diff --git a/dex2oat/linker/image_writer.h b/dex2oat/linker/image_writer.h
index 7bdaebe..c35dbaa 100644
--- a/dex2oat/linker/image_writer.h
+++ b/dex2oat/linker/image_writer.h
@@ -26,6 +26,7 @@
 #include <set>
 #include <stack>
 #include <string>
+#include <unordered_map>
 
 #include "art_method.h"
 #include "base/bit_utils.h"
diff --git a/runtime/class_linker-inl.h b/runtime/class_linker-inl.h
index 0dc62d3..7e01137 100644
--- a/runtime/class_linker-inl.h
+++ b/runtime/class_linker-inl.h
@@ -22,7 +22,6 @@
 #include "art_field-inl.h"
 #include "art_method-inl.h"
 #include "class_linker.h"
-#include "gc/heap-inl.h"
 #include "gc_root-inl.h"
 #include "handle_scope-inl.h"
 #include "mirror/class_loader.h"
@@ -439,6 +438,14 @@
   }
 }
 
+template <ReadBarrierOption kReadBarrierOption>
+inline ObjPtr<mirror::ObjectArray<mirror::Class>> ClassLinker::GetClassRoots() {
+  ObjPtr<mirror::ObjectArray<mirror::Class>> class_roots =
+      class_roots_.Read<kReadBarrierOption>();
+  DCHECK(class_roots != nullptr);
+  return class_roots;
+}
+
 }  // namespace art
 
 #endif  // ART_RUNTIME_CLASS_LINKER_INL_H_
diff --git a/runtime/class_linker.h b/runtime/class_linker.h
index cc17d57..0ae7a9b 100644
--- a/runtime/class_linker.h
+++ b/runtime/class_linker.h
@@ -556,12 +556,7 @@
       REQUIRES_SHARED(Locks::mutator_lock_);
 
   template <ReadBarrierOption kReadBarrierOption = kWithReadBarrier>
-  ObjPtr<mirror::ObjectArray<mirror::Class>> GetClassRoots() REQUIRES_SHARED(Locks::mutator_lock_) {
-    ObjPtr<mirror::ObjectArray<mirror::Class>> class_roots =
-        class_roots_.Read<kReadBarrierOption>();
-    DCHECK(class_roots != nullptr);
-    return class_roots;
-  }
+  ObjPtr<mirror::ObjectArray<mirror::Class>> GetClassRoots() REQUIRES_SHARED(Locks::mutator_lock_);
 
   // Move the class table to the pre-zygote table to reduce memory usage. This works by ensuring
   // that no more classes are ever added to the pre zygote table which makes it that the pages
diff --git a/runtime/class_root.h b/runtime/class_root.h
index 19a78b1..1cd135f 100644
--- a/runtime/class_root.h
+++ b/runtime/class_root.h
@@ -17,7 +17,8 @@
 #ifndef ART_RUNTIME_CLASS_ROOT_H_
 #define ART_RUNTIME_CLASS_ROOT_H_
 
-#include "class_linker.h"
+#include "class_linker-inl.h"
+#include "gc_root-inl.h"
 #include "mirror/class.h"
 #include "mirror/object_array-inl.h"
 #include "obj_ptr-inl.h"
diff --git a/runtime/class_table-inl.h b/runtime/class_table-inl.h
index 6b6fe34..a2cdb2c 100644
--- a/runtime/class_table-inl.h
+++ b/runtime/class_table-inl.h
@@ -19,6 +19,7 @@
 
 #include "class_table.h"
 
+#include "base/mutex-inl.h"
 #include "gc_root-inl.h"
 #include "mirror/class.h"
 #include "oat_file.h"
diff --git a/runtime/gc/allocator/rosalloc.cc b/runtime/gc/allocator/rosalloc.cc
index 8cc0c4e..4e2cf2b 100644
--- a/runtime/gc/allocator/rosalloc.cc
+++ b/runtime/gc/allocator/rosalloc.cc
@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "rosalloc.h"
+#include "rosalloc-inl.h"
 
 #include <list>
 #include <map>
diff --git a/runtime/gc/collector/concurrent_copying-inl.h b/runtime/gc/collector/concurrent_copying-inl.h
index 8fd235f..6b394c7 100644
--- a/runtime/gc/collector/concurrent_copying-inl.h
+++ b/runtime/gc/collector/concurrent_copying-inl.h
@@ -22,7 +22,7 @@
 #include "gc/accounting/atomic_stack.h"
 #include "gc/accounting/space_bitmap-inl.h"
 #include "gc/heap.h"
-#include "gc/space/region_space.h"
+#include "gc/space/region_space-inl.h"
 #include "gc/verification.h"
 #include "lock_word.h"
 #include "mirror/class.h"
diff --git a/runtime/gc/space/malloc_space.cc b/runtime/gc/space/malloc_space.cc
index be75efe..189aeb5 100644
--- a/runtime/gc/space/malloc_space.cc
+++ b/runtime/gc/space/malloc_space.cc
@@ -19,6 +19,7 @@
 #include "android-base/stringprintf.h"
 
