diff --git a/libc/bionic/ndk_cruft.cpp b/libc/bionic/ndk_cruft.cpp
index 109c523..8b34495 100644
--- a/libc/bionic/ndk_cruft.cpp
+++ b/libc/bionic/ndk_cruft.cpp
@@ -346,6 +346,14 @@
   return malloc(size);
 }
 
+#define __get_thread __real_get_thread
+#include "pthread_internal.h"
+#undef __get_thread
+// Various third-party apps contain a backport of our pthread_rwlock implementation that uses this.
+extern "C" pthread_internal_t* __get_thread() {
+  return __real_get_thread();
+}
+
 #endif // !defined(__LP64__)
 
 // This is never implemented in bionic, only needed for ABI compatibility with the NDK.
diff --git a/libc/bionic/pthread_internal.h b/libc/bionic/pthread_internal.h
index 8da99dd..99882ae 100644
--- a/libc/bionic/pthread_internal.h
+++ b/libc/bionic/pthread_internal.h
@@ -113,8 +113,10 @@
 __LIBC_HIDDEN__ void __init_alternate_signal_stack(pthread_internal_t*);
 __LIBC_HIDDEN__ void _pthread_internal_add(pthread_internal_t* thread);
 
-/* Various third-party apps contain a backport of our pthread_rwlock implementation that uses this. */
-extern "C" __LIBC64_HIDDEN__ pthread_internal_t* __get_thread(void);
+// Make __get_thread() inlined for performance reason. See http://b/19825434.
+static inline __always_inline pthread_internal_t* __get_thread() {
+  return reinterpret_cast<pthread_internal_t*>(__get_tls()[TLS_SLOT_THREAD_ID]);
+}
 
 __LIBC_HIDDEN__ void pthread_key_clean_all(void);
 __LIBC_HIDDEN__ void _pthread_internal_remove_locked(pthread_internal_t* thread, bool free_thread);
diff --git a/libc/bionic/pthread_internals.cpp b/libc/bionic/pthread_internals.cpp
index 14061d1..0dd88fe 100644
--- a/libc/bionic/pthread_internals.cpp
+++ b/libc/bionic/pthread_internals.cpp
@@ -68,7 +68,3 @@
   }
   g_thread_list = thread;
 }
-
-pthread_internal_t* __get_thread(void) {
-  return reinterpret_cast<pthread_internal_t*>(__get_tls()[TLS_SLOT_THREAD_ID]);
-}