingTable(mapping_table)),
+      vmap_table_(driver->GetCompiledMethodStorage()->DeduplicateVMapTable(vmap_table)),
+      gc_map_(driver->GetCompiledMethodStorage()->DeduplicateGCMap(native_gc_map)),
+      cfi_info_(driver->GetCompiledMethodStorage()->DeduplicateCFIInfo(cfi_info)),
+      patches_(driver->GetCompiledMethodStorage()->DeduplicateLinkerPatches(patches)) {
 }
 
 CompiledMethod* CompiledMethod::SwapAllocCompiledMethod(
@@ -166,13 +130,13 @@
     const size_t frame_size_in_bytes,
     const uint32_t core_spill_mask,
     const uint32_t fp_spill_mask,
-    DefaultSrcMap* src_mapping_table,
+    const ArrayRef<const SrcMapElem>& src_mapping_table,
     const ArrayRef<const uint8_t>& mapping_table,
     const ArrayRef<const uint8_t>& vmap_table,
     const ArrayRef<const uint8_t>& native_gc_map,
     const ArrayRef<const uint8_t>& cfi_info,
     const ArrayRef<const LinkerPatch>& patches) {
-  SwapAllocator<CompiledMethod> alloc(driver->GetSwapSpaceAllocator());
+  SwapAllocator<CompiledMethod> alloc(driver->GetCompiledMethodStorage()->GetSwapSpaceAllocator());
   CompiledMethod* ret = alloc.allocate(1);
   alloc.construct(ret, driver, instruction_set, quick_code, frame_size_in_bytes, core_spill_mask,
                   fp_spill_mask, src_mapping_table, mapping_table, vmap_table, native_gc_map,
@@ -180,22 +144,20 @@
   return ret;
 }
 
-
-
 void CompiledMethod::ReleaseSwapAllocatedCompiledMethod(CompilerDriver* driver, CompiledMethod* m) {
-  SwapAllocator<CompiledMethod> alloc(driver->GetSwapSpaceAllocator());
+  SwapAllocator<CompiledMethod> alloc(driver->GetCompiledMethodStorage()->GetSwapSpaceAllocator());
   alloc.destroy(m);
   alloc.deallocate(m, 1);
 }
 
 CompiledMethod::~CompiledMethod() {
-  if (owns_arrays_) {
-    delete src_mapping_table_;
-    delete mapping_table_;
-    delete vmap_table_;
-    delete gc_map_;
-    delete cfi_info_;
-  }
+  CompiledMethodStorage* storage = GetCompilerDriver()->GetCompiledMethodStorage();
+  storage->ReleaseLinkerPatches(patches_);
+  storage->ReleaseCFIInfo(cfi_info_);
+  storage->ReleaseGCMap(gc_map_);
+  storage->ReleaseVMapTable(vmap_table_);
+  storage->ReleaseMappingTable(mapping_table_);
+  storage->ReleaseSrcMappingTable(src_mapping_table_);
 }
 
 }  // namespace art
diff --git a/compiler/compiled_method.h b/compiler/compiled_method.h
index a4d2387..15a4ba0 100644
--- a/compiler/compiled_method.h
+++ b/compiler/compiled_method.h
@@ -23,19 +23,20 @@
 
 #include "arch/instruction_set.h"
 #include "base/bit_utils.h"
+#include "length_prefixed_array.h"
 #include "method_reference.h"
 #include "utils/array_ref.h"
-#include "utils/swap_space.h"
 
 namespace art {
 
 class CompilerDriver;
+class CompiledMethodStorage;
 
 class CompiledCode {
  public:
   // For Quick to supply an code blob
   CompiledCode(CompilerDriver* compiler_driver, InstructionSet instruction_set,
-               const ArrayRef<const uint8_t>& quick_code, bool owns_code_array);
+               const ArrayRef<const uint8_t>& quick_code);
 
   virtual ~CompiledCode();
 
@@ -43,8 +44,8 @@
     return instruction_set_;
   }
 
-  const SwapVector<uint8_t>* GetQuickCode() const {
-    return quick_code_;
+  ArrayRef<const uint8_t> GetQuickCode() const {
+    return GetArray(quick_code_);
   }
 
   bool operator==(const CompiledCode& rhs) const;
@@ -66,41 +67,46 @@
   static const void* CodePointer(const void* code_pointer,
                                  InstructionSet instruction_set);
 
-  const std::vector<uint32_t>& GetOatdataOffsetsToCompliledCodeOffset() const;
-  void AddOatdataOffsetToCompliledCodeOffset(uint32_t offset);
+ protected:
+  template <typename T>
+  static ArrayRef<const T> GetArray(const LengthPrefixedArray<T>* array) {
+    if (array == nullptr) {
+      return ArrayRef<const T>();
+    }
+    DCHECK_NE(array->size(), 0u);
+    return ArrayRef<const T>(&array->At(0), array->size());
+  }
+
+  CompilerDriver* GetCompilerDriver() {
+    return compiler_driver_;
+  }
 
  private:
   CompilerDriver* const compiler_driver_;
 
   const InstructionSet instruction_set_;
 
-  // If we own the code array (means that we free in destructor).
-  const bool owns_code_array_;
-
   // Used to store the PIC code for Quick.
-  SwapVector<uint8_t>* quick_code_;
-
-  // There are offsets from the oatdata symbol to where the offset to
-  // the compiled method will be found. These are computed by the
-  // OatWriter and then used by the ElfWriter to add relocations so
-  // that MCLinker can update the values to the location in the linked .so.
-  std::vector<uint32_t> oatdata_offsets_to_compiled_code_offset_;
+  const LengthPrefixedArray<uint8_t>* const quick_code_;
 };
 
 class SrcMapElem {
  public:
   uint32_t from_;
   int32_t to_;
-
-  // Lexicographical compare.
-  bool operator<(const SrcMapElem& other) const {
-    if (from_ != other.from_) {
-      return from_ < other.from_;
-    }
-    return to_ < other.to_;
-  }
 };
 
+inline bool operator<(const SrcMapElem& lhs, const SrcMapElem& rhs) {
+  if (lhs.from_ != rhs.from_) {
+    return lhs.from_ < rhs.from_;
+  }
+  return lhs.to_ < rhs.to_;
+}
+
+inline bool operator==(const SrcMapElem& lhs, const SrcMapElem& rhs) {
+  return lhs.from_ == rhs.from_ && lhs.to_ == rhs.to_;
+}
+
 template <class Allocator>
 class SrcMap FINAL : public std::vector<SrcMapElem, Allocator> {
  public:
@@ -151,7 +157,6 @@
 };
 
 using DefaultSrcMap = SrcMap<std::allocator<SrcMapElem>>;
-using SwapSrcMap = SrcMap<SwapAllocator<SrcMapElem>>;
 
 
 enum LinkerPatchType {
@@ -273,6 +278,9 @@
     uint32_t method_idx_;       // Method index for Call/Method patches.
     uint32_t type_idx_;         // Type index for Type patches.
     uint32_t element_offset_;   // Element offset in the dex cache arrays.
+    static_assert(sizeof(method_idx_) == sizeof(cmp1_), "needed by relational operators");
+    static_assert(sizeof(type_idx_) == sizeof(cmp1_), "needed by relational operators");
+    static_assert(sizeof(element_offset_) == sizeof(cmp1_), "needed by relational operators");
   };
   union {
     uint32_t cmp2_;             // Used for relational operators.
@@ -313,7 +321,7 @@
                  const size_t frame_size_in_bytes,
                  const uint32_t core_spill_mask,
                  const uint32_t fp_spill_mask,
-                 DefaultSrcMap* src_mapping_table,
+                 const ArrayRef<const SrcMapElem>& src_mapping_table,
                  const ArrayRef<const uint8_t>& mapping_table,
                  const ArrayRef<const uint8_t>& vmap_table,
                  const ArrayRef<const uint8_t>& native_gc_map,
@@ -329,7 +337,7 @@
       const size_t frame_size_in_bytes,
       const uint32_t core_spill_mask,
       const uint32_t fp_spill_mask,
-      DefaultSrcMap* src_mapping_table,
+      const ArrayRef<const SrcMapElem>& src_mapping_table,
       const ArrayRef<const uint8_t>& mapping_table,
       const ArrayRef<const uint8_t>& vmap_table,
       const ArrayRef<const uint8_t>& native_gc_map,
@@ -350,35 +358,31 @@
     return fp_spill_mask_;
   }
 
-  const SwapSrcMap& GetSrcMappingTable() const {
-    DCHECK(src_mapping_table_ != nullptr);
-    return *src_mapping_table_;
+  ArrayRef<const SrcMapElem> GetSrcMappingTable() const {
+    return GetArray(src_mapping_table_);
   }
 
-  SwapVector<uint8_t> const* GetMappingTable() const {
-    return mapping_table_;
+  ArrayRef<const uint8_t> GetMappingTable() const {
+    return GetArray(mapping_table_);
   }
 
-  const SwapVector<uint8_t>* GetVmapTable() const {
-    DCHECK(vmap_table_ != nullptr);
-    return vmap_table_;
+  ArrayRef<const uint8_t> GetVmapTable() const {
+    return GetArray(vmap_table_);
   }
 
-  SwapVector<uint8_t> const* GetGcMap() const {
-    return gc_map_;
+  ArrayRef<const uint8_t> GetGcMap() const {
+    return GetArray(gc_map_);
   }
 
-  const SwapVector<uint8_t>* GetCFIInfo() const {
-    return cfi_info_;
+  ArrayRef<const uint8_t> GetCFIInfo() const {
+    return GetArray(cfi_info_);
   }
 
   ArrayRef<const LinkerPatch> GetPatches() const {
-    return ArrayRef<const LinkerPatch>(patches_);
+    return GetArray(patches_);
   }
 
  private:
-  // Whether or not the arrays are owned by the compiled method or dedupe sets.
-  const bool owns_arrays_;
   // For quick code, the size of the activation used by the code.
   const size_t frame_size_in_bytes_;
   // For quick code, a bit mask describing spilled GPR callee-save registers.
@@ -386,19 +390,19 @@
   // For quick code, a bit mask describing spilled FPR callee-save registers.
   const uint32_t fp_spill_mask_;
   // For quick code, a set of pairs (PC, DEX) mapping from native PC offset to DEX offset.
-  SwapSrcMap* src_mapping_table_;
+  const LengthPrefixedArray<SrcMapElem>* const src_mapping_table_;
   // For quick code, a uleb128 encoded map from native PC offset to dex PC aswell as dex PC to
   // native PC offset. Size prefixed.
-  SwapVector<uint8_t>* mapping_table_;
+  const LengthPrefixedArray<uint8_t>* const mapping_table_;
   // For quick code, a uleb128 encoded map from GPR/FPR register to dex register. Size prefixed.
-  SwapVector<uint8_t>* vmap_table_;
+  const LengthPrefixedArray<uint8_t>* const vmap_table_;
   // For quick code, a map keyed by native PC indices to bitmaps describing what dalvik registers
   // are live.
-  SwapVector<uint8_t>* gc_map_;
+  const LengthPrefixedArray<uint8_t>* const gc_map_;
   // For quick code, a FDE entry for the debug_frame section.
-  SwapVector<uint8_t>* cfi_info_;
+  const LengthPrefixedArray<uint8_t>* const cfi_info_;
   // For quick code, linker patches needed by the method.
-  const SwapVector<LinkerPatch> patches_;
+  const LengthPrefixedArray<LinkerPatch>* const patches_;
 };
 
 }  // namespace art
diff --git a/compiler/compiled_method_test.cc b/compiler/compiled_method_test.cc
new file mode 100644
index 0000000..99ee875
--- /dev/null
+++ b/compiler/compiled_method_test.cc
@@ -0,0 +1,120 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <gtest/gtest.h>
+
+#include "compiled_method.h"
+
+namespace art {
+
+TEST(CompiledMethod, SrcMapElemOperators) {
+  SrcMapElem elems[] = {
+      { 1u, -1 },
+      { 1u, 0 },
+      { 1u, 1 },
+      { 2u, -1 },
+      { 2u, 0 },    // Index 4.
+      { 2u, 1 },
+      { 2u, 0u },   // Index 6: Arbitrarily add identical SrcMapElem with index 4.
+  };
+
+  for (size_t i = 0; i != arraysize(elems); ++i) {
+    for (size_t j = 0; j != arraysize(elems); ++j) {
+      bool expected = (i != 6u ? i : 4u) == (j != 6u ? j : 4u);
+      EXPECT_EQ(expected, elems[i] == elems[j]) << i << " " << j;
+    }
+  }
+
+  for (size_t i = 0; i != arraysize(elems); ++i) {
+    for (size_t j = 0; j != arraysize(elems); ++j) {
+      bool expected = (i != 6u ? i : 4u) < (j != 6u ? j : 4u);
+      EXPECT_EQ(expected, elems[i] < elems[j]) << i << " " << j;
+    }
+  }
+}
+
+TEST(CompiledMethod, LinkerPatchOperators) {
+  const DexFile* dex_file1 = reinterpret_cast<const DexFile*>(1);
+  const DexFile* dex_file2 = reinterpret_cast<const DexFile*>(2);
+  LinkerPatch patches[] = {
+      LinkerPatch::MethodPatch(16u, dex_file1, 1000u),
+      LinkerPatch::MethodPatch(16u, dex_file1, 1001u),
+      LinkerPatch::MethodPatch(16u, dex_file2, 1000u),
+      LinkerPatch::MethodPatch(16u, dex_file2, 1001u),  // Index 3.
+      LinkerPatch::CodePatch(16u, dex_file1, 1000u),
+      LinkerPatch::CodePatch(16u, dex_file1, 1001u),
+      LinkerPatch::CodePatch(16u, dex_file2, 1000u),
+      LinkerPatch::CodePatch(16u, dex_file2, 1001u),
+      LinkerPatch::RelativeCodePatch(16u, dex_file1, 1000u),
+      LinkerPatch::RelativeCodePatch(16u, dex_file1, 1001u),
+      LinkerPatch::RelativeCodePatch(16u, dex_file2, 1000u),
+      LinkerPatch::RelativeCodePatch(16u, dex_file2, 1001u),
+      LinkerPatch::TypePatch(16u, dex_file1, 1000u),
+      LinkerPatch::TypePatch(16u, dex_file1, 1001u),
+      LinkerPatch::TypePatch(16u, dex_file2, 1000u),
+      LinkerPatch::TypePatch(16u, dex_file2, 1001u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file1, 3000u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file1, 3001u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file1, 3000u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file1, 3001u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file2, 3000u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file2, 3001u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file2, 3000u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(16u, dex_file2, 3001u, 2001u),
+      LinkerPatch::MethodPatch(32u, dex_file1, 1000u),
+      LinkerPatch::MethodPatch(32u, dex_file1, 1001u),
+      LinkerPatch::MethodPatch(32u, dex_file2, 1000u),
+      LinkerPatch::MethodPatch(32u, dex_file2, 1001u),
+      LinkerPatch::CodePatch(32u, dex_file1, 1000u),
+      LinkerPatch::CodePatch(32u, dex_file1, 1001u),
+      LinkerPatch::CodePatch(32u, dex_file2, 1000u),
+      LinkerPatch::CodePatch(32u, dex_file2, 1001u),
+      LinkerPatch::RelativeCodePatch(32u, dex_file1, 1000u),
+      LinkerPatch::RelativeCodePatch(32u, dex_file1, 1001u),
+      LinkerPatch::RelativeCodePatch(32u, dex_file2, 1000u),
+      LinkerPatch::RelativeCodePatch(32u, dex_file2, 1001u),
+      LinkerPatch::TypePatch(32u, dex_file1, 1000u),
+      LinkerPatch::TypePatch(32u, dex_file1, 1001u),
+      LinkerPatch::TypePatch(32u, dex_file2, 1000u),
+      LinkerPatch::TypePatch(32u, dex_file2, 1001u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file1, 3000u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file1, 3001u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file1, 3000u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file1, 3001u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file2, 3000u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file2, 3001u, 2000u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file2, 3000u, 2001u),
+      LinkerPatch::DexCacheArrayPatch(32u, dex_file2, 3001u, 2001u),
+      LinkerPatch::MethodPatch(16u, dex_file2, 1001u),  // identical with patch as index 3.
+  };
+  constexpr size_t last_index = arraysize(patches) - 1u;
+
+  for (size_t i = 0; i != arraysize(patches); ++i) {
+    for (size_t j = 0; j != arraysize(patches); ++j) {
+      bool expected = (i != last_index ? i : 3u) == (j != last_index ? j : 3u);
+      EXPECT_EQ(expected, patches[i] == patches[j]) << i << " " << j;
+    }
+  }
+
+  for (size_t i = 0; i != arraysize(patches); ++i) {
+    for (size_t j = 0; j != arraysize(patches); ++j) {
+      bool expected = (i != last_index ? i : 3u) < (j != last_index ? j : 3u);
+      EXPECT_EQ(expected, patches[i] < patches[j]) << i << " " << j;
+    }
+  }
+}
+
+}  // namespace art
diff --git a/compiler/dex/dex_to_dex_compiler.cc b/compiler/dex/dex_to_dex_compiler.cc
index ff7ddc1..4836041 100644
--- a/compiler/dex/dex_to_dex_compiler.cc
+++ b/compiler/dex/dex_to_dex_compiler.cc
@@ -356,7 +356,7 @@
         0,
         0,
         0,
-        nullptr,                                     // src_mapping_table
+        ArrayRef<const SrcMapElem>(),                // src_mapping_table
         ArrayRef<const uint8_t>(),                   // mapping_table
         ArrayRef<const uint8_t>(builder.GetData()),  // vmap_table
         ArrayRef<const uint8_t>(),                   // gc_map
diff --git a/compiler/dex/quick/codegen_util.cc b/compiler/dex/quick/codegen_util.cc
index cde99b3..d68835a 100644
--- a/compiler/dex/quick/codegen_util.cc
+++ b/compiler/dex/quick/codegen_util.cc
@@ -22,6 +22,7 @@
 #endif
 
 #include "base/bit_vector-inl.h"
+#include "base/stringprintf.h"
 #include "dex/mir_graph.h"
 #include "driver/compiler_driver.h"
 #include "driver/compiler_options.h"
@@ -1165,7 +1166,7 @@
       cu_->compiler_driver, cu_->instruction_set,
       ArrayRef<const uint8_t>(code_buffer_),
       frame_size_, core_spill_mask_, fp_spill_mask_,
-      &src_mapping_table_,
+      ArrayRef<const SrcMapElem>(src_mapping_table_),
       ArrayRef<const uint8_t>(encoded_mapping_table_),
       ArrayRef<const uint8_t>(vmap_encoder.GetData()),
       ArrayRef<const uint8_t>(native_gc_map_),
diff --git a/compiler/dex/quick/ralloc_util.cc b/compiler/dex/quick/ralloc_util.cc
index d9d0434..dceb118 100644
--- a/compiler/dex/quick/ralloc_util.cc
+++ b/compiler/dex/quick/ralloc_util.cc
@@ -18,6 +18,7 @@
 
 #include "mir_to_lir-inl.h"
 
+#include "base/stringprintf.h"
 #include "dex/compiler_ir.h"
 #include "dex/dataflow_iterator-inl.h"
 #include "dex/mir_graph.h"
diff --git a/compiler/driver/compiled_method_storage.cc b/compiler/driver/compiled_method_storage.cc
new file mode 100644
index 0000000..bc5c6ca
--- /dev/null
+++ b/compiler/driver/compiled_method_storage.cc
@@ -0,0 +1,269 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <algorithm>
+#include <ostream>
+
+#include "compiled_method_storage.h"
+
+#include "base/logging.h"
+#include "compiled_method.h"
+#include "thread-inl.h"
+#include "utils.h"
+#include "utils/dedupe_set-inl.h"
+#include "utils/swap_space.h"
+
+namespace art {
+
+namespace {  // anonymous namespace
+
+template <typename T>
+const LengthPrefixedArray<T>* CopyArray(SwapSpace* swap_space, const ArrayRef<const T>& array) {
+  DCHECK(!array.empty());
+  SwapAllocator<uint8_t> allocator(swap_space);
+  void* storage = allocator.allocate(LengthPrefixedArray<T>::ComputeSize(array.size()));
+  LengthPrefixedArray<T>* array_copy = new(storage) LengthPrefixedArray<T>(array.size());
+  std::copy(array.begin(), array.end(), array_copy->begin());
+  return array_copy;
+}
+
+template <typename T>
+void ReleaseArray(SwapSpace* swap_space, const LengthPrefixedArray<T>* array) {
+  SwapAllocator<uint8_t> allocator(swap_space);
+  size_t size = LengthPrefixedArray<T>::ComputeSize(array->size());
+  array->~LengthPrefixedArray<T>();
+  allocator.deallocate(const_cast<uint8_t*>(reinterpret_cast<const uint8_t*>(array)), size);
+}
+
+}  // anonymous namespace
+
+template <typename T, typename DedupeSetType>
+inline const LengthPrefixedArray<T>* CompiledMethodStorage::AllocateOrDeduplicateArray(
+    const ArrayRef<const T>& data,
+    DedupeSetType* dedupe_set) {
+  if (data.empty()) {
+    return nullptr;
+  } else if (!DedupeEnabled()) {
+    return CopyArray(swap_space_.get(), data);
+  } else {
+    return dedupe_set->Add(Thread::Current(), data);
+  }
+}
+
+template <typename T>
+inline void CompiledMethodStorage::ReleaseArrayIfNotDeduplicated(
+    const LengthPrefixedArray<T>* array) {
+  if (array != nullptr && !DedupeEnabled()) {
+    ReleaseArray(swap_space_.get(), array);
+  }
+}
+
+template <typename ContentType>
+class CompiledMethodStorage::DedupeHashFunc {
+ private:
+  static constexpr bool kUseMurmur3Hash = true;
+
+ public:
+  size_t operator()(const ArrayRef<ContentType>& array) const {
+    const uint8_t* data = reinterpret_cast<const uint8_t*>(array.data());
+    // TODO: More reasonable assertion.
+    // static_assert(IsPowerOfTwo(sizeof(ContentType)),
+    //    "ContentType is not power of two, don't know whether array layout is as assumed");
+    uint32_t len = sizeof(ContentType) * array.size();
+    if (kUseMurmur3Hash) {
+      static constexpr uint32_t c1 = 0xcc9e2d51;
+      static constexpr uint32_t c2 = 0x1b873593;
+      static constexpr uint32_t r1 = 15;
+      static constexpr uint32_t r2 = 13;
+      static constexpr uint32_t m = 5;
+      static constexpr uint32_t n = 0xe6546b64;
+
+      uint32_t hash = 0;
+
+      const int nblocks = len / 4;
+      typedef __attribute__((__aligned__(1))) uint32_t unaligned_uint32_t;
+      const unaligned_uint32_t *blocks = reinterpret_cast<const uint32_t*>(data);
+      int i;
+      for (i = 0; i < nblocks; i++) {
+        uint32_t k = blocks[i];
+        k *= c1;
+        k = (k << r1) | (k >> (32 - r1));
+        k *= c2;
+
+        hash ^= k;
+        hash = ((hash << r2) | (hash >> (32 - r2))) * m + n;
+      }
+
+      const uint8_t *tail = reinterpret_cast<const uint8_t*>(data + nblocks * 4);
+      uint32_t k1 = 0;
+
+      switch (len & 3) {
+        case 3:
+          k1 ^= tail[2] << 16;
+          FALLTHROUGH_INTENDED;
+        case 2:
+          k1 ^= tail[1] << 8;
+          FALLTHROUGH_INTENDED;
+        case 1:
+          k1 ^= tail[0];
+
+          k1 *= c1;
+          k1 = (k1 << r1) | (k1 >> (32 - r1));
+          k1 *= c2;
+          hash ^= k1;
+      }
+
+      hash ^= len;
+      hash ^= (hash >> 16);
+      hash *= 0x85ebca6b;
+      hash ^= (hash >> 13);
+      hash *= 0xc2b2ae35;
+      hash ^= (hash >> 16);
+
+      return hash;
+    } else {
+      size_t hash = 0x811c9dc5;
+      for (uint32_t i = 0; i < len; ++i) {
+        hash = (hash * 16777619) ^ data[i];
+      }
+      hash += hash << 13;
+      hash ^= hash >> 7;
+      hash += hash << 3;
+      hash ^= hash >> 17;
+      hash += hash << 5;
+      return hash;
+    }
+  }
+};
+
+template <typename T>
+class CompiledMethodStorage::LengthPrefixedArrayAlloc {
+ public:
+  explicit LengthPrefixedArrayAlloc(SwapSpace* swap_space)
+      : swap_space_(swap_space) {
+  }
+
+  const LengthPrefixedArray<T>* Copy(const ArrayRef<const T>& array) {
+    return CopyArray(swap_space_, array);
+  }
+
+  void Destroy(const LengthPrefixedArray<T>* array) {
+    ReleaseArray(swap_space_, array);
+  }
+
+ private:
+  SwapSpace* const swap_space_;
+};
+
+CompiledMethodStorage::CompiledMethodStorage(int swap_fd)
+    : swap_space_(swap_fd == -1 ? nullptr : new SwapSpace(swap_fd, 10 * MB)),
+      dedupe_enabled_(true),
+      dedupe_code_("dedupe code", LengthPrefixedArrayAlloc<uint8_t>(swap_space_.get())),
+      dedupe_src_mapping_table_("dedupe source mapping table",
+                                LengthPrefixedArrayAlloc<SrcMapElem>(swap_space_.get())),
+      dedupe_mapping_table_("dedupe mapping table",
+                            LengthPrefixedArrayAlloc<uint8_t>(swap_space_.get())),
+      dedupe_vmap_table_("dedupe vmap table",
+                         LengthPrefixedArrayAlloc<uint8_t>(swap_space_.get())),
+      dedupe_gc_map_("dedupe gc map", LengthPrefixedArrayAlloc<uint8_t>(swap_space_.get())),
+      dedupe_cfi_info_("dedupe cfi info", LengthPrefixedArrayAlloc<uint8_t>(swap_space_.get())),
+      dedupe_linker_patches_("dedupe cfi info",
+                             LengthPrefixedArrayAlloc<LinkerPatch>(swap_space_.get())) {
+}
+
+CompiledMethodStorage::~CompiledMethodStorage() {
+  // All done by member destructors.
+}
+
+void CompiledMethodStorage::DumpMemoryUsage(std::ostream& os, bool extended) const {
+  if (swap_space_.get() != nullptr) {
+    os << " swap=" << PrettySize(swap_space_->GetSize());
+  }
+  if (extended) {
+    Thread* self = Thread::Current();
+    os << "\nCode dedupe: " << dedupe_code_.DumpStats(self);
+    os << "\nMapping table dedupe: " << dedupe_mapping_table_.DumpStats(self);
+    os << "\nVmap table dedupe: " << dedupe_vmap_table_.DumpStats(self);
+    os << "\nGC map dedupe: " << dedupe_gc_map_.DumpStats(self);
+    os << "\nCFI info dedupe: " << dedupe_cfi_info_.DumpStats(self);
+  }
+}
+
+const LengthPrefixedArray<uint8_t>* CompiledMethodStorage::DeduplicateCode(
+    const ArrayRef<const uint8_t>& code) {
+  return AllocateOrDeduplicateArray(code, &dedupe_code_);
+}
+
+void CompiledMethodStorage::ReleaseCode(const LengthPrefixedArray<uint8_t>* code) {
+  ReleaseArrayIfNotDeduplicated(code);
+}
+
+const LengthPrefixedArray<SrcMapElem>* CompiledMethodStorage::DeduplicateSrcMappingTable(
+    const ArrayRef<const SrcMapElem>& src_map) {
+  return AllocateOrDeduplicateArray(src_map, &dedupe_src_mapping_table_);
+}
+
+void CompiledMethodStorage::ReleaseSrcMappingTable(const LengthPrefixedArray<SrcMapElem>* src_map) {
+  ReleaseArrayIfNotDeduplicated(src_map);
+}
+
+const LengthPrefixedArray<uint8_t>* CompiledMethodStorage::DeduplicateMappingTable(
+    const ArrayRef<const uint8_t>& table) {
+  return AllocateOrDeduplicateArray(table, &dedupe_mapping_table_);
+}
+
+void CompiledMethodStorage::ReleaseMappingTable(const LengthPrefixedArray<uint8_t>* table) {
+  ReleaseArrayIfNotDeduplicated(table);
+}
+
+const LengthPrefixedArray<uint8_t>* CompiledMethodStorage::DeduplicateVMapTable(
+    const ArrayRef<const uint8_t>& table) {
+  return AllocateOrDeduplicateArray(table, &dedupe_vmap_table_);
+}
+
+void CompiledMethodStorage::ReleaseVMapTable(const LengthPrefixedArray<uint8_t>* table) {
+  ReleaseArrayIfNotDeduplicated(table);
+}
+
+const LengthPrefixedArray<uint8_t>* CompiledMethodStorage::DeduplicateGCMap(
+    const ArrayRef<const uint8_t>& gc_map) {
+  return AllocateOrDeduplicateArray(gc_map, &dedupe_gc_map_);
+}
+
+void CompiledMethodStorage::ReleaseGCMap(const LengthPrefixedArray<uint8_t>* gc_map) {
+  ReleaseArrayIfNotDeduplicated(gc_map);
+}
+
+const LengthPrefixedArray<uint8_t>* CompiledMethodStorage::DeduplicateCFIInfo(
+    const ArrayRef<const uint8_t>& cfi_info) {
+  return AllocateOrDeduplicateArray(cfi_info, &dedupe_cfi_info_);
+}
+
+void CompiledMethodStorage::ReleaseCFIInfo(const LengthPrefixedArray<uint8_t>* cfi_info) {
+  ReleaseArrayIfNotDeduplicated(cfi_info);
+}
+
+const LengthPrefixedArray<LinkerPatch>* CompiledMethodStorage::DeduplicateLinkerPatches(
+    const ArrayRef<const LinkerPatch>& linker_patches) {
+  return AllocateOrDeduplicateArray(linker_patches, &dedupe_linker_patches_);
+}
+
+void CompiledMethodStorage::ReleaseLinkerPatches(
+    const LengthPrefixedArray<LinkerPatch>* linker_patches) {
+  ReleaseArrayIfNotDeduplicated(linker_patches);
+}
+
+}  // namespace art
diff --git a/compiler/driver/compiled_method_storage.h b/compiler/driver/compiled_method_storage.h
new file mode 100644
index 0000000..ef10b67
--- /dev/null
+++ b/compiler/driver/compiled_method_storage.h
@@ -0,0 +1,117 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_DRIVER_COMPILED_METHOD_STORAGE_H_
+#define ART_COMPILER_DRIVER_COMPILED_METHOD_STORAGE_H_
+
+#include <iosfwd>
+#include <memory>
+
+#include "base/macros.h"
+#include "length_prefixed_array.h"
+#include "utils/array_ref.h"
+#include "utils/dedupe_set.h"
+#include "utils/swap_space.h"
+
+namespace art {
+
+class LinkerPatch;
+class SrcMapElem;
+
+class CompiledMethodStorage {
+ public:
+  explicit CompiledMethodStorage(int swap_fd);
+  ~CompiledMethodStorage();
+
+  void DumpMemoryUsage(std::ostream& os, bool extended) const;
+
+  void SetDedupeEnabled(bool dedupe_enabled) {
+    dedupe_enabled_ = dedupe_enabled;
+  }
+  bool DedupeEnabled() const {
+    return dedupe_enabled_;
+  }
+
+  SwapAllocator<void> GetSwapSpaceAllocator() {
+    return SwapAllocator<void>(swap_space_.get());
+  }
+
+  const LengthPrefixedArray<uint8_t>* DeduplicateCode(const ArrayRef<const uint8_t>& code);
+  void ReleaseCode(const LengthPrefixedArray<uint8_t>* code);
+
+  const LengthPrefixedArray<SrcMapElem>* DeduplicateSrcMappingTable(
+      const ArrayRef<const SrcMapElem>& src_map);
+  void ReleaseSrcMappingTable(const LengthPrefixedArray<SrcMapElem>* src_map);
+
+  const LengthPrefixedArray<uint8_t>* DeduplicateMappingTable(const ArrayRef<const uint8_t>& table);
+  void ReleaseMappingTable(const LengthPrefixedArray<uint8_t>* table);
+
+  const LengthPrefixedArray<uint8_t>* DeduplicateVMapTable(const ArrayRef<const uint8_t>& table);
+  void ReleaseVMapTable(const LengthPrefixedArray<uint8_t>* table);
+
+  const LengthPrefixedArray<uint8_t>* DeduplicateGCMap(const ArrayRef<const uint8_t>& gc_map);
+  void ReleaseGCMap(const LengthPrefixedArray<uint8_t>* gc_map);
+
+  const LengthPrefixedArray<uint8_t>* DeduplicateCFIInfo(const ArrayRef<const uint8_t>& cfi_info);
+  void ReleaseCFIInfo(const LengthPrefixedArray<uint8_t>* cfi_info);
+
+  const LengthPrefixedArray<LinkerPatch>* DeduplicateLinkerPatches(
+      const ArrayRef<const LinkerPatch>& linker_patches);
+  void ReleaseLinkerPatches(const LengthPrefixedArray<LinkerPatch>* linker_patches);
+
+ private:
+  template <typename T, typename DedupeSetType>
+  const LengthPrefixedArray<T>* AllocateOrDeduplicateArray(const ArrayRef<const T>& data,
+                                                           DedupeSetType* dedupe_set);
+
+  template <typename T>
+  void ReleaseArrayIfNotDeduplicated(const LengthPrefixedArray<T>* array);
+
+  // DeDuplication data structures.
+  template <typename ContentType>
+  class DedupeHashFunc;
+
+  template <typename T>
+  class LengthPrefixedArrayAlloc;
+
+  template <typename T>
+  using ArrayDedupeSet = DedupeSet<ArrayRef<const T>,
+                                   LengthPrefixedArray<T>,
+                                   LengthPrefixedArrayAlloc<T>,
+                                   size_t,
+                                   DedupeHashFunc<const T>,
+                                   4>;
+
+  // Swap pool and allocator used for native allocations. May be file-backed. Needs to be first
+  // as other fields rely on this.
+  std::unique_ptr<SwapSpace> swap_space_;
+
+  bool dedupe_enabled_;
+
+  ArrayDedupeSet<uint8_t> dedupe_code_;
+  ArrayDedupeSet<SrcMapElem> dedupe_src_mapping_table_;
+  ArrayDedupeSet<uint8_t> dedupe_mapping_table_;
+  ArrayDedupeSet<uint8_t> dedupe_vmap_table_;
+  ArrayDedupeSet<uint8_t> dedupe_gc_map_;
+  ArrayDedupeSet<uint8_t> dedupe_cfi_info_;
+  ArrayDedupeSet<LinkerPatch> dedupe_linker_patches_;
+
+  DISALLOW_COPY_AND_ASSIGN(CompiledMethodStorage);
+};
+
+}  // namespace art
+
+#endif  // ART_COMPILER_DRIVER_COMPILED_METHOD_STORAGE_H_
diff --git a/compiler/driver/compiled_method_storage_test.cc b/compiler/driver/compiled_method_storage_test.cc
new file mode 100644
index 0000000..c6dbd24
--- /dev/null
+++ b/compiler/driver/compiled_method_storage_test.cc
@@ -0,0 +1,160 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <gtest/gtest.h>
+
+#include "compiled_method_storage.h"
+#include "compiled_method.h"
+#include "compiler_driver.h"
+#include "compiler_options.h"
+#include "dex/verification_results.h"
+#include "dex/quick/dex_file_to_method_inliner_map.h"
+
+namespace art {
+
+TEST(CompiledMethodStorage, Deduplicate) {
+  CompilerOptions compiler_options;
+  VerificationResults verification_results(&compiler_options);
+  DexFileToMethodInlinerMap method_inliner_map;
+  CompilerDriver driver(&compiler_options,
+                        &verification_results,
+                        &method_inliner_map,
+                        Compiler::kOptimizing, kNone,
+                        nullptr,
+                        false,
+                        nullptr,
+                        nullptr,
+                        nullptr,
+                        1u,
+                        false,
+                        false,
+                        "",
+                        false,
+                        nullptr,
+                        -1,
+                        "");
+  CompiledMethodStorage* storage = driver.GetCompiledMethodStorage();
+
+  ASSERT_TRUE(storage->DedupeEnabled());  // The default.
+
+  const uint8_t raw_code1[] = { 1u, 2u, 3u };
+  const uint8_t raw_code2[] = { 4u, 3u, 2u, 1u };
+  ArrayRef<const uint8_t> code[] = {
+      ArrayRef<const uint8_t>(raw_code1),
+      ArrayRef<const uint8_t>(raw_code2),
+  };
+  const SrcMapElem raw_src_map1[] = { { 1u, 2u }, { 3u, 4u }, { 5u, 6u } };
+  const SrcMapElem raw_src_map2[] = { { 8u, 7u }, { 6u, 5u }, { 4u, 3u }, { 2u, 1u } };
+  ArrayRef<const SrcMapElem> src_map[] = {
+      ArrayRef<const SrcMapElem>(raw_src_map1),
+      ArrayRef<const SrcMapElem>(raw_src_map2),
+  };
+  const uint8_t raw_mapping_table1[] = { 5, 6, 7 };
+  const uint8_t raw_mapping_table2[] = { 7, 6, 5, 4 };
+  ArrayRef<const uint8_t> mapping_table[] = {
+      ArrayRef<const uint8_t>(raw_mapping_table1),
+      ArrayRef<const uint8_t>(raw_mapping_table2),
+  };
+  const uint8_t raw_vmap_table1[] = { 2, 4, 6 };
+  const uint8_t raw_vmap_table2[] = { 7, 5, 3, 1 };
+  ArrayRef<const uint8_t> vmap_table[] = {
+      ArrayRef<const uint8_t>(raw_vmap_table1),
+      ArrayRef<const uint8_t>(raw_vmap_table2),
+  };
+  const uint8_t raw_gc_map1[] = { 9, 8, 7 };
+  const uint8_t raw_gc_map2[] = { 6, 7, 8, 9 };
+  ArrayRef<const uint8_t> gc_map[] = {
+      ArrayRef<const uint8_t>(raw_gc_map1),
+      ArrayRef<const uint8_t>(raw_gc_map2),
+  };
+  const uint8_t raw_cfi_info1[] = { 1, 3, 5 };
+  const uint8_t raw_cfi_info2[] = { 8, 6, 4, 2 };
+  ArrayRef<const uint8_t> cfi_info[] = {
+      ArrayRef<const uint8_t>(raw_cfi_info1),
+      ArrayRef<const uint8_t>(raw_cfi_info2),
+  };
+  const LinkerPatch raw_patches1[] = {
+      LinkerPatch::CodePatch(0u, nullptr, 1u),
+      LinkerPatch::MethodPatch(4u, nullptr, 1u),
+  };
+  const LinkerPatch raw_patches2[] = {
+      LinkerPatch::CodePatch(0u, nullptr, 1u),
+      LinkerPatch::MethodPatch(4u, nullptr, 2u),
+  };
+  ArrayRef<const LinkerPatch> patches[] = {
+      ArrayRef<const LinkerPatch>(raw_patches1),
+      ArrayRef<const LinkerPatch>(raw_patches2),
+  };
+
+  std::vector<CompiledMethod*> compiled_methods;
+  compiled_methods.reserve(1u << 7);
+  for (auto&& c : code) {
+    for (auto&& s : src_map) {
+      for (auto&& m : mapping_table) {
+        for (auto&& v : vmap_table) {
+          for (auto&& g : gc_map) {
+            for (auto&& f : cfi_info) {
+              for (auto&& p : patches) {
+                compiled_methods.push_back(CompiledMethod::SwapAllocCompiledMethod(
+                        &driver, kNone, c, 0u, 0u, 0u, s, m, v, g, f, p));
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  constexpr size_t code_bit = 1u << 6;
+  constexpr size_t src_map_bit = 1u << 5;
+  constexpr size_t mapping_table_bit = 1u << 4;
+  constexpr size_t vmap_table_bit = 1u << 3;
+  constexpr size_t gc_map_bit = 1u << 2;
+  constexpr size_t cfi_info_bit = 1u << 1;
+  constexpr size_t patches_bit = 1u << 0;
+  CHECK_EQ(compiled_methods.size(), 1u << 7);
+  for (size_t i = 0; i != compiled_methods.size(); ++i) {
+    for (size_t j = 0; j != compiled_methods.size(); ++j) {
+      CompiledMethod* lhs = compiled_methods[i];
+      CompiledMethod* rhs = compiled_methods[j];
+      bool same_code = ((i ^ j) & code_bit) == 0u;
+      bool same_src_map = ((i ^ j) & src_map_bit) == 0u;
+      bool same_mapping_table = ((i ^ j) & mapping_table_bit) == 0u;
+      bool same_vmap_table = ((i ^ j) & vmap_table_bit) == 0u;
+      bool same_gc_map = ((i ^ j) & gc_map_bit) == 0u;
+      bool same_cfi_info = ((i ^ j) & cfi_info_bit) == 0u;
+      bool same_patches = ((i ^ j) & patches_bit) == 0u;
+      ASSERT_EQ(same_code, lhs->GetQuickCode().data() == rhs->GetQuickCode().data())
+          << i << " " << j;
+      ASSERT_EQ(same_src_map, lhs->GetSrcMappingTable().data() == rhs->GetSrcMappingTable().data())
+          << i << " " << j;
+      ASSERT_EQ(same_mapping_table, lhs->GetMappingTable().data() == rhs->GetMappingTable().data())
+          << i << " " << j;
+      ASSERT_EQ(same_vmap_table, lhs->GetVmapTable().data() == rhs->GetVmapTable().data())
+          << i << " " << j;
+      ASSERT_EQ(same_gc_map, lhs->GetGcMap().data() == rhs->GetGcMap().data())
+          << i << " " << j;
+      ASSERT_EQ(same_cfi_info, lhs->GetCFIInfo().data() == rhs->GetCFIInfo().data())
+          << i << " " << j;
+      ASSERT_EQ(same_patches, lhs->GetPatches().data() == rhs->GetPatches().data())
+          << i << " " << j;
+    }
+  }
+  for (CompiledMethod* method : compiled_methods) {
+    CompiledMethod::ReleaseSwapAllocatedCompiledMethod(&driver, method);
+  }
+}
+
+}  // namespace art
diff --git a/compiler/driver/compiler_driver.cc b/compiler/driver/compiler_driver.cc
index b956584..8750aa8 100644
--- a/compiler/driver/compiler_driver.cc
+++ b/compiler/driver/compiler_driver.cc
@@ -348,9 +348,8 @@
                                const std::string& dump_cfg_file_name, bool dump_cfg_append,
                                CumulativeLogger* timer, int swap_fd,
                                const std::string& profile_file)
-    : swap_space_(swap_fd == -1 ? nullptr : new SwapSpace(swap_fd, 10 * MB)),
-      swap_space_allocator_(new SwapAllocator<void>(swap_space_.get())),
-      profile_present_(false), compiler_options_(compiler_options),
+    : profile_present_(false),
+      compiler_options_(compiler_options),
       verification_results_(verification_results),
       method_inliner_map_(method_inliner_map),
       compiler_(Compiler::Create(this, compiler_kind)),
@@ -369,7 +368,6 @@
       had_hard_verifier_failure_(false),
       thread_count_(thread_count),
       stats_(new AOTCompilationStats),
-      dedupe_enabled_(true),
       dump_stats_(dump_stats),
       dump_passes_(dump_passes),
       dump_cfg_file_name_(dump_cfg_file_name),
@@ -377,12 +375,7 @@
       timings_logger_(timer),
       compiler_context_(nullptr),
       support_boot_image_fixup_(instruction_set != kMips && instruction_set != kMips64),
-      dedupe_code_("dedupe code", *swap_space_allocator_),
-      dedupe_src_mapping_table_("dedupe source mapping table", *swap_space_allocator_),
-      dedupe_mapping_table_("dedupe mapping table", *swap_space_allocator_),
-      dedupe_vmap_table_("dedupe vmap table", *swap_space_allocator_),
-      dedupe_gc_map_("dedupe gc map", *swap_space_allocator_),
-      dedupe_cfi_info_("dedupe cfi info", *swap_space_allocator_) {
+      compiled_method_storage_(swap_fd) {
   DCHECK(compiler_options_ != nullptr);
   DCHECK(verification_results_ != nullptr);
   DCHECK(method_inliner_map_ != nullptr);
@@ -402,36 +395,6 @@
   }
 }
 