 #include "base/logging.h"  // For VLOG
+#include "base/mutex-inl.h"
 #include "base/utils.h"
 #include "gc/accounting/card_table-inl.h"
 #include "gc/accounting/space_bitmap-inl.h"
diff --git a/runtime/gc/space/region_space-inl.h b/runtime/gc/space/region_space-inl.h
index 8cb079d..5ff1270 100644
--- a/runtime/gc/space/region_space-inl.h
+++ b/runtime/gc/space/region_space-inl.h
@@ -18,6 +18,10 @@
 #define ART_RUNTIME_GC_SPACE_REGION_SPACE_INL_H_
 
 #include "region_space.h"
+
+#include "base/mutex-inl.h"
+#include "mirror/object-inl.h"
+#include "region_space.h"
 #include "thread-current-inl.h"
 
 namespace art {
@@ -236,6 +240,15 @@
   }
 }
 
+template <typename Visitor>
+inline void RegionSpace::Walk(Visitor&& visitor) {
+  WalkInternal</* kToSpaceOnly= */ false>(visitor);
+}
+template <typename Visitor>
+inline void RegionSpace::WalkToSpace(Visitor&& visitor) {
+  WalkInternal</* kToSpaceOnly= */ true>(visitor);
+}
+
 inline mirror::Object* RegionSpace::GetNextObject(mirror::Object* obj) {
   const uintptr_t position = reinterpret_cast<uintptr_t>(obj) + obj->SizeOf();
   return reinterpret_cast<mirror::Object*>(RoundUp(position, kAlignment));
diff --git a/runtime/gc/space/region_space.h b/runtime/gc/space/region_space.h
index cc371b8..8810f8c 100644
--- a/runtime/gc/space/region_space.h
+++ b/runtime/gc/space/region_space.h
@@ -205,14 +205,9 @@
 
   // Go through all of the blocks and visit the continuous objects.
   template <typename Visitor>
-  ALWAYS_INLINE void Walk(Visitor&& visitor) REQUIRES(Locks::mutator_lock_) {
-    WalkInternal</* kToSpaceOnly= */ false>(visitor);
-  }
+  ALWAYS_INLINE void Walk(Visitor&& visitor) REQUIRES(Locks::mutator_lock_);
   template <typename Visitor>
-  ALWAYS_INLINE void WalkToSpace(Visitor&& visitor)
-      REQUIRES(Locks::mutator_lock_) {
-    WalkInternal</* kToSpaceOnly= */ true>(visitor);
-  }
+  ALWAYS_INLINE void WalkToSpace(Visitor&& visitor) REQUIRES(Locks::mutator_lock_);
 
   accounting::ContinuousSpaceBitmap::SweepCallback* GetSweepCallback() override {
     return nullptr;
diff --git a/runtime/gc/space/rosalloc_space_random_test.cc b/runtime/gc/space/rosalloc_space_random_test.cc
index ca3aff4..b50859b 100644
--- a/runtime/gc/space/rosalloc_space_random_test.cc
+++ b/runtime/gc/space/rosalloc_space_random_test.cc
@@ -16,6 +16,8 @@
 
 #include "space_test.h"
 
+#include "rosalloc_space.h"
+
 namespace art {
 namespace gc {
 namespace space {
diff --git a/runtime/gc/space/rosalloc_space_static_test.cc b/runtime/gc/space/rosalloc_space_static_test.cc
index a78623e..5e7ced6 100644
--- a/runtime/gc/space/rosalloc_space_static_test.cc
+++ b/runtime/gc/space/rosalloc_space_static_test.cc
@@ -16,6 +16,8 @@
 
 #include "space_test.h"
 
+#include "rosalloc_space.h"
+
 namespace art {
 namespace gc {
 namespace space {
diff --git a/runtime/gc/space/space_test.h b/runtime/gc/space/space_test.h
index c94b666..5aac217 100644
--- a/runtime/gc/space/space_test.h
+++ b/runtime/gc/space/space_test.h
@@ -22,6 +22,7 @@
 
