Replace call_once usage in libc++ with atomics

This avoid a dependency on pthread API (which are used to implement
call_once under the hood) which makes libc++ much more usable from
Wasm Workers (there pthreads are not available).

The downside here is that when thread race to be first one to use a
given facet we would "leak" an ID which could cause the `facets_` vector
to become more spare that it otherwise would be.

IIUC this `facets_` vector can already be sparse so it not clear to
me what impact this would have in practice.

We could make this change only for the Wasm Workers build of libc++
but I'd like to get as much test coverage of it as I can.

See: #26375

Author: sbc100

system/lib/libcxx/include/__locale

@@ -146,7 +146,9 @@ private:
 };
 
 class _LIBCPP_EXPORTED_FROM_ABI locale::id {
+#if !defined(__EMSCRIPTEN__) || !defined(__EMSCRIPTEN_WASM_WORKERS__)
   once_flag __flag_;
+#endif
   int32_t __id_;
 
   static int32_t __next_id;

system/lib/libcxx/src/locale.cpp

@@ -589,7 +589,24 @@ void locale::facet::__on_zero_shared() noexcept { delete this; }
 constinit int32_t locale::id::__next_id = 0;
 
 long locale::id::__get() {
+#if defined(__EMSCRIPTEN__) && defined(__EMSCRIPTEN_WASM_WORKERS__)
+  // Avoid `call_once` under Emscripten since we want to avoid the pthread
+  // dependenency that it comes with.
+  // TODO(https://github.com/emscripten-core/emscripten/issues/26426):
+  // Remove this patch once we have some kind of locking primitive
+  // in Wasm Worker that can be used to implement call_once (or
+  // __libcpp_mutex_t/__libcpp_condvar_t).
+  if (__libcpp_atomic_load(&__id_) == 0) {
+    int32_t proposed_id = __libcpp_atomic_add(&__next_id, 1);
+    int32_t expected = 0;
+    // If we race with another thread here the CAS will fail and
+    // the proposed_id will be leaked, but __id_ will be non-zero
+    // in either case.
+    __libcpp_atomic_compare_exchange(&__id_, &expected, proposed_id);
+  }
+#else
   call_once(__flag_, [&] { __id_ = __libcpp_atomic_add(&__next_id, 1); });
+#endif
   return __id_ - 1;
 }
 

test/codesize/hello_libcxx.cpp

@@ -9,4 +9,3 @@ int main() {
   std::cout << "hello, world!" << std::endl;
   return 0;
 }
-

test/test_other.py

@@ -13537,6 +13537,9 @@ def test_wasm_worker_pthread_api_usage(self):
   def test_wasm_worker_cxx_init(self):
     self.do_run_in_out_file_test('wasm_worker/wasm_worker_cxx_init.cpp', cflags=['-sWASM_WORKERS'])
 
+  def test_wasm_worker_hello_libcxx(self):
+    self.do_runf('codesize/hello_libcxx.cpp', 'hello, world!\n', cflags=['-sWASM_WORKERS'])
+
   @parameterized({
     # we will warn here since -O2 runs the optimizer and -g enables DWARF
     'O2_g': (True, ['-O2', '-g']),