-SwapVector<uint8_t>* CompilerDriver::DeduplicateCode(const ArrayRef<const uint8_t>& code) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_code_.Add(Thread::Current(), code);
-}
-
-SwapSrcMap* CompilerDriver::DeduplicateSrcMappingTable(const ArrayRef<SrcMapElem>& src_map) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_src_mapping_table_.Add(Thread::Current(), src_map);
-}
-
-SwapVector<uint8_t>* CompilerDriver::DeduplicateMappingTable(const ArrayRef<const uint8_t>& code) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_mapping_table_.Add(Thread::Current(), code);
-}
-
-SwapVector<uint8_t>* CompilerDriver::DeduplicateVMapTable(const ArrayRef<const uint8_t>& code) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_vmap_table_.Add(Thread::Current(), code);
-}
-
-SwapVector<uint8_t>* CompilerDriver::DeduplicateGCMap(const ArrayRef<const uint8_t>& code) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_gc_map_.Add(Thread::Current(), code);
-}
-
-SwapVector<uint8_t>* CompilerDriver::DeduplicateCFIInfo(const ArrayRef<const uint8_t>& cfi_info) {
-  DCHECK(dedupe_enabled_);
-  return dedupe_cfi_info_.Add(Thread::Current(), cfi_info);
-}
-
 CompilerDriver::~CompilerDriver() {
   Thread* self = Thread::Current();
   {
@@ -447,6 +410,7 @@
   compiler_->UnInit();
 }
 
+
 #define CREATE_TRAMPOLINE(type, abi, offset) \
     if (Is64BitInstructionSet(instruction_set_)) { \
       return CreateTrampoline64(instruction_set_, abi, \
@@ -2642,16 +2606,7 @@
   oss << " native alloc=" << PrettySize(allocated_space) << " free="
       << PrettySize(free_space);
 #endif
-  if (swap_space_.get() != nullptr) {
-    oss << " swap=" << PrettySize(swap_space_->GetSize());
-  }
-  if (extended) {
-    oss << "\nCode dedupe: " << dedupe_code_.DumpStats();
-    oss << "\nMapping table dedupe: " << dedupe_mapping_table_.DumpStats();
-    oss << "\nVmap table dedupe: " << dedupe_vmap_table_.DumpStats();
-    oss << "\nGC map dedupe: " << dedupe_gc_map_.DumpStats();
-    oss << "\nCFI info dedupe: " << dedupe_cfi_info_.DumpStats();
-  }
+  compiled_method_storage_.DumpMemoryUsage(oss, extended);
   return oss.str();
 }
 
diff --git a/compiler/driver/compiler_driver.h b/compiler/driver/compiler_driver.h
index 0dc8261..485cdcf 100644
--- a/compiler/driver/compiler_driver.h
+++ b/compiler/driver/compiler_driver.h
@@ -30,6 +30,7 @@
 #include "class_reference.h"
 #include "compiler.h"
 #include "dex_file.h"
+#include "driver/compiled_method_storage.h"
 #include "invoke_type.h"
 #include "method_reference.h"
 #include "mirror/class.h"  // For mirror::Class::Status.
@@ -38,10 +39,7 @@
 #include "runtime.h"
 #include "safe_map.h"
 #include "thread_pool.h"
-#include "utils/array_ref.h"
-#include "utils/dedupe_set.h"
 #include "utils/dex_cache_arrays_layout.h"
-#include "utils/swap_space.h"
 
 namespace art {
 
@@ -80,8 +78,6 @@
   kQuickAbi
 };
 
-static constexpr bool kUseMurmur3Hash = true;
-
 class CompilerDriver {
  public:
   // Create a compiler targeting the requested "instruction_set".
@@ -388,10 +384,6 @@
     support_boot_image_fixup_ = support_boot_image_fixup;
   }
 
-  SwapAllocator<void>& GetSwapSpaceAllocator() {
-    return *swap_space_allocator_.get();
-  }
-
   bool WriteElf(const std::string& android_root,
                 bool is_host,
                 const std::vector<const DexFile*>& dex_files,
@@ -431,10 +423,10 @@
   }
 
   void SetDedupeEnabled(bool dedupe_enabled) {
-    dedupe_enabled_ = dedupe_enabled;
+    compiled_method_storage_.SetDedupeEnabled(dedupe_enabled);
   }
   bool DedupeEnabled() const {
-    return dedupe_enabled_;
+    return compiled_method_storage_.DedupeEnabled();
   }
 
   // Checks if class specified by type_idx is one of the image_classes_
@@ -455,13 +447,6 @@
                                        uint16_t class_def_idx,
                                        const DexFile& dex_file) const;
 
-  SwapVector<uint8_t>* DeduplicateCode(const ArrayRef<const uint8_t>& code);
-  SwapSrcMap* DeduplicateSrcMappingTable(const ArrayRef<SrcMapElem>& src_map);
-  SwapVector<uint8_t>* DeduplicateMappingTable(const ArrayRef<const uint8_t>& code);
-  SwapVector<uint8_t>* DeduplicateVMapTable(const ArrayRef<const uint8_t>& code);
-  SwapVector<uint8_t>* DeduplicateGCMap(const ArrayRef<const uint8_t>& code);
-  SwapVector<uint8_t>* DeduplicateCFIInfo(const ArrayRef<const uint8_t>& cfi_info);
-
   // Should the compiler run on this method given profile information?
   bool SkipCompilation(const std::string& method_name);
 
@@ -479,6 +464,10 @@
     return compiler_kind_;
   }
 
+  CompiledMethodStorage* GetCompiledMethodStorage() {
+    return &compiled_method_storage_;
+  }
+
  private:
   // Return whether the declaring class of `resolved_member` is
   // available to `referrer_class` for read or write access using two
@@ -599,11 +588,6 @@
                       ThreadPool* thread_pool, TimingLogger* timings)
       REQUIRES(!Locks::mutator_lock_);
 
-  // Swap pool and allocator used for native allocations. May be file-backed. Needs to be first
-  // as other fields rely on this.
-  std::unique_ptr<SwapSpace> swap_space_;
-  std::unique_ptr<SwapAllocator<void> > swap_space_allocator_;
-
   ProfileFile profile_file_;
   bool profile_present_;
 
@@ -663,7 +647,6 @@
   class AOTCompilationStats;
   std::unique_ptr<AOTCompilationStats> stats_;
 
-  bool dedupe_enabled_;
   bool dump_stats_;
   const bool dump_passes_;
   const std::string dump_cfg_file_name_;
@@ -678,93 +661,7 @@
 
   bool support_boot_image_fixup_;
 
-  // DeDuplication data structures, these own the corresponding byte arrays.
-  template <typename ContentType>
-  class DedupeHashFunc {
-   public:
-    size_t operator()(const ArrayRef<ContentType>& array) const {
-      const uint8_t* data = reinterpret_cast<const uint8_t*>(array.data());
-      static_assert(IsPowerOfTwo(sizeof(ContentType)),
-          "ContentType is not power of two, don't know whether array layout is as assumed");
-      uint32_t len = sizeof(ContentType) * array.size();
-      if (kUseMurmur3Hash) {
-        static constexpr uint32_t c1 = 0xcc9e2d51;
-        static constexpr uint32_t c2 = 0x1b873593;
-        static constexpr uint32_t r1 = 15;
-        static constexpr uint32_t r2 = 13;
-        static constexpr uint32_t m = 5;
-        static constexpr uint32_t n = 0xe6546b64;
-
-        uint32_t hash = 0;
-
-        const int nblocks = len / 4;
-        typedef __attribute__((__aligned__(1))) uint32_t unaligned_uint32_t;
-        const unaligned_uint32_t *blocks = reinterpret_cast<const uint32_t*>(data);
-        int i;
-        for (i = 0; i < nblocks; i++) {
-          uint32_t k = blocks[i];
-          k *= c1;
-          k = (k << r1) | (k >> (32 - r1));
-          k *= c2;
-
-          hash ^= k;
-          hash = ((hash << r2) | (hash >> (32 - r2))) * m + n;
-        }
-
-        const uint8_t *tail = reinterpret_cast<const uint8_t*>(data + nblocks * 4);
-        uint32_t k1 = 0;
-
-        switch (len & 3) {
-          case 3:
-            k1 ^= tail[2] << 16;
-            FALLTHROUGH_INTENDED;
-          case 2:
-            k1 ^= tail[1] << 8;
-            FALLTHROUGH_INTENDED;
-          case 1:
-            k1 ^= tail[0];
-
-            k1 *= c1;
-            k1 = (k1 << r1) | (k1 >> (32 - r1));
-            k1 *= c2;
-            hash ^= k1;
-        }
-
-        hash ^= len;
-        hash ^= (hash >> 16);
-        hash *= 0x85ebca6b;
-        hash ^= (hash >> 13);
-        hash *= 0xc2b2ae35;
-        hash ^= (hash >> 16);
-
-        return hash;
-      } else {
-        size_t hash = 0x811c9dc5;
-        for (uint32_t i = 0; i < len; ++i) {
-          hash = (hash * 16777619) ^ data[i];
-        }
-        hash += hash << 13;
-        hash ^= hash >> 7;
-        hash += hash << 3;
-        hash ^= hash >> 17;
-        hash += hash << 5;
-        return hash;
-      }
-    }
-  };
-
-  DedupeSet<ArrayRef<const uint8_t>,
-            SwapVector<uint8_t>, size_t, DedupeHashFunc<const uint8_t>, 4> dedupe_code_;
-  DedupeSet<ArrayRef<SrcMapElem>,
-            SwapSrcMap, size_t, DedupeHashFunc<SrcMapElem>, 4> dedupe_src_mapping_table_;
-  DedupeSet<ArrayRef<const uint8_t>,
-            SwapVector<uint8_t>, size_t, DedupeHashFunc<const uint8_t>, 4> dedupe_mapping_table_;
-  DedupeSet<ArrayRef<const uint8_t>,
-            SwapVector<uint8_t>, size_t, DedupeHashFunc<const uint8_t>, 4> dedupe_vmap_table_;
-  DedupeSet<ArrayRef<const uint8_t>,
-            SwapVector<uint8_t>, size_t, DedupeHashFunc<const uint8_t>, 4> dedupe_gc_map_;
-  DedupeSet<ArrayRef<const uint8_t>,
-            SwapVector<uint8_t>, size_t, DedupeHashFunc<const uint8_t>, 4> dedupe_cfi_info_;
+  CompiledMethodStorage compiled_method_storage_;
 
   friend class CompileClassVisitor;
   DISALLOW_COPY_AND_ASSIGN(CompilerDriver);
diff --git a/compiler/dwarf/dwarf_test.cc b/compiler/dwarf/dwarf_test.cc
index a07d27c..3ba380e 100644
--- a/compiler/dwarf/dwarf_test.cc
+++ b/compiler/dwarf/dwarf_test.cc
@@ -126,7 +126,7 @@
                      initial_opcodes, kCFIFormat, &debug_frame_data_);
   std::vector<uintptr_t> debug_frame_patches;
   std::vector<uintptr_t> expected_patches { 28 };  // NOLINT
-  WriteDebugFrameFDE(is64bit, 0, 0x01000000, 0x01000000, opcodes.data(),
+  WriteDebugFrameFDE(is64bit, 0, 0x01000000, 0x01000000, ArrayRef<const uint8_t>(*opcodes.data()),
                      kCFIFormat, &debug_frame_data_, &debug_frame_patches);
 
   EXPECT_EQ(expected_patches, debug_frame_patches);
@@ -142,7 +142,8 @@
   std::vector<uintptr_t> debug_frame_patches;
   std::vector<uintptr_t> expected_patches { 32 };  // NOLINT
   WriteDebugFrameFDE(is64bit, 0, 0x0100000000000000, 0x0200000000000000,
-                     opcodes.data(), kCFIFormat, &debug_frame_data_, &debug_frame_patches);
+                     ArrayRef<const uint8_t>(*opcodes.data()),
+                     kCFIFormat, &debug_frame_data_, &debug_frame_patches);
   DW_CHECK("FDE cie=00000000 pc=100000000000000..300000000000000");
 
   EXPECT_EQ(expected_patches, debug_frame_patches);
@@ -179,7 +180,8 @@
                      initial_opcodes, kCFIFormat, &debug_frame_data_);
   std::vector<uintptr_t> debug_frame_patches;
   WriteDebugFrameFDE(is64bit, 0, 0x0100000000000000, 0x0200000000000000,
-                     opcodes.data(), kCFIFormat, &debug_frame_data_, &debug_frame_patches);
+                     ArrayRef<const uint8_t>(*opcodes.data()),
+                     kCFIFormat, &debug_frame_data_, &debug_frame_patches);
 
   CheckObjdumpOutput(is64bit, "-W");
 }
diff --git a/compiler/dwarf/headers.h b/compiler/dwarf/headers.h
index b7eff19..f3fba4b 100644
--- a/compiler/dwarf/headers.h
+++ b/compiler/dwarf/headers.h
@@ -25,6 +25,7 @@
 #include "dwarf/dwarf_constants.h"
 #include "dwarf/register.h"
 #include "dwarf/writer.h"
+#include "utils/array_ref.h"
 
 namespace art {
 namespace dwarf {
@@ -70,21 +71,19 @@
       writer.PushUint8(DW_EH_PE_absptr | DW_EH_PE_udata4);  // R: Pointer encoding.
     }
   }
-  writer.PushData(opcodes.data());
+  writer.PushData(*opcodes.data());
   writer.Pad(is64bit ? 8 : 4);
   writer.UpdateUint32(cie_header_start_, writer.data()->size() - cie_header_start_ - 4);
 }
 
 // Write frame description entry (FDE) to .debug_frame or .eh_frame section.
-template<typename Vector>
+inline
 void WriteDebugFrameFDE(bool is64bit, size_t cie_offset,
                         uint64_t initial_address, uint64_t address_range,
-                        const Vector* opcodes,
+                        const ArrayRef<const uint8_t>& opcodes,
                         CFIFormat format,
                         std::vector<uint8_t>* debug_frame,
                         std::vector<uintptr_t>* debug_frame_patches) {
-  static_assert(std::is_same<typename Vector::value_type, uint8_t>::value, "Invalid value type");
-
   Writer<> writer(debug_frame);
   size_t fde_header_start = writer.data()->size();
   writer.PushUint32(0);  // Length placeholder.
@@ -125,7 +124,7 @@
   writer.PushUint32(debug_abbrev_offset);
   writer.PushUint8(entries.Is64bit() ? 8 : 4);
   size_t entries_offset = writer.data()->size();
-  writer.PushData(entries.data());
+  writer.PushData(*entries.data());
   writer.UpdateUint32(start, writer.data()->size() - start - 4);
   // Copy patch locations and make them relative to .debug_info section.
   for (uintptr_t patch_location : entries.GetPatchLocations()) {
@@ -181,7 +180,7 @@
   writer.PushUint8(0);  // Terminate file list.
   writer.UpdateUint32(header_length_pos, writer.data()->size() - header_length_pos - 4);
   size_t opcodes_offset = writer.data()->size();
-  writer.PushData(opcodes.data());
+  writer.PushData(*opcodes.data());
   writer.UpdateUint32(header_start, writer.data()->size() - header_start - 4);
   // Copy patch locations and make them relative to .debug_line section.
   for (uintptr_t patch_location : opcodes.GetPatchLocations()) {
diff --git a/compiler/dwarf/writer.h b/compiler/dwarf/writer.h
index 42c32c4..00b9dfa 100644
--- a/compiler/dwarf/writer.h
+++ b/compiler/dwarf/writer.h
@@ -17,6 +17,7 @@
 #ifndef ART_COMPILER_DWARF_WRITER_H_
 #define ART_COMPILER_DWARF_WRITER_H_
 
+#include <type_traits>
 #include <vector>
 #include "base/bit_utils.h"
 #include "base/logging.h"
@@ -119,9 +120,10 @@
   }
 
   template<typename Vector2>
-  void PushData(const Vector2* buffer) {
-    static_assert(std::is_same<typename Vector2::value_type, uint8_t>::value, "Invalid value type");
-    data_->insert(data_->end(), buffer->begin(), buffer->end());
+  void PushData(const Vector2& buffer) {
+    static_assert(std::is_same<typename std::add_const<typename Vector::value_type>::type,
+                               const uint8_t>::value, "Invalid value type");
+    data_->insert(data_->end(), buffer.begin(), buffer.end());
   }
 
   void UpdateUint32(size_t offset, uint32_t value) {
diff --git a/compiler/elf_writer_debug.cc b/compiler/elf_writer_debug.cc
index c10ffeb..3a9e312 100644
--- a/compiler/elf_writer_debug.cc
+++ b/compiler/elf_writer_debug.cc
@@ -182,8 +182,8 @@
   WriteDebugFrameCIE(isa, address_type, format, debug_frame);
   for (const OatWriter::DebugInfo& mi : method_infos) {
     if (!mi.deduped_) {  // Only one FDE per unique address.
-      const SwapVector<uint8_t>* opcodes = mi.compiled_method_->GetCFIInfo();
-      if (opcodes != nullptr) {
+      ArrayRef<const uint8_t> opcodes = mi.compiled_method_->GetCFIInfo();
+      if (!opcodes.empty()) {
         address_to_fde_offset_map.emplace(mi.low_pc_, debug_frame->size());
         WriteDebugFrameFDE(Is64BitInstructionSet(isa), cie_offset,
                            mi.low_pc_, mi.high_pc_ - mi.low_pc_,
diff --git a/compiler/image_writer.cc b/compiler/image_writer.cc
index 4310be6..0e5a97f 100644
--- a/compiler/image_writer.cc
+++ b/compiler/image_writer.cc
@@ -796,7 +796,7 @@
                   offset, kNativeObjectRelocationTypeArtFieldArray });
           offset += header_size;
           // Forward individual fields so that we can quickly find where they belong.
-          for (size_t i = 0, count = cur_fields->Length(); i < count; ++i) {
+          for (size_t i = 0, count = cur_fields->size(); i < count; ++i) {
             // Need to forward arrays separate of fields.
             ArtField* field = &cur_fields->At(i);
             auto it2 = native_object_relocations_.find(field);
diff --git a/compiler/jit/jit_compiler.cc b/compiler/jit/jit_compiler.cc
index 3d1b42f..b563c80 100644
--- a/compiler/jit/jit_compiler.cc
+++ b/compiler/jit/jit_compiler.cc
@@ -160,7 +160,7 @@
   Runtime* runtime = Runtime::Current();
 
   // Check if the method is already compiled.
-  if (runtime->GetJit()->GetCodeCache()->ContainsMethod(method)) {
+  if (runtime->GetJit()->GetCodeCache()->ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
     VLOG(jit) << "Already compiled " << PrettyMethod(method);
     return true;
   }
@@ -207,10 +207,7 @@
     result = true;
   } else {
     TimingLogger::ScopedTiming t2("LinkCode", &logger);
-    OatFile::OatMethod oat_method(nullptr, 0);
-    if (AddToCodeCache(method, compiled_method, &oat_method)) {
-      oat_method.LinkMethod(method);
-      CHECK(runtime->GetJit()->GetCodeCache()->ContainsMethod(method)) << PrettyMethod(method);
+    if (AddToCodeCache(method, compiled_method)) {
       result = true;
     }
   }
@@ -227,57 +224,57 @@
 }
 
 bool JitCompiler::AddToCodeCache(ArtMethod* method,
-                                 const CompiledMethod* compiled_method,
-                                 OatFile::OatMethod* out_method) {
+                                 const CompiledMethod* compiled_method) {
   Runtime* runtime = Runtime::Current();
   JitCodeCache* const code_cache = runtime->GetJit()->GetCodeCache();
-  const auto* quick_code = compiled_method->GetQuickCode();
-  if (quick_code == nullptr) {
+  auto const quick_code = compiled_method->GetQuickCode();
+  if (quick_code.empty()) {
     return false;
   }
-  const auto code_size = quick_code->size();
+  const auto code_size = quick_code.size();
   Thread* const self = Thread::Current();
-  auto* const mapping_table = compiled_method->GetMappingTable();
-  auto* const vmap_table = compiled_method->GetVmapTable();
-  auto* const gc_map = compiled_method->GetGcMap();
+  auto const mapping_table = compiled_method->GetMappingTable();
+  auto const vmap_table = compiled_method->GetVmapTable();
+  auto const gc_map = compiled_method->GetGcMap();
   uint8_t* mapping_table_ptr = nullptr;
   uint8_t* vmap_table_ptr = nullptr;
   uint8_t* gc_map_ptr = nullptr;
 
-  if (mapping_table != nullptr) {
+  if (!mapping_table.empty()) {
     // Write out pre-header stuff.
     mapping_table_ptr = code_cache->AddDataArray(
-        self, mapping_table->data(), mapping_table->data() + mapping_table->size());
+        self, mapping_table.data(), mapping_table.data() + mapping_table.size());
     if (mapping_table_ptr == nullptr) {
       return false;  // Out of data cache.
     }
   }
 
-  if (vmap_table != nullptr) {
+  if (!vmap_table.empty()) {
     vmap_table_ptr = code_cache->AddDataArray(
-        self, vmap_table->data(), vmap_table->data() + vmap_table->size());
+        self, vmap_table.data(), vmap_table.data() + vmap_table.size());
     if (vmap_table_ptr == nullptr) {
       return false;  // Out of data cache.
     }
   }
 
-  if (gc_map != nullptr) {
+  if (!gc_map.empty()) {
     gc_map_ptr = code_cache->AddDataArray(
-        self, gc_map->data(), gc_map->data() + gc_map->size());
+        self, gc_map.data(), gc_map.data() + gc_map.size());
     if (gc_map_ptr == nullptr) {
       return false;  // Out of data cache.
     }
   }
 
   uint8_t* const code = code_cache->CommitCode(self,
+                                               method,
                                                mapping_table_ptr,
                                                vmap_table_ptr,
                                                gc_map_ptr,
                                                compiled_method->GetFrameSizeInBytes(),
                                                compiled_method->GetCoreSpillMask(),
                                                compiled_method->GetFpSpillMask(),
-                                               compiled_method->GetQuickCode()->data(),
-                                               compiled_method->GetQuickCode()->size());
+                                               compiled_method->GetQuickCode().data(),
+                                               compiled_method->GetQuickCode().size());
 
   if (code == nullptr) {
     return false;
@@ -285,13 +282,6 @@
 
   const size_t thumb_offset = compiled_method->CodeDelta();
   const uint32_t code_offset = sizeof(OatQuickMethodHeader) + thumb_offset;
-  *out_method = OatFile::OatMethod(code, code_offset);
-  DCHECK_EQ(out_method->GetGcMap(), gc_map_ptr);
-  DCHECK_EQ(out_method->GetMappingTable(), mapping_table_ptr);
-  DCHECK_EQ(out_method->GetVmapTable(), vmap_table_ptr);
-  DCHECK_EQ(out_method->GetFrameSizeInBytes(), compiled_method->GetFrameSizeInBytes());
-  DCHECK_EQ(out_method->GetCoreSpillMask(), compiled_method->GetCoreSpillMask());
-  DCHECK_EQ(out_method->GetFpSpillMask(), compiled_method->GetFpSpillMask());
   VLOG(jit)
       << "JIT added "
       << PrettyMethod(method) << "@" << method
diff --git a/compiler/jit/jit_compiler.h b/compiler/jit/jit_compiler.h
index 757f3f3..913a6d0 100644
--- a/compiler/jit/jit_compiler.h
+++ b/compiler/jit/jit_compiler.h
@@ -59,8 +59,8 @@
   // This is in the compiler since the runtime doesn't have access to the compiled method
   // structures.
   bool AddToCodeCache(ArtMethod* method,
-                      const CompiledMethod* compiled_method,
-                      OatFile::OatMethod* out_method) SHARED_REQUIRES(Locks::mutator_lock_);
+                      const CompiledMethod* compiled_method)
+      SHARED_REQUIRES(Locks::mutator_lock_);
 
   DISALLOW_COPY_AND_ASSIGN(JitCompiler);
 };
diff --git a/compiler/jni/quick/jni_compiler.cc b/compiler/jni/quick/jni_compiler.cc
index 34f0802..52a2382 100644
--- a/compiler/jni/quick/jni_compiler.cc
+++ b/compiler/jni/quick/jni_compiler.cc
@@ -487,7 +487,7 @@
                                                  frame_size,
                                                  main_jni_conv->CoreSpillMask(),
                                                  main_jni_conv->FpSpillMask(),
-                                                 nullptr,  // src_mapping_table.
+                                                 ArrayRef<const SrcMapElem>(),
                                                  ArrayRef<const uint8_t>(),  // mapping_table.
                                                  ArrayRef<const uint8_t>(),  // vmap_table.
                                                  ArrayRef<const uint8_t>(),  // native_gc_map.
diff --git a/compiler/linker/arm/relative_patcher_arm_base.cc b/compiler/linker/arm/relative_patcher_arm_base.cc
index cb9ea38..ac38f3d 100644
--- a/compiler/linker/arm/relative_patcher_arm_base.cc
+++ b/compiler/linker/arm/relative_patcher_arm_base.cc
@@ -85,8 +85,7 @@
                                                       const CompiledMethod* compiled_method,
                                                       MethodReference method_ref,
                                                       uint32_t max_extra_space) {
-  DCHECK(compiled_method->GetQuickCode() != nullptr);
-  uint32_t quick_code_size = compiled_method->GetQuickCode()->size();
+  uint32_t quick_code_size = compiled_method->GetQuickCode().size();
   uint32_t quick_code_offset = compiled_method->AlignCode(offset) + sizeof(OatQuickMethodHeader);
   uint32_t next_aligned_offset = compiled_method->AlignCode(quick_code_offset + quick_code_size);
   // Adjust for extra space required by the subclass.
diff --git a/compiler/linker/arm64/relative_patcher_arm64.cc b/compiler/linker/arm64/relative_patcher_arm64.cc
index 6f234a8..57018af 100644
--- a/compiler/linker/arm64/relative_patcher_arm64.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64.cc
@@ -74,7 +74,7 @@
   // Now that we have the actual offset where the code will be placed, locate the ADRP insns
   // that actually require the thunk.
   uint32_t quick_code_offset = compiled_method->AlignCode(offset) + sizeof(OatQuickMethodHeader);
-  ArrayRef<const uint8_t> code(*compiled_method->GetQuickCode());
+  ArrayRef<const uint8_t> code = compiled_method->GetQuickCode();
   uint32_t thunk_offset = compiled_method->AlignCode(quick_code_offset + code.size());
   DCHECK(compiled_method != nullptr);
   for (const LinkerPatch& patch : compiled_method->GetPatches()) {
diff --git a/compiler/linker/arm64/relative_patcher_arm64_test.cc b/compiler/linker/arm64/relative_patcher_arm64_test.cc
index 857d584..2a426b5 100644
--- a/compiler/linker/arm64/relative_patcher_arm64_test.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64_test.cc
@@ -237,7 +237,7 @@
     CHECK(!compiled_method_refs_.empty());
     CHECK_EQ(compiled_method_refs_[0].dex_method_index, 1u);
     CHECK_EQ(compiled_method_refs_.size(), compiled_methods_.size());
-    uint32_t method1_size = compiled_methods_[0]->GetQuickCode()->size();
+    uint32_t method1_size = compiled_methods_[0]->GetQuickCode().size();
     uint32_t thunk_offset = CompiledCode::AlignCode(method1_offset + method1_size, kArm64);
     uint32_t b_diff = thunk_offset - (method1_offset + num_nops * 4u);
     ASSERT_EQ(b_diff & 3u, 0u);
diff --git a/compiler/linker/relative_patcher_test.h b/compiler/linker/relative_patcher_test.h
index e357662..92cf8ca 100644
--- a/compiler/linker/relative_patcher_test.h
+++ b/compiler/linker/relative_patcher_test.h
@@ -74,8 +74,8 @@
     compiled_method_refs_.push_back(method_ref);
     compiled_methods_.emplace_back(new CompiledMethod(
         &driver_, instruction_set_, code,
-        0u, 0u, 0u, nullptr, ArrayRef<const uint8_t>(), ArrayRef<const uint8_t>(),
-        ArrayRef<const uint8_t>(), ArrayRef<const uint8_t>(),
+        0u, 0u, 0u, ArrayRef<const SrcMapElem>(), ArrayRef<const uint8_t>(),
+        ArrayRef<const uint8_t>(), ArrayRef<const uint8_t>(), ArrayRef<const uint8_t>(),
         patches));
   }
 
@@ -93,7 +93,7 @@
 
       offset += sizeof(OatQuickMethodHeader);
       uint32_t quick_code_offset = offset + compiled_method->CodeDelta();
-      const auto& code = *compiled_method->GetQuickCode();
+      const auto code = compiled_method->GetQuickCode();
       offset += code.size();
 
       method_offset_map_.map.Put(compiled_method_refs_[idx], quick_code_offset);
@@ -125,7 +125,7 @@
 
       out_.WriteFully(dummy_header, sizeof(OatQuickMethodHeader));
       offset += sizeof(OatQuickMethodHeader);
-      ArrayRef<const uint8_t> code(*compiled_method->GetQuickCode());
+      ArrayRef<const uint8_t> code = compiled_method->GetQuickCode();
       if (!compiled_method->GetPatches().empty()) {
         patched_code_.assign(code.begin(), code.end());
         code = ArrayRef<const uint8_t>(patched_code_);
@@ -164,7 +164,7 @@
       ++idx;
     }
     CHECK_NE(idx, compiled_method_refs_.size());
-    CHECK_EQ(compiled_methods_[idx]->GetQuickCode()->size(), expected_code.size());
+    CHECK_EQ(compiled_methods_[idx]->GetQuickCode().size(), expected_code.size());
 
     auto result = method_offset_map_.FindMethodOffset(method_ref);
     CHECK(result.first);  // Must have been linked.
diff --git a/compiler/oat_test.cc b/compiler/oat_test.cc
index 2d9d91a..06576cc 100644
--- a/compiler/oat_test.cc
+++ b/compiler/oat_test.cc
@@ -63,9 +63,9 @@
       EXPECT_EQ(oat_method.GetFpSpillMask(), compiled_method->GetFpSpillMask());
       uintptr_t oat_code_aligned = RoundDown(reinterpret_cast<uintptr_t>(quick_oat_code), 2);
       quick_oat_code = reinterpret_cast<const void*>(oat_code_aligned);
-      const SwapVector<uint8_t>* quick_code = compiled_method->GetQuickCode();
-      EXPECT_TRUE(quick_code != nullptr);
-      size_t code_size = quick_code->size() * sizeof(quick_code[0]);
+      ArrayRef<const uint8_t> quick_code = compiled_method->GetQuickCode();
+      EXPECT_FALSE(quick_code.empty());
+      size_t code_size = quick_code.size() * sizeof(quick_code[0]);
       EXPECT_EQ(0, memcmp(quick_oat_code, &quick_code[0], code_size))
           << PrettyMethod(method) << " " << code_size;
       CHECK_EQ(0, memcmp(quick_oat_code, &quick_code[0], code_size));
diff --git a/compiler/oat_writer.cc b/compiler/oat_writer.cc
index 640698b..dcb23bf 100644
--- a/compiler/oat_writer.cc
+++ b/compiler/oat_writer.cc
@@ -172,7 +172,7 @@
 }
 
 struct OatWriter::GcMapDataAccess {
-  static const SwapVector<uint8_t>* GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
+  static ArrayRef<const uint8_t> GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
     return compiled_method->GetGcMap();
   }
 
@@ -194,7 +194,7 @@
 };
 
 struct OatWriter::MappingTableDataAccess {
-  static const SwapVector<uint8_t>* GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
+  static ArrayRef<const uint8_t> GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
     return compiled_method->GetMappingTable();
   }
 
@@ -216,7 +216,7 @@
 };
 
 struct OatWriter::VmapTableDataAccess {
-  static const SwapVector<uint8_t>* GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
+  static ArrayRef<const uint8_t> GetData(const CompiledMethod* compiled_method) ALWAYS_INLINE {
     return compiled_method->GetVmapTable();
   }
 
@@ -388,8 +388,8 @@
       // Derived from CompiledMethod.
       uint32_t quick_code_offset = 0;
 
-      const SwapVector<uint8_t>* quick_code = compiled_method->GetQuickCode();
-      uint32_t code_size = quick_code->size() * sizeof(uint8_t);
+      ArrayRef<const uint8_t> quick_code = compiled_method->GetQuickCode();
+      uint32_t code_size = quick_code.size() * sizeof(uint8_t);
       uint32_t thumb_offset = compiled_method->CodeDelta();
 
       // Deduplicate code arrays if we are not producing debuggable code.
@@ -428,7 +428,7 @@
       uint32_t vmap_table_offset = method_header->vmap_table_offset_;
       // If we don't have quick code, then we must have a vmap, as that is how the dex2dex
       // compiler records its transformations.
-      DCHECK(quick_code != nullptr || vmap_table_offset != 0);
+      DCHECK(!quick_code.empty() || vmap_table_offset != 0);
       uint32_t gc_map_offset = method_header->gc_map_offset_;
       // The code offset was 0 when the mapping/vmap table offset was set, so it's set
       // to 0-offset and we need to adjust it by code_offset.
@@ -496,12 +496,12 @@
         } else {
           status = mirror::Class::kStatusNotReady;
         }
-        const SwapVector<uint8_t>* gc_map = compiled_method->GetGcMap();
-        if (gc_map != nullptr) {
-          size_t gc_map_size = gc_map->size() * sizeof(gc_map[0]);
+        ArrayRef<const uint8_t> gc_map = compiled_method->GetGcMap();
+        if (!gc_map.empty()) {
+          size_t gc_map_size = gc_map.size() * sizeof(gc_map[0]);
           bool is_native = it.MemberIsNative();
           CHECK(gc_map_size != 0 || is_native || status < mirror::Class::kStatusVerified)
-              << gc_map << " " << gc_map_size << " " << (is_native ? "true" : "false") << " "
+              << gc_map_size << " " << (is_native ? "true" : "false") << " "
               << (status < mirror::Class::kStatusVerified) << " " << status << " "
               << PrettyMethod(it.GetMemberIndex(), *dex_file_);
         }
@@ -519,30 +519,22 @@
  private:
   struct CodeOffsetsKeyComparator {
     bool operator()(const CompiledMethod* lhs, const CompiledMethod* rhs) const {
-      if (lhs->GetQuickCode() != rhs->GetQuickCode()) {
-        return lhs->GetQuickCode() < rhs->GetQuickCode();
+      // Code is deduplicated by CompilerDriver, compare only data pointers.
+      if (lhs->GetQuickCode().data() != rhs->GetQuickCode().data()) {
+        return lhs->GetQuickCode().data() < rhs->GetQuickCode().data();
       }
       // If the code is the same, all other fields are likely to be the same as well.
-      if (UNLIKELY(lhs->GetMappingTable() != rhs->GetMappingTable())) {
-        return lhs->GetMappingTable() < rhs->GetMappingTable();
+      if (UNLIKELY(lhs->GetMappingTable().data() != rhs->GetMappingTable().data())) {
+        return lhs->GetMappingTable().data() < rhs->GetMappingTable().data();
       }
-      if (UNLIKELY(lhs->GetVmapTable() != rhs->GetVmapTable())) {
-        return lhs->GetVmapTable() < rhs->GetVmapTable();
+      if (UNLIKELY(lhs->GetVmapTable().data() != rhs->GetVmapTable().data())) {
+        return lhs->GetVmapTable().data() < rhs->GetVmapTable().data();
       }
-      if (UNLIKELY(lhs->GetGcMap() != rhs->GetGcMap())) {
-        return lhs->GetGcMap() < rhs->GetGcMap();
+      if (UNLIKELY(lhs->GetGcMap().data() != rhs->GetGcMap().data())) {
+        return lhs->GetGcMap().data() < rhs->GetGcMap().data();
       }
-      const auto& lhs_patches = lhs->GetPatches();
-      const auto& rhs_patches = rhs->GetPatches();
-      if (UNLIKELY(lhs_patches.size() != rhs_patches.size())) {
-        return lhs_patches.size() < rhs_patches.size();
-      }
-      auto rit = rhs_patches.begin();
-      for (const LinkerPatch& lpatch : lhs_patches) {
-        if (UNLIKELY(!(lpatch == *rit))) {
-          return lpatch < *rit;
-        }
-        ++rit;
+      if (UNLIKELY(lhs->GetPatches().data() != rhs->GetPatches().data())) {
+        return lhs->GetPatches().data() < rhs->GetPatches().data();
       }
       return false;
     }
@@ -583,17 +575,17 @@
       DCHECK_LT(method_offsets_index_, oat_class->method_offsets_.size());
       DCHECK_EQ(DataAccess::GetOffset(oat_class, method_offsets_index_), 0u);
 
