Barrier deadlocks in GCC 11.1.0

Consider the following C++ code:

#include <barrier>
#include <iostream>
#include <string>
#include <thread>

int main() 
{
  constexpr unsigned N = 2;
  std::barrier barrier(N);
  auto const run = [&barrier]() 
  {
    std::string buff;
    for(unsigned i = 0; i < 1000; i++) 
    {
      std::cout << ((buff = "A ") + std::to_string(i) + "\n");
      barrier.arrive_and_wait();
      std::cout << ((buff = "B ") + std::to_string(i) + "\n");
      barrier.arrive_and_wait();
      std::cout << ((buff = "C ") + std::to_string(i) + "\n");
    }
  };

  std::thread ts[N];
  for(unsigned j = 0; j < N; j++) ts[j] = std::thread(run);
  for(unsigned j = 0; j < N; j++) ts[j].join();
}

When I compile/run it using Clang 12.0.1 it behaves as expected. Same expected behavior when I use GCC 11.1 (C++ 20) available online at en.cppreference.com. However, when I use my local GCC 11.1.0, it often deadlocks after printing the following lines:

A 0
A 0
B 0
B 0
C 0
A 1

1. Is this code well-defined?
2. Is it supposed to be deadlock-free?