 #include "base/globals.h"
 #include "common_runtime_test.h"
+#include "handle_scope-inl.h"
 #include "mirror/array-inl.h"
 #include "mirror/class-inl.h"
 #include "mirror/class_loader.h"
diff --git a/runtime/jit/jit_code_cache.cc b/runtime/jit/jit_code_cache.cc
index 8600b41..8239602 100644
--- a/runtime/jit/jit_code_cache.cc
+++ b/runtime/jit/jit_code_cache.cc
@@ -39,6 +39,7 @@
 #include "dex/method_reference.h"
 #include "entrypoints/runtime_asm_entrypoints.h"
 #include "gc/accounting/bitmap-inl.h"
+#include "gc/allocator/dlmalloc.h"
 #include "gc/scoped_gc_critical_section.h"
 #include "handle.h"
 #include "instrumentation.h"
diff --git a/runtime/method_handles.cc b/runtime/method_handles.cc
index 86ad32e..7d889c0 100644
--- a/runtime/method_handles.cc
+++ b/runtime/method_handles.cc
@@ -22,6 +22,7 @@
 #include "common_dex_operations.h"
 #include "interpreter/shadow_frame-inl.h"
 #include "jvalue-inl.h"
+#include "mirror/class-inl.h"
 #include "mirror/emulated_stack_frame.h"
 #include "mirror/method_handle_impl-inl.h"
 #include "mirror/method_type.h"
diff --git a/runtime/mirror/array-inl.h b/runtime/mirror/array-inl.h
index 704fb11..d806e40 100644
--- a/runtime/mirror/array-inl.h
+++ b/runtime/mirror/array-inl.h
@@ -27,6 +27,7 @@
 #include "class.h"
 #include "gc/heap-inl.h"
 #include "obj_ptr-inl.h"
+#include "runtime.h"
 #include "thread-current-inl.h"
 
 namespace art {
diff --git a/runtime/mirror/dex_cache-inl.h b/runtime/mirror/dex_cache-inl.h
index 6efb747..13eaf3d 100644
--- a/runtime/mirror/dex_cache-inl.h
+++ b/runtime/mirror/dex_cache-inl.h
@@ -27,12 +27,12 @@
 #include "base/enums.h"
 #include "class_linker.h"
 #include "dex/dex_file.h"
-#include "gc/heap-inl.h"
 #include "gc_root-inl.h"
 #include "mirror/call_site.h"
 #include "mirror/class.h"
 #include "mirror/method_type.h"
 #include "obj_ptr.h"
+#include "object-inl.h"
 #include "runtime.h"
 #include "write_barrier-inl.h"
 
@@ -42,6 +42,27 @@
 namespace mirror {
 
 template <typename T>
+inline DexCachePair<T>::DexCachePair(ObjPtr<T> object, uint32_t index)
+    : object(object), index(index) {}
+
+template <typename T>
+inline void DexCachePair<T>::Initialize(std::atomic<DexCachePair<T>>* dex_cache) {
+  DexCachePair<T> first_elem;
+  first_elem.object = GcRoot<T>(nullptr);
+  first_elem.index = InvalidIndexForSlot(0);
+  dex_cache[0].store(first_elem, std::memory_order_relaxed);
+}
+
+template <typename T>
+inline T* DexCachePair<T>::GetObjectForIndex(uint32_t idx) {
+  if (idx != index) {
+    return nullptr;
+  }
+  DCHECK(!object.IsNull());
+  return object.Read();
+}
+
+template <typename T>
 inline void NativeDexCachePair<T>::Initialize(std::atomic<NativeDexCachePair<T>>* dex_cache,
                                               PointerSize pointer_size) {
   NativeDexCachePair<T> first_elem;
diff --git a/runtime/mirror/dex_cache.h b/runtime/mirror/dex_cache.h
index da1cd3f..6149f9c 100644
--- a/runtime/mirror/dex_cache.h
+++ b/runtime/mirror/dex_cache.h
@@ -21,7 +21,7 @@
 #include "base/bit_utils.h"
 #include "base/mutex.h"
 #include "dex/dex_file_types.h"
-#include "gc_root-inl.h"
+#include "gc_root.h"  // Note: must not use -inl here to avoid circular dependency.
 #include "object.h"
 #include "object_array.h"
 
@@ -67,19 +67,12 @@
   // it's always non-null if the id branch succeeds (except for the 0th id).
   // Set the initial state for the 0th entry to be {0,1} which is guaranteed to fail
   // the lookup id == stored id branch.
-  DexCachePair(ObjPtr<T> object, uint32_t index)
-      : object(object),
-        index(index) {}
+  DexCachePair(ObjPtr<T> object, uint32_t index);
   DexCachePair() : index(0) {}
   DexCachePair(const DexCachePair<T>&) = default;
   DexCachePair& operator=(const DexCachePair<T>&) = default;
 