-      const SwapVector<uint8_t>* map = DataAccess::GetData(compiled_method);
-      uint32_t map_size = map == nullptr ? 0 : map->size() * sizeof((*map)[0]);
+      ArrayRef<const uint8_t> map = DataAccess::GetData(compiled_method);
+      uint32_t map_size = map.size() * sizeof(map[0]);
       if (map_size != 0u) {
-        auto lb = dedupe_map_.lower_bound(map);
-        if (lb != dedupe_map_.end() && !dedupe_map_.key_comp()(map, lb->first)) {
+        auto lb = dedupe_map_.lower_bound(map.data());
+        if (lb != dedupe_map_.end() && !dedupe_map_.key_comp()(map.data(), lb->first)) {
           DataAccess::SetOffset(oat_class, method_offsets_index_, lb->second);
         } else {
           DataAccess::SetOffset(oat_class, method_offsets_index_, offset_);
-          dedupe_map_.PutBefore(lb, map, offset_);
+          dedupe_map_.PutBefore(lb, map.data(), offset_);
           offset_ += map_size;
-          writer_->oat_header_->UpdateChecksum(&(*map)[0], map_size);
+          writer_->oat_header_->UpdateChecksum(&map[0], map_size);
         }
       }
       ++method_offsets_index_;
@@ -605,7 +597,7 @@
  private:
   // Deduplication is already done on a pointer basis by the compiler driver,
   // so we can simply compare the pointers to find out if things are duplicated.
-  SafeMap<const SwapVector<uint8_t>*, uint32_t> dedupe_map_;
+  SafeMap<const uint8_t*, uint32_t> dedupe_map_;
 };
 
 class OatWriter::InitImageMethodVisitor : public OatDexMethodVisitor {
@@ -647,7 +639,7 @@
       UNREACHABLE();
     }
 
-    if (compiled_method != nullptr && compiled_method->GetQuickCode()->size() != 0) {
+    if (compiled_method != nullptr && compiled_method->GetQuickCode().size() != 0) {
       method->SetEntryPointFromQuickCompiledCodePtrSize(
           reinterpret_cast<void*>(offsets.code_offset_), pointer_size_);
     }
@@ -713,10 +705,8 @@
       size_t file_offset = file_offset_;
       OutputStream* out = out_;
 
-      const SwapVector<uint8_t>* quick_code = compiled_method->GetQuickCode();
-      // Need a wrapper if we create a copy for patching.
-      ArrayRef<const uint8_t> wrapped(*quick_code);
-      uint32_t code_size = quick_code->size() * sizeof(uint8_t);
+      ArrayRef<const uint8_t> quick_code = compiled_method->GetQuickCode();
+      uint32_t code_size = quick_code.size() * sizeof(uint8_t);
 
       // Deduplicate code arrays.
       const OatMethodOffsets& method_offsets = oat_class->method_offsets_[method_offsets_index_];
@@ -753,8 +743,8 @@
         DCHECK_OFFSET_();
 
         if (!compiled_method->GetPatches().empty()) {
-          patched_code_.assign(quick_code->begin(), quick_code->end());
-          wrapped = ArrayRef<const uint8_t>(patched_code_);
+          patched_code_.assign(quick_code.begin(), quick_code.end());
+          quick_code = ArrayRef<const uint8_t>(patched_code_);
           for (const LinkerPatch& patch : compiled_method->GetPatches()) {
             if (patch.Type() == kLinkerPatchCallRelative) {
               // NOTE: Relative calls across oat files are not supported.
@@ -781,8 +771,8 @@
           }
         }
 
-        writer_->oat_header_->UpdateChecksum(wrapped.data(), code_size);
-        if (!out->WriteFully(wrapped.data(), code_size)) {
+        writer_->oat_header_->UpdateChecksum(quick_code.data(), code_size);
+        if (!out->WriteFully(quick_code.data(), code_size)) {
           ReportWriteFailure("method code", it);
           return false;
         }
@@ -947,14 +937,14 @@
       ++method_offsets_index_;
 
       // Write deduplicated map.
-      const SwapVector<uint8_t>* map = DataAccess::GetData(compiled_method);
-      size_t map_size = map == nullptr ? 0 : map->size() * sizeof((*map)[0]);
+      ArrayRef<const uint8_t> map = DataAccess::GetData(compiled_method);
+      size_t map_size = map.size() * sizeof(map[0]);
       DCHECK((map_size == 0u && map_offset == 0u) ||
             (map_size != 0u && map_offset != 0u && map_offset <= offset_))
           << map_size << " " << map_offset << " " << offset_ << " "
           << PrettyMethod(it.GetMemberIndex(), *dex_file_) << " for " << DataAccess::Name();
       if (map_size != 0u && map_offset == offset_) {
-        if (UNLIKELY(!out->WriteFully(&(*map)[0], map_size))) {
+        if (UNLIKELY(!out->WriteFully(&map[0], map_size))) {
           ReportWriteFailure(it);
           return false;
         }
diff --git a/compiler/optimizing/builder.cc b/compiler/optimizing/builder.cc
index 8ca352f..ed193c7 100644
--- a/compiler/optimizing/builder.cc
+++ b/compiler/optimizing/builder.cc
@@ -774,11 +774,12 @@
                                                        &string_init_offset);
   // Replace calls to String.<init> with StringFactory.
   if (is_string_init) {
-    HInvokeStaticOrDirect::DispatchInfo dispatch_info = ComputeDispatchInfo(is_string_init,
-                                                                            string_init_offset,
-                                                                            target_method,
-                                                                            direct_method,
-                                                                            direct_code);
+    HInvokeStaticOrDirect::DispatchInfo dispatch_info = {
+        HInvokeStaticOrDirect::MethodLoadKind::kStringInit,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        dchecked_integral_cast<uint64_t>(string_init_offset),
+        0U
+    };
     HInvoke* invoke = new (arena_) HInvokeStaticOrDirect(
         arena_,
         number_of_arguments - 1,
@@ -841,11 +842,12 @@
       clinit_check = ProcessClinitCheckForInvoke(dex_pc, method_idx, &clinit_check_requirement);
     }
 
-    HInvokeStaticOrDirect::DispatchInfo dispatch_info = ComputeDispatchInfo(is_string_init,
-                                                                            string_init_offset,
-                                                                            target_method,
-                                                                            direct_method,
-                                                                            direct_code);
+    HInvokeStaticOrDirect::DispatchInfo dispatch_info = {
+        HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        0u,
+        0U
+    };
     invoke = new (arena_) HInvokeStaticOrDirect(arena_,
                                                 number_of_arguments,
                                                 return_type,
@@ -958,77 +960,6 @@
   return clinit_check;
 }
 
-HInvokeStaticOrDirect::DispatchInfo HGraphBuilder::ComputeDispatchInfo(
-    bool is_string_init,
-    int32_t string_init_offset,
-    MethodReference target_method,
-    uintptr_t direct_method,
-    uintptr_t direct_code) {
-  HInvokeStaticOrDirect::MethodLoadKind method_load_kind;
-  HInvokeStaticOrDirect::CodePtrLocation code_ptr_location;
-  uint64_t method_load_data = 0u;
-  uint64_t direct_code_ptr = 0u;
-
-  if (is_string_init) {
-    // TODO: Use direct_method and direct_code for the appropriate StringFactory method.
-    method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kStringInit;
-    code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
-    method_load_data = string_init_offset;
-  } else if (target_method.dex_file == outer_compilation_unit_->GetDexFile() &&
-      target_method.dex_method_index == outer_compilation_unit_->GetDexMethodIndex()) {
-    method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kRecursive;
-    code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallSelf;
-  } else {
-    if (direct_method != 0u) {  // Should we use a direct pointer to the method?
-      if (direct_method != static_cast<uintptr_t>(-1)) {  // Is the method pointer known now?
-        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress;
-        method_load_data = direct_method;
-      } else {  // The direct pointer will be known at link time.
-        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup;
-      }
-    } else {  // Use dex cache.
-      DCHECK(target_method.dex_file == dex_compilation_unit_->GetDexFile());
-      DexCacheArraysLayout layout =
-          compiler_driver_->GetDexCacheArraysLayout(target_method.dex_file);
-      if (layout.Valid()) {  // Can we use PC-relative access to the dex cache arrays?
-        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative;
-        method_load_data = layout.MethodOffset(target_method.dex_method_index);
-      } else {  // We must go through the ArtMethod's pointer to resolved methods.
-        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod;
-      }
-    }
-    if (direct_code != 0u) {  // Should we use a direct pointer to the code?
-      if (direct_code != static_cast<uintptr_t>(-1)) {  // Is the code pointer known now?
-        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallDirect;
-        direct_code_ptr = direct_code;
-      } else if (compiler_driver_->IsImage() ||
-          target_method.dex_file == dex_compilation_unit_->GetDexFile()) {
-        // Use PC-relative calls for invokes within a multi-dex oat file.
-        // TODO: Recognize when the target dex file is within the current oat file for
-        // app compilation. At the moment we recognize only the boot image as multi-dex.
-        // NOTE: This will require changing the ARM backend which currently falls
-        // through from kCallPCRelative to kDirectCodeFixup for different dex files.
-        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative;
-      } else {  // The direct pointer will be known at link time.
-        // NOTE: This is used for app->boot calls when compiling an app against
-        // a relocatable but not yet relocated image.
-        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup;
-      }
-    } else {  // We must use the code pointer from the ArtMethod.
-      code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
-    }
-  }
-
-  if (graph_->IsDebuggable()) {
-    // For debuggable apps always use the code pointer from ArtMethod
-    // so that we don't circumvent instrumentation stubs if installed.
-    code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
-  }
-
-  return HInvokeStaticOrDirect::DispatchInfo {
-    method_load_kind, code_ptr_location, method_load_data, direct_code_ptr };
-}
-
 bool HGraphBuilder::SetupInvokeArguments(HInvoke* invoke,
                                          uint32_t number_of_vreg_arguments,
                                          uint32_t* args,
diff --git a/compiler/optimizing/builder.h b/compiler/optimizing/builder.h
index 6910d51..9eaa4b6 100644
--- a/compiler/optimizing/builder.h
+++ b/compiler/optimizing/builder.h
@@ -276,12 +276,6 @@
                                      uint32_t dex_pc,
                                      HInvoke* invoke);
 
-  HInvokeStaticOrDirect::DispatchInfo ComputeDispatchInfo(bool is_string_init,
-                                                          int32_t string_init_offset,
-                                                          MethodReference target_method,
-                                                          uintptr_t direct_method,
-                                                          uintptr_t direct_code);
-
   bool SetupInvokeArguments(HInvoke* invoke,
                             uint32_t number_of_vreg_arguments,
                             uint32_t* args,
diff --git a/compiler/optimizing/code_generator.cc b/compiler/optimizing/code_generator.cc
index 1c62dfa..a1bb5e0 100644
--- a/compiler/optimizing/code_generator.cc
+++ b/compiler/optimizing/code_generator.cc
@@ -379,13 +379,17 @@
 
   if (invoke->IsInvokeStaticOrDirect()) {
     HInvokeStaticOrDirect* call = invoke->AsInvokeStaticOrDirect();
-    if (call->IsStringInit()) {
-      locations->AddTemp(visitor->GetMethodLocation());
-    } else if (call->IsRecursive()) {
-      locations->SetInAt(call->GetCurrentMethodInputIndex(), visitor->GetMethodLocation());
-    } else {
-      locations->AddTemp(visitor->GetMethodLocation());
-      locations->SetInAt(call->GetCurrentMethodInputIndex(), Location::RequiresRegister());
+    switch (call->GetMethodLoadKind()) {
+      case HInvokeStaticOrDirect::MethodLoadKind::kRecursive:
+        locations->SetInAt(call->GetCurrentMethodInputIndex(), visitor->GetMethodLocation());
+        break;
+      case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod:
+        locations->AddTemp(visitor->GetMethodLocation());
+        locations->SetInAt(call->GetCurrentMethodInputIndex(), Location::RequiresRegister());
+        break;
+      default:
+        locations->AddTemp(visitor->GetMethodLocation());
+        break;
     }
   } else {
     locations->AddTemp(visitor->GetMethodLocation());
diff --git a/compiler/optimizing/code_generator.h b/compiler/optimizing/code_generator.h
index b04dfc0..47b6f30 100644
--- a/compiler/optimizing/code_generator.h
+++ b/compiler/optimizing/code_generator.h
@@ -172,6 +172,7 @@
                                OptimizingCompilerStats* stats = nullptr);
   virtual ~CodeGenerator() {}
 
+  // Get the graph. This is the outermost graph, never the graph of a method being inlined.
   HGraph* GetGraph() const { return graph_; }
 
   HBasicBlock* GetNextBlockToEmit() const;
@@ -431,6 +432,12 @@
                              uint32_t dex_pc,
                              SlowPathCode* slow_path) = 0;
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  virtual HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) = 0;
+
   // Generate a call to a static or direct method.
   virtual void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) = 0;
   // Generate a call to a virtual method.
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 92a5878..8d9794b 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -5155,26 +5155,51 @@
   }
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) {
+  if (desired_dispatch_info.method_load_kind ==
+      HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative) {
+    // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
+    return HInvokeStaticOrDirect::DispatchInfo {
+      HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod,
+      HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+      0u,
+      0u
+    };
+  }
+  if (desired_dispatch_info.code_ptr_location ==
+      HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative) {
+    const DexFile& outer_dex_file = GetGraph()->GetDexFile();
+    if (&outer_dex_file != target_method.dex_file) {
+      // Calls across dex files are more likely to exceed the available BL range,
+      // so use absolute patch with fixup if available and kCallArtMethod otherwise.
+      HInvokeStaticOrDirect::CodePtrLocation code_ptr_location =
+          (desired_dispatch_info.method_load_kind ==
+           HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup)
+          ? HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup
+          : HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
+      return HInvokeStaticOrDirect::DispatchInfo {
+        desired_dispatch_info.method_load_kind,
+        code_ptr_location,
+        desired_dispatch_info.method_load_data,
+        0u
+      };
+    }
+  }
+  return desired_dispatch_info;
+}
+
 void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) {
   // For better instruction scheduling we load the direct code pointer before the method pointer.
-  bool direct_code_loaded = false;
   switch (invoke->GetCodePtrLocation()) {
-    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
-      if (IsSameDexFile(*invoke->GetTargetMethod().dex_file, GetGraph()->GetDexFile())) {
-        break;
-      }
-      // Calls across dex files are more likely to exceed the available BL range,
-      // so use absolute patch by falling through to kDirectCodeFixup.
-      FALLTHROUGH_INTENDED;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
       // LR = code address from literal pool with link-time patch.
       __ LoadLiteral(LR, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod()));
-      direct_code_loaded = true;
       break;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
       // LR = invoke->GetDirectCodePtr();
       __ LoadImmediate(LR, invoke->GetDirectCodePtr());
-      direct_code_loaded = true;
       break;
     default:
       break;
@@ -5197,8 +5222,10 @@
                      DeduplicateMethodAddressLiteral(invoke->GetTargetMethod()));
       break;
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
-      // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
-      FALLTHROUGH_INTENDED;
+      // TODO: Implement this type.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: {
       Location current_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodInputIndex());
       Register method_reg;
@@ -5227,20 +5254,14 @@
       __ bl(GetFrameEntryLabel());
       break;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
-      if (!direct_code_loaded) {
-        relative_call_patches_.emplace_back(invoke->GetTargetMethod());
-        __ Bind(&relative_call_patches_.back().label);
-        Label label;
-        __ bl(&label);  // Arbitrarily branch to the instruction after BL, override at link time.
-        __ Bind(&label);
-        break;
-      }
-      // If we loaded the direct code above, fall through.
-      FALLTHROUGH_INTENDED;
+      relative_call_patches_.emplace_back(invoke->GetTargetMethod());
+      __ Bind(&relative_call_patches_.back().label);
+      // Arbitrarily branch to the BL itself, override at link time.
+      __ bl(&relative_call_patches_.back().label);
+      break;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
       // LR prepared above for better instruction scheduling.
-      DCHECK(direct_code_loaded);
       // LR()
       __ blx(LR);
       break;
diff --git a/compiler/optimizing/code_generator_arm.h b/compiler/optimizing/code_generator_arm.h
index 6900933..cef1095 100644
--- a/compiler/optimizing/code_generator_arm.h
+++ b/compiler/optimizing/code_generator_arm.h
@@ -362,6 +362,12 @@
 
   Label* GetFrameEntryLabel() { return &frame_entry_label_; }
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
   void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
 
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 1773c06..b0be446 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -2826,6 +2826,13 @@
   return false;
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM64::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method ATTRIBUTE_UNUSED) {
+  // On arm64 we support all dispatch types.
+  return desired_dispatch_info;
+}
+
 void CodeGeneratorARM64::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) {
   // For better instruction scheduling we load the direct code pointer before the method pointer.
   bool direct_code_loaded = false;
diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index 799f1bd..ab684ea 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -388,6 +388,12 @@
     return false;
   }
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
   void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
 
diff --git a/compiler/optimizing/code_generator_mips.cc b/compiler/optimizing/code_generator_mips.cc
index 4404aa3..6aed444 100644
--- a/compiler/optimizing/code_generator_mips.cc
+++ b/compiler/optimizing/code_generator_mips.cc
@@ -1000,6 +1000,9 @@
 void CodeGeneratorMIPS::AddLocationAsTemp(Location location, LocationSummary* locations) {
   if (location.IsRegister()) {
     locations->AddTemp(location);
+  } else if (location.IsRegisterPair()) {
+    locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>()));
+    locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>()));
   } else {
     UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location;
   }
@@ -2951,6 +2954,37 @@
   return false;
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorMIPS::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method ATTRIBUTE_UNUSED) {
+  switch (desired_dispatch_info.method_load_kind) {
+    case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup:
+    case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
+      // TODO: Implement these types. For the moment, we fall back to kDexCacheViaMethod.
+      return HInvokeStaticOrDirect::DispatchInfo {
+        HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        0u,
+        0u
+      };
+    default:
+      break;
+  }
+  switch (desired_dispatch_info.code_ptr_location) {
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
+      // TODO: Implement these types. For the moment, we fall back to kCallArtMethod.
+      return HInvokeStaticOrDirect::DispatchInfo {
+        desired_dispatch_info.method_load_kind,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        desired_dispatch_info.method_load_data,
+        0u
+      };
+    default:
+      return desired_dispatch_info;
+  }
+}
+
 void CodeGeneratorMIPS::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) {
   // All registers are assumed to be correctly set up per the calling convention.
 
@@ -2970,13 +3004,11 @@
       __ LoadConst32(temp.AsRegister<Register>(), invoke->GetMethodAddress());
       break;
     case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup:
-      // TODO: Implement this type. (Needs literal support.) At the moment, the
-      // CompilerDriver will not direct the backend to use this type for MIPS.
-      LOG(FATAL) << "Unsupported!";
-      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
-      // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
-      FALLTHROUGH_INTENDED;
+      // TODO: Implement these types.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: {
       Location current_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodInputIndex());
       Register reg = temp.AsRegister<Register>();
@@ -3017,15 +3049,15 @@
       __ Jalr(T9);
       __ Nop();
       break;
-    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
-      // TODO: Implement kCallPCRelative. For the moment, we fall back to kMethodCode.
-      FALLTHROUGH_INTENDED;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
-      // TODO: Implement kDirectCodeFixup. For the moment, we fall back to kMethodCode.
-      FALLTHROUGH_INTENDED;
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
+      // TODO: Implement these types.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod:
       // T9 = callee_method->entry_point_from_quick_compiled_code_;
-      __ LoadFromOffset(kLoadDoubleword,
+      __ LoadFromOffset(kLoadWord,
                         T9,
                         callee_method.AsRegister<Register>(),
                         ArtMethod::EntryPointFromQuickCompiledCodeOffset(
diff --git a/compiler/optimizing/code_generator_mips.h b/compiler/optimizing/code_generator_mips.h
index a571e76..059131d 100644
--- a/compiler/optimizing/code_generator_mips.h
+++ b/compiler/optimizing/code_generator_mips.h
@@ -332,6 +332,12 @@
     return type == Primitive::kPrimLong;
   }
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
   void GenerateVirtualCall(HInvokeVirtual* invoke ATTRIBUTE_UNUSED,
                            Location temp ATTRIBUTE_UNUSED) OVERRIDE {
diff --git a/compiler/optimizing/code_generator_mips64.cc b/compiler/optimizing/code_generator_mips64.cc
index 5f78285..55efd5f 100644
--- a/compiler/optimizing/code_generator_mips64.cc
+++ b/compiler/optimizing/code_generator_mips64.cc
@@ -2528,6 +2528,37 @@
   return false;
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorMIPS64::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method ATTRIBUTE_UNUSED) {
+  switch (desired_dispatch_info.method_load_kind) {
+    case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup:
+    case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
+      // TODO: Implement these types. For the moment, we fall back to kDexCacheViaMethod.
+      return HInvokeStaticOrDirect::DispatchInfo {
+        HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        0u,
+        0u
+      };
+    default:
+      break;
+  }
+  switch (desired_dispatch_info.code_ptr_location) {
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
+      // TODO: Implement these types. For the moment, we fall back to kCallArtMethod.
+      return HInvokeStaticOrDirect::DispatchInfo {
+        desired_dispatch_info.method_load_kind,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        desired_dispatch_info.method_load_data,
+        0u
+      };
+    default:
+      return desired_dispatch_info;
+  }
+}
+
 void CodeGeneratorMIPS64::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) {
   // All registers are assumed to be correctly set up per the calling convention.
 
@@ -2547,13 +2578,11 @@
       __ LoadConst64(temp.AsRegister<GpuRegister>(), invoke->GetMethodAddress());
       break;
     case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup:
-      // TODO: Implement this type. (Needs literal support.) At the moment, the
-      // CompilerDriver will not direct the backend to use this type for MIPS.
-      LOG(FATAL) << "Unsupported!";
-      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
-      // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
-      FALLTHROUGH_INTENDED;
+      // TODO: Implement these types.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: {
       Location current_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodInputIndex());
       GpuRegister reg = temp.AsRegister<GpuRegister>();
@@ -2593,12 +2622,12 @@
       // LR()
       __ Jalr(T9);
       break;
-    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
-      // TODO: Implement kCallPCRelative. For the moment, we fall back to kMethodCode.
-      FALLTHROUGH_INTENDED;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
-      // TODO: Implement kDirectCodeFixup. For the moment, we fall back to kMethodCode.
-      FALLTHROUGH_INTENDED;
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
+      // TODO: Implement these types.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod:
       // T9 = callee_method->entry_point_from_quick_compiled_code_;
       __ LoadFromOffset(kLoadDoubleword,
diff --git a/compiler/optimizing/code_generator_mips64.h b/compiler/optimizing/code_generator_mips64.h
index df3fc0d..9bbd027 100644
--- a/compiler/optimizing/code_generator_mips64.h
+++ b/compiler/optimizing/code_generator_mips64.h
@@ -326,6 +326,12 @@
 
   bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const { return false; }
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
   void GenerateVirtualCall(HInvokeVirtual* invoke ATTRIBUTE_UNUSED,
                            Location temp ATTRIBUTE_UNUSED) OVERRIDE {
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 963eec2..0df7e3b 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -3757,6 +3757,34 @@
   }
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorX86::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method ATTRIBUTE_UNUSED) {
+  if (desired_dispatch_info.method_load_kind ==
+      HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative) {
+    // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
+    return HInvokeStaticOrDirect::DispatchInfo {
+      HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod,
+      HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+      0u,
+      0u
+    };
+  }
+  switch (desired_dispatch_info.code_ptr_location) {
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
+      // For direct code, we actually prefer to call via the code pointer from ArtMethod*.
+      // (Though the direct CALL ptr16:32 is available for consideration).
+      return HInvokeStaticOrDirect::DispatchInfo {
+        desired_dispatch_info.method_load_kind,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        desired_dispatch_info.method_load_data,
+        0u
+      };
+    default:
+      return desired_dispatch_info;
+  }
+}
 
 void CodeGeneratorX86::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) {
   Location callee_method = temp;  // For all kinds except kRecursive, callee will be in temp.
@@ -3777,8 +3805,10 @@
       __ Bind(&method_patches_.back().label);  // Bind the label at the end of the "movl" insn.
       break;
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative:
-      // TODO: Implement this type. For the moment, we fall back to kDexCacheViaMethod.
-      FALLTHROUGH_INTENDED;
+      // TODO: Implement this type.
+      // Currently filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: {
       Location current_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodInputIndex());
       Register method_reg;
@@ -3814,9 +3844,9 @@
     }
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
-      // For direct code, we actually prefer to call via the code pointer from ArtMethod*.
-      // (Though the direct CALL ptr16:32 is available for consideration).
-      FALLTHROUGH_INTENDED;
+      // Filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod:
       // (callee_method + offset_of_quick_compiled_code)()
       __ call(Address(callee_method.AsRegister<Register>(),
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index fdfc5ab..ac3d06c 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -333,6 +333,12 @@
   // Helper method to move a 64bits value between two locations.
   void Move64(Location destination, Location source);
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   // Generate a call to a static or direct method.
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
   // Generate a call to a virtual method.
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index ed2e4ca..5218d70 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -473,6 +473,24 @@
   UNREACHABLE();
 }
 
+HInvokeStaticOrDirect::DispatchInfo CodeGeneratorX86_64::GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method ATTRIBUTE_UNUSED) {
+  switch (desired_dispatch_info.code_ptr_location) {
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
+    case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
+      // For direct code, we actually prefer to call via the code pointer from ArtMethod*.
+      return HInvokeStaticOrDirect::DispatchInfo {
+        desired_dispatch_info.method_load_kind,
+        HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod,
+        desired_dispatch_info.method_load_data,
+        0u
+      };
+    default:
+      return desired_dispatch_info;
+  }
+}
+
 void CodeGeneratorX86_64::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke,
                                                      Location temp) {
   // All registers are assumed to be correctly set up.
@@ -539,8 +557,9 @@
     }
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
-      // For direct code, we actually prefer to call via the code pointer from ArtMethod*.
-      FALLTHROUGH_INTENDED;
+      // Filtered out by GetSupportedInvokeStaticOrDirectDispatch().
+      LOG(FATAL) << "Unsupported";
+      UNREACHABLE();
     case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod:
       // (callee_method + offset_of_quick_compiled_code)()
       __ call(Address(callee_method.AsRegister<CpuRegister>(),
diff --git a/compiler/optimizing/code_generator_x86_64.h b/compiler/optimizing/code_generator_x86_64.h
index dc86a48..fc485f5 100644
--- a/compiler/optimizing/code_generator_x86_64.h
+++ b/compiler/optimizing/code_generator_x86_64.h
@@ -335,6 +335,12 @@
     return false;
   }
 
+  // Check if the desired_dispatch_info is supported. If it is, return it,
+  // otherwise return a fall-back info that should be used instead.
+  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
+      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
+      MethodReference target_method) OVERRIDE;
+
   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
   void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
 
diff --git a/compiler/optimizing/inliner.cc b/compiler/optimizing/inliner.cc
index e2aca30..0aaa6b3 100644
--- a/compiler/optimizing/inliner.cc
+++ b/compiler/optimizing/inliner.cc
@@ -32,6 +32,7 @@
 #include "optimizing_compiler.h"
 #include "reference_type_propagation.h"
 #include "register_allocator.h"
+#include "sharpening.h"
 #include "ssa_phi_elimination.h"
 #include "scoped_thread_state_change.h"
 #include "thread.h"
@@ -396,12 +397,14 @@
   HDeadCodeElimination dce(callee_graph, stats_);
   HConstantFolding fold(callee_graph);
   ReferenceTypePropagation type_propagation(callee_graph, handles_);
+  HSharpening sharpening(callee_graph, codegen_, dex_compilation_unit, compiler_driver_);
   InstructionSimplifier simplify(callee_graph, stats_);
   IntrinsicsRecognizer intrinsics(callee_graph, compiler_driver_);
 
   HOptimization* optimizations[] = {
     &intrinsics,
     &type_propagation,
+    &sharpening,
     &simplify,
     &dce,
     &fold,
@@ -415,6 +418,7 @@
   size_t number_of_instructions_budget = kMaximumNumberOfHInstructions;
   if (depth_ + 1 < compiler_driver_->GetCompilerOptions().GetInlineDepthLimit()) {
     HInliner inliner(callee_graph,
+                     codegen_,
                      outer_compilation_unit_,
                      dex_compilation_unit,
                      compiler_driver_,
@@ -484,7 +488,7 @@
         return false;
       }
 
-      if (!same_dex_file && current->NeedsDexCache()) {
+      if (!same_dex_file && current->NeedsDexCacheOfDeclaringClass()) {
         VLOG(compiler) << "Method " << PrettyMethod(method_index, callee_dex_file)
                        << " could not be inlined because " << current->DebugName()
                        << " it is in a different dex file and requires access to the dex cache";
diff --git a/compiler/optimizing/inliner.h b/compiler/optimizing/inliner.h
index bce5915..0f6a945 100644
--- a/compiler/optimizing/inliner.h
+++ b/compiler/optimizing/inliner.h
@@ -22,6 +22,7 @@
 
 namespace art {
 
+class CodeGenerator;
 class CompilerDriver;
 class DexCompilationUnit;
 class HGraph;
@@ -31,6 +32,7 @@
 class HInliner : public HOptimization {
  public:
   HInliner(HGraph* outer_graph,
+           CodeGenerator* codegen,
            const DexCompilationUnit& outer_compilation_unit,
            const DexCompilationUnit& caller_compilation_unit,
            CompilerDriver* compiler_driver,
@@ -40,6 +42,7 @@
       : HOptimization(outer_graph, kInlinerPassName, stats),
         outer_compilation_unit_(outer_compilation_unit),
         caller_compilation_unit_(caller_compilation_unit),
+        codegen_(codegen),
         compiler_driver_(compiler_driver),
         depth_(depth),
         number_of_inlined_instructions_(0),
@@ -57,6 +60,7 @@
 
   const DexCompilationUnit& outer_compilation_unit_;
   const DexCompilationUnit& caller_compilation_unit_;
+  CodeGenerator* const codegen_;
   CompilerDriver* const compiler_driver_;
   const size_t depth_;
   size_t number_of_inlined_instructions_;
diff --git a/compiler/optimizing/intrinsics_mips64.cc b/compiler/optimizing/intrinsics_mips64.cc
index 56c4177..0ab0b80 100644
--- a/compiler/optimizing/intrinsics_mips64.cc
+++ b/compiler/optimizing/intrinsics_mips64.cc
@@ -272,7 +272,9 @@
   GenReverseBytes(invoke->GetLocations(), Primitive::kPrimShort, GetAssembler());
 }
 
-static void GenNumberOfLeadingZeroes(LocationSummary* locations, bool is64bit, Mips64Assembler* assembler) {
+static void GenNumberOfLeadingZeroes(LocationSummary* locations,
+                                     bool is64bit,
+                                     Mips64Assembler* assembler) {
   GpuRegister in  = locations->InAt(0).AsRegister<GpuRegister>();
   GpuRegister out = locations->Out().AsRegister<GpuRegister>();
 
@@ -301,7 +303,9 @@
   GenNumberOfLeadingZeroes(invoke->GetLocations(), true, GetAssembler());
 }
 
-static void GenNumberOfTrailingZeroes(LocationSummary* locations, bool is64bit, Mips64Assembler* assembler) {
+static void GenNumberOfTrailingZeroes(LocationSummary* locations,
+                                      bool is64bit,
+                                      Mips64Assembler* assembler) {
   Location in = locations->InAt(0);
   Location out = locations->Out();
 
@@ -383,7 +387,7 @@
   GenRotateRight(invoke, Primitive::kPrimInt, GetAssembler());
 }
 
-// int java.lang.Long.rotateRight(long i, int distance)
+// long java.lang.Long.rotateRight(long i, int distance)
 void IntrinsicLocationsBuilderMIPS64::VisitLongRotateRight(HInvoke* invoke) {
   LocationSummary* locations = new (arena_) LocationSummary(invoke,
                                                            LocationSummary::kNoCall,
@@ -446,7 +450,7 @@
   GenRotateLeft(invoke, Primitive::kPrimInt, GetAssembler());
 }
 
-// int java.lang.Long.rotateLeft(long i, int distance)
+// long java.lang.Long.rotateLeft(long i, int distance)
 void IntrinsicLocationsBuilderMIPS64::VisitLongRotateLeft(HInvoke* invoke) {
   LocationSummary* locations = new (arena_) LocationSummary(invoke,
                                                            LocationSummary::kNoCall,
@@ -754,17 +758,19 @@
   __ SqrtD(out, in);
 }
 
-static void CreateFPToFP(ArenaAllocator* arena, HInvoke* invoke) {
+static void CreateFPToFP(ArenaAllocator* arena,
+                         HInvoke* invoke,
+                         Location::OutputOverlap overlaps = Location::kOutputOverlap) {
   LocationSummary* locations = new (arena) LocationSummary(invoke,
                                                            LocationSummary::kNoCall,
                                                            kIntrinsified);
   locations->SetInAt(0, Location::RequiresFpuRegister());
-  locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
+  locations->SetOut(Location::RequiresFpuRegister(), overlaps);
 }
 
 // double java.lang.Math.rint(double)
 void IntrinsicLocationsBuilderMIPS64::VisitMathRint(HInvoke* invoke) {
-  CreateFPToFP(arena_, invoke);
+  CreateFPToFP(arena_, invoke, Location::kNoOutputOverlap);
 }
 
 void IntrinsicCodeGeneratorMIPS64::VisitMathRint(HInvoke* invoke) {
@@ -788,15 +794,22 @@
                                              kQuietNaN |
                                              kSignalingNaN;
 
-void IntrinsicCodeGeneratorMIPS64::VisitMathFloor(HInvoke* invoke) {
-  LocationSummary* locations = invoke->GetLocations();
-  Mips64Assembler* assembler = GetAssembler();
+enum FloatRoundingMode {
+  kFloor,
+  kCeil,
+};
+
+static void GenRoundingMode(LocationSummary* locations,
+                            FloatRoundingMode mode,
+                            Mips64Assembler* assembler) {
   FpuRegister in = locations->InAt(0).AsFpuRegister<FpuRegister>();
   FpuRegister out = locations->Out().AsFpuRegister<FpuRegister>();
 
+  DCHECK_NE(in, out);
+
   Label done;
 
-  // double floor(double in) {
+  // double floor/ceil(double in) {
   //     if in.isNaN || in.isInfinite || in.isZero {
   //         return in;
   //     }
@@ -806,19 +819,23 @@
   __ MovD(out, in);
   __ Bnezc(AT, &done);
 
-  //     Long outLong = floor(in);
+  //     Long outLong = floor/ceil(in);
   //     if outLong == Long.MAX_VALUE {
-  //         // floor() has almost certainly returned a value which
-  //         // can't be successfully represented as a signed 64-bit
-  //         // number.  Java expects that the input value will be
-  //         // returned in these cases.
-  //         // There is also a small probability that floor(in)
-  //         // correctly truncates the input value to Long.MAX_VALUE.  In
-  //         // that case, this exception handling code still does the
-  //         // correct thing.
+  //         // floor()/ceil() has almost certainly returned a value
+  //         // which can't be successfully represented as a signed
+  //         // 64-bit number.  Java expects that the input value will
+  //         // be returned in these cases.
+  //         // There is also a small probability that floor(in)/ceil(in)
+  //         // correctly truncates/rounds up the input value to
+  //         // Long.MAX_VALUE.  In that case, this exception handling
+  //         // code still does the correct thing.
   //         return in;
   //     }
-  __ FloorLD(out, in);
+  if (mode == kFloor) {
+    __ FloorLD(out, in);
+  } else  if (mode == kCeil) {
+    __ CeilLD(out, in);
+  }
   __ Dmfc1(AT, out);
   __ MovD(out, in);
   __ LoadConst64(TMP, kPrimLongMax);
@@ -832,53 +849,17 @@
   // }
 }
 
+void IntrinsicCodeGeneratorMIPS64::VisitMathFloor(HInvoke* invoke) {
+  GenRoundingMode(invoke->GetLocations(), kFloor, GetAssembler());
+}
+
 // double java.lang.Math.ceil(double)
 void IntrinsicLocationsBuilderMIPS64::VisitMathCeil(HInvoke* invoke) {
   CreateFPToFP(arena_, invoke);
 }
 
 void IntrinsicCodeGeneratorMIPS64::VisitMathCeil(HInvoke* invoke) {
-  LocationSummary* locations = invoke->GetLocations();
-  Mips64Assembler* assembler = GetAssembler();
-  FpuRegister in = locations->InAt(0).AsFpuRegister<FpuRegister>();
-  FpuRegister out = locations->Out().AsFpuRegister<FpuRegister>();
-
-  Label done;
-
-  // double ceil(double in) {
-  //     if in.isNaN || in.isInfinite || in.isZero {
-  //         return in;
-  //     }
-  __ ClassD(out, in);
-  __ Dmfc1(AT, out);
-  __ Andi(AT, AT, kFPLeaveUnchanged);   // +0.0 | +Inf | -0.0 | -Inf | qNaN | sNaN
-  __ MovD(out, in);
-  __ Bnezc(AT, &done);
-
-  //     Long outLong = ceil(in);
-  //     if outLong == Long.MAX_VALUE {
-  //         // ceil() has almost certainly returned a value which
-  //         // can't be successfully represented as a signed 64-bit
-  //         // number.  Java expects that the input value will be
-  //         // returned in these cases.
-  //         // There is also a small probability that ceil(in)
-  //         // correctly rounds up the input value to Long.MAX_VALUE.  In
-  //         // that case, this exception handling code still does the
-  //         // correct thing.
-  //         return in;
-  //     }
-  __ CeilLD(out, in);
-  __ Dmfc1(AT, out);
-  __ MovD(out, in);
-  __ LoadConst64(TMP, kPrimLongMax);
-  __ Beqc(AT, TMP, &done);
-
-  //     double out = outLong;
-  //     return out;
-  __ Dmtc1(AT, out);
-  __ Cvtdl(out, out);
-  __ Bind(&done);
-  // }
+  GenRoundingMode(invoke->GetLocations(), kCeil, GetAssembler());
 }
 
 // byte libcore.io.Memory.peekByte(long address)
diff --git a/compiler/optimizing/load_store_elimination.cc b/compiler/optimizing/load_store_elimination.cc
index 68d02a4..90f28e5 100644
--- a/compiler/optimizing/load_store_elimination.cc
+++ b/compiler/optimizing/load_store_elimination.cc
@@ -695,12 +695,15 @@
       } else {
         redundant_store = true;
       }
+      // TODO: eliminate the store if the singleton object is not finalizable.
+      redundant_store = false;
     }
     if (redundant_store) {
       removed_instructions_.push_back(instruction);
       substitute_instructions_.push_back(nullptr);
       TryRemovingNullCheck(instruction);
     }
+
     heap_values[idx] = value;
     // This store may kill values in other heap locations due to aliasing.
     for (size_t i = 0; i < heap_values.size(); i++) {
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index 98c3096..3480265 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -1911,8 +1911,8 @@
   return !opt.GetDoesNotNeedEnvironment();
 }
 
