diff --git a/adbconnection/adbconnection.cc b/adbconnection/adbconnection.cc
index cf35914..b6d6600 100644
--- a/adbconnection/adbconnection.cc
+++ b/adbconnection/adbconnection.cc
@@ -476,7 +476,6 @@
   int rc = TEMP_FAILURE_RETRY(recvmsg(control_sock_, &msg, 0));
 
   if (rc <= 0) {
-    PLOG(WARNING) << "Receiving file descriptor from ADB failed (socket " << control_sock_ << ")";
     return android::base::unique_fd(-1);
   } else {
     VLOG(jdwp) << "Fds have been received from ADB!";
@@ -624,7 +623,6 @@
           android::base::unique_fd new_fd(ReadFdFromAdb());
           if (new_fd == -1) {
             // Something went wrong. We need to retry getting the control socket.
-            PLOG(ERROR) << "Something went wrong getting fds from adb. Retry!";
             control_sock_.reset();
             break;
           } else if (adb_connection_socket_ != -1) {
diff --git a/build/Android.common_path.mk b/build/Android.common_path.mk
index 03e68ae..1880726 100644
--- a/build/Android.common_path.mk
+++ b/build/Android.common_path.mk
@@ -73,8 +73,21 @@
 HOST_CORE_IMG_LOCATION := $(HOST_OUT_JAVA_LIBRARIES)/core.art
 TARGET_CORE_IMG_LOCATION := $(ART_TARGET_TEST_OUT)/core.art
 
-# Jar files for core.art.
-TEST_CORE_JARS := core-oj core-libart core-simple conscrypt okhttp bouncycastle
+# Modules to compile for core.art.
+CORE_IMG_JARS := core-oj core-libart core-simple okhttp bouncycastle
+HOST_CORE_IMG_JARS   := $(addsuffix -hostdex,$(CORE_IMG_JARS))
+TARGET_CORE_IMG_JARS := $(addsuffix -testdex,$(CORE_IMG_JARS))
+HOST_CORE_IMG_DEX_LOCATIONS   := $(foreach jar,$(HOST_CORE_IMG_JARS),  $(HOST_OUT_JAVA_LIBRARIES)/$(jar).jar)
+ifeq ($(ART_TEST_ANDROID_ROOT),)
+TARGET_CORE_IMG_DEX_LOCATIONS := $(foreach jar,$(TARGET_CORE_IMG_JARS),/$(DEXPREOPT_BOOT_JAR_DIR)/$(jar).jar)
+else
+TARGET_CORE_IMG_DEX_LOCATIONS := $(foreach jar,$(TARGET_CORE_IMG_JARS),$(ART_TEST_ANDROID_ROOT)/$(jar).jar)
+endif
+HOST_CORE_IMG_DEX_FILES   := $(foreach jar,$(HOST_CORE_IMG_JARS),  $(call intermediates-dir-for,JAVA_LIBRARIES,$(jar),t,COMMON)/javalib.jar)
+TARGET_CORE_IMG_DEX_FILES := $(foreach jar,$(TARGET_CORE_IMG_JARS),$(call intermediates-dir-for,JAVA_LIBRARIES,$(jar), ,COMMON)/javalib.jar)
+
+# Jar files for the boot class path for testing. Must start with CORE_IMG_JARS.
+TEST_CORE_JARS := $(CORE_IMG_JARS) conscrypt
 HOST_TEST_CORE_JARS   := $(addsuffix -hostdex,$(TEST_CORE_JARS))
 TARGET_TEST_CORE_JARS := $(addsuffix -testdex,$(TEST_CORE_JARS))
 HOST_CORE_DEX_LOCATIONS   := $(foreach jar,$(HOST_TEST_CORE_JARS),  $(HOST_OUT_JAVA_LIBRARIES)/$(jar).jar)
@@ -83,7 +96,6 @@
 else
 TARGET_CORE_DEX_LOCATIONS := $(foreach jar,$(TARGET_TEST_CORE_JARS),$(ART_TEST_ANDROID_ROOT)/framework/$(jar).jar)
 endif
-
 HOST_CORE_DEX_FILES   := $(foreach jar,$(HOST_TEST_CORE_JARS),  $(call intermediates-dir-for,JAVA_LIBRARIES,$(jar),t,COMMON)/javalib.jar)
 TARGET_CORE_DEX_FILES := $(foreach jar,$(TARGET_TEST_CORE_JARS),$(call intermediates-dir-for,JAVA_LIBRARIES,$(jar), ,COMMON)/javalib.jar)
 
diff --git a/build/Android.gtest.mk b/build/Android.gtest.mk
index 6885946..1bcca7c 100644
--- a/build/Android.gtest.mk
+++ b/build/Android.gtest.mk
@@ -452,12 +452,7 @@
     $$(gtest_exe) \
     $$($(3)TARGET_OUT_SHARED_LIBRARIES)/libjavacore.so \
     $$($(3)TARGET_OUT_SHARED_LIBRARIES)/libopenjdkd.so \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/core-libart-testdex.jar \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/core-oj-testdex.jar \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/core-simple-testdex.jar \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/conscrypt-testdex.jar \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/okhttp-testdex.jar \
-    $$(TARGET_OUT_JAVA_LIBRARIES)/bouncycastle-testdex.jar
+    $$(foreach jar,$$(TARGET_TEST_CORE_JARS),$$(TARGET_OUT_JAVA_LIBRARIES)/$$(jar).jar)
 
   ART_TEST_TARGET_GTEST_DEPENDENCIES += $$(gtest_deps)
 
@@ -515,7 +510,8 @@
     $$($(3)ART_HOST_OUT_SHARED_LIBRARIES)/libopenjdkd$$(ART_HOST_SHLIB_EXTENSION) \
     $$(gtest_exe) \
     $$(ART_GTEST_$(1)_HOST_DEPS) \
-    $(foreach file,$(ART_GTEST_$(1)_DEX_DEPS),$(ART_TEST_HOST_GTEST_$(file)_DEX))
+    $(foreach file,$(ART_GTEST_$(1)_DEX_DEPS),$(ART_TEST_HOST_GTEST_$(file)_DEX)) \
+    $(HOST_OUT_EXECUTABLES)/timeout_dumper
 
   ART_TEST_HOST_GTEST_DEPENDENCIES += $$(gtest_deps)
 
@@ -528,7 +524,9 @@
 $$(gtest_output): NAME := $$(gtest_rule)
 ifeq (,$(SANITIZE_HOST))
 $$(gtest_output): $$(gtest_exe) $$(gtest_deps)
-	$(hide) ($$(call ART_TEST_SKIP,$$(NAME)) && $$< --gtest_output=xml:$$@ && \
+	$(hide) ($$(call ART_TEST_SKIP,$$(NAME)) && \
+		timeout -k 120s -s SIGRTMIN+2 2400s $(HOST_OUT_EXECUTABLES)/timeout_dumper \
+			$$< --gtest_output=xml:$$@ && \
 		$$(call ART_TEST_PASSED,$$(NAME))) || $$(call ART_TEST_FAILED,$$(NAME))
 else
 # Note: envsetup currently exports ASAN_OPTIONS=detect_leaks=0 to suppress leak detection, as some
@@ -540,7 +538,9 @@
 # under ASAN.
 $$(gtest_output): $$(gtest_exe) $$(gtest_deps)
 	$(hide) ($$(call ART_TEST_SKIP,$$(NAME)) && set -o pipefail && \
-		ASAN_OPTIONS=detect_leaks=1 $$< --gtest_output=xml:$$@ 2>&1 | tee $$<.tmp.out >&2 && \
+		ASAN_OPTIONS=detect_leaks=1 timeout -k 120s -s SIGRTMIN+2 3600s \
+			$(HOST_OUT_EXECUTABLES)/timeout_dumper \
+				$$< --gtest_output=xml:$$@ 2>&1 | tee $$<.tmp.out >&2 && \
 		{ $$(call ART_TEST_PASSED,$$(NAME)) ; rm $$<.tmp.out ; }) || \
 		( grep -q AddressSanitizer $$<.tmp.out && export ANDROID_BUILD_TOP=`pwd` && \
 			{ echo "ABI: 'x86_64'" | cat - $$<.tmp.out | development/scripts/stack | tail -n 3000 ; } ; \
diff --git a/build/Android.oat.mk b/build/Android.oat.mk
index e2adac1..2ad1143 100644
--- a/build/Android.oat.mk
+++ b/build/Android.oat.mk
@@ -39,8 +39,6 @@
 # Use dex2oat debug version for better error reporting
 # $(1): compiler - optimizing, interpreter or interp-ac (interpreter-access-checks).
 # $(2): 2ND_ or undefined, 2ND_ for 32-bit host builds.
-# NB depending on HOST_CORE_DEX_LOCATIONS so we are sure to have the dex files in frameworks for
-# run-test --no-image
 define create-core-oat-host-rules
   core_compile_options :=
   core_image_name :=
@@ -80,13 +78,15 @@
 $$(core_image_name): PRIVATE_CORE_COMPILE_OPTIONS := $$(core_compile_options)
 $$(core_image_name): PRIVATE_CORE_IMG_NAME := $$(core_image_name)
 $$(core_image_name): PRIVATE_CORE_OAT_NAME := $$(core_oat_name)
-$$(core_image_name): $$(HOST_CORE_DEX_LOCATIONS) $$(core_dex2oat_dependency)
+$$(core_image_name): $$(HOST_CORE_IMG_DEX_LOCATIONS) $$(core_dex2oat_dependency)
 	@echo "host dex2oat: $$@"
 	@mkdir -p $$(dir $$@)
 	$$(hide) ANDROID_LOG_TAGS="*:e" $$(DEX2OAT) --runtime-arg -Xms$(DEX2OAT_IMAGE_XMS) \
 	  --runtime-arg -Xmx$(DEX2OAT_IMAGE_XMX) \
-	  --image-classes=$$(PRELOADED_CLASSES) $$(addprefix --dex-file=,$$(HOST_CORE_DEX_FILES)) \
-	  $$(addprefix --dex-location=,$$(HOST_CORE_DEX_LOCATIONS)) --oat-file=$$(PRIVATE_CORE_OAT_NAME) \
+	  --image-classes=$$(PRELOADED_CLASSES) \
+	  $$(addprefix --dex-file=,$$(HOST_CORE_IMG_DEX_FILES)) \
+	  $$(addprefix --dex-location=,$$(HOST_CORE_IMG_DEX_LOCATIONS)) \
+	  --oat-file=$$(PRIVATE_CORE_OAT_NAME) \
 	  --oat-location=$$(PRIVATE_CORE_OAT_NAME) --image=$$(PRIVATE_CORE_IMG_NAME) \
 	  --base=$$(LIBART_IMG_HOST_BASE_ADDRESS) --instruction-set=$$($(2)ART_HOST_ARCH) \
 	  $$(LOCAL_$(2)DEX2OAT_HOST_INSTRUCTION_SET_FEATURES_OPTION) \
@@ -169,13 +169,15 @@
 $$(core_image_name): PRIVATE_CORE_COMPILE_OPTIONS := $$(core_compile_options)
 $$(core_image_name): PRIVATE_CORE_IMG_NAME := $$(core_image_name)
 $$(core_image_name): PRIVATE_CORE_OAT_NAME := $$(core_oat_name)
-$$(core_image_name): $$(TARGET_CORE_DEX_FILES) $$(core_dex2oat_dependency)
+$$(core_image_name): $$(TARGET_CORE_IMG_DEX_FILES) $$(core_dex2oat_dependency)
 	@echo "target dex2oat: $$@"
 	@mkdir -p $$(dir $$@)
 	$$(hide) $$(DEX2OAT) --runtime-arg -Xms$(DEX2OAT_IMAGE_XMS) \
 	  --runtime-arg -Xmx$(DEX2OAT_IMAGE_XMX) \
-	  --image-classes=$$(PRELOADED_CLASSES) $$(addprefix --dex-file=,$$(TARGET_CORE_DEX_FILES)) \
-	  $$(addprefix --dex-location=,$$(TARGET_CORE_DEX_LOCATIONS)) --oat-file=$$(PRIVATE_CORE_OAT_NAME) \
+	  --image-classes=$$(PRELOADED_CLASSES) \
+	  $$(addprefix --dex-file=,$$(TARGET_CORE_IMG_DEX_FILES)) \
+	  $$(addprefix --dex-location=,$$(TARGET_CORE_IMG_DEX_LOCATIONS)) \
+	  --oat-file=$$(PRIVATE_CORE_OAT_NAME) \
 	  --oat-location=$$(PRIVATE_CORE_OAT_NAME) --image=$$(PRIVATE_CORE_IMG_NAME) \
 	  --base=$$(LIBART_IMG_TARGET_BASE_ADDRESS) --instruction-set=$$($(2)TARGET_ARCH) \
 	  --instruction-set-variant=$$($(2)DEX2OAT_TARGET_CPU_VARIANT) \
diff --git a/build/apex/Android.bp b/build/apex/Android.bp
index f2e12f6..0ec0a15 100644
--- a/build/apex/Android.bp
+++ b/build/apex/Android.bp
@@ -19,6 +19,11 @@
     "libopenjdkjvmti",
     "libadbconnection",
 ]
+bionic_native_shared_libs = [
+    "libc",
+    "libm",
+    "libdl",
+]
 // - Fake library that avoids namespace issues and gives some warnings for nosy apps.
 art_runtime_fake_native_shared_libs = [
      // FIXME: Does not work as-is, because `libart_fake` is defined in libart_fake/Android.mk,
@@ -102,7 +107,8 @@
     compile_multilib: "both",
     manifest: "manifest.json",
     native_shared_libs: art_runtime_base_native_shared_libs
-        + art_runtime_fake_native_shared_libs,
+        + art_runtime_fake_native_shared_libs
+        + bionic_native_shared_libs,
     multilib: {
         both: {
             // TODO: Add logic to create a `dalvikvm` symlink to `dalvikvm32` or `dalvikvm64`
@@ -130,7 +136,8 @@
     manifest: "manifest.json",
     native_shared_libs: art_runtime_base_native_shared_libs
         + art_runtime_fake_native_shared_libs
-        + art_runtime_debug_native_shared_libs,
+        + art_runtime_debug_native_shared_libs
+        + bionic_native_shared_libs,
     multilib: {
         both: {
             // TODO: Add logic to create a `dalvikvm` symlink to `dalvikvm32` or `dalvikvm64`
diff --git a/build/apex/runtests.sh b/build/apex/runtests.sh
index 924c44b..b5e8d8b 100755
--- a/build/apex/runtests.sh
+++ b/build/apex/runtests.sh
@@ -33,11 +33,7 @@
 
    sudo apt-get install libguestfs-tools
 "
-which tree > /dev/null || die "This script requires the 'tree' tool.
-On Debian-based systems, this can be installed with:
 
-   sudo apt-get install tree
-"
 [[ -n "$ANDROID_PRODUCT_OUT" ]] \
   || die "You need to source and lunch before you can use this script."
 
@@ -46,6 +42,7 @@
 
 build_apex_p=true
 list_image_files_p=false
+print_image_tree_p=false
 
 function usage {
   cat <<EOF
@@ -53,7 +50,8 @@
 Build (optional) and run tests on Android Runtime APEX package (on host).
 
   -s, --skip-build    skip the build step
-  -l, --list-files    list the contents of the ext4 image
+  -l, --list-files    list the contents of the ext4 image using `find`
+  -t, --print-tree    list the contents of the ext4 image using `tree`
   -h, --help          display this help and exit
 
 EOF
@@ -64,6 +62,7 @@
   case "$1" in
     (-s|--skip-build) build_apex_p=false;;
     (-l|--list-files) list_image_files_p=true;;
+    (-t|--print-tree) print_image_tree_p=true;;
     (-h|--help) usage;;
     (*) die "Unknown option: '$1'
 Try '$0 --help' for more information.";;
@@ -71,6 +70,14 @@
   shift
 done
 
+if $print_image_tree_p; then
+  which tree >/dev/null || die "This script requires the 'tree' tool.
+On Debian-based systems, this can be installed with:
+
+   sudo apt-get install tree
+"
+fi
+
 
 # build_apex APEX_MODULE
 # ----------------------
@@ -82,6 +89,24 @@
   fi
 }
 
+# maybe_list_apex_contents MOUNT_POINT
+# ------------------------------------
+# If any listing/printing option was used, honor them and display the contents
+# of the APEX payload at MOUNT_POINT.
+function maybe_list_apex_contents {
+  local mount_point=$1
+
+  # List the contents of the mounted image using `find` (optional).
+  if $list_image_files_p; then
+    say "Listing image files" && find "$mount_point"
+  fi
+
+  # List the contents of the mounted image using `tree` (optional).
+  if $print_image_tree_p; then
+    say "Printing image tree" && ls -ld "$mount_point" && tree -aph --du "$mount_point"
+  fi
+}
+
 function check_binary {
   [[ -x "$mount_point/bin/$1" ]] || die "Cannot find binary '$1' in mounted image"
 }
@@ -218,10 +243,6 @@
 
   # Mount the image from the Android Runtime APEX.
   guestmount -a "$image_file" -m "$partition" "$mount_point"
-
-  # List the contents of the mounted image (optional).
-  $list_image_files_p \
-    && say "Listing image files" && ls -ld "$mount_point" && tree -ap "$mount_point"
 }
 
 # Testing release APEX package (com.android.runtime.release).
@@ -229,6 +250,8 @@
 
 apex_module="com.android.runtime.release"
 
+say "Processing APEX package $apex_module"
+
 work_dir=$(mktemp -d)
 mount_point="$work_dir/image"
 
@@ -240,6 +263,9 @@
 # Set up APEX package.
 setup_target_apex "$apex_module" "$mount_point"
 
+# List the contents of the APEX image (optional).
+maybe_list_apex_contents "$mount_point"
+
 # Run tests on APEX package.
 say "Checking APEX package $apex_module"
 check_release_contents
@@ -249,12 +275,15 @@
 cleanup_target
 
 say "$apex_module tests passed"
+echo
 
 # Testing debug APEX package (com.android.runtime.debug).
 # -------------------------------------------------------
 
 apex_module="com.android.runtime.debug"
 
+say "Processing APEX package $apex_module"
+
 work_dir=$(mktemp -d)
 mount_point="$work_dir/image"
 
@@ -266,6 +295,9 @@
 # Set up APEX package.
 setup_target_apex "$apex_module" "$mount_point"
 
+# List the contents of the APEX image (optional).
+maybe_list_apex_contents "$mount_point"
+
 # Run tests on APEX package.
 say "Checking APEX package $apex_module"
 check_release_contents
@@ -279,6 +311,7 @@
 cleanup_target
 
 say "$apex_module tests passed"
+echo
 
 
 # Testing host APEX package (com.android.runtime.host).
@@ -319,6 +352,8 @@
 
 apex_module="com.android.runtime.host"
 
+say "Processing APEX package $apex_module"
+
 work_dir=$(mktemp -d)
 mount_point="$work_dir/zip"
 
@@ -330,6 +365,9 @@
 # Set up APEX package.
 setup_host_apex "$apex_module" "$mount_point"
 
+# List the contents of the APEX image (optional).
+maybe_list_apex_contents "$mount_point"
+
 # Run tests on APEX package.
 say "Checking APEX package $apex_module"
 check_release_contents
diff --git a/compiler/driver/compiler_driver.cc b/compiler/driver/compiler_driver.cc
index 18f7105..0039be0 100644
--- a/compiler/driver/compiler_driver.cc
+++ b/compiler/driver/compiler_driver.cc
@@ -1742,6 +1742,9 @@
     if (&cls->GetDexFile() == &accessor.GetDexFile()) {
       ObjectLock<mirror::Class> lock(self, cls);
       mirror::Class::SetStatus(cls, status, self);
+      if (status >= ClassStatus::kVerified) {
+        cls->SetVerificationAttempted();
+      }
     }
   } else {
     DCHECK(self->IsExceptionPending());
diff --git a/compiler/jit/jit_compiler.cc b/compiler/jit/jit_compiler.cc
index 9b8bb3e..e57bbfa 100644
--- a/compiler/jit/jit_compiler.cc
+++ b/compiler/jit/jit_compiler.cc
@@ -126,11 +126,11 @@
 }
 
 extern "C" bool jit_compile_method(
-    void* handle, ArtMethod* method, Thread* self, bool osr)
+    void* handle, ArtMethod* method, Thread* self, bool baseline, bool osr)
     REQUIRES_SHARED(Locks::mutator_lock_) {
   auto* jit_compiler = reinterpret_cast<JitCompiler*>(handle);
   DCHECK(jit_compiler != nullptr);
-  return jit_compiler->CompileMethod(self, method, osr);
+  return jit_compiler->CompileMethod(self, method, baseline, osr);
 }
 
 extern "C" void jit_types_loaded(void* handle, mirror::Class** types, size_t count)
@@ -181,7 +181,7 @@
   }
 }
 
-bool JitCompiler::CompileMethod(Thread* self, ArtMethod* method, bool osr) {
+bool JitCompiler::CompileMethod(Thread* self, ArtMethod* method, bool baseline, bool osr) {
   SCOPED_TRACE << "JIT compiling " << method->PrettyMethod();
 
   DCHECK(!method->IsProxyMethod());
@@ -198,7 +198,7 @@
     TimingLogger::ScopedTiming t2("Compiling", &logger);
     JitCodeCache* const code_cache = runtime->GetJit()->GetCodeCache();
     success = compiler_driver_->GetCompiler()->JitCompile(
-        self, code_cache, method, /* baseline= */ false, osr, jit_logger_.get());
+        self, code_cache, method, baseline, osr, jit_logger_.get());
   }
 
   // Trim maps to reduce memory usage.
diff --git a/compiler/jit/jit_compiler.h b/compiler/jit/jit_compiler.h
index d201611..29d2761 100644
--- a/compiler/jit/jit_compiler.h
+++ b/compiler/jit/jit_compiler.h
@@ -37,7 +37,7 @@
   virtual ~JitCompiler();
 
   // Compilation entrypoint. Returns whether the compilation succeeded.
-  bool CompileMethod(Thread* self, ArtMethod* method, bool osr)
+  bool CompileMethod(Thread* self, ArtMethod* method, bool baseline, bool osr)
       REQUIRES_SHARED(Locks::mutator_lock_);
 
   const CompilerOptions& GetCompilerOptions() const {
diff --git a/compiler/jni/jni_compiler_test.cc b/compiler/jni/jni_compiler_test.cc
index 92b9543..bd4304c 100644
--- a/compiler/jni/jni_compiler_test.cc
+++ b/compiler/jni/jni_compiler_test.cc
@@ -1300,15 +1300,15 @@
   EXPECT_TRUE(env->IsInstanceOf(JniCompilerTest::jobj_, klass));
   EXPECT_TRUE(env->IsSameObject(JniCompilerTest::jobj_, obj1));
   EXPECT_TRUE(env->IsSameObject(JniCompilerTest::jobj_, obj2));
-  EXPECT_EQ(0x12345678ABCDEF88ll, val1);
-  EXPECT_EQ(0x7FEDCBA987654321ll, val2);
+  EXPECT_EQ(0x12345678ABCDEF88LL, val1);
+  EXPECT_EQ(0x7FEDCBA987654321LL, val2);
   return 42;
 }
 
 void JniCompilerTest::GetTextImpl() {
   SetUpForTest(true, "getText", "(JLjava/lang/Object;JLjava/lang/Object;)I",
                CURRENT_JNI_WRAPPER(my_gettext));
-  jint result = env_->CallStaticIntMethod(jklass_, jmethod_, 0x12345678ABCDEF88ll, jobj_,
+  jint result = env_->CallStaticIntMethod(jklass_, jmethod_, 0x12345678ABCDEF88LL, jobj_,
                                           INT64_C(0x7FEDCBA987654321), jobj_);
   EXPECT_EQ(result, 42);
 }
diff --git a/compiler/optimizing/intrinsics.h b/compiler/optimizing/intrinsics.h
index 5bd11226..50b13c8 100644
--- a/compiler/optimizing/intrinsics.h
+++ b/compiler/optimizing/intrinsics.h
@@ -243,7 +243,8 @@
 // compilation.
 #define UNREACHABLE_INTRINSIC(Arch, Name)                                \
 void IntrinsicLocationsBuilder ## Arch::Visit ## Name(HInvoke* invoke) { \
-  if (!codegen_->GetCompilerOptions().IsBaseline()) {                    \
+  if (Runtime::Current()->IsAotCompiler() &&                             \
+      !codegen_->GetCompilerOptions().IsBaseline()) {                    \
     LOG(FATAL) << "Unreachable: intrinsic " << invoke->GetIntrinsic()    \
                << " should have been converted to HIR";                  \
   }                                                                      \
diff --git a/dex2oat/dex2oat.cc b/dex2oat/dex2oat.cc
index a5bba9b..872fab3 100644
--- a/dex2oat/dex2oat.cc
+++ b/dex2oat/dex2oat.cc
@@ -624,7 +624,6 @@
   explicit Dex2Oat(TimingLogger* timings) :
       compiler_kind_(Compiler::kOptimizing),
       // Take the default set of instruction features from the build.
-      boot_image_checksum_(0),
       key_value_store_(nullptr),
       verification_results_(nullptr),
       runtime_(nullptr),
@@ -1437,17 +1436,22 @@
 
     if (!IsBootImage()) {
       // When compiling an app, create the runtime early to retrieve
-      // the image location key needed for the oat header.
+      // the boot image checksums needed for the oat header.
       if (!CreateRuntime(std::move(runtime_options))) {
         return dex2oat::ReturnCode::kCreateRuntime;
       }
 
       if (CompilerFilter::DependsOnImageChecksum(compiler_options_->GetCompilerFilter())) {
         TimingLogger::ScopedTiming t3("Loading image checksum", timings_);
-        std::vector<ImageSpace*> image_spaces = Runtime::Current()->GetHeap()->GetBootImageSpaces();
-        boot_image_checksum_ = image_spaces[0]->GetImageHeader().GetImageChecksum();
-      } else {
-        boot_image_checksum_ = 0u;
+        Runtime* runtime = Runtime::Current();
+        key_value_store_->Put(OatHeader::kBootClassPathKey,
+                              android::base::Join(runtime->GetBootClassPathLocations(), ':'));
+        std::vector<ImageSpace*> image_spaces = runtime->GetHeap()->GetBootImageSpaces();
+        const std::vector<const DexFile*>& bcp_dex_files =
+            runtime->GetClassLinker()->GetBootClassPath();
+        key_value_store_->Put(
+            OatHeader::kBootClassPathChecksumsKey,
+            gc::space::ImageSpace::GetBootClassPathChecksums(image_spaces, bcp_dex_files));
       }
 
       // Open dex files for class path.
@@ -2015,7 +2019,7 @@
           elf_writer->EndDataBimgRelRo(data_bimg_rel_ro);
         }
 
-        if (!oat_writer->WriteHeader(elf_writer->GetStream(), boot_image_checksum_)) {
+        if (!oat_writer->WriteHeader(elf_writer->GetStream())) {
           LOG(ERROR) << "Failed to write oat header to the ELF file " << oat_file->GetPath();
           return false;
         }
@@ -2646,7 +2650,6 @@
   std::unique_ptr<CompilerOptions> compiler_options_;
   Compiler::Kind compiler_kind_;
 
-  uint32_t boot_image_checksum_;
   std::unique_ptr<SafeMap<std::string, std::string> > key_value_store_;
 
   std::unique_ptr<VerificationResults> verification_results_;
diff --git a/dex2oat/linker/image_test.h b/dex2oat/linker/image_test.h
index bd8cf5a..fa0a3d4 100644
--- a/dex2oat/linker/image_test.h
+++ b/dex2oat/linker/image_test.h
@@ -326,8 +326,7 @@
           elf_writer->EndDataBimgRelRo(data_bimg_rel_ro);
         }
 
-        bool header_ok = oat_writer->WriteHeader(elf_writer->GetStream(),
-                                                 /*boot_image_checksum=*/ 0u);
+        bool header_ok = oat_writer->WriteHeader(elf_writer->GetStream());
         ASSERT_TRUE(header_ok);
 
         writer->UpdateOatFileHeader(i, oat_writer->GetOatHeader());
diff --git a/dex2oat/linker/image_writer.cc b/dex2oat/linker/image_writer.cc
index 61d105f..e4e4b13 100644
--- a/dex2oat/linker/image_writer.cc
+++ b/dex2oat/linker/image_writer.cc
@@ -32,6 +32,7 @@
 #include "base/enums.h"
 #include "base/globals.h"
 #include "base/logging.h"  // For VLOG.
+#include "base/stl_util.h"
 #include "base/unix_file/fd_file.h"
 #include "class_linker-inl.h"
 #include "class_root.h"
@@ -153,6 +154,26 @@
       : nullptr;
 }
 
+bool ImageWriter::IsImageObject(ObjPtr<mirror::Object> obj) const {
+  // For boot image, we keep all objects remaining after the GC in PrepareImageAddressSpace().
+  if (compiler_options_.IsBootImage()) {
+    return true;
+  }
+  // Objects already in the boot image do not belong to the image being written.
+  if (IsInBootImage(obj.Ptr())) {
+    return false;
+  }
+  // DexCaches for the boot class path components that are not a part of the boot image
+  // cannot be garbage collected in PrepareImageAddressSpace() but we do not want to
+  // include them in the app image. So make sure we include only the app DexCaches.
+  if (obj->IsDexCache() &&
+      !ContainsElement(compiler_options_.GetDexFilesForOatFile(),
+                       obj->AsDexCache()->GetDexFile())) {
+    return false;
+  }
+  return true;
+}
+
 // Return true if an object is already in an image space.
 bool ImageWriter::IsInBootImage(const void* obj) const {
   gc::Heap* const heap = Runtime::Current()->GetHeap();
@@ -437,7 +458,7 @@
    */
   heap->VisitObjects([this, &visitor](ObjPtr<mirror::Object> object)
       REQUIRES_SHARED(Locks::mutator_lock_) {
-    if (!IsInBootImage(object.Ptr())) {
+    if (IsImageObject(object)) {
       visitor.SetObject(object);
 
       if (object->IsDexCache()) {
@@ -680,7 +701,7 @@
     ObjPtr<mirror::ClassLoader> class_loader = GetAppClassLoader();
     std::vector<ObjPtr<mirror::DexCache>> dex_caches = FindDexCaches(self);
     for (ObjPtr<mirror::DexCache> dex_cache : dex_caches) {
-      if (IsInBootImage(dex_cache.Ptr())) {
+      if (!IsImageObject(dex_cache)) {
         continue;  // Boot image DexCache is not written to the app image.
       }
       PreloadDexCache(dex_cache, class_loader);
@@ -989,7 +1010,7 @@
   for (const ClassLinker::DexCacheData& data : class_linker->GetDexCachesData()) {
     ObjPtr<mirror::DexCache> dex_cache =
         ObjPtr<mirror::DexCache>::DownCast(self->DecodeJObject(data.weak_root));
-    if (dex_cache == nullptr || IsInBootImage(dex_cache.Ptr())) {
+    if (dex_cache == nullptr || !IsImageObject(dex_cache)) {
       continue;
     }
     const DexFile* dex_file = dex_cache->GetDexFile();
@@ -1758,7 +1779,8 @@
   for (ObjPtr<mirror::DexCache> dex_cache : dex_caches) {
     // Pass the class loader associated with the DexCache. This can either be
     // the app's `class_loader` or `nullptr` if boot class loader.
-    PruneDexCache(dex_cache, IsInBootImage(dex_cache.Ptr()) ? nullptr : GetAppClassLoader());
+    bool is_app_image_dex_cache = compiler_options_.IsAppImage() && IsImageObject(dex_cache);
+    PruneDexCache(dex_cache, is_app_image_dex_cache ? GetAppClassLoader() : nullptr);
   }
 
