diff --git a/runtime/src/main/cpp/Porting.cpp b/runtime/src/main/cpp/Porting.cpp
index 4158125e6aa..797764e6df7 100644
--- a/runtime/src/main/cpp/Porting.cpp
+++ b/runtime/src/main/cpp/Porting.cpp
@@ -22,6 +22,9 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#if !KONAN_NO_THREADS
+#include <pthread.h>
+#endif
 #include <unistd.h>
 #if KONAN_WINDOWS
 #include <windows.h>
@@ -90,6 +93,53 @@ void consolePrintf(const char* format, ...) {
 }
 
 
+// Thread execution.
+#if !KONAN_NO_THREADS
+
+pthread_key_t terminationKey;
+pthread_once_t terminationKeyOnceControl =  PTHREAD_ONCE_INIT;
+
+typedef void (*destructor_t)();
+
+struct DestructorRecord {
+  struct DestructorRecord* next;
+  destructor_t destructor;
+};
+
+static void onThreadExitCallback(void* value) {
+  DestructorRecord* record = reinterpret_cast<DestructorRecord*>(value);
+  while (record != nullptr) {
+    record->destructor();
+    auto next = record->next;
+    free(record);
+    record = next;
+  }
+}
+
+static void onThreadExitInit() {
+  pthread_key_create(&terminationKey, onThreadExitCallback);
+}
+
+#endif  // !KONAN_NO_THREADS
+
+void onThreadExit(void (*destructor)()) {
+#if KONAN_NO_THREADS
+#ifdef KONAN_WASM
+  // No way to do that.
+#else
+  ::atexit(destructor);
+#endif
+#else  // !KONAN_NO_THREADS
+  // We cannot use pthread_cleanup_push() as it is lexical scope bound.
+  pthread_once(&terminationKeyOnceControl, onThreadExitInit);
+  DestructorRecord* destructorRecord = (DestructorRecord*)calloc(1, sizeof(DestructorRecord));
+  destructorRecord->destructor = destructor;
+  destructorRecord->next =
+      reinterpret_cast<DestructorRecord*>(pthread_getspecific(terminationKey));
+  pthread_setspecific(terminationKey, destructorRecord);
+#endif  // !KONAN_NO_THREADS
+}
+
 // Process execution.
 void abort() {
   ::abort();
diff --git a/runtime/src/main/cpp/Porting.h b/runtime/src/main/cpp/Porting.h
index b683bcab6d4..9c6664409f3 100644
--- a/runtime/src/main/cpp/Porting.h
+++ b/runtime/src/main/cpp/Porting.h
@@ -33,6 +33,9 @@ uint32_t consoleReadUtf8(void* utf8, uint32_t maxSizeBytes);
 void abort();
 void exit(int32_t status);
 
+// Thread control.
+void onThreadExit(void (*destructor)());
+
 // String/byte operations.
 // memcpy/memmove/memcmp are not here intentionally, as frequently implemented/optimized
 // by C compiler.
diff --git a/runtime/src/main/cpp/Runtime.cpp b/runtime/src/main/cpp/Runtime.cpp
index cea7cbc2f12..351a918f4ed 100644
--- a/runtime/src/main/cpp/Runtime.cpp
+++ b/runtime/src/main/cpp/Runtime.cpp
@@ -14,8 +14,6 @@
  * limitations under the License.
  */
 
-#include <stdio.h>
-
 #include "Alloc.h"
 #include "Exceptions.h"
 #include "Memory.h"
@@ -82,8 +80,11 @@ void DeinitRuntime(RuntimeState* state) {
 }
 
 void Kotlin_initRuntimeIfNeeded() {
-  if (runtimeState == nullptr)
+  if (runtimeState == nullptr) {
     runtimeState = InitRuntime();
+    // Register runtime deinit function at thread cleanup.
+    konan::onThreadExit(Kotlin_deinitRuntimeIfNeeded);
+  }
 }
 
 void Kotlin_deinitRuntimeIfNeeded() {
diff --git a/runtime/src/main/cpp/Runtime.h b/runtime/src/main/cpp/Runtime.h
index 1d7605fc636..59e237c8134 100644
--- a/runtime/src/main/cpp/Runtime.h
+++ b/runtime/src/main/cpp/Runtime.h
@@ -26,6 +26,8 @@ extern "C" {
 
 RuntimeState* InitRuntime();
 void DeinitRuntime(RuntimeState* state);
+void Kotlin_initRuntimeIfNeeded();
+void Kotlin_deinitRuntimeIfNeeded();
 
 // Appends given node to an initializer list.
 void AppendToInitializersTail(struct InitNode*);