-bool HInvokeStaticOrDirect::NeedsDexCache() const {
-  if (IsRecursive() || IsStringInit()) {
+bool HInvokeStaticOrDirect::NeedsDexCacheOfDeclaringClass() const {
+  if (GetMethodLoadKind() != MethodLoadKind::kDexCacheViaMethod) {
     return false;
   }
   if (!IsIntrinsic()) {
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 7aa933d..6028d4b 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1980,7 +1980,9 @@
     return NeedsEnvironment() || IsLoadClass() || IsLoadString();
   }
 
-  virtual bool NeedsDexCache() const { return false; }
+  // Returns whether the code generation of the instruction will require to have access
+  // to the dex cache of the current method's declaring class via the current method.
+  virtual bool NeedsDexCacheOfDeclaringClass() const { return false; }
 
   // Does this instruction have any use in an environment before
   // control flow hits 'other'?
@@ -3368,15 +3370,15 @@
   };
 
   struct DispatchInfo {
-    const MethodLoadKind method_load_kind;
-    const CodePtrLocation code_ptr_location;
+    MethodLoadKind method_load_kind;
+    CodePtrLocation code_ptr_location;
     // The method load data holds
     //   - thread entrypoint offset for kStringInit method if this is a string init invoke.
     //     Note that there are multiple string init methods, each having its own offset.
     //   - the method address for kDirectAddress
     //   - the dex cache arrays offset for kDexCachePcRel.
-    const uint64_t method_load_data;
-    const uint64_t direct_code_ptr;
+    uint64_t method_load_data;
+    uint64_t direct_code_ptr;
   };
 
   HInvokeStaticOrDirect(ArenaAllocator* arena,
@@ -3405,6 +3407,10 @@
         target_method_(target_method),
         dispatch_info_(dispatch_info) {}
 
+  void SetDispatchInfo(const DispatchInfo& dispatch_info) {
+    dispatch_info_ = dispatch_info;
+  }
+
   bool CanDoImplicitNullCheckOn(HInstruction* obj ATTRIBUTE_UNUSED) const OVERRIDE {
     // We access the method via the dex cache so we can't do an implicit null check.
     // TODO: for intrinsics we can generate implicit null checks.
@@ -3419,11 +3425,13 @@
   MethodLoadKind GetMethodLoadKind() const { return dispatch_info_.method_load_kind; }
   CodePtrLocation GetCodePtrLocation() const { return dispatch_info_.code_ptr_location; }
   bool IsRecursive() const { return GetMethodLoadKind() == MethodLoadKind::kRecursive; }
-  bool NeedsDexCache() const OVERRIDE;
+  bool NeedsDexCacheOfDeclaringClass() const OVERRIDE;
   bool IsStringInit() const { return GetMethodLoadKind() == MethodLoadKind::kStringInit; }
   uint32_t GetCurrentMethodInputIndex() const { return GetNumberOfArguments(); }
   bool HasMethodAddress() const { return GetMethodLoadKind() == MethodLoadKind::kDirectAddress; }
-  bool HasPcRelDexCache() const { return GetMethodLoadKind() == MethodLoadKind::kDexCachePcRelative; }
+  bool HasPcRelDexCache() const {
+    return GetMethodLoadKind() == MethodLoadKind::kDexCachePcRelative;
+  }
   bool HasDirectCodePtr() const { return GetCodePtrLocation() == CodePtrLocation::kCallDirect; }
   MethodReference GetTargetMethod() const { return target_method_; }
 
@@ -4765,7 +4773,7 @@
 
   const DexFile& GetDexFile() { return dex_file_; }
 
-  bool NeedsDexCache() const OVERRIDE { return !is_referrers_class_; }
+  bool NeedsDexCacheOfDeclaringClass() const OVERRIDE { return !is_referrers_class_; }
 
   static SideEffects SideEffectsForArchRuntimeCalls() {
     return SideEffects::CanTriggerGC();
@@ -4807,7 +4815,7 @@
 
   // TODO: Can we deopt or debug when we resolve a string?
   bool NeedsEnvironment() const OVERRIDE { return false; }
-  bool NeedsDexCache() const OVERRIDE { return true; }
+  bool NeedsDexCacheOfDeclaringClass() const OVERRIDE { return true; }
   bool CanBeNull() const OVERRIDE { return false; }
 
   static SideEffects SideEffectsForArchRuntimeCalls() {
diff --git a/compiler/optimizing/optimizing_compiler.cc b/compiler/optimizing/optimizing_compiler.cc
index d6f2543..6632f95 100644
--- a/compiler/optimizing/optimizing_compiler.cc
+++ b/compiler/optimizing/optimizing_compiler.cc
@@ -63,6 +63,7 @@
 #include "prepare_for_register_allocation.h"
 #include "reference_type_propagation.h"
 #include "register_allocator.h"
+#include "sharpening.h"
 #include "side_effects_analysis.h"
 #include "ssa_builder.h"
 #include "ssa_phi_elimination.h"
@@ -378,6 +379,7 @@
 }
 
 static void MaybeRunInliner(HGraph* graph,
+                            CodeGenerator* codegen,
                             CompilerDriver* driver,
                             OptimizingCompilerStats* stats,
                             const DexCompilationUnit& dex_compilation_unit,
@@ -392,7 +394,7 @@
 
   ArenaAllocator* arena = graph->GetArena();
   HInliner* inliner = new (arena) HInliner(
-    graph, dex_compilation_unit, dex_compilation_unit, driver, handles, stats);
+    graph, codegen, dex_compilation_unit, dex_compilation_unit, driver, handles, stats);
   ReferenceTypePropagation* type_propagation =
     new (arena) ReferenceTypePropagation(graph, handles,
         "reference_type_propagation_after_inlining");
@@ -445,6 +447,7 @@
 }
 
 static void RunOptimizations(HGraph* graph,
+                             CodeGenerator* codegen,
                              CompilerDriver* driver,
                              OptimizingCompilerStats* stats,
                              const DexCompilationUnit& dex_compilation_unit,
@@ -467,6 +470,7 @@
   BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, induction);
   ReferenceTypePropagation* type_propagation =
       new (arena) ReferenceTypePropagation(graph, handles);
+  HSharpening* sharpening = new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver);
   InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier(
       graph, stats, "instruction_simplifier_after_types");
   InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier(
@@ -481,6 +485,7 @@
     fold1,
     simplify1,
     type_propagation,
+    sharpening,
     dce1,
     simplify2
   };
@@ -502,7 +507,7 @@
 
     RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer);
   } else {
-    MaybeRunInliner(graph, driver, stats, dex_compilation_unit, pass_observer, handles);
+    MaybeRunInliner(graph, codegen, driver, stats, dex_compilation_unit, pass_observer, handles);
 
     HOptimization* optimizations2[] = {
       // BooleanSimplifier depends on the InstructionSimplifier removing
@@ -577,8 +582,13 @@
   ScopedObjectAccess soa(Thread::Current());
   StackHandleScopeCollection handles(soa.Self());
   soa.Self()->TransitionFromRunnableToSuspended(kNative);
-  RunOptimizations(graph, compiler_driver, compilation_stats_.get(),
-                   dex_compilation_unit, pass_observer, &handles);
+  RunOptimizations(graph,
+                   codegen,
+                   compiler_driver,
+                   compilation_stats_.get(),
+                   dex_compilation_unit,
+                   pass_observer,
+                   &handles);
 
   AllocateRegisters(graph, codegen, pass_observer);
 
@@ -607,7 +617,7 @@
       codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
       codegen->GetCoreSpillMask(),
       codegen->GetFpuSpillMask(),
-      &src_mapping_table,
+      ArrayRef<const SrcMapElem>(src_mapping_table),
       ArrayRef<const uint8_t>(),  // mapping_table.
       ArrayRef<const uint8_t>(stack_map),
       ArrayRef<const uint8_t>(),  // native_gc_map.
@@ -652,7 +662,7 @@
       codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
       codegen->GetCoreSpillMask(),
       codegen->GetFpuSpillMask(),
-      &src_mapping_table,
+      ArrayRef<const SrcMapElem>(src_mapping_table),
       AlignVectorSize(mapping_table),
       AlignVectorSize(vmap_table),
       AlignVectorSize(gc_map),
diff --git a/compiler/optimizing/sharpening.cc b/compiler/optimizing/sharpening.cc
new file mode 100644
index 0000000..6494964
--- /dev/null
+++ b/compiler/optimizing/sharpening.cc
@@ -0,0 +1,134 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "sharpening.h"
+
+#include "code_generator.h"
+#include "utils/dex_cache_arrays_layout-inl.h"
+#include "driver/compiler_driver.h"
+#include "nodes.h"
+
+namespace art {
+
+void HSharpening::Run() {
+  // We don't care about the order of the blocks here.
+  for (HBasicBlock* block : graph_->GetReversePostOrder()) {
+    for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
+      HInstruction* instruction = it.Current();
+      if (instruction->IsInvokeStaticOrDirect()) {
+        ProcessInvokeStaticOrDirect(instruction->AsInvokeStaticOrDirect());
+      }
+      // TODO: Move the sharpening of invoke-virtual/-interface/-super from HGraphBuilder
+      //       here. Rewrite it to avoid the CompilerDriver's reliance on verifier data
+      //       because we know the type better when inlining.
+      // TODO: HLoadClass, HLoadString - select PC relative dex cache array access if
+      //       available.
+    }
+  }
+}
+
+void HSharpening::ProcessInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+  if (invoke->IsStringInit()) {
+    // Not using the dex cache arrays. But we could still try to use a better dispatch...
+    // TODO: Use direct_method and direct_code for the appropriate StringFactory method.
+    return;
+  }
+
+  // TODO: Avoid CompilerDriver.
+  InvokeType invoke_type = invoke->GetOriginalInvokeType();
+  MethodReference target_method(&graph_->GetDexFile(), invoke->GetDexMethodIndex());
+  int vtable_idx;
+  uintptr_t direct_code, direct_method;
+  bool success = compiler_driver_->ComputeInvokeInfo(
+      &compilation_unit_,
+      invoke->GetDexPc(),
+      false /* update_stats: already updated in builder */,
+      true /* enable_devirtualization */,
+      &invoke_type,
+      &target_method,
+      &vtable_idx,
+      &direct_code,
+      &direct_method);
+  DCHECK(success);
+  DCHECK_EQ(invoke_type, invoke->GetInvokeType());
+  DCHECK_EQ(target_method.dex_file, invoke->GetTargetMethod().dex_file);
+  DCHECK_EQ(target_method.dex_method_index, invoke->GetTargetMethod().dex_method_index);
+
+  HInvokeStaticOrDirect::MethodLoadKind method_load_kind;
+  HInvokeStaticOrDirect::CodePtrLocation code_ptr_location;
+  uint64_t method_load_data = 0u;
+  uint64_t direct_code_ptr = 0u;
+
+  HGraph* outer_graph = codegen_->GetGraph();
+  if (target_method.dex_file == &outer_graph->GetDexFile() &&
+      target_method.dex_method_index == outer_graph->GetMethodIdx()) {
+    method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kRecursive;
+    code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallSelf;
+  } else {
+    if (direct_method != 0u) {  // Should we use a direct pointer to the method?
+      if (direct_method != static_cast<uintptr_t>(-1)) {  // Is the method pointer known now?
+        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress;
+        method_load_data = direct_method;
+      } else {  // The direct pointer will be known at link time.
+        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup;
+      }
+    } else {  // Use dex cache.
+      DCHECK_EQ(target_method.dex_file, &graph_->GetDexFile());
+      DexCacheArraysLayout layout =
+          compiler_driver_->GetDexCacheArraysLayout(target_method.dex_file);
+      if (layout.Valid()) {  // Can we use PC-relative access to the dex cache arrays?
+        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative;
+        method_load_data = layout.MethodOffset(target_method.dex_method_index);
+      } else {  // We must go through the ArtMethod's pointer to resolved methods.
+        method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod;
+      }
+    }
+    if (direct_code != 0u) {  // Should we use a direct pointer to the code?
+      if (direct_code != static_cast<uintptr_t>(-1)) {  // Is the code pointer known now?
+        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallDirect;
+        direct_code_ptr = direct_code;
+      } else if (compiler_driver_->IsImage() ||
+          target_method.dex_file == &graph_->GetDexFile()) {
+        // Use PC-relative calls for invokes within a multi-dex oat file.
+        // TODO: Recognize when the target dex file is within the current oat file for
+        // app compilation. At the moment we recognize only the boot image as multi-dex.
+        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative;
+      } else {  // The direct pointer will be known at link time.
+        // NOTE: This is used for app->boot calls when compiling an app against
+        // a relocatable but not yet relocated image.
+        code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup;
+      }
+    } else {  // We must use the code pointer from the ArtMethod.
+      code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
+    }
+  }
+
+  if (graph_->IsDebuggable()) {
+    // For debuggable apps always use the code pointer from ArtMethod
+    // so that we don't circumvent instrumentation stubs if installed.
+    code_ptr_location = HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod;
+  }
+
+  HInvokeStaticOrDirect::DispatchInfo desired_dispatch_info = {
+      method_load_kind, code_ptr_location, method_load_data, direct_code_ptr
+  };
+  HInvokeStaticOrDirect::DispatchInfo dispatch_info =
+      codegen_->GetSupportedInvokeStaticOrDirectDispatch(desired_dispatch_info,
+                                                         invoke->GetTargetMethod());
+  invoke->SetDispatchInfo(dispatch_info);
+}
+
+}  // namespace art
diff --git a/compiler/optimizing/sharpening.h b/compiler/optimizing/sharpening.h
new file mode 100644
index 0000000..adae700
--- /dev/null
+++ b/compiler/optimizing/sharpening.h
@@ -0,0 +1,58 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_SHARPENING_H_
+#define ART_COMPILER_OPTIMIZING_SHARPENING_H_
+
+#include "optimization.h"
+
+namespace art {
+
+class CodeGenerator;
+class CompilerDriver;
+class DexCompilationUnit;
+class HInvokeStaticOrDirect;
+
+// Optimization that tries to improve the way we dispatch methods and access types,
+// fields, etc. Besides actual method sharpening based on receiver type (for example
+// virtual->direct), this includes selecting the best available dispatch for
+// invoke-static/-direct based on code generator support.
+class HSharpening : public HOptimization {
+ public:
+  HSharpening(HGraph* graph,
+              CodeGenerator* codegen,
+              const DexCompilationUnit& compilation_unit,
+              CompilerDriver* compiler_driver)
+      : HOptimization(graph, kSharpeningPassName),
+        codegen_(codegen),
+        compilation_unit_(compilation_unit),
+        compiler_driver_(compiler_driver) { }
+
+  void Run() OVERRIDE;
+
+  static constexpr const char* kSharpeningPassName = "sharpening";
+
+ private:
+  void ProcessInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke);
+
+  CodeGenerator* codegen_;
+  const DexCompilationUnit& compilation_unit_;
+  CompilerDriver* compiler_driver_;
+};
+
+}  // namespace art
+
+#endif  // ART_COMPILER_OPTIMIZING_SHARPENING_H_
diff --git a/compiler/utils/dedupe_set-inl.h b/compiler/utils/dedupe_set-inl.h
new file mode 100644
index 0000000..ac54813
--- /dev/null
+++ b/compiler/utils/dedupe_set-inl.h
@@ -0,0 +1,253 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_UTILS_DEDUPE_SET_INL_H_
+#define ART_COMPILER_UTILS_DEDUPE_SET_INL_H_
+
+#include "dedupe_set.h"
+
+#include <algorithm>
+#include <inttypes.h>
+#include <unordered_map>
+
+#include "base/mutex.h"
+#include "base/hash_set.h"
+#include "base/stl_util.h"
+#include "base/stringprintf.h"
+#include "base/time_utils.h"
+
+namespace art {
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+struct DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::Stats {
+  size_t collision_sum = 0u;
+  size_t collision_max = 0u;
+  size_t total_probe_distance = 0u;
+  size_t total_size = 0u;
+};
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+class DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::Shard {
+ public:
+  Shard(const Alloc& alloc, const std::string& lock_name)
+      : alloc_(alloc),
+        lock_name_(lock_name),
+        lock_(lock_name_.c_str()),
+        keys_() {
+  }
+
+  ~Shard() {
+    for (const HashedKey<StoreKey>& key : keys_) {
+      DCHECK(key.Key() != nullptr);
+      alloc_.Destroy(key.Key());
+    }
+  }
+
+  const StoreKey* Add(Thread* self, size_t hash, const InKey& in_key) REQUIRES(!lock_) {
+    MutexLock lock(self, lock_);
+    HashedKey<InKey> hashed_in_key(hash, &in_key);
+    auto it = keys_.Find(hashed_in_key);
+    if (it != keys_.end()) {
+      DCHECK(it->Key() != nullptr);
+      return it->Key();
+    }
+    const StoreKey* store_key = alloc_.Copy(in_key);
+    keys_.Insert(HashedKey<StoreKey> { hash, store_key });
+    return store_key;
+  }
+
+  void UpdateStats(Thread* self, Stats* global_stats) REQUIRES(!lock_) {
+    // HashSet<> doesn't keep entries ordered by hash, so we actually allocate memory
+    // for bookkeeping while collecting the stats.
+    std::unordered_map<HashType, size_t> stats;
+    {
+      MutexLock lock(self, lock_);
+      // Note: The total_probe_distance will be updated with the current state.
+      // It may have been higher before a re-hash.
+      global_stats->total_probe_distance += keys_.TotalProbeDistance();
+      global_stats->total_size += keys_.Size();
+      for (const HashedKey<StoreKey>& key : keys_) {
+        auto it = stats.find(key.Hash());
+        if (it == stats.end()) {
+          stats.insert({key.Hash(), 1u});
+        } else {
+          ++it->second;
+        }
+      }
+    }
+    for (const auto& entry : stats) {
+      size_t number_of_entries = entry.second;
+      if (number_of_entries > 1u) {
+        global_stats->collision_sum += number_of_entries - 1u;
+        global_stats->collision_max = std::max(global_stats->collision_max, number_of_entries);
+      }
+    }
+  }
+
+ private:
+  template <typename T>
+  class HashedKey {
+   public:
+    HashedKey() : hash_(0u), key_(nullptr) { }
+    HashedKey(size_t hash, const T* key) : hash_(hash), key_(key) { }
+
+    size_t Hash() const {
+      return hash_;
+    }
+
+    const T* Key() const {
+      return key_;
+    }
+
+    bool IsEmpty() const {
+      return Key() == nullptr;
+    }
+
+    void MakeEmpty() {
+      key_ = nullptr;
+    }
+
+   private:
+    size_t hash_;
+    const T* key_;
+  };
+
+  class ShardEmptyFn {
+   public:
+    bool IsEmpty(const HashedKey<StoreKey>& key) const {
+      return key.IsEmpty();
+    }
+
+    void MakeEmpty(HashedKey<StoreKey>& key) {
+      key.MakeEmpty();
+    }
+  };
+
+  struct ShardHashFn {
+    template <typename T>
+    size_t operator()(const HashedKey<T>& key) const {
+      return key.Hash();
+    }
+  };
+
+  struct ShardPred {
+    typename std::enable_if<!std::is_same<StoreKey, InKey>::value, bool>::type
+    operator()(const HashedKey<StoreKey>& lhs, const HashedKey<StoreKey>& rhs) const {
+      DCHECK(lhs.Key() != nullptr);
+      DCHECK(rhs.Key() != nullptr);
+      // Rehashing: stored keys are already deduplicated, so we can simply compare key pointers.
+      return lhs.Key() == rhs.Key();
+    }
+
+    template <typename LeftT, typename RightT>
+    bool operator()(const HashedKey<LeftT>& lhs, const HashedKey<RightT>& rhs) const {
+      DCHECK(lhs.Key() != nullptr);
+      DCHECK(rhs.Key() != nullptr);
+      return lhs.Hash() == rhs.Hash() &&
+          lhs.Key()->size() == rhs.Key()->size() &&
+          std::equal(lhs.Key()->begin(), lhs.Key()->end(), rhs.Key()->begin());
+    }
+  };
+
+  Alloc alloc_;
+  const std::string lock_name_;
+  Mutex lock_;
+  HashSet<HashedKey<StoreKey>, ShardEmptyFn, ShardHashFn, ShardPred> keys_ GUARDED_BY(lock_);
+};
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+const StoreKey* DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::Add(
+    Thread* self, const InKey& key) {
+  uint64_t hash_start;
+  if (kIsDebugBuild) {
+    hash_start = NanoTime();
+  }
+  HashType raw_hash = HashFunc()(key);
+  if (kIsDebugBuild) {
+    uint64_t hash_end = NanoTime();
+    hash_time_ += hash_end - hash_start;
+  }
+  HashType shard_hash = raw_hash / kShard;
+  HashType shard_bin = raw_hash % kShard;
+  return shards_[shard_bin]->Add(self, shard_hash, key);
+}
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::DedupeSet(const char* set_name,
+                                                                         const Alloc& alloc)
+    : hash_time_(0) {
+  for (HashType i = 0; i < kShard; ++i) {
+    std::ostringstream oss;
+    oss << set_name << " lock " << i;
+    shards_[i].reset(new Shard(alloc, oss.str()));
+  }
+}
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::~DedupeSet() {
+  // Everything done by member destructors.
+}
+
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
+          HashType kShard>
+std::string DedupeSet<InKey, StoreKey, Alloc, HashType, HashFunc, kShard>::DumpStats(
+    Thread* self) const {
+  Stats stats;
+  for (HashType shard = 0; shard < kShard; ++shard) {
+    shards_[shard]->UpdateStats(self, &stats);
+  }
+  return StringPrintf("%zu collisions, %zu max hash collisions, "
+                      "%zu/%zu probe distance, %" PRIu64 " ns hash time",
+                      stats.collision_sum,
+                      stats.collision_max,
+                      stats.total_probe_distance,
+                      stats.total_size,
+                      hash_time_);
+}
+
+
+}  // namespace art
+
+#endif  // ART_COMPILER_UTILS_DEDUPE_SET_INL_H_
diff --git a/compiler/utils/dedupe_set.h b/compiler/utils/dedupe_set.h
index 2c4a689..b62f216 100644
--- a/compiler/utils/dedupe_set.h
+++ b/compiler/utils/dedupe_set.h
@@ -17,151 +17,41 @@
 #ifndef ART_COMPILER_UTILS_DEDUPE_SET_H_
 #define ART_COMPILER_UTILS_DEDUPE_SET_H_
 
-#include <algorithm>
-#include <inttypes.h>
 #include <memory>
-#include <set>
+#include <stdint.h>
 #include <string>
 
-#include "base/mutex.h"
-#include "base/stl_util.h"
-#include "base/stringprintf.h"
-#include "base/time_utils.h"
-#include "utils/swap_space.h"
+#include "base/macros.h"
 
 namespace art {
 
+class Thread;
+
 // A set of Keys that support a HashFunc returning HashType. Used to find duplicates of Key in the
 // Add method. The data-structure is thread-safe through the use of internal locks, it also
 // supports the lock being sharded.
-template <typename InKey, typename StoreKey, typename HashType, typename HashFunc,
+template <typename InKey,
+          typename StoreKey,
+          typename Alloc,
+          typename HashType,
+          typename HashFunc,
           HashType kShard = 1>
 class DedupeSet {
-  typedef std::pair<HashType, const InKey*> HashedInKey;
-  struct HashedKey {
-    StoreKey* store_ptr;
-    union {
-      HashType store_hash;        // Valid if store_ptr != null.
-      const HashedInKey* in_key;  // Valid if store_ptr == null.
-    };
-  };
-
-  class Comparator {
-   public:
-    bool operator()(const HashedKey& a, const HashedKey& b) const {
-      HashType a_hash = (a.store_ptr != nullptr) ? a.store_hash : a.in_key->first;
-      HashType b_hash = (b.store_ptr != nullptr) ? b.store_hash : b.in_key->first;
-      if (a_hash != b_hash) {
-        return a_hash < b_hash;
-      }
-      if (a.store_ptr != nullptr && b.store_ptr != nullptr) {
-        return std::lexicographical_compare(a.store_ptr->begin(), a.store_ptr->end(),
-                                            b.store_ptr->begin(), b.store_ptr->end());
-      } else if (a.store_ptr != nullptr && b.store_ptr == nullptr) {
-        return std::lexicographical_compare(a.store_ptr->begin(), a.store_ptr->end(),
-                                            b.in_key->second->begin(), b.in_key->second->end());
-      } else if (a.store_ptr == nullptr && b.store_ptr != nullptr) {
-        return std::lexicographical_compare(a.in_key->second->begin(), a.in_key->second->end(),
-                                            b.store_ptr->begin(), b.store_ptr->end());
-      } else {
-        return std::lexicographical_compare(a.in_key->second->begin(), a.in_key->second->end(),
-                                            b.in_key->second->begin(), b.in_key->second->end());
-      }
-    }
-  };
-
  public:
-  StoreKey* Add(Thread* self, const InKey& key) {
-    uint64_t hash_start;
-    if (kIsDebugBuild) {
-      hash_start = NanoTime();
-    }
-    HashType raw_hash = HashFunc()(key);
-    if (kIsDebugBuild) {
-      uint64_t hash_end = NanoTime();
-      hash_time_ += hash_end - hash_start;
-    }
-    HashType shard_hash = raw_hash / kShard;
-    HashType shard_bin = raw_hash % kShard;
-    HashedInKey hashed_in_key(shard_hash, &key);
-    HashedKey hashed_key;
-    hashed_key.store_ptr = nullptr;
-    hashed_key.in_key = &hashed_in_key;
-    MutexLock lock(self, *lock_[shard_bin]);
-    auto it = keys_[shard_bin].find(hashed_key);
-    if (it != keys_[shard_bin].end()) {
-      DCHECK(it->store_ptr != nullptr);
-      return it->store_ptr;
-    }
-    hashed_key.store_ptr = CreateStoreKey(key);
-    hashed_key.store_hash = shard_hash;
-    keys_[shard_bin].insert(hashed_key);
-    return hashed_key.store_ptr;
-  }
+  // Add a new key to the dedupe set if not present. Return the equivalent deduplicated stored key.
+  const StoreKey* Add(Thread* self, const InKey& key);
 
-  DedupeSet(const char* set_name, SwapAllocator<void>& alloc)
-      : allocator_(alloc), hash_time_(0) {
-    for (HashType i = 0; i < kShard; ++i) {
-      std::ostringstream oss;
-      oss << set_name << " lock " << i;
-      lock_name_[i] = oss.str();
-      lock_[i].reset(new Mutex(lock_name_[i].c_str()));
-    }
-  }
+  DedupeSet(const char* set_name, const Alloc& alloc);
 
-  ~DedupeSet() {
-    // Have to manually free all pointers.
-    for (auto& shard : keys_) {
-      for (const auto& hashed_key : shard) {
-        DCHECK(hashed_key.store_ptr != nullptr);
-        DeleteStoreKey(hashed_key.store_ptr);
-      }
-    }
-  }
+  ~DedupeSet();
 
-  std::string DumpStats() const {
-    size_t collision_sum = 0;
-    size_t collision_max = 0;
-    for (HashType shard = 0; shard < kShard; ++shard) {
-      HashType last_hash = 0;
-      size_t collision_cur_max = 0;
-      for (const HashedKey& key : keys_[shard]) {
-        DCHECK(key.store_ptr != nullptr);
-        if (key.store_hash == last_hash) {
-          collision_cur_max++;
-          if (collision_cur_max > 1) {
-            collision_sum++;
-            if (collision_cur_max > collision_max) {
-              collision_max = collision_cur_max;
-            }
-          }
-        } else {
-          collision_cur_max = 1;
-          last_hash = key.store_hash;
-        }
-      }
-    }
-    return StringPrintf("%zu collisions, %zu max bucket size, %" PRIu64 " ns hash time",
-                        collision_sum, collision_max, hash_time_);
-  }
+  std::string DumpStats(Thread* self) const;
 
  private:
-  StoreKey* CreateStoreKey(const InKey& key) {
-    StoreKey* ret = allocator_.allocate(1);
-    allocator_.construct(ret, key.begin(), key.end(), allocator_);
-    return ret;
-  }
+  struct Stats;
+  class Shard;
 
-  void DeleteStoreKey(StoreKey* key) {
-    SwapAllocator<StoreKey> alloc(allocator_);
-    alloc.destroy(key);
-    alloc.deallocate(key, 1);
-  }
-
-  std::string lock_name_[kShard];
-  std::unique_ptr<Mutex> lock_[kShard];
-  std::set<HashedKey, Comparator> keys_[kShard];
-  SwapAllocator<StoreKey> allocator_;
+  std::unique_ptr<Shard> shards_[kShard];
   uint64_t hash_time_;
 
   DISALLOW_COPY_AND_ASSIGN(DedupeSet);
diff --git a/compiler/utils/dedupe_set_test.cc b/compiler/utils/dedupe_set_test.cc
index 637964e..60a891d 100644
--- a/compiler/utils/dedupe_set_test.cc
+++ b/compiler/utils/dedupe_set_test.cc
@@ -18,15 +18,18 @@
 
 #include <algorithm>
 #include <cstdio>
+#include <vector>
 
+#include "dedupe_set-inl.h"
 #include "gtest/gtest.h"
 #include "thread-inl.h"
+#include "utils/array_ref.h"
 
 namespace art {
 
-class DedupeHashFunc {
+class DedupeSetTestHashFunc {
  public:
-  size_t operator()(const std::vector<uint8_t>& array) const {
+  size_t operator()(const ArrayRef<const uint8_t>& array) const {
     size_t hash = 0;
     for (uint8_t c : array) {
       hash += c;
@@ -36,46 +39,52 @@
     return hash;
   }
 };
+
+class DedupeSetTestAlloc {
+ public:
+  const std::vector<uint8_t>* Copy(const ArrayRef<const uint8_t>& src) {
+    return new std::vector<uint8_t>(src.begin(), src.end());
+  }
+
+  void Destroy(const std::vector<uint8_t>* key) {
+    delete key;
+  }
+};
+
 TEST(DedupeSetTest, Test) {
   Thread* self = Thread::Current();
-  typedef std::vector<uint8_t> ByteArray;
-  SwapAllocator<void> swap(nullptr);
-  DedupeSet<ByteArray, SwapVector<uint8_t>, size_t, DedupeHashFunc> deduplicator("test", swap);
-  SwapVector<uint8_t>* array1;
+  DedupeSetTestAlloc alloc;
+  DedupeSet<ArrayRef<const uint8_t>,
+            std::vector<uint8_t>,
+            DedupeSetTestAlloc,
+            size_t,
+            DedupeSetTestHashFunc> deduplicator("test", alloc);
+  const std::vector<uint8_t>* array1;
   {
-    ByteArray test1;
-    test1.push_back(10);
-    test1.push_back(20);
-    test1.push_back(30);
-    test1.push_back(45);
-
+    uint8_t raw_test1[] = { 10u, 20u, 30u, 45u };
+    ArrayRef<const uint8_t> test1(raw_test1);
     array1 = deduplicator.Add(self, test1);
     ASSERT_NE(array1, nullptr);
     ASSERT_TRUE(std::equal(test1.begin(), test1.end(), array1->begin()));
   }
 
-  SwapVector<uint8_t>* array2;
+  const std::vector<uint8_t>* array2;
   {
-    ByteArray test1;
-    test1.push_back(10);
-    test1.push_back(20);
-    test1.push_back(30);
-    test1.push_back(45);
-    array2 = deduplicator.Add(self, test1);
+    uint8_t raw_test2[] = { 10u, 20u, 30u, 45u };
+    ArrayRef<const uint8_t> test2(raw_test2);
+    array2 = deduplicator.Add(self, test2);
     ASSERT_EQ(array2, array1);
-    ASSERT_TRUE(std::equal(test1.begin(), test1.end(), array2->begin()));
+    ASSERT_TRUE(std::equal(test2.begin(), test2.end(), array2->begin()));
   }
 
-  SwapVector<uint8_t>* array3;
+  const std::vector<uint8_t>* array3;
   {
-    ByteArray test1;
-    test1.push_back(10);
-    test1.push_back(22);
-    test1.push_back(30);
-    test1.push_back(47);
-    array3 = deduplicator.Add(self, test1);
+    uint8_t raw_test3[] = { 10u, 22u, 30u, 47u };
+    ArrayRef<const uint8_t> test3(raw_test3);
+    array3 = deduplicator.Add(self, test3);
     ASSERT_NE(array3, nullptr);
-    ASSERT_TRUE(std::equal(test1.begin(), test1.end(), array3->begin()));
+    ASSERT_NE(array3, array1);
+    ASSERT_TRUE(std::equal(test3.begin(), test3.end(), array3->begin()));
   }
 }
 
diff --git a/dexdump/dexdump.cc b/dexdump/dexdump.cc
index 282db5d..52e6c02 100644
--- a/dexdump/dexdump.cc
+++ b/dexdump/dexdump.cc
@@ -775,7 +775,7 @@
     // case Instruction::k35ms:       // [opt] invoke-virtual+super
     // case Instruction::k35mi:       // [opt] inline invoke
       {
-        u4 arg[5];
+        u4 arg[Instruction::kMaxVarArgRegs];
         pDecInsn->GetVarArgs(arg);
         fputs(" {", gOutFile);
         for (int i = 0, n = pDecInsn->VRegA(); i < n; i++) {
@@ -788,6 +788,21 @@
         fprintf(gOutFile, "}, %s", indexBuf);
       }
       break;
+    case Instruction::k25x:        // op vC, {vD, vE, vF, vG} (B: count)
+      {
+        u4 arg[Instruction::kMaxVarArgRegs25x];
+        pDecInsn->GetAllArgs25x(arg);
+        fprintf(gOutFile, " v%d, {", arg[0]);
+        for (int i = 0, n = pDecInsn->VRegB(); i < n; i++) {
+          if (i == 0) {
+            fprintf(gOutFile, "v%d", arg[Instruction::kLambdaVirtualRegisterWidth + i]);
+          } else {
+            fprintf(gOutFile, ", v%d", arg[Instruction::kLambdaVirtualRegisterWidth + i]);
+          }
+        }  // for
+        fputc('}', gOutFile);
+      }
+      break;
     case Instruction::k3rc:        // op {vCCCC .. v(CCCC+AA-1)}, thing@BBBB
     // NOT SUPPORTED:
     // case Instruction::k3rms:       // [opt] invoke-virtual+super/range
diff --git a/runtime/arch/mips/context_mips.cc b/runtime/arch/mips/context_mips.cc
index 4dedb33..375a03a 100644
--- a/runtime/arch/mips/context_mips.cc
+++ b/runtime/arch/mips/context_mips.cc
@@ -28,11 +28,11 @@
   std::fill_n(gprs_, arraysize(gprs_), nullptr);
   std::fill_n(fprs_, arraysize(fprs_), nullptr);
   gprs_[SP] = &sp_;
-  gprs_[RA] = &ra_;
+  gprs_[T9] = &t9_;
   gprs_[A0] = &arg0_;
   // Initialize registers with easy to spot debug values.
   sp_ = MipsContext::kBadGprBase + SP;
-  ra_ = MipsContext::kBadGprBase + RA;
+  t9_ = MipsContext::kBadGprBase + T9;
   arg0_ = 0;
 }
 
diff --git a/runtime/arch/mips/context_mips.h b/runtime/arch/mips/context_mips.h
index f1e2905..7dcff63 100644
--- a/runtime/arch/mips/context_mips.h
+++ b/runtime/arch/mips/context_mips.h
@@ -41,7 +41,7 @@
   }
 
   void SetPC(uintptr_t new_pc) OVERRIDE {
-    SetGPR(RA, new_pc);
+    SetGPR(T9, new_pc);
   }
 
   bool IsAccessibleGPR(uint32_t reg) OVERRIDE {
@@ -86,9 +86,10 @@
   // Pointers to registers in the stack, initialized to null except for the special cases below.
   uintptr_t* gprs_[kNumberOfCoreRegisters];
   uint32_t* fprs_[kNumberOfFRegisters];
-  // Hold values for sp and ra (return address) if they are not located within a stack frame, as
-  // well as the first argument.
-  uintptr_t sp_, ra_, arg0_;
+  // Hold values for sp and t9 if they are not located within a stack frame. We use t9 for the
+  // PC (as ra is required to be valid for single-frame deopt and must not be clobbered). We
+  // also need the first argument for single-frame deopt.
+  uintptr_t sp_, t9_, arg0_;
 };
 }  // namespace mips
 }  // namespace art
diff --git a/runtime/arch/mips/quick_entrypoints_mips.S b/runtime/arch/mips/quick_entrypoints_mips.S
index ba58c3f..0691f2a 100644
--- a/runtime/arch/mips/quick_entrypoints_mips.S
+++ b/runtime/arch/mips/quick_entrypoints_mips.S
@@ -374,7 +374,7 @@
     lw      $ra, 124($a0)
     lw      $a0, 16($a0)
     move    $v0, $zero          # clear result registers r0 and r1
-    jalr    $zero, $ra          # do long jump
+    jalr    $zero, $t9          # do long jump
     move    $v1, $zero
 END art_quick_do_long_jump
 
diff --git a/runtime/art_method.cc b/runtime/art_method.cc
index f5befdf..a10d7af 100644
--- a/runtime/art_method.cc
+++ b/runtime/art_method.cc
@@ -390,21 +390,70 @@
   }
 
   Runtime* runtime = Runtime::Current();
-  const void* code = runtime->GetInstrumentation()->GetQuickCodeFor(this, sizeof(void*));
-  DCHECK(code != nullptr);
+  const void* existing_entry_point = GetEntryPointFromQuickCompiledCode();
+  DCHECK(existing_entry_point != nullptr);
+  ClassLinker* class_linker = runtime->GetClassLinker();
 
-  if (runtime->GetClassLinker()->IsQuickGenericJniStub(code)) {
+  if (class_linker->IsQuickGenericJniStub(existing_entry_point)) {
     // The generic JNI does not have any method header.
     return nullptr;
   }
 
-  code = EntryPointToCodePointer(code);
-  OatQuickMethodHeader* method_header = reinterpret_cast<OatQuickMethodHeader*>(
-      reinterpret_cast<uintptr_t>(code) - sizeof(OatQuickMethodHeader));
+  // Check whether the current entry point contains this pc.
+  if (!class_linker->IsQuickResolutionStub(existing_entry_point) &&
+      !class_linker->IsQuickToInterpreterBridge(existing_entry_point)) {
+    OatQuickMethodHeader* method_header =
+        OatQuickMethodHeader::FromEntryPoint(existing_entry_point);
 
-  // TODO(ngeoffray): validate the pc. Note that unit tests can give unrelated pcs (for
-  // example arch_test).
-  UNUSED(pc);
+    if (method_header->Contains(pc)) {
+      return method_header;
+    }
+  }
+
+  // Check whether the pc is in the JIT code cache.
+  jit::Jit* jit = Runtime::Current()->GetJit();
+  if (jit != nullptr) {
+    jit::JitCodeCache* code_cache = jit->GetCodeCache();
+    OatQuickMethodHeader* method_header = code_cache->LookupMethodHeader(pc, this);
+    if (method_header != nullptr) {
+      DCHECK(method_header->Contains(pc));
+      return method_header;
+    } else {
+      DCHECK(!code_cache->ContainsPc(reinterpret_cast<const void*>(pc))) << std::hex << pc;
+    }
+  }
+
+  // The code has to be in an oat file.
+  bool found;
+  OatFile::OatMethod oat_method = class_linker->FindOatMethodFor(this, &found);
+  if (!found) {
+    // Only for unit tests.
+    // TODO(ngeoffray): Update these tests to pass the right pc?
+    return OatQuickMethodHeader::FromEntryPoint(existing_entry_point);
+  }
+  const void* oat_entry_point = oat_method.GetQuickCode();
+  if (oat_entry_point == nullptr || class_linker->IsQuickGenericJniStub(oat_entry_point)) {
+    DCHECK(IsNative());
+    return nullptr;
+  }
+
+  OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromEntryPoint(oat_entry_point);
+  if (pc == 0) {
+    // This is a downcall, it can only happen for a native method.
+    DCHECK(IsNative());
+    return method_header;
+  }
+
+  if (pc == reinterpret_cast<uintptr_t>(GetQuickInstrumentationExitPc())) {
+    // If we're instrumenting, just return the compiled OAT code.
+    // TODO(ngeoffray): Avoid this call path.
+    return method_header;
+  }
+
+  DCHECK(method_header->Contains(pc))
+      << PrettyMethod(this)
+      << std::hex << pc << " " << oat_entry_point
+      << " " << (uintptr_t)(method_header->code_ + method_header->code_size_);
   return method_header;
 }
 
diff --git a/runtime/art_method.h b/runtime/art_method.h
index 9f1495c..bb9804e 100644
--- a/runtime/art_method.h
+++ b/runtime/art_method.h
@@ -433,6 +433,10 @@
     return ++hotness_count_;
   }
 