   // Drop the array class cache in the ClassLinker, as these are roots holding those classes live.
@@ -1856,7 +1878,7 @@
         continue;
       }
       const DexFile* dex_file = dex_cache->GetDexFile();
-      if (!IsInBootImage(dex_cache.Ptr())) {
+      if (IsImageObject(dex_cache)) {
         dex_cache_count += image_dex_files.find(dex_file) != image_dex_files.end() ? 1u : 0u;
       }
     }
@@ -1875,7 +1897,7 @@
         continue;
       }
       const DexFile* dex_file = dex_cache->GetDexFile();
-      if (!IsInBootImage(dex_cache.Ptr())) {
+      if (IsImageObject(dex_cache)) {
         non_image_dex_caches += image_dex_files.find(dex_file) != image_dex_files.end() ? 1u : 0u;
       }
     }
@@ -1889,7 +1911,7 @@
         continue;
       }
       const DexFile* dex_file = dex_cache->GetDexFile();
-      if (!IsInBootImage(dex_cache.Ptr()) &&
+      if (IsImageObject(dex_cache) &&
           image_dex_files.find(dex_file) != image_dex_files.end()) {
         dex_caches->Set<false>(i, dex_cache.Ptr());
         ++i;
@@ -1942,7 +1964,7 @@
 mirror::Object* ImageWriter::TryAssignBinSlot(WorkStack& work_stack,
                                               mirror::Object* obj,
                                               size_t oat_index) {
-  if (obj == nullptr || IsInBootImage(obj)) {
+  if (obj == nullptr || !IsImageObject(obj)) {
     // Object is null or already in the image, there is no work to do.
     return obj;
   }
@@ -2373,7 +2395,7 @@
   {
     auto ensure_bin_slots_assigned = [&](mirror::Object* obj)
         REQUIRES_SHARED(Locks::mutator_lock_) {
-      if (!Runtime::Current()->GetHeap()->ObjectIsInBootImageSpace(obj)) {
+      if (IsImageObject(obj)) {
         CHECK(IsImageBinSlotAssigned(obj)) << mirror::Object::PrettyTypeOf(obj) << " " << obj;
       }
     };
@@ -2444,7 +2466,7 @@
   {
     auto unbin_objects_into_offset = [&](mirror::Object* obj)
         REQUIRES_SHARED(Locks::mutator_lock_) {
-      if (!IsInBootImage(obj)) {
+      if (IsImageObject(obj)) {
         UnbinObjectsIntoOffset(obj);
       }
     };
@@ -2909,7 +2931,7 @@
 }
 
 void ImageWriter::CopyAndFixupObject(Object* obj) {
-  if (IsInBootImage(obj)) {
+  if (!IsImageObject(obj)) {
     return;
   }
   size_t offset = GetImageOffset(obj);
diff --git a/dex2oat/linker/image_writer.h b/dex2oat/linker/image_writer.h
index 33bacf8..b680265 100644
--- a/dex2oat/linker/image_writer.h
+++ b/dex2oat/linker/image_writer.h
@@ -674,7 +674,12 @@
   template <typename T>
   T* NativeCopyLocation(T* obj) REQUIRES_SHARED(Locks::mutator_lock_);
 
-  // Return true of obj is inside of the boot image space. This may only return true if we are
+  // Return true if `obj` belongs to the image we're writing.
+  // For a boot image, this is true for all objects.
+  // For an app image, boot image objects and boot class path dex caches are excluded.
+  bool IsImageObject(ObjPtr<mirror::Object> obj) const REQUIRES_SHARED(Locks::mutator_lock_);
+
+  // Return true if `obj` is inside of the boot image space. This may only return true if we are
   // compiling an app image.
   bool IsInBootImage(const void* obj) const;
 
diff --git a/dex2oat/linker/oat_writer.cc b/dex2oat/linker/oat_writer.cc
index d045698..e2a9ac2 100644
--- a/dex2oat/linker/oat_writer.cc
+++ b/dex2oat/linker/oat_writer.cc
@@ -2808,11 +2808,9 @@
   return true;
 }
 
-bool OatWriter::WriteHeader(OutputStream* out, uint32_t boot_image_checksum) {
+bool OatWriter::WriteHeader(OutputStream* out) {
   CHECK(write_state_ == WriteState::kWriteHeader);
 
-  oat_header_->SetBootImageChecksum(boot_image_checksum);
-
   // Update checksum with header data.
   DCHECK_EQ(oat_header_->GetChecksum(), 0u);  // For checksum calculation.
   const uint8_t* header_begin = reinterpret_cast<const uint8_t*>(oat_header_.get());
diff --git a/dex2oat/linker/oat_writer.h b/dex2oat/linker/oat_writer.h
index 9cd2fd0..cc0e83a 100644
--- a/dex2oat/linker/oat_writer.h
+++ b/dex2oat/linker/oat_writer.h
@@ -198,7 +198,7 @@
   // Check the size of the written oat file.
   bool CheckOatSize(OutputStream* out, size_t file_offset, size_t relative_offset);
   // Write the oat header. This finalizes the oat file.
-  bool WriteHeader(OutputStream* out, uint32_t boot_image_checksum);
+  bool WriteHeader(OutputStream* out);
 
   // Returns whether the oat file has an associated image.
   bool HasImage() const {
diff --git a/dex2oat/linker/oat_writer_test.cc b/dex2oat/linker/oat_writer_test.cc
index 5de1540..ecf9db8 100644
--- a/dex2oat/linker/oat_writer_test.cc
+++ b/dex2oat/linker/oat_writer_test.cc
@@ -240,7 +240,7 @@
       elf_writer->EndDataBimgRelRo(data_bimg_rel_ro);
     }
 
-    if (!oat_writer.WriteHeader(elf_writer->GetStream(), /*boot_image_checksum=*/ 42u)) {
+    if (!oat_writer.WriteHeader(elf_writer->GetStream())) {
       return false;
     }
 
@@ -396,6 +396,7 @@
 
   ScratchFile tmp_base, tmp_oat(tmp_base, ".oat"), tmp_vdex(tmp_base, ".vdex");
   SafeMap<std::string, std::string> key_value_store;
+  key_value_store.Put(OatHeader::kBootClassPathChecksumsKey, "testkey");
   bool success = WriteElf(tmp_vdex.GetFile(),
                           tmp_oat.GetFile(),
                           class_linker->GetBootClassPath(),
@@ -418,7 +419,8 @@
   const OatHeader& oat_header = oat_file->GetOatHeader();
   ASSERT_TRUE(oat_header.IsValid());
   ASSERT_EQ(class_linker->GetBootClassPath().size(), oat_header.GetDexFileCount());  // core
-  ASSERT_EQ(42u, oat_header.GetBootImageChecksum());
+  ASSERT_TRUE(oat_header.GetStoreValueByKey(OatHeader::kBootClassPathChecksumsKey) != nullptr);
+  ASSERT_STREQ("testkey", oat_header.GetStoreValueByKey(OatHeader::kBootClassPathChecksumsKey));
 
   ASSERT_TRUE(java_lang_dex_file_ != nullptr);
   const DexFile& dex_file = *java_lang_dex_file_;
@@ -464,7 +466,7 @@
 TEST_F(OatTest, OatHeaderSizeCheck) {
   // If this test is failing and you have to update these constants,
   // it is time to update OatHeader::kOatVersion
-  EXPECT_EQ(68U, sizeof(OatHeader));
+  EXPECT_EQ(64U, sizeof(OatHeader));
   EXPECT_EQ(4U, sizeof(OatMethodOffsets));
   EXPECT_EQ(8U, sizeof(OatQuickMethodHeader));
   EXPECT_EQ(166 * static_cast<size_t>(GetInstructionSetPointerSize(kRuntimeISA)),
diff --git a/libartbase/Android.bp b/libartbase/Android.bp
index 58d12a1..c3fb5fd 100644
--- a/libartbase/Android.bp
+++ b/libartbase/Android.bp
@@ -177,9 +177,6 @@
     header_libs: [
         "libnativehelper_header_only",
     ],
-    include_dirs: [
-        "external/icu/icu4c/source/common",
-    ],
 }
 
 art_cc_test {
diff --git a/libartbase/base/common_art_test.cc b/libartbase/base/common_art_test.cc
index 278203d..46724ee 100644
--- a/libartbase/base/common_art_test.cc
+++ b/libartbase/base/common_art_test.cc
@@ -26,7 +26,6 @@
 #include "android-base/stringprintf.h"
 #include "android-base/strings.h"
 #include "android-base/unique_fd.h"
-#include <unicode/uvernum.h>
 
 #include "art_field-inl.h"
 #include "base/file_utils.h"
@@ -329,15 +328,18 @@
 }
 
 std::vector<std::string> CommonArtTestImpl::GetLibCoreDexFileNames() {
-  // Note: This must match the TEST_CORE_JARS in Android.common_path.mk
+  // Note: This must start with the CORE_IMG_JARS in Android.common_path.mk
   // because that's what we use for compiling the core.art image.
+  // It may contain additional modules from TEST_CORE_JARS.
   static const char* const kLibcoreModules[] = {
+      // CORE_IMG_JARS modules.
       "core-oj",
       "core-libart",
       "core-simple",
-      "conscrypt",
       "okhttp",
       "bouncycastle",
+      // Additional modules.
+      "conscrypt",
   };
 
   std::vector<std::string> result;
diff --git a/libdexfile/Android.bp b/libdexfile/Android.bp
index 7f25f02..4d6aa5c 100644
--- a/libdexfile/Android.bp
+++ b/libdexfile/Android.bp
@@ -131,6 +131,71 @@
     ],
 }
 
+cc_library_headers {
+    name: "libdexfile_external_headers",
+    host_supported: true,
+    header_libs: ["libbase_headers"],
+    export_header_lib_headers: ["libbase_headers"],
+    export_include_dirs: ["external/include"],
+
+    target: {
+        windows: {
+            enabled: true,
+        },
+    },
+}
+
+cc_library {
+    name: "libdexfile_external",
+    host_supported: true,
+    srcs: [
+        "external/dex_file_ext.cc",
+    ],
+    header_libs: ["libdexfile_external_headers"],
+    shared_libs: [
+        "libbase",
+        "libdexfile",
+    ],
+
+    // TODO(b/120670568): Enable this when linking bug is fixed.
+    // stubs: {
+    //     symbol_file: "external/libdexfile_external.map.txt",
+    //     versions: ["1"],
+    // },
+
+    // Hide symbols using version scripts for targets that support it, i.e. all
+    // but Darwin.
+    // TODO(b/120670568): Clean this up when stubs above is enabled.
+    target: {
+        android: {
+            version_script: "external/libdexfile_external.map.txt",
+        },
+        linux_bionic: {
+            version_script: "external/libdexfile_external.map.txt",
+        },
+        linux_glibc: {
+            version_script: "external/libdexfile_external.map.txt",
+        },
+        windows: {
+            version_script: "external/libdexfile_external.map.txt",
+        },
+    },
+}
+
+// Support library with a C++ API for accessing the libdexfile API for external
+// (non-ART) users. They should link to their own instance of this (either
+// statically or through linker namespaces).
+cc_library {
+    name: "libdexfile_support",
+    host_supported: true,
+    srcs: [
+        "external/dex_file_supp.cc",
+    ],
+    header_libs: ["libdexfile_external_headers"],
+    shared_libs: ["libdexfile_external"],
+    export_header_lib_headers: ["libdexfile_external_headers"],
+}
+
 art_cc_test {
     name: "art_libdexfile_tests",
     defaults: [
diff --git a/libdexfile/dex/art_dex_file_loader.cc b/libdexfile/dex/art_dex_file_loader.cc
index 20a519b..ae1322d 100644
--- a/libdexfile/dex/art_dex_file_loader.cc
+++ b/libdexfile/dex/art_dex_file_loader.cc
@@ -156,14 +156,16 @@
   return false;
 }
 
-std::unique_ptr<const DexFile> ArtDexFileLoader::Open(const uint8_t* base,
-                                                      size_t size,
-                                                      const std::string& location,
-                                                      uint32_t location_checksum,
-                                                      const OatDexFile* oat_dex_file,
-                                                      bool verify,
-                                                      bool verify_checksum,
-                                                      std::string* error_msg) const {
+std::unique_ptr<const DexFile> ArtDexFileLoader::Open(
+    const uint8_t* base,
+    size_t size,
+    const std::string& location,
+    uint32_t location_checksum,
+    const OatDexFile* oat_dex_file,
+    bool verify,
+    bool verify_checksum,
+    std::string* error_msg,
+    std::unique_ptr<DexFileContainer> container) const {
   ScopedTrace trace(std::string("Open dex file from RAM ") + location);
   return OpenCommon(base,
                     size,
@@ -175,7 +177,7 @@
                     verify,
                     verify_checksum,
                     error_msg,
-                    /*container=*/ nullptr,
+                    std::move(container),
                     /*verify_result=*/ nullptr);
 }
 
diff --git a/libdexfile/dex/art_dex_file_loader.h b/libdexfile/dex/art_dex_file_loader.h
index 40d4673..d41eac5 100644
--- a/libdexfile/dex/art_dex_file_loader.h
+++ b/libdexfile/dex/art_dex_file_loader.h
@@ -54,14 +54,16 @@
                             bool* only_contains_uncompressed_dex = nullptr) const override;
 
   // Opens .dex file, backed by existing memory
-  std::unique_ptr<const DexFile> Open(const uint8_t* base,
-                                      size_t size,
-                                      const std::string& location,
-                                      uint32_t location_checksum,
-                                      const OatDexFile* oat_dex_file,
-                                      bool verify,
-                                      bool verify_checksum,
-                                      std::string* error_msg) const override;
+  std::unique_ptr<const DexFile> Open(
+      const uint8_t* base,
+      size_t size,
+      const std::string& location,
+      uint32_t location_checksum,
+      const OatDexFile* oat_dex_file,
+      bool verify,
+      bool verify_checksum,
+      std::string* error_msg,
+      std::unique_ptr<DexFileContainer> container = nullptr) const override;
 
   // Opens .dex file that has been memory-mapped by the caller.
   std::unique_ptr<const DexFile> Open(const std::string& location,
diff --git a/libdexfile/dex/dex_file_loader.cc b/libdexfile/dex/dex_file_loader.cc
index 3667c8c..1884bcf 100644
--- a/libdexfile/dex/dex_file_loader.cc
+++ b/libdexfile/dex/dex_file_loader.cc
@@ -212,14 +212,16 @@
   return false;
 }
 
-std::unique_ptr<const DexFile> DexFileLoader::Open(const uint8_t* base,
-                                                   size_t size,
-                                                   const std::string& location,
-                                                   uint32_t location_checksum,
-                                                   const OatDexFile* oat_dex_file,
-                                                   bool verify,
-                                                   bool verify_checksum,
-                                                   std::string* error_msg) const {
+std::unique_ptr<const DexFile> DexFileLoader::Open(
+    const uint8_t* base,
+    size_t size,
+    const std::string& location,
+    uint32_t location_checksum,
+    const OatDexFile* oat_dex_file,
+    bool verify,
+    bool verify_checksum,
+    std::string* error_msg,
+    std::unique_ptr<DexFileContainer> container) const {
   return OpenCommon(base,
                     size,
                     /*data_base=*/ nullptr,
@@ -230,7 +232,7 @@
                     verify,
                     verify_checksum,
                     error_msg,
-                    /*container=*/ nullptr,
+                    std::move(container),
                     /*verify_result=*/ nullptr);
 }
 
diff --git a/libdexfile/dex/dex_file_loader.h b/libdexfile/dex/dex_file_loader.h
index 8fc836e..49e177f 100644
--- a/libdexfile/dex/dex_file_loader.h
+++ b/libdexfile/dex/dex_file_loader.h
@@ -121,14 +121,16 @@
                                     bool* zip_file_only_contains_uncompress_dex = nullptr) const;
 
   // Opens .dex file, backed by existing memory
-  virtual std::unique_ptr<const DexFile> Open(const uint8_t* base,
-                                              size_t size,
-                                              const std::string& location,
-                                              uint32_t location_checksum,
-                                              const OatDexFile* oat_dex_file,
-                                              bool verify,
-                                              bool verify_checksum,
-                                              std::string* error_msg) const;
+  virtual std::unique_ptr<const DexFile> Open(
+      const uint8_t* base,
+      size_t size,
+      const std::string& location,
+      uint32_t location_checksum,
+      const OatDexFile* oat_dex_file,
+      bool verify,
+      bool verify_checksum,
+      std::string* error_msg,
+      std::unique_ptr<DexFileContainer> container = nullptr) const;
 
   // Open a dex file with a separate data section.
   virtual std::unique_ptr<const DexFile> OpenWithDataSection(
diff --git a/libdexfile/external/dex_file_ext.cc b/libdexfile/external/dex_file_ext.cc
new file mode 100644
index 0000000..3c193f4
--- /dev/null
+++ b/libdexfile/external/dex_file_ext.cc
@@ -0,0 +1,346 @@
+/*
+ * Copyright (C) 2018 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 <inttypes.h>
+#include <stdint.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include <cerrno>
+#include <cstring>
+#include <map>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <android-base/logging.h>
+#include <android-base/macros.h>
+#include <android-base/mapped_file.h>
+#include <android-base/stringprintf.h>
+
+#include <dex/class_accessor-inl.h>
+#include <dex/code_item_accessors-inl.h>
+#include <dex/dex_file-inl.h>
+#include <dex/dex_file_loader.h>
+
+#include "art_api/ext_dex_file.h"
+
+extern "C" class ExtDexFileString {
+ public:
+  const std::string str_;
+};
+
+namespace art {
+namespace {
+
+const ExtDexFileString empty_string{""};
+
+struct MethodCacheEntry {
+  int32_t offset;  // Offset relative to the start of the dex file header.
+  int32_t len;
+  int32_t index;  // Method index.
+  std::string name;  // Method name. Not filled in for all cache entries.
+};
+
+class MappedFileContainer : public DexFileContainer {
+ public:
+  explicit MappedFileContainer(std::unique_ptr<android::base::MappedFile>&& map)
+      : map_(std::move(map)) {}
+  ~MappedFileContainer() override {}
+  int GetPermissions() override { return 0; }
+  bool IsReadOnly() override { return true; }
+  bool EnableWrite() override { return false; }
+  bool DisableWrite() override { return false; }
+
+ private:
+  std::unique_ptr<android::base::MappedFile> map_;
+  DISALLOW_COPY_AND_ASSIGN(MappedFileContainer);
+};
+
+}  // namespace
+}  // namespace art
+
+extern "C" {
+
+const ExtDexFileString* ExtDexFileMakeString(const char* str) {
+  if (str[0] == '\0') {
+    return &art::empty_string;
+  }
+  return new ExtDexFileString{str};
+}
+
+const char* ExtDexFileGetString(const ExtDexFileString* ext_string, /*out*/ size_t* size) {
+  DCHECK(ext_string != nullptr);
+  *size = ext_string->str_.size();
+  return ext_string->str_.data();
+}
+
+void ExtDexFileFreeString(const ExtDexFileString* ext_string) {
+  DCHECK(ext_string != nullptr);
+  if (ext_string != &art::empty_string) {
+    delete (ext_string);
+  }
+}
+
+// Wraps DexFile to add the caching needed by the external interface. This is
+// what gets passed over as ExtDexFile*.
+class ExtDexFile {
+  // Method cache for GetMethodInfoForOffset. This is populated as we iterate
+  // sequentially through the class defs. MethodCacheEntry.name is only set for
+  // methods returned by GetMethodInfoForOffset.
+  std::map<int32_t, art::MethodCacheEntry> method_cache_;
+
+  // Index of first class def for which method_cache_ isn't complete.
+  uint32_t class_def_index_ = 0;
+
+ public:
+  std::unique_ptr<const art::DexFile> dex_file_;
+  explicit ExtDexFile(std::unique_ptr<const art::DexFile>&& dex_file)
+      : dex_file_(std::move(dex_file)) {}
+
+  art::MethodCacheEntry* GetMethodCacheEntryForOffset(int64_t dex_offset) {
+    // First look in the method cache.
+    auto it = method_cache_.upper_bound(dex_offset);
+    if (it != method_cache_.end() && dex_offset >= it->second.offset) {
+      return &it->second;
+    }
+
+    for (; class_def_index_ < dex_file_->NumClassDefs(); class_def_index_++) {
+      art::ClassAccessor accessor(*dex_file_, class_def_index_);
+
+      for (const art::ClassAccessor::Method& method : accessor.GetMethods()) {
+        art::CodeItemInstructionAccessor code = method.GetInstructions();
+        if (!code.HasCodeItem()) {
+          continue;
+        }
+
+        int32_t offset = reinterpret_cast<const uint8_t*>(code.Insns()) - dex_file_->Begin();
+        int32_t len = code.InsnsSizeInBytes();
+        int32_t index = method.GetIndex();
+        auto res =
+            method_cache_.emplace(offset + len, art::MethodCacheEntry{offset, len, index, ""});
+        if (offset <= dex_offset && dex_offset < offset + len) {
+          return &res.first->second;
+        }
+      }
+    }
+
+    return nullptr;
+  }
+
+  const std::string& GetMethodName(art::MethodCacheEntry& entry) {
+    if (entry.name.empty()) {
+      entry.name = dex_file_->PrettyMethod(entry.index, false);
+    }
+    return entry.name;
+  }
+};
+
+int ExtDexFileOpenFromMemory(const void* addr,
+                             /*inout*/ size_t* size,
+                             const char* location,
+                             /*out*/ const ExtDexFileString** ext_error_msg,
+                             /*out*/ ExtDexFile** ext_dex_file) {
+  if (*size < sizeof(art::DexFile::Header)) {
+    *size = sizeof(art::DexFile::Header);
+    *ext_error_msg = nullptr;
+    return false;
+  }
+
+  const art::DexFile::Header* header = reinterpret_cast<const art::DexFile::Header*>(addr);
+  uint32_t file_size = header->file_size_;
+  if (art::CompactDexFile::IsMagicValid(header->magic_)) {
+    // Compact dex files store the data section separately so that it can be shared.
+    // Therefore we need to extend the read memory range to include it.
+    // TODO: This might be wasteful as we might read data in between as well.
+    //       In practice, this should be fine, as such sharing only happens on disk.
+    uint32_t computed_file_size;
+    if (__builtin_add_overflow(header->data_off_, header->data_size_, &computed_file_size)) {
+      *ext_error_msg = new ExtDexFileString{
+          android::base::StringPrintf("Corrupt CompactDexFile header in '%s'", location)};
+      return false;
+    }
+    if (computed_file_size > file_size) {
+      file_size = computed_file_size;
+    }
+  } else if (!art::StandardDexFile::IsMagicValid(header->magic_)) {
+    *ext_error_msg = new ExtDexFileString{
+        android::base::StringPrintf("Unrecognized dex file header in '%s'", location)};
+    return false;
+  }
+
+  if (*size < file_size) {
+    *size = file_size;
+    *ext_error_msg = nullptr;
+    return false;
+  }
+
+  std::string loc_str(location);
+  art::DexFileLoader loader;
+  std::string error_msg;
+  std::unique_ptr<const art::DexFile> dex_file = loader.Open(static_cast<const uint8_t*>(addr),
+                                                             *size,
+                                                             loc_str,
+                                                             header->checksum_,
+                                                             /*oat_dex_file=*/nullptr,
+                                                             /*verify=*/false,
+                                                             /*verify_checksum=*/false,
+                                                             &error_msg);
+  if (dex_file == nullptr) {
+    *ext_error_msg = new ExtDexFileString{std::move(error_msg)};
+    return false;
+  }
+
+  *ext_dex_file = new ExtDexFile(std::move(dex_file));
+  return true;
+}
+
+int ExtDexFileOpenFromFd(int fd,
+                         off_t offset,
+                         const char* location,
+                         /*out*/ const ExtDexFileString** ext_error_msg,
+                         /*out*/ ExtDexFile** ext_dex_file) {
+  size_t length;
+  {
+    struct stat sbuf;
+    std::memset(&sbuf, 0, sizeof(sbuf));
+    if (fstat(fd, &sbuf) == -1) {
+      *ext_error_msg = new ExtDexFileString{
+          android::base::StringPrintf("fstat '%s' failed: %s", location, std::strerror(errno))};
+      return false;
+    }
+    if (S_ISDIR(sbuf.st_mode)) {
+      *ext_error_msg = new ExtDexFileString{
+          android::base::StringPrintf("Attempt to mmap directory '%s'", location)};
+      return false;
+    }
+    length = sbuf.st_size;
+  }
+
+  if (length < offset + sizeof(art::DexFile::Header)) {
+    *ext_error_msg = new ExtDexFileString{android::base::StringPrintf(
+        "Offset %" PRId64 " too large for '%s' of size %zu", int64_t{offset}, location, length)};
+    return false;
+  }
+
+  // Cannot use MemMap in libartbase here, because it pulls in dlopen which we
+  // can't have when being compiled statically.
+  std::unique_ptr<android::base::MappedFile> map =
+      android::base::MappedFile::FromFd(fd, offset, length, PROT_READ);
+  if (map == nullptr) {
+    *ext_error_msg = new ExtDexFileString{
+        android::base::StringPrintf("mmap '%s' failed: %s", location, std::strerror(errno))};
+    return false;
+  }
+
+  const art::DexFile::Header* header = reinterpret_cast<const art::DexFile::Header*>(map->data());
+  uint32_t file_size;
+  if (__builtin_add_overflow(offset, header->file_size_, &file_size)) {
+    *ext_error_msg =
+        new ExtDexFileString{android::base::StringPrintf("Corrupt header in '%s'", location)};
+    return false;
+  }
+  if (length < file_size) {
+    *ext_error_msg = new ExtDexFileString{
+        android::base::StringPrintf("Dex file '%s' too short: expected %" PRIu32 ", got %" PRIu64,
+                                    location,
+                                    file_size,
+                                    uint64_t{length})};
+    return false;
+  }
+
+  void* addr = map->data();
+  size_t size = map->size();
+  auto container = std::make_unique<art::MappedFileContainer>(std::move(map));
+
+  std::string loc_str(location);
+  std::string error_msg;
+  art::DexFileLoader loader;
+  std::unique_ptr<const art::DexFile> dex_file = loader.Open(reinterpret_cast<const uint8_t*>(addr),
+                                                             size,
+                                                             loc_str,
+                                                             header->checksum_,
+                                                             /*oat_dex_file=*/nullptr,
+                                                             /*verify=*/false,
+                                                             /*verify_checksum=*/false,
+                                                             &error_msg,
+                                                             std::move(container));
+  if (dex_file == nullptr) {
+    *ext_error_msg = new ExtDexFileString{std::move(error_msg)};
+    return false;
+  }
+  *ext_dex_file = new ExtDexFile(std::move(dex_file));
+  return true;
+}
+
+int ExtDexFileGetMethodInfoForOffset(ExtDexFile* ext_dex_file,
+                                     int64_t dex_offset,
+                                     /*out*/ ExtDexFileMethodInfo* method_info) {
+  if (!ext_dex_file->dex_file_->IsInDataSection(ext_dex_file->dex_file_->Begin() + dex_offset)) {
+    return false;  // The DEX offset is not within the bytecode of this dex file.
+  }
+
+  if (ext_dex_file->dex_file_->IsCompactDexFile()) {
+    // The data section of compact dex files might be shared.
+    // Check the subrange unique to this compact dex.
+    const art::CompactDexFile::Header& cdex_header =
+        ext_dex_file->dex_file_->AsCompactDexFile()->GetHeader();
+    uint32_t begin = cdex_header.data_off_ + cdex_header.OwnedDataBegin();
+    uint32_t end = cdex_header.data_off_ + cdex_header.OwnedDataEnd();
+    if (dex_offset < begin || dex_offset >= end) {
+      return false;  // The DEX offset is not within the bytecode of this dex file.
+    }
+  }
+
+  art::MethodCacheEntry* entry = ext_dex_file->GetMethodCacheEntryForOffset(dex_offset);
+  if (entry != nullptr) {
+    method_info->offset = entry->offset;
+    method_info->len = entry->len;
+    method_info->name = new ExtDexFileString{ext_dex_file->GetMethodName(*entry)};
+    return true;
+  }
+
+  return false;
+}
+
+void ExtDexFileGetAllMethodInfos(ExtDexFile* ext_dex_file,
+                                 int with_signature,
+                                 ExtDexFileMethodInfoCallback* method_info_cb,
+                                 void* user_data) {
+  for (art::ClassAccessor accessor : ext_dex_file->dex_file_->GetClasses()) {
+    for (const art::ClassAccessor::Method& method : accessor.GetMethods()) {
+      art::CodeItemInstructionAccessor code = method.GetInstructions();
+      if (!code.HasCodeItem()) {
+        continue;
+      }
+
+      ExtDexFileMethodInfo method_info;
+      method_info.offset = static_cast<int32_t>(reinterpret_cast<const uint8_t*>(code.Insns()) -
+                                                ext_dex_file->dex_file_->Begin());
+      method_info.len = code.InsnsSizeInBytes();
+      method_info.name = new ExtDexFileString{
+          ext_dex_file->dex_file_->PrettyMethod(method.GetIndex(), with_signature)};
+      method_info_cb(&method_info, user_data);
+    }
+  }
+}
+
+void ExtDexFileFree(ExtDexFile* ext_dex_file) { delete (ext_dex_file); }
+
+}  // extern "C"
diff --git a/libdexfile/external/dex_file_supp.cc b/libdexfile/external/dex_file_supp.cc
new file mode 100644
index 0000000..6514c8a
--- /dev/null
+++ b/libdexfile/external/dex_file_supp.cc
@@ -0,0 +1,34 @@
+/*
+ * Copyright (C) 2018 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 "art_api/ext_dex_file.h"
+
+namespace art_api {
+namespace dex {
+
+DexFile::~DexFile() { ExtDexFileFree(ext_dex_file_); }
+
+MethodInfo DexFile::AbsorbMethodInfo(const ExtDexFileMethodInfo& ext_method_info) {
+  return {ext_method_info.offset, ext_method_info.len, DexString(ext_method_info.name)};
+}
+
+void DexFile::AddMethodInfoCallback(const ExtDexFileMethodInfo* ext_method_info, void* ctx) {
+  auto vect = static_cast<MethodInfoVector*>(ctx);
+  vect->emplace_back(AbsorbMethodInfo(*ext_method_info));
+}
+
+}  // namespace dex
+}  // namespace art_api
diff --git a/libdexfile/external/include/art_api/ext_dex_file.h b/libdexfile/external/include/art_api/ext_dex_file.h
new file mode 100644
index 0000000..5f64ab1
--- /dev/null
+++ b/libdexfile/external/include/art_api/ext_dex_file.h
@@ -0,0 +1,248 @@
+/*
+ * Copyright (C) 2018 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_LIBDEXFILE_EXTERNAL_INCLUDE_ART_API_EXT_DEX_FILE_H_
+#define ART_LIBDEXFILE_EXTERNAL_INCLUDE_ART_API_EXT_DEX_FILE_H_
+
+// Dex file external API
+
+#include <sys/types.h>
+
+#include <cstring>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <vector>
+
+#include <android-base/macros.h>
+
+extern "C" {
+
+// This is the stable C ABI that backs art_api::dex below. Structs and functions
+// may only be added here.
+// TODO(b/120978655): Move this to a separate pure C header.
+//
+// Clients should use the C++ wrappers in art_api::dex instead.
+
+// Opaque wrapper for an std::string allocated in libdexfile which must be freed
+// using ExtDexFileFreeString.
+class ExtDexFileString;
+
+// Returns an ExtDexFileString initialized to the given string.
+const ExtDexFileString* ExtDexFileMakeString(const char* str);
+
+// Returns a pointer to the underlying null-terminated character array and its
+// size for the given ExtDexFileString.
+const char* ExtDexFileGetString(const ExtDexFileString* ext_string, /*out*/ size_t* size);
+
+// Frees an ExtDexFileString.
+void ExtDexFileFreeString(const ExtDexFileString* ext_string);
+
+struct ExtDexFileMethodInfo {
+  int32_t offset;
+  int32_t len;
+  const ExtDexFileString* name;
+};
+
+class ExtDexFile;
+
+// See art_api::dex::DexFile::OpenFromMemory. Returns true on success.
+int ExtDexFileOpenFromMemory(const void* addr,
+                             /*inout*/ size_t* size,
+                             const char* location,
+                             /*out*/ const ExtDexFileString** error_msg,
+                             /*out*/ ExtDexFile** ext_dex_file);
+
+// See art_api::dex::DexFile::OpenFromFd. Returns true on success.
+int ExtDexFileOpenFromFd(int fd,
+                         off_t offset,
+                         const char* location,
+                         /*out*/ const ExtDexFileString** error_msg,
+                         /*out*/ ExtDexFile** ext_dex_file);
+
+// See art_api::dex::DexFile::GetMethodInfoForOffset. Returns true on success.
+int ExtDexFileGetMethodInfoForOffset(ExtDexFile* ext_dex_file,
+                                     int64_t dex_offset,
+                                     /*out*/ ExtDexFileMethodInfo* method_info);
+
+typedef void ExtDexFileMethodInfoCallback(const ExtDexFileMethodInfo* ext_method_info,
+                                          void* user_data);
+
+// See art_api::dex::DexFile::GetAllMethodInfos.
+void ExtDexFileGetAllMethodInfos(ExtDexFile* ext_dex_file,
+                                 int with_signature,
+                                 ExtDexFileMethodInfoCallback* method_info_cb,
+                                 void* user_data);
+
+// Frees an ExtDexFile.
+void ExtDexFileFree(ExtDexFile* ext_dex_file);
+
+}  // extern "C"
+
+namespace art_api {
+namespace dex {
+
+// Minimal std::string look-alike for a string returned from libdexfile.
+class DexString final {
+ public:
+  DexString(DexString&& dex_str) { ReplaceExtString(std::move(dex_str)); }
+  explicit DexString(const char* str = "") : ext_string_(ExtDexFileMakeString(str)) {}
+  ~DexString() { ExtDexFileFreeString(ext_string_); }
+
+  DexString& operator=(DexString&& dex_str) {
+    ReplaceExtString(std::move(dex_str));
+    return *this;
+  }
+
+  const char* data() const {
+    size_t ignored;
+    return ExtDexFileGetString(ext_string_, &ignored);
+  }
+  const char* c_str() const { return data(); }
+
+  size_t size() const {
+    size_t len;
+    (void)ExtDexFileGetString(ext_string_, &len);
+    return len;
+  }
+  size_t length() const { return size(); }
+
+  operator std::string_view() const {
+    size_t len;
+    const char* chars = ExtDexFileGetString(ext_string_, &len);
+    return std::string_view(chars, len);
+  }
+
+ private:
+  friend class DexFile;
+  friend bool operator==(const DexString&, const DexString&);
+  explicit DexString(const ExtDexFileString* ext_string) : ext_string_(ext_string) {}
+  const ExtDexFileString* ext_string_;  // Owned instance. Never nullptr.
+
+  void ReplaceExtString(DexString&& dex_str) {
+    ext_string_ = dex_str.ext_string_;
+    dex_str.ext_string_ = ExtDexFileMakeString("");
+  }
+
+  DISALLOW_COPY_AND_ASSIGN(DexString);
+};
+
+inline bool operator==(const DexString& s1, const DexString& s2) {
+  size_t l1, l2;
+  const char* str1 = ExtDexFileGetString(s1.ext_string_, &l1);
+  const char* str2 = ExtDexFileGetString(s2.ext_string_, &l2);
+  // Use memcmp to avoid assumption about absence of null characters in the strings.
+  return l1 == l2 && !std::memcmp(str1, str2, l1);
+}
+
+struct MethodInfo {
+  int32_t offset;  // Code offset relative to the start of the dex file header
+  int32_t len;  // Code length
+  DexString name;
+};
+
+inline bool operator==(const MethodInfo& s1, const MethodInfo& s2) {
+  return s1.offset == s2.offset && s1.len == s2.len && s1.name == s2.name;
+}
+
+// External stable API to access ordinary dex files and CompactDex. This wraps
+// the stable C ABI and handles instance ownership. Thread-compatible but not
+// thread-safe.
+class DexFile {
+ public:
+  DexFile(DexFile&& dex_file) {
+    ext_dex_file_ = dex_file.ext_dex_file_;
+    dex_file.ext_dex_file_ = nullptr;
+  }
+  virtual ~DexFile();
+
+  // Interprets a chunk of memory as a dex file. As long as *size is too small,
+  // returns nullptr, sets *size to a new size to try again with, and sets
+  // *error_msg to "". That might happen repeatedly. Also returns nullptr
+  // on error in which case *error_msg is set to a nonempty string.
+  //
+  // location is a string that describes the dex file, and is preferably its
+  // path. It is mostly used to make error messages better, and may be "".
+  //
+  // The caller must retain the memory.
+  static std::unique_ptr<DexFile> OpenFromMemory(const void* addr,
+                                                 size_t* size,
+                                                 const std::string& location,
+                                                 /*out*/ std::string* error_msg) {
+    ExtDexFile* ext_dex_file;
+    const ExtDexFileString* ext_error_msg = nullptr;
+    if (ExtDexFileOpenFromMemory(addr, size, location.c_str(), &ext_error_msg, &ext_dex_file)) {
+      return std::unique_ptr<DexFile>(new DexFile(ext_dex_file));
+    }
+    *error_msg = (ext_error_msg == nullptr) ? "" : std::string(DexString(ext_error_msg));
+    return nullptr;
+  }
+
+  // mmaps the given file offset in the open fd and reads a dexfile from there.
+  // Returns nullptr on error in which case *error_msg is set.
+  //
+  // location is a string that describes the dex file, and is preferably its
+  // path. It is mostly used to make error messages better, and may be "".
+  static std::unique_ptr<DexFile> OpenFromFd(int fd,
+                                             off_t offset,
+                                             const std::string& location,
+                                             /*out*/ std::string* error_msg) {
+    ExtDexFile* ext_dex_file;
+    const ExtDexFileString* ext_error_msg = nullptr;
+    if (ExtDexFileOpenFromFd(fd, offset, location.c_str(), &ext_error_msg, &ext_dex_file)) {
+      return std::unique_ptr<DexFile>(new DexFile(ext_dex_file));
+    }
+    *error_msg = std::string(DexString(ext_error_msg));
+    return nullptr;
+  }
+
+  // Given an offset relative to the start of the dex file header, if there is a
+  // method whose instruction range includes that offset then returns info about
+  // it, otherwise returns a struct with offset == 0.
+  MethodInfo GetMethodInfoForOffset(int64_t dex_offset) {
+    ExtDexFileMethodInfo ext_method_info;
+    if (ExtDexFileGetMethodInfoForOffset(ext_dex_file_, dex_offset, &ext_method_info)) {
+      return AbsorbMethodInfo(ext_method_info);
+    }
+    return {/*offset=*/0, /*len=*/0, /*name=*/DexString()};
+  }
+
+  // Returns info structs about all methods in the dex file. MethodInfo.name
+  // receives the full function signature if with_signature is set, otherwise it
+  // gets the class and method name only.
+  std::vector<MethodInfo> GetAllMethodInfos(bool with_signature = true) {
+    MethodInfoVector res;
+    ExtDexFileGetAllMethodInfos(
+        ext_dex_file_, with_signature, AddMethodInfoCallback, static_cast<void*>(&res));
+    return res;
+  }
+
+ private:
+  explicit DexFile(ExtDexFile* ext_dex_file) : ext_dex_file_(ext_dex_file) {}
+  ExtDexFile* ext_dex_file_;  // Owned instance. nullptr only in moved-from zombies.
+
+  typedef std::vector<MethodInfo> MethodInfoVector;
+
+  static MethodInfo AbsorbMethodInfo(const ExtDexFileMethodInfo& ext_method_info);
+  static void AddMethodInfoCallback(const ExtDexFileMethodInfo* ext_method_info, void* user_data);
+
+  DISALLOW_COPY_AND_ASSIGN(DexFile);
+};
+
+}  // namespace dex
+}  // namespace art_api
+
+#endif  // ART_LIBDEXFILE_EXTERNAL_INCLUDE_ART_API_EXT_DEX_FILE_H_
diff --git a/libdexfile/external/libdexfile_external.map.txt b/libdexfile/external/libdexfile_external.map.txt
new file mode 100644
index 0000000..450b633
--- /dev/null
+++ b/libdexfile/external/libdexfile_external.map.txt
@@ -0,0 +1,13 @@
+LIBDEXFILE_EXTERNAL_1 {
+  global:
+    ExtDexFileFree;
+    ExtDexFileFreeString;
+    ExtDexFileGetAllMethodInfos;
+    ExtDexFileGetMethodInfoForOffset;
+    ExtDexFileGetString;
+    ExtDexFileMakeString;
+    ExtDexFileOpenFromFd;
+    ExtDexFileOpenFromMemory;
+  local:
+    *;
+};
diff --git a/oatdump/oatdump.cc b/oatdump/oatdump.cc
index 1c74a92..d2a5bb8 100644
--- a/oatdump/oatdump.cc
+++ b/oatdump/oatdump.cc
@@ -473,9 +473,6 @@
                            GetQuickToInterpreterBridgeOffset);
 #undef DUMP_OAT_HEADER_OFFSET
 