-  static void Initialize(std::atomic<DexCachePair<T>>* dex_cache) {
-    DexCachePair<T> first_elem;
-    first_elem.object = GcRoot<T>(nullptr);
-    first_elem.index = InvalidIndexForSlot(0);
-    dex_cache[0].store(first_elem, std::memory_order_relaxed);
-  }
+  static void Initialize(std::atomic<DexCachePair<T>>* dex_cache);
 
   static uint32_t InvalidIndexForSlot(uint32_t slot) {
     // Since the cache size is a power of two, 0 will always map to slot 0.
@@ -87,13 +80,7 @@
     return (slot == 0) ? 1u : 0u;
   }
 
-  T* GetObjectForIndex(uint32_t idx) REQUIRES_SHARED(Locks::mutator_lock_) {
-    if (idx != index) {
-      return nullptr;
-    }
-    DCHECK(!object.IsNull());
-    return object.Read();
-  }
+  T* GetObjectForIndex(uint32_t idx) REQUIRES_SHARED(Locks::mutator_lock_);
 };
 
 template <typename T> struct PACKED(2 * __SIZEOF_POINTER__) NativeDexCachePair {
diff --git a/runtime/mirror/object.cc b/runtime/mirror/object.cc
index 025c10b..9b38576 100644
--- a/runtime/mirror/object.cc
+++ b/runtime/mirror/object.cc
@@ -27,7 +27,7 @@
 #include "dex/descriptors_names.h"
 #include "dex/dex_file-inl.h"
 #include "gc/accounting/card_table-inl.h"
-#include "gc/heap.h"
+#include "gc/heap-inl.h"
 #include "handle_scope-inl.h"
 #include "iftable-inl.h"
 #include "monitor.h"
diff --git a/runtime/native/dalvik_system_DexFile.cc b/runtime/native/dalvik_system_DexFile.cc
index 69f7648..f9f87d8 100644
--- a/runtime/native/dalvik_system_DexFile.cc
+++ b/runtime/native/dalvik_system_DexFile.cc
@@ -35,6 +35,7 @@
 #include "dex/descriptors_names.h"
 #include "dex/dex_file-inl.h"
 #include "dex/dex_file_loader.h"
+#include "handle_scope-inl.h"
 #include "jit/debugger_interface.h"
 #include "jni/jni_internal.h"
 #include "mirror/class_loader.h"
diff --git a/runtime/native/java_lang_VMClassLoader.cc b/runtime/native/java_lang_VMClassLoader.cc
index 1ad233a..46162c1 100644
--- a/runtime/native/java_lang_VMClassLoader.cc
+++ b/runtime/native/java_lang_VMClassLoader.cc
@@ -21,6 +21,7 @@
 #include "dex/descriptors_names.h"
 #include "dex/dex_file_loader.h"
 #include "dex/utf.h"
+#include "handle_scope-inl.h"
 #include "jni/jni_internal.h"
 #include "mirror/class_loader.h"
 #include "mirror/object-inl.h"
diff --git a/runtime/non_debuggable_classes.cc b/runtime/non_debuggable_classes.cc
index f42a2d6..8b6c2ed 100644
--- a/runtime/non_debuggable_classes.cc
+++ b/runtime/non_debuggable_classes.cc
@@ -16,6 +16,7 @@
 
 #include "non_debuggable_classes.h"
 
+#include "jni/jni_env_ext.h"
 #include "jni/jni_internal.h"
 #include "mirror/class-inl.h"
 #include "nativehelper/scoped_local_ref.h"
diff --git a/runtime/oat.h b/runtime/oat.h
index 5c5a02d..2a6d738 100644
--- a/runtime/oat.h
+++ b/runtime/oat.h
@@ -23,7 +23,6 @@
 #include "base/macros.h"
 #include "base/safe_map.h"
 #include "compiler_filter.h"
-#include "dex/dex_file.h"
 
 namespace art {
 
diff --git a/runtime/transaction.cc b/runtime/transaction.cc
index 1e5b2bb..e4bf447 100644
--- a/runtime/transaction.cc
+++ b/runtime/transaction.cc
@@ -18,6 +18,7 @@
 
 #include <android-base/logging.h>
 
+#include "base/mutex-inl.h"
 #include "base/stl_util.h"
 #include "gc/accounting/card_table-inl.h"
 #include "gc_root-inl.h"
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 96ef266..52367c7 100644
--- a/test/626-const-class-linking/clear_dex_cache_types.cc
+++ b/test/626-const-class-linking/clear_dex_cache_types.cc
@@ -15,6 +15,7 @@
  */
 
 #include "jni.h"
+#include "handle_scope-inl.h"
 #include "mirror/class-inl.h"
 #include "mirror/class_loader.h"
 #include "mirror/dex_cache-inl.h"