+  void ClearCounter() {
+    hotness_count_ = 0;
+  }
+
   const uint8_t* GetQuickenedInfo() SHARED_REQUIRES(Locks::mutator_lock_);
 
   // Returns the method header for the compiled code containing 'pc'. Note that runtime
diff --git a/runtime/base/arena_allocator.cc b/runtime/base/arena_allocator.cc
index 1704688..71afa0f 100644
--- a/runtime/base/arena_allocator.cc
+++ b/runtime/base/arena_allocator.cc
@@ -94,6 +94,8 @@
   "CodeGen      ",
   "ParallelMove ",
   "GraphChecker ",
+  "LSE          ",
+  "Verifier     ",
 };
 
 template <bool kCount>
diff --git a/runtime/base/arena_allocator.h b/runtime/base/arena_allocator.h
index 7747bdf..ace6c38 100644
--- a/runtime/base/arena_allocator.h
+++ b/runtime/base/arena_allocator.h
@@ -107,6 +107,7 @@
   kArenaAllocParallelMoveResolver,
   kArenaAllocGraphChecker,
   kArenaAllocLSE,
+  kArenaAllocVerifier,
   kNumArenaAllocKinds
 };
 
diff --git a/runtime/base/dchecked_vector.h b/runtime/base/dchecked_vector.h
index 6ec573a..2bd12df 100644
--- a/runtime/base/dchecked_vector.h
+++ b/runtime/base/dchecked_vector.h
@@ -59,8 +59,10 @@
       : Base() { }
   explicit dchecked_vector(const allocator_type& alloc)
       : Base(alloc) { }
+  // Note that we cannot forward to std::vector(size_type, const allocator_type&) because it is not
+  // available in C++11, which is the latest GCC can support. http://b/25022512
   explicit dchecked_vector(size_type n, const allocator_type& alloc = allocator_type())
-      : Base(n, alloc) { }
+      : Base(alloc) { resize(n); }
   dchecked_vector(size_type n,
                   const value_type& value,
                   const allocator_type& alloc = allocator_type())
diff --git a/runtime/base/hash_set.h b/runtime/base/hash_set.h
index 4819f06..95baa82 100644
--- a/runtime/base/hash_set.h
+++ b/runtime/base/hash_set.h
@@ -420,6 +420,19 @@
     Resize(Size() / max_load_factor_);
   }
 
+  // Reserve enough room to insert until Size() == num_elements without requiring to grow the hash
+  // set. No-op if the hash set is already large enough to do this.
+  void Reserve(size_t num_elements) {
+    size_t num_buckets = num_elements / max_load_factor_;
+    // Deal with rounding errors. Add one for rounding.
+    while (static_cast<size_t>(num_buckets * max_load_factor_) <= num_elements + 1u) {
+      ++num_buckets;
+    }
+    if (num_buckets > NumBuckets()) {
+      Resize(num_buckets);
+    }
+  }
+
   // To distance that inserted elements were probed. Used for measuring how good hash functions
   // are.
   size_t TotalProbeDistance() const {
@@ -488,6 +501,15 @@
     }
   }
 
+  // The hash set expands when Size() reaches ElementsUntilExpand().
+  size_t ElementsUntilExpand() const {
+    return elements_until_expand_;
+  }
+
+  size_t NumBuckets() const {
+    return num_buckets_;
+  }
+
  private:
   T& ElementForIndex(size_t index) {
     DCHECK_LT(index, NumBuckets());
@@ -543,10 +565,6 @@
     return emptyfn_.IsEmpty(ElementForIndex(index));
   }
 
-  size_t NumBuckets() const {
-    return num_buckets_;
-  }
-
   // Allocate a number of buckets.
   void AllocateStorage(size_t num_buckets) {
     num_buckets_ = num_buckets;
diff --git a/runtime/base/hash_set_test.cc b/runtime/base/hash_set_test.cc
index 743e98e..8254063 100644
--- a/runtime/base/hash_set_test.cc
+++ b/runtime/base/hash_set_test.cc
@@ -333,4 +333,25 @@
   ASSERT_NE(hash_set.end(), hash_set.Find(std::forward_list<int>({1, 2, 3, 4})));
 }
 
+TEST_F(HashSetTest, TestReserve) {
+  HashSet<std::string, IsEmptyFnString> hash_set;
+  std::vector<size_t> sizes = {1, 10, 25, 55, 128, 1024, 4096};
+  for (size_t size : sizes) {
+    hash_set.Reserve(size);
+    const size_t buckets_before = hash_set.NumBuckets();
+    // Check that we expanded enough.
+    CHECK_GE(hash_set.ElementsUntilExpand(), size);
+    // Try inserting elements until we are at our reserve size and ensure the hash set did not
+    // expand.
+    while (hash_set.Size() < size) {
+      hash_set.Insert(std::to_string(hash_set.Size()));
+    }
+    CHECK_EQ(hash_set.NumBuckets(), buckets_before);
+  }
+  // Check the behaviour for shrinking, it does not necessarily resize down.
+  constexpr size_t size = 100;
+  hash_set.Reserve(size);
+  CHECK_GE(hash_set.ElementsUntilExpand(), size);
+}
+
 }  // namespace art
diff --git a/runtime/base/scoped_arena_containers.h b/runtime/base/scoped_arena_containers.h
index b74aef1..1236585 100644
--- a/runtime/base/scoped_arena_containers.h
+++ b/runtime/base/scoped_arena_containers.h
@@ -20,6 +20,7 @@
 #include <deque>
 #include <queue>
 #include <set>
+#include <type_traits>
 #include <unordered_map>
 #include <utility>
 
@@ -218,12 +219,19 @@
   }
 };
 
-// Declare but do not define a partial specialization for T[].
-// This is to prevent accidental use of this unsupported use case.
+// In general we lack support for arrays. We would need to call the destructor on each element,
+// which requires access to the array size. Support for that is future work.
+//
+// However, we can support trivially destructible component types, as then a destructor doesn't
+// need to be called.
 template <typename T>
 class ArenaDelete<T[]> {
  public:
-  void operator()(T* ptr) const = delete;
+  void operator()(T* ptr ATTRIBUTE_UNUSED) const {
+    static_assert(std::is_trivially_destructible<T>::value,
+                  "ArenaUniquePtr does not support non-trivially-destructible arrays.");
+    // TODO: Implement debug checks, and MEMORY_TOOL support.
+  }
 };
 
 // Arena unique ptr that only calls the destructor of the element.
diff --git a/runtime/base/stringpiece.h b/runtime/base/stringpiece.h
index 9c83cf5..46743e9 100644
--- a/runtime/base/stringpiece.h
+++ b/runtime/base/stringpiece.h
@@ -165,7 +165,7 @@
   // Pointer to char data, not necessarily zero terminated.
   const char* ptr_;
   // Length of data.
-  size_type   length_;
+  size_type length_;
 };
 
 // This large function is defined inline so that in a fairly common case where
diff --git a/runtime/class_linker.cc b/runtime/class_linker.cc
index 81622e1..69d0799 100644
--- a/runtime/class_linker.cc
+++ b/runtime/class_linker.cc
@@ -1174,15 +1174,26 @@
   mirror::LongArray::ResetArrayClass();
   mirror::ShortArray::ResetArrayClass();
   Thread* const self = Thread::Current();
-  JavaVMExt* const vm = Runtime::Current()->GetJavaVM();
   for (const ClassLoaderData& data : class_loaders_) {
-    vm->DeleteWeakGlobalRef(self, data.weak_root);
-    delete data.allocator;
-    delete data.class_table;
+    DeleteClassLoader(self, data);
   }
   class_loaders_.clear();
 }
 
+void ClassLinker::DeleteClassLoader(Thread* self, const ClassLoaderData& data) {
+  Runtime* const runtime = Runtime::Current();
+  JavaVMExt* const vm = runtime->GetJavaVM();
+  vm->DeleteWeakGlobalRef(self, data.weak_root);
+  if (runtime->GetJit() != nullptr) {
+    jit::JitCodeCache* code_cache = runtime->GetJit()->GetCodeCache();
+    if (code_cache != nullptr) {
+      code_cache->RemoveMethodsIn(self, *data.allocator);
+    }
+  }
+  delete data.allocator;
+  delete data.class_table;
+}
+
 mirror::PointerArray* ClassLinker::AllocPointerArray(Thread* self, size_t length) {
   return down_cast<mirror::PointerArray*>(image_pointer_size_ == 8u ?
       static_cast<mirror::Array*>(mirror::LongArray::Alloc(self, length)) :
@@ -1833,13 +1844,6 @@
       return code;
     }
   }
-  jit::Jit* const jit = Runtime::Current()->GetJit();
-  if (jit != nullptr) {
-    auto* code = jit->GetCodeCache()->GetCodeFor(method);
-    if (code != nullptr) {
-      return code;
-    }
-  }
   if (method->IsNative()) {
     // No code and native? Use generic trampoline.
     return GetQuickGenericJniStub();
@@ -1856,13 +1860,6 @@
   if (found) {
     return oat_method.GetQuickCode();
   }
-  jit::Jit* jit = Runtime::Current()->GetJit();
-  if (jit != nullptr) {
-    auto* code = jit->GetCodeCache()->GetCodeFor(method);
-    if (code != nullptr) {
-      return code;
-    }
-  }
   return nullptr;
 }
 
@@ -5325,7 +5322,7 @@
     ScopedArenaUnorderedMap<ArtMethod*, ArtMethod*> move_table(allocator.Adapter());
     if (virtuals != old_virtuals) {
       // Maps from heap allocated miranda method to linear alloc miranda method.
-      StrideIterator<ArtMethod> out = virtuals->Begin(method_size, method_alignment);
+      StrideIterator<ArtMethod> out = virtuals->begin(method_size, method_alignment);
       // Copy over the old methods + miranda methods.
       for (auto& m : klass->GetVirtualMethods(image_pointer_size_)) {
         move_table.emplace(&m, &*out);
@@ -5335,7 +5332,7 @@
         ++out;
       }
     }
-    StrideIterator<ArtMethod> out(virtuals->Begin(method_size, method_alignment)
+    StrideIterator<ArtMethod> out(virtuals->begin(method_size, method_alignment)
                                       + old_method_count);
     // Copy over miranda methods before copying vtable since CopyOf may cause thread suspension and
     // we want the roots of the miranda methods to get visited.
@@ -5367,7 +5364,7 @@
       move_table.emplace(def_method, &new_method);
       ++out;
     }
-    virtuals->SetLength(new_method_count);
+    virtuals->SetSize(new_method_count);
     UpdateClassVirtualMethods(klass.Get(), virtuals);
     // Done copying methods, they are all roots in the class now, so we can end the no thread
     // suspension assert.
@@ -5382,7 +5379,7 @@
       self->AssertPendingOOMException();
       return false;
     }
-    out = virtuals->Begin(method_size, method_alignment) + old_method_count;
+    out = virtuals->begin(method_size, method_alignment) + old_method_count;
     size_t vtable_pos = old_vtable_count;
     for (size_t i = old_method_count; i < new_method_count; ++i) {
       // Leave the declaring class alone as type indices are relative to it
@@ -6387,7 +6384,6 @@
 void ClassLinker::CleanupClassLoaders() {
   Thread* const self = Thread::Current();
   WriterMutexLock mu(self, *Locks::classlinker_classes_lock_);
-  JavaVMExt* const vm = Runtime::Current()->GetJavaVM();
   for (auto it = class_loaders_.begin(); it != class_loaders_.end(); ) {
     const ClassLoaderData& data = *it;
     // Need to use DecodeJObject so that we get null for cleared JNI weak globals.
@@ -6395,10 +6391,7 @@
     if (class_loader != nullptr) {
       ++it;
     } else {
-      // Weak reference was cleared, delete the data associated with this class loader.
-      delete data.class_table;
-      delete data.allocator;
-      vm->DeleteWeakGlobalRef(self, data.weak_root);
+      DeleteClassLoader(self, data);
       it = class_loaders_.erase(it);
     }
   }
diff --git a/runtime/class_linker.h b/runtime/class_linker.h
index a2d38ac..392efd2 100644
--- a/runtime/class_linker.h
+++ b/runtime/class_linker.h
@@ -551,6 +551,10 @@
     LinearAlloc* allocator;
   };
 
+  static void DeleteClassLoader(Thread* self, const ClassLoaderData& data)
+      REQUIRES(Locks::classlinker_classes_lock_)
+      SHARED_REQUIRES(Locks::mutator_lock_);
+
   void VisitClassLoaders(ClassLoaderVisitor* visitor) const
       SHARED_REQUIRES(Locks::classlinker_classes_lock_, Locks::mutator_lock_);
 
diff --git a/runtime/dex_instruction.h b/runtime/dex_instruction.h
index 48a12e5..035230e 100644
--- a/runtime/dex_instruction.h
+++ b/runtime/dex_instruction.h
@@ -185,6 +185,7 @@
 
   static constexpr uint32_t kMaxVarArgRegs = 5;
   static constexpr uint32_t kMaxVarArgRegs25x = 6;  // lambdas are 2 registers.
+  static constexpr uint32_t kLambdaVirtualRegisterWidth = 2;
 
   // Returns the size (in 2 byte code units) of this instruction.
   size_t SizeInCodeUnits() const {
@@ -248,7 +249,7 @@
 
   // VRegA
   bool HasVRegA() const;
-  int32_t VRegA() const;
+  ALWAYS_INLINE int32_t VRegA() const;
 
   int8_t VRegA_10t() const {
     return VRegA_10t(Fetch16(0));
diff --git a/runtime/exception_test.cc b/runtime/exception_test.cc
index 4de8a8e..b1d4d35 100644
--- a/runtime/exception_test.cc
+++ b/runtime/exception_test.cc
@@ -93,7 +93,7 @@
                                       fake_code_.begin(), fake_code_.end());
 
     // NOTE: Don't align the code (it will not be executed) but check that the Thumb2
-    // adjustment will be a NOP, see ArtMethod::EntryPointToCodePointer().
+    // adjustment will be a NOP, see EntryPointToCodePointer().
     CHECK_ALIGNED(mapping_table_offset, 2);
     const uint8_t* code_ptr = &fake_header_code_and_maps_[gc_map_offset];
 
diff --git a/runtime/gc/accounting/bitmap.cc b/runtime/gc/accounting/bitmap.cc
index fdded02..380cb8e 100644
--- a/runtime/gc/accounting/bitmap.cc
+++ b/runtime/gc/accounting/bitmap.cc
@@ -18,6 +18,7 @@
 
 #include "base/bit_utils.h"
 #include "card_table.h"
+#include "jit/jit_code_cache.h"
 #include "mem_map.h"
 
 namespace art {
@@ -91,6 +92,7 @@
 }
 
 template class MemoryRangeBitmap<CardTable::kCardSize>;
+template class MemoryRangeBitmap<jit::kJitCodeAlignment>;
 
 }  // namespace accounting
 }  // namespace gc
diff --git a/runtime/gc/heap.cc b/runtime/gc/heap.cc
index 657fcb5..1d38525 100644
--- a/runtime/gc/heap.cc
+++ b/runtime/gc/heap.cc
@@ -26,6 +26,7 @@
 
 #include "art_field-inl.h"
 #include "base/allocator.h"
+#include "base/arena_allocator.h"
 #include "base/dumpable.h"
 #include "base/histogram-inl.h"
 #include "base/stl_util.h"
@@ -1258,11 +1259,11 @@
 }
 
 void Heap::Trim(Thread* self) {
+  Runtime* const runtime = Runtime::Current();
   if (!CareAboutPauseTimes()) {
     ATRACE_BEGIN("Deflating monitors");
     // Deflate the monitors, this can cause a pause but shouldn't matter since we don't care
     // about pauses.
-    Runtime* runtime = Runtime::Current();
     {
       ScopedSuspendAll ssa(__FUNCTION__);
       uint64_t start_time = NanoTime();
@@ -1274,6 +1275,10 @@
   }
   TrimIndirectReferenceTables(self);
   TrimSpaces(self);
+  // Trim arenas that may have been used by JIT or verifier.
+  ATRACE_BEGIN("Trimming arena maps");
+  runtime->GetArenaPool()->TrimMaps();
+  ATRACE_END();
 }
 
 class TrimIndirectReferenceTableClosure : public Closure {
diff --git a/runtime/image.cc b/runtime/image.cc
index 192371f..1bc19ff 100644
--- a/runtime/image.cc
+++ b/runtime/image.cc
@@ -150,10 +150,10 @@
 void ImageSection::VisitPackedArtFields(ArtFieldVisitor* visitor, uint8_t* base) const {
   for (size_t pos = 0; pos < Size(); ) {
     auto* array = reinterpret_cast<LengthPrefixedArray<ArtField>*>(base + Offset() + pos);
-    for (size_t i = 0; i < array->Length(); ++i) {
+    for (size_t i = 0; i < array->size(); ++i) {
       visitor->Visit(&array->At(i, sizeof(ArtField)));
     }
-    pos += array->ComputeSize(array->Length());
+    pos += array->ComputeSize(array->size());
   }
 }
 
@@ -164,10 +164,10 @@
   const size_t method_size = ArtMethod::Size(pointer_size);
   for (size_t pos = 0; pos < Size(); ) {
     auto* array = reinterpret_cast<LengthPrefixedArray<ArtMethod>*>(base + Offset() + pos);
-    for (size_t i = 0; i < array->Length(); ++i) {
+    for (size_t i = 0; i < array->size(); ++i) {
       visitor->Visit(&array->At(i, method_size, method_alignment));
     }
-    pos += array->ComputeSize(array->Length(), method_size, method_alignment);
+    pos += array->ComputeSize(array->size(), method_size, method_alignment);
   }
 }
 
diff --git a/runtime/instrumentation.cc b/runtime/instrumentation.cc
index ed64d7e..4db37e6 100644
--- a/runtime/instrumentation.cc
+++ b/runtime/instrumentation.cc
@@ -97,16 +97,6 @@
 
 static void UpdateEntrypoints(ArtMethod* method, const void* quick_code)
     SHARED_REQUIRES(Locks::mutator_lock_) {
-  Runtime* const runtime = Runtime::Current();
-  jit::Jit* jit = runtime->GetJit();
-  if (jit != nullptr) {
-    const void* old_code_ptr = method->GetEntryPointFromQuickCompiledCode();
-    jit::JitCodeCache* code_cache = jit->GetCodeCache();
-    if (code_cache->ContainsCodePtr(old_code_ptr)) {
-      // Save the old compiled code since we need it to implement ClassLinker::GetQuickOatCodeFor.
-      code_cache->SaveCompiledCode(method, old_code_ptr);
-    }
-  }
   method->SetEntryPointFromQuickCompiledCode(quick_code);
 }
 
diff --git a/runtime/jit/jit.cc b/runtime/jit/jit.cc
index 0607493..5afd28e 100644
--- a/runtime/jit/jit.cc
+++ b/runtime/jit/jit.cc
@@ -49,7 +49,7 @@
 void Jit::DumpInfo(std::ostream& os) {
   os << "Code cache size=" << PrettySize(code_cache_->CodeCacheSize())
      << " data cache size=" << PrettySize(code_cache_->DataCacheSize())
-     << " num methods=" << code_cache_->NumMethods()
+     << " number of compiled code=" << code_cache_->NumberOfCompiledCode()
      << "\n";
   cumulative_timings_.Dump(os);
 }
diff --git a/runtime/jit/jit_code_cache.cc b/runtime/jit/jit_code_cache.cc
index 4187358..2d0a2a5 100644
--- a/runtime/jit/jit_code_cache.cc
+++ b/runtime/jit/jit_code_cache.cc
@@ -19,8 +19,12 @@
 #include <sstream>
 
 #include "art_method-inl.h"
+#include "entrypoints/runtime_asm_entrypoints.h"
+#include "gc/accounting/bitmap-inl.h"
+#include "linear_alloc.h"
 #include "mem_map.h"
 #include "oat_file-inl.h"
+#include "thread_list.h"
 
 namespace art {
 namespace jit {
@@ -74,14 +78,10 @@
 
 JitCodeCache::JitCodeCache(MemMap* code_map, MemMap* data_map)
     : lock_("Jit code cache", kJitCodeCacheLock),
+      lock_cond_("Jit code cache variable", lock_),
+      collection_in_progress_(false),
       code_map_(code_map),
-      data_map_(data_map),
-      num_methods_(0) {
-
-  VLOG(jit) << "Created jit code cache: data size="
-            << PrettySize(data_map_->Size())
-            << ", code size="
-            << PrettySize(code_map_->Size());
+      data_map_(data_map) {
 
   code_mspace_ = create_mspace_with_base(code_map_->Begin(), code_map_->Size(), false /*locked*/);
   data_mspace_ = create_mspace_with_base(data_map_->Begin(), data_map_->Size(), false /*locked*/);
@@ -96,13 +96,22 @@
 
   CHECKED_MPROTECT(code_map_->Begin(), code_map_->Size(), kProtCode);
   CHECKED_MPROTECT(data_map_->Begin(), data_map_->Size(), kProtData);
+
+  live_bitmap_.reset(CodeCacheBitmap::Create("code-cache-bitmap",
+                                             reinterpret_cast<uintptr_t>(code_map_->Begin()),
+                                             reinterpret_cast<uintptr_t>(code_map_->End())));
+
+  if (live_bitmap_.get() == nullptr) {
+    PLOG(FATAL) << "creating bitmaps for the JIT code cache failed";
+  }
+
+  VLOG(jit) << "Created jit code cache: data size="
+            << PrettySize(data_map_->Size())
+            << ", code size="
+            << PrettySize(code_map_->Size());
 }
 
-bool JitCodeCache::ContainsMethod(ArtMethod* method) const {
-  return ContainsCodePtr(method->GetEntryPointFromQuickCompiledCode());
-}
-
-bool JitCodeCache::ContainsCodePtr(const void* ptr) const {
+bool JitCodeCache::ContainsPc(const void* ptr) const {
   return code_map_->Begin() <= ptr && ptr < code_map_->End();
 }
 
@@ -121,6 +130,7 @@
 };
 
 uint8_t* JitCodeCache::CommitCode(Thread* self,
+                                  ArtMethod* method,
                                   const uint8_t* mapping_table,
                                   const uint8_t* vmap_table,
                                   const uint8_t* gc_map,
@@ -129,6 +139,93 @@
                                   size_t fp_spill_mask,
                                   const uint8_t* code,
                                   size_t code_size) {
+  uint8_t* result = CommitCodeInternal(self,
+                                       method,
+                                       mapping_table,
+                                       vmap_table,
+                                       gc_map,
+                                       frame_size_in_bytes,
+                                       core_spill_mask,
+                                       fp_spill_mask,
+                                       code,
+                                       code_size);
+  if (result == nullptr) {
+    // Retry.
+    GarbageCollectCache(self);
+    result = CommitCodeInternal(self,
+                                method,
+                                mapping_table,
+                                vmap_table,
+                                gc_map,
+                                frame_size_in_bytes,
+                                core_spill_mask,
+                                fp_spill_mask,
+                                code,
+                                code_size);
+  }
+  return result;
+}
+
+bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
+  bool in_collection = false;
+  while (collection_in_progress_) {
+    in_collection = true;
+    lock_cond_.Wait(self);
+  }
+  return in_collection;
+}
+
+static uintptr_t FromCodeToAllocation(const void* code) {
+  size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
+  return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
+}
+
+void JitCodeCache::FreeCode(const void* code_ptr, ArtMethod* method ATTRIBUTE_UNUSED) {
+  uintptr_t allocation = FromCodeToAllocation(code_ptr);
+  const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
+  const uint8_t* data = method_header->GetNativeGcMap();
+  if (data != nullptr) {
+    mspace_free(data_mspace_, const_cast<uint8_t*>(data));
+  }
+  data = method_header->GetMappingTable();
+  if (data != nullptr) {
+    mspace_free(data_mspace_, const_cast<uint8_t*>(data));
+  }
+  // Use the offset directly to prevent sanity check that the method is
+  // compiled with optimizing.
+  // TODO(ngeoffray): Clean up.
+  if (method_header->vmap_table_offset_ != 0) {
+    data = method_header->code_ - method_header->vmap_table_offset_;
+    mspace_free(data_mspace_, const_cast<uint8_t*>(data));
+  }
+  mspace_free(code_mspace_, reinterpret_cast<uint8_t*>(allocation));
+}
+
+void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
+  MutexLock mu(self, lock_);
+  // We do not check if a code cache GC is in progress, as this method comes
+  // with the classlinker_classes_lock_ held, and suspending ourselves could
+  // lead to a deadlock.
+  for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
+    if (alloc.ContainsUnsafe(it->second)) {
+      FreeCode(it->first, it->second);
+      it = method_code_map_.erase(it);
+    } else {
+      ++it;
+    }
+  }
+}
+
+uint8_t* JitCodeCache::CommitCodeInternal(Thread* self,
+                                          ArtMethod* method,
+                                          const uint8_t* mapping_table,
+                                          const uint8_t* vmap_table,
+                                          const uint8_t* gc_map,
+                                          size_t frame_size_in_bytes,
+                                          size_t core_spill_mask,
+                                          size_t fp_spill_mask,
+                                          const uint8_t* code,
+                                          size_t code_size) {
   size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
   // Ensure the header ends up at expected instruction alignment.
   size_t header_size = RoundUp(sizeof(OatQuickMethodHeader), alignment);
@@ -137,7 +234,9 @@
   OatQuickMethodHeader* method_header = nullptr;
   uint8_t* code_ptr = nullptr;
 
+  ScopedThreadSuspension sts(self, kSuspended);
   MutexLock mu(self, lock_);
+  WaitForPotentialCollectionToComplete(self);
   {
     ScopedCodeCacheWrite scc(code_map_.get());
     uint8_t* result = reinterpret_cast<uint8_t*>(
@@ -149,7 +248,7 @@
     DCHECK_ALIGNED_PARAM(reinterpret_cast<uintptr_t>(code_ptr), alignment);
 
     std::copy(code, code + code_size, code_ptr);
-    method_header = reinterpret_cast<OatQuickMethodHeader*>(code_ptr) - 1;
+    method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
     new (method_header) OatQuickMethodHeader(
         (mapping_table == nullptr) ? 0 : code_ptr - mapping_table,
         (vmap_table == nullptr) ? 0 : code_ptr - vmap_table,
@@ -162,8 +261,12 @@
 
   __builtin___clear_cache(reinterpret_cast<char*>(code_ptr),
                           reinterpret_cast<char*>(code_ptr + code_size));
-
-  ++num_methods_;  // TODO: This is hacky but works since each method has exactly one code region.
+  method_code_map_.Put(code_ptr, method);
+  // We have checked there was no collection in progress earlier. If we
+  // were, setting the entry point of a method would be unsafe, as the collection
+  // could delete it.
+  DCHECK(!collection_in_progress_);
+  method->SetEntryPointFromQuickCompiledCode(method_header->GetEntryPoint());
   return reinterpret_cast<uint8_t*>(method_header);
 }
 
@@ -181,10 +284,32 @@
   return bytes_allocated;
 }
 
+size_t JitCodeCache::NumberOfCompiledCode() {
+  MutexLock mu(Thread::Current(), lock_);
+  return method_code_map_.size();
+}
+
 uint8_t* JitCodeCache::ReserveData(Thread* self, size_t size) {
   size = RoundUp(size, sizeof(void*));
-  MutexLock mu(self, lock_);
-  return reinterpret_cast<uint8_t*>(mspace_malloc(data_mspace_, size));
+  uint8_t* result = nullptr;
+
+  {
+    ScopedThreadSuspension sts(self, kSuspended);
+    MutexLock mu(self, lock_);
+    WaitForPotentialCollectionToComplete(self);
+    result = reinterpret_cast<uint8_t*>(mspace_malloc(data_mspace_, size));
+  }
+
+  if (result == nullptr) {
+    // Retry.
+    GarbageCollectCache(self);
+    ScopedThreadSuspension sts(self, kSuspended);
+    MutexLock mu(self, lock_);
+    WaitForPotentialCollectionToComplete(self);
+    result = reinterpret_cast<uint8_t*>(mspace_malloc(data_mspace_, size));
+  }
+
+  return result;
 }
 
 uint8_t* JitCodeCache::AddDataArray(Thread* self, const uint8_t* begin, const uint8_t* end) {
@@ -196,29 +321,143 @@
   return result;
 }
 
-const void* JitCodeCache::GetCodeFor(ArtMethod* method) {
-  const void* code = method->GetEntryPointFromQuickCompiledCode();
-  if (ContainsCodePtr(code)) {
-    return code;
+class MarkCodeVisitor FINAL : public StackVisitor {
+ public:
+  MarkCodeVisitor(Thread* thread_in, JitCodeCache* code_cache_in)
+      : StackVisitor(thread_in, nullptr, StackVisitor::StackWalkKind::kSkipInlinedFrames),
+        code_cache_(code_cache_in),
+        bitmap_(code_cache_->GetLiveBitmap()) {}
+
+  bool VisitFrame() OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_) {
+    const OatQuickMethodHeader* method_header = GetCurrentOatQuickMethodHeader();
+    if (method_header == nullptr) {
+      return true;
+    }
+    const void* code = method_header->GetCode();
+    if (code_cache_->ContainsPc(code)) {
+      // Use the atomic set version, as multiple threads are executing this code.
+      bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
+    }
+    return true;
   }
-  MutexLock mu(Thread::Current(), lock_);
-  auto it = method_code_map_.find(method);
-  if (it != method_code_map_.end()) {
-    return it->second;
+
+ private:
+  JitCodeCache* const code_cache_;
+  CodeCacheBitmap* const bitmap_;
+};
+
+class MarkCodeClosure FINAL : public Closure {
+ public:
+  MarkCodeClosure(JitCodeCache* code_cache, Barrier* barrier)
+      : code_cache_(code_cache), barrier_(barrier) {}
+
+  void Run(Thread* thread) OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_) {
+    DCHECK(thread == Thread::Current() || thread->IsSuspended());
+    MarkCodeVisitor visitor(thread, code_cache_);
+    visitor.WalkStack();
+    if (thread->GetState() == kRunnable) {
+      barrier_->Pass(Thread::Current());
+    }
   }
-  return nullptr;
+
+ private:
+  JitCodeCache* const code_cache_;
+  Barrier* const barrier_;
+};
+
+void JitCodeCache::GarbageCollectCache(Thread* self) {
+  if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
+    LOG(INFO) << "Clearing code cache, code="
+              << PrettySize(CodeCacheSize())
+              << ", data=" << PrettySize(DataCacheSize());
+  }
+
+  size_t map_size = 0;
+  ScopedThreadSuspension sts(self, kSuspended);
+
+  // Walk over all compiled methods and set the entry points of these
+  // methods to interpreter.
+  {
+    MutexLock mu(self, lock_);
+    if (WaitForPotentialCollectionToComplete(self)) {
+      return;
+    }
+    collection_in_progress_ = true;
+    map_size = method_code_map_.size();
+    for (auto& it : method_code_map_) {
+      it.second->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
+    }
+  }
+
+  // Run a checkpoint on all threads to mark the JIT compiled code they are running.
+  {
+    Barrier barrier(0);
+    MarkCodeClosure closure(this, &barrier);
+    size_t threads_running_checkpoint =
+        Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
+    if (threads_running_checkpoint != 0) {
+      barrier.Increment(self, threads_running_checkpoint);
+    }
+  }
+
+  // Free unused compiled code, and restore the entry point of used compiled code.
+  {
+    MutexLock mu(self, lock_);
+    DCHECK_EQ(map_size, method_code_map_.size());
+    ScopedCodeCacheWrite scc(code_map_.get());
+    for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
+      const void* code_ptr = it->first;
+      ArtMethod* method = it->second;
+      uintptr_t allocation = FromCodeToAllocation(code_ptr);
+      const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
+      if (GetLiveBitmap()->Test(allocation)) {
+        method->SetEntryPointFromQuickCompiledCode(method_header->GetEntryPoint());
+        ++it;
+      } else {
+        method->ClearCounter();
+        DCHECK_NE(method->GetEntryPointFromQuickCompiledCode(), method_header->GetEntryPoint());
+        FreeCode(code_ptr, method);
+        it = method_code_map_.erase(it);
+      }
+    }
+    GetLiveBitmap()->Bitmap::Clear();
+    collection_in_progress_ = false;
+    lock_cond_.Broadcast(self);
+  }
+
+  if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
+    LOG(INFO) << "After clearing code cache, code="
+              << PrettySize(CodeCacheSize())
+              << ", data=" << PrettySize(DataCacheSize());
+  }
 }
 
-void JitCodeCache::SaveCompiledCode(ArtMethod* method, const void* old_code_ptr) {
-  DCHECK_EQ(method->GetEntryPointFromQuickCompiledCode(), old_code_ptr);
-  DCHECK(ContainsCodePtr(old_code_ptr)) << PrettyMethod(method) << " old_code_ptr="
-      << old_code_ptr;
-  MutexLock mu(Thread::Current(), lock_);
-  auto it = method_code_map_.find(method);
-  if (it != method_code_map_.end()) {
-    return;
+
+OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
+  static_assert(kRuntimeISA != kThumb2, "kThumb2 cannot be a runtime ISA");
+  if (kRuntimeISA == kArm) {
+    // On Thumb-2, the pc is offset by one.
+    --pc;
   }
-  method_code_map_.Put(method, old_code_ptr);
+  if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
+    return nullptr;
+  }
+
+  MutexLock mu(Thread::Current(), lock_);
+  if (method_code_map_.empty()) {
+    return nullptr;
+  }
+  auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
+  --it;
+
+  const void* code_ptr = it->first;
+  OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
+  if (!method_header->Contains(pc)) {
+    return nullptr;
+  }
+  DCHECK_EQ(it->second, method)
+      << PrettyMethod(method) << " " << PrettyMethod(it->second) << " " << std::hex << pc;
+  return method_header;
 }
 
 }  // namespace jit
diff --git a/runtime/jit/jit_code_cache.h b/runtime/jit/jit_code_cache.h
index fa90c18..4e415b8 100644
--- a/runtime/jit/jit_code_cache.h
+++ b/runtime/jit/jit_code_cache.h
@@ -22,6 +22,7 @@
 #include "atomic.h"
 #include "base/macros.h"
 #include "base/mutex.h"
+#include "gc/accounting/bitmap.h"
 #include "gc/allocator/dlmalloc.h"
 #include "gc_root.h"
 #include "jni.h"
@@ -33,32 +34,40 @@
 namespace art {
 
 class ArtMethod;
-class CompiledMethod;
-class CompilerCallbacks;
+class LinearAlloc;
 
 namespace jit {
 
 class JitInstrumentationCache;
 
+// Alignment that will suit all architectures.
+static constexpr int kJitCodeAlignment = 16;
+using CodeCacheBitmap = gc::accounting::MemoryRangeBitmap<kJitCodeAlignment>;
+
 class JitCodeCache {
  public:
   static constexpr size_t kMaxCapacity = 1 * GB;
-  static constexpr size_t kDefaultCapacity = 2 * MB;
+  // Put the default to a very low amount for debug builds to stress the code cache
+  // collection.
+  static constexpr size_t kDefaultCapacity = kIsDebugBuild ? 20 * KB : 2 * MB;
 
   // Create the code cache with a code + data capacity equal to "capacity", error message is passed
   // in the out arg error_msg.
   static JitCodeCache* Create(size_t capacity, std::string* error_msg);
 
-  size_t NumMethods() const {
-    return num_methods_;
-  }
-
+  // Number of bytes allocated in the code cache.
   size_t CodeCacheSize() REQUIRES(!lock_);
 
+  // Number of bytes allocated in the data cache.
   size_t DataCacheSize() REQUIRES(!lock_);
 
+  // Number of compiled code in the code cache. Note that this is not the number
+  // of methods that got JIT compiled, as we might have collected some.
+  size_t NumberOfCompiledCode() REQUIRES(!lock_);
+
   // Allocate and write code and its metadata to the code cache.
   uint8_t* CommitCode(Thread* self,
+                      ArtMethod* method,
                       const uint8_t* mapping_table,
                       const uint8_t* vmap_table,
                       const uint8_t* gc_map,
@@ -67,51 +76,89 @@
                       size_t fp_spill_mask,
                       const uint8_t* code,
                       size_t code_size)
+      SHARED_REQUIRES(Locks::mutator_lock_)
       REQUIRES(!lock_);
 
-  // Return true if the code cache contains the code pointer which si the entrypoint of the method.
-  bool ContainsMethod(ArtMethod* method) const
-      SHARED_REQUIRES(Locks::mutator_lock_);
-
-  // Return true if the code cache contains a code ptr.
-  bool ContainsCodePtr(const void* ptr) const;
+  // Return true if the code cache contains this pc.
+  bool ContainsPc(const void* pc) const;
 
   // Reserve a region of data of size at least "size". Returns null if there is no more room.
-  uint8_t* ReserveData(Thread* self, size_t size) REQUIRES(!lock_);
+  uint8_t* ReserveData(Thread* self, size_t size)
+      SHARED_REQUIRES(Locks::mutator_lock_)
+      REQUIRES(!lock_);
 
   // Add a data array of size (end - begin) with the associated contents, returns null if there
   // is no more room.
   uint8_t* AddDataArray(Thread* self, const uint8_t* begin, const uint8_t* end)
+      SHARED_REQUIRES(Locks::mutator_lock_)
       REQUIRES(!lock_);
 
-  // Get code for a method, returns null if it is not in the jit cache.
-  const void* GetCodeFor(ArtMethod* method)
-      SHARED_REQUIRES(Locks::mutator_lock_) REQUIRES(!lock_);
+  CodeCacheBitmap* GetLiveBitmap() const {
+    return live_bitmap_.get();
+  }
 
-  // Save the compiled code for a method so that GetCodeFor(method) will return old_code_ptr if the
-  // entrypoint isn't within the cache.
-  void SaveCompiledCode(ArtMethod* method, const void* old_code_ptr)
-      SHARED_REQUIRES(Locks::mutator_lock_) REQUIRES(!lock_);
+  // Perform a collection on the code cache.
+  void GarbageCollectCache(Thread* self)
+      REQUIRES(!lock_)
+      SHARED_REQUIRES(Locks::mutator_lock_);
+
+  // Given the 'pc', try to find the JIT compiled code associated with it.
+  // Return null if 'pc' is not in the code cache. 'method' is passed for
+  // sanity check.
+  OatQuickMethodHeader* LookupMethodHeader(uintptr_t pc, ArtMethod* method)
+      REQUIRES(!lock_)
+      SHARED_REQUIRES(Locks::mutator_lock_);
+
+  void RemoveMethodsIn(Thread* self, const LinearAlloc& alloc)
+      REQUIRES(!lock_)
+      REQUIRES(Locks::classlinker_classes_lock_)
+      SHARED_REQUIRES(Locks::mutator_lock_);
 
  private:
-  // Takes ownership of code_mem_map.
+  // Take ownership of code_mem_map.
   JitCodeCache(MemMap* code_map, MemMap* data_map);
 