-    os << "BOOT IMAGE CHECKSUM:\n";
-    os << StringPrintf("0x%08x\n\n", oat_header.GetBootImageChecksum());
-
     // Print the key-value store.
     {
       os << "KEY VALUE STORE:\n";
diff --git a/runtime/Android.bp b/runtime/Android.bp
index b03ef60..4780f16 100644
--- a/runtime/Android.bp
+++ b/runtime/Android.bp
@@ -420,6 +420,8 @@
         "libbacktrace",
         "libbase",
         "libcutils",
+        "libdexfile_external",  // libunwindstack dependency
+        "libdexfile_support",  // libunwindstack dependency
         "liblog",
         "libnativebridge",
         "libnativeloader",
@@ -553,9 +555,6 @@
     header_libs: [
         "libnativehelper_header_only",
     ],
-    include_dirs: [
-        "external/icu/icu4c/source/common",
-    ],
 }
 
 art_cc_test {
diff --git a/runtime/class_linker.cc b/runtime/class_linker.cc
index d33541c..3b92e2c 100644
--- a/runtime/class_linker.cc
+++ b/runtime/class_linker.cc
@@ -212,6 +212,22 @@
   self->AssertPendingException();
 }
 
+// Ensures that methods have the kAccSkipAccessChecks bit set. We use the
+// kAccVerificationAttempted bit on the class access flags to determine whether this has been done
+// before.
+template <bool kNeedsVerified = false>
+static void EnsureSkipAccessChecksMethods(Handle<mirror::Class> klass, PointerSize pointer_size)
+    REQUIRES_SHARED(Locks::mutator_lock_) {
+  if (kNeedsVerified) {
+    // To not fail access-flags access checks, push a minimal state.
+    mirror::Class::SetStatus(klass, ClassStatus::kVerified, Thread::Current());
+  }
+  if (!klass->WasVerificationAttempted()) {
+    klass->SetSkipAccessChecksFlagOnAllMethods(pointer_size);
+    klass->SetVerificationAttempted();
+  }
+}
+
 void ClassLinker::ThrowEarlierClassFailure(ObjPtr<mirror::Class> c, bool wrap_in_no_class_def) {
   // The class failed to initialize on a previous attempt, so we want to throw
   // a NoClassDefFoundError (v2 2.17.5).  The exception to this rule is if we
@@ -1037,20 +1053,15 @@
   runtime->SetSentinel(heap->AllocNonMovableObject<true>(
       self, java_lang_Object, java_lang_Object->GetObjectSize(), VoidFunctor()));
 
-  const std::vector<std::string>& boot_class_path = runtime->GetBootClassPath();
-  if (boot_class_path.size() != spaces.size()) {
-    *error_msg = StringPrintf("Boot class path has %zu components but there are %zu image spaces.",
-                              boot_class_path.size(),
-                              spaces.size());
-    return false;
-  }
+  const std::vector<std::string>& boot_class_path_locations = runtime->GetBootClassPathLocations();
+  CHECK_LE(spaces.size(), boot_class_path_locations.size());
   for (size_t i = 0u, size = spaces.size(); i != size; ++i) {
     // Boot class loader, use a null handle.
     std::vector<std::unique_ptr<const DexFile>> dex_files;
     if (!AddImageSpace(spaces[i],
                        ScopedNullHandle<mirror::ClassLoader>(),
                        /*dex_elements=*/ nullptr,
-                       /*dex_location=*/ boot_class_path[i].c_str(),
+                       /*dex_location=*/ boot_class_path_locations[i].c_str(),
                        /*out*/&dex_files,
                        error_msg)) {
       return false;
@@ -1069,6 +1080,15 @@
   return true;
 }
 
+void ClassLinker::AddExtraBootDexFiles(
+    Thread* self,
+    std::vector<std::unique_ptr<const DexFile>>&& additional_dex_files) {
+  for (std::unique_ptr<const DexFile>& dex_file : additional_dex_files) {
+    AppendToBootClassPath(self, *dex_file);
+    boot_dex_files_.push_back(std::move(dex_file));
+  }
+}
+
 bool ClassLinker::IsBootClassLoader(ScopedObjectAccessAlreadyRunnable& soa,
                                     ObjPtr<mirror::ClassLoader> class_loader) {
   return class_loader == nullptr ||
@@ -3946,6 +3966,7 @@
   h_class->SetAccessFlags(kAccPublic | kAccFinal | kAccAbstract);
   h_class->SetPrimitiveType(type);
   h_class->SetIfTable(GetClassRoot<mirror::Object>(this)->GetIfTable());
+  EnsureSkipAccessChecksMethods</* kNeedsVerified= */ true>(h_class, image_pointer_size_);
   mirror::Class::SetStatus(h_class, ClassStatus::kInitialized, self);
   const char* descriptor = Primitive::Descriptor(type);
   ObjPtr<mirror::Class> existing = InsertClass(descriptor,
@@ -4093,6 +4114,7 @@
   new_class->PopulateEmbeddedVTable(image_pointer_size_);
   ImTable* object_imt = java_lang_Object->GetImt(image_pointer_size_);
   new_class->SetImt(object_imt, image_pointer_size_);
+  EnsureSkipAccessChecksMethods</* kNeedsVerified= */ true>(new_class, image_pointer_size_);
   mirror::Class::SetStatus(new_class, ClassStatus::kInitialized, self);
   // don't need to set new_class->SetObjectSize(..)
   // because Object::SizeOf delegates to Array::SizeOf
@@ -4123,6 +4145,8 @@
   // and remove "interface".
   access_flags |= kAccAbstract | kAccFinal;
   access_flags &= ~kAccInterface;
+  // Arrays are access-checks-clean and preverified.
+  access_flags |= kAccVerificationAttempted;
 
   new_class->SetAccessFlags(access_flags);
 
@@ -4357,17 +4381,6 @@
   return false;
 }
 
-// Ensures that methods have the kAccSkipAccessChecks bit set. We use the
-// kAccVerificationAttempted bit on the class access flags to determine whether this has been done
-// before.
-static void EnsureSkipAccessChecksMethods(Handle<mirror::Class> klass, PointerSize pointer_size)
-    REQUIRES_SHARED(Locks::mutator_lock_) {
-  if (!klass->WasVerificationAttempted()) {
-    klass->SetSkipAccessChecksFlagOnAllMethods(pointer_size);
-    klass->SetVerificationAttempted();
-  }
-}
-
 verifier::FailureKind ClassLinker::VerifyClass(
     Thread* self, Handle<mirror::Class> klass, verifier::HardFailLogMode log_level) {
   {
@@ -4844,6 +4857,7 @@
   {
     // Lock on klass is released. Lock new class object.
     ObjectLock<mirror::Class> initialization_lock(self, klass);
+    EnsureSkipAccessChecksMethods(klass, image_pointer_size_);
     mirror::Class::SetStatus(klass, ClassStatus::kInitialized, self);
   }
 
@@ -5594,8 +5608,7 @@
   DCHECK(c != nullptr);
 
   if (c->IsInitialized()) {
-    EnsureSkipAccessChecksMethods(c, image_pointer_size_);
-    self->AssertNoPendingException();
+    DCHECK(c->WasVerificationAttempted()) << c->PrettyClassAndClassLoader();
     return true;
   }
   // SubtypeCheckInfo::Initialized must happen-before any new-instance for that type.
diff --git a/runtime/class_linker.h b/runtime/class_linker.h
index d3eab7c..d0a7c9b 100644
--- a/runtime/class_linker.h
+++ b/runtime/class_linker.h
@@ -127,6 +127,12 @@
       REQUIRES_SHARED(Locks::mutator_lock_)
       REQUIRES(!Locks::dex_lock_);
 
+  // Add boot class path dex files that were not included in the boot image.
+  // ClassLinker takes ownership of these dex files.
+  void AddExtraBootDexFiles(Thread* self,
+                            std::vector<std::unique_ptr<const DexFile>>&& additional_dex_files)
+      REQUIRES_SHARED(Locks::mutator_lock_);
+
   // Add an image space to the class linker, may fix up classloader fields and dex cache fields.
   // The dex files that were newly opened for the space are placed in the out argument
   // out_dex_files. Returns true if the operation succeeded.
diff --git a/runtime/class_linker_test.cc b/runtime/class_linker_test.cc
index fe45b9e..061c788 100644
--- a/runtime/class_linker_test.cc
+++ b/runtime/class_linker_test.cc
@@ -114,7 +114,8 @@
     EXPECT_EQ(0, primitive->GetIfTableCount());
     EXPECT_TRUE(primitive->GetIfTable() != nullptr);
     EXPECT_EQ(primitive->GetIfTable()->Count(), 0u);
-    EXPECT_EQ(kAccPublic | kAccFinal | kAccAbstract, primitive->GetAccessFlags());
+    EXPECT_EQ(kAccPublic | kAccFinal | kAccAbstract | kAccVerificationAttempted,
+              primitive->GetAccessFlags());
   }
 
   void AssertObjectClass(ObjPtr<mirror::Class> JavaLangObject)
diff --git a/runtime/common_runtime_test.cc b/runtime/common_runtime_test.cc
index a101976..a20baa0 100644
--- a/runtime/common_runtime_test.cc
+++ b/runtime/common_runtime_test.cc
@@ -24,7 +24,6 @@
 #include "nativehelper/scoped_local_ref.h"
 
 #include "android-base/stringprintf.h"
-#include <unicode/uvernum.h>
 
 #include "art_field-inl.h"
 #include "base/file_utils.h"
diff --git a/runtime/dexopt_test.cc b/runtime/dexopt_test.cc
index b46c933..7f697d1 100644
--- a/runtime/dexopt_test.cc
+++ b/runtime/dexopt_test.cc
@@ -116,19 +116,19 @@
   ASSERT_TRUE(odex_file.get() != nullptr) << error_msg;
   EXPECT_EQ(filter, odex_file->GetCompilerFilter());
 
-  std::unique_ptr<ImageHeader> image_header(
-          gc::space::ImageSpace::ReadImageHeader(image_location.c_str(),
-                                                 kRuntimeISA,
-                                                 &error_msg));
-  ASSERT_TRUE(image_header != nullptr) << error_msg;
+  std::string boot_image_checksums = gc::space::ImageSpace::GetBootClassPathChecksums(
+      Runtime::Current()->GetBootClassPath(), image_location, kRuntimeISA, &error_msg);
+  ASSERT_FALSE(boot_image_checksums.empty()) << error_msg;
+
   const OatHeader& oat_header = odex_file->GetOatHeader();
-  uint32_t boot_image_checksum = image_header->GetImageChecksum();
 
   if (CompilerFilter::DependsOnImageChecksum(filter)) {
+    const char* checksums = oat_header.GetStoreValueByKey(OatHeader::kBootClassPathChecksumsKey);
+    ASSERT_TRUE(checksums != nullptr);
     if (with_alternate_image) {
-      EXPECT_NE(boot_image_checksum, oat_header.GetBootImageChecksum());
+      EXPECT_NE(boot_image_checksums, checksums);
     } else {
-      EXPECT_EQ(boot_image_checksum, oat_header.GetBootImageChecksum());
+      EXPECT_EQ(boot_image_checksums, checksums);
     }
   }
 }
diff --git a/runtime/gc/collector/concurrent_copying-inl.h b/runtime/gc/collector/concurrent_copying-inl.h
index 3160422..1014c0e 100644
--- a/runtime/gc/collector/concurrent_copying-inl.h
+++ b/runtime/gc/collector/concurrent_copying-inl.h
@@ -37,14 +37,15 @@
     mirror::Object* ref,
     accounting::ContinuousSpaceBitmap* bitmap) {
   if (kEnableGenerationalConcurrentCopyingCollection
-      && young_gen_
       && !done_scanning_.load(std::memory_order_acquire)) {
-    // Everything in the unevac space should be marked for generational CC except for large objects.
-    DCHECK(region_space_bitmap_->Test(ref) || region_space_->IsLargeObject(ref)) << ref << " "
+    // Everything in the unevac space should be marked for young generation CC,
+    // except for large objects.
+    DCHECK(!young_gen_ || region_space_bitmap_->Test(ref) || region_space_->IsLargeObject(ref))
+        << ref << " "
         << ref->GetClass<kVerifyNone, kWithoutReadBarrier>()->PrettyClass();
-    // Since the mark bitmap is still filled in from last GC, we can not use that or else the
-    // mutator may see references to the from space. Instead, use the baker pointer itself as
-    // the mark bit.
+    // Since the mark bitmap is still filled in from last GC (or from marking phase of 2-phase CC,
+    // we can not use that or else the mutator may see references to the from space. Instead, use
+    // the baker pointer itself as the mark bit.
     if (ref->AtomicSetReadBarrierState(ReadBarrier::NonGrayState(), ReadBarrier::GrayState())) {
       // TODO: We don't actually need to scan this object later, we just need to clear the gray
       // bit.
@@ -244,7 +245,7 @@
   DCHECK(region_space_->IsInUnevacFromSpace(from_ref));
   if (kUseBakerReadBarrier && from_ref->GetReadBarrierStateAcquire() == ReadBarrier::GrayState()) {
     return true;
-  } else if (!(kEnableGenerationalConcurrentCopyingCollection && young_gen_)
+  } else if (!kEnableGenerationalConcurrentCopyingCollection
              || done_scanning_.load(std::memory_order_acquire)) {
     // If the card table scanning is not finished yet, then only read-barrier
     // state should be checked. Checking the mark bitmap is unreliable as there
diff --git a/runtime/gc/collector/concurrent_copying.cc b/runtime/gc/collector/concurrent_copying.cc
index 7736568..861f0d3 100644
--- a/runtime/gc/collector/concurrent_copying.cc
+++ b/runtime/gc/collector/concurrent_copying.cc
@@ -188,6 +188,11 @@
   {
     ReaderMutexLock mu(self, *Locks::mutator_lock_);
     InitializePhase();
+    // In case of forced evacuation, all regions are evacuated and hence no
+    // need to compute live_bytes.
+    if (kEnableGenerationalConcurrentCopyingCollection && !young_gen_ && !force_evacuate_all_) {
+      MarkingPhase();
+    }
   }
   if (kUseBakerReadBarrier && kGrayDirtyImmuneObjects) {
     // Switch to read barrier mark entrypoints before we gray the objects. This is required in case
@@ -201,7 +206,7 @@
   FlipThreadRoots();
   {
     ReaderMutexLock mu(self, *Locks::mutator_lock_);
-    MarkingPhase();
+    CopyingPhase();
   }
   // Verify no from space refs. This causes a pause.
   if (kEnableNoFromSpaceRefsVerification) {
@@ -299,12 +304,22 @@
           DCHECK_EQ(space->GetGcRetentionPolicy(), space::kGcRetentionPolicyAlwaysCollect);
           space->AsContinuousMemMapAllocSpace()->BindLiveToMarkBitmap();
         }
-        // Age all of the cards for the region space so that we know which evac regions to scan.
-        Runtime::Current()->GetHeap()->GetCardTable()->ModifyCardsAtomic(
-            space->Begin(),
-            space->End(),
-            AgeCardVisitor(),
-            VoidFunctor());
+        if (young_gen_) {
+          // Age all of the cards for the region space so that we know which evac regions to scan.
+          heap_->GetCardTable()->ModifyCardsAtomic(space->Begin(),
+                                                   space->End(),
+                                                   AgeCardVisitor(),
+                                                   VoidFunctor());
+        } else {
+          // In a full-heap GC cycle, the card-table corresponding to region-space and
+          // non-moving space can be cleared, because this cycle only needs to
+          // capture writes during the marking phase of this cycle to catch
+          // objects that skipped marking due to heap mutation. Furthermore,
+          // if the next GC is a young-gen cycle, then it only needs writes to
+          // be captured after the thread-flip of this GC cycle, as that is when
+          // the young-gen for the next GC cycle starts getting populated.
+          heap_->GetCardTable()->ClearCardRange(space->Begin(), space->Limit());
+        }
       } else {
         if (space == region_space_) {
           // It is OK to clear the bitmap with mutators running since the only place it is read is
@@ -381,6 +396,7 @@
   if (kEnableGenerationalConcurrentCopyingCollection && !young_gen_) {
     region_space_bitmap_->Clear();
   }
+  mark_stack_mode_.store(ConcurrentCopying::kMarkStackModeThreadLocal, std::memory_order_relaxed);
   // Mark all of the zygote large objects without graying them.
   MarkZygoteLargeObjects();
 }
@@ -471,7 +487,7 @@
     TimingLogger::ScopedTiming split("(Paused)FlipCallback", cc->GetTimings());
     // Note: self is not necessarily equal to thread since thread may be suspended.
     Thread* self = Thread::Current();
-    if (kVerifyNoMissingCardMarks) {
+    if (kVerifyNoMissingCardMarks && cc->young_gen_) {
       cc->VerifyNoMissingCardMarks();
     }
     CHECK_EQ(thread, self);
@@ -485,9 +501,11 @@
     }
     {
       TimingLogger::ScopedTiming split2("(Paused)SetFromSpace", cc->GetTimings());
-      // Only change live bytes for full CC.
+      // Only change live bytes for 1-phase full heap CC.
       cc->region_space_->SetFromSpace(
-          cc->rb_table_, evac_mode, /*clear_live_bytes=*/ !cc->young_gen_);
+          cc->rb_table_,
+          evac_mode,
+          /*clear_live_bytes=*/ !kEnableGenerationalConcurrentCopyingCollection);
     }
     cc->SwapStacks();
     if (ConcurrentCopying::kEnableFromSpaceAccountingCheck) {
@@ -496,9 +514,7 @@
       cc->from_space_num_bytes_at_first_pause_ = cc->region_space_->GetBytesAllocated();
     }
     cc->is_marking_ = true;
-    cc->mark_stack_mode_.store(ConcurrentCopying::kMarkStackModeThreadLocal,
-                               std::memory_order_relaxed);
-    if (kIsDebugBuild && !cc->young_gen_) {
+    if (kIsDebugBuild && !kEnableGenerationalConcurrentCopyingCollection) {
       cc->region_space_->AssertAllRegionLiveBytesZeroOrCleared();
     }
     if (UNLIKELY(Runtime::Current()->IsActiveTransaction())) {
@@ -860,13 +876,483 @@
   ConcurrentCopying* const collector_;
 };
 
