Mutex is not thread-safe on ARM processors

[jgaskins]

The following code generates a 1M-element array across multiple fibers and then prints its size, so the output should be 1000000:

mutex = Mutex.new
numbers = Array(Int32).new(initial_capacity: 1_000_000)
done = Channel(Nil).new
concurrency = 20
iterations = 1_000_000 // concurrency
concurrency.times do
  spawn do
    iterations.times { mutex.synchronize { numbers << 0 } }
  ensure
    done.send nil
  end
end

# Wait for all fibers to complete
concurrency.times { done.receive }
sleep 100.milliseconds # Wait just a bit longer to be sure the discrepancy isn't due to a *different* race condition
pp numbers.size

On Intel processors on both macOS and Linux, it gives the correct output reliably. On ARM processors on both macOS and Linux, it comes up anywhere from a few dozen to a few thousand short.

If I use Crystal::SpinLock#sync instead of Mutex#synchronize it displays the correct output, so it seems it's something in between Mutex and SpinLock. This is good news because it means that Channel should still be thread-safe on ARM since it uses SpinLock directly.

[yxhuvud]

Just to check, is this with or without -Dpreview_mt ?

[jgaskins]

With.

[straight-shoota]

Probably related to/duplicate of #13055

[jgaskins]

@straight-shoota Yep, even the repro is almost the same. 👍🏼