-  // Lock which guards.
+  // Internal version of 'CommitCode' that will not retry if the
+  // allocation fails. Return null if the allocation fails.
+  uint8_t* CommitCodeInternal(Thread* self,
+                              ArtMethod* method,
+                              const uint8_t* mapping_table,
+                              const uint8_t* vmap_table,
+                              const uint8_t* gc_map,
+                              size_t frame_size_in_bytes,
+                              size_t core_spill_mask,
+                              size_t fp_spill_mask,
+                              const uint8_t* code,
+                              size_t code_size)
+      REQUIRES(!lock_)
+      SHARED_REQUIRES(Locks::mutator_lock_);
+
+  // If a collection is in progress, wait for it to finish. Return
+  // whether the thread actually waited.
+  bool WaitForPotentialCollectionToComplete(Thread* self)
+      REQUIRES(lock_) REQUIRES(!Locks::mutator_lock_);
+
+  // Free in the mspace allocations taken by 'method'.
+  void FreeCode(const void* code_ptr, ArtMethod* method) REQUIRES(lock_);
+
+  // Lock for guarding allocations, collections, and the method_code_map_.
   Mutex lock_;
+  // Condition to wait on during collection.
+  ConditionVariable lock_cond_ GUARDED_BY(lock_);
+  // Whether there is a code cache collection in progress.
+  bool collection_in_progress_ GUARDED_BY(lock_);
   // Mem map which holds code.
   std::unique_ptr<MemMap> code_map_;
   // Mem map which holds data (stack maps and profiling info).
   std::unique_ptr<MemMap> data_map_;
   // The opaque mspace for allocating code.
-  void* code_mspace_;
+  void* code_mspace_ GUARDED_BY(lock_);
   // The opaque mspace for allocating data.
-  void* data_mspace_;
-  // Number of compiled methods.
-  size_t num_methods_;
-  // This map holds code for methods if they were deoptimized by the instrumentation stubs. This is
-  // required since we have to implement ClassLinker::GetQuickOatCodeFor for walking stacks.
-  SafeMap<ArtMethod*, const void*> method_code_map_ GUARDED_BY(lock_);
+  void* data_mspace_ GUARDED_BY(lock_);
+  // Bitmap for collecting code and data.
+  std::unique_ptr<CodeCacheBitmap> live_bitmap_;
+  // This map holds compiled code associated to the ArtMethod
+  SafeMap<const void*, ArtMethod*> method_code_map_ GUARDED_BY(lock_);
 
   DISALLOW_IMPLICIT_CONSTRUCTORS(JitCodeCache);
 };
diff --git a/runtime/jit/jit_instrumentation.cc b/runtime/jit/jit_instrumentation.cc
index 9b9c5d2..666b8e7 100644
--- a/runtime/jit/jit_instrumentation.cc
+++ b/runtime/jit/jit_instrumentation.cc
@@ -76,8 +76,7 @@
   ScopedObjectAccessUnchecked soa(self);
   // Since we don't have on-stack replacement, some methods can remain in the interpreter longer
   // than we want resulting in samples even after the method is compiled.
-  if (method->IsClassInitializer() || method->IsNative() ||
-      Runtime::Current()->GetJit()->GetCodeCache()->ContainsMethod(method)) {
+  if (method->IsClassInitializer() || method->IsNative()) {
     return;
   }
   if (thread_pool_.get() == nullptr) {
diff --git a/runtime/jit/profiling_info.cc b/runtime/jit/profiling_info.cc
index 0c039f2..7c5f78e 100644
--- a/runtime/jit/profiling_info.cc
+++ b/runtime/jit/profiling_info.cc
@@ -28,15 +28,10 @@
 ProfilingInfo* ProfilingInfo::Create(ArtMethod* method) {
   // Walk over the dex instructions of the method and keep track of
   // instructions we are interested in profiling.
-  const uint16_t* code_ptr = nullptr;
-  const uint16_t* code_end = nullptr;
-  {
-    ScopedObjectAccess soa(Thread::Current());
-    DCHECK(!method->IsNative());
-    const DexFile::CodeItem& code_item = *method->GetCodeItem();
-    code_ptr = code_item.insns_;
-    code_end = code_item.insns_ + code_item.insns_size_in_code_units_;
-  }
+  DCHECK(!method->IsNative());
+  const DexFile::CodeItem& code_item = *method->GetCodeItem();
+  const uint16_t* code_ptr = code_item.insns_;
+  const uint16_t* code_end = code_item.insns_ + code_item.insns_size_in_code_units_;
 
   uint32_t dex_pc = 0;
   std::vector<uint32_t> entries;
@@ -91,7 +86,7 @@
 
   ScopedObjectAccess soa(self);
   for (size_t i = 0; i < InlineCache::kIndividualCacheSize; ++i) {
-    mirror::Class* existing = cache->classes_[i].Read<kWithoutReadBarrier>();
+    mirror::Class* existing = cache->classes_[i].Read();
     if (existing == cls) {
       // Receiver type is already in the cache, nothing else to do.
       return;
diff --git a/runtime/jit/profiling_info.h b/runtime/jit/profiling_info.h
index 73ca41a..7a2d1a8 100644
--- a/runtime/jit/profiling_info.h
+++ b/runtime/jit/profiling_info.h
@@ -36,7 +36,7 @@
  */
 class ProfilingInfo {
  public:
-  static ProfilingInfo* Create(ArtMethod* method);
+  static ProfilingInfo* Create(ArtMethod* method) SHARED_REQUIRES(Locks::mutator_lock_);
 
   // Add information from an executed INVOKE instruction to the profile.
   void AddInvokeInfo(Thread* self, uint32_t dex_pc, mirror::Class* cls);
diff --git a/runtime/length_prefixed_array.h b/runtime/length_prefixed_array.h
index 0ff6d7a..e01b6cc 100644
--- a/runtime/length_prefixed_array.h
+++ b/runtime/length_prefixed_array.h
@@ -30,19 +30,34 @@
 class LengthPrefixedArray {
  public:
   explicit LengthPrefixedArray(size_t length)
-      : length_(dchecked_integral_cast<uint32_t>(length)) {}
+      : size_(dchecked_integral_cast<uint32_t>(length)) {}
 
   T& At(size_t index, size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
-    DCHECK_LT(index, length_);
+    DCHECK_LT(index, size_);
     return AtUnchecked(index, element_size, alignment);
   }
 
-  StrideIterator<T> Begin(size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
+  const T& At(size_t index, size_t element_size = sizeof(T), size_t alignment = alignof(T)) const {
+    DCHECK_LT(index, size_);
+    return AtUnchecked(index, element_size, alignment);
+  }
+
+  StrideIterator<T> begin(size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
     return StrideIterator<T>(&AtUnchecked(0, element_size, alignment), element_size);
   }
 
-  StrideIterator<T> End(size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
-    return StrideIterator<T>(&AtUnchecked(length_, element_size, alignment), element_size);
+  StrideIterator<const T> begin(size_t element_size = sizeof(T),
+                                size_t alignment = alignof(T)) const {
+    return StrideIterator<const T>(&AtUnchecked(0, element_size, alignment), element_size);
+  }
+
+  StrideIterator<T> end(size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
+    return StrideIterator<T>(&AtUnchecked(size_, element_size, alignment), element_size);
+  }
+
+  StrideIterator<const T> end(size_t element_size = sizeof(T),
+                              size_t alignment = alignof(T)) const {
+    return StrideIterator<const T>(&AtUnchecked(size_, element_size, alignment), element_size);
   }
 
   static size_t OffsetOfElement(size_t index,
@@ -60,13 +75,13 @@
     return result;
   }
 
-  uint64_t Length() const {
-    return length_;
+  size_t size() const {
+    return size_;
   }
 
   // Update the length but does not reallocate storage.
-  void SetLength(size_t length) {
-    length_ = dchecked_integral_cast<uint32_t>(length);
+  void SetSize(size_t length) {
+    size_ = dchecked_integral_cast<uint32_t>(length);
   }
 
  private:
@@ -75,7 +90,12 @@
         reinterpret_cast<uintptr_t>(this) + OffsetOfElement(index, element_size, alignment));
   }
 
-  uint32_t length_;
+  const T& AtUnchecked(size_t index, size_t element_size, size_t alignment) const {
+    return *reinterpret_cast<T*>(
+        reinterpret_cast<uintptr_t>(this) + OffsetOfElement(index, element_size, alignment));
+  }
+
+  uint32_t size_;
   uint8_t data[0];
 };
 
@@ -84,7 +104,7 @@
 IterationRange<StrideIterator<T>> MakeIterationRangeFromLengthPrefixedArray(
     LengthPrefixedArray<T>* arr, size_t element_size = sizeof(T), size_t alignment = alignof(T)) {
   return arr != nullptr ?
-      MakeIterationRange(arr->Begin(element_size, alignment), arr->End(element_size, alignment)) :
+      MakeIterationRange(arr->begin(element_size, alignment), arr->end(element_size, alignment)) :
       MakeEmptyIterationRange(StrideIterator<T>(nullptr, 0));
 }
 
diff --git a/runtime/linear_alloc.cc b/runtime/linear_alloc.cc
index 43e81d9..f91b0ed 100644
--- a/runtime/linear_alloc.cc
+++ b/runtime/linear_alloc.cc
@@ -48,4 +48,8 @@
   return allocator_.Contains(ptr);
 }
 
+bool LinearAlloc::ContainsUnsafe(void* ptr) const {
+  return allocator_.Contains(ptr);
+}
+
 }  // namespace art
diff --git a/runtime/linear_alloc.h b/runtime/linear_alloc.h
index 1b21527..df7f17d 100644
--- a/runtime/linear_alloc.h
+++ b/runtime/linear_alloc.h
@@ -47,6 +47,10 @@
   // Return true if the linear alloc contrains an address.
   bool Contains(void* ptr) const REQUIRES(!lock_);
 
+  // Unsafe version of 'Contains' only to be used when the allocator is going
+  // to be deleted.
+  bool ContainsUnsafe(void* ptr) const NO_THREAD_SAFETY_ANALYSIS;
+
  private:
   mutable Mutex lock_ DEFAULT_MUTEX_ACQUIRED_AFTER;
   ArenaAllocator allocator_ GUARDED_BY(lock_);
diff --git a/runtime/mirror/class-inl.h b/runtime/mirror/class-inl.h
index a528c3b..19ee7f4 100644
--- a/runtime/mirror/class-inl.h
+++ b/runtime/mirror/class-inl.h
@@ -928,22 +928,22 @@
 
 inline uint32_t Class::NumDirectMethods() {
   LengthPrefixedArray<ArtMethod>* arr = GetDirectMethodsPtrUnchecked();
-  return arr != nullptr ? arr->Length() : 0u;
+  return arr != nullptr ? arr->size() : 0u;
 }
 
 inline uint32_t Class::NumVirtualMethods() {
   LengthPrefixedArray<ArtMethod>* arr = GetVirtualMethodsPtrUnchecked();
-  return arr != nullptr ? arr->Length() : 0u;
+  return arr != nullptr ? arr->size() : 0u;
 }
 
 inline uint32_t Class::NumInstanceFields() {
   LengthPrefixedArray<ArtField>* arr = GetIFieldsPtrUnchecked();
-  return arr != nullptr ? arr->Length() : 0u;
+  return arr != nullptr ? arr->size() : 0u;
 }
 
 inline uint32_t Class::NumStaticFields() {
   LengthPrefixedArray<ArtField>* arr = GetSFieldsPtrUnchecked();
-  return arr != nullptr ? arr->Length() : 0u;
+  return arr != nullptr ? arr->size() : 0u;
 }
 
 }  // namespace mirror
diff --git a/runtime/mirror/class.cc b/runtime/mirror/class.cc
index 53fedab..9d01a1d 100644
--- a/runtime/mirror/class.cc
+++ b/runtime/mirror/class.cc
@@ -574,7 +574,7 @@
     return nullptr;
   }
   size_t low = 0;
-  size_t high = fields->Length();
+  size_t high = fields->size();
   ArtField* ret = nullptr;
   while (low < high) {
     size_t mid = (low + high) / 2;
diff --git a/runtime/native/java_lang_Class.cc b/runtime/native/java_lang_Class.cc
index 3a73900..5e42392 100644
--- a/runtime/native/java_lang_Class.cc
+++ b/runtime/native/java_lang_Class.cc
@@ -190,7 +190,7 @@
     return nullptr;
   }
   size_t low = 0;
-  size_t high = fields->Length();
+  size_t high = fields->size();
   const uint16_t* const data = name->GetValue();
   const size_t length = name->GetLength();
   while (low < high) {
diff --git a/runtime/oat_quick_method_header.h b/runtime/oat_quick_method_header.h
index 6eadd87..c9a2cfb 100644
--- a/runtime/oat_quick_method_header.h
+++ b/runtime/oat_quick_method_header.h
@@ -21,6 +21,7 @@
 #include "base/macros.h"
 #include "quick/quick_method_frame_info.h"
 #include "stack_map.h"
+#include "utils.h"
 
 namespace art {
 
@@ -39,6 +40,16 @@
 
   ~OatQuickMethodHeader();
 
+  static OatQuickMethodHeader* FromCodePointer(const void* code_ptr) {
+    uintptr_t code = reinterpret_cast<uintptr_t>(code_ptr);
+    uintptr_t header = code - OFFSETOF_MEMBER(OatQuickMethodHeader, code_);
+    return reinterpret_cast<OatQuickMethodHeader*>(header);
+  }
+
+  static OatQuickMethodHeader* FromEntryPoint(const void* entry_point) {
+    return FromCodePointer(EntryPointToCodePointer(entry_point));
+  }
+
   OatQuickMethodHeader& operator=(const OatQuickMethodHeader&) = default;
 
   uintptr_t NativeQuickPcOffset(const uintptr_t pc) const {
@@ -74,6 +85,11 @@
 
   bool Contains(uintptr_t pc) const {
     uintptr_t code_start = reinterpret_cast<uintptr_t>(code_);
+    static_assert(kRuntimeISA != kThumb2, "kThumb2 cannot be a runtime ISA");
+    if (kRuntimeISA == kArm) {
+      // On Thumb-2, the pc is offset by one.
+      code_start++;
+    }
     return code_start <= pc && pc <= (code_start + code_size_);
   }
 
diff --git a/runtime/proxy_test.cc b/runtime/proxy_test.cc
index bc9ba37..57472ad 100644
--- a/runtime/proxy_test.cc
+++ b/runtime/proxy_test.cc
@@ -216,10 +216,10 @@
 
   LengthPrefixedArray<ArtField>* static_fields0 = proxyClass0->GetSFieldsPtr();
   ASSERT_TRUE(static_fields0 != nullptr);
-  ASSERT_EQ(2u, static_fields0->Length());
+  ASSERT_EQ(2u, static_fields0->size());
   LengthPrefixedArray<ArtField>* static_fields1 = proxyClass1->GetSFieldsPtr();
   ASSERT_TRUE(static_fields1 != nullptr);
-  ASSERT_EQ(2u, static_fields1->Length());
+  ASSERT_EQ(2u, static_fields1->size());
 
   EXPECT_EQ(static_fields0->At(0).GetDeclaringClass(), proxyClass0.Get());
   EXPECT_EQ(static_fields0->At(1).GetDeclaringClass(), proxyClass0.Get());
diff --git a/runtime/quick/inline_method_analyser.cc b/runtime/quick/inline_method_analyser.cc
index 99e262e..6554394 100644
--- a/runtime/quick/inline_method_analyser.cc
+++ b/runtime/quick/inline_method_analyser.cc
@@ -73,7 +73,6 @@
 bool InlineMethodAnalyser::AnalyseMethodCode(verifier::MethodVerifier* verifier,
                                              InlineMethod* method) {
   DCHECK(verifier != nullptr);
-  DCHECK_EQ(Runtime::Current()->IsCompiler(), method != nullptr);
   if (!Runtime::Current()->UseJit()) {
     DCHECK_EQ(verifier->CanLoadClasses(), method != nullptr);
   }
diff --git a/runtime/stack.cc b/runtime/stack.cc
index 9359d27..b0727da 100644
--- a/runtime/stack.cc
+++ b/runtime/stack.cc
@@ -856,13 +856,11 @@
   // If we are the JIT then we may have just compiled the method after the
   // IsQuickToInterpreterBridge check.
   jit::Jit* const jit = Runtime::Current()->GetJit();
-  if (jit != nullptr &&
-      jit->GetCodeCache()->ContainsCodePtr(reinterpret_cast<const void*>(code))) {
+  if (jit != nullptr && jit->GetCodeCache()->ContainsPc(code)) {
     return;
   }
 
-  uint32_t code_size = reinterpret_cast<const OatQuickMethodHeader*>(
-      EntryPointToCodePointer(code))[-1].code_size_;
+  uint32_t code_size = OatQuickMethodHeader::FromEntryPoint(code)->code_size_;
   uintptr_t code_start = reinterpret_cast<uintptr_t>(code);
   CHECK(code_start <= pc && pc <= (code_start + code_size))
       << PrettyMethod(method)
diff --git a/runtime/stride_iterator.h b/runtime/stride_iterator.h
index a9da51b..ac04c3b 100644
--- a/runtime/stride_iterator.h
+++ b/runtime/stride_iterator.h
@@ -19,6 +19,8 @@
 
 #include <iterator>
 
+#include "base/logging.h"
+
 namespace art {
 
 template<typename T>
diff --git a/runtime/verifier/method_verifier-inl.h b/runtime/verifier/method_verifier-inl.h
index 2d9fd53..f52d011 100644
--- a/runtime/verifier/method_verifier-inl.h
+++ b/runtime/verifier/method_verifier-inl.h
@@ -38,6 +38,10 @@
   return insn_flags_[index];
 }
 
+inline InstructionFlags& MethodVerifier::GetInstructionFlags(size_t index) {
+  return insn_flags_[index];
+}
+
 inline mirror::ClassLoader* MethodVerifier::GetClassLoader() {
   return class_loader_.Get();
 }
diff --git a/runtime/verifier/method_verifier.cc b/runtime/verifier/method_verifier.cc
index 4051a1c..e1d4160 100644
--- a/runtime/verifier/method_verifier.cc
+++ b/runtime/verifier/method_verifier.cc
@@ -58,12 +58,14 @@
 // On VLOG(verifier), should we dump the whole state when we run into a hard failure?
 static constexpr bool kDumpRegLinesOnHardFailureIfVLOG = true;
 
+PcToRegisterLineTable::PcToRegisterLineTable(ScopedArenaAllocator& arena)
+    : register_lines_(arena.Adapter(kArenaAllocVerifier)) {}
+
 void PcToRegisterLineTable::Init(RegisterTrackingMode mode, InstructionFlags* flags,
                                  uint32_t insns_size, uint16_t registers_size,
                                  MethodVerifier* verifier) {
   DCHECK_GT(insns_size, 0U);
-  register_lines_.reset(new RegisterLine*[insns_size]());
-  size_ = insns_size;
+  register_lines_.resize(insns_size);
   for (uint32_t i = 0; i < insns_size; i++) {
     bool interesting = false;
     switch (mode) {
@@ -80,19 +82,12 @@
         break;
     }
     if (interesting) {
-      register_lines_[i] = RegisterLine::Create(registers_size, verifier);
+      register_lines_[i].reset(RegisterLine::Create(registers_size, verifier));
     }
   }
 }
 
-PcToRegisterLineTable::~PcToRegisterLineTable() {
-  for (size_t i = 0; i < size_; i++) {
-    delete register_lines_[i];
-    if (kIsDebugBuild) {
-      register_lines_[i] = nullptr;
-    }
-  }
-}
+PcToRegisterLineTable::~PcToRegisterLineTable() {}
 
 // Note: returns true on failure.
 ALWAYS_INLINE static inline bool FailOrAbort(MethodVerifier* verifier, bool condition,
@@ -398,7 +393,10 @@
                                bool need_precise_constants, bool verify_to_dump,
                                bool allow_thread_suspension)
     : self_(self),
-      reg_types_(can_load_classes),
+      arena_stack_(Runtime::Current()->GetArenaPool()),
+      arena_(&arena_stack_),
+      reg_types_(can_load_classes, arena_),
+      reg_table_(arena_),
       work_insn_idx_(DexFile::kDexNoIndex),
       dex_method_idx_(dex_method_idx),
       mirror_method_(method),
@@ -702,7 +700,11 @@
   }
 
   // Allocate and initialize an array to hold instruction data.
-  insn_flags_.reset(new InstructionFlags[code_item_->insns_size_in_code_units_]());
+  insn_flags_.reset(arena_.AllocArray<InstructionFlags>(code_item_->insns_size_in_code_units_));
+  DCHECK(insn_flags_ != nullptr);
+  std::uninitialized_fill_n(insn_flags_.get(),
+                            code_item_->insns_size_in_code_units_,
+                            InstructionFlags());
   // Run through the instructions and see if the width checks out.
   bool result = ComputeWidthsAndCountOps();
   // Flag instructions guarded by a "try" block and check exception handlers.
@@ -848,7 +850,7 @@
         break;
     }
     size_t inst_size = inst->SizeInCodeUnits();
-    insn_flags_[dex_pc].SetIsOpcode();
+    GetInstructionFlags(dex_pc).SetIsOpcode();
     dex_pc += inst_size;
     inst = inst->RelativeAt(inst_size);
   }
@@ -881,7 +883,7 @@
                                         << " endAddr=" << end << " (size=" << insns_size << ")";
       return false;
     }
-    if (!insn_flags_[start].IsOpcode()) {
+    if (!GetInstructionFlags(start).IsOpcode()) {
       Fail(VERIFY_ERROR_BAD_CLASS_HARD)
           << "'try' block starts inside an instruction (" << start << ")";
       return false;
@@ -889,7 +891,7 @@
     uint32_t dex_pc = start;
     const Instruction* inst = Instruction::At(code_item_->insns_ + dex_pc);
     while (dex_pc < end) {
-      insn_flags_[dex_pc].SetInTry();
+      GetInstructionFlags(dex_pc).SetInTry();
       size_t insn_size = inst->SizeInCodeUnits();
       dex_pc += insn_size;
       inst = inst->RelativeAt(insn_size);
@@ -903,7 +905,7 @@
     CatchHandlerIterator iterator(handlers_ptr);
     for (; iterator.HasNext(); iterator.Next()) {
       uint32_t dex_pc= iterator.GetHandlerAddress();
-      if (!insn_flags_[dex_pc].IsOpcode()) {
+      if (!GetInstructionFlags(dex_pc).IsOpcode()) {
         Fail(VERIFY_ERROR_BAD_CLASS_HARD)
             << "exception handler starts at bad address (" << dex_pc << ")";
         return false;
@@ -913,7 +915,7 @@
             << "exception handler begins with move-result* (" << dex_pc << ")";
         return false;
       }
-      insn_flags_[dex_pc].SetBranchTarget();
+      GetInstructionFlags(dex_pc).SetBranchTarget();
       // Ensure exception types are resolved so that they don't need resolution to be delivered,
       // unresolved exception types will be ignored by exception delivery
       if (iterator.GetHandlerTypeIndex() != DexFile::kDexNoIndex16) {
@@ -935,8 +937,8 @@
   const Instruction* inst = Instruction::At(code_item_->insns_);
 
   /* Flag the start of the method as a branch target, and a GC point due to stack overflow errors */
-  insn_flags_[0].SetBranchTarget();
-  insn_flags_[0].SetCompileTimeInfoPoint();
+  GetInstructionFlags(0).SetBranchTarget();
+  GetInstructionFlags(0).SetCompileTimeInfoPoint();
 
   uint32_t insns_size = code_item_->insns_size_in_code_units_;
   for (uint32_t dex_pc = 0; dex_pc < insns_size;) {
@@ -948,18 +950,18 @@
     // All invoke points are marked as "Throw" points already.
     // We are relying on this to also count all the invokes as interesting.
     if (inst->IsBranch()) {
-      insn_flags_[dex_pc].SetCompileTimeInfoPoint();
+      GetInstructionFlags(dex_pc).SetCompileTimeInfoPoint();
       // The compiler also needs safepoints for fall-through to loop heads.
       // Such a loop head must be a target of a branch.
       int32_t offset = 0;
       bool cond, self_ok;
       bool target_ok = GetBranchOffset(dex_pc, &offset, &cond, &self_ok);
       DCHECK(target_ok);
-      insn_flags_[dex_pc + offset].SetCompileTimeInfoPoint();
+      GetInstructionFlags(dex_pc + offset).SetCompileTimeInfoPoint();
     } else if (inst->IsSwitch() || inst->IsThrow()) {
-      insn_flags_[dex_pc].SetCompileTimeInfoPoint();
+      GetInstructionFlags(dex_pc).SetCompileTimeInfoPoint();
     } else if (inst->IsReturn()) {
-      insn_flags_[dex_pc].SetCompileTimeInfoPointAndReturn();
+      GetInstructionFlags(dex_pc).SetCompileTimeInfoPointAndReturn();
     }
     dex_pc += inst->SizeInCodeUnits();
     inst = inst->Next();
@@ -1202,7 +1204,7 @@
   }
   // Make sure the array-data is marked as an opcode. This ensures that it was reached when
   // traversing the code item linearly. It is an approximation for a by-spec padding value.
-  if (!insn_flags_[cur_offset + array_data_offset].IsOpcode()) {
+  if (!GetInstructionFlags(cur_offset + array_data_offset).IsOpcode()) {
     Fail(VERIFY_ERROR_BAD_CLASS_HARD) << "array data table at " << cur_offset
                                       << ", data offset " << array_data_offset
                                       << " not correctly visited, probably bad padding.";
@@ -1245,13 +1247,13 @@
   int32_t abs_offset = cur_offset + offset;
   if (abs_offset < 0 ||
       (uint32_t) abs_offset >= insn_count ||
-      !insn_flags_[abs_offset].IsOpcode()) {
+      !GetInstructionFlags(abs_offset).IsOpcode()) {
     Fail(VERIFY_ERROR_BAD_CLASS_HARD) << "invalid branch target " << offset << " (-> "
                                       << reinterpret_cast<void*>(abs_offset) << ") at "
                                       << reinterpret_cast<void*>(cur_offset);
     return false;
   }
-  insn_flags_[abs_offset].SetBranchTarget();
+  GetInstructionFlags(abs_offset).SetBranchTarget();
   return true;
 }
 
@@ -1315,7 +1317,7 @@
   }
   // Make sure the switch data is marked as an opcode. This ensures that it was reached when
   // traversing the code item linearly. It is an approximation for a by-spec padding value.
-  if (!insn_flags_[cur_offset + switch_offset].IsOpcode()) {
+  if (!GetInstructionFlags(cur_offset + switch_offset).IsOpcode()) {
     Fail(VERIFY_ERROR_BAD_CLASS_HARD) << "switch table at " << cur_offset
                                       << ", switch offset " << switch_offset
                                       << " not correctly visited, probably bad padding.";
@@ -1387,14 +1389,14 @@
     int32_t abs_offset = cur_offset + offset;
     if (abs_offset < 0 ||
         abs_offset >= static_cast<int32_t>(insn_count) ||
-        !insn_flags_[abs_offset].IsOpcode()) {
+        !GetInstructionFlags(abs_offset).IsOpcode()) {
       Fail(VERIFY_ERROR_BAD_CLASS_HARD) << "invalid switch target " << offset
                                         << " (-> " << reinterpret_cast<void*>(abs_offset) << ") at "
                                         << reinterpret_cast<void*>(cur_offset)
                                         << "[" << targ << "]";
       return false;
     }
-    insn_flags_[abs_offset].SetBranchTarget();
+    GetInstructionFlags(abs_offset).SetBranchTarget();
   }
   return true;
 }
@@ -1435,7 +1437,6 @@
                   registers_size,
                   this);
 
-
   work_line_.reset(RegisterLine::Create(registers_size, this));
   saved_line_.reset(RegisterLine::Create(registers_size, this));
 
@@ -1491,7 +1492,7 @@
       vios->Stream() << reg_line->Dump(this) << "\n";
     }
     vios->Stream()
-        << StringPrintf("0x%04zx", dex_pc) << ": " << insn_flags_[dex_pc].ToString() << " ";
+        << StringPrintf("0x%04zx", dex_pc) << ": " << GetInstructionFlags(dex_pc).ToString() << " ";
     const bool kDumpHexOfInstruction = false;
     if (kDumpHexOfInstruction) {
       vios->Stream() << inst->DumpHex(5) << " ";
@@ -1677,7 +1678,7 @@
   const uint32_t insns_size = code_item_->insns_size_in_code_units_;
 
   /* Begin by marking the first instruction as "changed". */
-  insn_flags_[0].SetChanged();
+  GetInstructionFlags(0).SetChanged();
   uint32_t start_guess = 0;
 
   /* Continue until no instructions are marked "changed". */
@@ -1688,7 +1689,7 @@
     // Find the first marked one. Use "start_guess" as a way to find one quickly.
     uint32_t insn_idx = start_guess;
     for (; insn_idx < insns_size; insn_idx++) {
-      if (insn_flags_[insn_idx].IsChanged())
+      if (GetInstructionFlags(insn_idx).IsChanged())
         break;
     }
     if (insn_idx == insns_size) {
@@ -1708,7 +1709,7 @@
     // situation where we have a stray "changed" flag set on an instruction that isn't a branch
     // target.
     work_insn_idx_ = insn_idx;
-    if (insn_flags_[insn_idx].IsBranchTarget()) {
+    if (GetInstructionFlags(insn_idx).IsBranchTarget()) {
       work_line_->CopyFromLine(reg_table_.GetLine(insn_idx));
     } else if (kIsDebugBuild) {
       /*
@@ -1734,8 +1735,8 @@
       return false;
     }
     /* Clear "changed" and mark as visited. */
-    insn_flags_[insn_idx].SetVisited();
-    insn_flags_[insn_idx].ClearChanged();
+    GetInstructionFlags(insn_idx).SetVisited();
+    GetInstructionFlags(insn_idx).ClearChanged();
   }
 
   if (gDebugVerify) {
@@ -1762,10 +1763,10 @@
            (insns[insn_idx + 1] == Instruction::kPackedSwitchSignature ||
             insns[insn_idx + 1] == Instruction::kSparseSwitchSignature ||
             insns[insn_idx + 1] == Instruction::kArrayDataSignature))) {
-        insn_flags_[insn_idx].SetVisited();
+        GetInstructionFlags(insn_idx).SetVisited();
       }
 
-      if (!insn_flags_[insn_idx].IsVisited()) {
+      if (!GetInstructionFlags(insn_idx).IsVisited()) {
         if (dead_start < 0)
           dead_start = insn_idx;
       } else if (dead_start >= 0) {
@@ -1895,8 +1896,8 @@
   // We need to ensure the work line is consistent while performing validation. When we spot a
   // peephole pattern we compute a new line for either the fallthrough instruction or the
   // branch target.
-  std::unique_ptr<RegisterLine> branch_line;
-  std::unique_ptr<RegisterLine> fallthrough_line;
+  ArenaUniquePtr<RegisterLine> branch_line;
+  ArenaUniquePtr<RegisterLine> fallthrough_line;
 
   switch (inst->Opcode()) {
     case Instruction::NOP:
@@ -2144,9 +2145,9 @@
       work_line_->PushMonitor(this, inst->VRegA_11x(), work_insn_idx_);
       // Check whether the previous instruction is a move-object with vAA as a source, creating
       // untracked lock aliasing.
-      if (0 != work_insn_idx_ && !insn_flags_[work_insn_idx_].IsBranchTarget()) {
+      if (0 != work_insn_idx_ && !GetInstructionFlags(work_insn_idx_).IsBranchTarget()) {
         uint32_t prev_idx = work_insn_idx_ - 1;
-        while (0 != prev_idx && !insn_flags_[prev_idx].IsOpcode()) {
+        while (0 != prev_idx && !GetInstructionFlags(prev_idx).IsOpcode()) {
           prev_idx--;
         }
         const Instruction* prev_inst = Instruction::At(code_item_->insns_ + prev_idx);
@@ -2427,10 +2428,10 @@
       uint32_t instance_of_idx = 0;
       if (0 != work_insn_idx_) {
         instance_of_idx = work_insn_idx_ - 1;
-        while (0 != instance_of_idx && !insn_flags_[instance_of_idx].IsOpcode()) {
+        while (0 != instance_of_idx && !GetInstructionFlags(instance_of_idx).IsOpcode()) {
           instance_of_idx--;
         }
-        if (FailOrAbort(this, insn_flags_[instance_of_idx].IsOpcode(),
+        if (FailOrAbort(this, GetInstructionFlags(instance_of_idx).IsOpcode(),
                         "Unable to get previous instruction of if-eqz/if-nez for work index ",
                         work_insn_idx_)) {
           break;
@@ -2486,15 +2487,15 @@
           update_line->SetRegisterType<LockOp::kKeep>(this,
                                                       instance_of_inst->VRegB_22c(),
                                                       cast_type);
-          if (!insn_flags_[instance_of_idx].IsBranchTarget() && 0 != instance_of_idx) {
+          if (!GetInstructionFlags(instance_of_idx).IsBranchTarget() && 0 != instance_of_idx) {
             // See if instance-of was preceded by a move-object operation, common due to the small
             // register encoding space of instance-of, and propagate type information to the source
             // of the move-object.
             uint32_t move_idx = instance_of_idx - 1;
-            while (0 != move_idx && !insn_flags_[move_idx].IsOpcode()) {
+            while (0 != move_idx && !GetInstructionFlags(move_idx).IsOpcode()) {
               move_idx--;
             }
-            if (FailOrAbort(this, insn_flags_[move_idx].IsOpcode(),
+            if (FailOrAbort(this, GetInstructionFlags(move_idx).IsOpcode(),
                             "Unable to get previous instruction of if-eqz/if-nez for work index ",
                             work_insn_idx_)) {
               break;
@@ -2786,8 +2787,7 @@
         work_line_->MarkRefsAsInitialized(this, this_type, this_reg, work_insn_idx_);
       }
       if (return_type == nullptr) {
-        return_type = &reg_types_.FromDescriptor(GetClassLoader(), return_type_descriptor,
-                                                 false);
+        return_type = &reg_types_.FromDescriptor(GetClassLoader(), return_type_descriptor, false);
       }
       if (!return_type->IsLowHalf()) {
         work_line_->SetResultRegisterType(this, *return_type);
@@ -2860,7 +2860,7 @@
         uint32_t method_idx = (is_range) ? inst->VRegB_3rc() : inst->VRegB_35c();
         const DexFile::MethodId& method_id = dex_file_->GetMethodId(method_idx);
         uint32_t return_type_idx = dex_file_->GetProtoId(method_id.proto_idx_).return_type_idx_;
-        descriptor =  dex_file_->StringByTypeIdx(return_type_idx);
+        descriptor = dex_file_->StringByTypeIdx(return_type_idx);
       } else {
         descriptor = abs_method->GetReturnTypeDescriptor();
       }
@@ -3309,7 +3309,7 @@
       return false;
     }
     /* update branch target, set "changed" if appropriate */
-    if (nullptr != branch_line.get()) {
+    if (nullptr != branch_line) {
       if (!UpdateRegisters(work_insn_idx_ + branch_target, branch_line.get(), false)) {
         return false;
       }
@@ -3364,7 +3364,7 @@
    * Handle instructions that can throw and that are sitting in a "try" block. (If they're not in a
    * "try" block when they throw, control transfers out of the method.)
    */
-  if ((opcode_flags & Instruction::kThrow) != 0 && insn_flags_[work_insn_idx_].IsInTry()) {
+  if ((opcode_flags & Instruction::kThrow) != 0 && GetInstructionFlags(work_insn_idx_).IsInTry()) {
     bool has_catch_all_handler = false;
     CatchHandlerIterator iterator(*code_item_, work_insn_idx_);
 
@@ -3434,11 +3434,11 @@
     if (!CheckNotMoveException(code_item_->insns_, next_insn_idx)) {
       return false;
     }
-    if (nullptr != fallthrough_line.get()) {
+    if (nullptr != fallthrough_line) {
       // Make workline consistent with fallthrough computed from peephole optimization.
       work_line_->CopyFromLine(fallthrough_line.get());
     }
-    if (insn_flags_[next_insn_idx].IsReturn()) {
+    if (GetInstructionFlags(next_insn_idx).IsReturn()) {
       // For returns we only care about the operand to the return, all other registers are dead.
       const Instruction* ret_inst = Instruction::At(code_item_->insns_ + next_insn_idx);
       AdjustReturnLine(this, ret_inst, work_line_.get());
@@ -3456,7 +3456,7 @@
        * We're not recording register data for the next instruction, so we don't know what the
        * prior state was. We have to assume that something has changed and re-evaluate it.
        */
-      insn_flags_[next_insn_idx].SetChanged();
+      GetInstructionFlags(next_insn_idx).SetChanged();
     }
   }
 
@@ -3480,7 +3480,7 @@
   }
 
   DCHECK_LT(*start_guess, code_item_->insns_size_in_code_units_);
-  DCHECK(insn_flags_[*start_guess].IsOpcode());
+  DCHECK(GetInstructionFlags(*start_guess).IsOpcode());
 
   if (have_pending_runtime_throw_failure_) {
     have_any_pending_runtime_throw_failure_ = true;
@@ -3491,30 +3491,55 @@
   return true;
 }  // NOLINT(readability/fn_size)
 
+void MethodVerifier::UninstantiableError(const char* descriptor) {
+  Fail(VerifyError::VERIFY_ERROR_NO_CLASS) << "Could not create precise reference for "
+                                           << "non-instantiable klass " << descriptor;
+}
+
+inline bool MethodVerifier::IsInstantiableOrPrimitive(mirror::Class* klass) {
+  return klass->IsInstantiable() || klass->IsPrimitive();
+}
+
 const RegType& MethodVerifier::ResolveClassAndCheckAccess(uint32_t class_idx) {
-  const char* descriptor = dex_file_->StringByTypeIdx(class_idx);
-  const RegType& referrer = GetDeclaringClass();
   mirror::Class* klass = dex_cache_->GetResolvedType(class_idx);
-  const RegType& result = klass != nullptr ?
-      FromClass(descriptor, klass, klass->CannotBeAssignedFromOtherTypes()) :
-      reg_types_.FromDescriptor(GetClassLoader(), descriptor, false);
-  if (result.IsConflict()) {
-    Fail(VERIFY_ERROR_BAD_CLASS_SOFT) << "accessing broken descriptor '" << descriptor
-        << "' in " << referrer;
-    return result;
+  const RegType* result = nullptr;
+  if (klass != nullptr) {
+    bool precise = klass->CannotBeAssignedFromOtherTypes();
+    if (precise && !IsInstantiableOrPrimitive(klass)) {
+      const char* descriptor = dex_file_->StringByTypeIdx(class_idx);
+      UninstantiableError(descriptor);
+      precise = false;
+    }
+    result = reg_types_.FindClass(klass, precise);
+    if (result == nullptr) {
+      const char* descriptor = dex_file_->StringByTypeIdx(class_idx);
+      result = reg_types_.InsertClass(descriptor, klass, precise);
+    }
+  } else {
+    const char* descriptor = dex_file_->StringByTypeIdx(class_idx);
+    result = &reg_types_.FromDescriptor(GetClassLoader(), descriptor, false);
   }
-  if (klass == nullptr && !result.IsUnresolvedTypes()) {
-    dex_cache_->SetResolvedType(class_idx, result.GetClass());
+  DCHECK(result != nullptr);
+  if (result->IsConflict()) {
+    const char* descriptor = dex_file_->StringByTypeIdx(class_idx);
+    Fail(VERIFY_ERROR_BAD_CLASS_SOFT) << "accessing broken descriptor '" << descriptor
+        << "' in " << GetDeclaringClass();
+    return *result;
+  }
+  if (klass == nullptr && !result->IsUnresolvedTypes()) {
+    dex_cache_->SetResolvedType(class_idx, result->GetClass());
   }
   // Check if access is allowed. Unresolved types use xxxWithAccessCheck to
   // check at runtime if access is allowed and so pass here. If result is
   // primitive, skip the access check.
-  if (result.IsNonZeroReferenceTypes() && !result.IsUnresolvedTypes() &&
-      !referrer.IsUnresolvedTypes() && !referrer.CanAccess(result)) {
-    Fail(VERIFY_ERROR_ACCESS_CLASS) << "illegal class access: '"
-                                    << referrer << "' -> '" << result << "'";
+  if (result->IsNonZeroReferenceTypes() && !result->IsUnresolvedTypes()) {
+    const RegType& referrer = GetDeclaringClass();
+    if (!referrer.IsUnresolvedTypes() && !referrer.CanAccess(*result)) {
+      Fail(VERIFY_ERROR_ACCESS_CLASS) << "illegal class access: '"
+                                      << referrer << "' -> '" << result << "'";
+    }
   }
-  return result;
+  return *result;
 }
 
 const RegType& MethodVerifier::GetCaughtExceptionType() {
@@ -3720,9 +3745,10 @@
       } else {
         const uint32_t method_idx = (is_range) ? inst->VRegB_3rc() : inst->VRegB_35c();
         const uint16_t class_idx = dex_file_->GetMethodId(method_idx).class_idx_;
-        res_method_class = &reg_types_.FromDescriptor(GetClassLoader(),
-                                                      dex_file_->StringByTypeIdx(class_idx),
-                                                      false);
+        res_method_class = &reg_types_.FromDescriptor(
+            GetClassLoader(),
+            dex_file_->StringByTypeIdx(class_idx),
+            false);
       }
       if (!res_method_class->IsAssignableFrom(actual_arg_type)) {
         Fail(actual_arg_type.IsUnresolvedTypes() ? VERIFY_ERROR_NO_CLASS:
@@ -4476,14 +4502,16 @@
         field->GetType<false>();
 
     if (field_type_class != nullptr) {
-      field_type = &FromClass(field->GetTypeDescriptor(), field_type_class,
+      field_type = &FromClass(field->GetTypeDescriptor(),
+                              field_type_class,
                               field_type_class->CannotBeAssignedFromOtherTypes());
     } else {
       Thread* self = Thread::Current();
       DCHECK(!can_load_classes_ || self->IsExceptionPending());
       self->ClearException();
       field_type = &reg_types_.FromDescriptor(field->GetDeclaringClass()->GetClassLoader(),
-                                              field->GetTypeDescriptor(), false);
+                                              field->GetTypeDescriptor(),
+                                              false);
     }
     if (field_type == nullptr) {
       Fail(VERIFY_ERROR_BAD_CLASS_HARD) << "Cannot infer field type from " << inst->Name();
@@ -4604,14 +4632,14 @@
                                      bool update_merge_line) {
   bool changed = true;
   RegisterLine* target_line = reg_table_.GetLine(next_insn);
-  if (!insn_flags_[next_insn].IsVisitedOrChanged()) {
+  if (!GetInstructionFlags(next_insn).IsVisitedOrChanged()) {
     /*
      * We haven't processed this instruction before, and we haven't touched the registers here, so
      * there's nothing to "merge". Copy the registers over and mark it as changed. (This is the
      * only way a register can transition out of "unknown", so this is not just an optimization.)
      */
     target_line->CopyFromLine(merge_line);
-    if (insn_flags_[next_insn].IsReturn()) {
+    if (GetInstructionFlags(next_insn).IsReturn()) {
       // Verify that the monitor stack is empty on return.
       merge_line->VerifyMonitorStackEmpty(this);
 