-// Concurrently mark roots that are guarded by read barriers and process the mark stack.
+template <bool kAtomicTestAndSet>
+class ConcurrentCopying::CaptureRootsForMarkingVisitor : public RootVisitor {
+ public:
+  explicit CaptureRootsForMarkingVisitor(ConcurrentCopying* cc, Thread* self)
+      : collector_(cc), self_(self) {}
+
+  void VisitRoots(mirror::Object*** roots,
+                  size_t count,
+                  const RootInfo& info ATTRIBUTE_UNUSED) override
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    for (size_t i = 0; i < count; ++i) {
+      mirror::Object** root = roots[i];
+      mirror::Object* ref = *root;
+      if (ref != nullptr && !collector_->TestAndSetMarkBitForRef<kAtomicTestAndSet>(ref)) {
+        collector_->PushOntoMarkStack(self_, ref);
+      }
+    }
+  }
+
+  void VisitRoots(mirror::CompressedReference<mirror::Object>** roots,
+                  size_t count,
+                  const RootInfo& info ATTRIBUTE_UNUSED) override
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    for (size_t i = 0; i < count; ++i) {
+      mirror::CompressedReference<mirror::Object>* const root = roots[i];
+      if (!root->IsNull()) {
+        mirror::Object* ref = root->AsMirrorPtr();
+        if (!collector_->TestAndSetMarkBitForRef<kAtomicTestAndSet>(ref)) {
+          collector_->PushOntoMarkStack(self_, ref);
+        }
+      }
+    }
+  }
+
+ private:
+  ConcurrentCopying* const collector_;
+  Thread* const self_;
+};
+
+class ConcurrentCopying::RevokeThreadLocalMarkStackCheckpoint : public Closure {
+ public:
+  RevokeThreadLocalMarkStackCheckpoint(ConcurrentCopying* concurrent_copying,
+                                       bool disable_weak_ref_access)
+      : concurrent_copying_(concurrent_copying),
+        disable_weak_ref_access_(disable_weak_ref_access) {
+  }
+
+  void Run(Thread* thread) override NO_THREAD_SAFETY_ANALYSIS {
+    // Note: self is not necessarily equal to thread since thread may be suspended.
+    Thread* const self = Thread::Current();
+    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
+        << thread->GetState() << " thread " << thread << " self " << self;
+    // Revoke thread local mark stacks.
+    accounting::AtomicStack<mirror::Object>* tl_mark_stack = thread->GetThreadLocalMarkStack();
+    if (tl_mark_stack != nullptr) {
+      MutexLock mu(self, concurrent_copying_->mark_stack_lock_);
+      concurrent_copying_->revoked_mark_stacks_.push_back(tl_mark_stack);
+      thread->SetThreadLocalMarkStack(nullptr);
+    }
+    // Disable weak ref access.
+    if (disable_weak_ref_access_) {
+      thread->SetWeakRefAccessEnabled(false);
+    }
+    // If thread is a running mutator, then act on behalf of the garbage collector.
+    // See the code in ThreadList::RunCheckpoint.
+    concurrent_copying_->GetBarrier().Pass(self);
+  }
+
+ protected:
+  ConcurrentCopying* const concurrent_copying_;
+
+ private:
+  const bool disable_weak_ref_access_;
+};
+
+class ConcurrentCopying::CaptureThreadRootsForMarkingAndCheckpoint :
+  public RevokeThreadLocalMarkStackCheckpoint {
+ public:
+  explicit CaptureThreadRootsForMarkingAndCheckpoint(ConcurrentCopying* cc) :
+    RevokeThreadLocalMarkStackCheckpoint(cc, /* disable_weak_ref_access */ false) {}
+
+  void Run(Thread* thread) override
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    Thread* const self = Thread::Current();
+    ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
+    // We can use the non-CAS VisitRoots functions below because we update thread-local GC roots
+    // only.
+    CaptureRootsForMarkingVisitor</*kAtomicTestAndSet*/ true> visitor(concurrent_copying_, self);
+    thread->VisitRoots(&visitor, kVisitRootFlagAllRoots);
+    // Barrier handling is done in the base class' Run() below.
+    RevokeThreadLocalMarkStackCheckpoint::Run(thread);
+  }
+};
+
+void ConcurrentCopying::CaptureThreadRootsForMarking() {
+  TimingLogger::ScopedTiming split("CaptureThreadRootsForMarking", GetTimings());
+  if (kVerboseMode) {
+    LOG(INFO) << "time=" << region_space_->Time();
+    region_space_->DumpNonFreeRegions(LOG_STREAM(INFO));
+  }
+  Thread* const self = Thread::Current();
+  CaptureThreadRootsForMarkingAndCheckpoint check_point(this);
+  ThreadList* thread_list = Runtime::Current()->GetThreadList();
+  gc_barrier_->Init(self, 0);
+  size_t barrier_count = thread_list->RunCheckpoint(&check_point, /* callback */ nullptr);
+  // If there are no threads to wait which implys that all the checkpoint functions are finished,
+  // then no need to release the mutator lock.
+  if (barrier_count == 0) {
+    return;
+  }
+  Locks::mutator_lock_->SharedUnlock(self);
+  {
+    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
+    gc_barrier_->Increment(self, barrier_count);
+  }
+  Locks::mutator_lock_->SharedLock(self);
+  if (kVerboseMode) {
+    LOG(INFO) << "time=" << region_space_->Time();
+    region_space_->DumpNonFreeRegions(LOG_STREAM(INFO));
+    LOG(INFO) << "GC end of CaptureThreadRootsForMarking";
+  }
+}
+
+// Used to scan ref fields of an object.
+template <bool kHandleInterRegionRefs>
+class ConcurrentCopying::ComputeLiveBytesAndMarkRefFieldsVisitor {
+ public:
+  explicit ComputeLiveBytesAndMarkRefFieldsVisitor(ConcurrentCopying* collector,
+                                                   size_t obj_region_idx)
+      : collector_(collector),
+      obj_region_idx_(obj_region_idx),
+      contains_inter_region_idx_(false) {}
+
+  void operator()(mirror::Object* obj, MemberOffset offset, bool /* is_static */) const
+      ALWAYS_INLINE
+      REQUIRES_SHARED(Locks::mutator_lock_)
+      REQUIRES_SHARED(Locks::heap_bitmap_lock_) {
+    DCHECK_EQ(collector_->RegionSpace()->RegionIdxForRef(obj), obj_region_idx_);
+    DCHECK(kHandleInterRegionRefs || collector_->immune_spaces_.ContainsObject(obj));
+    CheckReference(obj->GetFieldObject<mirror::Object, kVerifyNone, kWithoutReadBarrier>(offset));
+  }
+
+  void operator()(ObjPtr<mirror::Class> klass, ObjPtr<mirror::Reference> ref) const
+      REQUIRES_SHARED(Locks::mutator_lock_) ALWAYS_INLINE {
+    DCHECK(klass->IsTypeOfReferenceClass());
+    // If the referent is not null, then we must re-visit the object during
+    // copying phase to enqueue it for delayed processing and setting
+    // read-barrier state to gray to ensure that call to GetReferent() triggers
+    // the read-barrier. We use same data structure that is used to remember
+    // objects with inter-region refs for this purpose too.
+    if (kHandleInterRegionRefs
+        && !contains_inter_region_idx_
+        && ref->AsReference()->GetReferent<kWithoutReadBarrier>() != nullptr) {
+      contains_inter_region_idx_ = true;
+    }
+  }
+
+  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
+      ALWAYS_INLINE
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    if (!root->IsNull()) {
+      VisitRoot(root);
+    }
+  }
+
+  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
+      ALWAYS_INLINE
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    CheckReference(root->AsMirrorPtr());
+  }
+
+  bool ContainsInterRegionRefs() const ALWAYS_INLINE REQUIRES_SHARED(Locks::mutator_lock_) {
+    return contains_inter_region_idx_;
+  }
+
+ private:
+  void CheckReference(mirror::Object* ref) const
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    if (ref == nullptr) {
+      // Nothing to do.
+      return;
+    }
+    if (!collector_->TestAndSetMarkBitForRef(ref)) {
+      collector_->PushOntoLocalMarkStack(ref);
+    }
+    if (kHandleInterRegionRefs && !contains_inter_region_idx_) {
+      size_t ref_region_idx = collector_->RegionSpace()->RegionIdxForRef(ref);
+      // If a region-space object refers to an outside object, we will have a
+      // mismatch of region idx, but the object need not be re-visited in
+      // copying phase.
+      if (ref_region_idx != static_cast<size_t>(-1) && obj_region_idx_ != ref_region_idx) {
+        contains_inter_region_idx_ = true;
+      }
+    }
+  }
+
+  ConcurrentCopying* const collector_;
+  const size_t obj_region_idx_;
+  mutable bool contains_inter_region_idx_;
+};
+
+void ConcurrentCopying::AddLiveBytesAndScanRef(mirror::Object* ref) {
+  DCHECK(ref != nullptr);
+  DCHECK(!immune_spaces_.ContainsObject(ref));
+  DCHECK(TestMarkBitmapForRef(ref));
+  size_t obj_region_idx = static_cast<size_t>(-1);
+  if (LIKELY(region_space_->HasAddress(ref))) {
+    obj_region_idx = region_space_->RegionIdxForRefUnchecked(ref);
+    // Add live bytes to the corresponding region
+    if (!region_space_->IsRegionNewlyAllocated(obj_region_idx)) {
+      // Newly Allocated regions are always chosen for evacuation. So no need
+      // to update live_bytes_.
+      size_t obj_size = ref->SizeOf<kDefaultVerifyFlags>();
+      size_t alloc_size = RoundUp(obj_size, space::RegionSpace::kAlignment);
+      region_space_->AddLiveBytes(ref, alloc_size);
+    }
+  }
+  ComputeLiveBytesAndMarkRefFieldsVisitor</*kHandleInterRegionRefs*/ true>
+      visitor(this, obj_region_idx);
+  ref->VisitReferences</*kVisitNativeRoots=*/true, kDefaultVerifyFlags, kWithoutReadBarrier>(
+      visitor, visitor);
+  // Mark the corresponding card dirty if the object contains any
+  // inter-region reference.
+  if (visitor.ContainsInterRegionRefs()) {
+    heap_->GetCardTable()->MarkCard(ref);
+  }
+}
+
+template <bool kAtomic>
+bool ConcurrentCopying::TestAndSetMarkBitForRef(mirror::Object* ref) {
+  accounting::ContinuousSpaceBitmap* bitmap = nullptr;
+  accounting::LargeObjectBitmap* los_bitmap = nullptr;
+  if (LIKELY(region_space_->HasAddress(ref))) {
+    bitmap = region_space_bitmap_;
+  } else if (heap_->GetNonMovingSpace()->HasAddress(ref)) {
+    bitmap = heap_->GetNonMovingSpace()->GetMarkBitmap();
+  } else if (immune_spaces_.ContainsObject(ref)) {
+    // References to immune space objects are always live.
+    DCHECK(heap_mark_bitmap_->GetContinuousSpaceBitmap(ref)->Test(ref));
+    return true;
+  } else {
+    // Should be a large object. Must be page aligned and the LOS must exist.
+    if (kIsDebugBuild
+        && (!IsAligned<kPageSize>(ref) || heap_->GetLargeObjectsSpace() == nullptr)) {
+      // It must be heap corruption. Remove memory protection and dump data.
+      region_space_->Unprotect();
+      heap_->GetVerification()->LogHeapCorruption(/* obj */ nullptr,
+                                                  MemberOffset(0),
+                                                  ref,
+                                                  /* fatal */ true);
+    }
+    los_bitmap = heap_->GetLargeObjectsSpace()->GetMarkBitmap();
+  }
+  if (kAtomic) {
+    return (bitmap != nullptr) ? bitmap->AtomicTestAndSet(ref) : los_bitmap->AtomicTestAndSet(ref);
+  } else {
+    return (bitmap != nullptr) ? bitmap->Set(ref) : los_bitmap->Set(ref);
+  }
+}
+
+bool ConcurrentCopying::TestMarkBitmapForRef(mirror::Object* ref) {
+  if (LIKELY(region_space_->HasAddress(ref))) {
+    return region_space_bitmap_->Test(ref);
+  } else if (heap_->GetNonMovingSpace()->HasAddress(ref)) {
+    return heap_->GetNonMovingSpace()->GetMarkBitmap()->Test(ref);
+  } else if (immune_spaces_.ContainsObject(ref)) {
+    // References to immune space objects are always live.
+    DCHECK(heap_mark_bitmap_->GetContinuousSpaceBitmap(ref)->Test(ref));
+    return true;
+  } else {
+    // Should be a large object. Must be page aligned and the LOS must exist.
+    if (kIsDebugBuild
+        && (!IsAligned<kPageSize>(ref) || heap_->GetLargeObjectsSpace() == nullptr)) {
+      // It must be heap corruption. Remove memory protection and dump data.
+      region_space_->Unprotect();
+      heap_->GetVerification()->LogHeapCorruption(/* obj */ nullptr,
+                                                  MemberOffset(0),
+                                                  ref,
+                                                  /* fatal */ true);
+    }
+    return heap_->GetLargeObjectsSpace()->GetMarkBitmap()->Test(ref);
+  }
+}
+
+void ConcurrentCopying::PushOntoLocalMarkStack(mirror::Object* ref) {
+  if (kIsDebugBuild) {
+    Thread *self = Thread::Current();
+    DCHECK_EQ(thread_running_gc_, self);
+    DCHECK(self->GetThreadLocalMarkStack() == nullptr);
+  }
+  DCHECK_EQ(mark_stack_mode_.load(std::memory_order_relaxed), kMarkStackModeThreadLocal);
+  gc_mark_stack_->PushBack(ref);
+}
+
+void ConcurrentCopying::ProcessMarkStackForMarkingAndComputeLiveBytes() {
+  // Process thread-local mark stack containing thread roots
+  ProcessThreadLocalMarkStacks(/* disable_weak_ref_access */ false,
+                               /* checkpoint_callback */ nullptr,
+                               [this] (mirror::Object* ref)
+                                   REQUIRES_SHARED(Locks::mutator_lock_) {
+                                 AddLiveBytesAndScanRef(ref);
+                               });
+
+  while (!gc_mark_stack_->IsEmpty()) {
+    mirror::Object* ref = gc_mark_stack_->PopBack();
+    AddLiveBytesAndScanRef(ref);
+  }
+}
+
+class ConcurrentCopying::ImmuneSpaceCaptureRefsVisitor {
+ public:
+  explicit ImmuneSpaceCaptureRefsVisitor(ConcurrentCopying* cc) : collector_(cc) {}
+
+  ALWAYS_INLINE void operator()(mirror::Object* obj) const REQUIRES_SHARED(Locks::mutator_lock_) {
+    ComputeLiveBytesAndMarkRefFieldsVisitor</*kHandleInterRegionRefs*/ false>
+        visitor(collector_, /*obj_region_idx*/ static_cast<size_t>(-1));
+    obj->VisitReferences</*kVisitNativeRoots=*/true, kDefaultVerifyFlags, kWithoutReadBarrier>(
+        visitor, visitor);
+  }
+
+  static void Callback(mirror::Object* obj, void* arg) REQUIRES_SHARED(Locks::mutator_lock_) {
+    reinterpret_cast<ImmuneSpaceScanObjVisitor*>(arg)->operator()(obj);
+  }
+
+ private:
+  ConcurrentCopying* const collector_;
+};
+
+/* Invariants for two-phase CC
+ * ===========================
+ * A) Definitions
+ * ---------------
+ * 1) Black: marked in bitmap, rb_state is non-gray, and not in mark stack
+ * 2) Black-clean: marked in bitmap, and corresponding card is clean/aged
+ * 3) Black-dirty: marked in bitmap, and corresponding card is dirty
+ * 4) Gray: marked in bitmap, and exists in mark stack
+ * 5) Gray-dirty: marked in bitmap, rb_state is gray, corresponding card is
+ *    dirty, and exists in mark stack
+ * 6) White: unmarked in bitmap, rb_state is non-gray, and not in mark stack
+ *
+ * B) Before marking phase
+ * -----------------------
+ * 1) All objects are white
+ * 2) Cards are either clean or aged (cannot be asserted without a STW pause)
+ * 3) Mark bitmap is cleared
+ * 4) Mark stack is empty
+ *
+ * C) During marking phase
+ * ------------------------
+ * 1) If a black object holds an inter-region or white reference, then its
+ *    corresponding card is dirty. In other words, it changes from being
+ *    black-clean to black-dirty
+ * 2) No black-clean object points to a white object
+ *
+ * D) After marking phase
+ * -----------------------
+ * 1) There are no gray objects
+ * 2) All newly allocated objects are in from space
+ * 3) No white object can be reachable, directly or otherwise, from a
+ *    black-clean object
+ *
+ * E) During copying phase
+ * ------------------------
+ * 1) Mutators cannot observe white and black-dirty objects
+ * 2) New allocations are in to-space (newly allocated regions are part of to-space)
+ * 3) An object in mark stack must have its rb_state = Gray
+ *
+ * F) During card table scan
+ * --------------------------
+ * 1) Referents corresponding to root references are gray or in to-space
+ * 2) Every path from an object that is read or written by a mutator during
+ *    this period to a dirty black object goes through some gray object.
+ *    Mutators preserve this by graying black objects as needed during this
+ *    period. Ensures that a mutator never encounters a black dirty object.
+ *
+ * G) After card table scan
+ * ------------------------
+ * 1) There are no black-dirty objects
+ * 2) Referents corresponding to root references are gray, black-clean or in
+ *    to-space
+ *
+ * H) After copying phase
+ * -----------------------
+ * 1) Mark stack is empty
+ * 2) No references into evacuated from-space
+ * 3) No reference to an object which is unmarked and is also not in newly
+ *    allocated region. In other words, no reference to white objects.
+*/
+
 void ConcurrentCopying::MarkingPhase() {
   TimingLogger::ScopedTiming split("MarkingPhase", GetTimings());
   if (kVerboseMode) {
     LOG(INFO) << "GC MarkingPhase";
   }
+  accounting::CardTable* const card_table = heap_->GetCardTable();
+  Thread* const self = Thread::Current();
+  // Clear live_bytes_ of every non-free region, except the ones that are newly
+  // allocated.
+  region_space_->SetAllRegionLiveBytesZero();
+  if (kIsDebugBuild) {
+    region_space_->AssertAllRegionLiveBytesZeroOrCleared();
+  }
+  // Scan immune spaces
+  {
+    TimingLogger::ScopedTiming split2("ScanImmuneSpaces", GetTimings());
+    for (auto& space : immune_spaces_.GetSpaces()) {
+      DCHECK(space->IsImageSpace() || space->IsZygoteSpace());
+      accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap();
+      accounting::ModUnionTable* table = heap_->FindModUnionTableFromSpace(space);
+      ImmuneSpaceCaptureRefsVisitor visitor(this);
+      if (table != nullptr) {
+        table->VisitObjects(ImmuneSpaceCaptureRefsVisitor::Callback, &visitor);
+      } else {
+        WriterMutexLock rmu(Thread::Current(), *Locks::heap_bitmap_lock_);
+        card_table->Scan<false>(
+            live_bitmap,
+            space->Begin(),
+            space->Limit(),
+            visitor,
+            accounting::CardTable::kCardDirty - 1);
+      }
+    }
+  }
+  // Scan runtime roots
+  {
+    TimingLogger::ScopedTiming split2("VisitConcurrentRoots", GetTimings());
+    CaptureRootsForMarkingVisitor visitor(this, self);
+    Runtime::Current()->VisitConcurrentRoots(&visitor, kVisitRootFlagAllRoots);
+  }
+  {
+    // TODO: don't visit the transaction roots if it's not active.
+    TimingLogger::ScopedTiming split2("VisitNonThreadRoots", GetTimings());
+    CaptureRootsForMarkingVisitor visitor(this, self);
+    Runtime::Current()->VisitNonThreadRoots(&visitor);
+  }
+  // Capture thread roots
+  CaptureThreadRootsForMarking();
+  // Process mark stack
+  ProcessMarkStackForMarkingAndComputeLiveBytes();
+
+  // Age the cards.
+  for (space::ContinuousSpace* space : GetHeap()->GetContinuousSpaces()) {
+    if (space->IsImageSpace() || space->IsZygoteSpace()) {
+      // Image and zygote spaces are already handled since we gray the objects in the pause.
+      continue;
+    }
+    card_table->ModifyCardsAtomic(space->Begin(), space->End(), AgeCardVisitor(), VoidFunctor());
+  }
+
+  if (kVerboseMode) {
+    LOG(INFO) << "GC end of MarkingPhase";
+  }
+}
+
+template <bool kNoUnEvac>
+void ConcurrentCopying::ScanDirtyObject(mirror::Object* obj) {
+  Scan<kNoUnEvac>(obj);
+  // Set the read-barrier state of a reference-type object to gray if its
+  // referent is not marked yet. This is to ensure that if GetReferent() is
+  // called, it triggers the read-barrier to process the referent before use.
+  if (UNLIKELY((obj->GetClass<kVerifyNone, kWithoutReadBarrier>()->IsTypeOfReferenceClass()))) {
+    mirror::Object* referent =
+        obj->AsReference<kVerifyNone, kWithoutReadBarrier>()->GetReferent<kWithoutReadBarrier>();
+    if (referent != nullptr && !IsInToSpace(referent)) {
+      obj->AtomicSetReadBarrierState(ReadBarrier::NonGrayState(), ReadBarrier::GrayState());
+    }
+  }
+}
+
+// Concurrently mark roots that are guarded by read barriers and process the mark stack.
+void ConcurrentCopying::CopyingPhase() {
+  TimingLogger::ScopedTiming split("CopyingPhase", GetTimings());
+  if (kVerboseMode) {
+    LOG(INFO) << "GC CopyingPhase";
+  }
   Thread* self = Thread::Current();
+  accounting::CardTable* const card_table = heap_->GetCardTable();
   if (kIsDebugBuild) {
     MutexLock mu(self, *Locks::thread_list_lock_);
     CHECK(weak_ref_access_enabled_);
@@ -879,7 +1365,7 @@
   if (kUseBakerReadBarrier) {
     gc_grays_immune_objects_ = false;
   }
-  if (kEnableGenerationalConcurrentCopyingCollection && young_gen_) {
+  if (kEnableGenerationalConcurrentCopyingCollection) {
     if (kVerboseMode) {
       LOG(INFO) << "GC ScanCardsForSpace";
     }
@@ -897,34 +1383,45 @@
         continue;
       }
       // Scan all of the objects on dirty cards in unevac from space, and non moving space. These
-      // are from previous GCs and may reference things in the from space.
+      // are from previous GCs (or from marking phase of 2-phase full GC) and may reference things
+      // in the from space.
       //
       // Note that we do not need to process the large-object space (the only discontinuous space)
       // as it contains only large string objects and large primitive array objects, that have no
       // reference to other objects, except their class. There is no need to scan these large
       // objects, as the String class and the primitive array classes are expected to never move
-      // during a minor (young-generation) collection:
+      // during a collection:
       // - In the case where we run with a boot image, these classes are part of the image space,
       //   which is an immune space.
       // - In the case where we run without a boot image, these classes are allocated in the
       //   non-moving space (see art::ClassLinker::InitWithoutImage).
-      Runtime::Current()->GetHeap()->GetCardTable()->Scan<false>(
+      card_table->Scan<false>(
           space->GetMarkBitmap(),
           space->Begin(),
           space->End(),
           [this, space](mirror::Object* obj)
               REQUIRES(Locks::heap_bitmap_lock_)
               REQUIRES_SHARED(Locks::mutator_lock_) {
-            // Don't push or gray unevac refs.
-            if (kIsDebugBuild && space == region_space_) {
-              // We may get unevac large objects.
-              if (!region_space_->IsInUnevacFromSpace(obj)) {
-                CHECK(region_space_bitmap_->Test(obj));
-                region_space_->DumpRegionForObject(LOG_STREAM(FATAL_WITHOUT_ABORT), obj);
-                LOG(FATAL) << "Scanning " << obj << " not in unevac space";
+            // TODO: This code may be refactored to avoid scanning object while
+            // done_scanning_ is false by setting rb_state to gray, and pushing the
+            // object on mark stack. However, it will also require clearing the
+            // corresponding mark-bit and, for region space objects,
+            // decrementing the object's size from the corresponding region's
+            // live_bytes.
+            if (young_gen_) {
+              // Don't push or gray unevac refs.
+              if (kIsDebugBuild && space == region_space_) {
+                // We may get unevac large objects.
+                if (!region_space_->IsInUnevacFromSpace(obj)) {
+                  CHECK(region_space_bitmap_->Test(obj));
+                  region_space_->DumpRegionForObject(LOG_STREAM(FATAL_WITHOUT_ABORT), obj);
+                  LOG(FATAL) << "Scanning " << obj << " not in unevac space";
+                }
               }
+              ScanDirtyObject</*kNoUnEvac*/ true>(obj);
+            } else if (space != region_space_ || region_space_->IsInUnevacFromSpace(obj)) {
+              ScanDirtyObject</*kNoUnEvac*/ false>(obj);
             }
-            Scan<true>(obj);
           },
           accounting::CardTable::kCardDirty - 1);
     }
@@ -947,10 +1444,13 @@
       if (kUseBakerReadBarrier && kGrayDirtyImmuneObjects && table != nullptr) {
         table->VisitObjects(ImmuneSpaceScanObjVisitor::Callback, &visitor);
       } else {
-        // TODO: Scan only the aged cards.
-        live_bitmap->VisitMarkedRange(reinterpret_cast<uintptr_t>(space->Begin()),
-                                      reinterpret_cast<uintptr_t>(space->Limit()),
-                                      visitor);
+        WriterMutexLock rmu(Thread::Current(), *Locks::heap_bitmap_lock_);
+        card_table->Scan<false>(
+            live_bitmap,
+            space->Begin(),
+            space->Limit(),
+            visitor,
+            accounting::CardTable::kCardDirty - 1);
       }
     }
   }
@@ -1059,7 +1559,7 @@
     CHECK(weak_ref_access_enabled_);
   }
   if (kVerboseMode) {
-    LOG(INFO) << "GC end of MarkingPhase";
+    LOG(INFO) << "GC end of CopyingPhase";
   }
 }
 
@@ -1419,40 +1919,6 @@
   ConcurrentCopying* const collector_;
 };
 
-class ConcurrentCopying::RevokeThreadLocalMarkStackCheckpoint : public Closure {
- public:
-  RevokeThreadLocalMarkStackCheckpoint(ConcurrentCopying* concurrent_copying,
-                                       bool disable_weak_ref_access)
-      : concurrent_copying_(concurrent_copying),
-        disable_weak_ref_access_(disable_weak_ref_access) {
-  }
-
-  void Run(Thread* thread) override NO_THREAD_SAFETY_ANALYSIS {
-    // Note: self is not necessarily equal to thread since thread may be suspended.
-    Thread* self = Thread::Current();
-    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
-        << thread->GetState() << " thread " << thread << " self " << self;
-    // Revoke thread local mark stacks.
-    accounting::AtomicStack<mirror::Object>* tl_mark_stack = thread->GetThreadLocalMarkStack();
-    if (tl_mark_stack != nullptr) {
-      MutexLock mu(self, concurrent_copying_->mark_stack_lock_);
-      concurrent_copying_->revoked_mark_stacks_.push_back(tl_mark_stack);
-      thread->SetThreadLocalMarkStack(nullptr);
-    }
-    // Disable weak ref access.
-    if (disable_weak_ref_access_) {
-      thread->SetWeakRefAccessEnabled(false);
-    }
-    // If thread is a running mutator, then act on behalf of the garbage collector.
-    // See the code in ThreadList::RunCheckpoint.
-    concurrent_copying_->GetBarrier().Pass(self);
-  }
-
- private:
-  ConcurrentCopying* const concurrent_copying_;
-  const bool disable_weak_ref_access_;
-};
-
 void ConcurrentCopying::RevokeThreadLocalMarkStacks(bool disable_weak_ref_access,
                                                     Closure* checkpoint_callback) {
   Thread* self = Thread::Current();
@@ -1510,7 +1976,11 @@
   if (mark_stack_mode == kMarkStackModeThreadLocal) {
     // Process the thread-local mark stacks and the GC mark stack.
     count += ProcessThreadLocalMarkStacks(/* disable_weak_ref_access= */ false,
-                                          /* checkpoint_callback= */ nullptr);
+                                          /* checkpoint_callback= */ nullptr,
+                                          [this] (mirror::Object* ref)
+                                              REQUIRES_SHARED(Locks::mutator_lock_) {
+                                            ProcessMarkStackRef(ref);
+                                          });
     while (!gc_mark_stack_->IsEmpty()) {
       mirror::Object* to_ref = gc_mark_stack_->PopBack();
       ProcessMarkStackRef(to_ref);
@@ -1566,8 +2036,10 @@
   return count == 0;
 }
 
+template <typename Processor>
 size_t ConcurrentCopying::ProcessThreadLocalMarkStacks(bool disable_weak_ref_access,
-                                                       Closure* checkpoint_callback) {
+                                                       Closure* checkpoint_callback,
+                                                       const Processor& processor) {
   // Run a checkpoint to collect all thread local mark stacks and iterate over them all.
   RevokeThreadLocalMarkStacks(disable_weak_ref_access, checkpoint_callback);
   size_t count = 0;
@@ -1581,7 +2053,7 @@
   for (accounting::AtomicStack<mirror::Object>* mark_stack : mark_stacks) {
     for (StackReference<mirror::Object>* p = mark_stack->Begin(); p != mark_stack->End(); ++p) {
       mirror::Object* to_ref = p->AsMirrorPtr();
-      ProcessMarkStackRef(to_ref);
+      processor(to_ref);
       ++count;
     }
     {
@@ -1632,6 +2104,12 @@
         perform_scan = true;
         // Only add to the live bytes if the object was not already marked and we are not the young
         // GC.
+        // Why add live bytes even after 2-phase GC?
+        // We need to ensure that if there is a unevac region with any live
+        // objects, then its live_bytes must be non-zero. Otherwise,
+        // ClearFromSpace() will clear the region. Considering, that we may skip
+        // live objects during marking phase of 2-phase GC, we have to take care
+        // of such objects here.
         add_to_live_bytes = true;
       }
       break;
@@ -1773,7 +2251,12 @@
   DisableWeakRefAccessCallback dwrac(this);
   // Process the thread local mark stacks one last time after switching to the shared mark stack
   // mode and disable weak ref accesses.
-  ProcessThreadLocalMarkStacks(/* disable_weak_ref_access= */ true, &dwrac);
+  ProcessThreadLocalMarkStacks(/* disable_weak_ref_access= */ true,
+                               &dwrac,
+                               [this] (mirror::Object* ref)
+                                   REQUIRES_SHARED(Locks::mutator_lock_) {
+                                 ProcessMarkStackRef(ref);
+                               });
   if (kVerboseMode) {
     LOG(INFO) << "Switched to shared mark stack mode and disabled weak ref access";
   }
@@ -2039,7 +2522,7 @@
     uint64_t cleared_objects;
     {
       TimingLogger::ScopedTiming split4("ClearFromSpace", GetTimings());
-      region_space_->ClearFromSpace(&cleared_bytes, &cleared_objects);
+      region_space_->ClearFromSpace(&cleared_bytes, &cleared_objects, /*clear_bitmap*/ !young_gen_);
       // `cleared_bytes` and `cleared_objects` may be greater than the from space equivalents since
       // RegionSpace::ClearFromSpace may clear empty unevac regions.
       CHECK_GE(cleared_bytes, from_bytes);
@@ -2348,7 +2831,7 @@
   DCHECK(!immune_spaces_.ContainsObject(from_ref)) << "ref=" << from_ref;
   if (kUseBakerReadBarrier && from_ref->GetReadBarrierStateAcquire() == ReadBarrier::GrayState()) {
     return true;
-  } else if (!(kEnableGenerationalConcurrentCopyingCollection && young_gen_)
+  } else if (!kEnableGenerationalConcurrentCopyingCollection
              || done_scanning_.load(std::memory_order_acquire)) {
     // Read the comment in IsMarkedInUnevacFromSpace()
     accounting::ContinuousSpaceBitmap* mark_bitmap = heap_->GetNonMovingSpace()->GetMarkBitmap();
@@ -2939,7 +3422,7 @@
     los_bitmap = heap_->GetLargeObjectsSpace()->GetMarkBitmap();
     DCHECK(los_bitmap->HasAddress(ref));
   }
-  if (kEnableGenerationalConcurrentCopyingCollection && young_gen_) {
+  if (kEnableGenerationalConcurrentCopyingCollection) {
     // The sticky-bit CC collector is only compatible with Baker-style read barriers.
     DCHECK(kUseBakerReadBarrier);
     // Not done scanning, use AtomicSetReadBarrierPointer.
diff --git a/runtime/gc/collector/concurrent_copying.h b/runtime/gc/collector/concurrent_copying.h
index 237e070..4442ad5 100644
--- a/runtime/gc/collector/concurrent_copying.h
+++ b/runtime/gc/collector/concurrent_copying.h
@@ -79,6 +79,8 @@
   void InitializePhase() REQUIRES_SHARED(Locks::mutator_lock_)
       REQUIRES(!mark_stack_lock_, !immune_gray_stack_lock_);
   void MarkingPhase() REQUIRES_SHARED(Locks::mutator_lock_)
+      REQUIRES(!mark_stack_lock_);
+  void CopyingPhase() REQUIRES_SHARED(Locks::mutator_lock_)
       REQUIRES(!mark_stack_lock_, !skipped_blocks_lock_, !immune_gray_stack_lock_);
   void ReclaimPhase() REQUIRES_SHARED(Locks::mutator_lock_) REQUIRES(!mark_stack_lock_);
   void FinishPhase() REQUIRES(!mark_stack_lock_,
@@ -161,6 +163,13 @@
   template <bool kNoUnEvac>
   void Scan(mirror::Object* to_ref) REQUIRES_SHARED(Locks::mutator_lock_)
       REQUIRES(!mark_stack_lock_);
+  // Scan the reference fields of object 'obj' in the dirty cards during
+  // card-table scan. In addition to visiting the references, it also sets the
+  // read-barrier state to gray for Reference-type objects to ensure that
+  // GetReferent() called on these objects calls the read-barrier on the referent.
+  template <bool kNoUnEvac>
+  void ScanDirtyObject(mirror::Object* obj) REQUIRES_SHARED(Locks::mutator_lock_)
+      REQUIRES(!mark_stack_lock_);
   // Process a field.
   template <bool kNoUnEvac>
   void Process(mirror::Object* obj, MemberOffset offset)
@@ -198,7 +207,10 @@
   void VerifyNoMissingCardMarks()
       REQUIRES(Locks::mutator_lock_)
       REQUIRES(!mark_stack_lock_);
-  size_t ProcessThreadLocalMarkStacks(bool disable_weak_ref_access, Closure* checkpoint_callback)
+  template <typename Processor>
+  size_t ProcessThreadLocalMarkStacks(bool disable_weak_ref_access,
+                                      Closure* checkpoint_callback,
+                                      const Processor& processor)
       REQUIRES_SHARED(Locks::mutator_lock_) REQUIRES(!mark_stack_lock_);
   void RevokeThreadLocalMarkStacks(bool disable_weak_ref_access, Closure* checkpoint_callback)
       REQUIRES_SHARED(Locks::mutator_lock_);
@@ -295,6 +307,15 @@
   // Set the read barrier mark entrypoints to non-null.
   void ActivateReadBarrierEntrypoints();
 
+  void CaptureThreadRootsForMarking() REQUIRES_SHARED(Locks::mutator_lock_);
+  void AddLiveBytesAndScanRef(mirror::Object* ref) REQUIRES_SHARED(Locks::mutator_lock_);
+  bool TestMarkBitmapForRef(mirror::Object* ref) REQUIRES_SHARED(Locks::mutator_lock_);
+  template <bool kAtomic = false>
+  bool TestAndSetMarkBitForRef(mirror::Object* ref) REQUIRES_SHARED(Locks::mutator_lock_);
+  void PushOntoLocalMarkStack(mirror::Object* ref) REQUIRES_SHARED(Locks::mutator_lock_);
+  void ProcessMarkStackForMarkingAndComputeLiveBytes() REQUIRES_SHARED(Locks::mutator_lock_)
+      REQUIRES(!mark_stack_lock_);
+
   space::RegionSpace* region_space_;      // The underlying region space.
   std::unique_ptr<Barrier> gc_barrier_;
   std::unique_ptr<accounting::ObjectStack> gc_mark_stack_;
@@ -375,7 +396,7 @@
   // Generational "sticky", only trace through dirty objects in region space.
   const bool young_gen_;
   // If true, the GC thread is done scanning marked objects on dirty and aged
-  // card (see ConcurrentCopying::MarkingPhase).
+  // card (see ConcurrentCopying::CopyingPhase).
   Atomic<bool> done_scanning_;
 
   // The skipped blocks are memory blocks/chucks that were copies of
@@ -441,6 +462,10 @@
   class VerifyNoFromSpaceRefsFieldVisitor;
   class VerifyNoFromSpaceRefsVisitor;
   class VerifyNoMissingCardMarkVisitor;
+  class ImmuneSpaceCaptureRefsVisitor;
+  template <bool kAtomicTestAndSet = false> class CaptureRootsForMarkingVisitor;
+  class CaptureThreadRootsForMarkingAndCheckpoint;
+  template <bool kHandleInterRegionRefs> class ComputeLiveBytesAndMarkRefFieldsVisitor;
 
   DISALLOW_IMPLICIT_CONSTRUCTORS(ConcurrentCopying);
 };
diff --git a/runtime/gc/heap-inl.h b/runtime/gc/heap-inl.h
index 9e1ba35..1c09b5c 100644
--- a/runtime/gc/heap-inl.h
+++ b/runtime/gc/heap-inl.h
@@ -214,7 +214,7 @@
   if (AllocatorMayHaveConcurrentGC(allocator) && IsGcConcurrent()) {
     // New_num_bytes_allocated is zero if we didn't update num_bytes_allocated_.
     // That's fine.
-    CheckConcurrentGC(self, new_num_bytes_allocated, &obj);
+    CheckConcurrentGCForJava(self, new_num_bytes_allocated, &obj);
   }
   VerifyObject(obj);
   self->VerifyStack();
@@ -254,8 +254,8 @@
                                            size_t* bytes_allocated,
                                            size_t* usable_size,
                                            size_t* bytes_tl_bulk_allocated) {
-  if (allocator_type != kAllocatorTypeTLAB &&
-      allocator_type != kAllocatorTypeRegionTLAB &&
+  if (allocator_type != kAllocatorTypeRegionTLAB &&
+      allocator_type != kAllocatorTypeTLAB &&
       allocator_type != kAllocatorTypeRosAlloc &&
       UNLIKELY(IsOutOfMemoryOnAllocation(allocator_type, alloc_size, kGrow))) {
     return nullptr;
@@ -396,30 +396,46 @@
 inline bool Heap::IsOutOfMemoryOnAllocation(AllocatorType allocator_type,
                                             size_t alloc_size,
                                             bool grow) {
-  size_t new_footprint = num_bytes_allocated_.load(std::memory_order_relaxed) + alloc_size;
-  if (UNLIKELY(new_footprint > max_allowed_footprint_)) {
-    if (UNLIKELY(new_footprint > growth_limit_)) {
+  size_t old_target = target_footprint_.load(std::memory_order_relaxed);
+  while (true) {
+    size_t old_allocated = num_bytes_allocated_.load(std::memory_order_relaxed);
+    size_t new_footprint = old_allocated + alloc_size;
+    // Tests against heap limits are inherently approximate, since multiple allocations may
+    // race, and this is not atomic with the allocation.
+    if (UNLIKELY(new_footprint <= old_target)) {
+      return false;
+    } else if (UNLIKELY(new_footprint > growth_limit_)) {
       return true;
     }
-    if (!AllocatorMayHaveConcurrentGC(allocator_type) || !IsGcConcurrent()) {
-      if (!grow) {
+    // We are between target_footprint_ and growth_limit_ .
+    if (AllocatorMayHaveConcurrentGC(allocator_type) && IsGcConcurrent()) {
+      return false;
+    } else {
+      if (grow) {
+        if (target_footprint_.compare_exchange_weak(/*inout ref*/old_target, new_footprint,
+                                                    std::memory_order_relaxed)) {
+          VlogHeapGrowth(old_target, new_footprint, alloc_size);
+          return false;
+        }  // else try again.
+      } else {
         return true;
       }
-      // TODO: Grow for allocation is racy, fix it.
-      VlogHeapGrowth(max_allowed_footprint_, new_footprint, alloc_size);
-      max_allowed_footprint_ = new_footprint;
     }
   }
-  return false;
 }
 
