MicroSpinLock: Use atomic_exchange instead of `atomic_compare_excha…

…nge` (#1971)

Summary:
The lock only has two-states, so `atomic_exchange` works and is as or more performant in any situation.

Pull Request resolved: #1971

Reviewed By: Gownta

Differential Revision: D47343961

Pulled By: Orvid

fbshipit-source-id: 83d56cccc2617bfbd0c27d2464f498ecf1b1ed5e

folly/synchronization/MicroSpinLock.h

@@ -73,15 +73,15 @@ struct MicroSpinLock {
  void init() noexcept { payload()->store(FREE); }
 
  bool try_lock() noexcept {
- bool ret = cas(FREE, LOCKED);
+ bool ret = xchg(LOCKED) == FREE;
  annotate_rwlock_try_acquired(
  this, annotate_rwlock_level::wrlock, ret, __FILE__, __LINE__);
  return ret;
  }
 
  void lock() noexcept {
  detail::Sleeper sleeper;
- while (!cas(FREE, LOCKED)) {
+ while (xchg(LOCKED) != FREE) {
  do {
  sleeper.wait();
  } while (payload()->load(std::memory_order_relaxed) == LOCKED);
@@ -103,13 +103,9 @@ struct MicroSpinLock {
  return reinterpret_cast<std::atomic<uint8_t>*>(&this->lock_);
  }
 
- bool cas(uint8_t compare, uint8_t newVal) noexcept {
- return std::atomic_compare_exchange_strong_explicit(
- payload(),
- &compare,
- newVal,
- std::memory_order_acquire,
- std::memory_order_relaxed);
+ uint8_t xchg(uint8_t newVal) noexcept {
+ return std::atomic_exchange_explicit(
+ payload(), newVal, std::memory_order_acq_rel);
  }
 };
 static_assert(