@@ -4621,10 +4649,9 @@
       AdjustReturnLine(this, ret_inst, target_line);
     }
   } else {
-    std::unique_ptr<RegisterLine> copy(gDebugVerify ?
-                                 RegisterLine::Create(target_line->NumRegs(), this) :
-                                 nullptr);
+    ArenaUniquePtr<RegisterLine> copy;
     if (gDebugVerify) {
+      copy.reset(RegisterLine::Create(target_line->NumRegs(), this));
       copy->CopyFromLine(target_line);
     }
     changed = target_line->MergeRegisters(this, merge_line);
@@ -4643,13 +4670,13 @@
     }
   }
   if (changed) {
-    insn_flags_[next_insn].SetChanged();
+    GetInstructionFlags(next_insn).SetChanged();
   }
   return true;
 }
 
 InstructionFlags* MethodVerifier::CurrentInsnFlags() {
-  return &insn_flags_[work_insn_idx_];
+  return &GetInstructionFlags(work_insn_idx_);
 }
 
 const RegType& MethodVerifier::GetMethodReturnType() {
@@ -4685,8 +4712,7 @@
         = dex_file_->GetTypeDescriptor(dex_file_->GetTypeId(method_id.class_idx_));
     if (mirror_method_ != nullptr) {
       mirror::Class* klass = mirror_method_->GetDeclaringClass();
-      declaring_class_ = &FromClass(descriptor, klass,
-                                    klass->CannotBeAssignedFromOtherTypes());
+      declaring_class_ = &FromClass(descriptor, klass, klass->CannotBeAssignedFromOtherTypes());
     } else {
       declaring_class_ = &reg_types_.FromDescriptor(GetClassLoader(), descriptor, false);
     }
diff --git a/runtime/verifier/method_verifier.h b/runtime/verifier/method_verifier.h
index ba694b7..7b51d6e 100644
--- a/runtime/verifier/method_verifier.h
+++ b/runtime/verifier/method_verifier.h
@@ -21,7 +21,10 @@
 #include <sstream>
 #include <vector>
 
+#include "base/arena_allocator.h"
 #include "base/macros.h"
+#include "base/scoped_arena_containers.h"
+#include "base/stl_util.h"
 #include "dex_file.h"
 #include "handle.h"
 #include "instruction_flags.h"
@@ -107,7 +110,7 @@
 // execution of that instruction.
 class PcToRegisterLineTable {
  public:
-  PcToRegisterLineTable() : size_(0) {}
+  explicit PcToRegisterLineTable(ScopedArenaAllocator& arena);
   ~PcToRegisterLineTable();
 
   // Initialize the RegisterTable. Every instruction address can have a different set of information
@@ -116,14 +119,12 @@
   void Init(RegisterTrackingMode mode, InstructionFlags* flags, uint32_t insns_size,
             uint16_t registers_size, MethodVerifier* verifier);
 
-  RegisterLine* GetLine(size_t idx) {
-    DCHECK_LT(idx, size_);
-    return register_lines_[idx];
+  RegisterLine* GetLine(size_t idx) const {
+    return register_lines_[idx].get();
   }
 
  private:
-  std::unique_ptr<RegisterLine*[]> register_lines_;
-  size_t size_;
+  ScopedArenaVector<ArenaUniquePtr<RegisterLine>> register_lines_;
 
   DISALLOW_COPY_AND_ASSIGN(PcToRegisterLineTable);
 };
@@ -240,7 +241,8 @@
   // Accessors used by the compiler via CompilerCallback
   const DexFile::CodeItem* CodeItem() const;
   RegisterLine* GetRegLine(uint32_t dex_pc);
-  const InstructionFlags& GetInstructionFlags(size_t index) const;
+  ALWAYS_INLINE const InstructionFlags& GetInstructionFlags(size_t index) const;
+  ALWAYS_INLINE InstructionFlags& GetInstructionFlags(size_t index);
   mirror::ClassLoader* GetClassLoader() SHARED_REQUIRES(Locks::mutator_lock_);
   mirror::DexCache* GetDexCache() SHARED_REQUIRES(Locks::mutator_lock_);
   MethodReference GetMethodReference() const;
@@ -275,7 +277,14 @@
     return IsConstructor() && !IsStatic();
   }
 
+  ScopedArenaAllocator& GetArena() {
+    return arena_;
+  }
+
  private:
+  void UninstantiableError(const char* descriptor);
+  static bool IsInstantiableOrPrimitive(mirror::Class* klass) SHARED_REQUIRES(Locks::mutator_lock_);
+
   // Is the method being verified a constructor? See the comment on the field.
   bool IsConstructor() const {
     return is_constructor_;
@@ -687,19 +696,23 @@
   // The thread we're verifying on.
   Thread* const self_;
 
+  // Arena allocator.
+  ArenaStack arena_stack_;
+  ScopedArenaAllocator arena_;
+
   RegTypeCache reg_types_;
 
   PcToRegisterLineTable reg_table_;
 
   // Storage for the register status we're currently working on.
-  std::unique_ptr<RegisterLine> work_line_;
+  ArenaUniquePtr<RegisterLine> work_line_;
 
   // The address of the instruction we're currently working on, note that this is in 2 byte
   // quantities
   uint32_t work_insn_idx_;
 
   // Storage for the register status we're saving for later.
-  std::unique_ptr<RegisterLine> saved_line_;
+  ArenaUniquePtr<RegisterLine> saved_line_;
 
   const uint32_t dex_method_idx_;  // The method we're working on.
   // Its object representation if known.
@@ -715,7 +728,8 @@
   const DexFile::CodeItem* const code_item_;  // The code item containing the code for the method.
   const RegType* declaring_class_;  // Lazily computed reg type of the method's declaring class.
   // Instruction widths and flags, one entry per code unit.
-  std::unique_ptr<InstructionFlags[]> insn_flags_;
+  // Owned, but not unique_ptr since insn_flags_ are allocated in arenas.
+  ArenaUniquePtr<InstructionFlags[]> insn_flags_;
   // The dex PC of a FindLocksAtDexPc request, -1 otherwise.
   uint32_t interesting_dex_pc_;
   // The container into which FindLocksAtDexPc should write the registers containing held locks,
diff --git a/runtime/verifier/reg_type-inl.h b/runtime/verifier/reg_type-inl.h
index f445132..11a53e5 100644
--- a/runtime/verifier/reg_type-inl.h
+++ b/runtime/verifier/reg_type-inl.h
@@ -20,6 +20,7 @@
 #include "reg_type.h"
 
 #include "base/casts.h"
+#include "base/scoped_arena_allocator.h"
 #include "mirror/class.h"
 
 namespace art {
@@ -180,6 +181,10 @@
   return instance_;
 }
 
+inline void* RegType::operator new(size_t size, ScopedArenaAllocator* arena) {
+  return arena->Alloc(size, kArenaAllocMisc);
+}
+
 }  // namespace verifier
 }  // namespace art
 
diff --git a/runtime/verifier/reg_type.cc b/runtime/verifier/reg_type.cc
index b86a4c8..16cab03 100644
--- a/runtime/verifier/reg_type.cc
+++ b/runtime/verifier/reg_type.cc
@@ -16,6 +16,7 @@
 
 #include "reg_type-inl.h"
 
+#include "base/arena_bit_vector.h"
 #include "base/bit_vector-inl.h"
 #include "base/casts.h"
 #include "class_linker-inl.h"
@@ -46,20 +47,17 @@
 const DoubleHiType* DoubleHiType::instance_ = nullptr;
 const IntegerType* IntegerType::instance_ = nullptr;
 
-PrimitiveType::PrimitiveType(mirror::Class* klass, const std::string& descriptor, uint16_t cache_id)
-    SHARED_REQUIRES(Locks::mutator_lock_)
+PrimitiveType::PrimitiveType(mirror::Class* klass, const StringPiece& descriptor, uint16_t cache_id)
     : RegType(klass, descriptor, cache_id) {
   CHECK(klass != nullptr);
   CHECK(!descriptor.empty());
 }
 
-Cat1Type::Cat1Type(mirror::Class* klass, const std::string& descriptor, uint16_t cache_id)
-    SHARED_REQUIRES(Locks::mutator_lock_)
+Cat1Type::Cat1Type(mirror::Class* klass, const StringPiece& descriptor, uint16_t cache_id)
     : PrimitiveType(klass, descriptor, cache_id) {
 }
 
-Cat2Type::Cat2Type(mirror::Class* klass, const std::string& descriptor, uint16_t cache_id)
-    SHARED_REQUIRES(Locks::mutator_lock_)
+Cat2Type::Cat2Type(mirror::Class* klass, const StringPiece& descriptor, uint16_t cache_id)
     : PrimitiveType(klass, descriptor, cache_id) {
 }
 
@@ -121,11 +119,11 @@
 }
 
 std::string IntegerType::Dump() const {
-    return "Integer";
+  return "Integer";
 }
 
 const DoubleHiType* DoubleHiType::CreateInstance(mirror::Class* klass,
-                                                 const std::string& descriptor,
+                                                 const StringPiece& descriptor,
                                                  uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new DoubleHiType(klass, descriptor, cache_id);
@@ -140,7 +138,7 @@
 }
 
 const DoubleLoType* DoubleLoType::CreateInstance(mirror::Class* klass,
-                                                 const std::string& descriptor,
+                                                 const StringPiece& descriptor,
                                                  uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new DoubleLoType(klass, descriptor, cache_id);
@@ -154,14 +152,14 @@
   }
 }
 
-const LongLoType* LongLoType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const LongLoType* LongLoType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                              uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new LongLoType(klass, descriptor, cache_id);
   return instance_;
 }
 
-const LongHiType* LongHiType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const LongHiType* LongHiType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                              uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new LongHiType(klass, descriptor, cache_id);
@@ -182,7 +180,7 @@
   }
 }
 
-const FloatType* FloatType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const FloatType* FloatType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                            uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new FloatType(klass, descriptor, cache_id);
@@ -196,7 +194,7 @@
   }
 }
 
-const CharType* CharType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const CharType* CharType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                          uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new CharType(klass, descriptor, cache_id);
@@ -210,7 +208,7 @@
   }
 }
 
-const ShortType* ShortType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const ShortType* ShortType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                            uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new ShortType(klass, descriptor, cache_id);
@@ -224,7 +222,7 @@
   }
 }
 
-const ByteType* ByteType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const ByteType* ByteType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                          uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new ByteType(klass, descriptor, cache_id);
@@ -238,7 +236,7 @@
   }
 }
 
-const IntegerType* IntegerType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const IntegerType* IntegerType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                                uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new IntegerType(klass, descriptor, cache_id);
@@ -253,7 +251,7 @@
 }
 
 const ConflictType* ConflictType::CreateInstance(mirror::Class* klass,
-                                                 const std::string& descriptor,
+                                                 const StringPiece& descriptor,
                                                  uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new ConflictType(klass, descriptor, cache_id);
@@ -267,7 +265,7 @@
   }
 }
 
-const BooleanType* BooleanType::CreateInstance(mirror::Class* klass, const std::string& descriptor,
+const BooleanType* BooleanType::CreateInstance(mirror::Class* klass, const StringPiece& descriptor,
                                          uint16_t cache_id) {
   CHECK(BooleanType::instance_ == nullptr);
   instance_ = new BooleanType(klass, descriptor, cache_id);
@@ -286,7 +284,7 @@
 }
 
 const UndefinedType* UndefinedType::CreateInstance(mirror::Class* klass,
-                                                   const std::string& descriptor,
+                                                   const StringPiece& descriptor,
                                                    uint16_t cache_id) {
   CHECK(instance_ == nullptr);
   instance_ = new UndefinedType(klass, descriptor, cache_id);
@@ -300,7 +298,7 @@
   }
 }
 
-PreciseReferenceType::PreciseReferenceType(mirror::Class* klass, const std::string& descriptor,
+PreciseReferenceType::PreciseReferenceType(mirror::Class* klass, const StringPiece& descriptor,
                                            uint16_t cache_id)
     : RegType(klass, descriptor, cache_id) {
   // Note: no check for IsInstantiable() here. We may produce this in case an InstantiationError
@@ -335,14 +333,14 @@
 
 std::string UnresolvedReferenceType::Dump() const {
   std::stringstream result;
-  result << "Unresolved Reference" << ": " << PrettyDescriptor(GetDescriptor().c_str());
+  result << "Unresolved Reference" << ": " << PrettyDescriptor(GetDescriptor().as_string().c_str());
   return result.str();
 }
 
 std::string UnresolvedUninitializedRefType::Dump() const {
   std::stringstream result;
   result << "Unresolved And Uninitialized Reference" << ": "
-      << PrettyDescriptor(GetDescriptor().c_str())
+      << PrettyDescriptor(GetDescriptor().as_string().c_str())
       << " Allocation PC: " << GetAllocationPc();
   return result.str();
 }
@@ -350,7 +348,7 @@
 std::string UnresolvedUninitializedThisRefType::Dump() const {
   std::stringstream result;
   result << "Unresolved And Uninitialized This Reference"
-      << PrettyDescriptor(GetDescriptor().c_str());
+      << PrettyDescriptor(GetDescriptor().as_string().c_str());
   return result.str();
 }
 
diff --git a/runtime/verifier/reg_type.h b/runtime/verifier/reg_type.h
index 2834a9a..80b751c 100644
--- a/runtime/verifier/reg_type.h
+++ b/runtime/verifier/reg_type.h
@@ -22,9 +22,11 @@
 #include <set>
 #include <string>
 
+#include "base/arena_object.h"
 #include "base/bit_vector.h"
 #include "base/macros.h"
 #include "base/mutex.h"
+#include "base/stringpiece.h"
 #include "gc_root.h"
 #include "handle_scope.h"
 #include "object_callbacks.h"
@@ -35,6 +37,9 @@
 class Class;
 }  // namespace mirror
 
+class ArenaBitVector;
+class ScopedArenaAllocator;
+
 namespace verifier {
 
 class RegTypeCache;
@@ -173,7 +178,7 @@
   bool IsJavaLangObjectArray() const
       SHARED_REQUIRES(Locks::mutator_lock_);
   bool IsInstantiableTypes() const SHARED_REQUIRES(Locks::mutator_lock_);
-  const std::string& GetDescriptor() const {
+  const StringPiece& GetDescriptor() const {
     DCHECK(HasClass() ||
            (IsUnresolvedTypes() && !IsUnresolvedMergedReference() &&
             !IsUnresolvedSuperClass()));
@@ -274,10 +279,20 @@
   void VisitRoots(RootVisitor* visitor, const RootInfo& root_info) const
       SHARED_REQUIRES(Locks::mutator_lock_);
 
+  static void* operator new(size_t size) noexcept {
+    return ::operator new(size);
+  }
+
+  static void* operator new(size_t size, ArenaAllocator* arena) = delete;
+  static void* operator new(size_t size, ScopedArenaAllocator* arena);
+
  protected:
-  RegType(mirror::Class* klass, const std::string& descriptor,
+  RegType(mirror::Class* klass,
+          const StringPiece& descriptor,
           uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
-      : descriptor_(descriptor), klass_(klass), cache_id_(cache_id) {
+      : descriptor_(descriptor),
+        klass_(klass),
+        cache_id_(cache_id) {
     if (kIsDebugBuild) {
       CheckInvariants();
     }
@@ -285,9 +300,8 @@
 
   void CheckInvariants() const SHARED_REQUIRES(Locks::mutator_lock_);
 
-  const std::string descriptor_;
-  mutable GcRoot<mirror::Class>
-      klass_;  // Non-const only due to moving classes.
+  const StringPiece descriptor_;
+  mutable GcRoot<mirror::Class> klass_;  // Non-const only due to moving classes.
   const uint16_t cache_id_;
 
   friend class RegTypeCache;
@@ -311,7 +325,7 @@
 
   // Create the singleton instance.
   static const ConflictType* CreateInstance(mirror::Class* klass,
-                                            const std::string& descriptor,
+                                            const StringPiece& descriptor,
                                             uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
 
@@ -319,7 +333,7 @@
   static void Destroy();
 
  private:
-  ConflictType(mirror::Class* klass, const std::string& descriptor,
+  ConflictType(mirror::Class* klass, const StringPiece& descriptor,
                uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : RegType(klass, descriptor, cache_id) {}
 
@@ -340,7 +354,7 @@
 
   // Create the singleton instance.
   static const UndefinedType* CreateInstance(mirror::Class* klass,
-                                             const std::string& descriptor,
+                                             const StringPiece& descriptor,
                                              uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
 
@@ -348,7 +362,7 @@
   static void Destroy();
 
  private:
-  UndefinedType(mirror::Class* klass, const std::string& descriptor,
+  UndefinedType(mirror::Class* klass, const StringPiece& descriptor,
                 uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : RegType(klass, descriptor, cache_id) {}
 
@@ -357,7 +371,7 @@
 
 class PrimitiveType : public RegType {
  public:
-  PrimitiveType(mirror::Class* klass, const std::string& descriptor,
+  PrimitiveType(mirror::Class* klass, const StringPiece& descriptor,
                 uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_);
 
   bool HasClassVirtual() const OVERRIDE { return true; }
@@ -365,7 +379,7 @@
 
 class Cat1Type : public PrimitiveType {
  public:
-  Cat1Type(mirror::Class* klass, const std::string& descriptor,
+  Cat1Type(mirror::Class* klass, const StringPiece& descriptor,
            uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_);
 };
 
@@ -374,14 +388,14 @@
   bool IsInteger() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const IntegerType* CreateInstance(mirror::Class* klass,
-                                           const std::string& descriptor,
+                                           const StringPiece& descriptor,
                                            uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const IntegerType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  IntegerType(mirror::Class* klass, const std::string& descriptor,
+  IntegerType(mirror::Class* klass, const StringPiece& descriptor,
               uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
   static const IntegerType* instance_;
@@ -392,14 +406,14 @@
   bool IsBoolean() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const BooleanType* CreateInstance(mirror::Class* klass,
-                                           const std::string& descriptor,
+                                           const StringPiece& descriptor,
                                            uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const BooleanType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  BooleanType(mirror::Class* klass, const std::string& descriptor,
+  BooleanType(mirror::Class* klass, const StringPiece& descriptor,
               uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
 
@@ -411,14 +425,14 @@
   bool IsByte() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const ByteType* CreateInstance(mirror::Class* klass,
-                                        const std::string& descriptor,
+                                        const StringPiece& descriptor,
                                         uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const ByteType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  ByteType(mirror::Class* klass, const std::string& descriptor,
+  ByteType(mirror::Class* klass, const StringPiece& descriptor,
            uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
   static const ByteType* instance_;
@@ -429,14 +443,14 @@
   bool IsShort() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const ShortType* CreateInstance(mirror::Class* klass,
-                                         const std::string& descriptor,
+                                         const StringPiece& descriptor,
                                          uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const ShortType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  ShortType(mirror::Class* klass, const std::string& descriptor,
+  ShortType(mirror::Class* klass, const StringPiece& descriptor,
             uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
   static const ShortType* instance_;
@@ -447,14 +461,14 @@
   bool IsChar() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const CharType* CreateInstance(mirror::Class* klass,
-                                        const std::string& descriptor,
+                                        const StringPiece& descriptor,
                                         uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const CharType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  CharType(mirror::Class* klass, const std::string& descriptor,
+  CharType(mirror::Class* klass, const StringPiece& descriptor,
            uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
   static const CharType* instance_;
@@ -465,14 +479,14 @@
   bool IsFloat() const OVERRIDE { return true; }
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   static const FloatType* CreateInstance(mirror::Class* klass,
-                                         const std::string& descriptor,
+                                         const StringPiece& descriptor,
                                          uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const FloatType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  FloatType(mirror::Class* klass, const std::string& descriptor,
+  FloatType(mirror::Class* klass, const StringPiece& descriptor,
             uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat1Type(klass, descriptor, cache_id) {}
   static const FloatType* instance_;
@@ -480,7 +494,7 @@
 
 class Cat2Type : public PrimitiveType {
  public:
-  Cat2Type(mirror::Class* klass, const std::string& descriptor,
+  Cat2Type(mirror::Class* klass, const StringPiece& descriptor,
            uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_);
 };
 
@@ -490,14 +504,14 @@
   bool IsLongLo() const OVERRIDE { return true; }
   bool IsLong() const OVERRIDE { return true; }
   static const LongLoType* CreateInstance(mirror::Class* klass,
-                                          const std::string& descriptor,
+                                          const StringPiece& descriptor,
                                           uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const LongLoType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  LongLoType(mirror::Class* klass, const std::string& descriptor,
+  LongLoType(mirror::Class* klass, const StringPiece& descriptor,
              uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat2Type(klass, descriptor, cache_id) {}
   static const LongLoType* instance_;
@@ -508,14 +522,14 @@
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   bool IsLongHi() const OVERRIDE { return true; }
   static const LongHiType* CreateInstance(mirror::Class* klass,
-                                          const std::string& descriptor,
+                                          const StringPiece& descriptor,
                                           uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const LongHiType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  LongHiType(mirror::Class* klass, const std::string& descriptor,
+  LongHiType(mirror::Class* klass, const StringPiece& descriptor,
              uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat2Type(klass, descriptor, cache_id) {}
   static const LongHiType* instance_;
@@ -527,14 +541,14 @@
   bool IsDoubleLo() const OVERRIDE { return true; }
   bool IsDouble() const OVERRIDE { return true; }
   static const DoubleLoType* CreateInstance(mirror::Class* klass,
-                                            const std::string& descriptor,
+                                            const StringPiece& descriptor,
                                             uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const DoubleLoType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  DoubleLoType(mirror::Class* klass, const std::string& descriptor,
+  DoubleLoType(mirror::Class* klass, const StringPiece& descriptor,
                uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat2Type(klass, descriptor, cache_id) {}
   static const DoubleLoType* instance_;
@@ -545,14 +559,14 @@
   std::string Dump() const OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_);
   virtual bool IsDoubleHi() const OVERRIDE { return true; }
   static const DoubleHiType* CreateInstance(mirror::Class* klass,
-                                      const std::string& descriptor,
+                                      const StringPiece& descriptor,
                                       uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
   static const DoubleHiType* GetInstance() PURE;
   static void Destroy();
 
  private:
-  DoubleHiType(mirror::Class* klass, const std::string& descriptor,
+  DoubleHiType(mirror::Class* klass, const StringPiece& descriptor,
                uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : Cat2Type(klass, descriptor, cache_id) {}
   static const DoubleHiType* instance_;
@@ -677,7 +691,7 @@
 // instructions and must be passed to a constructor.
 class UninitializedType : public RegType {
  public:
-  UninitializedType(mirror::Class* klass, const std::string& descriptor,
+  UninitializedType(mirror::Class* klass, const StringPiece& descriptor,
                     uint32_t allocation_pc, uint16_t cache_id)
       : RegType(klass, descriptor, cache_id), allocation_pc_(allocation_pc) {}
 
@@ -697,7 +711,7 @@
 class UninitializedReferenceType FINAL : public UninitializedType {
  public:
   UninitializedReferenceType(mirror::Class* klass,
-                             const std::string& descriptor,
+                             const StringPiece& descriptor,
                              uint32_t allocation_pc, uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : UninitializedType(klass, descriptor, allocation_pc, cache_id) {}
@@ -713,7 +727,7 @@
 // constructor.
 class UnresolvedUninitializedRefType FINAL : public UninitializedType {
  public:
-  UnresolvedUninitializedRefType(const std::string& descriptor,
+  UnresolvedUninitializedRefType(const StringPiece& descriptor,
                                  uint32_t allocation_pc, uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : UninitializedType(nullptr, descriptor, allocation_pc, cache_id) {
@@ -737,7 +751,7 @@
 class UninitializedThisReferenceType FINAL : public UninitializedType {
  public:
   UninitializedThisReferenceType(mirror::Class* klass,
-                                 const std::string& descriptor,
+                                 const StringPiece& descriptor,
                                  uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : UninitializedType(klass, descriptor, 0, cache_id) {
@@ -758,7 +772,7 @@
 
 class UnresolvedUninitializedThisRefType FINAL : public UninitializedType {
  public:
-  UnresolvedUninitializedThisRefType(const std::string& descriptor,
+  UnresolvedUninitializedThisRefType(const StringPiece& descriptor,
                                      uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : UninitializedType(nullptr, descriptor, 0, cache_id) {
@@ -781,7 +795,7 @@
 // sub-class.
 class ReferenceType FINAL : public RegType {
  public:
-  ReferenceType(mirror::Class* klass, const std::string& descriptor,
+  ReferenceType(mirror::Class* klass, const StringPiece& descriptor,
                 uint16_t cache_id) SHARED_REQUIRES(Locks::mutator_lock_)
       : RegType(klass, descriptor, cache_id) {}
 
@@ -799,7 +813,7 @@
 // type.
 class PreciseReferenceType FINAL : public RegType {
  public:
-  PreciseReferenceType(mirror::Class* klass, const std::string& descriptor,
+  PreciseReferenceType(mirror::Class* klass, const StringPiece& descriptor,
                        uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
 
@@ -815,7 +829,7 @@
 // Common parent of unresolved types.
 class UnresolvedType : public RegType {
  public:
-  UnresolvedType(const std::string& descriptor, uint16_t cache_id)
+  UnresolvedType(const StringPiece& descriptor, uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : RegType(nullptr, descriptor, cache_id) {}
 
@@ -827,7 +841,7 @@
 // of this type must be conservative.
 class UnresolvedReferenceType FINAL : public UnresolvedType {
  public:
-  UnresolvedReferenceType(const std::string& descriptor, uint16_t cache_id)
+  UnresolvedReferenceType(const StringPiece& descriptor, uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_)
       : UnresolvedType(descriptor, cache_id) {
     if (kIsDebugBuild) {
@@ -882,8 +896,10 @@
 class UnresolvedMergedType FINAL : public UnresolvedType {
  public:
   // Note: the constructor will copy the unresolved BitVector, not use it directly.
-  UnresolvedMergedType(const RegType& resolved, const BitVector& unresolved,
-                       const RegTypeCache* reg_type_cache, uint16_t cache_id)
+  UnresolvedMergedType(const RegType& resolved,
+                       const BitVector& unresolved,
+                       const RegTypeCache* reg_type_cache,
+                       uint16_t cache_id)
       SHARED_REQUIRES(Locks::mutator_lock_);
 
   // The resolved part. See description below.
diff --git a/runtime/verifier/reg_type_cache-inl.h b/runtime/verifier/reg_type_cache-inl.h
index b6f253b..68af62e 100644
--- a/runtime/verifier/reg_type_cache-inl.h
+++ b/runtime/verifier/reg_type_cache-inl.h
@@ -118,6 +118,18 @@
   }
 }
 
+template <class RegTypeType>
+inline RegTypeType& RegTypeCache::AddEntry(RegTypeType* new_entry) {
+  DCHECK(new_entry != nullptr);
+  entries_.push_back(new_entry);
+  if (new_entry->HasClass()) {
+    mirror::Class* klass = new_entry->GetClass();
+    DCHECK(!klass->IsPrimitive());
+    klass_entries_.push_back(std::make_pair(GcRoot<mirror::Class>(klass), new_entry));
+  }
+  return *new_entry;
+}
+
 }  // namespace verifier
 }  // namespace art
 #endif  // ART_RUNTIME_VERIFIER_REG_TYPE_CACHE_INL_H_
diff --git a/runtime/verifier/reg_type_cache.cc b/runtime/verifier/reg_type_cache.cc
index bb756e9..71ed4a2 100644
--- a/runtime/verifier/reg_type_cache.cc
+++ b/runtime/verifier/reg_type_cache.cc
@@ -16,7 +16,9 @@
 
 #include "reg_type_cache-inl.h"
 
+#include "base/arena_bit_vector.h"
 #include "base/casts.h"
+#include "base/scoped_arena_allocator.h"
 #include "base/stl_util.h"
 #include "class_linker-inl.h"
 #include "dex_file-inl.h"
@@ -29,9 +31,10 @@
 
 bool RegTypeCache::primitive_initialized_ = false;
 uint16_t RegTypeCache::primitive_count_ = 0;
-const PreciseConstType* RegTypeCache::small_precise_constants_[kMaxSmallConstant - kMinSmallConstant + 1];
+const PreciseConstType* RegTypeCache::small_precise_constants_[kMaxSmallConstant -
+                                                               kMinSmallConstant + 1];
 
-static bool MatchingPrecisionForClass(const RegType* entry, bool precise)
+ALWAYS_INLINE static inline bool MatchingPrecisionForClass(const RegType* entry, bool precise)
     SHARED_REQUIRES(Locks::mutator_lock_) {
   if (entry->IsPreciseReference() == precise) {
     // We were or weren't looking for a precise reference and we found what we need.
@@ -67,7 +70,8 @@
   DCHECK_EQ(entries_.size(), primitive_count_);
 }
 
-const RegType& RegTypeCache::FromDescriptor(mirror::ClassLoader* loader, const char* descriptor,
+const RegType& RegTypeCache::FromDescriptor(mirror::ClassLoader* loader,
+                                            const char* descriptor,
                                             bool precise) {
   DCHECK(RegTypeCache::primitive_initialized_);
   if (descriptor[1] == '\0') {
@@ -159,13 +163,20 @@
   return klass;
 }
 
-const RegType& RegTypeCache::From(mirror::ClassLoader* loader, const char* descriptor,
+StringPiece RegTypeCache::AddString(const StringPiece& string_piece) {
+  char* ptr = arena_.AllocArray<char>(string_piece.length());
+  memcpy(ptr, string_piece.data(), string_piece.length());
+  return StringPiece(ptr, string_piece.length());
+}
+
+const RegType& RegTypeCache::From(mirror::ClassLoader* loader,
+                                  const char* descriptor,
                                   bool precise) {
+  StringPiece sp_descriptor(descriptor);
   // Try looking up the class in the cache first. We use a StringPiece to avoid continual strlen
   // operations on the descriptor.
-  StringPiece descriptor_sp(descriptor);
   for (size_t i = primitive_count_; i < entries_.size(); i++) {
-    if (MatchDescriptor(i, descriptor_sp, precise)) {
+    if (MatchDescriptor(i, sp_descriptor, precise)) {
       return *(entries_[i]);
     }
   }
@@ -186,12 +197,11 @@
     if (klass->CannotBeAssignedFromOtherTypes() || precise) {
       DCHECK(!(klass->IsAbstract()) || klass->IsArrayClass());
       DCHECK(!klass->IsInterface());
-      entry = new PreciseReferenceType(klass, descriptor_sp.as_string(), entries_.size());
+      entry = new (&arena_) PreciseReferenceType(klass, AddString(sp_descriptor), entries_.size());
     } else {
-      entry = new ReferenceType(klass, descriptor_sp.as_string(), entries_.size());
+      entry = new (&arena_) ReferenceType(klass, AddString(sp_descriptor), entries_.size());
     }
-    AddEntry(entry);
-    return *entry;
+    return AddEntry(entry);
   } else {  // Class not resolved.
     // We tried loading the class and failed, this might get an exception raised
     // so we want to clear it before we go on.
@@ -202,9 +212,8 @@
       DCHECK(!Thread::Current()->IsExceptionPending());
     }
     if (IsValidDescriptor(descriptor)) {
-      RegType* entry = new UnresolvedReferenceType(descriptor_sp.as_string(), entries_.size());
-      AddEntry(entry);
-      return *entry;
+      return AddEntry(
+          new (&arena_) UnresolvedReferenceType(AddString(sp_descriptor), entries_.size()));
     } else {
       // The descriptor is broken return the unknown type as there's nothing sensible that
       // could be done at runtime
@@ -213,50 +222,65 @@
   }
 }
 
-const RegType& RegTypeCache::FromClass(const char* descriptor, mirror::Class* klass, bool precise) {
+const RegType* RegTypeCache::FindClass(mirror::Class* klass, bool precise) const {
   DCHECK(klass != nullptr);
   if (klass->IsPrimitive()) {
     // Note: precise isn't used for primitive classes. A char is assignable to an int. All
     // primitive classes are final.
-    return RegTypeFromPrimitiveType(klass->GetPrimitiveType());
-  } else {
-    // Look for the reference in the list of entries to have.
-    for (size_t i = primitive_count_; i < entries_.size(); i++) {
-      const RegType* cur_entry = entries_[i];
-      if (cur_entry->klass_.Read() == klass && MatchingPrecisionForClass(cur_entry, precise)) {
-        return *cur_entry;
+    return &RegTypeFromPrimitiveType(klass->GetPrimitiveType());
+  }
+  for (auto& pair : klass_entries_) {
+    mirror::Class* const reg_klass = pair.first.Read();
+    if (reg_klass == klass) {
+      const RegType* reg_type = pair.second;
+      if (MatchingPrecisionForClass(reg_type, precise)) {
+        return reg_type;
       }
     }
-    // No reference to the class was found, create new reference.
-    RegType* entry;
-    if (precise) {
-      entry = new PreciseReferenceType(klass, descriptor, entries_.size());
-    } else {
-      entry = new ReferenceType(klass, descriptor, entries_.size());
-    }
-    AddEntry(entry);
-    return *entry;
   }
+  return nullptr;
 }
 
-RegTypeCache::RegTypeCache(bool can_load_classes) : can_load_classes_(can_load_classes) {
+const RegType* RegTypeCache::InsertClass(const StringPiece& descriptor,
+                                         mirror::Class* klass,
+                                         bool precise) {
+  // No reference to the class was found, create new reference.
+  DCHECK(FindClass(klass, precise) == nullptr);
+  RegType* const reg_type = precise
+      ? static_cast<RegType*>(
+          new (&arena_) PreciseReferenceType(klass, descriptor, entries_.size()))
+      : new (&arena_) ReferenceType(klass, descriptor, entries_.size());
+  return &AddEntry(reg_type);
+}
+
+const RegType& RegTypeCache::FromClass(const char* descriptor, mirror::Class* klass, bool precise) {
+  DCHECK(klass != nullptr);
+  const RegType* reg_type = FindClass(klass, precise);
+  if (reg_type == nullptr) {
+    reg_type = InsertClass(AddString(StringPiece(descriptor)), klass, precise);
+  }
+  return *reg_type;
+}
+
+RegTypeCache::RegTypeCache(bool can_load_classes, ScopedArenaAllocator& arena)
+    : entries_(arena.Adapter(kArenaAllocVerifier)),
+      klass_entries_(arena.Adapter(kArenaAllocVerifier)),
+      can_load_classes_(can_load_classes),
+      arena_(arena) {
   if (kIsDebugBuild) {
     Thread::Current()->AssertThreadSuspensionIsAllowable(gAborting == 0);
   }
-  entries_.reserve(64);
+  // The klass_entries_ array does not have primitives or small constants.
+  static constexpr size_t kNumReserveEntries = 32;
+  klass_entries_.reserve(kNumReserveEntries);
+  // We want to have room for additional entries after inserting primitives and small
+  // constants.
+  entries_.reserve(kNumReserveEntries + kNumPrimitivesAndSmallConstants);
   FillPrimitiveAndSmallConstantTypes();
 }
 