-// Request a GC if new_num_bytes_allocated is sufficiently large.
-// A call with new_num_bytes_allocated == 0 is a fast no-op.
-inline void Heap::CheckConcurrentGC(Thread* self,
+inline bool Heap::ShouldConcurrentGCForJava(size_t new_num_bytes_allocated) {
+  // For a Java allocation, we only check whether the number of Java allocated bytes excceeds a
+  // threshold. By not considering native allocation here, we (a) ensure that Java heap bounds are
+  // maintained, and (b) reduce the cost of the check here.
+  return new_num_bytes_allocated >= concurrent_start_bytes_;
+}
+
+inline void Heap::CheckConcurrentGCForJava(Thread* self,
                                     size_t new_num_bytes_allocated,
                                     ObjPtr<mirror::Object>* obj) {
-  if (UNLIKELY(new_num_bytes_allocated >= concurrent_start_bytes_)) {
-    RequestConcurrentGCAndSaveObject(self, false, obj);
+  if (UNLIKELY(ShouldConcurrentGCForJava(new_num_bytes_allocated))) {
+    RequestConcurrentGCAndSaveObject(self, false /* force_full */, obj);
   }
 }
 
diff --git a/runtime/gc/heap.cc b/runtime/gc/heap.cc
index dc79731..d47aca9 100644
--- a/runtime/gc/heap.cc
+++ b/runtime/gc/heap.cc
@@ -17,6 +17,9 @@
 #include "heap.h"
 
 #include <limits>
+#if defined(__BIONIC__) || defined(__GLIBC__)
+#include <malloc.h>  // For mallinfo()
+#endif
 #include <memory>
 #include <vector>
 
@@ -187,7 +190,7 @@
            bool low_memory_mode,
            size_t long_pause_log_threshold,
            size_t long_gc_log_threshold,
-           bool ignore_max_footprint,
+           bool ignore_target_footprint,
            bool use_tlab,
            bool verify_pre_gc_heap,
            bool verify_pre_sweeping_heap,
@@ -218,7 +221,7 @@
       post_gc_last_process_cpu_time_ns_(process_cpu_start_time_ns_),
       pre_gc_weighted_allocated_bytes_(0.0),
       post_gc_weighted_allocated_bytes_(0.0),
-      ignore_max_footprint_(ignore_max_footprint),
+      ignore_target_footprint_(ignore_target_footprint),
       zygote_creation_lock_("zygote creation lock", kZygoteCreationLock),
       zygote_space_(nullptr),
       large_object_threshold_(large_object_threshold),
@@ -231,13 +234,14 @@
       next_gc_type_(collector::kGcTypePartial),
       capacity_(capacity),
       growth_limit_(growth_limit),
-      max_allowed_footprint_(initial_size),
+      target_footprint_(initial_size),
       concurrent_start_bytes_(std::numeric_limits<size_t>::max()),
       total_bytes_freed_ever_(0),
       total_objects_freed_ever_(0),
       num_bytes_allocated_(0),
-      new_native_bytes_allocated_(0),
+      native_bytes_registered_(0),
       old_native_bytes_allocated_(0),
+      native_objects_notified_(0),
       num_bytes_freed_revoke_(0),
       verify_missing_card_marks_(false),
       verify_system_weaks_(false),
@@ -616,11 +620,11 @@
   task_processor_.reset(new TaskProcessor());
   reference_processor_.reset(new ReferenceProcessor());
   pending_task_lock_ = new Mutex("Pending task lock");
-  if (ignore_max_footprint_) {
+  if (ignore_target_footprint_) {
     SetIdealFootprint(std::numeric_limits<size_t>::max());
     concurrent_start_bytes_ = std::numeric_limits<size_t>::max();
   }
-  CHECK_NE(max_allowed_footprint_, 0U);
+  CHECK_NE(target_footprint_.load(std::memory_order_relaxed), 0U);
   // Create our garbage collectors.
   for (size_t i = 0; i < 2; ++i) {
     const bool concurrent = i != 0;
@@ -1158,10 +1162,11 @@
     rosalloc_space_->DumpStats(os);
   }
 
-  os << "Registered native bytes allocated: "
-     << (old_native_bytes_allocated_.load(std::memory_order_relaxed) +
-         new_native_bytes_allocated_.load(std::memory_order_relaxed))
-     << "\n";
+  os << "Native bytes total: " << GetNativeBytes()
+     << " registered: " << native_bytes_registered_.load(std::memory_order_relaxed) << "\n";
+
+  os << "Total native bytes at last GC: "
+     << old_native_bytes_allocated_.load(std::memory_order_relaxed) << "\n";
 
   BaseMutex::DumpAll(os);
 }
@@ -1337,7 +1342,8 @@
   size_t total_bytes_free = GetFreeMemory();
   oss << "Failed to allocate a " << byte_count << " byte allocation with " << total_bytes_free
       << " free bytes and " << PrettySize(GetFreeMemoryUntilOOME()) << " until OOM,"
-      << " max allowed footprint " << max_allowed_footprint_ << ", growth limit "
+      << " target footprint " << target_footprint_.load(std::memory_order_relaxed)
+      << ", growth limit "
       << growth_limit_;
   // If the allocation failed due to fragmentation, print out the largest continuous allocation.
   if (total_bytes_free >= byte_count) {
@@ -1872,7 +1878,7 @@
 }
 
 void Heap::SetTargetHeapUtilization(float target) {
-  DCHECK_GT(target, 0.0f);  // asserted in Java code
+  DCHECK_GT(target, 0.1f);  // asserted in Java code
   DCHECK_LT(target, 1.0f);
   target_utilization_ = target;
 }
@@ -2286,8 +2292,8 @@
     }
     if (IsGcConcurrent()) {
       concurrent_start_bytes_ =
-          std::max(max_allowed_footprint_, kMinConcurrentRemainingBytes) -
-          kMinConcurrentRemainingBytes;
+          UnsignedDifference(target_footprint_.load(std::memory_order_relaxed),
+                             kMinConcurrentRemainingBytes);
     } else {
       concurrent_start_bytes_ = std::numeric_limits<size_t>::max();
     }
@@ -2616,6 +2622,39 @@
   ATRACE_INT("Heap size (KB)", heap_size / KB);
 }
 
+size_t Heap::GetNativeBytes() {
+  size_t malloc_bytes;
+#if defined(__BIONIC__) || defined(__GLIBC__)
+  size_t mmapped_bytes;
+  struct mallinfo mi = mallinfo();
+  // In spite of the documentation, the jemalloc version of this call seems to do what we want,
+  // and it is thread-safe.
+  if (sizeof(size_t) > sizeof(mi.uordblks) && sizeof(size_t) > sizeof(mi.hblkhd)) {
+    // Shouldn't happen, but glibc declares uordblks as int.
+    // Avoiding sign extension gets us correct behavior for another 2 GB.
+    malloc_bytes = (unsigned int)mi.uordblks;
+    mmapped_bytes = (unsigned int)mi.hblkhd;
+  } else {
+    malloc_bytes = mi.uordblks;
+    mmapped_bytes = mi.hblkhd;
+  }
+  // From the spec, we clearly have mmapped_bytes <= malloc_bytes. Reality is sometimes
+  // dramatically different. (b/119580449) If so, fudge it.
+  if (mmapped_bytes > malloc_bytes) {
+    malloc_bytes = mmapped_bytes;
+  }
+#else
+  // We should hit this case only in contexts in which GC triggering is not critical. Effectively
+  // disable GC triggering based on malloc().
+  malloc_bytes = 1000;
+#endif
+  return malloc_bytes + native_bytes_registered_.load(std::memory_order_relaxed);
+  // An alternative would be to get RSS from /proc/self/statm. Empirically, that's no
+  // more expensive, and it would allow us to count memory allocated by means other than malloc.
+  // However it would change as pages are unmapped and remapped due to memory pressure, among
+  // other things. It seems risky to trigger GCs as a result of such changes.
+}
+
 collector::GcType Heap::CollectGarbageInternal(collector::GcType gc_type,
                                                GcCause gc_cause,
                                                bool clear_soft_references) {
@@ -2666,16 +2705,7 @@
     ++runtime->GetStats()->gc_for_alloc_count;
     ++self->GetStats()->gc_for_alloc_count;
   }
-  const uint64_t bytes_allocated_before_gc = GetBytesAllocated();
-
-  if (gc_type == NonStickyGcType()) {
-    // Move all bytes from new_native_bytes_allocated_ to
-    // old_native_bytes_allocated_ now that GC has been triggered, resetting
-    // new_native_bytes_allocated_ to zero in the process.
-    old_native_bytes_allocated_.fetch_add(
-        new_native_bytes_allocated_.exchange(0, std::memory_order_relaxed),
-        std::memory_order_relaxed);
-  }
+  const size_t bytes_allocated_before_gc = GetBytesAllocated();
 
   DCHECK_LT(gc_type, collector::kGcTypeMax);
   DCHECK_NE(gc_type, collector::kGcTypeNone);
@@ -2747,6 +2777,9 @@
   FinishGC(self, gc_type);
   // Inform DDMS that a GC completed.
   Dbg::GcDidFinish();
+
+  old_native_bytes_allocated_.store(GetNativeBytes());
+
   // Unload native libraries for class unloading. We do this after calling FinishGC to prevent
   // deadlocks in case the JNI_OnUnload function does allocations.
   {
@@ -3521,16 +3554,17 @@
 }
 
 size_t Heap::GetPercentFree() {
-  return static_cast<size_t>(100.0f * static_cast<float>(GetFreeMemory()) / max_allowed_footprint_);
+  return static_cast<size_t>(100.0f * static_cast<float>(
+      GetFreeMemory()) / target_footprint_.load(std::memory_order_relaxed));
 }
 
-void Heap::SetIdealFootprint(size_t max_allowed_footprint) {
-  if (max_allowed_footprint > GetMaxMemory()) {
-    VLOG(gc) << "Clamp target GC heap from " << PrettySize(max_allowed_footprint) << " to "
+void Heap::SetIdealFootprint(size_t target_footprint) {
+  if (target_footprint > GetMaxMemory()) {
+    VLOG(gc) << "Clamp target GC heap from " << PrettySize(target_footprint) << " to "
              << PrettySize(GetMaxMemory());
-    max_allowed_footprint = GetMaxMemory();
+    target_footprint = GetMaxMemory();
   }
-  max_allowed_footprint_ = max_allowed_footprint;
+  target_footprint_.store(target_footprint, std::memory_order_relaxed);
 }
 
 bool Heap::IsMovableObject(ObjPtr<mirror::Object> obj) const {
@@ -3563,10 +3597,10 @@
 }
 
 void Heap::GrowForUtilization(collector::GarbageCollector* collector_ran,
-                              uint64_t bytes_allocated_before_gc) {
+                              size_t bytes_allocated_before_gc) {
   // We know what our utilization is at this moment.
   // This doesn't actually resize any memory. It just lets the heap grow more when necessary.
-  const uint64_t bytes_allocated = GetBytesAllocated();
+  const size_t bytes_allocated = GetBytesAllocated();
   // Trace the new heap size after the GC is finished.
   TraceHeapSize(bytes_allocated);
   uint64_t target_size;
@@ -3574,16 +3608,18 @@
   // Use the multiplier to grow more for foreground.
   const double multiplier = HeapGrowthMultiplier();  // Use the multiplier to grow more for
   // foreground.
-  const uint64_t adjusted_min_free = static_cast<uint64_t>(min_free_ * multiplier);
-  const uint64_t adjusted_max_free = static_cast<uint64_t>(max_free_ * multiplier);
+  const size_t adjusted_min_free = static_cast<size_t>(min_free_ * multiplier);
+  const size_t adjusted_max_free = static_cast<size_t>(max_free_ * multiplier);
   if (gc_type != collector::kGcTypeSticky) {
     // Grow the heap for non sticky GC.
-    ssize_t delta = bytes_allocated / GetTargetHeapUtilization() - bytes_allocated;
-    CHECK_GE(delta, 0) << "bytes_allocated=" << bytes_allocated
-                       << " target_utilization_=" << target_utilization_;
+    uint64_t delta = bytes_allocated * (1.0 / GetTargetHeapUtilization() - 1.0);
+    DCHECK_LE(delta, std::numeric_limits<size_t>::max()) << "bytes_allocated=" << bytes_allocated
+        << " target_utilization_=" << target_utilization_;
     target_size = bytes_allocated + delta * multiplier;
-    target_size = std::min(target_size, bytes_allocated + adjusted_max_free);
-    target_size = std::max(target_size, bytes_allocated + adjusted_min_free);
+    target_size = std::min(target_size,
+                           static_cast<uint64_t>(bytes_allocated + adjusted_max_free));
+    target_size = std::max(target_size,
+                           static_cast<uint64_t>(bytes_allocated + adjusted_min_free));
     next_gc_type_ = collector::kGcTypeSticky;
   } else {
     collector::GcType non_sticky_gc_type = NonStickyGcType();
@@ -3600,22 +3636,24 @@
     // We also check that the bytes allocated aren't over the footprint limit in order to prevent a
     // pathological case where dead objects which aren't reclaimed by sticky could get accumulated
     // if the sticky GC throughput always remained >= the full/partial throughput.
+    size_t target_footprint = target_footprint_.load(std::memory_order_relaxed);
     if (current_gc_iteration_.GetEstimatedThroughput() * kStickyGcThroughputAdjustment >=
         non_sticky_collector->GetEstimatedMeanThroughput() &&
         non_sticky_collector->NumberOfIterations() > 0 &&
-        bytes_allocated <= max_allowed_footprint_) {
+        bytes_allocated <= target_footprint) {
       next_gc_type_ = collector::kGcTypeSticky;
     } else {
       next_gc_type_ = non_sticky_gc_type;
     }
     // If we have freed enough memory, shrink the heap back down.
-    if (bytes_allocated + adjusted_max_free < max_allowed_footprint_) {
+    if (bytes_allocated + adjusted_max_free < target_footprint) {
       target_size = bytes_allocated + adjusted_max_free;
     } else {
-      target_size = std::max(bytes_allocated, static_cast<uint64_t>(max_allowed_footprint_));
+      target_size = std::max(bytes_allocated, target_footprint);
     }
   }
-  if (!ignore_max_footprint_) {
+  CHECK_LE(target_size, std::numeric_limits<size_t>::max());
+  if (!ignore_target_footprint_) {
     SetIdealFootprint(target_size);
     if (IsGcConcurrent()) {
       const uint64_t freed_bytes = current_gc_iteration_.GetFreedBytes() +
@@ -3624,26 +3662,25 @@
       // Bytes allocated will shrink by freed_bytes after the GC runs, so if we want to figure out
       // how many bytes were allocated during the GC we need to add freed_bytes back on.
       CHECK_GE(bytes_allocated + freed_bytes, bytes_allocated_before_gc);
-      const uint64_t bytes_allocated_during_gc = bytes_allocated + freed_bytes -
+      const size_t bytes_allocated_during_gc = bytes_allocated + freed_bytes -
           bytes_allocated_before_gc;
       // Calculate when to perform the next ConcurrentGC.
       // Estimate how many remaining bytes we will have when we need to start the next GC.
       size_t remaining_bytes = bytes_allocated_during_gc;
       remaining_bytes = std::min(remaining_bytes, kMaxConcurrentRemainingBytes);
       remaining_bytes = std::max(remaining_bytes, kMinConcurrentRemainingBytes);
-      if (UNLIKELY(remaining_bytes > max_allowed_footprint_)) {
+      size_t target_footprint = target_footprint_.load(std::memory_order_relaxed);
+      if (UNLIKELY(remaining_bytes > target_footprint)) {
         // A never going to happen situation that from the estimated allocation rate we will exceed
         // the applications entire footprint with the given estimated allocation rate. Schedule
         // another GC nearly straight away.
-        remaining_bytes = kMinConcurrentRemainingBytes;
+        remaining_bytes = std::min(kMinConcurrentRemainingBytes, target_footprint);
       }
-      DCHECK_LE(remaining_bytes, max_allowed_footprint_);
-      DCHECK_LE(max_allowed_footprint_, GetMaxMemory());
+      DCHECK_LE(target_footprint_.load(std::memory_order_relaxed), GetMaxMemory());
       // Start a concurrent GC when we get close to the estimated remaining bytes. When the
       // allocation rate is very high, remaining_bytes could tell us that we should start a GC
       // right away.
-      concurrent_start_bytes_ = std::max(max_allowed_footprint_ - remaining_bytes,
-                                         static_cast<size_t>(bytes_allocated));
+      concurrent_start_bytes_ = std::max(target_footprint - remaining_bytes, bytes_allocated);
     }
   }
 }
@@ -3671,11 +3708,11 @@
 }
 
 void Heap::ClearGrowthLimit() {
-  if (max_allowed_footprint_ == growth_limit_ && growth_limit_ < capacity_) {
-    max_allowed_footprint_ = capacity_;
+  if (target_footprint_.load(std::memory_order_relaxed) == growth_limit_
+      && growth_limit_ < capacity_) {
+    target_footprint_.store(capacity_, std::memory_order_relaxed);
     concurrent_start_bytes_ =
-         std::max(max_allowed_footprint_, kMinConcurrentRemainingBytes) -
-         kMinConcurrentRemainingBytes;
+        UnsignedDifference(capacity_, kMinConcurrentRemainingBytes);
   }
   growth_limit_ = capacity_;
   ScopedObjectAccess soa(Thread::Current());
@@ -3915,40 +3952,101 @@
                             static_cast<jlong>(timeout));
 }
 
-void Heap::RegisterNativeAllocation(JNIEnv* env, size_t bytes) {
-  size_t old_value = new_native_bytes_allocated_.fetch_add(bytes, std::memory_order_relaxed);
+// For GC triggering purposes, we count old (pre-last-GC) and new native allocations as
+// different fractions of Java allocations.
+// For now, we essentially do not count old native allocations at all, so that we can preserve the
+// existing behavior of not limiting native heap size. If we seriously considered it, we would
+// have to adjust collection thresholds when we encounter large amounts of old native memory,
+// and handle native out-of-memory situations.
 
-  if (old_value > NativeAllocationGcWatermark() * HeapGrowthMultiplier() &&
-             !IsGCRequestPending()) {
-    // Trigger another GC because there have been enough native bytes
-    // allocated since the last GC.
+static constexpr size_t kOldNativeDiscountFactor = 65536;  // Approximately infinite for now.
+static constexpr size_t kNewNativeDiscountFactor = 2;
+
+// If weighted java + native memory use exceeds our target by kStopForNativeFactor, and
+// newly allocated memory exceeds kHugeNativeAlloc, we wait for GC to complete to avoid
+// running out of memory.
+static constexpr float kStopForNativeFactor = 2.0;
+static constexpr size_t kHugeNativeAllocs = 200*1024*1024;
+
+// Return the ratio of the weighted native + java allocated bytes to its target value.
+// A return value > 1.0 means we should collect. Significantly larger values mean we're falling
+// behind.
+inline float Heap::NativeMemoryOverTarget(size_t current_native_bytes) {
+  // Collection check for native allocation. Does not enforce Java heap bounds.
+  // With adj_start_bytes defined below, effectively checks
+  // <java bytes allocd> + c1*<old native allocd> + c2*<new native allocd) >= adj_start_bytes,
+  // where c3 > 1, and currently c1 and c2 are 1 divided by the values defined above.
+  size_t old_native_bytes = old_native_bytes_allocated_.load(std::memory_order_relaxed);
+  if (old_native_bytes > current_native_bytes) {
+    // Net decrease; skip the check, but update old value.
+    // It's OK to lose an update if two stores race.
+    old_native_bytes_allocated_.store(current_native_bytes, std::memory_order_relaxed);
+    return 0.0;
+  } else {
+    size_t new_native_bytes = UnsignedDifference(current_native_bytes, old_native_bytes);
+    size_t weighted_native_bytes = new_native_bytes / kNewNativeDiscountFactor
+        + old_native_bytes / kOldNativeDiscountFactor;
+    size_t adj_start_bytes = concurrent_start_bytes_
+        + NativeAllocationGcWatermark() / kNewNativeDiscountFactor;
+    return static_cast<float>(GetBytesAllocated() + weighted_native_bytes)
+         / static_cast<float>(adj_start_bytes);
+  }
+}
+
+inline void Heap::CheckConcurrentGCForNative(Thread* self) {
+  size_t current_native_bytes = GetNativeBytes();
+  float gc_urgency = NativeMemoryOverTarget(current_native_bytes);
+  if (UNLIKELY(gc_urgency >= 1.0)) {
     if (IsGcConcurrent()) {
-      RequestConcurrentGC(ThreadForEnv(env), kGcCauseForNativeAlloc, /*force_full=*/true);
+      RequestConcurrentGC(self, kGcCauseForNativeAlloc, /*force_full=*/true);
+      if (gc_urgency > kStopForNativeFactor
+          && current_native_bytes > kHugeNativeAllocs) {
+        // We're in danger of running out of memory due to rampant native allocation.
+        if (VLOG_IS_ON(heap) || VLOG_IS_ON(startup)) {
+          LOG(INFO) << "Stopping for native allocation, urgency: " << gc_urgency;
+        }
+        WaitForGcToComplete(kGcCauseForAlloc, self);
+      }
     } else {
       CollectGarbageInternal(NonStickyGcType(), kGcCauseForNativeAlloc, false);
     }
   }
 }
 
-void Heap::RegisterNativeFree(JNIEnv*, size_t bytes) {
-  // Take the bytes freed out of new_native_bytes_allocated_ first. If
-  // new_native_bytes_allocated_ reaches zero, take the remaining bytes freed
-  // out of old_native_bytes_allocated_ to ensure all freed bytes are
-  // accounted for.
-  size_t allocated;
-  size_t new_freed_bytes;
-  do {
-    allocated = new_native_bytes_allocated_.load(std::memory_order_relaxed);
-    new_freed_bytes = std::min(allocated, bytes);
-  } while (!new_native_bytes_allocated_.CompareAndSetWeakRelaxed(allocated,
-                                                                   allocated - new_freed_bytes));
-  if (new_freed_bytes < bytes) {
-    old_native_bytes_allocated_.fetch_sub(bytes - new_freed_bytes, std::memory_order_relaxed);
+// About kNotifyNativeInterval allocations have occurred. Check whether we should garbage collect.
+void Heap::NotifyNativeAllocations(JNIEnv* env) {
+  native_objects_notified_.fetch_add(kNotifyNativeInterval, std::memory_order_relaxed);
+  CheckConcurrentGCForNative(ThreadForEnv(env));
+}
+
+// Register a native allocation with an explicit size.
+// This should only be done for large allocations of non-malloc memory, which we wouldn't
+// otherwise see.
+void Heap::RegisterNativeAllocation(JNIEnv* env, size_t bytes) {
+  native_bytes_registered_.fetch_add(bytes, std::memory_order_relaxed);
+  uint32_t objects_notified =
+      native_objects_notified_.fetch_add(1, std::memory_order_relaxed);
+  if (objects_notified % kNotifyNativeInterval == kNotifyNativeInterval - 1
+      || bytes > kCheckImmediatelyThreshold) {
+    CheckConcurrentGCForNative(ThreadForEnv(env));
   }
 }
 
+void Heap::RegisterNativeFree(JNIEnv*, size_t bytes) {
+  size_t allocated;
+  size_t new_freed_bytes;
+  do {
+    allocated = native_bytes_registered_.load(std::memory_order_relaxed);
+    new_freed_bytes = std::min(allocated, bytes);
+    // We should not be registering more free than allocated bytes.
+    // But correctly keep going in non-debug builds.
+    DCHECK_EQ(new_freed_bytes, bytes);
+  } while (!native_bytes_registered_.CompareAndSetWeakRelaxed(allocated,
+                                                              allocated - new_freed_bytes));
+}
+
 size_t Heap::GetTotalMemory() const {
-  return std::max(max_allowed_footprint_, GetBytesAllocated());
+  return std::max(target_footprint_.load(std::memory_order_relaxed), GetBytesAllocated());
 }
 
 void Heap::AddModUnionTable(accounting::ModUnionTable* mod_union_table) {
@@ -4250,8 +4348,8 @@
   return verification_.get();
 }
 
-void Heap::VlogHeapGrowth(size_t max_allowed_footprint, size_t new_footprint, size_t alloc_size) {
-  VLOG(heap) << "Growing heap from " << PrettySize(max_allowed_footprint) << " to "
+void Heap::VlogHeapGrowth(size_t old_footprint, size_t new_footprint, size_t alloc_size) {
+  VLOG(heap) << "Growing heap from " << PrettySize(old_footprint) << " to "
              << PrettySize(new_footprint) << " for a " << PrettySize(alloc_size) << " allocation";
 }
 
@@ -4262,20 +4360,21 @@
     gc::Heap* heap = Runtime::Current()->GetHeap();
     // Trigger a GC, if not already done. The first GC after fork, whenever it
     // takes place, will adjust the thresholds to normal levels.
-    if (heap->max_allowed_footprint_ == heap->growth_limit_) {
+    if (heap->target_footprint_.load(std::memory_order_relaxed) == heap->growth_limit_) {
       heap->RequestConcurrentGC(self, kGcCauseBackground, false);
     }
   }
 };
 
 void Heap::PostForkChildAction(Thread* self) {
-  // Temporarily increase max_allowed_footprint_ and concurrent_start_bytes_ to
+  // Temporarily increase target_footprint_ and concurrent_start_bytes_ to
   // max values to avoid GC during app launch.
   if (collector_type_ == kCollectorTypeCC && !IsLowMemoryMode()) {
-    // Set max_allowed_footprint_ to the largest allowed value.
+    // Set target_footprint_ to the largest allowed value.
     SetIdealFootprint(growth_limit_);
     // Set concurrent_start_bytes_ to half of the heap size.
-    concurrent_start_bytes_ = std::max(max_allowed_footprint_ / 2, GetBytesAllocated());
+    size_t target_footprint = target_footprint_.load(std::memory_order_relaxed);
+    concurrent_start_bytes_ = std::max(target_footprint / 2, GetBytesAllocated());
 
     GetTaskProcessor()->AddTask(
         self, new TriggerPostForkCCGcTask(NanoTime() + MsToNs(kPostForkMaxHeapDurationMS)));
diff --git a/runtime/gc/heap.h b/runtime/gc/heap.h
index 504eff2..de65f023 100644
--- a/runtime/gc/heap.h
+++ b/runtime/gc/heap.h
@@ -126,7 +126,6 @@
 
 class Heap {
  public:
-  // If true, measure the total allocation time.
   static constexpr size_t kDefaultStartingSize = kPageSize;
   static constexpr size_t kDefaultInitialSize = 2 * MB;
   static constexpr size_t kDefaultMaximumSize = 256 * MB;
@@ -155,6 +154,16 @@
   // Used so that we don't overflow the allocation time atomic integer.
   static constexpr size_t kTimeAdjust = 1024;
 
+  // Client should call NotifyNativeAllocation every kNotifyNativeInterval allocations.
+  // Should be chosen so that time_to_call_mallinfo / kNotifyNativeInterval is on the same order
+  // as object allocation time. time_to_call_mallinfo seems to be on the order of 1 usec.
+  static constexpr uint32_t kNotifyNativeInterval = 32;
+
+  // RegisterNativeAllocation checks immediately whether GC is needed if size exceeds the
+  // following. kCheckImmediatelyThreshold * kNotifyNativeInterval should be small enough to
+  // make it safe to allocate that many bytes between checks.
+  static constexpr size_t kCheckImmediatelyThreshold = 300000;
+
   // How often we allow heap trimming to happen (nanoseconds).
   static constexpr uint64_t kHeapTrimWait = MsToNs(5000);
   // How long we wait after a transition request to perform a collector transition (nanoseconds).
@@ -187,7 +196,7 @@
        bool low_memory_mode,
        size_t long_pause_threshold,
        size_t long_gc_threshold,
-       bool ignore_max_footprint,
+       bool ignore_target_footprint,
        bool use_tlab,
        bool verify_pre_gc_heap,
        bool verify_pre_sweeping_heap,
@@ -269,10 +278,22 @@
   void CheckPreconditionsForAllocObject(ObjPtr<mirror::Class> c, size_t byte_count)
       REQUIRES_SHARED(Locks::mutator_lock_);
 
+  // Inform the garbage collector of a non-malloc allocated native memory that might become
+  // reclaimable in the future as a result of Java garbage collection.
   void RegisterNativeAllocation(JNIEnv* env, size_t bytes)
       REQUIRES(!*gc_complete_lock_, !*pending_task_lock_);
   void RegisterNativeFree(JNIEnv* env, size_t bytes);
 
+  // Notify the garbage collector of malloc allocations that might be reclaimable
+  // as a result of Java garbage collection. Each such call represents approximately
+  // kNotifyNativeInterval such allocations.
+  void NotifyNativeAllocations(JNIEnv* env)
+      REQUIRES(!*gc_complete_lock_, !*pending_task_lock_);
+
+  uint32_t GetNotifyNativeInterval() {
+    return kNotifyNativeInterval;
+  }
+
   // Change the allocator, updates entrypoints.
   void ChangeAllocator(AllocatorType allocator)
       REQUIRES(Locks::mutator_lock_, !Locks::runtime_shutdown_lock_);
@@ -536,21 +557,20 @@
 
   // Returns approximately how much free memory we have until the next GC happens.
   size_t GetFreeMemoryUntilGC() const {
-    return max_allowed_footprint_ - GetBytesAllocated();
+    return UnsignedDifference(target_footprint_.load(std::memory_order_relaxed),
+                              GetBytesAllocated());
   }
 
   // Returns approximately how much free memory we have until the next OOME happens.
   size_t GetFreeMemoryUntilOOME() const {
-    return growth_limit_ - GetBytesAllocated();
+    return UnsignedDifference(growth_limit_, GetBytesAllocated());
   }
 
   // Returns how much free memory we have until we need to grow the heap to perform an allocation.
   // Similar to GetFreeMemoryUntilGC. Implements java.lang.Runtime.freeMemory.
   size_t GetFreeMemory() const {
-    size_t byte_allocated = num_bytes_allocated_.load(std::memory_order_relaxed);
-    size_t total_memory = GetTotalMemory();
-    // Make sure we don't get a negative number.
-    return total_memory - std::min(total_memory, byte_allocated);
+    return UnsignedDifference(GetTotalMemory(),
+                              num_bytes_allocated_.load(std::memory_order_relaxed));
   }
 
   // Get the space that corresponds to an object's address. Current implementation searches all
@@ -877,12 +897,16 @@
     return main_space_backup_ != nullptr;
   }
 
+  static ALWAYS_INLINE size_t UnsignedDifference(size_t x, size_t y) {
+    return x > y ? x - y : 0;
+  }
+
   static ALWAYS_INLINE bool AllocatorHasAllocationStack(AllocatorType allocator_type) {
     return
+        allocator_type != kAllocatorTypeRegionTLAB &&
         allocator_type != kAllocatorTypeBumpPointer &&
         allocator_type != kAllocatorTypeTLAB &&
-        allocator_type != kAllocatorTypeRegion &&
-        allocator_type != kAllocatorTypeRegionTLAB;
+        allocator_type != kAllocatorTypeRegion;
   }
   static ALWAYS_INLINE bool AllocatorMayHaveConcurrentGC(AllocatorType allocator_type) {
     if (kUseReadBarrier) {
@@ -890,24 +914,30 @@
       return true;
     }
     return
-        allocator_type != kAllocatorTypeBumpPointer &&
-        allocator_type != kAllocatorTypeTLAB;
+        allocator_type != kAllocatorTypeTLAB &&
+        allocator_type != kAllocatorTypeBumpPointer;
   }
   static bool IsMovingGc(CollectorType collector_type) {
     return
+        collector_type == kCollectorTypeCC ||
         collector_type == kCollectorTypeSS ||
         collector_type == kCollectorTypeGSS ||
-        collector_type == kCollectorTypeCC ||
         collector_type == kCollectorTypeCCBackground ||
         collector_type == kCollectorTypeHomogeneousSpaceCompact;
   }
   bool ShouldAllocLargeObject(ObjPtr<mirror::Class> c, size_t byte_count) const
       REQUIRES_SHARED(Locks::mutator_lock_);
-  ALWAYS_INLINE void CheckConcurrentGC(Thread* self,
-                                       size_t new_num_bytes_allocated,
-                                       ObjPtr<mirror::Object>* obj)
+
+  // Checks whether we should garbage collect:
+  ALWAYS_INLINE bool ShouldConcurrentGCForJava(size_t new_num_bytes_allocated);
+  float NativeMemoryOverTarget(size_t current_native_bytes);
+  ALWAYS_INLINE void CheckConcurrentGCForJava(Thread* self,
+                                              size_t new_num_bytes_allocated,
+                                              ObjPtr<mirror::Object>* obj)
       REQUIRES_SHARED(Locks::mutator_lock_)
       REQUIRES(!*pending_task_lock_, !*gc_complete_lock_);
+  void CheckConcurrentGCForNative(Thread* self)
+      REQUIRES(!*pending_task_lock_, !*gc_complete_lock_);
 
   accounting::ObjectStack* GetMarkStack() {
     return mark_stack_.get();
@@ -968,6 +998,11 @@
   void ThrowOutOfMemoryError(Thread* self, size_t byte_count, AllocatorType allocator_type)
       REQUIRES_SHARED(Locks::mutator_lock_);
 
