Avoid using __cxa_thread_atexit for pthread_cleanup_push.

These is no need to use `__cxa_thread_atexit` here, we can just call
these functions during thread exit (which is what musl does).

This also avoids references `__cxa_thread_atexit` which is normally a
libc++abi symbol from C programs.

Needed by (split out from) #14489.

Author: sbc100

system/lib/pthread/emscripten_tls_init.c

@@ -19,8 +19,6 @@
 // linker-generated symbol that loads static TLS data at the given location.
 extern void __wasm_init_tls(void *memory);
 
-extern int __cxa_thread_atexit(void (*)(void *), void *, void *);
-
 extern int __dso_handle;
 
 void* emscripten_tls_init(void) {

system/lib/pthread/pthread_create.c

@@ -19,7 +19,6 @@
 
 // See musl's pthread_create.c
 
-extern int __cxa_thread_atexit(void (*)(void *), void *, void *);
 extern int __pthread_create_js(struct pthread *thread, const pthread_attr_t *attr, void *(*start_routine) (void *), void *arg);
 extern void _emscripten_thread_init(int, int, int);
 extern void __pthread_exit_run_handlers();
@@ -32,7 +31,7 @@ static void dummy_0()
 }
 weak_alias(dummy_0, __pthread_tsd_run_dtors);
 
-void __run_cleanup_handlers(void* _unused) {
+static void __run_cleanup_handlers() {
   pthread_t self = __pthread_self();
   while (self->cancelbuf) {
     void (*f)(void *) = self->cancelbuf->__f;
@@ -46,11 +45,6 @@ void __do_cleanup_push(struct __ptcb *cb) {
   struct pthread *self = __pthread_self();
   cb->__next = self->cancelbuf;
   self->cancelbuf = cb;
-  static thread_local bool registered = false;
-  if (!registered) {
-    __cxa_thread_atexit(__run_cleanup_handlers, NULL, &__dso_handle);
-    registered = true;
-  }
 }
 
 void __do_cleanup_pop(struct __ptcb *cb) {
@@ -128,6 +122,11 @@ void _emscripten_thread_exit(void* result) {
   self->cancelasync = PTHREAD_CANCEL_DEFERRED;
   self->result = result;
 
+  // Run any handlers registered with pthread_cleanup_push
+  __run_cleanup_handlers();
+
+  // Run any JS thread exit handlers (for C++ programs this includes any
+  // functions registered with __cxa_thread_atexit).
   __pthread_exit_run_handlers();
 
   // Call into the musl function that runs destructors of all thread-specific data.

tests/other/metadce/minimal_main_Oz_USE_PTHREADS_PROXY_TO_PTHREAD.exports

@@ -13,7 +13,7 @@ L
 M
 N
 O
-P
+u
 v
 w
 x

tests/other/metadce/minimal_main_Oz_USE_PTHREADS_PROXY_TO_PTHREAD.imports

@@ -18,4 +18,3 @@ a.q
 a.r
 a.s
 a.t
-a.u

tests/other/metadce/minimal_main_Oz_USE_PTHREADS_PROXY_TO_PTHREAD.sent

@@ -18,4 +18,3 @@ q
 r
 s
 t
-u