 RegTypeCache::~RegTypeCache() {
-  CHECK_LE(primitive_count_, entries_.size());
-  // Delete only the non primitive types.
-  if (entries_.size() == kNumPrimitivesAndSmallConstants) {
-    // All entries are from the global pool, nothing to delete.
-    return;
-  }
-  std::vector<const RegType*>::iterator non_primitive_begin = entries_.begin();
-  std::advance(non_primitive_begin, kNumPrimitivesAndSmallConstants);
-  STLDeleteContainerPointers(non_primitive_begin, entries_.end());
+  DCHECK_LE(primitive_count_, entries_.size());
 }
 
 void RegTypeCache::ShutDown() {
@@ -318,9 +342,9 @@
 }
 
 const RegType& RegTypeCache::FromUnresolvedMerge(const RegType& left, const RegType& right) {
-  BitVector types(1,                                    // Allocate at least a word.
-                  true,                                 // Is expandable.
-                  Allocator::GetMallocAllocator());     // TODO: Arenas in the verifier.
+  ArenaBitVector types(&arena_,
+                       kDefaultArenaBitVectorBytes * kBitsPerByte,  // Allocate at least 8 bytes.
+                       true);                                       // Is expandable.
   const RegType* left_resolved;
   if (left.IsUnresolvedMergedReference()) {
     const UnresolvedMergedType* left_merge = down_cast<const UnresolvedMergedType*>(&left);
@@ -361,20 +385,15 @@
       const BitVector& unresolved_part = cmp_type->GetUnresolvedTypes();
       // Use SameBitsSet. "types" is expandable to allow merging in the components, but the
       // BitVector in the final RegType will be made non-expandable.
-      if (&resolved_part == &resolved_parts_merged &&
-              types.SameBitsSet(&unresolved_part)) {
+      if (&resolved_part == &resolved_parts_merged && types.SameBitsSet(&unresolved_part)) {
         return *cur_entry;
       }
     }
   }
-
-  // Create entry.
-  RegType* entry = new UnresolvedMergedType(resolved_parts_merged,
-                                            types,
-                                            this,
-                                            entries_.size());
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(new (&arena_) UnresolvedMergedType(resolved_parts_merged,
+                                                     types,
+                                                     this,
+                                                     entries_.size()));
 }
 
 const RegType& RegTypeCache::FromUnresolvedSuperClass(const RegType& child) {
@@ -391,14 +410,12 @@
       }
     }
   }
-  RegType* entry = new UnresolvedSuperClass(child.GetId(), this, entries_.size());
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(new (&arena_) UnresolvedSuperClass(child.GetId(), this, entries_.size()));
 }
 
 const UninitializedType& RegTypeCache::Uninitialized(const RegType& type, uint32_t allocation_pc) {
   UninitializedType* entry = nullptr;
-  const std::string& descriptor(type.GetDescriptor());
+  const StringPiece& descriptor(type.GetDescriptor());
   if (type.IsUnresolvedTypes()) {
     for (size_t i = primitive_count_; i < entries_.size(); i++) {
       const RegType* cur_entry = entries_[i];
@@ -409,7 +426,9 @@
         return *down_cast<const UnresolvedUninitializedRefType*>(cur_entry);
       }
     }
-    entry = new UnresolvedUninitializedRefType(descriptor, allocation_pc, entries_.size());
+    entry = new (&arena_) UnresolvedUninitializedRefType(descriptor,
+                                                         allocation_pc,
+                                                         entries_.size());
   } else {
     mirror::Class* klass = type.GetClass();
     for (size_t i = primitive_count_; i < entries_.size(); i++) {
@@ -421,17 +440,19 @@
         return *down_cast<const UninitializedReferenceType*>(cur_entry);
       }
     }
-    entry = new UninitializedReferenceType(klass, descriptor, allocation_pc, entries_.size());
+    entry = new (&arena_) UninitializedReferenceType(klass,
+                                                     descriptor,
+                                                     allocation_pc,
+                                                     entries_.size());
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const RegType& RegTypeCache::FromUninitialized(const RegType& uninit_type) {
   RegType* entry;
 
   if (uninit_type.IsUnresolvedTypes()) {
-    const std::string& descriptor(uninit_type.GetDescriptor());
+    const StringPiece& descriptor(uninit_type.GetDescriptor());
     for (size_t i = primitive_count_; i < entries_.size(); i++) {
       const RegType* cur_entry = entries_[i];
       if (cur_entry->IsUnresolvedReference() &&
@@ -439,7 +460,7 @@
         return *cur_entry;
       }
     }
-    entry = new UnresolvedReferenceType(descriptor, entries_.size());
+    entry = new (&arena_) UnresolvedReferenceType(descriptor, entries_.size());
   } else {
     mirror::Class* klass = uninit_type.GetClass();
     if (uninit_type.IsUninitializedThisReference() && !klass->IsFinal()) {
@@ -450,7 +471,7 @@
           return *cur_entry;
         }
       }
-      entry = new ReferenceType(klass, "", entries_.size());
+      entry = new (&arena_) ReferenceType(klass, "", entries_.size());
     } else if (!klass->IsPrimitive()) {
       // We're uninitialized because of allocation, look or create a precise type as allocations
       // may only create objects of that type.
@@ -469,18 +490,19 @@
           return *cur_entry;
         }
       }
-      entry = new PreciseReferenceType(klass, uninit_type.GetDescriptor(), entries_.size());
+      entry = new (&arena_) PreciseReferenceType(klass,
+                                                 uninit_type.GetDescriptor(),
+                                                 entries_.size());
     } else {
       return Conflict();
     }
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const UninitializedType& RegTypeCache::UninitializedThisArgument(const RegType& type) {
   UninitializedType* entry;
-  const std::string& descriptor(type.GetDescriptor());
+  const StringPiece& descriptor(type.GetDescriptor());
   if (type.IsUnresolvedTypes()) {
     for (size_t i = primitive_count_; i < entries_.size(); i++) {
       const RegType* cur_entry = entries_[i];
@@ -489,7 +511,7 @@
         return *down_cast<const UninitializedType*>(cur_entry);
       }
     }
-    entry = new UnresolvedUninitializedThisRefType(descriptor, entries_.size());
+    entry = new (&arena_) UnresolvedUninitializedThisRefType(descriptor, entries_.size());
   } else {
     mirror::Class* klass = type.GetClass();
     for (size_t i = primitive_count_; i < entries_.size(); i++) {
@@ -498,10 +520,9 @@
         return *down_cast<const UninitializedType*>(cur_entry);
       }
     }
-    entry = new UninitializedThisReferenceType(klass, descriptor, entries_.size());
+    entry = new (&arena_) UninitializedThisReferenceType(klass, descriptor, entries_.size());
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const ConstantType& RegTypeCache::FromCat1NonSmallConstant(int32_t value, bool precise) {
@@ -515,12 +536,11 @@
   }
   ConstantType* entry;
   if (precise) {
-    entry = new PreciseConstType(value, entries_.size());
+    entry = new (&arena_) PreciseConstType(value, entries_.size());
   } else {
-    entry = new ImpreciseConstType(value, entries_.size());
+    entry = new (&arena_) ImpreciseConstType(value, entries_.size());
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const ConstantType& RegTypeCache::FromCat2ConstLo(int32_t value, bool precise) {
@@ -533,12 +553,11 @@
   }
   ConstantType* entry;
   if (precise) {
-    entry = new PreciseConstLoType(value, entries_.size());
+    entry = new (&arena_) PreciseConstLoType(value, entries_.size());
   } else {
-    entry = new ImpreciseConstLoType(value, entries_.size());
+    entry = new (&arena_) ImpreciseConstLoType(value, entries_.size());
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const ConstantType& RegTypeCache::FromCat2ConstHi(int32_t value, bool precise) {
@@ -551,32 +570,30 @@
   }
   ConstantType* entry;
   if (precise) {
-    entry = new PreciseConstHiType(value, entries_.size());
+    entry = new (&arena_) PreciseConstHiType(value, entries_.size());
   } else {
-    entry = new ImpreciseConstHiType(value, entries_.size());
+    entry = new (&arena_) ImpreciseConstHiType(value, entries_.size());
   }
-  AddEntry(entry);
-  return *entry;
+  return AddEntry(entry);
 }
 
 const RegType& RegTypeCache::GetComponentType(const RegType& array, mirror::ClassLoader* loader) {
   if (!array.IsArrayTypes()) {
     return Conflict();
   } else if (array.IsUnresolvedTypes()) {
-    const std::string& descriptor(array.GetDescriptor());
-    const std::string component(descriptor.substr(1, descriptor.size() - 1));
-    return FromDescriptor(loader, component.c_str(), false);
+    const std::string descriptor(array.GetDescriptor().as_string());
+    return FromDescriptor(loader, descriptor.c_str() + 1, false);
   } else {
     mirror::Class* klass = array.GetClass()->GetComponentType();
     std::string temp;
+    const char* descriptor = klass->GetDescriptor(&temp);
     if (klass->IsErroneous()) {
       // Arrays may have erroneous component types, use unresolved in that case.
       // We assume that the primitive classes are not erroneous, so we know it is a
       // reference type.
-      return FromDescriptor(loader, klass->GetDescriptor(&temp), false);
+      return FromDescriptor(loader, descriptor, false);
     } else {
-      return FromClass(klass->GetDescriptor(&temp), klass,
-                       klass->CannotBeAssignedFromOtherTypes());
+      return FromClass(descriptor, klass, klass->CannotBeAssignedFromOtherTypes());
     }
   }
 }
@@ -618,10 +635,10 @@
   for (size_t i = primitive_count_; i < entries_.size(); ++i) {
     entries_[i]->VisitRoots(visitor, root_info);
   }
-}
-
-void RegTypeCache::AddEntry(RegType* new_entry) {
-  entries_.push_back(new_entry);
+  for (auto& pair : klass_entries_) {
+    GcRoot<mirror::Class>& root = pair.first;
+    root.VisitRoot(visitor, root_info);
+  }
 }
 
 }  // namespace verifier
diff --git a/runtime/verifier/reg_type_cache.h b/runtime/verifier/reg_type_cache.h
index 93948a1..6f9a04e 100644
--- a/runtime/verifier/reg_type_cache.h
+++ b/runtime/verifier/reg_type_cache.h
@@ -19,6 +19,7 @@
 
 #include "base/casts.h"
 #include "base/macros.h"
+#include "base/scoped_arena_containers.h"
 #include "object_callbacks.h"
 #include "reg_type.h"
 #include "runtime.h"
@@ -31,15 +32,19 @@
   class Class;
   class ClassLoader;
 }  // namespace mirror
+class ScopedArenaAllocator;
 class StringPiece;
 
 namespace verifier {
 
 class RegType;
 
+// Use 8 bytes since that is the default arena allocator alignment.
+static constexpr size_t kDefaultArenaBitVectorBytes = 8;
+
 class RegTypeCache {
  public:
-  explicit RegTypeCache(bool can_load_classes);
+  explicit RegTypeCache(bool can_load_classes, ScopedArenaAllocator& arena);
   ~RegTypeCache();
   static void Init() SHARED_REQUIRES(Locks::mutator_lock_) {
     if (!RegTypeCache::primitive_initialized_) {
@@ -53,6 +58,13 @@
   const art::verifier::RegType& GetFromId(uint16_t id) const;
   const RegType& From(mirror::ClassLoader* loader, const char* descriptor, bool precise)
       SHARED_REQUIRES(Locks::mutator_lock_);
+  // Find a RegType, returns null if not found.
+  const RegType* FindClass(mirror::Class* klass, bool precise) const
+      SHARED_REQUIRES(Locks::mutator_lock_);
+  // Insert a new class with a specified descriptor, must not already be in the cache.
+  const RegType* InsertClass(const StringPiece& descriptor, mirror::Class* klass, bool precise)
+      SHARED_REQUIRES(Locks::mutator_lock_);
+  // Get or insert a reg type for a description, klass, and precision.
   const RegType& FromClass(const char* descriptor, mirror::Class* klass, bool precise)
       SHARED_REQUIRES(Locks::mutator_lock_);
   const ConstantType& FromCat1Const(int32_t value, bool precise)
@@ -150,7 +162,13 @@
   const ConstantType& FromCat1NonSmallConstant(int32_t value, bool precise)
       SHARED_REQUIRES(Locks::mutator_lock_);
 
-  void AddEntry(RegType* new_entry);
+  // Returns the pass in RegType.
+  template <class RegTypeType>
+  RegTypeType& AddEntry(RegTypeType* new_entry) SHARED_REQUIRES(Locks::mutator_lock_);
+
+  // Add a string piece to the arena allocator so that it stays live for the lifetime of the
+  // verifier.
+  StringPiece AddString(const StringPiece& string_piece);
 
   template <class Type>
   static const Type* CreatePrimitiveTypeInstance(const std::string& descriptor)
@@ -160,7 +178,8 @@
   // A quick look up for popular small constants.
   static constexpr int32_t kMinSmallConstant = -1;
   static constexpr int32_t kMaxSmallConstant = 4;
-  static const PreciseConstType* small_precise_constants_[kMaxSmallConstant - kMinSmallConstant + 1];
+  static const PreciseConstType* small_precise_constants_[kMaxSmallConstant -
+                                                          kMinSmallConstant + 1];
 
   static constexpr size_t kNumPrimitivesAndSmallConstants =
       12 + (kMaxSmallConstant - kMinSmallConstant + 1);
@@ -172,11 +191,17 @@
   static uint16_t primitive_count_;
 
   // The actual storage for the RegTypes.
-  std::vector<const RegType*> entries_;
+  ScopedArenaVector<const RegType*> entries_;
+
+  // Fast lookup for quickly finding entries that have a matching class.
+  ScopedArenaVector<std::pair<GcRoot<mirror::Class>, const RegType*>> klass_entries_;
 
   // Whether or not we're allowed to load classes.
   const bool can_load_classes_;
 
+  // Arena allocator.
+  ScopedArenaAllocator& arena_;
+
   DISALLOW_COPY_AND_ASSIGN(RegTypeCache);
 };
 
diff --git a/runtime/verifier/reg_type_test.cc b/runtime/verifier/reg_type_test.cc
index 971b1f5..22ac7e4 100644
--- a/runtime/verifier/reg_type_test.cc
+++ b/runtime/verifier/reg_type_test.cc
@@ -20,6 +20,7 @@
 
 #include "base/bit_vector.h"
 #include "base/casts.h"
+#include "base/scoped_arena_allocator.h"
 #include "common_runtime_test.h"
 #include "reg_type_cache-inl.h"
 #include "reg_type-inl.h"
@@ -29,12 +30,23 @@
 namespace art {
 namespace verifier {
 
-class RegTypeTest : public CommonRuntimeTest {};
+class BaseRegTypeTest : public CommonRuntimeTest {
+ public:
+  void PostRuntimeCreate() OVERRIDE {
+    stack.reset(new ArenaStack(Runtime::Current()->GetArenaPool()));
+    allocator.reset(new ScopedArenaAllocator(stack.get()));
+  }
+
+  std::unique_ptr<ArenaStack> stack;
+  std::unique_ptr<ScopedArenaAllocator> allocator;
+};
+
+class RegTypeTest : public BaseRegTypeTest {};
 
 TEST_F(RegTypeTest, ConstLoHi) {
   // Tests creating primitive types types.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& ref_type_const_0 = cache.FromCat1Const(10, true);
   const RegType& ref_type_const_1 = cache.FromCat1Const(10, true);
   const RegType& ref_type_const_2 = cache.FromCat1Const(30, true);
@@ -56,7 +68,7 @@
 
 TEST_F(RegTypeTest, Pairs) {
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   int64_t val = static_cast<int32_t>(1234);
   const RegType& precise_lo = cache.FromCat2ConstLo(static_cast<int32_t>(val), true);
   const RegType& precise_hi = cache.FromCat2ConstHi(static_cast<int32_t>(val >> 32), true);
@@ -80,7 +92,7 @@
 
 TEST_F(RegTypeTest, Primitives) {
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
 
   const RegType& bool_reg_type = cache.Boolean();
   EXPECT_FALSE(bool_reg_type.IsUndefined());
@@ -347,13 +359,13 @@
   EXPECT_TRUE(double_reg_type.HasClass());
 }
 
-class RegTypeReferenceTest : public CommonRuntimeTest {};
+class RegTypeReferenceTest : public BaseRegTypeTest {};
 
 TEST_F(RegTypeReferenceTest, JavalangObjectImprecise) {
   // Tests matching precisions. A reference type that was created precise doesn't
   // match the one that is imprecise.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& imprecise_obj = cache.JavaLangObject(false);
   const RegType& precise_obj = cache.JavaLangObject(true);
   const RegType& precise_obj_2 = cache.FromDescriptor(nullptr, "Ljava/lang/Object;", true);
@@ -368,7 +380,7 @@
   // Tests creating unresolved types. Miss for the first time asking the cache and
   // a hit second time.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& ref_type_0 = cache.FromDescriptor(nullptr, "Ljava/lang/DoesNotExist;", true);
   EXPECT_TRUE(ref_type_0.IsUnresolvedReference());
   EXPECT_TRUE(ref_type_0.IsNonZeroReferenceTypes());
@@ -384,7 +396,7 @@
 TEST_F(RegTypeReferenceTest, UnresolvedUnintializedType) {
   // Tests creating types uninitialized types from unresolved types.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& ref_type_0 = cache.FromDescriptor(nullptr, "Ljava/lang/DoesNotExist;", true);
   EXPECT_TRUE(ref_type_0.IsUnresolvedReference());
   const RegType& ref_type = cache.FromDescriptor(nullptr, "Ljava/lang/DoesNotExist;", true);
@@ -406,7 +418,7 @@
 TEST_F(RegTypeReferenceTest, Dump) {
   // Tests types for proper Dump messages.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& unresolved_ref = cache.FromDescriptor(nullptr, "Ljava/lang/DoesNotExist;", true);
   const RegType& unresolved_ref_another = cache.FromDescriptor(nullptr, "Ljava/lang/DoesNotExistEither;", true);
   const RegType& resolved_ref = cache.JavaLangString();
@@ -431,7 +443,7 @@
   // Hit the second time. Then check for the same effect when using
   // The JavaLangObject method instead of FromDescriptor. String class is final.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& ref_type = cache.JavaLangString();
   const RegType& ref_type_2 = cache.JavaLangString();
   const RegType& ref_type_3 = cache.FromDescriptor(nullptr, "Ljava/lang/String;", true);
@@ -451,7 +463,7 @@
   // Hit the second time. Then I am checking for the same effect when using
   // The JavaLangObject method instead of FromDescriptor. Object Class in not final.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache(true);
+  RegTypeCache cache(true, *allocator);
   const RegType& ref_type = cache.JavaLangObject(true);
   const RegType& ref_type_2 = cache.JavaLangObject(true);
   const RegType& ref_type_3 = cache.FromDescriptor(nullptr, "Ljava/lang/Object;", true);
@@ -464,7 +476,7 @@
   // Tests merging logic
   // String and object , LUB is object.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache_new(true);
+  RegTypeCache cache_new(true, *allocator);
   const RegType& string = cache_new.JavaLangString();
   const RegType& Object = cache_new.JavaLangObject(true);
   EXPECT_TRUE(string.Merge(Object, &cache_new).IsJavaLangObject());
@@ -487,7 +499,7 @@
 TEST_F(RegTypeTest, MergingFloat) {
   // Testing merging logic with float and float constants.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache_new(true);
+  RegTypeCache cache_new(true, *allocator);
 
   constexpr int32_t kTestConstantValue = 10;
   const RegType& float_type = cache_new.Float();
@@ -518,7 +530,7 @@
 TEST_F(RegTypeTest, MergingLong) {
   // Testing merging logic with long and long constants.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache_new(true);
+  RegTypeCache cache_new(true, *allocator);
 
   constexpr int32_t kTestConstantValue = 10;
   const RegType& long_lo_type = cache_new.LongLo();
@@ -572,7 +584,7 @@
 TEST_F(RegTypeTest, MergingDouble) {
   // Testing merging logic with double and double constants.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache_new(true);
+  RegTypeCache cache_new(true, *allocator);
 
   constexpr int32_t kTestConstantValue = 10;
   const RegType& double_lo_type = cache_new.DoubleLo();
@@ -626,7 +638,7 @@
 TEST_F(RegTypeTest, ConstPrecision) {
   // Tests creating primitive types types.
   ScopedObjectAccess soa(Thread::Current());
-  RegTypeCache cache_new(true);
+  RegTypeCache cache_new(true, *allocator);
   const RegType& imprecise_const = cache_new.FromCat1Const(10, false);
   const RegType& precise_const = cache_new.FromCat1Const(10, true);
 
diff --git a/runtime/verifier/register_line-inl.h b/runtime/verifier/register_line-inl.h
index 1df2428..57fb701 100644
--- a/runtime/verifier/register_line-inl.h
+++ b/runtime/verifier/register_line-inl.h
@@ -182,6 +182,21 @@
   }
 }
 
+inline RegisterLine* RegisterLine::Create(size_t num_regs, MethodVerifier* verifier) {
+  void* memory = verifier->GetArena().Alloc(OFFSETOF_MEMBER(RegisterLine, line_) +
+                                                (num_regs * sizeof(uint16_t)));
+  return new (memory) RegisterLine(num_regs, verifier);
+}
+
+inline RegisterLine::RegisterLine(size_t num_regs, MethodVerifier* verifier)
+    : num_regs_(num_regs),
+      monitors_(verifier->GetArena().Adapter(kArenaAllocVerifier)),
+      reg_to_lock_depths_(std::less<uint32_t>(), verifier->GetArena().Adapter(kArenaAllocVerifier)),
+      this_initialized_(false) {
+  std::uninitialized_fill_n(line_, num_regs_, 0u);
+  SetResultTypeToUnknown(verifier);
+}
+
 }  // namespace verifier
 }  // namespace art
 
diff --git a/runtime/verifier/register_line.cc b/runtime/verifier/register_line.cc
index f48b1e1..37343b5 100644
--- a/runtime/verifier/register_line.cc
+++ b/runtime/verifier/register_line.cc
@@ -412,12 +412,9 @@
   }
 }
 
-// Check whether there is another register in the search map that is locked the same way as the
-// register in the src map. This establishes an alias.
-static bool FindLockAliasedRegister(
-    uint32_t src,
-    const AllocationTrackingSafeMap<uint32_t, uint32_t, kAllocatorTagVerifier>& src_map,
-    const AllocationTrackingSafeMap<uint32_t, uint32_t, kAllocatorTagVerifier>& search_map) {
+bool FindLockAliasedRegister(uint32_t src,
+                             const RegisterLine::RegToLockDepthsMap& src_map,
+                             const RegisterLine::RegToLockDepthsMap& search_map) {
   auto it = src_map.find(src);
   if (it == src_map.end()) {
     // "Not locked" is trivially aliased.
diff --git a/runtime/verifier/register_line.h b/runtime/verifier/register_line.h
index 46db1c6..b2f5555 100644
--- a/runtime/verifier/register_line.h
+++ b/runtime/verifier/register_line.h
@@ -20,6 +20,7 @@
 #include <memory>
 #include <vector>
 
+#include "base/scoped_arena_containers.h"
 #include "safe_map.h"
 
 namespace art {
@@ -58,11 +59,11 @@
 // stack of entered monitors (identified by code unit offset).
 class RegisterLine {
  public:
-  static RegisterLine* Create(size_t num_regs, MethodVerifier* verifier) {
-    void* memory = operator new(sizeof(RegisterLine) + (num_regs * sizeof(uint16_t)));
-    RegisterLine* rl = new (memory) RegisterLine(num_regs, verifier);
-    return rl;
-  }
+  // A map from register to a bit vector of indices into the monitors_ stack.
+  using RegToLockDepthsMap = ScopedArenaSafeMap<uint32_t, uint32_t>;
+
+  // Create a register line of num_regs registers.
+  static RegisterLine* Create(size_t num_regs, MethodVerifier* verifier);
 
   // Implement category-1 "move" instructions. Copy a 32-bit value from "vsrc" to "vdst".
   void CopyRegister1(MethodVerifier* verifier, uint32_t vdst, uint32_t vsrc, TypeCategory cat)
@@ -311,11 +312,11 @@
   // Write a bit at each register location that holds a reference.
   void WriteReferenceBitMap(MethodVerifier* verifier, std::vector<uint8_t>* data, size_t max_bytes);
 
-  size_t GetMonitorEnterCount() {
+  size_t GetMonitorEnterCount() const {
     return monitors_.size();
   }
 
-  uint32_t GetMonitorEnterDexPc(size_t i) {
+  uint32_t GetMonitorEnterDexPc(size_t i) const {
     return monitors_[i];
   }
 
@@ -375,11 +376,7 @@
     reg_to_lock_depths_.erase(reg);
   }
 
-  RegisterLine(size_t num_regs, MethodVerifier* verifier)
-      : num_regs_(num_regs), this_initialized_(false) {
-    memset(&line_, 0, num_regs_ * sizeof(uint16_t));
-    SetResultTypeToUnknown(verifier);
-  }
+  RegisterLine(size_t num_regs, MethodVerifier* verifier);
 
   // Storage for the result register's type, valid after an invocation.
   uint16_t result_[2];
@@ -388,17 +385,18 @@
   const uint32_t num_regs_;
 
   // A stack of monitor enter locations.
-  std::vector<uint32_t, TrackingAllocator<uint32_t, kAllocatorTagVerifier>> monitors_;
+  ScopedArenaVector<uint32_t> monitors_;
+
   // A map from register to a bit vector of indices into the monitors_ stack. As we pop the monitor
   // stack we verify that monitor-enter/exit are correctly nested. That is, if there was a
   // monitor-enter on v5 and then on v6, we expect the monitor-exit to be on v6 then on v5.
-  AllocationTrackingSafeMap<uint32_t, uint32_t, kAllocatorTagVerifier> reg_to_lock_depths_;
+  RegToLockDepthsMap reg_to_lock_depths_;
 
   // Whether "this" initialization (a constructor supercall) has happened.
   bool this_initialized_;
 
   // An array of RegType Ids associated with each dex register.
-  uint16_t line_[0];
+  uint16_t line_[1];
 
   DISALLOW_COPY_AND_ASSIGN(RegisterLine);
 };
diff --git a/test/530-checker-lse/src/Main.java b/test/530-checker-lse/src/Main.java
index ea35b37..c766aaa 100644
--- a/test/530-checker-lse/src/Main.java
+++ b/test/530-checker-lse/src/Main.java
@@ -56,7 +56,7 @@
 
   /// CHECK-START: double Main.calcCircleArea(double) load_store_elimination (after)
   /// CHECK: NewInstance
-  /// CHECK-NOT: InstanceFieldSet
+  /// CHECK: InstanceFieldSet
   /// CHECK-NOT: InstanceFieldGet
 
   static double calcCircleArea(double radius) {
@@ -117,8 +117,7 @@
   /// CHECK: InstanceFieldGet
   /// CHECK: InstanceFieldSet
   /// CHECK: NewInstance
-  /// CHECK-NOT: NullCheck
-  /// CHECK-NOT: InstanceFieldSet
+  /// CHECK: InstanceFieldSet
   /// CHECK-NOT: InstanceFieldGet
 
   // A new allocation shouldn't alias with pre-existing values.
@@ -224,7 +223,7 @@
 
   /// CHECK-START: int Main.test8() load_store_elimination (after)
   /// CHECK: NewInstance
-  /// CHECK-NOT: InstanceFieldSet
+  /// CHECK: InstanceFieldSet
   /// CHECK: InvokeVirtual
   /// CHECK-NOT: NullCheck
   /// CHECK-NOT: InstanceFieldGet
diff --git a/test/955-lambda-smali/run b/test/955-lambda-smali/run
index b754680..2fb2f89 100755
--- a/test/955-lambda-smali/run
+++ b/test/955-lambda-smali/run
@@ -15,4 +15,4 @@
 # limitations under the License.
 
 # Ensure that the lambda experimental opcodes are turned on for dalvikvm and dex2oat
-${RUN} "$@" --runtime-option -Xexperimental:lambdas -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:lambdas
+${RUN} "$@" --experimental lambdas
diff --git a/test/960-default-smali/build b/test/960-default-smali/build
index c786687..06692f9 100755
--- a/test/960-default-smali/build
+++ b/test/960-default-smali/build
@@ -18,7 +18,7 @@
 set -e
 
 # Generate the smali Main.smali file or fail
-./util-src/generate_smali.py ./smali
+${ANDROID_BUILD_TOP}/art/test/utils/python/generate_smali_main.py ./smali
 
 if [[ $@ == *"--jvm"* ]]; then
   # Build the Java files if we are running a --jvm test
@@ -29,5 +29,5 @@
 fi
 
 # Build the smali files and make a dex
-${SMALI} -JXmx256m --experimental --api-level 23 --output classes.dex $(find smali -name '*.smali')
+${SMALI} -JXmx256m ${SMALI_ARGS} --output classes.dex $(find smali -name '*.smali')
 zip "$TEST_NAME.jar" classes.dex
diff --git a/test/960-default-smali/info.txt b/test/960-default-smali/info.txt
index eb596e2..9583abb 100644
--- a/test/960-default-smali/info.txt
+++ b/test/960-default-smali/info.txt
@@ -2,15 +2,16 @@
 
 Obviously needs to run under ART or a Java 8 Language runtime and compiler.
 
-When run a Main.smali file will be generated by the util-src/generate_smali.py
-script. If we run with --jvm we will use the tools/extract-embedded-java script to
-turn the smali into equivalent Java using the embedded Java code.
+When run a Main.smali file will be generated by the
+test/utils/python/generate_smali_main.py script. If we run with --jvm we will
+use the tools/extract-embedded-java script to turn the smali into equivalent
+Java using the embedded Java code.
 
 When updating be sure to write the equivalent Java code in comments of the smali
 files.
 
-Care should be taken when updating the generate_smali.py script. It must always
-return equivalent output when run multiple times.
+Care should be taken when updating the generate_smali_main.py script. It must
+always return equivalent output when run multiple times.
 
 To update the test files do the following steps:
     <Add new classes/interfaces>
diff --git a/test/960-default-smali/run b/test/960-default-smali/run
index e378b06..22f6800 100755
--- a/test/960-default-smali/run
+++ b/test/960-default-smali/run
@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if echo $@ | grep -q -- "--jvm"; then
-  ${RUN} "$@"
-else
-  ${RUN} "$@" --runtime-option -Xexperimental:default-methods -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:default-methods
-fi
+${RUN} --experimental default-methods "$@"
diff --git a/test/961-default-iface-resolution-generated/build b/test/961-default-iface-resolution-generated/build
index 707c17e..5eb851f 100755
--- a/test/961-default-iface-resolution-generated/build
+++ b/test/961-default-iface-resolution-generated/build
@@ -40,7 +40,7 @@
 fi
 
 # Build the smali files and make a dex
-${SMALI} -JXmx512m --experimental --api-level 23 --output classes.dex $(find smali -name '*.smali')
+${SMALI} -JXmx512m ${SMALI_ARGS} --output classes.dex $(find smali -name '*.smali')
 zip $TEST_NAME.jar classes.dex
 
 # Reset the ulimit back to its initial value
diff --git a/test/961-default-iface-resolution-generated/run b/test/961-default-iface-resolution-generated/run
index e378b06..22f6800 100755
--- a/test/961-default-iface-resolution-generated/run
+++ b/test/961-default-iface-resolution-generated/run
@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if echo $@ | grep -q -- "--jvm"; then
-  ${RUN} "$@"
-else
-  ${RUN} "$@" --runtime-option -Xexperimental:default-methods -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:default-methods
-fi
+${RUN} --experimental default-methods "$@"
diff --git a/test/962-iface-static/build b/test/962-iface-static/build
index 5ad82f7..06bb3bd 100755
--- a/test/962-iface-static/build
+++ b/test/962-iface-static/build
@@ -26,5 +26,5 @@
 fi
 
 # Build the smali files and make a dex
-${SMALI} -JXmx512m --experimental --api-level 23 --output classes.dex $(find smali -name '*.smali')
+${SMALI} -JXmx512m ${SMALI_ARGS} --output classes.dex $(find smali -name '*.smali')
 zip $TEST_NAME.jar classes.dex
diff --git a/test/962-iface-static/run b/test/962-iface-static/run
index e713708..d37737f 100755
--- a/test/962-iface-static/run
+++ b/test/962-iface-static/run
@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if echo $@ | grep -q -- "--jvm"; then
-  ${RUN} "$@"
-else
-  ${RUN} "$@" --runtime-option -Xexperimental:default-methods -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:default-methods
-fi
+${RUN} --experimental default-methods "$@"
diff --git a/test/963-default-range-smali/build b/test/963-default-range-smali/build
index 5ad82f7..06bb3bd 100755
--- a/test/963-default-range-smali/build
+++ b/test/963-default-range-smali/build
@@ -26,5 +26,5 @@
 fi
 
 # Build the smali files and make a dex
-${SMALI} -JXmx512m --experimental --api-level 23 --output classes.dex $(find smali -name '*.smali')
+${SMALI} -JXmx512m ${SMALI_ARGS} --output classes.dex $(find smali -name '*.smali')
 zip $TEST_NAME.jar classes.dex
diff --git a/test/963-default-range-smali/run b/test/963-default-range-smali/run
index e713708..d37737f 100755
--- a/test/963-default-range-smali/run
+++ b/test/963-default-range-smali/run
@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if echo $@ | grep -q -- "--jvm"; then
-  ${RUN} "$@"
-else
-  ${RUN} "$@" --runtime-option -Xexperimental:default-methods -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:default-methods
-fi
+${RUN} --experimental default-methods "$@"
diff --git a/test/964-default-iface-init-generated/build b/test/964-default-iface-init-generated/build
index deef803..b0fbe4b 100755
--- a/test/964-default-iface-init-generated/build
+++ b/test/964-default-iface-init-generated/build
@@ -38,7 +38,7 @@
 fi
 
 # Build the smali files and make a dex
-${SMALI} -JXmx512m --experimental --api-level 23 --output classes.dex $(find smali -name '*.smali')
+${SMALI} -JXmx512m ${SMALI_ARGS} --output classes.dex $(find smali -name '*.smali')
 zip $TEST_NAME.jar classes.dex
 
 # Reset the ulimit back to its initial value
diff --git a/test/964-default-iface-init-generated/run b/test/964-default-iface-init-generated/run
index e378b06..22f6800 100755
--- a/test/964-default-iface-init-generated/run
+++ b/test/964-default-iface-init-generated/run
@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if echo $@ | grep -q -- "--jvm"; then
-  ${RUN} "$@"
-else
-  ${RUN} "$@" --runtime-option -Xexperimental:default-methods -Xcompiler-option --runtime-arg -Xcompiler-option -Xexperimental:default-methods
-fi
+${RUN} --experimental default-methods "$@"
diff --git a/test/Android.run-test.mk b/test/Android.run-test.mk
index 9c04135..6ce3d94 100644
--- a/test/Android.run-test.mk
+++ b/test/Android.run-test.mk
@@ -441,55 +441,9 @@
 # Known broken tests for the mips32 optimizing compiler backend.
 TEST_ART_BROKEN_OPTIMIZING_MIPS_RUN_TESTS := \
     441-checker-inliner \
-    442-checker-constant-folding \
-    444-checker-nce \
-    445-checker-licm \
-    446-checker-inliner2 \
-    447-checker-inliner3 \
-    449-checker-bce \
-    450-checker-types \
-    455-checker-gvn \
-    458-checker-instruction-simplification \
-    462-checker-inlining-across-dex-files \
-    463-checker-boolean-simplifier \
-    464-checker-inline-sharpen-calls \
-    465-checker-clinit-gvn \
-    468-checker-bool-simplifier-regression \
-    473-checker-inliner-constants \
-    474-checker-boolean-input \
-    476-checker-ctor-memory-barrier \
-    477-checker-bound-type \
-    478-checker-clinit-check-pruning \
-    478-checker-inliner-nested-loop \
-    480-checker-dead-blocks \
-    482-checker-loop-back-edge-use \
-    484-checker-register-hints \
-    485-checker-dce-loop-update \
-    485-checker-dce-switch \
-    486-checker-must-do-null-check \
-    487-checker-inline-calls \
-    488-checker-inline-recursive-calls \
-    490-checker-inline \
-    492-checker-inline-invoke-interface \
-    493-checker-inline-invoke-interface \
-    494-checker-instanceof-tests \
-    495-checker-checkcast-tests \
-    496-checker-inlining-and-class-loader \
-    508-checker-disassembly \
     510-checker-try-catch \
-    517-checker-builder-fallthrough \
     521-checker-array-set-null \
-    522-checker-regression-monitor-exit \
-    523-checker-can-throw-regression \
-    525-checker-arrays-and-fields \
-    526-checker-caller-callee-regs \
-    529-checker-unresolved \
-    530-checker-loops \
-    530-checker-regression-reftype-final \
-    532-checker-nonnull-arrayset \
-    534-checker-bce-deoptimization \
     536-checker-intrinsic-optimization \
-    537-checker-debuggable \
 
 ifeq (mips,$(TARGET_ARCH))
   ifneq (,$(filter optimizing,$(COMPILER_TYPES)))
diff --git a/test/etc/default-build b/test/etc/default-build
index c92402b..4743216 100755
--- a/test/etc/default-build
+++ b/test/etc/default-build
@@ -96,7 +96,7 @@
 
 if [ -d smali ]; then
   # Compile Smali classes
-  ${SMALI} -JXmx256m --experimental --api-level 23 --output smali_classes.dex `find smali -name '*.smali'`
+  ${SMALI} -JXmx256m ${SMALI_ARGS} --output smali_classes.dex `find smali -name '*.smali'`
 
   # Don't bother with dexmerger if we provide our own main function in a smali file.
   if [ ${SKIP_DX_MERGER} = "false" ]; then
diff --git a/test/etc/run-test-jar b/test/etc/run-test-jar
index fbefa07..280b4bc 100755
--- a/test/etc/run-test-jar
+++ b/test/etc/run-test-jar
@@ -18,6 +18,7 @@
 DEBUGGER="n"
 DEV_MODE="n"
 DEX2OAT=""
+EXPERIMENTAL=""
 FALSE_BIN="/system/bin/false"
 FLAGS=""
 GDB=""
@@ -196,6 +197,13 @@
         FLAGS="${FLAGS} -Xcompiler-option --compile-pic"
         COMPILE_FLAGS="${COMPILE_FLAGS} --compile-pic"
         shift
+    elif [ "x$1" = "x--experimental" ]; then
+        if [ "$#" -lt 2 ]; then
+            echo "missing --experimental option" 1>&2
+            exit 1
+        fi
+        EXPERIMENTAL="$EXPERIMENTAL $2"
+        shift 2
     elif expr "x$1" : "x--" >/dev/null 2>&1; then
         echo "unknown $0 option: $1" 1>&2
         exit 1
@@ -204,6 +212,13 @@
     fi
 done
 
+if [ "$USE_JVM" = "n" ]; then
+    for feature in ${EXPERIMENTAL}; do
+        FLAGS="${FLAGS} -Xexperimental:${feature}"
+        COMPILE_FLAGS="${COMPILE_FLAGS} --runtime-arg -Xexperimental:${feature}"
+    done
+fi
+
 if [ "x$1" = "x" ] ; then
   MAIN="Main"
 else
diff --git a/test/run-test b/test/run-test
index 293779f..5a43fb0 100755
--- a/test/run-test
+++ b/test/run-test
@@ -46,6 +46,7 @@
 export DEX_LOCATION=/data/run-test/${test_dir}
 export NEED_DEX="true"
 export USE_JACK="false"
+export SMALI_ARGS="--experimental --api-level 23"
 
 # If dx was not set by the environment variable, assume it is in the path.
 if [ -z "$DX" ]; then
diff --git a/test/960-default-smali/util-src/generate_smali.py b/test/utils/python/generate_smali_main.py
similarity index 98%
rename from test/960-default-smali/util-src/generate_smali.py
rename to test/utils/python/generate_smali_main.py
index b2bf1f0..d796d31 100755
--- a/test/960-default-smali/util-src/generate_smali.py
+++ b/test/utils/python/generate_smali_main.py
@@ -15,7 +15,7 @@
 # limitations under the License.
 
 """
-Generate Smali Main file for test 960
+Generate Smali Main file from a classes.xml file.
 """
 
 import os
diff --git a/tools/libcore_failures.txt b/tools/libcore_failures.txt
index 81ea79a..9a8b462 100644
--- a/tools/libcore_failures.txt
+++ b/tools/libcore_failures.txt
@@ -159,13 +159,6 @@
   bug: 22786792
 },
 {
-  description: "Formatting failures",
-  result: EXEC_FAILED,
-  names: ["libcore.java.text.NumberFormatTest#test_currencyFromLocale",
-          "libcore.java.text.NumberFormatTest#test_currencyWithPatternDigits"],
-  bug: 25136848
-},
-{
   description: "Lack of IPv6 on some buildbot slaves",
   result: EXEC_FAILED,
   names: ["libcore.io.OsTest#test_byteBufferPositions_sendto_recvfrom_af_inet6",
diff --git a/tools/run-jdwp-tests.sh b/tools/run-jdwp-tests.sh
index edec362..9aed271 100755
--- a/tools/run-jdwp-tests.sh
+++ b/tools/run-jdwp-tests.sh
@@ -88,7 +88,8 @@
   fi
 done
 
-vm_args="--vm-arg $image"
+vm_args="--vm-arg $image --vm-arg -Xusejit:true"
+debuggee_args="$debuggee_args -Xusejit:true"
 if [[ $debug == "yes" ]]; then
   art="$art -d"
   art_debugee="$art_debugee -d"
diff --git a/tools/run-libcore-tests.sh b/tools/run-libcore-tests.sh
index 80f7a37..67a7983 100755
--- a/tools/run-libcore-tests.sh
+++ b/tools/run-libcore-tests.sh
@@ -102,4 +102,4 @@
 # Run the tests using vogar.
 echo "Running tests for the following test packages:"
 echo ${working_packages[@]} | tr " " "\n"
-vogar $vogar_args --expectations art/tools/libcore_failures.txt --classpath $jsr166_test_jar --classpath $test_jar ${working_packages[@]}
+vogar $vogar_args --vm-arg -Xusejit:true --expectations art/tools/libcore_failures.txt --classpath $jsr166_test_jar --classpath $test_jar ${working_packages[@]}