+  // Are we out of memory, and thus should force a GC or fail?
+  // For concurrent collectors, out of memory is defined by growth_limit_.
+  // For nonconcurrent collectors it is defined by target_footprint_ unless grow is
+  // set. If grow is set, the limit is growth_limit_ and we adjust target_footprint_
+  // to accomodate the allocation.
   ALWAYS_INLINE bool IsOutOfMemoryOnAllocation(AllocatorType allocator_type,
                                                size_t alloc_size,
                                                bool grow);
@@ -1031,7 +1066,7 @@
   // collection. bytes_allocated_before_gc is used to measure bytes / second for the period which
   // the GC was run.
   void GrowForUtilization(collector::GarbageCollector* collector_ran,
-                          uint64_t bytes_allocated_before_gc = 0);
+                          size_t bytes_allocated_before_gc = 0);
 
   size_t GetPercentFree();
 
@@ -1065,8 +1100,8 @@
   // What kind of concurrency behavior is the runtime after? Currently true for concurrent mark
   // sweep GC, false for other GC types.
   bool IsGcConcurrent() const ALWAYS_INLINE {
-    return collector_type_ == kCollectorTypeCMS ||
-        collector_type_ == kCollectorTypeCC ||
+    return collector_type_ == kCollectorTypeCC ||
+        collector_type_ == kCollectorTypeCMS ||
         collector_type_ == kCollectorTypeCCBackground;
   }
 
@@ -1095,15 +1130,19 @@
     return HasZygoteSpace() ? collector::kGcTypePartial : collector::kGcTypeFull;
   }
 
-  // How large new_native_bytes_allocated_ can grow before we trigger a new
-  // GC.
+  // Return the amount of space we allow for native memory when deciding whether to
+  // collect. We collect when a weighted sum of Java memory plus native memory exceeds
+  // the similarly weighted sum of the Java heap size target and this value.
   ALWAYS_INLINE size_t NativeAllocationGcWatermark() const {
-    // Reuse max_free_ for the native allocation gc watermark, so that the
-    // native heap is treated in the same way as the Java heap in the case
-    // where the gc watermark update would exceed max_free_. Using max_free_
-    // instead of the target utilization means the watermark doesn't depend on
-    // the current number of registered native allocations.
-    return max_free_;
+    // It probably makes most sense to use a constant multiple of target_footprint_ .
+    // This is a good indication of the live data size, together with the
+    // intended space-time trade-off, as expressed by SetTargetHeapUtilization.
+    // For a fixed target utilization, the amount of GC effort per native
+    // allocated byte remains roughly constant as the Java heap size changes.
+    // But we previously triggered on max_free_ native allocation which is often much
+    // smaller. To avoid unexpected growth, we partially keep that limit in place for now.
+    // TODO: Consider HeapGrowthMultiplier(). Maybe.
+    return std::min(target_footprint_.load(std::memory_order_relaxed), 2 * max_free_);
   }
 
   ALWAYS_INLINE void IncrementNumberOfBytesFreedRevoke(size_t freed_bytes_revoke);
@@ -1113,6 +1152,11 @@
   // Remove a vlog code from heap-inl.h which is transitively included in half the world.
   static void VlogHeapGrowth(size_t max_allowed_footprint, size_t new_footprint, size_t alloc_size);
 
+  // Return our best approximation of the number of bytes of native memory that
+  // are currently in use, and could possibly be reclaimed as an indirect result
+  // of a garbage collection.
+  size_t GetNativeBytes();
+
   // All-known continuous spaces, where objects lie within fixed bounds.
   std::vector<space::ContinuousSpace*> continuous_spaces_ GUARDED_BY(Locks::mutator_lock_);
 
@@ -1192,9 +1236,9 @@
   double pre_gc_weighted_allocated_bytes_;
   double post_gc_weighted_allocated_bytes_;
 
-  // If we ignore the max footprint it lets the heap grow until it hits the heap capacity, this is
-  // useful for benchmarking since it reduces time spent in GC to a low %.
-  const bool ignore_max_footprint_;
+  // If we ignore the target footprint it lets the heap grow until it hits the heap capacity, this
+  // is useful for benchmarking since it reduces time spent in GC to a low %.
+  const bool ignore_target_footprint_;
 
   // Lock which guards zygote space creation.
   Mutex zygote_creation_lock_;
@@ -1243,14 +1287,18 @@
 
   // The size the heap is limited to. This is initially smaller than capacity, but for largeHeap
   // programs it is "cleared" making it the same as capacity.
+  // Only weakly enforced for simultaneous allocations.
   size_t growth_limit_;
 
-  // When the number of bytes allocated exceeds the footprint TryAllocate returns null indicating
-  // a GC should be triggered.
-  size_t max_allowed_footprint_;
+  // Target size (as in maximum allocatable bytes) for the heap. Weakly enforced as a limit for
+  // non-concurrent GC. Used as a guideline for computing concurrent_start_bytes_ in the
+  // concurrent GC case.
+  Atomic<size_t> target_footprint_;
 
   // When num_bytes_allocated_ exceeds this amount then a concurrent GC should be requested so that
   // it completes ahead of an allocation failing.
+  // A multiple of this is also used to determine when to trigger a GC in response to native
+  // allocation.
   size_t concurrent_start_bytes_;
 
   // Since the heap was created, how many bytes have been freed.
@@ -1263,19 +1311,18 @@
   // TLABS in their entirety, even if they have not yet been parceled out.
   Atomic<size_t> num_bytes_allocated_;
 
-  // Number of registered native bytes allocated since the last time GC was
-  // triggered. Adjusted after each RegisterNativeAllocation and
-  // RegisterNativeFree. Used to determine when to trigger GC for native
-  // allocations.
-  // See the REDESIGN section of go/understanding-register-native-allocation.
-  Atomic<size_t> new_native_bytes_allocated_;
+  // Number of registered native bytes allocated. Adjusted after each RegisterNativeAllocation and
+  // RegisterNativeFree. Used to  help determine when to trigger GC for native allocations. Should
+  // not include bytes allocated through the system malloc, since those are implicitly included.
+  Atomic<size_t> native_bytes_registered_;
 
-  // Number of registered native bytes allocated prior to the last time GC was
-  // triggered, for debugging purposes. The current number of registered
-  // native bytes is determined by taking the sum of
-  // old_native_bytes_allocated_ and new_native_bytes_allocated_.
+  // Approximately the smallest value of GetNativeBytes() we've seen since the last GC.
   Atomic<size_t> old_native_bytes_allocated_;
 
+  // Total number of native objects of which we were notified since the beginning of time, mod 2^32.
+  // Allows us to check for GC only roughly every kNotifyNativeInterval allocations.
+  Atomic<uint32_t> native_objects_notified_;
+
   // Number of bytes freed by thread local buffer revokes. This will
   // cancel out the ahead-of-time bulk counting of bytes allocated in
   // rosalloc thread-local buffers.  It is temporarily accumulated
@@ -1360,10 +1407,10 @@
 
   // Minimum free guarantees that you always have at least min_free_ free bytes after growing for
   // utilization, regardless of target utilization ratio.
-  size_t min_free_;
+  const size_t min_free_;
 
   // The ideal maximum free size, when we grow the heap for utilization.
-  size_t max_free_;
+  const size_t max_free_;
 
   // Target ideal heap utilization ratio.
   double target_utilization_;
diff --git a/runtime/gc/space/image_space.cc b/runtime/gc/space/image_space.cc
index 4c2074d..66db063 100644
--- a/runtime/gc/space/image_space.cc
+++ b/runtime/gc/space/image_space.cc
@@ -28,6 +28,7 @@
 
 #include "art_field-inl.h"
 #include "art_method-inl.h"
+#include "base/array_ref.h"
 #include "base/bit_memory_region.h"
 #include "base/callee_save_type.h"
 #include "base/enums.h"
@@ -44,6 +45,7 @@
 #include "dex/dex_file_loader.h"
 #include "exec_utils.h"
 #include "gc/accounting/space_bitmap-inl.h"
+#include "gc/task_processor.h"
 #include "image-inl.h"
 #include "image_space_fs.h"
 #include "intern_table-inl.h"
@@ -59,6 +61,7 @@
 namespace gc {
 namespace space {
 
+using android::base::StringAppendF;
 using android::base::StringPrintf;
 
 Atomic<uint32_t> ImageSpace::bitmap_index_(0);
@@ -654,6 +657,22 @@
   const CodeVisitor code_visitor_;
 };
 
+template <typename ReferenceVisitor>
+class ImageSpace::ClassTableVisitor final {
+ public:
+  explicit ClassTableVisitor(const ReferenceVisitor& reference_visitor)
+      : reference_visitor_(reference_visitor) {}
+
+  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
+      REQUIRES_SHARED(Locks::mutator_lock_) {
+    DCHECK(root->AsMirrorPtr() != nullptr);
+    root->Assign(reference_visitor_(root->AsMirrorPtr()));
+  }
+
+ private:
+  ReferenceVisitor reference_visitor_;
+};
+
 // Helper class encapsulating loading, so we can access private ImageSpace members (this is a
 // nested class), but not declare functions in the header.
 class ImageSpace::Loader {
@@ -666,30 +685,39 @@
       REQUIRES_SHARED(Locks::mutator_lock_) {
     TimingLogger logger(__PRETTY_FUNCTION__, /*precise=*/ true, VLOG_IS_ON(image));
 
-    const bool create_thread_pool = true;
     std::unique_ptr<ThreadPool> thread_pool;
-    if (create_thread_pool) {
-      TimingLogger::ScopedTiming timing("CreateThreadPool", &logger);
-      ScopedThreadStateChange stsc(Thread::Current(), kNative);
-      constexpr size_t kStackSize = 64 * KB;
-      constexpr size_t kMaxRuntimeWorkers = 4u;
-      const size_t num_workers =
-          std::min(static_cast<size_t>(std::thread::hardware_concurrency()), kMaxRuntimeWorkers);
-      thread_pool.reset(new ThreadPool("Runtime", num_workers, /*create_peers=*/false, kStackSize));
-      thread_pool->StartWorkers(Thread::Current());
-    }
-
     std::unique_ptr<ImageSpace> space = Init(image_filename,
                                              image_location,
                                              oat_file,
                                              &logger,
-                                             thread_pool.get(),
+                                             &thread_pool,
                                              image_reservation,
                                              error_msg);
     if (thread_pool != nullptr) {
-      TimingLogger::ScopedTiming timing("CreateThreadPool", &logger);
-      ScopedThreadStateChange stsc(Thread::Current(), kNative);
-      thread_pool.reset();
+      // Delay the thread pool deletion to prevent the deletion slowing down the startup by causing
+      // preemption. TODO: Just do this in heap trim.
+      static constexpr uint64_t kThreadPoolDeleteDelay = MsToNs(5000);
+
+      class DeleteThreadPoolTask : public HeapTask {
+       public:
+        explicit DeleteThreadPoolTask(std::unique_ptr<ThreadPool>&& thread_pool)
+            : HeapTask(NanoTime() + kThreadPoolDeleteDelay), thread_pool_(std::move(thread_pool)) {}
+
+        void Run(Thread* self) override {
+          ScopedTrace trace("DestroyThreadPool");
+          ScopedThreadStateChange stsc(self, kNative);
+          thread_pool_.reset();
+        }
+
+       private:
+        std::unique_ptr<ThreadPool> thread_pool_;
+      };
+      gc::TaskProcessor* const processor = Runtime::Current()->GetHeap()->GetTaskProcessor();
+      // The thread pool is already done being used since Init has finished running. Deleting the
+      // thread pool is done async since it takes a non-trivial amount of time to do.
+      if (processor != nullptr) {
+        processor->AddTask(Thread::Current(), new DeleteThreadPoolTask(std::move(thread_pool)));
+      }
     }
     if (space != nullptr) {
       uint32_t expected_reservation_size =
@@ -701,11 +729,22 @@
 
       TimingLogger::ScopedTiming timing("RelocateImage", &logger);
       ImageHeader* image_header = reinterpret_cast<ImageHeader*>(space->GetMemMap()->Begin());
-      if (!RelocateInPlace(*image_header,
-                           space->GetMemMap()->Begin(),
-                           space->GetLiveBitmap(),
-                           oat_file,
-                           error_msg)) {
+      const PointerSize pointer_size = image_header->GetPointerSize();
+      bool result;
+      if (pointer_size == PointerSize::k64) {
+        result = RelocateInPlace<PointerSize::k64>(*image_header,
+                                                   space->GetMemMap()->Begin(),
+                                                   space->GetLiveBitmap(),
+                                                   oat_file,
+                                                   error_msg);
+      } else {
+        result = RelocateInPlace<PointerSize::k32>(*image_header,
+                                                   space->GetMemMap()->Begin(),
+                                                   space->GetLiveBitmap(),
+                                                   oat_file,
+                                                   error_msg);
+      }
+      if (!result) {
         return nullptr;
       }
       Runtime* runtime = Runtime::Current();
@@ -740,7 +779,7 @@
                                           const char* image_location,
                                           const OatFile* oat_file,
                                           TimingLogger* logger,
-                                          ThreadPool* thread_pool,
+                                          std::unique_ptr<ThreadPool>* thread_pool,
                                           /*inout*/MemMap* image_reservation,
                                           /*out*/std::string* error_msg)
       REQUIRES_SHARED(Locks::mutator_lock_) {
@@ -817,6 +856,18 @@
       return nullptr;
     }
 
+    const size_t kMinBlocks = 2;
+    if (thread_pool != nullptr && image_header->GetBlockCount() >= kMinBlocks) {
+      TimingLogger::ScopedTiming timing("CreateThreadPool", logger);
+      ScopedThreadStateChange stsc(Thread::Current(), kNative);
+      constexpr size_t kStackSize = 64 * KB;
+      constexpr size_t kMaxRuntimeWorkers = 4u;
+      const size_t num_workers =
+          std::min(static_cast<size_t>(std::thread::hardware_concurrency()), kMaxRuntimeWorkers);
+      thread_pool->reset(new ThreadPool("Image", num_workers, /*create_peers=*/false, kStackSize));
+      thread_pool->get()->StartWorkers(Thread::Current());
+    }
+
     // GetImageBegin is the preferred address to map the image. If we manage to map the
     // image at the image begin, the amount of fixup work required is minimized.
     // If it is pic we will retry with error_msg for the failure case. Pass a null error_msg to
@@ -829,7 +880,7 @@
         *image_header,
         file->Fd(),
         logger,
-        thread_pool,
+        thread_pool != nullptr ? thread_pool->get() : nullptr,
         image_reservation,
         error_msg);
     if (!map.IsValid()) {
@@ -966,8 +1017,7 @@
 
       const uint64_t start = NanoTime();
       Thread* const self = Thread::Current();
-      const size_t kMinBlocks = 2;
-      const bool use_parallel = pool != nullptr && image_header.GetBlockCount() >= kMinBlocks;
+      const bool use_parallel = pool != nullptr;
       for (const ImageHeader::Block& block : image_header.GetBlocks(temp_map.Begin())) {
         auto function = [&](Thread*) {
           const uint64_t start2 = NanoTime();
@@ -1089,11 +1139,8 @@
   class FixupObjectVisitor : public FixupVisitor {
    public:
     template<typename... Args>
-    explicit FixupObjectVisitor(gc::accounting::ContinuousSpaceBitmap* visited,
-                                const PointerSize pointer_size,
-                                Args... args)
+    explicit FixupObjectVisitor(gc::accounting::ContinuousSpaceBitmap* visited, Args... args)
         : FixupVisitor(args...),
-          pointer_size_(pointer_size),
           visited_(visited) {}
 
     // Fix up separately since we also need to fix up method entrypoints.
@@ -1105,39 +1152,14 @@
 
     ALWAYS_INLINE void operator()(ObjPtr<mirror::Object> obj,
                                   MemberOffset offset,
-
                                   bool is_static ATTRIBUTE_UNUSED) const
         NO_THREAD_SAFETY_ANALYSIS {
-      // There could be overlap between ranges, we must avoid visiting the same reference twice.
-      // Avoid the class field since we already fixed it up in FixupClassVisitor.
-      if (offset.Uint32Value() != mirror::Object::ClassOffset().Uint32Value()) {
-        // Space is not yet added to the heap, don't do a read barrier.
-        mirror::Object* ref = obj->GetFieldObject<mirror::Object, kVerifyNone, kWithoutReadBarrier>(
-            offset);
-        // Use SetFieldObjectWithoutWriteBarrier to avoid card marking since we are writing to the
-        // image.
-        obj->SetFieldObjectWithoutWriteBarrier<false, true, kVerifyNone>(offset, ForwardObject(ref));
-      }
-    }
-
-    // Visit a pointer array and forward corresponding native data. Ignores pointer arrays in the
-    // boot image. Uses the bitmap to ensure the same array is not visited multiple times.
-    template <typename Visitor>
-    void UpdatePointerArrayContents(mirror::PointerArray* array, const Visitor& visitor) const
-        NO_THREAD_SAFETY_ANALYSIS {
-      DCHECK(array != nullptr);
-      DCHECK(visitor.IsInAppImage(array));
-      // The bit for the array contents is different than the bit for the array. Since we may have
-      // already visited the array as a long / int array from walking the bitmap without knowing it
-      // was a pointer array.
-      static_assert(kObjectAlignment == 8u, "array bit may be in another object");
-      mirror::Object* const contents_bit = reinterpret_cast<mirror::Object*>(
-          reinterpret_cast<uintptr_t>(array) + kObjectAlignment);
-      // If the bit is not set then the contents have not yet been updated.
-      if (!visited_->Test(contents_bit)) {
-        array->Fixup<kVerifyNone>(array, pointer_size_, visitor);
-        visited_->Set(contents_bit);
-      }
+      // Space is not yet added to the heap, don't do a read barrier.
+      mirror::Object* ref = obj->GetFieldObject<mirror::Object, kVerifyNone, kWithoutReadBarrier>(
+          offset);
+      // Use SetFieldObjectWithoutWriteBarrier to avoid card marking since we are writing to the
+      // image.
+      obj->SetFieldObjectWithoutWriteBarrier<false, true, kVerifyNone>(offset, ForwardObject(ref));
     }
 
     // java.lang.ref.Reference visitor.
@@ -1152,81 +1174,16 @@
 
     void operator()(mirror::Object* obj) const
         NO_THREAD_SAFETY_ANALYSIS {
-      if (visited_->Test(obj)) {
-        // Already visited.
-        return;
-      }
-      visited_->Set(obj);
-
-      // Handle class specially first since we need it to be updated to properly visit the rest of
-      // the instance fields.
-      {
-        mirror::Class* klass = obj->GetClass<kVerifyNone, kWithoutReadBarrier>();
-        DCHECK(klass != nullptr) << "Null class in image";
-        // No AsClass since our fields aren't quite fixed up yet.
-        mirror::Class* new_klass = down_cast<mirror::Class*>(ForwardObject(klass));
-        if (klass != new_klass) {
-          obj->SetClass<kVerifyNone>(new_klass);
-        }
-        if (new_klass != klass && IsInAppImage(new_klass)) {
-          // Make sure the klass contents are fixed up since we depend on it to walk the fields.
-          operator()(new_klass);
-        }
-      }
-
-      if (obj->IsClass()) {
-        mirror::Class* klass = obj->AsClass<kVerifyNone>();
-        // Fixup super class before visiting instance fields which require
-        // information from their super class to calculate offsets.
-        mirror::Class* super_class =
-            klass->GetSuperClass<kVerifyNone, kWithoutReadBarrier>().Ptr();
-        if (super_class != nullptr) {
-          mirror::Class* new_super_class = down_cast<mirror::Class*>(ForwardObject(super_class));
-          if (new_super_class != super_class && IsInAppImage(new_super_class)) {
-            // Recursively fix all dependencies.
-            operator()(new_super_class);
-          }
-        }
-      }
-
-      obj->VisitReferences</*visit native roots*/false, kVerifyNone, kWithoutReadBarrier>(
-          *this,
-          *this);
-      // Note that this code relies on no circular dependencies.
-      // We want to use our own class loader and not the one in the image.
-      if (obj->IsClass<kVerifyNone>()) {
-        mirror::Class* as_klass = obj->AsClass<kVerifyNone>();
-        FixupObjectAdapter visitor(boot_image_, app_image_, app_oat_);
-        as_klass->FixupNativePointers<kVerifyNone>(as_klass, pointer_size_, visitor);
-        // Deal with the pointer arrays. Use the helper function since multiple classes can reference
-        // the same arrays.
-        mirror::PointerArray* const vtable = as_klass->GetVTable<kVerifyNone, kWithoutReadBarrier>();
-        if (vtable != nullptr && IsInAppImage(vtable)) {
-          operator()(vtable);
-          UpdatePointerArrayContents(vtable, visitor);
-        }
-        mirror::IfTable* iftable = as_klass->GetIfTable<kVerifyNone, kWithoutReadBarrier>();
-        // Ensure iftable arrays are fixed up since we need GetMethodArray to return the valid
-        // contents.
-        if (IsInAppImage(iftable)) {
-          operator()(iftable);
-          for (int32_t i = 0, count = iftable->Count(); i < count; ++i) {
-            if (iftable->GetMethodArrayCount<kVerifyNone, kWithoutReadBarrier>(i) > 0) {
-              mirror::PointerArray* methods =
-                  iftable->GetMethodArray<kVerifyNone, kWithoutReadBarrier>(i);
-              if (visitor.IsInAppImage(methods)) {
-                operator()(methods);
-                DCHECK(methods != nullptr);
-                UpdatePointerArrayContents(methods, visitor);
-              }
-            }
-          }
-        }
+      if (!visited_->Set(obj)) {
+        // Not already visited.
+        obj->VisitReferences</*visit native roots*/false, kVerifyNone, kWithoutReadBarrier>(
+            *this,
+            *this);
+        CHECK(!obj->IsClass());
       }
     }
 
    private:
-    const PointerSize pointer_size_;
     gc::accounting::ContinuousSpaceBitmap* const visited_;
   };
 
@@ -1306,6 +1263,7 @@
   // Relocate an image space mapped at target_base which possibly used to be at a different base
   // address. In place means modifying a single ImageSpace in place rather than relocating from
   // one ImageSpace to another.
+  template <PointerSize kPointerSize>
   static bool RelocateInPlace(ImageHeader& image_header,
                               uint8_t* target_base,
                               accounting::ContinuousSpaceBitmap* bitmap,
@@ -1317,7 +1275,6 @@
     uint32_t boot_image_end = 0;
     uint32_t boot_oat_begin = 0;
     uint32_t boot_oat_end = 0;
-    const PointerSize pointer_size = image_header.GetPointerSize();
     gc::Heap* const heap = Runtime::Current()->GetHeap();
     heap->GetBootImagesSize(&boot_image_begin, &boot_image_end, &boot_oat_begin, &boot_oat_end);
     if (boot_image_begin == boot_image_end) {
@@ -1359,11 +1316,8 @@
       return true;
     }
     ScopedDebugDisallowReadBarriers sddrb(Thread::Current());
-    // Need to update the image to be at the target base.
-    const ImageSection& objects_section = image_header.GetObjectsSection();
-    uintptr_t objects_begin = reinterpret_cast<uintptr_t>(target_base + objects_section.Offset());
-    uintptr_t objects_end = reinterpret_cast<uintptr_t>(target_base + objects_section.End());
     FixupObjectAdapter fixup_adapter(boot_image, app_image, app_oat);
+    PatchObjectVisitor<kPointerSize, FixupObjectAdapter> patch_object_visitor(fixup_adapter);
     if (fixup_image) {
       // Two pass approach, fix up all classes first, then fix up non class-objects.
       // The visited bitmap is used to ensure that pointer arrays are not forwarded twice.
@@ -1371,16 +1325,64 @@
           gc::accounting::ContinuousSpaceBitmap::Create("Relocate bitmap",
                                                         target_base,
                                                         image_header.GetImageSize()));
-      FixupObjectVisitor fixup_object_visitor(visited_bitmap.get(),
-                                              pointer_size,
-                                              boot_image,
-                                              app_image,
-                                              app_oat);
-      TimingLogger::ScopedTiming timing("Fixup classes", &logger);
-      // Fixup objects may read fields in the boot image, use the mutator lock here for sanity. Though
-      // its probably not required.
+      FixupObjectVisitor fixup_object_visitor(visited_bitmap.get(), boot_image, app_image, app_oat);
+      {
+        TimingLogger::ScopedTiming timing("Fixup classes", &logger);
+        const auto& class_table_section = image_header.GetClassTableSection();
+        if (class_table_section.Size() > 0u) {
+          ScopedObjectAccess soa(Thread::Current());
+          ClassTableVisitor class_table_visitor(fixup_adapter);
+          size_t read_count = 0u;
+          const uint8_t* data = target_base + class_table_section.Offset();
+          // We avoid making a copy of the data since we want modifications to be propagated to the
+          // memory map.
+          ClassTable::ClassSet temp_set(data, /*make_copy_of_data=*/ false, &read_count);
+          for (ClassTable::TableSlot& slot : temp_set) {
+            slot.VisitRoot(class_table_visitor);
+            mirror::Class* klass = slot.Read<kWithoutReadBarrier>();
+            if (!fixup_adapter.IsInAppImage(klass)) {
+              continue;
+            }
+            const bool already_marked = visited_bitmap->Set(klass);
+            CHECK(!already_marked) << "App image class already visited";
+            patch_object_visitor.VisitClass(klass);
+            // Then patch the non-embedded vtable and iftable.
+            mirror::PointerArray* vtable = klass->GetVTable<kVerifyNone, kWithoutReadBarrier>();
+            if (vtable != nullptr &&
+                fixup_object_visitor.IsInAppImage(vtable) &&
+                !visited_bitmap->Set(vtable)) {
+              patch_object_visitor.VisitPointerArray(vtable);
+            }
+            auto* iftable = klass->GetIfTable<kVerifyNone, kWithoutReadBarrier>();
+            if (iftable != nullptr && fixup_object_visitor.IsInAppImage(iftable)) {
+              // Avoid processing the fields of iftable since we will process them later anyways
+              // below.
+              int32_t ifcount = klass->GetIfTableCount<kVerifyNone>();
+              for (int32_t i = 0; i != ifcount; ++i) {
+                mirror::PointerArray* unpatched_ifarray =
+                    iftable->GetMethodArrayOrNull<kVerifyNone, kWithoutReadBarrier>(i);
+                if (unpatched_ifarray != nullptr) {
+                  // The iftable has not been patched, so we need to explicitly adjust the pointer.
+                  mirror::PointerArray* ifarray = fixup_adapter(unpatched_ifarray);
+                  if (fixup_object_visitor.IsInAppImage(ifarray) &&
+                      !visited_bitmap->Set(ifarray)) {
+                    patch_object_visitor.VisitPointerArray(ifarray);
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+
+      // Fixup objects may read fields in the boot image, use the mutator lock here for sanity.
+      // Though its probably not required.
+      TimingLogger::ScopedTiming timing("Fixup cobjects", &logger);
       ScopedObjectAccess soa(Thread::Current());
-      timing.NewTiming("Fixup objects");
+      // Need to update the image to be at the target base.
+      const ImageSection& objects_section = image_header.GetObjectsSection();
+      uintptr_t objects_begin = reinterpret_cast<uintptr_t>(target_base + objects_section.Offset());
+      uintptr_t objects_end = reinterpret_cast<uintptr_t>(target_base + objects_section.End());
       bitmap->VisitMarkedRange(objects_begin, objects_end, fixup_object_visitor);
       // Fixup image roots.
       CHECK(app_image.InSource(reinterpret_cast<uintptr_t>(
@@ -1392,96 +1394,19 @@
           AsObjectArray<mirror::DexCache, kVerifyNone>();
       for (int32_t i = 0, count = dex_caches->GetLength(); i < count; ++i) {
         mirror::DexCache* dex_cache = dex_caches->Get<kVerifyNone, kWithoutReadBarrier>(i);
-        // Fix up dex cache pointers.
-        mirror::StringDexCacheType* strings = dex_cache->GetStrings();
-        if (strings != nullptr) {
-          mirror::StringDexCacheType* new_strings = fixup_adapter.ForwardObject(strings);
-          if (strings != new_strings) {
-            dex_cache->SetStrings(new_strings);
-          }
-          dex_cache->FixupStrings<kWithoutReadBarrier>(new_strings, fixup_adapter);
-        }
-        mirror::TypeDexCacheType* types = dex_cache->GetResolvedTypes();
-        if (types != nullptr) {
-          mirror::TypeDexCacheType* new_types = fixup_adapter.ForwardObject(types);
-          if (types != new_types) {
-            dex_cache->SetResolvedTypes(new_types);
-          }
-          dex_cache->FixupResolvedTypes<kWithoutReadBarrier>(new_types, fixup_adapter);
-        }
-        mirror::MethodDexCacheType* methods = dex_cache->GetResolvedMethods();
-        if (methods != nullptr) {
-          mirror::MethodDexCacheType* new_methods = fixup_adapter.ForwardObject(methods);
-          if (methods != new_methods) {
-            dex_cache->SetResolvedMethods(new_methods);
-          }
-          for (size_t j = 0, num = dex_cache->NumResolvedMethods(); j != num; ++j) {
-            auto pair = mirror::DexCache::GetNativePairPtrSize(new_methods, j, pointer_size);
-            ArtMethod* orig = pair.object;
-            ArtMethod* copy = fixup_adapter.ForwardObject(orig);
-            if (orig != copy) {
-              pair.object = copy;
-              mirror::DexCache::SetNativePairPtrSize(new_methods, j, pair, pointer_size);
-            }
-          }
-        }
-        mirror::FieldDexCacheType* fields = dex_cache->GetResolvedFields();
-        if (fields != nullptr) {
-          mirror::FieldDexCacheType* new_fields = fixup_adapter.ForwardObject(fields);
-          if (fields != new_fields) {
-            dex_cache->SetResolvedFields(new_fields);
-          }
-          for (size_t j = 0, num = dex_cache->NumResolvedFields(); j != num; ++j) {
-            mirror::FieldDexCachePair orig =
-                mirror::DexCache::GetNativePairPtrSize(new_fields, j, pointer_size);
-            mirror::FieldDexCachePair copy(fixup_adapter.ForwardObject(orig.object), orig.index);
-            if (orig.object != copy.object) {
-              mirror::DexCache::SetNativePairPtrSize(new_fields, j, copy, pointer_size);
-            }
-          }
-        }
-
-        mirror::MethodTypeDexCacheType* method_types = dex_cache->GetResolvedMethodTypes();
-        if (method_types != nullptr) {
-          mirror::MethodTypeDexCacheType* new_method_types =
-              fixup_adapter.ForwardObject(method_types);
-          if (method_types != new_method_types) {
-            dex_cache->SetResolvedMethodTypes(new_method_types);
-          }
-          dex_cache->FixupResolvedMethodTypes<kWithoutReadBarrier>(new_method_types, fixup_adapter);
-        }
-        GcRoot<mirror::CallSite>* call_sites = dex_cache->GetResolvedCallSites();
-        if (call_sites != nullptr) {
-          GcRoot<mirror::CallSite>* new_call_sites = fixup_adapter.ForwardObject(call_sites);
-          if (call_sites != new_call_sites) {
-            dex_cache->SetResolvedCallSites(new_call_sites);
-          }
-          dex_cache->FixupResolvedCallSites<kWithoutReadBarrier>(new_call_sites, fixup_adapter);
-        }
-
-        GcRoot<mirror::String>* preresolved_strings = dex_cache->GetPreResolvedStrings();
-        if (preresolved_strings != nullptr) {
-          GcRoot<mirror::String>* new_array = fixup_adapter.ForwardObject(preresolved_strings);
-          if (preresolved_strings != new_array) {
-            dex_cache->SetPreResolvedStrings(new_array);
-          }
-          const size_t num_preresolved_strings = dex_cache->NumPreResolvedStrings();
-          for (size_t j = 0; j < num_preresolved_strings; ++j) {
-            new_array[j] = GcRoot<mirror::String>(
-                fixup_adapter(new_array[j].Read<kWithoutReadBarrier>()));
-          }
-        }
+        CHECK(dex_cache != nullptr);
+        patch_object_visitor.VisitDexCacheArrays(dex_cache);
       }
     }
     {
       // Only touches objects in the app image, no need for mutator lock.
       TimingLogger::ScopedTiming timing("Fixup methods", &logger);
       FixupArtMethodVisitor method_visitor(fixup_image,
-                                           pointer_size,
+                                           kPointerSize,
                                            boot_image,
                                            app_image,
                                            app_oat);
-      image_header.VisitPackedArtMethods(&method_visitor, target_base, pointer_size);
+      image_header.VisitPackedArtMethods(&method_visitor, target_base, kPointerSize);
     }
     if (fixup_image) {
       {
@@ -1492,26 +1417,14 @@
       }
       {
         TimingLogger::ScopedTiming timing("Fixup imt", &logger);
-        image_header.VisitPackedImTables(fixup_adapter, target_base, pointer_size);
+        image_header.VisitPackedImTables(fixup_adapter, target_base, kPointerSize);
       }
       {
         TimingLogger::ScopedTiming timing("Fixup conflict tables", &logger);
-        image_header.VisitPackedImtConflictTables(fixup_adapter, target_base, pointer_size);
+        image_header.VisitPackedImtConflictTables(fixup_adapter, target_base, kPointerSize);
       }
       // In the app image case, the image methods are actually in the boot image.
       image_header.RelocateImageMethods(boot_image.Delta());
-      const auto& class_table_section = image_header.GetClassTableSection();
-      if (class_table_section.Size() > 0u) {
-        // Note that we require that ReadFromMemory does not make an internal copy of the elements.
-        // This also relies on visit roots not doing any verification which could fail after we update
-        // the roots to be the image addresses.
-        ScopedObjectAccess soa(Thread::Current());
-        WriterMutexLock mu(Thread::Current(), *Locks::classlinker_classes_lock_);
-        ClassTable temp_table;
-        temp_table.ReadFromMemory(target_base + class_table_section.Offset());
-        FixupRootVisitor root_visitor(boot_image, app_image, app_oat);
-        temp_table.VisitRoots(root_visitor);
-      }
       // Fix up the intern table.
       const auto& intern_table_section = image_header.GetInternedStringsSection();
       if (intern_table_section.Size() > 0u) {
@@ -1654,8 +1567,10 @@
       *error_msg = StringPrintf("Cannot read header of %s", filename.c_str());
       return false;
     }
-    if (system_hdr.GetComponentCount() != boot_class_path_.size()) {
-      *error_msg = StringPrintf("Unexpected component count in %s, received %u, expected %zu",
+    if (system_hdr.GetComponentCount() == 0u ||
+        system_hdr.GetComponentCount() > boot_class_path_.size()) {
+      *error_msg = StringPrintf("Unexpected component count in %s, received %u, "
+                                    "expected non-zero and <= %zu",
                                 filename.c_str(),
                                 system_hdr.GetComponentCount(),
                                 boot_class_path_.size());
@@ -1672,10 +1587,12 @@
       return false;
     }
 
