Onde está a corrida neste aviso sanitzer de tópicos?

O código abaixo produz um aviso ao executar com desinfetante de threads no macOS. Não consigo ver onde está a corrida. O bloco de controle shared_ptr e weak_ptr é seguro para threads, e empurrar e saltar do std::queueé feito com um bloqueio mantido.

#include <future>
#include <memory>
#include <queue>

class Foo {
public:
  Foo() {
    fut = std::async(std::launch::async, [this] {
      while (!shouldStop) {
        std::scoped_lock lock(mut);
        while (!requests.empty()) {
          std::weak_ptr<float> requestData = requests.front();
          requests.pop();
          (void)requestData;
        }
      }
    });
  }

  ~Foo() {
    shouldStop.store(true);
    fut.get();
  }

  void add(const std::weak_ptr<float> subscriber) {
    std::scoped_lock lock(mut);
    requests.push(subscriber);
  }

private:
  std::atomic<bool> shouldStop = false;
  std::future<void> fut;
  std::queue<std::weak_ptr<float>> requests;
  std::mutex mut;
};

int main() {
  Foo foo;

  int numIterations = 100000;

  while (--numIterations) {
    auto subscriber = std::make_shared<float>();

    foo.add(subscriber);
    subscriber.reset();
  }

  std::this_thread::sleep_for(std::chrono::seconds(1));
}

Aviso com stacktrace:

WARNING: ThreadSanitizer: data race (pid=11176)
  Write of size 8 at 0x7b0800000368 by thread T1 (mutexes: write M16):
    #0 operator delete(void*) <null>:1062032 (libclang_rt.tsan_osx_dynamic.dylib:x86_64h+0x4f225)
    #1 std::__1::__shared_ptr_emplace<float, std::__1::allocator<float> >::__on_zero_shared_weak() new:272 (minimal:x86_64+0x1000113de)
    #2 std::__1::weak_ptr<float>::~weak_ptr() memory:5148 (minimal:x86_64+0x100010762)
    #3 std::__1::weak_ptr<float>::~weak_ptr() memory:5146 (minimal:x86_64+0x100002448)
    #4 Foo::Foo()::'lambda'()::operator()() const minimal_race.cpp:15 (minimal:x86_64+0x10000576e)
    #5 void std::__1::__async_func<Foo::Foo()::'lambda'()>::__execute<>(std::__1::__tuple_indices<>) type_traits:4345 (minimal:x86_64+0x1000052f0)
    #6 std::__1::__async_func<Foo::Foo()::'lambda'()>::operator()() future:2323 (minimal:x86_64+0x100005268)
    #7 std::__1::__async_assoc_state<void, std::__1::__async_func<Foo::Foo()::'lambda'()> >::__execute() future:1040 (minimal:x86_64+0x100005119)
    #8 void* std::__1::__thread_proxy<std::__1::tuple<std::__1::unique_ptr<std::__1::__thread_struct, std::__1::default_delete<std::__1::__thread_struct> >, void (std::__1::__async_assoc_state<void, std::__1::__async_func<Foo::Foo()::'lambda'()> >::*)(), std::__1::__async_assoc_state<void, std::__1::__async_func<Foo::Foo()::'lambda'()> >*> >(void*) type_traits:4286 (minimal:x86_64+0x10000717c)

  Previous atomic write of size 8 at 0x7b0800000368 by main thread:
    #0 __tsan_atomic64_fetch_add <null>:1062032 (libclang_rt.tsan_osx_dynamic.dylib:x86_64h+0x24cdd)
    #1 std::__1::shared_ptr<float>::~shared_ptr() memory:3472 (minimal:x86_64+0x1000114d4)
    #2 std::__1::shared_ptr<float>::~shared_ptr() memory:4502 (minimal:x86_64+0x100002488)
    #3 main memory:4639 (minimal:x86_64+0x10000210b)

SUMMARY: ThreadSanitizer: data race new:272 in std::__1::__shared_ptr_emplace<float, std::__1::allocator<float> >::__on_zero_shared_weak()

edit: Eu compilo com:

clang++ -std=c++17 -g -fsanitize=thread -o test  minimal_race.cpp

versão clang:

$ clang++ --version
clang version 7.1.0 (tags/RELEASE_710/final)
Target: x86_64-apple-darwin18.7.0
Thread model: posix
InstalledDir: /usr/local/opt/llvm@7/bin

Estou usando o macOS 10.14.6

Eu acho que isso pode ser um bug com libc ++ e clang em sua implementação fraca_ptr / shared_ptr ...

alterar a fila weak_ptr para um shared_ptr corrige o problema mesmo nas versões antigas do clang.

as mudanças são a linha 25:

  void add(const std::shared_ptr<float> subscriber) {

linha 33:

  std::queue<std::shared_ptr<float>> requests;

e reescrevendo o tempo na linha 42 como:

  while (--numIterations) {
    auto subscriber = std::make_shared<float>();

    foo.add(move(subscriber));
  }