+    ArrayRef<const std::string> provided_locations(boot_class_path_locations_.data(),
+                                                   system_hdr.GetComponentCount());
     std::vector<std::string> locations =
-        ExpandMultiImageLocations(boot_class_path_locations_, image_location_);
+        ExpandMultiImageLocations(provided_locations, image_location_);
     std::vector<std::string> filenames =
-        ExpandMultiImageLocations(boot_class_path_locations_, filename);
+        ExpandMultiImageLocations(provided_locations, filename);
     DCHECK_EQ(locations.size(), filenames.size());
     std::vector<std::unique_ptr<ImageSpace>> spaces;
     spaces.reserve(locations.size());
@@ -1694,7 +1611,7 @@
     }
     for (size_t i = 0u, size = spaces.size(); i != size; ++i) {
       std::string expected_boot_class_path =
-          (i == 0u) ? android::base::Join(boot_class_path_locations_, ':') : std::string();
+          (i == 0u) ? android::base::Join(provided_locations, ':') : std::string();
       if (!OpenOatFile(spaces[i].get(),
                        boot_class_path_[i],
                        expected_boot_class_path,
@@ -1766,22 +1683,6 @@
     BitMemoryRegion visited_objects_;
   };
 
-  template <typename ReferenceVisitor>
-  class ClassTableVisitor final {
-   public:
-    explicit ClassTableVisitor(const ReferenceVisitor& reference_visitor)
-        : reference_visitor_(reference_visitor) {}
-
-    void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
-        REQUIRES_SHARED(Locks::mutator_lock_) {
-      DCHECK(root->AsMirrorPtr() != nullptr);
-      root->Assign(reference_visitor_(root->AsMirrorPtr()));
-    }
-
-   private:
-    ReferenceVisitor reference_visitor_;
-  };
-
   template <PointerSize kPointerSize>
   static void DoRelocateSpaces(const std::vector<std::unique_ptr<ImageSpace>>& spaces,
                                uint32_t diff) REQUIRES_SHARED(Locks::mutator_lock_) {
@@ -2444,9 +2345,113 @@
   return true;
 }
 
+std::string ImageSpace::GetBootClassPathChecksums(const std::vector<std::string>& boot_class_path,
+                                                  const std::string& image_location,
+                                                  InstructionSet image_isa,
+                                                  /*out*/std::string* error_msg) {
+  std::string system_filename;
+  bool has_system = false;
+  std::string cache_filename;
+  bool has_cache = false;
+  bool dalvik_cache_exists = false;
+  bool is_global_cache = false;
+  if (!FindImageFilename(image_location.c_str(),
+                         image_isa,
+                         &system_filename,
+                         &has_system,
+                         &cache_filename,
+                         &dalvik_cache_exists,
+                         &has_cache,
+                         &is_global_cache)) {
+    *error_msg = StringPrintf("Unable to find image file for %s and %s",
+                              image_location.c_str(),
+                              GetInstructionSetString(image_isa));
+    return std::string();
+  }
+
+  DCHECK(has_system || has_cache);
+  const std::string& filename = has_system ? system_filename : cache_filename;
+  std::unique_ptr<ImageHeader> header = ReadSpecificImageHeader(filename.c_str(), error_msg);
+  if (header == nullptr) {
+    return std::string();
+  }
+  if (header->GetComponentCount() == 0u || header->GetComponentCount() > boot_class_path.size()) {
+    *error_msg = StringPrintf("Unexpected component count in %s, received %u, "
+                                  "expected non-zero and <= %zu",
+                              filename.c_str(),
+                              header->GetComponentCount(),
+                              boot_class_path.size());
+    return std::string();
+  }
+
+  std::string boot_image_checksum =
+      StringPrintf("i;%d/%08x", header->GetComponentCount(), header->GetImageChecksum());
+  ArrayRef<const std::string> boot_class_path_tail =
+      ArrayRef<const std::string>(boot_class_path).SubArray(header->GetComponentCount());
+  for (const std::string& bcp_filename : boot_class_path_tail) {
+    std::vector<std::unique_ptr<const DexFile>> dex_files;
+    const ArtDexFileLoader dex_file_loader;
+    if (!dex_file_loader.Open(bcp_filename.c_str(),
+                              bcp_filename,  // The location does not matter here.
+                              /*verify=*/ false,
+                              /*verify_checksum=*/ false,
+                              error_msg,
+                              &dex_files)) {
+      return std::string();
+    }
+    DCHECK(!dex_files.empty());
+    StringAppendF(&boot_image_checksum, ":d");
+    for (const std::unique_ptr<const DexFile>& dex_file : dex_files) {
+      StringAppendF(&boot_image_checksum, "/%08x", dex_file->GetLocationChecksum());
+    }
+  }
+  return boot_image_checksum;
+}
+
+std::string ImageSpace::GetBootClassPathChecksums(
+    const std::vector<ImageSpace*>& image_spaces,
+    const std::vector<const DexFile*>& boot_class_path) {
+  DCHECK(!image_spaces.empty());
+  const ImageHeader& primary_header = image_spaces.front()->GetImageHeader();
+  uint32_t component_count = primary_header.GetComponentCount();
+  DCHECK_EQ(component_count, image_spaces.size());
+  std::string boot_image_checksum =
+      StringPrintf("i;%d/%08x", component_count, primary_header.GetImageChecksum());
+  size_t pos = 0u;
+  for (const ImageSpace* space : image_spaces) {
+    size_t num_dex_files = space->oat_file_non_owned_->GetOatDexFiles().size();
+    if (kIsDebugBuild) {
+      CHECK_NE(num_dex_files, 0u);
+      CHECK_LE(space->oat_file_non_owned_->GetOatDexFiles().size(), boot_class_path.size() - pos);
+      for (size_t i = 0; i != num_dex_files; ++i) {
+        CHECK_EQ(space->oat_file_non_owned_->GetOatDexFiles()[i]->GetDexFileLocation(),
+                 boot_class_path[pos + i]->GetLocation());
+      }
+    }
+    pos += num_dex_files;
+  }
+  ArrayRef<const DexFile* const> boot_class_path_tail =
+      ArrayRef<const DexFile* const>(boot_class_path).SubArray(pos);
+  DCHECK(boot_class_path_tail.empty() ||
+         !DexFileLoader::IsMultiDexLocation(boot_class_path_tail.front()->GetLocation().c_str()));
+  for (const DexFile* dex_file : boot_class_path_tail) {
+    if (!DexFileLoader::IsMultiDexLocation(dex_file->GetLocation().c_str())) {
+      StringAppendF(&boot_image_checksum, ":d");
+    }
+    StringAppendF(&boot_image_checksum, "/%08x", dex_file->GetLocationChecksum());
+  }
+  return boot_image_checksum;
+}
+
 std::vector<std::string> ImageSpace::ExpandMultiImageLocations(
     const std::vector<std::string>& dex_locations,
     const std::string& image_location) {
+  return ExpandMultiImageLocations(ArrayRef<const std::string>(dex_locations), image_location);
+}
+
+std::vector<std::string> ImageSpace::ExpandMultiImageLocations(
+    ArrayRef<const std::string> dex_locations,
+    const std::string& image_location) {
   DCHECK(!dex_locations.empty());
 
   // Find the path.
diff --git a/runtime/gc/space/image_space.h b/runtime/gc/space/image_space.h
index dbc12d1..14e364a 100644
--- a/runtime/gc/space/image_space.h
+++ b/runtime/gc/space/image_space.h
@@ -24,6 +24,8 @@
 
 namespace art {
 
+template <typename T> class ArrayRef;
+class DexFile;
 class OatFile;
 
 namespace gc {
@@ -124,6 +126,19 @@
                                 bool* has_data,
                                 bool *is_global_cache);
 
+  // Returns the checksums for the boot image and extra boot class path dex files,
+  // based on the boot class path, image location and ISA (may differ from the ISA of an
+  // initialized Runtime). The boot image and dex files do not need to be loaded in memory.
+  static std::string GetBootClassPathChecksums(const std::vector<std::string>& boot_class_path,
+                                               const std::string& image_location,
+                                               InstructionSet image_isa,
+                                               /*out*/std::string* error_msg);
+
+  // Returns the checksums for the boot image and extra boot class path dex files,
+  // based on the boot image and boot class path dex files loaded in memory.
+  static std::string GetBootClassPathChecksums(const std::vector<ImageSpace*>& image_spaces,
+                                               const std::vector<const DexFile*>& boot_class_path);
+
   // Expand a single image location to multi-image locations based on the dex locations.
   static std::vector<std::string> ExpandMultiImageLocations(
       const std::vector<std::string>& dex_locations,
@@ -188,7 +203,14 @@
   friend class Space;
 
  private:
+  // Internal overload that takes ArrayRef<> instead of vector<>.
+  static std::vector<std::string> ExpandMultiImageLocations(
+      ArrayRef<const std::string> dex_locations,
+      const std::string& image_location);
+
   class BootImageLoader;
+  template <typename ReferenceVisitor>
+  class ClassTableVisitor;
   class Loader;
   template <typename PatchObjectVisitor>
   class PatchArtFieldVisitor;
diff --git a/runtime/gc/space/region_space-inl.h b/runtime/gc/space/region_space-inl.h
index 5ff1270..dbec4ea 100644
--- a/runtime/gc/space/region_space-inl.h
+++ b/runtime/gc/space/region_space-inl.h
@@ -205,9 +205,10 @@
       continue;
     }
     if (r->IsLarge()) {
-      // Avoid visiting dead large objects since they may contain dangling pointers to the
-      // from-space.
-      DCHECK_GT(r->LiveBytes(), 0u) << "Visiting dead large object";
+      // We may visit a large object with live_bytes = 0 here. However, it is
+      // safe as it cannot contain dangling pointers because corresponding regions
+      // (and regions corresponding to dead referents) cannot be allocated for new
+      // allocations without first clearing regions' live_bytes and state.
       mirror::Object* obj = reinterpret_cast<mirror::Object*>(r->Begin());
       DCHECK(obj->GetClass() != nullptr);
       visitor(obj);
diff --git a/runtime/gc/space/region_space.cc b/runtime/gc/space/region_space.cc
index 21cae93..98b140e 100644
--- a/runtime/gc/space/region_space.cc
+++ b/runtime/gc/space/region_space.cc
@@ -319,6 +319,7 @@
                 state == RegionState::kRegionStateLarge) &&
                type == RegionType::kRegionTypeToSpace);
         bool should_evacuate = r->ShouldBeEvacuated(evac_mode);
+        bool is_newly_allocated = r->IsNewlyAllocated();
         if (should_evacuate) {
           r->SetAsFromSpace();
           DCHECK(r->IsInFromSpace());
@@ -329,6 +330,17 @@
         if (UNLIKELY(state == RegionState::kRegionStateLarge &&
                      type == RegionType::kRegionTypeToSpace)) {
           prev_large_evacuated = should_evacuate;
+          // In 2-phase full heap GC, this function is called after marking is
+          // done. So, it is possible that some newly allocated large object is
+          // marked but its live_bytes is still -1. We need to clear the
+          // mark-bit otherwise the live_bytes will not be updated in
+          // ConcurrentCopying::ProcessMarkStackRef() and hence will break the
+          // logic.
+          if (kEnableGenerationalConcurrentCopyingCollection
+              && !should_evacuate
+              && is_newly_allocated) {
+            GetMarkBitmap()->Clear(reinterpret_cast<mirror::Object*>(r->Begin()));
+          }
           num_expected_large_tails = RoundUp(r->BytesAllocated(), kRegionSize) / kRegionSize - 1;
           DCHECK_GT(num_expected_large_tails, 0U);
         }
@@ -367,7 +379,8 @@
 }
 
 void RegionSpace::ClearFromSpace(/* out */ uint64_t* cleared_bytes,
-                                 /* out */ uint64_t* cleared_objects) {
+                                 /* out */ uint64_t* cleared_objects,
+                                 const bool clear_bitmap) {
   DCHECK(cleared_bytes != nullptr);
   DCHECK(cleared_objects != nullptr);
   *cleared_bytes = 0;
@@ -395,13 +408,18 @@
   // (see b/62194020).
   uint8_t* clear_block_begin = nullptr;
   uint8_t* clear_block_end = nullptr;
-  auto clear_region = [&clear_block_begin, &clear_block_end](Region* r) {
+  auto clear_region = [this, &clear_block_begin, &clear_block_end, clear_bitmap](Region* r) {
     r->Clear(/*zero_and_release_pages=*/false);
     if (clear_block_end != r->Begin()) {
       // Region `r` is not adjacent to the current clear block; zero and release
       // pages within the current block and restart a new clear block at the
       // beginning of region `r`.
       ZeroAndProtectRegion(clear_block_begin, clear_block_end);
+      if (clear_bitmap) {
+        GetLiveBitmap()->ClearRange(
+            reinterpret_cast<mirror::Object*>(clear_block_begin),
+            reinterpret_cast<mirror::Object*>(clear_block_end));
+      }
       clear_block_begin = r->Begin();
     }
     // Add region `r` to the clear block.
@@ -426,20 +444,23 @@
         // It is also better to clear these regions now instead of at the end of the next GC to
         // save RAM. If we don't clear the regions here, they will be cleared next GC by the normal
         // live percent evacuation logic.
+        *cleared_bytes += r->BytesAllocated();
+        *cleared_objects += r->ObjectsAllocated();
+        clear_region(r);
         size_t free_regions = 1;
         // Also release RAM for large tails.
         while (i + free_regions < num_regions_ && regions_[i + free_regions].IsLargeTail()) {
-          DCHECK(r->IsLarge());
           clear_region(&regions_[i + free_regions]);
           ++free_regions;
         }
-        *cleared_bytes += r->BytesAllocated();
-        *cleared_objects += r->ObjectsAllocated();
         num_non_free_regions_ -= free_regions;
-        clear_region(r);
-        GetLiveBitmap()->ClearRange(
-            reinterpret_cast<mirror::Object*>(r->Begin()),
-            reinterpret_cast<mirror::Object*>(r->Begin() + free_regions * kRegionSize));
+        // When clear_bitmap is true, this clearing of bitmap is taken care in
+        // clear_region().
+        if (!clear_bitmap) {
+          GetLiveBitmap()->ClearRange(
+              reinterpret_cast<mirror::Object*>(r->Begin()),
+              reinterpret_cast<mirror::Object*>(r->Begin() + free_regions * kRegionSize));
+        }
         continue;
       }
       r->SetUnevacFromSpaceAsToSpace();
@@ -519,6 +540,11 @@
   }
   // Clear pages for the last block since clearing happens when a new block opens.
   ZeroAndReleasePages(clear_block_begin, clear_block_end - clear_block_begin);
+  if (clear_bitmap) {
+    GetLiveBitmap()->ClearRange(
+        reinterpret_cast<mirror::Object*>(clear_block_begin),
+        reinterpret_cast<mirror::Object*>(clear_block_end));
+  }
   // Update non_free_region_index_limit_.
   SetNonFreeRegionLimit(new_non_free_region_index_limit);
   evac_region_ = nullptr;
diff --git a/runtime/gc/space/region_space.h b/runtime/gc/space/region_space.h
index 8810f8c..0d5ebcc 100644
--- a/runtime/gc/space/region_space.h
+++ b/runtime/gc/space/region_space.h
@@ -228,6 +228,11 @@
     return false;
   }
 
+  bool IsRegionNewlyAllocated(size_t idx) const NO_THREAD_SAFETY_ANALYSIS {
+    DCHECK_LT(idx, num_regions_);
+    return regions_[idx].IsNewlyAllocated();
+  }
+
   bool IsInNewlyAllocatedRegion(mirror::Object* ref) {
     if (HasAddress(ref)) {
       Region* r = RefToRegionUnlocked(ref);
@@ -291,7 +296,9 @@
   size_t FromSpaceSize() REQUIRES(!region_lock_);
   size_t UnevacFromSpaceSize() REQUIRES(!region_lock_);
   size_t ToSpaceSize() REQUIRES(!region_lock_);
-  void ClearFromSpace(/* out */ uint64_t* cleared_bytes, /* out */ uint64_t* cleared_objects)
+  void ClearFromSpace(/* out */ uint64_t* cleared_bytes,
+                      /* out */ uint64_t* cleared_objects,
+                      const bool clear_bitmap)
       REQUIRES(!region_lock_);
 
   void AddLiveBytes(mirror::Object* ref, size_t alloc_size) {
@@ -310,6 +317,40 @@
     }
   }
 
+  void SetAllRegionLiveBytesZero() REQUIRES(!region_lock_) {
+    MutexLock mu(Thread::Current(), region_lock_);
+    const size_t iter_limit = kUseTableLookupReadBarrier
+        ? num_regions_
+        : std::min(num_regions_, non_free_region_index_limit_);
+    for (size_t i = 0; i < iter_limit; ++i) {
+      Region* r = &regions_[i];
+      // Newly allocated regions don't need up-to-date live_bytes_ for deciding
+      // whether to be evacuated or not. See Region::ShouldBeEvacuated().
+      if (!r->IsFree() && !r->IsNewlyAllocated()) {
+        r->ZeroLiveBytes();
+      }
+    }
+  }
+
+  size_t RegionIdxForRefUnchecked(mirror::Object* ref) const NO_THREAD_SAFETY_ANALYSIS {
+    DCHECK(HasAddress(ref));
+    uintptr_t offset = reinterpret_cast<uintptr_t>(ref) - reinterpret_cast<uintptr_t>(Begin());
+    size_t reg_idx = offset / kRegionSize;
+    DCHECK_LT(reg_idx, num_regions_);
+    Region* reg = &regions_[reg_idx];
+    DCHECK_EQ(reg->Idx(), reg_idx);
+    DCHECK(reg->Contains(ref));
+    return reg_idx;
+  }
+  // Return -1 as region index for references outside this region space.
+  size_t RegionIdxForRef(mirror::Object* ref) const NO_THREAD_SAFETY_ANALYSIS {
+    if (HasAddress(ref)) {
+      return RegionIdxForRefUnchecked(ref);
+    } else {
+      return static_cast<size_t>(-1);
+    }
+  }
+
   void RecordAlloc(mirror::Object* ref) REQUIRES(!region_lock_);
   bool AllocNewTlab(Thread* self, size_t min_bytes) REQUIRES(!region_lock_);
 
@@ -515,11 +556,10 @@
     ALWAYS_INLINE bool ShouldBeEvacuated(EvacMode evac_mode);
 
     void AddLiveBytes(size_t live_bytes) {
-      DCHECK(IsInUnevacFromSpace());
+      DCHECK(kEnableGenerationalConcurrentCopyingCollection || IsInUnevacFromSpace());
       DCHECK(!IsLargeTail());
       DCHECK_NE(live_bytes_, static_cast<size_t>(-1));
-      // For large allocations, we always consider all bytes in the
-      // regions live.
+      // For large allocations, we always consider all bytes in the regions live.
       live_bytes_ += IsLarge() ? Top() - begin_ : live_bytes;
       DCHECK_LE(live_bytes_, BytesAllocated());
     }
diff --git a/runtime/interpreter/interpreter.cc b/runtime/interpreter/interpreter.cc
index d004d64..2e41a9d 100644
--- a/runtime/interpreter/interpreter.cc
+++ b/runtime/interpreter/interpreter.cc
@@ -366,6 +366,13 @@
     }
   } else {
     // Enter the "with access check" interpreter.
+
+    // The boot classpath should really not have to run access checks.
+    DCHECK(method->GetDeclaringClass()->GetClassLoader() != nullptr
+           || Runtime::Current()->IsVerificationSoftFail()
+           || Runtime::Current()->IsAotCompiler())
+        << method->PrettyMethod();
+
     if (kInterpreterImplKind == kMterpImplKind) {
       // No access check variants for Mterp.  Just use the switch version.
       if (transaction_active) {
diff --git a/runtime/jit/jit.cc b/runtime/jit/jit.cc
index e43d771..03c97f4 100644
--- a/runtime/jit/jit.cc
+++ b/runtime/jit/jit.cc
@@ -58,7 +58,7 @@
 void* Jit::jit_compiler_handle_ = nullptr;
 void* (*Jit::jit_load_)(void) = nullptr;
 void (*Jit::jit_unload_)(void*) = nullptr;
-bool (*Jit::jit_compile_method_)(void*, ArtMethod*, Thread*, bool) = nullptr;
+bool (*Jit::jit_compile_method_)(void*, ArtMethod*, Thread*, bool, bool) = nullptr;
 void (*Jit::jit_types_loaded_)(void*, mirror::Class**, size_t count) = nullptr;
 bool (*Jit::jit_generate_debug_info_)(void*) = nullptr;
 void (*Jit::jit_update_options_)(void*) = nullptr;
@@ -242,7 +242,7 @@
   return true;
 }
 
-bool Jit::CompileMethod(ArtMethod* method, Thread* self, bool osr) {
+bool Jit::CompileMethod(ArtMethod* method, Thread* self, bool baseline, bool osr) {
   DCHECK(Runtime::Current()->UseJitCompilation());
   DCHECK(!method->IsRuntimeMethod());
 
@@ -272,7 +272,7 @@
   VLOG(jit) << "Compiling method "
             << ArtMethod::PrettyMethod(method_to_compile)
             << " osr=" << std::boolalpha << osr;
-  bool success = jit_compile_method_(jit_compiler_handle_, method_to_compile, self, osr);
+  bool success = jit_compile_method_(jit_compiler_handle_, method_to_compile, self, baseline, osr);
   code_cache_->DoneCompiling(method_to_compile, self, osr);
   if (!success) {
     VLOG(jit) << "Failed to compile method "
@@ -291,6 +291,12 @@
   return success;
 }
 
+void Jit::WaitForWorkersToBeCreated() {
+  if (thread_pool_ != nullptr) {
+    thread_pool_->WaitForWorkersToBeCreated();
+  }
+}
+
 void Jit::DeleteThreadPool() {
   Thread* self = Thread::Current();
   DCHECK(Runtime::Current()->IsShuttingDown(self));
@@ -549,6 +555,7 @@
   enum class TaskKind {
     kAllocateProfile,
     kCompile,
+    kCompileBaseline,
     kCompileOsr,
   };
 
@@ -568,10 +575,12 @@
     ScopedObjectAccess soa(self);
     switch (kind_) {
       case TaskKind::kCompile:
+      case TaskKind::kCompileBaseline:
       case TaskKind::kCompileOsr: {
         Runtime::Current()->GetJit()->CompileMethod(
             method_,
             self,
+            /* baseline= */ (kind_ == TaskKind::kCompileBaseline),
             /* osr= */ (kind_ == TaskKind::kCompileOsr));
         break;
       }
diff --git a/runtime/jit/jit.h b/runtime/jit/jit.h
index 7ce5f07..e5c9766 100644
--- a/runtime/jit/jit.h
+++ b/runtime/jit/jit.h
@@ -161,7 +161,7 @@
   // Create JIT itself.
   static Jit* Create(JitCodeCache* code_cache, JitOptions* options);
 
-  bool CompileMethod(ArtMethod* method, Thread* self, bool osr)
+  bool CompileMethod(ArtMethod* method, Thread* self, bool baseline, bool osr)
       REQUIRES_SHARED(Locks::mutator_lock_);
 
   const JitCodeCache* GetCodeCache() const {
@@ -174,6 +174,7 @@
 
   void CreateThreadPool();
   void DeleteThreadPool();
+  void WaitForWorkersToBeCreated();
 
   // Dump interesting info: #methods compiled, code vs data size, compile / verify cumulative
   // loggers.
@@ -304,7 +305,7 @@
   static void* jit_compiler_handle_;
   static void* (*jit_load_)(void);
   static void (*jit_unload_)(void*);
-  static bool (*jit_compile_method_)(void*, ArtMethod*, Thread*, bool);
+  static bool (*jit_compile_method_)(void*, ArtMethod*, Thread*, bool, bool);
   static void (*jit_types_loaded_)(void*, mirror::Class**, size_t count);
   static void (*jit_update_options_)(void*);
   static bool (*jit_generate_debug_info_)(void*);
diff --git a/runtime/mirror/class-inl.h b/runtime/mirror/class-inl.h
index 185ae3b..679ca43 100644
--- a/runtime/mirror/class-inl.h
+++ b/runtime/mirror/class-inl.h
@@ -871,6 +871,9 @@
 }
 
 inline void Class::SetAccessFlags(uint32_t new_access_flags) {
+  if (kIsDebugBuild) {
+    SetAccessFlagsDCheck(new_access_flags);
+  }
   // Called inside a transaction when setting pre-verified flag during boot image compilation.
   if (Runtime::Current()->IsActiveTransaction()) {
     SetField32<true>(AccessFlagsOffset(), new_access_flags);
diff --git a/runtime/mirror/class.cc b/runtime/mirror/class.cc
index 83d76a9..c5ed1bf 100644
--- a/runtime/mirror/class.cc
+++ b/runtime/mirror/class.cc
@@ -206,6 +206,10 @@
     }
   }
 
+  if (kIsDebugBuild && new_status >= ClassStatus::kInitialized) {
+    CHECK(h_this->WasVerificationAttempted()) << h_this->PrettyClassAndClassLoader();
+  }
+
   if (!class_linker_initialized) {
     // When the class linker is being initialized its single threaded and by definition there can be
     // no waiters. During initialization classes may appear temporary but won't be retired as their
@@ -1461,5 +1465,12 @@
 template void Class::GetAccessFlagsDCheck<kVerifyWrites>();
 template void Class::GetAccessFlagsDCheck<kVerifyAll>();
 
+void Class::SetAccessFlagsDCheck(uint32_t new_access_flags) {
+  uint32_t old_access_flags = GetField32<kVerifyNone>(AccessFlagsOffset());
+  // kAccVerificationAttempted is retained.
+  CHECK((old_access_flags & kAccVerificationAttempted) == 0 ||
+        (new_access_flags & kAccVerificationAttempted) != 0);
+}
+
 }  // namespace mirror
 }  // namespace art
diff --git a/runtime/mirror/class.h b/runtime/mirror/class.h
index 66b1405..d5aa514 100644
--- a/runtime/mirror/class.h
+++ b/runtime/mirror/class.h
@@ -1306,6 +1306,8 @@
   template<VerifyObjectFlags kVerifyFlags>
   void GetAccessFlagsDCheck() REQUIRES_SHARED(Locks::mutator_lock_);
 
+  void SetAccessFlagsDCheck(uint32_t new_access_flags) REQUIRES_SHARED(Locks::mutator_lock_);
+
   // Check that the pointer size matches the one in the class linker.
   ALWAYS_INLINE static void CheckPointerSize(PointerSize pointer_size);
 
diff --git a/runtime/native/dalvik_system_VMRuntime.cc b/runtime/native/dalvik_system_VMRuntime.cc
index 3e5003c..892d4cc 100644
--- a/runtime/native/dalvik_system_VMRuntime.cc
+++ b/runtime/native/dalvik_system_VMRuntime.cc
@@ -271,7 +271,7 @@
 #endif
 }
 
-static void VMRuntime_registerNativeAllocation(JNIEnv* env, jobject, jint bytes) {
+static void VMRuntime_registerNativeAllocationInternal(JNIEnv* env, jobject, jint bytes) {
   if (UNLIKELY(bytes < 0)) {
     ScopedObjectAccess soa(env);
     ThrowRuntimeException("allocation size negative %d", bytes);
@@ -280,11 +280,7 @@
   Runtime::Current()->GetHeap()->RegisterNativeAllocation(env, static_cast<size_t>(bytes));
 }
 
-static void VMRuntime_registerSensitiveThread(JNIEnv*, jobject) {
-  Runtime::Current()->RegisterSensitiveThread();
-}
-
-static void VMRuntime_registerNativeFree(JNIEnv* env, jobject, jint bytes) {
+static void VMRuntime_registerNativeFreeInternal(JNIEnv* env, jobject, jint bytes) {
   if (UNLIKELY(bytes < 0)) {
     ScopedObjectAccess soa(env);
     ThrowRuntimeException("allocation size negative %d", bytes);
@@ -293,6 +289,18 @@
   Runtime::Current()->GetHeap()->RegisterNativeFree(env, static_cast<size_t>(bytes));
 }
 
+static jint VMRuntime_getNotifyNativeInterval(JNIEnv*, jclass) {
+  return Runtime::Current()->GetHeap()->GetNotifyNativeInterval();
+}
+
+static void VMRuntime_notifyNativeAllocationsInternal(JNIEnv* env, jobject) {
+  Runtime::Current()->GetHeap()->NotifyNativeAllocations(env);
+}
+
+static void VMRuntime_registerSensitiveThread(JNIEnv*, jobject) {
+  Runtime::Current()->RegisterSensitiveThread();
+}
+
 static void VMRuntime_updateProcessState(JNIEnv*, jobject, jint process_state) {
   Runtime* runtime = Runtime::Current();
   runtime->UpdateProcessState(static_cast<ProcessState>(process_state));
@@ -710,9 +718,11 @@
   FAST_NATIVE_METHOD(VMRuntime, newUnpaddedArray, "(Ljava/lang/Class;I)Ljava/lang/Object;"),
   NATIVE_METHOD(VMRuntime, properties, "()[Ljava/lang/String;"),
   NATIVE_METHOD(VMRuntime, setTargetSdkVersionNative, "(I)V"),
-  NATIVE_METHOD(VMRuntime, registerNativeAllocation, "(I)V"),
+  NATIVE_METHOD(VMRuntime, registerNativeAllocationInternal, "(I)V"),
+  NATIVE_METHOD(VMRuntime, registerNativeFreeInternal, "(I)V"),
+  NATIVE_METHOD(VMRuntime, getNotifyNativeInterval, "()I"),
+  NATIVE_METHOD(VMRuntime, notifyNativeAllocationsInternal, "()V"),
   NATIVE_METHOD(VMRuntime, registerSensitiveThread, "()V"),
-  NATIVE_METHOD(VMRuntime, registerNativeFree, "(I)V"),
   NATIVE_METHOD(VMRuntime, requestConcurrentGC, "()V"),
   NATIVE_METHOD(VMRuntime, requestHeapTrim, "()V"),
   NATIVE_METHOD(VMRuntime, runHeapTasks, "()V"),
diff --git a/runtime/oat.cc b/runtime/oat.cc
index e931b28..d7c968f 100644
--- a/runtime/oat.cc
+++ b/runtime/oat.cc
@@ -79,8 +79,7 @@
       quick_generic_jni_trampoline_offset_(0),
       quick_imt_conflict_trampoline_offset_(0),
       quick_resolution_trampoline_offset_(0),
-      quick_to_interpreter_bridge_offset_(0),
-      boot_image_checksum_(0) {
+      quick_to_interpreter_bridge_offset_(0) {
   // Don't want asserts in header as they would be checked in each file that includes it. But the
   // fields are private, so we check inside a method.
   static_assert(sizeof(magic_) == sizeof(kOatMagic),
@@ -316,16 +315,6 @@
   quick_to_interpreter_bridge_offset_ = offset;
 }
 
-uint32_t OatHeader::GetBootImageChecksum() const {
-  CHECK(IsValid());
-  return boot_image_checksum_;
-}
-
-void OatHeader::SetBootImageChecksum(uint32_t boot_image_checksum) {
-  CHECK(IsValid());
-  boot_image_checksum_ = boot_image_checksum;
-}
-
 uint32_t OatHeader::GetKeyValueStoreSize() const {
   CHECK(IsValid());
   return key_value_store_size_;
diff --git a/runtime/oat.h b/runtime/oat.h
index b09c81e..ded1489 100644
--- a/runtime/oat.h
+++ b/runtime/oat.h
@@ -31,8 +31,8 @@
 class PACKED(4) OatHeader {
  public:
   static constexpr uint8_t kOatMagic[] = { 'o', 'a', 't', '\n' };
-  // Last oat version changed reason: Pass boot class path to LoadBootImage.
-  static constexpr uint8_t kOatVersion[] = { '1', '6', '5', '\0' };
+  // Last oat version changed reason: Partial boot image.
+  static constexpr uint8_t kOatVersion[] = { '1', '6', '6', '\0' };
 
   static constexpr const char* kDex2OatCmdLineKey = "dex2oat-cmdline";
   static constexpr const char* kDebuggableKey = "debuggable";
@@ -40,6 +40,7 @@
   static constexpr const char* kCompilerFilter = "compiler-filter";
   static constexpr const char* kClassPathKey = "classpath";
   static constexpr const char* kBootClassPathKey = "bootclasspath";
+  static constexpr const char* kBootClassPathChecksumsKey = "bootclasspath-checksums";
   static constexpr const char* kConcurrentCopying = "concurrent-copying";
   static constexpr const char* kCompilationReasonKey = "compilation-reason";
 
@@ -93,9 +94,6 @@
   InstructionSet GetInstructionSet() const;
   uint32_t GetInstructionSetFeaturesBitmap() const;
 
-  uint32_t GetBootImageChecksum() const;
-  void SetBootImageChecksum(uint32_t boot_image_checksum);
-
   uint32_t GetKeyValueStoreSize() const;
   const uint8_t* GetKeyValueStore() const;
   const char* GetStoreValueByKey(const char* key) const;
@@ -137,8 +135,6 @@
   uint32_t quick_resolution_trampoline_offset_;
   uint32_t quick_to_interpreter_bridge_offset_;
 
-  uint32_t boot_image_checksum_;
-
   uint32_t key_value_store_size_;
   uint8_t key_value_store_[0];  // note variable width data at end
 
diff --git a/runtime/oat_file_assistant.cc b/runtime/oat_file_assistant.cc
index 6f32b98..8b81bb9 100644
--- a/runtime/oat_file_assistant.cc
+++ b/runtime/oat_file_assistant.cc
@@ -419,7 +419,7 @@
       // starts up.
       LOG(WARNING) << "Dex location " << dex_location_ << " does not seem to include dex file. "
         << "Allow oat file use. This is potentially dangerous.";
-    } else if (file.GetOatHeader().GetBootImageChecksum() != image_info->boot_image_checksum) {
+    } else if (!image_info->ValidateBootClassPathChecksums(file)) {
       VLOG(oat) << "Oat image checksum does not match image checksum.";
       return kOatBootImageOutOfDate;
     }
@@ -560,6 +560,13 @@
   return required_dex_checksums_found_ ? &cached_required_dex_checksums_ : nullptr;
 }
 
+bool OatFileAssistant::ImageInfo::ValidateBootClassPathChecksums(const OatFile& oat_file) const {
+  const char* oat_boot_class_path_checksums =
+      oat_file.GetOatHeader().GetStoreValueByKey(OatHeader::kBootClassPathChecksumsKey);
+  return oat_boot_class_path_checksums != nullptr &&
+         oat_boot_class_path_checksums == boot_class_path_checksums;
+}
+
 std::unique_ptr<OatFileAssistant::ImageInfo>
 OatFileAssistant::ImageInfo::GetRuntimeImageInfo(InstructionSet isa, std::string* error_msg) {
   CHECK(error_msg != nullptr);
@@ -567,14 +574,11 @@
   Runtime* runtime = Runtime::Current();
   std::unique_ptr<ImageInfo> info(new ImageInfo());
   info->location = runtime->GetImageLocation();
-
-  std::unique_ptr<ImageHeader> image_header(
-      gc::space::ImageSpace::ReadImageHeader(info->location.c_str(), isa, error_msg));
-  if (image_header == nullptr) {
+  info->boot_class_path_checksums = gc::space::ImageSpace::GetBootClassPathChecksums(
+      runtime->GetBootClassPath(), info->location, isa, error_msg);
+  if (info->boot_class_path_checksums.empty()) {
     return nullptr;
   }
-
-  info->boot_image_checksum = image_header->GetImageChecksum();
   return info;
 }
 
diff --git a/runtime/oat_file_assistant.h b/runtime/oat_file_assistant.h
index 09c9d3b..def55b8 100644
--- a/runtime/oat_file_assistant.h
+++ b/runtime/oat_file_assistant.h
@@ -246,8 +246,10 @@
 
  private:
   struct ImageInfo {
-    uint32_t boot_image_checksum = 0;
+    bool ValidateBootClassPathChecksums(const OatFile& oat_file) const;
+
     std::string location;
+    std::string boot_class_path_checksums;
 
     static std::unique_ptr<ImageInfo> GetRuntimeImageInfo(InstructionSet isa,
                                                           std::string* error_msg);
diff --git a/runtime/runtime.cc b/runtime/runtime.cc
index d2c915e..55bc2ec 100644
--- a/runtime/runtime.cc
+++ b/runtime/runtime.cc
@@ -349,6 +349,9 @@
   }
 
   if (jit_ != nullptr) {
+    // Wait for the workers to be created since there can't be any threads attaching during
+    // shutdown.
+    jit_->WaitForWorkersToBeCreated();
     // Stop the profile saver thread before marking the runtime as shutting down.
     // The saver will try to dump the profiles before being sopped and that
     // requires holding the mutator lock.
@@ -971,8 +974,8 @@
   VLOG(startup) << "Runtime::StartDaemonThreads exiting";
 }
 
-static size_t OpenDexFiles(const std::vector<std::string>& dex_filenames,
-                           const std::vector<std::string>& dex_locations,
+static size_t OpenDexFiles(ArrayRef<const std::string> dex_filenames,
+                           ArrayRef<const std::string> dex_locations,
                            std::vector<std::unique_ptr<const DexFile>>* dex_files) {
   DCHECK(dex_files != nullptr) << "OpenDexFiles: out-param is nullptr";
   size_t failure_count = 0;
@@ -1429,6 +1432,21 @@
         GetInternTable()->AddImageStringsToTable(image_space, VoidFunctor());
       }
     }
+    if (heap_->GetBootImageSpaces().size() != GetBootClassPath().size()) {
+      // The boot image did not contain all boot class path components. Load the rest.
+      DCHECK_LT(heap_->GetBootImageSpaces().size(), GetBootClassPath().size());
+      size_t start = heap_->GetBootImageSpaces().size();
+      DCHECK_LT(start, GetBootClassPath().size());
+      std::vector<std::unique_ptr<const DexFile>> extra_boot_class_path;
+      if (runtime_options.Exists(Opt::BootClassPathDexList)) {
+        extra_boot_class_path.swap(*runtime_options.GetOrDefault(Opt::BootClassPathDexList));
+      } else {
+        OpenDexFiles(ArrayRef<const std::string>(GetBootClassPath()).SubArray(start),
+                     ArrayRef<const std::string>(GetBootClassPathLocations()).SubArray(start),
+                     &extra_boot_class_path);
+      }
+      class_linker_->AddExtraBootDexFiles(self, std::move(extra_boot_class_path));
+    }
     if (IsJavaDebuggable()) {
       // Now that we have loaded the boot image, deoptimize its methods if we are running
       // debuggable, as the code may have been compiled non-debuggable.
@@ -1439,7 +1457,9 @@
     if (runtime_options.Exists(Opt::BootClassPathDexList)) {
       boot_class_path.swap(*runtime_options.GetOrDefault(Opt::BootClassPathDexList));
     } else {
-      OpenDexFiles(GetBootClassPath(), GetBootClassPathLocations(), &boot_class_path);
+      OpenDexFiles(ArrayRef<const std::string>(GetBootClassPath()),
+                   ArrayRef<const std::string>(GetBootClassPathLocations()),
+                   &boot_class_path);
     }
     if (!class_linker_->InitWithoutImage(std::move(boot_class_path), &error_msg)) {
       LOG(ERROR) << "Could not initialize without image: " << error_msg;
diff --git a/runtime/thread_pool.cc b/runtime/thread_pool.cc
index de698c2..e1c756d 100644
--- a/runtime/thread_pool.cc
+++ b/runtime/thread_pool.cc
@@ -1,3 +1,4 @@
+
 /*
  * Copyright (C) 2012 The Android Open Source Project
  *
@@ -86,7 +87,7 @@
 void ThreadPoolWorker::Run() {
   Thread* self = Thread::Current();
   Task* task = nullptr;
-  thread_pool_->creation_barier_.Wait(self);
+  thread_pool_->creation_barier_.Pass(self);
   while ((task = thread_pool_->GetTask(self)) != nullptr) {
     task->Run(self);
     task->Finalize();
@@ -150,7 +151,7 @@
     MutexLock mu(self, task_queue_lock_);
     shutting_down_ = false;
     // Add one since the caller of constructor waits on the barrier too.
-    creation_barier_.Init(self, max_active_workers_ + 1);
+    creation_barier_.Init(self, max_active_workers_);
     while (GetThreadCount() < max_active_workers_) {
       const std::string worker_name = StringPrintf("%s worker thread %zu", name_.c_str(),
                                                    GetThreadCount());
@@ -158,8 +159,16 @@
           new ThreadPoolWorker(this, worker_name, worker_stack_size_));
     }
   }
-  // Wait for all of the threads to attach.
-  creation_barier_.Wait(Thread::Current());
+}
+
+void ThreadPool::WaitForWorkersToBeCreated() {
+  creation_barier_.Increment(Thread::Current(), 0);
+}
+
+const std::vector<ThreadPoolWorker*>& ThreadPool::GetWorkers() {
+  // Wait for all the workers to be created before returning them.
+  WaitForWorkersToBeCreated();
+  return threads_;
 }
 
 void ThreadPool::DeleteThreads() {
diff --git a/runtime/thread_pool.h b/runtime/thread_pool.h
index f55d72e..0a2a50c 100644
--- a/runtime/thread_pool.h
+++ b/runtime/thread_pool.h
@@ -101,9 +101,7 @@
     return threads_.size();
   }
 
-  const std::vector<ThreadPoolWorker*>& GetWorkers() const {
-    return threads_;
-  }
+  const std::vector<ThreadPoolWorker*>& GetWorkers();
 
   // Broadcast to the workers and tell them to empty out the work queue.
   void StartWorkers(Thread* self) REQUIRES(!task_queue_lock_);
@@ -154,6 +152,9 @@
   // Set the "nice" priorty for threads in the pool.
   void SetPthreadPriority(int priority);
 
+  // Wait for workers to be created.
+  void WaitForWorkersToBeCreated();
+
  protected:
   // get a task to run, blocks if there are no tasks left
   virtual Task* GetTask(Thread* self) REQUIRES(!task_queue_lock_);
diff --git a/test/175-alloc-big-bignums/expected.txt b/test/175-alloc-big-bignums/expected.txt
new file mode 100644
index 0000000..f75da10
--- /dev/null
+++ b/test/175-alloc-big-bignums/expected.txt
@@ -0,0 +1 @@
+Test complete
diff --git a/test/175-alloc-big-bignums/info.txt b/test/175-alloc-big-bignums/info.txt
new file mode 100644
index 0000000..8f6bcc3
--- /dev/null
+++ b/test/175-alloc-big-bignums/info.txt
@@ -0,0 +1,11 @@
+Allocate large numbers of huge BigIntegers in rapid succession. Most of the
+associated memory will be in the C++ heap. This makes sure that we trigger
+the garbage collector often enough to prevent us from running out of memory.
+
+The test allocates roughly 10GB of native memory, approximately 1MB of which
+will be live at any point. Basically all native memory deallocation is
+triggered by Java garbage collection.
+
+This test is a lot nastier than it looks. In particular, failure on target tends
+to exhaust device memory, and kill off all processes on the device, including the
+adb daemon :-( .
diff --git a/test/175-alloc-big-bignums/src/Main.java b/test/175-alloc-big-bignums/src/Main.java
new file mode 100644
index 0000000..5fbeb46
--- /dev/null
+++ b/test/175-alloc-big-bignums/src/Main.java
@@ -0,0 +1,38 @@
+/*
+ * Copyright (C) 2018 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.
+ */
+
+import java.math.BigInteger;
+
+// This is motivated by the assumption that BigInteger allocates malloc memory
+// underneath. That's true (in 2018) on Android.
+
+public class Main {
+  public static void main(String[] args) throws Exception {
+    final int nIters = 20_000;  // Presumed < 1_000_000.
+    final BigInteger big2_20 = BigInteger.valueOf(1024*1024); // 2^20
+    BigInteger huge = BigInteger.valueOf(1).shiftLeft(4_000_000);  // ~0.5MB
+    for (int i = 0; i < nIters; ++i) {  // 10 GB total
+      huge = huge.add(BigInteger.ONE);
+    }
+    if (huge.bitLength() != 4_000_001) {
+      System.out.println("Wrong answer length: " + huge.bitLength());
+    } else if (huge.mod(big2_20).compareTo(BigInteger.valueOf(nIters)) != 0) {
+      System.out.println("Wrong answer: ..." + huge.mod(big2_20));
+    } else {
+      System.out.println("Test complete");
+    }
+  }
+}
diff --git a/test/1934-jvmti-signal-thread/signal_threads.cc b/test/1934-jvmti-signal-thread/signal_threads.cc
index 726a7a86..dfb08c1 100644
--- a/test/1934-jvmti-signal-thread/signal_threads.cc
+++ b/test/1934-jvmti-signal-thread/signal_threads.cc
@@ -47,19 +47,19 @@
                             jvmti_env,
                             jvmti_env->Allocate(sizeof(NativeMonitor),
                                                 reinterpret_cast<unsigned char**>(&mon)))) {
-    return -1l;
+    return -1L;
   }
   if (JvmtiErrorToException(env,
                             jvmti_env,
                             jvmti_env->CreateRawMonitor("test-1934 start",
                                                         &mon->start_monitor))) {
-    return -1l;
+    return -1L;
   }
   if (JvmtiErrorToException(env,
                             jvmti_env,
                             jvmti_env->CreateRawMonitor("test-1934 continue",
                                                         &mon->continue_monitor))) {
-    return -1l;
+    return -1L;
   }
   mon->should_continue = false;
   mon->should_start = false;
@@ -92,7 +92,7 @@
   while (!mon->should_continue) {
     if (JvmtiErrorToException(env,
                               jvmti_env,
-                              jvmti_env->RawMonitorWait(mon->continue_monitor, -1l))) {
+                              jvmti_env->RawMonitorWait(mon->continue_monitor, -1L))) {
       JvmtiErrorToException(env, jvmti_env, jvmti_env->RawMonitorExit(mon->continue_monitor));
       return;
     }
@@ -112,7 +112,7 @@
   while (!mon->should_start) {
     if (JvmtiErrorToException(env,
                               jvmti_env,
-                              jvmti_env->RawMonitorWait(mon->start_monitor, -1l))) {
+                              jvmti_env->RawMonitorWait(mon->start_monitor, -1L))) {
       return;
     }
   }
diff --git a/test/566-polymorphic-inlining/polymorphic_inline.cc b/test/566-polymorphic-inlining/polymorphic_inline.cc
index 17ccd9a..00827cf 100644
--- a/test/566-polymorphic-inlining/polymorphic_inline.cc
+++ b/test/566-polymorphic-inlining/polymorphic_inline.cc
@@ -46,7 +46,7 @@
       usleep(1000);
     }
     // Will either ensure it's compiled or do the compilation itself.
-    jit->CompileMethod(method, soa.Self(), /* osr */ false);
+    jit->CompileMethod(method, soa.Self(), /*baseline=*/ false, /*osr=*/ false);
   }
 
   CodeInfo info(header);
diff --git a/test/570-checker-osr/osr.cc b/test/570-checker-osr/osr.cc
index b2b3634..dc0e94c 100644
--- a/test/570-checker-osr/osr.cc
+++ b/test/570-checker-osr/osr.cc
@@ -128,7 +128,7 @@
           // Sleep to yield to the compiler thread.
           usleep(1000);
           // Will either ensure it's compiled or do the compilation itself.
-          jit->CompileMethod(m, Thread::Current(), /* osr */ true);
+          jit->CompileMethod(m, Thread::Current(), /*baseline=*/ false, /*osr=*/ true);
         }
       });
 }
diff --git a/test/626-const-class-linking/clear_dex_cache_types.cc b/test/626-const-class-linking/clear_dex_cache_types.cc
index 52367c7..82c82c6 100644
--- a/test/626-const-class-linking/clear_dex_cache_types.cc
+++ b/test/626-const-class-linking/clear_dex_cache_types.cc
@@ -41,6 +41,7 @@
   if (status == ClassStatus::kResolved) {
     ObjectLock<mirror::Class> lock(soa.Self(), klass);
     klass->SetStatus(klass, ClassStatus::kVerified, soa.Self());
+    klass->SetVerificationAttempted();
   } else {
     LOG(ERROR) << klass->PrettyClass() << " has unexpected status: " << status;
   }
diff --git a/test/Android.run-test.mk b/test/Android.run-test.mk
index 5d07601..e3157ef 100644
--- a/test/Android.run-test.mk
+++ b/test/Android.run-test.mk
@@ -47,12 +47,8 @@
 # Also need libopenjdkjvmti.
 ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += libopenjdkjvmti-target libopenjdkjvmtid-target
 
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/core-libart-testdex.jar
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/core-oj-testdex.jar
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/core-simple-testdex.jar
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/conscrypt-testdex.jar
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/okhttp-testdex.jar
-ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += $(TARGET_OUT_JAVA_LIBRARIES)/bouncycastle-testdex.jar
+ART_TEST_TARGET_RUN_TEST_DEPENDENCIES += \
+  $(foreach jar,$(TARGET_TEST_CORE_JARS),$(TARGET_OUT_JAVA_LIBRARIES)/$(jar).jar)
 
 # All tests require the host executables. The tests also depend on the core images, but on
 # specific version depending on the compiler.
@@ -74,6 +70,7 @@
   $(ART_HOST_OUT_SHARED_LIBRARIES)/libopenjdkd$(ART_HOST_SHLIB_EXTENSION) \
   $(ART_HOST_OUT_SHARED_LIBRARIES)/libopenjdkjvmti$(ART_HOST_SHLIB_EXTENSION) \
   $(ART_HOST_OUT_SHARED_LIBRARIES)/libopenjdkjvmtid$(ART_HOST_SHLIB_EXTENSION) \
+  $(HOST_CORE_DEX_LOCATIONS) \
 
 ifneq ($(HOST_PREFER_32_BIT),true)
 ART_TEST_HOST_RUN_TEST_DEPENDENCIES += \
diff --git a/test/common/runtime_state.cc b/test/common/runtime_state.cc
index 65127fc..55631a9 100644
--- a/test/common/runtime_state.cc
+++ b/test/common/runtime_state.cc
@@ -227,7 +227,7 @@
       // Make sure there is a profiling info, required by the compiler.
       ProfilingInfo::Create(self, method, /* retry_allocation */ true);
       // Will either ensure it's compiled or do the compilation itself.
-      jit->CompileMethod(method, self, /* osr */ false);
+      jit->CompileMethod(method, self, /*baseline=*/ false, /*osr=*/ false);
     }
   }
 }
diff --git a/test/etc/run-test-jar b/test/etc/run-test-jar
index 16106ab..4c31ee5 100755
--- a/test/etc/run-test-jar
+++ b/test/etc/run-test-jar
@@ -548,7 +548,10 @@
   exit
 fi
 
-bpath_modules="core-oj core-libart core-simple conscrypt okhttp bouncycastle"
+# Note: This must start with the CORE_IMG_JARS in Android.common_path.mk
+# because that's what we use for compiling the core.art image.
+# It may contain additional modules from TEST_CORE_JARS.
+bpath_modules="core-oj core-libart core-simple okhttp bouncycastle conscrypt"
 if [ "${HOST}" = "y" ]; then
     framework="${ANDROID_HOST_OUT}/framework"
     if [ "${ANDROID_HOST_OUT:0:${#ANDROID_BUILD_TOP}+1}" = "${ANDROID_BUILD_TOP}/" ]; then
diff --git a/test/knownfailures.json b/test/knownfailures.json
index ae20557..5bcd1c0 100644
--- a/test/knownfailures.json
+++ b/test/knownfailures.json
@@ -447,19 +447,63 @@
     },
     {
         "tests": [
-            "137-cfi",
-            "595-profile-saving",
-            "900-hello-plugin",
-            "909-attach-agent",
-            "981-dedup-original-dex",
-            "1900-track-alloc"
+            "004-ThreadStress",
+            "130-hprof",
+            "579-inline-infinite",
+            "1946-list-descriptors"
         ],
-        "description": ["Tests that require exact knowledge of the number of plugins and agents."],
+        "description": ["Too slow to finish in the timeout"],
         "variant": "jvmti-stress | redefine-stress | trace-stress | field-stress | step-stress"
     },
     {
         "tests": [
+            "911-get-stack-trace"
+        ],
+        "description": ["Tests that fail when run with step-stress for unknown reasons."],
+        "bug": "b/120995005",
+        "variant": "jvmti-stress | step-stress"
+    },
+    {
+        "tests": [
+            "004-SignalTest",
+            "004-StackWalk",
+            "064-field-access",
+            "083-compiler-regressions",
+            "098-ddmc",
+            "107-int-math2",
+            "129-ThreadGetId",
+            "135-MirandaDispatch",
             "132-daemon-locks-shutdown",
+            "163-app-image-methods",
+            "607-daemon-stress",
+            "674-hiddenapi",
+            "687-deopt",
+            "904-object-allocation"
+        ],
+        "description": ["Tests that sometimes fail when run with jvmti-stress for unknown reasons."],
+        "bug": "b/120995005",
+        "variant": "jvmti-stress | trace-stress | field-stress | step-stress"
+    },
+    {
+        "tests": [
+            "018-stack-overflow",
+            "137-cfi",
+            "595-profile-saving",
+            "597-deopt-busy-loop",
+            "597-deopt-new-string",
+            "660-clinit",
+            "900-hello-plugin",
+            "909-attach-agent",
+            "924-threads",
+            "981-dedup-original-dex",
+            "1900-track-alloc"
+        ],
+        "description": ["Tests that require exact knowledge of the deoptimization state, the ",
+                        "number of plugins and agents, or breaks other openjdkjvmti assumptions."],
+        "variant": "jvmti-stress | redefine-stress | trace-stress | field-stress | step-stress"
+    },
+    {
+        "tests": [
             "607-daemon-stress",
             "602-deoptimizeable",
             "121-simple-suspend-check",
@@ -569,6 +613,12 @@
         "env_vars": {"SANITIZE_HOST": "address"}
     },
     {
+        "tests": "175-alloc-big-bignums",
+        "description": "ASAN runs out of memory due to huge allocations.",
+        "variant": "host",
+        "env_vars": {"SANITIZE_HOST": "address"}
+    },
+    {
         "tests": "202-thread-oome",
         "description": "ASAN aborts when large thread stacks are requested.",
         "variant": "host",
diff --git a/test/ti-stress/stress.cc b/test/ti-stress/stress.cc
index e123e9f..cd7af10 100644
--- a/test/ti-stress/stress.cc
+++ b/test/ti-stress/stress.cc
@@ -157,14 +157,12 @@
       : jvmtienv_(jvmtienv),
         class_(c),
         name_(nullptr),
-        generic_(nullptr),
         file_(nullptr),
         debug_ext_(nullptr) {}
 
   ~ScopedClassInfo() {
     if (class_ != nullptr) {
       jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(name_));
-      jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(generic_));
       jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(file_));
       jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(debug_ext_));
     }
@@ -173,12 +171,11 @@
   bool Init() {
     if (class_ == nullptr) {
       name_ = const_cast<char*>("<NONE>");
-      generic_ = const_cast<char*>("<NONE>");
       return true;
     } else {
       jvmtiError ret1 = jvmtienv_->GetSourceFileName(class_, &file_);
       jvmtiError ret2 = jvmtienv_->GetSourceDebugExtension(class_, &debug_ext_);
-      return jvmtienv_->GetClassSignature(class_, &name_, &generic_) == JVMTI_ERROR_NONE &&
+      return jvmtienv_->GetClassSignature(class_, &name_, nullptr) == JVMTI_ERROR_NONE &&
           ret1 != JVMTI_ERROR_MUST_POSSESS_CAPABILITY &&
           ret1 != JVMTI_ERROR_INVALID_CLASS &&
           ret2 != JVMTI_ERROR_MUST_POSSESS_CAPABILITY &&
@@ -192,9 +189,6 @@
   const char* GetName() const {
     return name_;
   }
-  const char* GetGeneric() const {
-    return generic_;
-  }
   const char* GetSourceDebugExtension() const {
     if (debug_ext_ == nullptr) {
       return "<UNKNOWN_SOURCE_DEBUG_EXTENSION>";
@@ -214,7 +208,6 @@
   jvmtiEnv* jvmtienv_;
   jclass class_;
   char* name_;
-  char* generic_;
   char* file_;
   char* debug_ext_;
 };
@@ -229,14 +222,12 @@
         class_info_(nullptr),
         name_(nullptr),
         signature_(nullptr),
-        generic_(nullptr),
         first_line_(-1) {}
 
   ~ScopedMethodInfo() {
     DeleteLocalRef(env_, declaring_class_);
     jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(name_));
     jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(signature_));
-    jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(generic_));
   }
 
   bool Init() {
@@ -257,7 +248,7 @@
       return false;
     }
     return class_info_->Init() &&
-        (jvmtienv_->GetMethodName(method_, &name_, &signature_, &generic_) == JVMTI_ERROR_NONE);
+        (jvmtienv_->GetMethodName(method_, &name_, &signature_, nullptr) == JVMTI_ERROR_NONE);
   }
 
   const ScopedClassInfo& GetDeclaringClassInfo() const {
@@ -276,10 +267,6 @@
     return signature_;
   }
 
-  const char* GetGeneric() const {
-    return generic_;
-  }
-
   jint GetFirstLine() const {
     return first_line_;
   }
@@ -292,7 +279,6 @@
   std::unique_ptr<ScopedClassInfo> class_info_;
   char* name_;
   char* signature_;
-  char* generic_;
   jint first_line_;
 
   friend std::ostream& operator<<(std::ostream &os, ScopedMethodInfo const& m);
@@ -306,20 +292,18 @@
         field_(field),
         class_info_(nullptr),
         name_(nullptr),
-        type_(nullptr),
-        generic_(nullptr) {}
+        type_(nullptr) {}
 
   ~ScopedFieldInfo() {
     jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(name_));
     jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(type_));
-    jvmtienv_->Deallocate(reinterpret_cast<unsigned char*>(generic_));
   }
 
   bool Init() {
     class_info_.reset(new ScopedClassInfo(jvmtienv_, declaring_class_));
     return class_info_->Init() &&
         (jvmtienv_->GetFieldName(
-            declaring_class_, field_, &name_, &type_, &generic_) == JVMTI_ERROR_NONE);
+            declaring_class_, field_, &name_, &type_, nullptr) == JVMTI_ERROR_NONE);
   }
 
   const ScopedClassInfo& GetDeclaringClassInfo() const {
@@ -338,10 +322,6 @@
     return type_;
   }
 
-  const char* GetGeneric() const {
-    return generic_;
-  }
-
  private:
   jvmtiEnv* jvmtienv_;
   jclass declaring_class_;
@@ -349,7 +329,6 @@
   std::unique_ptr<ScopedClassInfo> class_info_;
   char* name_;
   char* type_;
-  char* generic_;
 
   friend std::ostream& operator<<(std::ostream &os, ScopedFieldInfo const& m);
 };
diff --git a/tools/bootjars.sh b/tools/bootjars.sh
index ad6ee6b..9f22827 100755
--- a/tools/bootjars.sh
+++ b/tools/bootjars.sh
@@ -72,8 +72,10 @@
   # FIXME: The soong invocation we're using for getting the variables does not give us anything
   # defined in Android.common_path.mk, otherwise we would just use HOST-/TARGET_TEST_CORE_JARS.
 
-  # The core_jars_list must match the TEST_CORE_JARS variable in the Android.common_path.mk .
-  core_jars_list="core-oj core-libart core-simple conscrypt okhttp bouncycastle"
+  # Note: This must start with the CORE_IMG_JARS in Android.common_path.mk
+  # because that's what we use for compiling the core.art image.
+  # It may contain additional modules from TEST_CORE_JARS.
+  core_jars_list="core-oj core-libart core-simple"
   core_jars_suffix=
   if [[ $mode == target ]]; then
     core_jars_suffix=-testdex
diff --git a/tools/dist_linunx_bionic.sh b/tools/dist_linux_bionic.sh
similarity index 100%
rename from tools/dist_linunx_bionic.sh
rename to tools/dist_linux_bionic.sh
diff --git a/tools/run-jdwp-tests.sh b/tools/run-jdwp-tests.sh
index f4a2dc1..c85a5ed 100755
--- a/tools/run-jdwp-tests.sh
+++ b/tools/run-jdwp-tests.sh
@@ -55,9 +55,10 @@
   done
 }
 
-# Note: This must match the TEST_CORE_JARS in Android.common_path.mk
+# Note: This must start with the CORE_IMG_JARS in Android.common_path.mk
 # because that's what we use for compiling the core.art image.
-BOOT_CLASSPATH_JARS="core-oj core-libart core-simple conscrypt okhttp bouncycastle"
+# It may contain additional modules from TEST_CORE_JARS.
+BOOT_CLASSPATH_JARS="core-oj core-libart core-simple okhttp bouncycastle conscrypt"
 
 vm_args=""
 art="$android_root/bin/art"
diff --git a/tools/run-libcore-tests.sh b/tools/run-libcore-tests.sh
index 63f1fce..63fe81b 100755
--- a/tools/run-libcore-tests.sh
+++ b/tools/run-libcore-tests.sh
@@ -57,9 +57,10 @@
   done
 }
 
-# Note: This must match the TEST_CORE_JARS in Android.common_path.mk
+# Note: This must start with the CORE_IMG_JARS in Android.common_path.mk
 # because that's what we use for compiling the core.art image.
-BOOT_CLASSPATH_JARS="core-oj core-libart core-simple conscrypt okhttp bouncycastle"
+# It may contain additional modules from TEST_CORE_JARS.
+BOOT_CLASSPATH_JARS="core-oj core-libart core-simple okhttp bouncycastle conscrypt"
 
 DEPS="core-tests jsr166-tests mockito-target"